content
stringlengths 1
15.9M
|
|---|
\section{Introduction}\label{Intro}
There has been a recent revival of interest in \textit{content caching}, particularly focusing on wireless networks. This interest stems from a very practical problem: exponential growth in mobile traffic cannot be matched by the increase in the spectral efficiency of wireless networks. This, in turn, leads to congestion in the radio access as well as the backhaul links, and increased delay and outages for users. \textit{Proactively caching} popular contents at the network edge during off-peak hours has been recently proposed as a potential remedy for this problem (see \cite{AlmerothCacing,GolrezaeiFemtocaching,MaddahAliCentralized,GregoryDtoD,JiArXivNonuniform}, and references therein). Proactive caching shifts traffic from peak to off-peak hours, reduces latency for users, and potentially provides energy savings. In this paper, we focus on the coded caching model proposed in \cite{MaddahAliCentralized}, which considers a single server holding a database of $N$ popular contents of equal size ($F$ bits), serving a set of $K$ users, which have local storage space, sufficient to hold $M$ files, that can be used to proactively cache content during off-peak hours.
Caching in this model consists of two distinct phases: In the first phase, which takes place during off-peak periods, i.e., when the network is not congested, caches at user terminals are filled by the server. This first phase is called the \emph{placement phase}. The only constraint on the data transmitted and stored in a user cache in this phase is the cache capacity. However, due to the ``proactive'' nature of the cache placement, it is carried out without the knowledge of the particular user requests. A shared communication channel is considered to be available from the server to all the users during the peak traffic period. Once the user demands are revealed in this period, the server broadcasts additional information over the common error-free channel in order to satisfy all the user requests simultaneously. This constitutes the \emph{delivery phase}. Since the delivery phase takes place during peak traffic period, the goal is to minimize the rate of transmission over the shared link, called the \textit{delivery rate}, by exploiting the contents that are available the caches.
Over the past decade, research on caching has mainly focused on the placement phase; the goal has been to decide which contents to cache, typically at a server that serves many users, by anticipating future demands based on the history (see \cite{baev2008approximation, BorstCaching, Blasco:ISIT:14}, and references therein). In our model, due to the uniform popularity of the files in the database, this \textit{uncoded} caching approach simply stores an equal portion of each file (i.e., fraction of $M/N$), in each cache. The delivery rate of the uncoded caching scheme, when the user requests are as distinct as possible, i.e., the worst-case rate, is given by
\begin{align}\label{rate_uncoded}
R_{\rm{U}}(M) &= K \cdot \left(1 - \frac{M}{N} \right) \cdot \min\left\{1, \frac{N}{K} \right\}.
\end{align}
The gain from this conventional uncoded caching approach, compared to not having any caches at the users, i.e., the factor $(1-M/N)$ in (\ref{rate_uncoded}), derives mainly from the availability of popular contents locally.
On the other hand, the \emph{coded caching} approach proposed in \cite{MaddahAliCentralized} jointly designs the placement and delivery phases in a \textit{centralized} manner; that is, a central server, which knows the number and identity of the users in the system, can arrange the placement and delivery phases in coordination across the users. This coded caching approach, thanks to the presence of a common broadcast channel to all the users, can significantly reduce the backhaul traffic over conventional uncoded caching by creating multicasting opportunities, even when users request different files. The following delivery rate is achieved by the centralized coded caching scheme proposed in~\cite{MaddahAliCentralized}, referred to as the MAN scheme in the rest of the paper:
\begin{align}\label{rate_coded_MAN}
R^{\rm{C}}_{\rm{MAN}}(M) &= K \cdot (1 - M/N) \cdot \min\left\{\frac{1}{1+KM/N}, \frac{N}{K}\right\}.
\end{align}
Following the seminal work of \cite{MaddahAliCentralized}, further research on centralized coded caching have been carried out recently to reduce the required delivery rate. Authors in \cite{ZhiChenXOR} have proposed a coded placement phase, referred to as the CFL scheme in this paper. When the number of users, $K$, is at least as large as the number of files in the system, i.e., $K \ge N$, the following delivery rate can be achieved for a cache capacity of $M=1/K$ by the CFL scheme:
\begin{align}\label{rate_coded_CFL}
R^{\rm{C}}_{\rm{CFL}} \left( \frac{1}{K} \right) &= N \left( 1 - \frac{1}{K} \right).
\end{align}
It is also shown in \cite{ZhiChenXOR} that the proposed scheme is indeed optimal for small cache capacities; that is, when $M \leq 1/K$.
By employing coded multicasting opportunities across users with the same demand, an improved centralized caching scheme for any cache capacity $0 \le M \le N$ is presented in \cite{KaiWanUncodedCaching} for the special case $K > N=2$. Again, considering more users in the system than the files in the database, i.e., $K> N$, the authors in \cite{MohammadDenizTCom} have further reduced the delivery rate by exploiting a novel coded placement scheme, when $4 \le N < K \le 3N/2$, and $N$ and $K$ are not relatively prime.
A theoretical lower bound on the delivery rate allows us to quantify how far the proposed caching schemes perform compared to the optimal performance. In addition to the cut-set bound studied in \cite{MaddahAliCentralized}, a tighter lower bound is derived in \cite{SenguptaCaching}. For the special case of $N=K=3$, a lower bound is derived through a novel computational approach in \cite{TianCaching}. A labeling problem in a directed tree, the complexity of which grows exponentially with the number of users, is considered in \cite{GhasemiCachingLowerBound}, to characterize another lower bound. However, none of these bounds are tight in general, and the optimal delivery rate-cache capacity trade-off for a caching system remains an open problem.
In contrast to the above centralized setting, in many practical scenarios, the identity and number of active users in the delivery phase are not known in advance, and there may not be a central authority to coordinate the cache contents across users. Maddah-Ali and Niesen considered the corresponding \textit{decentralized} coded caching problem, and showed that coded caching can still help reduce the delivery rate \cite{MaddahAliDecentralized}. In the decentralized scenario, while the multicasting opportunities cannot be designed and maximized in advance, they will still appear even if the bits of files are randomly cached in the placement phase. The following delivery rate is achievable by the proposed decentralized caching scheme in \cite{MaddahAliDecentralized}:
\begin{align}\label{rate_dec_MAN}
R^{\rm{D}}_{\rm{MAN}}(M) &= K \cdot \left(1 - \frac{M}{N} \right) \cdot \min\left\{\frac{N}{KM}\big(1-(1-M/N)^K\big), \frac{N}{K}\right\}.
\end{align}
We note that the delivery rate in (\ref{rate_dec_MAN}) provides the same local caching gain as in the centralized setting, i.e., the factor $(1-M/N)$, as well as an additional global gain thanks to coding. The centralized caching schemes proposed in \cite{KaiWanUncodedCaching} is also extended to the decentralized setting, and it is shown to reduce the required delivery rate.
Coded caching have been studied under many other network settings in the recent years, such as online coded caching \cite{PedarsaniOnlineCaching}, multi-layer coded caching \cite{KaramchandaniHierarchical}, caching files with nonuniform distributions \cite{NiesenNonuniform, JiArXivNonuniform} and distinct sizes \cite{ZhangDistinctFileSizes}, user with distinct cache capacities \cite{WangHeterogenous}. It has also been extended to lossy reconstructions of cached files in \cite{QianDenizLossy, ElzaDistortionMemoryTradeoff, TimoDistortionCaching}. Several works have considered noisy channels rather than an error-free shared link in the delivery phase \cite{MaddahAliInterferenceJournal,NaderializadehMaddahAliInterference, TimoErasureChannel}, as well as delivery over a fading channel \cite{HuangFadingChannelcodedcaching}.
In this paper, we study coded caching in both the centralized and decentralized settings. We propose a novel \textit{group-based caching scheme}, and show that it improves the caching gain in both scenarios, particularly when the number of users in the system is larger than the number of files, i.e., when $K > N$. Note that, this setting may be applicable for the distribution of extremely popular files that become viral over the Internet, and are requested by a large number of users over a relatively short period of time, or for the distribution of various software updates to users.
Our main contributions can be summarized as below:
\begin{enumerate}[label=\arabic*)]
\item In the centralized setting, we propose a novel \textit{group-based centralized} (GBC) coded caching scheme for a cache size of $M = N/K$ at the users. It is shown that the GBC scheme achieves a lower delivery rate compared to the state-of-the-art results when $K>N\geq3$. This improvement can be further extended to other cache sizes through memory-sharing with the schemes proposed in \cite{MaddahAliCentralized} and \cite{ZhiChenXOR}.
\item By employing the same group-based caching idea in the decentralized setting, we introduce the \textit{group-based decentralized caching} (GBD) scheme. The GBD scheme is shown to achieve a smaller delivery rate compared to the schemes presented in \cite{MaddahAliDecentralized} and \cite{KaiWanUncodedCaching} for $K>N$.
\item We provide numerical results validating the superiority of the proposed group-based coded caching schemes compared to the state-of-the-art in both the centralized and decentralized settings.
\end{enumerate}
The rest of the paper is organized as follows. We present
the system model and relevant previous results in Section \ref{SystemModel}. The group-based centralized coded caching scheme is introduced in Section \ref{CentralizedScheme} for the centralized setting. In Section \ref{DecentralizedScheme} it is extended to the decentralized scenario. In both Sections \ref{CentralizedScheme} and \ref{DecentralizedScheme}, the derivation of the corresponding delivery rates are complemented with the analytical and numerical comparison of the achieved delivery rates with the state-of-the-art results. The proofs of our main results can be found in Appendix. We conclude the paper in Section \ref{Conclusion}.
\textit{Notations:} The sets of integers, real numbers, and positive real numbers are denoted by $\mathcal Z$, $\mathcal R$, and $\mathcal R^+$, respectively. For two integers $i \le j$, $\left[ {i:j} \right]$ represents the set of integers $\left\{ {i, i + 1, \ldots , j} \right\}$. If $i>j$, then $\left[ {i:j} \right] \buildrel \Delta \over = \left\{ \emptyset \right\}$. The binomial coefficient ``$n$ choose $k$'' is denoted by $\binom{n}{k}$. Notation $ \oplus $ refers to the bitwise XOR operation. We use $\left| \cdot \right|$ to indicate the length of a binary sequence, or the cardinality of a set.
\section{System Model and Previous Results}\label{SystemModel}
Consider a server which has $N$ popular files, denoted by $W_1, W_2, ..., W_N$, each of length $F$ bits, in its database. The files are assumed to be independent of each other and distributed uniformly over the set $\left[ 1:2^F \right]$. There are $K$ users in the system, denoted by $U_1, U_2, ..., U_K$, and each user is equipped with a local cache of capacity $MF$ bits.
The caching system operates in two distinct phases. In the initial \textit{placement phase}, cache of each user is filled by the server. The contents of the cache of user $U_k$ at the end of the placement phase is denoted by $Z_k$, for $k=1,...,K$. The user demands are revealed after the placement phase. Each user requests a single file from among the available files in the database, and the demand of user $U_k$ is denoted by $d_k$, where ${d_k} \in \left[ 1:N \right]$, $\forall k$. In the \textit{delivery phase} that follows, a common message $X$ is transmitted by the server to the users over an error-free shared link to satisfy all the user requests simultaneously.
\textbf{Definition.} An $(M,R,F)$ \textit{caching and delivery code} consists of
\begin{enumerate}[label=\roman*.]
\item $K$ caching functions:
\begin{equation}\label{S1} {\phi _k}:{\left[ {1:{2^F}} \right]^N} \to \left[ {1:{2^{\left\lfloor {FM} \right\rfloor }}} \right], \quad k=1, ..., K,
\end{equation}
each of which maps the database $\left\{ W_1, W_2, ..., W_N \right\}$ to the cache content $Z_k$ of user $U_k$, i.e., $Z_k = \phi _k \left( W_1, W_2, ..., W_N \right)$;
\item one coded delivery function:
\begin{equation}\label{S2} \psi :{\left[ {1:{2^F}} \right]^N} \times {\left[ 1:N \right]^K} \to \left[ {1:{2^{\left\lfloor {FR} \right\rfloor }}} \right],
\end{equation}
which generates the common message $X$ to be delivered through the shared link as a function of the database $\left\{ W_1, W_2, ..., W_N \right\}$ and the user demands $\left\{ d_1, d_2, ..., d_K \right\}$, i.e., $X=\psi \left(W_1, W_2, ..., W_N, d_1, d_2, ..., d_K\right)$;
\item $K$ decoding functions:
\begin{equation}\label{S3} {\mu _k}:\left[ {1:{2^{\left\lfloor {FM} \right\rfloor }}} \right] \times \left[ {1:{2^{\left\lfloor {FR} \right\rfloor }}} \right] \times \left[ 1:N \right]^K \to \left[ {1:{2^F}} \right], \quad k=1, ..., K,
\end{equation}
each of which reconstructs file ${\hat W_{{d_k}}}$ at user $k$ as a function of the cache content $Z_k$, the common message $X$ delivered over the shared link, and the user demands $\left\{ d_1, d_2, ..., d_K \right\}$, i.e., ${\hat W_{{d_k}}} = {\mu _k}\left(Z_k, X, d_1, d_2, ..., d_K\right)$.
\end{enumerate}
In the above definition, $M$ represents the normalized cache capacity of each user (normalized by $F$), and $R$ denotes the transmission rate during the delivery phase.
\textbf{Definition.} The probability of error of an $(M,R,F)$ caching and delivery code is defined as
\begin{equation}\label{S4} {P_e} \buildrel \Delta \over = \mathop {\max }\limits_{\left( {{d_1},{d_2},...,{d_K}} \right)} \Pr \left\{ {\mathop \bigcup \limits_{k = 1}^K \left\{ {{{\hat W}_{d_k}} \ne {W_{{d_k}}}} \right\}} \right\}.
\end{equation}
\textbf{Definition.} The delivery rate-cache capacity pair $(M,R)$ is \textit{achievable} if for every $\varepsilon > 0$ an $(M,R,F)$ caching and delivery code can be found with significantly large $F$, that has a probability of error less than $\varepsilon$, i.e., $P_e < \varepsilon$.
\begin{figure}[!t]
\centering
\includegraphics[scale=0.7]{System_Model.pdf}
\caption{Illustration of a caching system including a server holding $N$ popular files, each $F$ bits long, in its database, serving contents to $K$ users, each equipped with a local cache of capacity $MF$ bits, through an error-free shared link.}
\label{System_Model}
\end{figure}
Naturally, there is a trade-off between the cache capacity of the users, and the required minimum delivery rate. Our goal is to define and characterize this trade-off rigorously. The delivery rate-cache capacity trade-off for the caching network depicted in Fig. \ref{System_Model}, with the caches of equal capacity $M$, is denoted by $R^{\rm{C}}(M)$, and defined as
\begin{equation}\label{S5} {R^{\rm{C}}}\left( M \right) \buildrel \Delta \over = \inf \left\{ {R:\left( {M,R} \right) \mbox{ is achievable} } \right\}.
\end{equation}
For example, when there is no cache available at the users, i.e., $M=0$, the worst case of user demands corresponds to users requesting as distinct files as possible; and all the requested files should be delivered by the server over the shared link; we have $R^{\rm{C}}(0)=\min \left\{ {N,K} \right\}$. On the other hand, when the cache capacity is large enough to store all the files, i.e., when $M=N$, all the files can be made available locally to all the users, and no content needs to be sent during the delivery phase, and $R^{\rm{C}}(N)=0$. Our goal is to characterize the delivery rate-cache capacity trade-off for all cache capacities $0 < M < N$ for any given number of users $K$.
For a centralized caching system, we denote the best achievable delivery rate-cache capacity trade-off in the literature by $R^{\rm{C}}_b(M)$. For $N > K$, the best known delivery rate is achieved by the scheme proposed in \cite{MaddahAliCentralized}, for any cache capacity satisfying $0 \le M \le N$, i.e., we have $R^{\rm{C}}_b(M) = R^{\rm{C}}_{\rm{MAN}}(M)$. On the other hand, for $N \le K$, we define the following set of $\left( N,K \right)$ corresponding to the scenario presented in \cite{MohammadDenizTCom} to find the best achievable scheme in the literature:
\begin{equation}\label{PaticularSetOurs} \zeta \buildrel \Delta \over = \left\{ {\left( {N,K} \right):4 \le N < K \le \frac{{3N}}{2},\mbox{$N$ and $K$ have a common divisor $c>1$}} \right\}.
\end{equation}
When $N \le K$, two following cases are considered to characterize $R_b^{\rm{C}}(M)$.
\begin{enumerate}[label=\bfseries Case \arabic*:,align=left]
\item $\left( {N,K} \right) \in \zeta$. The best delivery rate in the literature for Case 1 can be characterized as follows. For cache capacity $M = 1/K$, the CFL scheme achieves the optimal delivery rate \cite{ZhiChenXOR}, while for $M = (N-1)/K$, the coded caching scheme presented in \cite{MohammadDenizTCom}, which will be referred to as AG, achieves the best known delivery rate. For cache capacities $M=tN/K$, the MAN scheme should be utilized, where if $K < 3N/2$, then $t \in \left[ 1:K \right]$, and if $K = 3N/2$, then $t \in \left[ 2:K \right]$. For any other cache capacities $0 \le M \le N$, the lower convex envelope of the mentioned points can be achieved through memory-sharing.
\item $N \le K$ and $\left( {N,K} \right) \notin \zeta$. The best delivery rate in the literature is achieved by memory-sharing between the MAN and CFL schemes. In this case, the CFL scheme again achieves the optimal delivery rate for a cache capacity of $M = 1/K$. To find the best achievable scheme for $M = N/K$, we define, for $t \in \left[ 1:K \right]$,
\begin{align}\label{fNKt} f(N,K,t) \triangleq \frac{{\left( {N - 1} \right)\left( {K - t} \right)}}{{\left( {t + 1} \right)\left( {tN - 1} \right)}} + {N^2}\left( {1 - \frac{1}{K}} \right)\left( {\frac{{t - 1}}{{tN - 1}}} \right),
\end{align}
and
\begin{equation}\label{optimumt} {t^*} \buildrel \Delta \over = \mathop {\arg \min }\limits_{t \in \left[ {1:K} \right]} f(N,K,t).
\end{equation}
Then, we have
\begin{equation}\label{BestDeliveryRatePoint} R_b^{\rm{C}}\left(N/K \right) = f(N,K,t^*).
\end{equation}
For cache capacities $M = lN/K$, where $l \in \left[ t^*:K \right]$, the MAN scheme should be employed. The best delivery rate for other cache capacities is the lowest convex envelope of the points $M = l/K$ and $M = lN/K$, for $l \in \left[ t^*:K \right]$, which can be achieved by memory-sharing.
\end{enumerate}
\section{Centralized Coded Caching}\label{CentralizedScheme}
In this section, the placement and delivery phases of the proposed centralized coded caching scheme will be introduced. The delivery rate achieved by this scheme will be analyzed and compared with the existing schemes in the literature. We first illustrate the proposed caching scheme on an example.
\begin{figure}[!t]
\centering
\includegraphics[scale=0.7]{Cache_Content.pdf}
\caption{Cache contents at the end of the placement phase for the proposed GBC scheme when each of the $K=10$ users demand a single file from among $N=3$ files in the database. The worst-case user demand combination is assumed, and the users are grouped into three groups according to their requests.}
\label{CacheContent}
\end{figure}
\theoremstyle{definition}
\newtheorem{exmp}{Example}
\begin{exmp}\label{ExampleCentralized}
In this example, we consider $K = 10$ users, $N = 3$ files, and a cache capacity of $M = 3/10$. Each file $W_i$ is first divided into $10$ non-overlapping subfiles $W_{i,j}$, for $j \in \left[ 1:10 \right]$, and $i \in \left[ 1:3 \right]$, each of length $F/10$ bits. Then one subfile from each file is placed into each user's cache; that is, we have $Z_j = \left( {{W_{1,j}},{W_{2,j}},{W_{3,j}}} \right)$, for $j=1, \ldots, 10$. See Fig. \ref{CacheContent} for the cache contents at the end of the placement phase.
The worst-case of user demands is when the requested files are as distinct as possible. Without loss of generality, by re-ordering the users, we consider the following user demands:
\begin{equation}\label{DemandsExample}
{d_j} =
\begin{cases}
{{1},\quad 1 \le j \le 4},\\
{{2},\quad 5 \le j \le 7},\\
{{3},\quad 8 \le j \le 10}.
\end{cases}
\end{equation}
In the delivery phase, the users are grouped according to their demands. Users that request file $W_i$ from the server constitute group $G_i$, for $i \in \left[ 1:3 \right]$. See Fig. \ref{CacheContent} for the group formation corresponding to the demands in \eqref{DemandsExample}. The delivery phase is divided into two distinct parts, that are designed based on the group structure.
\begin{enumerate}[label=\bfseries Part \arabic*:,align=left]
\item The first part of the delivery phase is designed to enable each user to retrieve all the subfiles of its demand that have been placed in the caches of users in the same group. As an example, all users $U_1$, ..., $U_4$, i.e., members of group $G_1$, should be able to decode all the subfiles $W_{1,1}$, $W_{1,2}$, $W_{1,3}$, $W_{1,4}$, i.e., the subfiles stored in the cache of users in $G_1$, after receiving the message transmitted in part 1. Accordingly, in our example, in part 1 of the delivery phase the server sends the following coded subfiles over the shared link. ${W_{1,1}} \oplus {W_{1,2}}$, ${W_{1,2}} \oplus {W_{1,3}}$, ${W_{1,3}} \oplus {W_{1,4}}$, ${W_{2,5}} \oplus {W_{2,6}}$, ${W_{2,6}} \oplus {W_{2,7}}$, ${W_{3,8}} \oplus {W_{3,9}}$, ${W_{3,9}} \oplus {W_{3,10}}$.
\item The purpose of part 2 is to make sure that, each user can retrieve all the subfiles of its desired file, which have been placed in the cache of users in other groups. Hence, the server transmits the following coded subfiles in the second part of the delivery phase. ${W_{1,5}} \oplus {W_{1,6}}$, ${W_{1,6}} \oplus {W_{1,7}}$, ${W_{2,1}} \oplus {W_{2,2}}$, ${W_{2,2}} \oplus {W_{2,3}}$, ${W_{2,3}} \oplus {W_{2,4}}$, ${W_{1,7}} \oplus {W_{2,4}}$, ${W_{1,8}} \oplus {W_{1,9}}$, ${W_{1,9}} \oplus {W_{1,10}}$, ${W_{3,1}} \oplus {W_{3,2}}$, ${W_{3,2}} \oplus {W_{3,3}}$, ${W_{3,3}} \oplus {W_{3,4}}$, ${W_{1,10}} \oplus {W_{3,4}}$, ${W_{2,8}} \oplus {W_{2,9}}$, ${W_{2,9}} \oplus {W_{2,10}}$, ${W_{3,5}} \oplus {W_{3,6}}$, ${W_{3,6}} \oplus {W_{3,7}}$, ${W_{3,7}} \oplus {W_{2,10}}$.
\end{enumerate}
It can be easily verified that, together with the contents placed locally, user $U_k$ can decode its requested file $W_{d_k}$, $\forall k \in \left[ 1:10 \right]$, from the message transmitted in the delivery phase. As a result, by delivering a total of $12F/5$ bits, which corresponds to a delivery rate of $2.4$, all user demands are satisfied. The delivery rate of the best achievable scheme in the literature for cache capacity of $M = 3/10$ can be evaluated from \eqref{BestDeliveryRatePoint}, and it is given by $R_b^{\rm{C}}\left( 3/10 \right)=2.43$.
\qed
\end{exmp}
\subsection{Group-Based Centralized Coded Caching (GBC)}\label{GBCscheme}
Here we generalize the ideas introduced above in the example, and introduce the group-based coded caching (GBC) scheme for any $N$ and $K$ values. We consider a cache capacity of $M = N/K$, i.e., the aggregate size of the cache memories distributed across the network is equivalent to the total size of the database.
Note that, the server has no information about the user demands during the placement phase. Therefore, to satisfy all demand combinations efficiently, and to reduce the dependency of the delivery rate on the particular user demands as much as possible, we retain a symmetry among the subfiles of each file cached at each user. We employ the same placement phase proposed in \cite{MaddahAliCentralized} for $M = N/K$, in which each file $W_i$ is divided into $K$ non-overlapping subfiles $W_{i,j}$, for $i \in \left[ 1:N \right]$ and $j \in \left[ 1:K \right]$, each of the same size $F/K$ bits, and the cache contents of user $U_j$ is given by $Z_j = \left( {{W_{1,j}},{W_{2,j}}, \ldots ,{W_{N,j}}} \right)$, $\forall j \in \left[ 1:K \right]$. It is easy to see that the cache capacity constraint is satisfied.
Without loss of generality, by re-ordering the users, it is assumed that the first $K_1$ users, referred to as group $G_1$, have the same request $W_1$, the next $K_2$ users, i.e., group $G_2$, request the same file $W_2$, and so on so forth. Hence, there are $K_i \ge 0$ users in each group $G_i$, each with the same request $W_i$, for $i \in \left[ 1:N \right]$. We will see that the delivery rate for the proposed GBC scheme is independent of the values $K_1, K_2, ..., K_N$. An example of grouping users is illustrated in Fig. \ref{CacheContent}, where $K_1=4$, and $K_2=K_3=3$. We define the following variable which will help simplify some of the expressions:
\begin{align}\label{S_sumK}
S_i \buildrel \Delta \over = \sum\limits_{l = 1}^i {{K_l}}.
\end{align}
The coded delivery phase of the GBC scheme, for the user demands described above, is presented in Algorithm \ref{DeliveryCentralized}.
There are two parts in the delivery phase just as in Example \ref{ExampleCentralized}. Having received the contents delivered in the first part of Algorithm \ref{DeliveryCentralized}, each user can obtain the missing subfiles of its requested file, which are in the cache of users in the same group; that is, each user $U_k$ in group $G_i$ requesting file $W_i$ has access to subfile $W_{i,k}$, and can decode all subfiles $W_{i,l}$, $\forall l \in \left\{ S_{i-1}+1:S_i \right\}$, after receiving the contents delivered in line 3 of Algorithm \ref{DeliveryCentralized}, for $i=1, ..., N$ and $k = S_{i-1}+1, ..., S_i$. With the contents delivered in the second part, each user can decode the subfiles of its requested file which are in the cache of users in other groups. Note that, all users in group $G_i$ demanding file $W_i$ have decoded subfile $W_{j,S_i}$, and they can obtain all subfiles $W_{i,l}$, $\forall l \in \left\{ S_{j-1}+1:S_j \right\}$, i.e., subfiles of file $W_i$ having been cached by users in group $G_j$, after receiving $\mathop \bigcup \limits_{k = S_{j - 1} + 1}^{S_j - 1} \left( {{W_{i,k}} \oplus {W_{i,k + 1}}} \right)$ and $W_{i, S_j} \oplus W_{j, S_i}$, for $i=1, ..., N-1$ and $j=i+1, ..., N$. Similarly, all users in group $G_j$ can decode all subfiles of their requested file $W_j$, which are in the cache of users in group $G_i$ by receiving $\mathop \bigcup \limits_{k = S_{i - 1} + 1}^{S_i - 1} \left( {W_{j,k}\oplus W_{j,k+1}} \right)$ and $W_{i, S_j} \oplus W_{j, S_i}$, for $i=1, ..., N-1$ and $j=i+1, ..., N$. In this way, at the end of the proposed delivery phase, the users can recover all the bits of their requested files.
\begin{algorithm}
\caption{Coded Delivery Phase of the GBC Scheme}
\label{DeliveryCentralized}
\begin{algorithmic}[1]
\Statex
\State{\textbf{Part 1}: Exchanging contents between users in the same group}{}
\For {$i = 1, \ldots, N$}
\State {server delivers $\left( \mathop \bigcup \limits_{k = S_{i - 1} + 1}^{S_{i}-1} \left( {{W_{i,k}} \oplus {W_{i,k + 1}}} \right) \right)$.}
\EndFor
\Statex
\State{\textbf{Part 2}: Exchanging contents between users in different groups}{}
\For {$i = 1, \ldots, N-1$}
\For {$j = i+1, \ldots, N$}
\State {server delivers $\left( \mathop \bigcup \limits_{k = S_{j - 1} + 1}^{S_j - 1} \left( {{W_{i,k}} \oplus {W_{i,k + 1}}} \right), \mathop \bigcup \limits_{k = S_{i - 1} + 1}^{S_i - 1} \left( {W_{j,k}\oplus W_{j,k+1}} \right), W_{i, S_j} \oplus W_{j, S_i} \right)$.}
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
\subsection{Delivery Rate Analysis}\label{AnalysisGBC}
Note that, the delivery rate of the GBC scheme, $R^{\mathrm{C}}_{\rm{GBC}}$, should be evaluated for the worst-case user demands. It can be argued that when $N < K$, the worst-case user demands is when there is at least one user requesting each file, i.e., $K_i > 0$, $\forall i \in \left[ 1:N \right]$. On the other hand, when $N \ge K$, without loss of generality, by re-ordering the users, the worst-case user demands is assumed to happen when $K_i = 1$, if $i \in \left[ 1:K \right]$; and $K_i = 0$, otherwise.
When $N \ge K$, considering the worst-case user demands, the server transmits the contents $\left( \bigcup \limits_{i = 1}^{K - 1} \bigcup \limits_{j = i + 1}^K \left( W_{i,{S_j}} \oplus W_{j,{S_i}} \right) \right)$ using Algorithm \ref{DeliveryCentralized}, which are similar to the contents delivered using the delivery phase proposed in \cite[Algorithm 1]{MaddahAliCentralized} for a cache capacity of $M = N/K$. Thus, when $N \ge K$, the GBC scheme achieves the same delivery rate as the MAN scheme for $M = N/K$, i.e., $R_{\rm{GBC}}^{\rm{C}}(N/K) = R^{\rm{C}}_{\rm{MAN}}(N/K)$.
For $K>N$, the delivery rate of the GBC scheme is stated in the next theorem, whose proof can be found in Appendix \ref{ProofFirstTheorem}. We then compare the result with the best achievable scheme in the literature.
\begin{theorem}\label{TheoremGBC}
In a centralized coded caching system with $N$ files, each of size $F$ bits, $K$ users, each equipped with a cache of capacity $MF$ bits, where $M = N/K$, the following delivery rate is achievable by the proposed GBC scheme, if $N < K$:
\begin{equation}\label{OurDeliveryRatePointTheorem} {R_{\mathrm{GBC}}^{\rm{C}}}\left( {\frac{N}{K}} \right) = N - \frac{{N\left( {N + 1} \right)}}{{2K}}.
\end{equation}
\end{theorem}
\begin{remark}
It can be seen that the delivery rate of the GBC scheme depends only on $N$ and $K$, and is independent of the values of $K_1,..., K_{N}$, which implies the popularity of the files. The more distinct the files requested by the users, the higher the required delivery rate. Accordingly, the delivery rate given in \eqref{OurDeliveryRatePointTheorem} is obtained for the worst-case user demand combination, such that each file is requested by at least one user.
\end{remark}
\subsection{Comparison with the State-of-the-Art}
In this subsection, we compare the performance of the GBC scheme with other caching schemes in the literature both analytically and numerically. We first show that the delivery rate achieved for a cache capacity of $M = N/K$ in \eqref{OurDeliveryRatePointTheorem} is the best known result as long as $N < K$. We then extend this improvement to a wider range of cache capacities through memory-sharing.
\begin{corollary}
The achievable delivery rate of the GBC scheme for a cache capacity of $M = N/K$ given in \eqref{OurDeliveryRatePointTheorem} improves upon the best known delivery rate in the literature, $R^{\rm{C}}_b(N/K)$, when $K > N \ge 3$.
\end{corollary}
\begin{proof}
We compare the delivery rate of the GBC scheme with the best achievable scheme in the literature described in Section \ref{SystemModel} for both cases $\left( {N,K} \right) \in \zeta$ and $\left( {N,K} \right) \notin \zeta$. For the first case, the slope of the delivery rate for $1/K \le M \le (N-1)/K$, which is achieved by memory-sharing between the CFL and AG schemes for points $M = 1/K$ and $M = (N-1)/K$, respectively, is $\frac{(K - 2)(K - 2N)}{2(N - 2)}$. On the other hand, the slope of the delivery rate for $1/K \le M \le N/K$, achieved through memory-sharing between the CFL and GBC schemes for points $M = 1/K$ and $M = N/K$, respectively, is $- N/2$. Since for $\left( {N,K} \right) \in \zeta$, we have $N \ge 4$, and $N+2 \le K \le 3N/2$, it can be easily verified that the latter lies below the former; that is, by exploiting the GBC scheme rather than the AG scheme, a lower delivery rate can be achieved at $M = N/K$.
For a comparison of $R^{\rm{C}}_{\mathrm{GBC}}\left( N/K \right)$ and $R_b^{\rm{C}}(N/K)$ in the second case, we need to evaluate $R_b^{\rm{C}}\left( N/K \right)$ given in \eqref{BestDeliveryRatePoint}, which is not straightforward. Instead, in Appendix \ref{CentralizedImprovement}, we prove that for $K > N \ge 3$, ${R^{\rm{C}}_{\mathrm{GBC}}}\left( N/K \right) \le f\left( N,K,t \right)$, $\forall t \in \left\{ 1:K \right\}$. Since $R_b^{\rm{C}}\left( N/K \right) = f\left( N,K,t^* \right)$, where $t^* \in \left\{ 1:K \right\}$, it can be concluded that ${R^{\rm{C}}_{\mathrm{GBC}}}\left( N/K \right) \le {R^{\rm{C}}_b}\left( N/K \right)$. This completes the proof.
\end{proof}
\begin{remark}
We observe through simulations that, apart from the case $N=2$, the centralized scheme proposed in \cite{KaiWanUncodedCaching}, referred to as WTP, does not improve upon time-sharing between the MAN and CFL schemes. Nevertheless, below we explicitly demonstrate the superiority of the GBC scheme over WTP for a cache capacity of $M = N/K$, when $K > N \ge 2$. Let $R^{\rm{C}}_{\mathrm{WTP}}\left( M \right)$ denote the delivery rate-cache capacity trade-off of the WTP scheme in the centralized setting. According to \cite[Theorem 1]{KaiWanUncodedCaching}, we have
\begin{equation}\label{WTPDeliveryRate} R_{\mathrm{WTP}}^{\rm{C}}\left(\frac{N}{K} \right) \ge \min \left( {R_{{\mathrm{co}}_1}\left( {\frac{N}{K}} \right),R_{{\mathrm{co}}_2}\left( {\frac{N}{K}} \right)} \right),
\end{equation}
where
\begin{subequations}
\label{WTPDeliveryRateco}
\begin{align}
{R_{{\rm{c}}{{\rm{o}}_1}}}\left( M \right) &= N - M - \frac{{M\left( {N - 1} \right)K\left( {N - M} \right)}}{{{N^2}\left( {K - 1} \right)}},\\
{R_{{\rm{c}}{{\rm{o}}_2}}}\left( M \right) &= \frac{{K\left( {N - M} \right)}}{{N + KM}}.
\end{align}
\end{subequations}
When $K > N \ge 2$, it can be easily verified that
\begin{subequations}
\label{CompareOursWTPDeliveryRateco}
\begin{align}
{R^{\rm{C}}_{{\rm{GBC}}}}\left( {\frac{N}{K}} \right) &\le {R_{{\rm{c}}{{\rm{o}}_1}}}\left( {\frac{N}{K}} \right) = N + \frac{1}{K} - \frac{{2N}}{K},\\
{R^{\rm{C}}_{{\rm{GBC}}}}\left( {\frac{N}{K}} \right) &\le {R_{{\rm{c}}{{\rm{o}}_2}}}\left( {\frac{N}{K}} \right) = \frac{{K - 1}}{2},
\end{align}
\end{subequations}
which concludes that
\begin{equation}\label{CompareOursWTPDeliveryRate} {R^{\rm{C}}_{{\rm{GBC}}}}\left( {\frac{N}{K}} \right) \le {R^{\rm{C}}_{{\rm{WTP}}}}\left( {\frac{N}{K}} \right).
\end{equation}
\end{remark}
The improvement obtained by using the GBC scheme for $M = N/K$, when $N < K$, can be extended to any cache capacities $1/K < M < \hat tN/K$, where $\hat t \buildrel \Delta \over = \max \left\{ {2,{t^*}} \right\}$, through memory-sharing between the CFL scheme for $M = 1/K$, the GBC scheme for $M = N/K$, and the MAN scheme for cache capacity $M = \hat tN/K$. Note that, when $t^* = 1$, the improvement can be extended to the interval $1/K < M < 2N/K$, while for $t^* \ge 2$, the superiority of the GBC scheme holds for cache capacities $1/K < M < t^*N/K$.
\begin{corollary}
For $N$ files and $K$ users, each equipped with a cache of normalized capacity $M$, the following delivery rate-cache capacity trade-off is achievable in the centralized setting if $K>N$:
\begin{align}\label{OurDeliveryRateIntervalCorollary}
R^{\rm{C}}_{\mathrm{GBC}}\left( M \right) =
\begin{cases}
N \left( 1 - \frac{M}{2} - \frac{1}{2K} \right), \qquad \qquad \qquad \quad \; \mbox{ if }1/K \le M \le N/K,\\
\frac{K-\hat{t}}{\hat{t}^2-1}\left(\frac{KM}{N} -1\right) + \frac{K-N}{\hat{t}-1} \left( \hat{t}N/K-M \right), \mbox{ if } N/K \le M \le \hat{t}N/K,
\end{cases}
\end{align}
where $\hat t \buildrel \Delta \over = \max \left\{ {2,{t^*}} \right\}$, and $t^*$ is determined as in \eqref{optimumt}.
\end{corollary}
\begin{figure}[!t]
\centering
\includegraphics[scale=0.7]{N_10_K_15.pdf}
\caption{Delivery rate cache capacity trade-off of the proposed GBC scheme compared with the existing schemes in the literature for $N=10$ and $K=15$. This setting corresponds to Case 1 in Section \ref{SystemModel}. Since $K = 3N/2$, for $1/K \le M \le 2N/K$, $R_{\rm{b}}^{\rm{C}}$ is achieved by memory-sharing between CFL, AG, and MAN schemes.}
\label{N10K15}
\end{figure}
Next, we illustrate the improvements in the delivery rate offered by the proposed GBC scheme through numerical simulations. The delivery rate of known centralized caching schemes as a function of the cache capacity, $M$, are compared in Fig. \ref{N10K15}, for $N=10$ files and $K=15$ users. Note that, these $N$ and $K$ values correspond to Case 1 for the best achievable scheme in the literature, i.e., $\left( {N,K} \right) \in \zeta$. The best known achievable delivery rate in the literature for Case 1 is achieved by memory-sharing between the CFL, AG, and MAN schemes. Since we have $K = 3N/2$, the schemes are compared over the range $1/K \le M \le 2N/K$, where GBC provides an improvement. We also include the centralized WTP scheme, whose performance is slightly worse than that achieved by memory-sharing between the CFL, AG, and MAN schemes in this scenario. The information theory and cut-set lower bounds investigated in \cite[Theorem 1]{SenguptaCaching} and \cite[Theorem 2]{MaddahAliCentralized}, respectively, are also included in this figure for comparison. We observe a significant improvement in the delivery rate offered by the GBC scheme compared to all other schemes in the literature for all cache capacity values $1/K < M < 2N/K$. Despite the improvement upon other known schemes, there is still a gap between $R_{\mathrm{GBC}}$ and the theoretical lower bound; although this gap might as well be due to the looseness of the known lower bounds.
In Fig. \ref{N50K130}, we compare the delivery rate of the GBC scheme with the existing schemes for $N=50$ and $K=130$. This scenario corresponds to Case 2 of the best achievable schemes in the literature, i.e., $\left( {N,K} \right) \notin \zeta$, whose delivery rate is achieved through memory-sharing between the CFL and MAN schemes. According to \eqref{optimumt}, we have $t^*=4$, and hence, $\hat t = t^* = 4$. The centralized WTP scheme achieves the same performance as memory-sharing between the CFL and MAN schemes in this scenario. We again observe that the GBC scheme achieves a significantly lower delivery rate compared to the known caching schemes over the range of cache capacities $1/K \le M \le 4N/K$.
\begin{figure}[!t]
\centering
\includegraphics[scale=0.7]{N_50_K_130.pdf}
\caption{Delivery rate cache capacity trade-off of the proposed GBC scheme compared with the best achievable scheme in the literature for $N = 50$ and $K = 130$. This setting corresponds to Case 2 in Section \ref{SystemModel}. $R_{\rm{b}}^{\rm{C}}$ is achieved by memory-sharing between the CFL and MAN schemes for $1/K \le M \le 4N/K$.}
\label{N50K130}
\end{figure}
In Fig. \ref{NConstantKVariable}, the delivery rate of the proposed GBC scheme for cache capacity $M = N/K$, i.e., $R^{\rm{C}}_{\rm{GBC}}( N/K )$ given in \eqref{OurDeliveryRatePointTheorem}, is compared with the best achievable delivery rate in the literature for the same cache size, when the number of files is $N=100$, and the number of users varies from $K = 200$ to $K = 1000$. In this scenario, since $K \ge 2N$, the best achievable rate in the literature, $R_b^{\rm{C}}( N/K )$, is determined by \eqref{BestDeliveryRatePoint}. The WTP scheme is not included as it again achieves the same delivery rate as $R_b^{\rm{C}}( N/K )$. We observe that the GBC provides a significant gain in the delivery rate for the whole range of $K$ values; however, the improvement is more pronounced for smaller number of users, $K \sim 200$. For $K=200$ users in the system, the GBC scheme provides a $9.75\%$ reduction in the delivery rate. We also observe that the gap between the delivery rate of the GBC scheme and the lower bound decreases with increasing number of users in the system.
\begin{figure}[!t]
\centering
\includegraphics[scale=0.7]{N_100_K_Variable.pdf}
\caption{Delivery rate of the GBC scheme for cache capacity $M = N/K$ as a function of $K$, for $K \in \left[ 200:1000 \right]$, and $N=100$. The best achievable delivery rate in the literature is determined by \eqref{BestDeliveryRatePoint} for all values of $K$.}
\label{NConstantKVariable}
\end{figure}
\section{Decentralized Coded Caching}\label{DecentralizedScheme}
In this section, we consider a decentralized caching system, in which neither the number nor the identity of the users that participate in the delivery phase are known during the placement phase \cite{MaddahAliDecentralized}. Accordingly, in the decentralized setting, we cannot coordinate the cache contents among users in order to maximize the multicasting opportunities during the delivery phase. We apply the group-based caching ideas we have developed for the centralized scenario to decentralized caching. The corresponding caching scheme is called the group-based decentralized coded caching (GBD).
To simplify the notation, for $k=1, ..., K$, we define $W_{d_k,V}$ as the bits of file $W_{d_k}$, the file requested by user $U_k$, which have been placed exclusively in the caches of the users in set $V$, where $V \subset \left[ {1:K} \right]$. We also note that, all XOR operations used in this section are assumed to be zero-padded such that they all have the same length as the longest of their arguments. The GBD scheme is first demonstrated on a simple example.
\begin{exmp}
Consider a decentralized coded caching system with $K=5$ users and $N=3$ files. Due to lack of coordination in the placement phase, similarly to the scheme proposed in \cite{MaddahAliDecentralized}, $MF/3$ random bits of each file are cached independently at each user during the placement phase. During the delivery phase, without loss of generality, by re-ordering the users, the following worst-case user demands are considered:
\begin{equation}\label{DemandsExample2}
{d_k} =
\begin{cases}
{{1},\quad 1 \le k \le 2,}\\
{{2},\quad 3 \le k \le 4,}\\
{{3},\quad k = 5.}
\end{cases}
\end{equation}
The contents are sent by the server in three parts during the delivery phase to satisfy the above demand combination. Below, we explain each part in detail.
\begin{enumerate}[label=\bfseries Part \arabic*:,align=left]
\item In the first part, the bits of each requested file $W_i$, which are not cached anywhere, are delivered directly by the server, for $i \in \left[ 1:3 \right]$. Therefore, the server delivers the following bits in the first part of the delivery phase: ${W_{1,\left\{ \emptyset \right\} }}$, ${W_{2,\left\{ \emptyset \right\} }}$, ${W_{3,\left\{ \emptyset \right\} }}$.
\item In part 2, the bits of file $W_{d_k}$ requested by user $U_k$, which are only in the cache of user $U_j$ are delivered, for $k,j \in \left[ 1:5 \right]$ and $k \ne j$. In other words, the purpose of this part is to enable each user to decode the bits of its requested file which have been placed in the cache of only one other user. Accordingly, the delivery phase of the GBC scheme can be applied in this part. For the example under consideration, the following bits constitute the second part of the delivery phase: ${W_{1,\left\{1\right\}}} \oplus {W_{1,\left\{2\right\}}}$, ${W_{2,\left\{3\right\}}} \oplus {W_{2,\left\{4\right\}}}$, ${W_{1,\left\{3\right\}}} \oplus {W_{1,\left\{4\right\}}}$, ${W_{2,\left\{1\right\}}} \oplus {W_{2,\left\{2\right\}}}$, ${W_{1,\left\{4\right\}}} \oplus {W_{2,\left\{2\right\}}}$, ${W_{3,\left\{1\right\}}} \oplus {W_{3,\left\{2\right\}}}$, ${W_{1,\left\{5\right\}}} \oplus {W_{3,\left\{1\right\}}}$, ${W_{3,\left\{3\right\}}} \oplus {W_{3,\left\{4\right\}}}$, ${W_{2,\left\{5\right\}}} \oplus {W_{3,\left\{4\right\}}}$.
\item The bits of the file requested by each user which are in the cache of more than one other user are delivered in the third part by exploiting the procedure proposed in \cite[Algorithm 1]{MaddahAliDecentralized}. The server delivers the following in the third and last part of the delivery phase: ${W_{1,\left\{ {2,3} \right\}}} \oplus {W_{1,\left\{ {1,3} \right\}}} \oplus {W_{2,\left\{ {1,2} \right\}}}$, ${W_{1,\left\{ {2,4} \right\}}} \oplus {W_{1,\left\{ {1,4} \right\}}} \oplus {W_{2,\left\{ {1,2} \right\}}}$, ${W_{1,\left\{ {2,5} \right\}}} \oplus {W_{1,\left\{ {1,5} \right\}}} \oplus {W_{3,\left\{ {1,2} \right\}}}$, ${W_{1,\left\{ {3,4} \right\}}} \oplus {W_{2,\left\{ {1,4} \right\}}} \oplus {W_{2,\left\{ {1,3} \right\}}}$, ${W_{1,\left\{ {3,5} \right\}}} \oplus {W_{2,\left\{ {1,5} \right\}}} \oplus {W_{3,\left\{ {1,3} \right\}}}$, ${W_{1,\left\{ {4,5} \right\}}} \oplus {W_{2,\left\{ {1,5} \right\}}} \oplus {W_{3,\left\{ {1,4} \right\}}}$, ${W_{1,\left\{ {2,3,4} \right\}}} \oplus {W_{1,\left\{ {1,3,4} \right\}}} \oplus {W_{2,\left\{ {1,2,4} \right\}}} \oplus {W_{2,\left\{ {1,2,3} \right\}}}$, ${W_{1,\left\{ {2,3,5} \right\}}} \oplus {W_{1,\left\{ {1,3,5} \right\}}} \oplus {W_{2,\left\{ {1,2,5} \right\}}} \oplus {W_{3,\left\{ {1,2,3} \right\}}}$, ${W_{1,\left\{ {2,4,5} \right\}}} \oplus {W_{1,\left\{ {1,4,5} \right\}}} \oplus {W_{2,\left\{ {1,2,5} \right\}}} \oplus {W_{3,\left\{ {1,2,4} \right\}}}$, ${W_{1,\left\{ {3,4,5} \right\}}} \oplus {W_{2,\left\{ {1,4,5} \right\}}} \oplus {W_{2,\left\{ {1,3,5} \right\}}} \oplus {W_{3,\left\{ {1,3,4} \right\}}}$, ${W_{1,\left\{ {2,3,4,5} \right\}}} \oplus {W_{1,\left\{ {1,3,4,5} \right\}}} \oplus {W_{2,\left\{ {1,2,4,5} \right\}}} \oplus {W_{2,\left\{ {1,2,3,5} \right\}}} \oplus {W_{3,\left\{ {1,2,3,4} \right\}}}$, ${W_{1,\left\{ {3,4} \right\}}} \oplus {W_{2,\left\{ {2,4} \right\}}} \oplus {W_{2,\left\{ {2,3} \right\}}}$, ${W_{1,\left\{ {3,5} \right\}}} \oplus {W_{2,\left\{ {2,5} \right\}}} \oplus {W_{3,\left\{ {2,3} \right\}}}$, ${W_{1,\left\{ {4,5} \right\}}} \oplus {W_{2,\left\{ {2,5} \right\}}} \oplus {W_{3,\left\{ {2,4} \right\}}}$, ${W_{1,\left\{ {3,4,5} \right\}}} \oplus {W_{2,\left\{ {2,4,5} \right\}}} \oplus {W_{2,\left\{ {2,3,5} \right\}}} \oplus {W_{3,\left\{ {2,3,4} \right\}}}$, ${W_{2,\left\{ {4,5} \right\}}} \oplus {W_{2,\left\{ {3,5} \right\}}} \oplus {W_{3,\left\{ {3,4} \right\}}}$.
\end{enumerate}
In this case, using the law of large numbers, the size of $W_{i,V}$, for $F$ large enough, can be approximated as
\begin{equation}\label{SizePortion} \left| {{W_{i,V}}} \right| \approx {\left( {\frac{M}{3}} \right)^{\left| V \right|}}{\left( {1 - \frac{M}{3}} \right)^{5 - \left| V \right|}}F,
\end{equation}
for any set $V \subset \left[ {1:5} \right]$. For example, for a cache capacity of $M=1$, the delivery rate of the proposed coded caching scheme is $1.407$, while the schemes proposed in \cite{MaddahAliDecentralized} and \cite{KaiWanUncodedCaching}, can achieve the delivery rates $1.737$ and $1.473$, respectively.
\qed
\end{exmp}
\subsection{Group-Based Decentralized Coded Coding (GBD) Scheme}
Here, we generalize the proposed decentralized group-based coded caching (GBD) scheme. To simplify the notation, for $n,m \in \mathcal Z$, and $V \subset \left[1:K \right]$, we use the notation $\left\{ {V,n} \right\}\backslash \left\{ m \right\}$ to represent the set of integers $V \cup \{n\} \backslash \{m\}$.
In the placement phase, each user caches a random subset of $MF/N$ bits of each file independently. Since there are $N$ files, each of length $F$ bits, this placement phase satisfies the memory constraint.
\begin{algorithm}
\caption{Coded Delivery Phase of the GBD Scheme}
\label{DeliveryDecentralized}
\begin{algorithmic}[1]
\Statex
\Procedure {Delivery-Coded}{}
\State{\textbf{Part 1}: Delivering bits that are not in the cache of any user}
\For {$i = 1, 2, \ldots, N$}
\State{server delivers $\left( W_{{d_{S_{i-1} + 1}},\left\{ \emptyset \right\}} \right)$}
\EndFor
\Statex
\State{\textbf{Part 2}: Delivering bits that are in the cache of only one user}
\State{server delivers $\left( {\bigcup\limits_{i = 1}^N {\bigcup\limits_{k = {S_{i - 1}} + 1}^{{S_i} - 1} {\left( {{W_{i,\left\{ k \right\}}} \oplus {W_{i,\left\{ k + 1 \right\}}}} \right)} } } \right)$.}
\State{server delivers $\bigcup\limits_{i = 1}^{N - 1} \bigcup\limits_{j = i + 1}^N \left( \bigcup\limits_{k = {S_{j - 1}} + 1}^{{S_j} - 1} {\left( {{W_{i,\left\{ k \right\}}} \oplus {W_{i,\left\{ {k + 1} \right\}}}} \right)} ,\qquad \qquad \qquad \qquad \qquad \qquad \qquad \right.$ $ \left. \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \bigcup\limits_{k = {S_{i - 1}} + 1}^{{S_i} - 1} {\left( {{W_{j,\left\{ k \right\}}} \oplus {W_{j,\left\{ {k + 1} \right\}}}} \right)} , {W_{i,\left\{ {{S_j}} \right\}}} \oplus W_{j,\left\{ {{S_i}} \right\}} \right) $.}
\Statex
\State{\textbf{Part 3}: Delivering bits that are in the cache of more than one user}
\For{$i = 1, 2, \ldots, K - 2$ }
\For{$j = 2, 3, \ldots, K - i$ }
\For{$V \subset \left[ {i + 1:K} \right]: \left| V \right| = j$ }
\State server delivers $\left( \left( \mathop \oplus \limits_{v \in V} W_{d_{v},\left\{ {V,i} \right\}\backslash \left\{v\right\}} \right) \oplus W_{d_i, V } \right)$
\EndFor
\EndFor
\EndFor
\EndProcedure
\Statex
\Procedure {Delivery-Random}{}
\For{$i = 1, 2, \ldots, N$}
\State {server delivers enough random linear combination of the bits of the file $W_i$ to the users requesting it in order to decode it}
\EndFor
\EndProcedure
\end{algorithmic}
\end{algorithm}
Similarly to Section \ref{GBCscheme}, without loss of generality, we can re-order the users and re-label the files such that the first $K_1$ users, referred to as group $G_1$, have the same request $W_1$, the next $K_2$ users, group $G_2$, request $W_2$, and so on so forth. As a result, in the delivery phase, we have
\begin{equation}\label{DemandsDecentralized} d_k = i,\quad \mbox{for $i=1, ..., N$, and $k=S_{i-1} + 1, ..., S_i$},
\end{equation}
where $S_i$ is as defined in (\ref{S_sumK}).
There are two different procedures for the delivery phase, called DELIVERY-CODED and DELIVERY-RANDOM, presented in Algorithm \ref{DeliveryDecentralized}. The server follows either of the two, whichever achieves a smaller delivery rate.
Let us start with the DELIVERY-CODED procedure of Algorithm \ref{DeliveryDecentralized}, in which the contents are delivered in three distinct parts, as explained in the example above. The main idea behind the coded delivery phase is to deliver each user the missing bits of its requested file, that have been cached by $i$ user(s), $\forall i \in \left[ 0:K-1 \right]$.
In the first part, the bits of each requested file that are not in the cache of any user are directly delivered by the server. Each transmitted content is destined for all the users in a separate group, which have the same request.
In the second part, the bits of each requested file that have been cached by only one user are served to the users requesting the file by utilizing the GBC scheme developed for the centralized scenario.
Each user $U_k$ in group $G_i$ requests $W_i$ and has already cached $W_{i,\left\{ k \right\}}$ for $i \in \left[1:N \right]$ and $k \in \left[ S_{i-1}+1:S_i \right]$. Having received the bits delivered in line 7 of Algorithm \ref{DeliveryDecentralized}, $U_k$ can decode all bits $W_{i,\left\{ l \right\}}$, $\forall l \in \left[ S_{i-1}+1:S_i \right]$. The users also receive the missing bits of their requested files having been cached by a user in a different group; that is, by receiving $\bigcup\limits_{k = S_{i - 1} + 1}^{S_i - 1} \left( W_{j,\left\{k\right\}} \oplus W_{j,\left\{k + 1\right\}} \right)$, $\bigcup\limits_{k = {S_{j - 1}} + 1}^{{S_j} - 1} {\left( {{W_{i,\left\{k\right\}}} \oplus {W_{i,\left\{k + 1\right\}}}} \right)}$, and ${W_{i,\left\{S_j\right\}}} \oplus {W_{j,\left\{S_i\right\}}}$, each user in groups $G_i$ and $G_j$ can decode the bits of its request which have been placed in the cache of users in the other group, for $i=1, ..., N-1$ and $j=i+1, ..., N$.
In the last part, the same procedure as the one proposed in \cite{MaddahAliDecentralized} is performed for the missing bits of each file that have been cached by more than one user. Hence, following the DELIVERY-CODED procedure presented in Algorithm \ref{DeliveryDecentralized}, each user recovers all the bits of its desired file.
The second delivery procedure, DELIVERY-RANDOM, is as presented in In the last part, the same procedure as the one proposed in \cite{MaddahAliDecentralized}, and the server delivers enough random linear combinations of the bits of each requested file targeted for the users in the same group requesting that file to decode it.
\subsection{Delivery Rate Analysis}
In the following, we derive an expression for the delivery rate-cache capacity trade-off of the proposed GBD scheme, denoted by $R^{\rm{D}}_{\rm{GBD}}(M)$. All discussions in this section are stated assuming that $M \le N$, and $F$ is large enough. For each randomly chosen bit of each file, the probability of having been cached by each user is $M/N$. Since the contents are cached independently by each user in the placement phase, a random bit of each file is cached exclusively by the users in set $V \subset \left[ 1:K \right]$ (and no user outside this set) with probability $\left( M/N \right)^{\left| V \right|} \left( {1 - M/N} \right)^{K - \left| V \right|}$.
Similarly to the arguments presented in Section \ref{AnalysisGBC}, when $N < K$, the worst-case user demands correspond to the scenario in which each file is requested by at least one user, i.e., $K_i > 0$, $\forall i \in \left[ {1:N} \right]$. On the other hand, when $N \ge K$, without loss of generality, the worst-case user demands can be assumed as $d_k = k$, $\forall k \in \left[ {1:K} \right]$.
When $N \ge K$, for the worst-case user demands described above, similar conditions as the GBC scheme hold, and the GBD scheme achieves the same delivery rate as the decentralized caching scheme proposed in \cite{MaddahAliDecentralized}, called the decentralized MAN scheme, i.e., $R^{\rm{D}}_{\rm{GBD}}\left( M \right) = R^{\rm{D}}_{\rm{MAN}}\left( M \right)$.
Next we consider the more interesting $N < K$ case. We start with the first procedure of Algorithm \ref{DeliveryDecentralized}. In part 1, the server delivers $N$ groups of bits, each group corresponding to a different file, which have not been cached by any user. The delivery rate-cache capacity trade-off for this part of Algorithm \ref{DeliveryDecentralized}, $R^{\rm{D}}_{1}(M)$, can be evaluated as
\begin{equation}\label{OurDeliveryRateDecentralized1} {R^{\rm{D}}_{{1}}}\left( M \right) = N{\left( {1 - \frac{M}{N}} \right)^K}.
\end{equation}
For the second part, we first need to find the total number of XOR-ed contents delivered by the server, each of which has length $\left( M/N \right) \left( 1-M/N \right)^{K-1}F$ bits (the length of each XOR-ed content is equivalent to the number of bits of a file that have been cached by only one user). Since the delivery phase of the GBC scheme is applied for this part, based on \eqref{OurDeliveryRatePointTheorem}, it can be easily evaluated that $\left( {NK - N(N+1)/2} \right)$ XOR-ed contents are served\footnote{Note that, in the delivery phase of the GBC scheme for $M = N/K$, a total of $\left( NK - N(N+1)/2 \right)$ XOR-ed contents, each of size $F/K$ bits, are delivered, which results in $R^{\rm{C}}_{\rm{GBC}}( N/K )=N-N(N+1)/2K$.}. Thus, the delivery rate-cache capacity trade-off corresponding to the second part of the first procedure of Algorithm \ref{DeliveryDecentralized} is given by
\begin{equation}\label{OurDeliveryRateDecentralized2} {R^{\rm{D}}_{{2}}}\left( M \right) = \left( {NK - \frac{{N\left( {N + 1} \right)}}{2}} \right)\left( {\frac{M}{N}} \right){\left( {1 - \frac{M}{N}} \right)^{K - 1}}.
\end{equation}
The last part of the proposed delivery phase is equivalent to the first delivery phase procedure proposed in \cite[Algorithm 1]{MaddahAliDecentralized}, with which each user can decode the bits of its requested file, which have been cached by more than one user. Following the same technique as \cite{MaddahAliDecentralized}, the delivery rate corresponding to this part is derived as follows:
\begin{align}\label{OurDeliveryRateDecentralized3} {R^{\rm{D}}_{{3}}}\left( M \right) = & \sum\limits_{i = 1}^{K - 2} {\sum\limits_{j = 2}^{K - i} {\binom{K-i}{j}{{\left( {\frac{M}{N}} \right)}^j}{{\left( {1 - \frac{M}{N}} \right)}^{K - j}}} } \nonumber\\
= & - \left( {K - 2} \right){\left( {1 - \frac{M}{N}} \right)^K} - \frac{1}{2}\left( {K - 2} \right)\left( {K + 1} \right){\left( {\frac{M}{N}} \right)}{\left( {1 - \frac{M}{N}} \right)^{K - 1}} \nonumber\\
&~~~~~~ + \frac{N}{M}\left( {1 - {{\left( {1 - \frac{M}{N}} \right)}^{K - 1}}} \right) - 1.
\end{align}
The overall delivery rate-cache capacity trade-off for the first procedure of Algorithm \ref{DeliveryDecentralized}, $R^{\rm{D}}_{\rm{GBD}_1}(M)$, is the sum of the delivery rates of all three parts in (\ref{OurDeliveryRateDecentralized1}), (\ref{OurDeliveryRateDecentralized2}), and (\ref{OurDeliveryRateDecentralized3}), and is evaluated as
\begin{align} {R^{\rm{D}}_{{\rm{GBD}_1}}}\left( M \right) &= {R^{\rm{D}}_{{1}}}\left( M \right) + {R^{\rm{D}}_{{2}}}\left( M \right) + {R^{\rm{D}}_{{3}}}\left( M \right) \nonumber \\
&= \frac{N}{M} - 1 - \left[ {\left( {K - N - 2} \right)\left( {1 + \frac{1}{2}\left( {K - N - 1} \right)\frac{M}{N}} \right) + \frac{N}{M}} \right]{\left( {1 - \frac{M}{N}} \right)^{K - 1}}. \label{OurDeliveryRateDecentralizedFirstProcedure}
\end{align}
For the worst-case user demands, it is shown in \cite[Appendix A]{MaddahAliDecentralized} that the second delivery procedure achieves the same delivery rate-cache capacity trade-off as the uncoded scheme, given by
\begin{equation}\label{UncodedDeliveryRate} {R^{\rm{D}}_{\rm{GBD}_2}}\left( M \right) = K\left( {1 - \frac{M}{N}} \right)\min \left\{ {1,\frac{N}{K}} \right\}.
\end{equation}
The delivery rate of the proposed GBD scheme is evidently the minimum value of $R^{\rm{D}}_{\rm{GBD}_1}(M)$ and $R^{\rm{D}}_{\rm{GBD}_2}(M)$, which is presented in the following theorem.
\begin{theorem}
In a decentralized caching system with $K$ users requesting contents from a server with $N$ files in its database, when $N < K$, the following delivery rate-cache capacity trade-off is achievable by the GBD scheme:
\begin{align}\label{DeliveryRateAlgorithm2}
& R^{\rm{D}}_{\rm{GBD}}\left( M \right) = \left( {1 - \frac{M}{N}} \right)\nonumber\\
& \quad \; \times \min \left\{ \frac{N}{M} - \left[ {\left( {K - N - 2} \right)\left( {1 + \frac{1}{2}\left( {K - N - 1} \right)\frac{M}{N}} \right) + \frac{N}{M}} \right]{{\left( {1 - \frac{M}{N}} \right)}^{K - 2}},N \right\}.
\end{align}
\end{theorem}
\begin{remark}
We remark that in the decentralized caching model, it is assumed that each user sends its cache content together with its request to the server at the beginning of the delivery phase. In this way, by knowing the number of popular files in the database, when $N<K$, the server can decide to perform the delivery phase procedure that requires the smallest delivery rate.
\end{remark}
\subsection{Comparison with the State-of-the-Art}
The difference between the proposed GBD scheme and the scheme investigated in \cite[Algorithm 1]{MaddahAliDecentralized} lies in the first procedure of the delivery phase when $N < K$. As a result, to compare the two schemes, the delivery rate of the proposed GBD scheme for the DELIVERY-CODED procedure, i.e., $R_{\rm{GBD}_1}^{\rm{D}}(M)$ given in \eqref{OurDeliveryRateDecentralizedFirstProcedure}, should be compared with the delivery rate of the first delivery phase procedure stated in \cite[Algorithm 1]{MaddahAliDecentralized}, which is given by
\begin{equation}\label{DeliveryRateDecentralizedFirstProcedureMaddahAli} R_{\rm{MAN}_1}^{\rm{D}}\left( M \right) = \left( \frac{N}{M} - 1 \right) \left( 1 - \left( {1 - \frac{M}{N}} \right)^K \right).
\end{equation}
For $M \le N$, we have ${\left( {1 - M/N} \right)^K} \le {\left( {1 - M/N} \right)^{K - 1}}$. Hence, to show that ${R_{\rm{GBD}_1}^{\rm{D}}}\left( M \right) \le {R_{\rm{MAN}_1}^{\rm{D}}}\left( M \right)$, it suffices to prove that
\begin{align}\label{DeliveryRateDecentralizedFirstProcedureCompare} \left( {K - N - 2} \right)\left( {1 + \frac{1}{2}\left( {K - N - 1} \right)\frac{M}{N}} \right) \ge - 1,
\end{align}
which holds for $N < K$. Therefore, compared to the decentralized coded caching scheme proposed in \cite[Algorithm 1]{MaddahAliDecentralized}, the GBD scheme requires a smaller delivery rate, if $N < K$.
\begin{figure}[!t]
\centering
\includegraphics[scale=0.7]{Decentralized_N_30_K_50.pdf}
\caption{Delivery rate-cache capacity trade-off in the decentralized setting for the GBD, MAN and WTP schemes, for $N=30$ and $K=50$. In this scenario, we have $\hat t = t^* = 2$, and the best centralized performance can be achieved by memory-sharing between the CFL scheme for $M = 1/K$, the GBC scheme for $M = N/K$, and the MAN scheme for cache capacities $M = lN/K$, where $l \in \left[ 2:K \right]$.}
\label{DecentralizedN30K50}
\end{figure}
Next, we compare the GBD scheme with the state-of-the-art caching schemes numerically. In Fig. \ref{DecentralizedN30K50}, the delivery rate of the proposed GBD scheme is compared with the decentralized caching schemes proposed in \cite{MaddahAliDecentralized} and \cite{KaiWanUncodedCaching}, referred to as MAN and WTP, respectively, for $N=30$ and $K=50$. The superiority of the proposed GBD scheme over the state-of-the-art caching schemes, especially for relatively small cache capacities, is visible in the figure. We also include in the figure the best achievable centralized coded caching scheme for this setting, the delivery rate-cache capacity trade-off of which is denoted by $R_{\rm{Cent}}(M)$. This curve is obtained through memory-sharing between the CFL scheme for cache capacity $M = 1/K$, the proposed GBC scheme for $M = N/K$, and the centralized MAN scheme for $M = lN/K$, where $l \in \left[ \hat t:K \right]$ (note that, in this scenario, $\hat t = t^* = 2$). Although the delivery rate of the optimal decentralized caching scheme might not be lower bounded by that of the considered centralized caching scheme (since the centralized caching scheme under consideration is not optimal), the difference between the delivery rates of the centralized and decentralized schemes roughly indicates the loss due to the decentralization of the coded caching schemes considered here. We observe that the delivery rate of the proposed GBD scheme is very close to the performance of the best known centralized scheme, particularly for small cache capacities.
\section{Conclusion}\label{Conclusion}
We have considered a symmetric caching system with $K$ users and $N$ files, where each file has the same size of $F$ bits, and each user has the same cache capacity of $MF$ bits, that is, a cache that is sufficient to store $M$ files. The system considered here models wireless networks, in which the caches are filled over off-peak periods without any cost constraint or rate limitation (apart from the limited cache capacities), but without knowing the user demands; and all the user demands arrive (almost) simultaneously, and they are served simultaneously through an error-free shared link. We have proposed a novel group-based centralized (GBC) coded caching scheme for a cache capacity of $M = N/K$, which corresponds to the case in which the total cache capacity distributed across the network is sufficient to store all the files in the database. In the centralized scenario, each file is distributed evenly across all the users, and the users are grouped based on their requests. Our algorithm satisfies user demands in pairs, making sure that each transmitted information serves all the users in a group, that is, users requesting the same file. We have shown that the proposed caching scheme outperforms the best achievable schemes in the literature in terms of the delivery rate. Then the improvement has been extended to a larger range of cache capacities through memory-sharing with the existing achievable schemes in the literature.
We have next employed the GBC scheme in the decentralized setting. Since the user identities are not known in the decentralized caching scenario, it is not possible to distribute the contents among users. Allowing the users to cache bits of contents randomly, we will have different parts of each content cached by a different set of users. Our decentralized caching scheme, GBD, adopts the proposed group-based centralized caching scheme in order to efficiently deliver those pats of the files that have been cached by only a single user. We have shown that the GBD scheme also achieves a smaller delivery rate than any other known scheme in the literature. We believe that the proposed group-based delivery scheme is particularly appropriate for non-uniform user demands, which is currently under investigation.
\appendices
\section{Proof of Theorem \ref{TheoremGBC}}\label{ProofFirstTheorem}
To prove Theorem \ref{TheoremGBC}, we first go through the coded delivery phase presented in Algorithm \ref{DeliveryCentralized}, and show that all user requests are satisfied at the end of the delivery phase. First part of this algorithm enables each user to obtain the subfiles of its requested file which are in the cache of all other users in the same group. We consider the first group, i.e., $i=1$ in line 2 of Algorithm \ref{DeliveryCentralized}, which includes the users that demand $W_1$. In this case, the XOR-ed contents $W_{1,k} \oplus W_{1,k+1}$, for $k \in \left[ {1:{K_1} - 1} \right]$, are delivered by the server. Having access to the subfile $W_{1,k}$ locally in its cache, each user $U_k$, for $k \in \left[ {1:{K_1}} \right]$, can decode all the remaining subfiles $W_{1,j}$, for $j \in \left[ {1:{K_1}} \right]\backslash \left\{ {k} \right\}$. Thus, a total number of $(K_{1} - 1)$ XOR-ed contents, each of size $\frac{F}{K}$ bits, are delivered by the server for the users in group $G_1$. Similarly, the second group ($i=2$ in line 2 of Algorithm \ref{DeliveryCentralized}), containing the users requesting file $W_2$, the XOR-ed contents $W_{2,k} \oplus W_{2,k+1}$, for $k \in \left[ {K_{1} + 1:{K_1 + K_{2}} - 1} \right]$, are sent by the server. With subfile $W_{2,k}$ available locally at user $U_k$, for $k \in \left[ {K_{1} + 1:{K_1 + K_{2}}} \right]$, user $U_k$ can obtain the missing subfiles $W_{2,j}$, for $j \in \left[ {K_{1} + 1:{K_1 + K_{2}}} \right] \backslash \left\{ {k} \right\}$. Hence, a total of $(K_2 - 1)F/K$ bits are served for the users in $G_2$, and so on so forth. Accordingly, for the users belonging to group $G_i$, $(K_i - 1)F/K$ bits are delivered by the server, for $i=1, ..., N$, and the total number of bits transmitted by the server in the first part of the coded delivery phase presented in Algorithm \ref{DeliveryCentralized} is given by
\begin{equation}\label{DeliveryRateFirstProcedure} \frac{F}{K}\sum\limits_{i = 1}^N {\left( {{K_i} - 1} \right)} = \left( {K - N} \right)\frac{F}{K}.
\end{equation}
In the second part of Algorithm \ref{DeliveryCentralized}, each user in group $G_i$, for $i \in \left[ 1:N \right]$, will decode the missing subfiles of its requested file, which are in the cache of users belonging to groups $j \in \left[ 1:N \right] \backslash \left\{ {i} \right\}$. We first start with $i=1$ and $j=2$ in lines 7 and 8, respectively. The XOR-ed contents $W_{1,k} \oplus W_{1,k+1}$, for $k \in \left[ {{K_1} + 1:{K_1} + {K_2} - 1} \right]$, i.e., the subfiles of $W_1$ cached by users in group $G_2$, are delivered in line 9. In line 10, the XOR-ed contents $W_{2,k} \oplus W_{2,k+1}$, for $k \in \left[ 1:{K_1} - 1 \right]$, i.e., the subfiles of $W_2$ cached by users in group $G_1$, are delivered by the server. Finally, by delivering $W_{1,K_1 + K_2} \oplus W_{2,K_1}$ in line 11, and having already decoded $W_{2,k}$ ($W_{1,k}$), each user $U_k$ in $G_1$ ($G_2$) can recover the missing subfiles of its requested file $W_1$ ($W_2$) which are in the cache of users in $G_2$ ($G_1$), for $k \in \left[ 1:{K_1} \right]$ (for $k \in \left[ {{K_1} + 1:{K_1} + {K_2}} \right]$). In this particular case, the number of bits delivered by the server in lines 9, 10, and 11 are $(K_2 - 1)F/K$, $(K_1 - 1)F/K$, and $F/K$, respectively, which adds up to a total number of $(K_1 + K_2 - 1)F/K$ bits. In a similar manner, the subfiles can be exchanged between users in groups $G_i$ and $G_j$, for $i \in \left[ 1:N-1 \right]$ and $j \in \left[ i + 1:N \right]$, by delivering a total of $(K_i + K_j - 1)F/K$ bits through sending the XOR-ed contents stated in lines 9, 10, and 11 of Algorithm \ref{DeliveryCentralized}. Hence, the total number of bits delivered by the server in the second part of the coded delivery phase is given by
\begin{equation}\label{DeliveryRateSecondProcedure} \frac{F}{K}\sum\limits_{i = 1}^{N - 1} {\sum\limits_{j = i + 1}^N {\left( {{K_i} + {K_j} - 1} \right)} } = \left( {N - 1} \right)\left( {K - \frac{N}{2}} \right)\frac{F}{K}.
\end{equation}
By summing up \eqref{DeliveryRateFirstProcedure} and \eqref{DeliveryRateSecondProcedure}, the delivery rate of the GBC scheme is given by
\begin{equation}\label{DeliveryRateEndProcedure} {R^{\rm{C}}_{\rm{GBC}}}\left( {\frac{N}{K}} \right) = N - \frac{{N\left( {N + 1} \right)}}{{2K}}.
\end{equation}
\section{Proof of Inequality ${R_{\rm{GBC}}^{\rm{C}}}\left( N/K \right) \le f\left( N,K,t \right)$}
\label{CentralizedImprovement}
In order to show that ${R_{\rm{GBC}}^{\rm{C}}}\left( N/K \right) \le R_b^{\rm{C}} \left( N/K \right)$, it suffices to prove that ${R_{\rm{GBC}}^{\rm{C}}}\left( N/K \right) \le f\left( N,K,t \right)$, $\forall t \in \left[ 1:K \right]$. Thus, for $N < K$ and $t \in \left[ 1:K \right]$, we need to determine
\begin{equation}\label{AppendixCentralized1} N - \frac{{N\left( {N + 1} \right)}}{{2K}} \le {\frac{{\left( {N - 1} \right)\left( {K - t} \right)}}{{\left( {t + 1} \right)\left( {tN - 1} \right)}} + {N^2}\left( {1 - \frac{1}{K}} \right)\left( {\frac{{t - 1}}{{tN - 1}}} \right)}.
\end{equation}
After some mathematical manipulations, the inequality \eqref{AppendixCentralized1} can be simplified to
\begin{equation}\label{AppendixCentralized2} \left( {N - 1} \right)\left[ {{N^2}{t^2} + \left( {N + 1} \right)\left( {N - 2K} \right)t + 2{K^2} + N - 2KN} \right] \ge 0.
\end{equation}
We define function $h:\left( \left[ {1:K} \right] \to \mathcal R \right)$ as follows:
\begin{equation}\label{AppendixCentralized3} h(t) \buildrel \Delta \over = {{N^2}{t^2} + \left( {N + 1} \right)\left( {N - 2K} \right)t + 2{K^2} + N - 2KN},
\end{equation}
and we need to show that $(N-1)h(t) \ge 0$. We prove this inequality for two following cases: $K \le N(N+1)/2$, and $K > N(N+1)/2$. First, let us start with $K \le N(N+1)/2$. We can rewrite $h(t)$ as follows:
\begin{equation}\label{AppendixCentralized4} h\left( t \right) = {\left( {K - tN} \right)^2} + \left( {{N^2} + N - 2K} \right)t + {K^2} + N - 2KN.
\end{equation}
Since $t \ge 1$ and $K \le N(N+1)/2$, we have
\begin{align}\label{AppendixCentralized5} h\left( t \right) \ge& {\left( {K - tN} \right)^2} + {K^2} + {N^2} - 2KN + 2N + 2K = {\left( {K - tN} \right)^2} + \left( K - N \right) \left( K - N - 2 \right).
\end{align}
For $t \in \left[ {1:K} \right]$ and $K \ge N + 1$, using \eqref{AppendixCentralized5}, it can be easily verified that
\begin{equation}\label{AppendixCentralized6} \left( N - 1 \right) h\left( t \right) \ge \left( N - 1 \right) \left( \left( K - tN \right)^2 + \left( K - N \right) \left( K - N - 2 \right) \right) \ge 0.
\end{equation}
When $K > N(N+1)/2$, the proof is provided for $N \ge 3$. In this case, we have $K \ge 2N$. To prove that $h(t) \ge 0$, we first define a polynomial function of degree two $g: \left( \mathcal R^+ \to \mathcal R \right)$ as follows:
\begin{equation}\label{AppendixCentralized7} g(x) \buildrel \Delta \over = {{N^2}{x^2} + \left( {N + 1} \right)\left( {N - 2K} \right)x + 2{K^2} + N - 2KN},
\end{equation}
and then show that $g(x) \ge 0$, $\forall x > 0$. To have $g(x) \ge 0$, $\forall x > 0$, the following inequality should be held
\begin{equation}\label{AppendixCentralized8} {\left( {N + 1} \right)^2}{\left( {N - 2K} \right)^2} - 4{N^2}\left( {2{K^2} + N - 2KN} \right) \le 0,
\end{equation}
which is equivalent to
\begin{equation}\label{AppendixCentralized9} 4\left( {{{\left( {N + 1} \right)}^2} - 2{N^2}} \right)\left[ {{K^2} - KN - \frac{{{N^2}{{\left( {N - 1} \right)}^2}}}{{4\left( {2{N^2} - {{\left( {N + 1} \right)}^2}} \right)}}} \right] \le 0.
\end{equation}
For $N \ge 3$, we have
\begin{equation}\label{AppendixCentralized10} {\left( {N + 1} \right)^2} - 2{N^2} = 1 - N\left( {N - 2} \right) < 0.
\end{equation}
As a result, inequality \eqref{AppendixCentralized9} can be interpreted as
\begin{equation}\label{AppendixCentralized11} {{K^2} - KN - \frac{{{N^2}{{\left( {N - 1} \right)}^2}}}{{4\left( {2{N^2} - {{\left( {N + 1} \right)}^2}} \right)}}} \ge 0.
\end{equation}
It can be easily verified that
\newcommand\myfirstinequality{\mathrel{\overset{\makebox[0pt]{\mbox{\normalfont\tiny\sffamily (a)}}}{\le}}}
\newcommand\mysecondinequality{\mathrel{\overset{\makebox[0pt]{\mbox{\normalfont\tiny\sffamily (b)}}}{\le}}}
\begin{equation}\label{AppendixCentralized12}
\frac{{{N^2}{{\left( {N - 1} \right)}^2}}}{{4\left( {2{N^2} - {{\left( {N + 1} \right)}^2}} \right)}} \myfirstinequality 2{N^2} \mysecondinequality KN,
\end{equation}
where the inequalities (a) and (b) come from the fact that $N \ge 3$ and $K \ge 2N$, respectively. Using inequality \eqref{AppendixCentralized12} in \eqref{AppendixCentralized11}, we have
\begin{equation}\label{AppendixCentralized13} {K^2} - KN - \frac{{{N^2}{{\left( {N - 1} \right)}^2}}}{{4\left( {2{N^2} - {{\left( {N + 1} \right)}^2}} \right)}} \ge {K^2} - 2KN \ge 0,
\end{equation}
where the first inequality is due to the fact that $K \ge 2N$. Hence, when $N \ge 3$ and $K \ge N(N+1)/2$, we have $g(x) \ge 0$, $\forall x>0$, which we conclude $h(t) \ge 0$.
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{s:intro}
We are living in a society that cannot do without
computer systems. Services supplied by computer systems
have permeated our environments and deeply changed our
societies and the way we live in them. Computers pervade
our lives, integrating themselves in all environments. At
first confined in large control rooms, now they take the
form of tiny embedded systems soon to be ``sprayed'' on
physical entities so as to augment them with advanced
processing and communication capabilities. Thus it is
very much evident to what extent we depend on computers.
What is often overlooked by many is the fact
that most of the logics behind computer services
supporting and sustaining our societies lies in the
software layers. Software has become the point of
accumulation of a large amount of complexity~\cite{Sch06}. It is
ubiquitous, mobile, and has pervaded all aspects of our
lives. What is more important for this discussion, software
is the main culprit behind the majority of
computer failures~\cite{Lyu98a,Lyu98b,Lapr98}.
Among the reasons that brought to this state of things we
focus our attention here on a particular one. Clever
organizations and system structures allowed the
visible complexity of software development to be reduced---at first
through modules and layers, then by means of objects, and
more recently with services, components, aspects, and models. As
a result, we have been given tools to compose and
orchestrate complex, powerful, and flexible
software-intensive systems in a relatively short amount
of time. The inherently larger flexibility of software
development turned software into the ideal ``location''
where to store the bulk of the complexity of nowadays'
computer-based services. Unfortunately, this very same
characteristic of software makes it also considerably
\emph{fragile to changes}~\cite{Sch06}. In particular
software's flexibility
also means that most of the assumptions drawn at
design-time may get invalidated when the software system
is ported, reused, redeployed, or simply when it is
executed in a physical environment other than the
one originally meant for. This means that
truly resilient software systems demand special care to
\emph{assumption failures detection}, \emph{avoidance},
and \emph{recovery}. Despite this fact,
no systematic approach allows yet for the expression
and verification of hypotheses regarding the expected
properties and behaviors of
\begin{itemize}
\item the hardware components (e.g. the failure semantics
of the memory modules we depend on);
\item third-party software (e.g. the reliability of an open-source
software library we make use of);
\item the execution environment (e.g. the security provisions
offered by the Java execution environment we are currently using);
\item the physical environment (e.g., the characteristics of the faults
experienced in
a space-borne vehicle orbiting around the sun).
\end{itemize}
While several tools exist, in practice
most of the above assumptions often end up being either
sifted off or ``hardwired'' in the executable code. As
such, those removed or concealed hypotheses cannot be
easily inspected, verified, or maintained.
Despite the availability of several conceptual and
practical tools---a few examples of which are briefly discussed
in Sect.~\ref{s:reltech}---still
we are lacking methodologies and architectures to tackle this
problem in its complex entirety---from design-time to the
various aspects of the run-time. As a consequence, our software systems
often end up being entities whose structure, properties,
and dependencies are not completely known, hence at times
deviate from their intended goals.
Across the system layers, a complex and at times obscure
``web'' of software machines is being executed
concurrently by our computers. Their mutual dependencies
determine the quality of the match of our software with
its deployment platform(s) and run-time environment(s)
and, consequently, their performance, cost, and in general
their quality of service and experience. At our behest or
otherwise, a huge variety of design assumptions is
continuously matched with the truth of the current
conditions. A hardware component assumed to be available;
an expected feature in an OSGi bundle or in a web browser
platform; a memory management policy supported by
a mobile platform~\cite{mw09.12}, or ranges of operational
conditions taken for granted at all times---all are but
assumptions and all have a dynamically varying truth
value.
Our societies, our very lives, are often entrusted to
machines driven by software; weird as it may sound,
in some cases this
is done without question---as an act of faith as it
were. This is clearly unacceptable. The more we rely on
computer systems---the more we depend on their correct
functioning for our welfare, health, and economy---the
more it becomes important to design those systems with
architectural and structuring techniques that allow
software complexity to be decomposed, but without hiding
in the process those hypotheses and assumptions
pertaining e.g. the target execution environment and the
expected fault- and system models.
Our position is that existing tools will have to be
augmented so as to minimize the risks of assumption
failures e.g. when porting, deploying, or moving software
to a new machine. We envision novel autonomic run-time
executives that continuously verify those hypotheses and
assumptions by matching them with endogenous knowledge
deducted from the processing subsystems as well as
exogenous knowledge derived from their execution and
physical environments. Mechanisms for propagating such
knowledge through all stages of software development would
allow the chances of assumptions failures to be considerably reduced.
The ultimate result we envisage is the ability to express
truly assumption failure-tolerant software systems, i.e.,
software systems that endorse provisions to efficiently
and effectively tackle---to some agreed upon extent---the
problem of assumption failures.
This paper makes three main contributions.
A first one is exposing our
vision of assumption failure-tolerant software systems.
Such systems explicitly address three main ``hazards'' of
software development, which we call the Horning
syndrome, the Hidden Intelligence syndrome, and the
Boulding syndrome. Assumption failures and the three
syndromes are presented in Sect.~\ref{s:hazards}.
A second contribution is introducing the
concept of assumption failure-tolerant software systems
and providing three examples of strategies---one for
each of the above syndromes.
This is done in Sect.~\ref{s:aftss}.
A third contribution is our vision of a holistic approach
to resilient software development, where the concept of
assumption failure plays a pivotal role. Such vision---introduced
after a brief overview of related and complementary
technologies in Sect.~\ref{s:reltech}---is the subject
of Sect.~\ref{s:vision}.
The paper is concluded by Sect.~\ref{s:end} in which we
summarize our main lessons learned and provide our conclusions.
\section{Three Hazards of Software Development}\label{s:hazards}
As mentioned before, assumption failures may have dire
consequences on software dependability. In what follows we
consider two well known exemplary cases from which we
derive a base of three ``syndromes'' that we
deem as the main hazards of
assumption failures. We assume the reader to be already
familiar with the basic facts of those two cases. Furthermore, we shall
focus our attention only on a few aspects and
causes---namely those more closely related to the subject at hand.
\subsection{Case 1: Ariane 5 Flight 501 Failure}
On June 4, 1996, the maiden flight of the Ariane 5 rocket
ended in a failure just forty seconds after its lift-off.
At an altitude of about 3,700 meters, the launcher veered
off its flight path, broke up and exploded. After the
failure, the European Space Agency set up an independent
Inquiry Board to identify the causes of the failure. The
Inquiry Board unravelled several reasons, the most
important of which was a failure in the so-called
Inertial Reference System (IRS), a key component
responsible for flight attitude and movement control in
space. Being so critical for the success of the mission,
the IRS adopted a simple hardware fault-tolerance
design pattern: two identical replicas were operating in
parallel (hot standby), executing the same software system. As
mentioned before, we shall not focus here on all the
design faults of this scheme, e.g. its lack of design
diversity~\cite{Avi85}. Our focus will be on one of
the several concomitant causes, namely a software reuse error
in the IRS. The Ariane 5 software included software
modules that were originally developed and successfully
used in the Ariane 4 program. Such software was written
with a specific physical environment as its reference.
Such reference environment was characterized by well
defined ranges for several flight trajectory parameters.
One such parameter was the rocket's maximum horizontal
velocity. In the Ariane 4, horizontal velocity could be
represented as a 16-bit signed integer. The Ariane 5 was
a new generation, thus it was faster. In particular
horizontal velocity could not be represented in a signed
short integer, which caused an overflow in both IRS
replicas. This event triggered a chain of failures that
led the rocket to complete loss of guidance and attitude
information shortly after the start of the ignition
sequence. Now completely blind and unaware, the Ariane 5
committed self destruction as an ultimate means to
prevent any further catastrophic failures.
The Ariane 5 failure provides us with several lessons---in the
rest of this subsection we shall focus on two of them.
\subsubsection{Horning Syndrome.}
The Ariane 5 failure warns us of the fact that an assumption
regarding the target physical environment of a software
component may clash with a real life fact. In the case at
hand, the target physical environment was assumed to be
one where horizontal velocity would not exceed some
agreed upon threshold. This assumption clashed with the
characteristics of a new target environment.
The term we shall use to describe this event is
``assumption failure'' or ``assumption-versus-context
clash''. The key lesson in this case is then that the
physical environment can play a fundamental role in
determining software quality.
By paraphrasing a famous quote by Whorf, the environment
shapes the way our fault-tolerance software is constructed
and determines how dependable it will ultimately be.
James Horning described
this concept through his well known quote~\cite{Hor98}:
\begin{quote}
``What is the most often overlooked risk in
software engineering?\\
That the environment will do something the
designer never anticipated.''
\end{quote}
This is precisely what happened in the case of the
failure of the Ariane 5's IRS: new unanticipated
environmental conditions violated some design
assumptions. For this reason we call this class of
assumption failures hazards ``the Horning Syndrome'', or
\hbox{$\hbox{S}_H$}{} for brevity. For the same reason we shall use the
terms ``Horning Assumptions'' (\hbox{$\hbox{A}_H$}) and ``Horning
Failures'' (\hbox{$\hbox{F}_H$}) respectively to refer to this class of
assumptions and of failures.
In what follows we shall use lowercase letters in Italics to denote
assumptions. Given a letter representing an assumption, the same letter
in bold typeface shall
represent the true value for that assumption. As an
example, the Ariane-5 failure was caused (among other reasons) by a
clash between $f$: \{``Horizontal Velocity can be
represented by a short integer''\} and \textbf{f}:
\{``Horizontal velocity is now $n$''\},
where $n$ is larger than the maximum short integer.
\subsubsection{Hidden Intelligence Syndrome.}
The second aspect we deem important to highlight in the
context of the failure of the IRS is related to a lack of
propagation of knowledge. The Horning Assumption that led
to this Horning Failure originated at Ariane 4's design
time. On the other hand the software code that
implemented the Ariane 4 design did not include any
mechanism to store, inspect, or validate such assumption.
This vital piece of information was simply lost. This
loss of information made it more difficult to verify the
inadequacy of the Ariane 4 software to the new
environment it had been deployed. We call an accident
such as this a case of the Hidden Intelligence Syndrome
(\hbox{$\hbox{S}_{{H\!I}}$}). Consequently we use the terms Hidden Intelligence
Assumption (\hbox{$\hbox{A}_{{H\!I}}$}) and Hidden Intelligence Failure (\hbox{$\hbox{F}_{{H\!I}}$}).
Unfortunately accidents due to the \hbox{$\hbox{S}_H$}{} and the \hbox{$\hbox{S}_{{H\!I}}$}{}
are far from being uncommon---computer history is
crowded with examples, with a whole range of
consequences. In what follows we highlight this fact in
another well known case---the deadly Therac-25 failures.
\subsection{Case 2: The Therac-25 Accidents}
The Therac-25 accidents have been branded as ``the most
serious computer-related accidents to
date''~\cite{Lev95}. Several texts describe and analyze
them in detail---including the just cited one. As we did
for the Ariane 5, here we shall not provide yet another
summary of the case; rather, we shall highlight the
reasons why the Therac-25 is also a case of the
above assumption hazards and of a third class of
hazards.
The Therac-25 was a so-called ``linac,'' that is, a
medical linear accelerator that uses accelerated
electrons to create high-energy beams to destroy
tumors with minimal impact on the surrounding healthy
tissue. It was the latest member of a successful family
of linacs, which included the Therac-6 and the
Therac-20. Compared to its predecessors, model 25 was
more compact, cheaper and had more functional features. In
particular the cheaper cost was a result of several
modifications including a substantial redesign of the
embedded hardware-software platform. In the redesign,
some expensive hardware services were taken over by the
software layer. For instance it was decided to remove
hardware interlocks that would shut the machine down in
the face of certain exceptions.
There is evidence that several such exceptions had
occurred while previous models, e.g. the Therac-20, were
operative. Unfortunately, none of these occurrences were
reported or fed back to the design process of the
Therac-25. Had it been otherwise, they would have revealed that
certain rare combinations of events triggered the
emission of extraordinary high levels of energy
beams---were it not for the safety interlocks present in
the old models. History repeated itself with model 25,
only this time the killer doses of beams \emph{were\/}
emitted, resulting in the killing or serious injuring of
several people.
\subsubsection{Another Case of the Horning Syndrome.} We
observe how the Therac may be considered as a special
case of Horning Assumption failure in which the
``unanticipated behavior'' is due to endogenous causes
and Horning's ``environment'' is the hardware platform.
The ``culprit'' in this case is the clash between two
design assumptions and two indisputable facts.
Assumptions were fault assumption $f$:
$\{$``\emph{No residual fault exists}''$\}$ and hardware
component assumption $p$: $\{$``\emph{All exceptions are
caught by the hardware and the execution environment, and
result in shutting the machine down}''$\}$. The
corresponding facts were $\mathbf{f}$:
$\{$``\emph{Residual faults still exist}''$\}$, that is
$\neg f$, and $\mathbf{p}$: $\{$``\emph{Exceptions exist
that are not caught}''$\}$---that is, $\neg p$. The
unanticipated behavior is in this case the machine still
remaining operative in a faulty state, thus the violation
of the safety mission requirements.
\subsubsection{Another Case of Hidden Intelligence.} As
mentioned already, because of the failure-free behavior
of the Therac-20, its software was considered as
fault-free. Reusing that software on the new machine
model produced a clash. Thus we could say that, for the
Therac family of machines, a hardware fault-masking
scheme translated into software hidden
intelligence---that is, a case of the \hbox{$\hbox{S}_{{H\!I}}$}. Such hidden
intelligence made it more difficult to verify the
inadequacy of the new platform to its
operational specifications.
\subsubsection{Boulding Syndrome.} Finally we observe how the
Therac-25 software, despite its real-time design goals, was
basically structured as a quasi closed-world system. Such
systems are among the naivest classes of systems in
Kenneth Boulding's famous classification~\cite{Bou56}:
quoting from the cited article, they belong to
the categories of ``Clockworks'' (``simple dynamic system
with predetermined, necessary motions'') and
``Thermostats'' (``control mechanisms in which [\dots\kern-1.5pt] the
system will move to the maintenance of any given
equilibrium, \emph{within limits}''). Such systems are
characterized by predefined assumptions about their
platform, their internal state, and the environment they
are meant to be deployed in. They are closed, ``blind''
entities so to say, built from synchronous assumptions,
and designed so as to be plugged in well defined hardware
systems and environments whose changes, idiosyncrasies,
or fluctuations most of them deliberately ignore. Using a
well known English vernacular, they are ``\emph{sitting ducks}''
to change---they keep on doing their prescribed task, as
defined at design time, irrespective of environmental
conditions; that is, they lack the ability to detect and
respond to deployment- and run-time changes.
Clearly the Therac machines and their software comply to
this definition. In particular those machines were
missing introspection mechanisms (for instance,
self-tests) able to verify whether the target platform did
include the expected mechanisms and behaviors.
A case like the Therac's---that is, when a clash exists
between a system's Boulding category and the actual
characteristics of its operational environment---shall be
referred to in what follows as a case of the Boulding
Syndrome (\hbox{$\hbox{S}_B$}). The above mentioned Boulding categories
and clashes will also be respectively referred to as
Boulding Assumptions (\hbox{$\hbox{S}_B$}) and Boulding Failures (\hbox{$\hbox{S}_B$}).
\subsection{Preliminary Conclusions}
By means of two well known cases we have shown how
computer system failures can be the result of software
assumption failures. Moreover, in so doing we have introduced
three major hazards or syndromes requiring particular
attention:
\begin{description}
\item[Horning syndrome:] mistakenly not considering that the
physical environment may change and produce unprecedented or
unanticipated conditions;
\item[Hidden Intelligence syndrome:] mistakenly concealing or
discarding important knowledge for the sake of hiding
complexity;
\item[Boulding syndrome:] mistakenly designing a system with
insufficient context-awareness with respect to the current
environments.
\end{description}
In what follows we describe examples of strategies to
treat some cases of the three syndromes so as to decrease
the risk to trigger assumption failures.
\section{Assumption Failure-Tolerant Software Systems}\label{s:aftss}
The key strategy we adopt here is to offer the designer the
possibility to postpone the choice of one out of multiple alternative
design-time assumptions to a proper future time
(compile-time, deployment-time, run-time, etc.) In what
follows we shall describe how to do so for the following
classes of assumptions:
\begin{itemize}
\item Assumptions related to the failure semantics of hardware components.
\item Assumptions related to the fault-tolerance design
patterns to adopt.
\item Assumptions related to dimensioning of resources.
\end{itemize}
\subsection{Assumptions on Hardware Components'
Failure Semantics}
As we have already remarked, software depends on certain
behaviors expected from the underlying hardware
architecture. Hardware neutrality and the principles of
layered design dictate that most of the actual processes
and actors in the bare machine are not disclosed. Thus
for instance we rarely know (and often care about) the
particular technology of the main memory integrated
circuits our software is making use of.
This is a case of
the Hidden Intelligence syndrome.
By not expressing explicitly our requirements concerning the way hardware
(e.g., memory modules) should behave
we leave the door open to
dependability assumption failures.
As an example, while yesterday's software was running
atop CMOS chips, today a common choice e.g. for airborne
applications is SDRAM---because of speed, cost, weight,
power and simplicity of design~\cite{Lad02}. But CMOS
memories mostly experience single bit
errors~\cite{1981coae.conf...66O}, while SDRAM chips are
known to be subjected to several classes of severe
faults, including so-called ``single-event
effects''~\cite{Lad02}, i.e.,
a threat that can lead to total loss of a
whole chip.
Examples include:
\begin{enumerate}
\item Single-event latch-up (SEL),
a threat that can bring to the loss of all data stored on
chip~\cite{WikiLatchup}.
\item Single-event upset (SEU), leading to
frequent soft errors~\cite{WikiSoftError,WikiSEU}.
\item Single-event functional interrupt (SFI), i.e. a special
case of SEU that places the
device into a test mode, halt, or undefined state. The SFI
halts normal operations, and requires
a power reset to recover~\cite{Holbert}.
\end{enumerate}
Furthermore~\cite{Lad02} remarks
how even \emph{from lot to lot\/} error and failure rates
can vary more than one order of magnitude. In other
words, the superior performance of the new generation of
memories is paid with a higher instability and a trickier
failure semantics.
Let us suppose for the time being that the software
system at hand needs to be compiled
in order to be executed on the target
platform. The solution we propose
to alleviate this problem is as follows:
\begin{itemize}
\item First, we assume memory access is abstracted (for instance
through services, libraries, overloaded operators, or aspects).
This allows the actual memory access methods
to be specified in a second moment.
\item Secondly, a number of design-time hypotheses regarding the failure semantics
of the hardware memory subsystem are drawn. These may take the form
of fault/failure assumptions such as for instance:
\begin{description}
\item{$f_0$:} ``Memory is stable and unaffected by failures''.
\item{$f_1$:} ``Memory is affected by transient faults and CMOS-like
failure behaviors''.
\item{$f_2$:} ``Memory is affected by permanent stuck-at faults and CMOS-like
failure behaviors''.
\item{$f_3$:} ``Memory is affected by transient faults and SDRAM-like
failure behaviors, including SEL''.
\item{$f_4$:} ``Memory is affected by transient faults and SDRAM-like
failure behaviors, including SEL and SEU''.
\end{description}
\item For each assumption $f_i$ (in this case $0\le i\le4$) a diverse set of
memory access methods, $M_i$, is designed. With the exception of $M_0$,
each $M_i$ is a fault-tolerant version specifically designed to
tolerate the memory modules' failure modes assumed in $f_i$.
\item To compile the code on the target platform, an Autoconf-like
toolset~\cite{Cal10} is assumed to be available. Special checking rules are coded
in the toolset making use of e.g. Serial Presence Detect (see Fig.~\ref{f:spd}) to
get access to information related to the memory modules on the target computer.
For instance, Linux tools such as ``\texttt{lshw}'' provide higher-level access to information
such as the memory modules' manufacturer, models, and characteristics
(see an example in Fig.~\ref{f:lshw}).
Such rules could access local or remote, shared databases reporting
known failure behaviors for models and even specific lots thereof.
Once the most probable memory behavior \textbf{f} is retrieved,
a method $M_j$ is selected
to actually access memory on the target computer.
Selection is done as follows: first we isolate
those methods that are able to tolerate \textbf{f}, then we
arrange them into a list ordered according to some cost function (e.g.
proportional to the expenditure of resources);
finally we select the minimum element of that list.
\end{itemize}
\begin{figure}[t]
\centerline{\includegraphics[angle=90,width=0.5\textwidth]{MVC-008F.png}}
\caption{The
Serial Presence Detect (yellow circle) allows information
about a computer's memory module, e.g. its manufacturer, model, size,
and speed, to be accessed.}
\label{f:spd}
\end{figure}
\begin{figure}[t]
\begin{verbatim}
*-memory
description: System Memory
physical id: 1000
slot: System board or motherboard
size: 1536MiB
*-bank:0
description: DIMM DDR Synchronous 533 MHz (1.9 ns)
vendor: CE00000000000000
physical id: 0
serial: F504F679
slot: DIMM_A
size: 1GiB
width: 64 bits
clock: 533MHz (1.9ns)
*-bank:1
description: DIMM DDR Synchronous 667 MHz (1.5 ns)
vendor: CE00000000000000
physical id: 1
serial: F33DD2FD
slot: DIMM_B
size: 512MiB
width: 64 bits
clock: 667MHz (1.5ns)
\end{verbatim}
\caption{Excerpt from the output of command-line \texttt{sudo lshw}
on a Dell Inspiron 6000 laptop.}\label{f:lshw}
\end{figure}
The above strategy allows the designer to postpone the
choice between alternative design-time assumptions to the
right moment, that it, when the code is compiled on the
chosen target action. A similar strategy could be
embedded in the execution environment, e.g. a Web browser
or a Java Virtual Machine. Such strategy could
selectively provide access at deployment time to
knowledge necessary to choose which of the design-time
alternative assumptions has the highest chance to match
reality. Note that our strategy helps avoiding \hbox{$\hbox{S}_{{H\!I}}$}{}
and brings the designer to explicitly deal with the problem
of assumption failures. Furthermore this is done with full
separation of the design concerns.
\subsubsection{Comparison with existing strategy.}
A somewhat similar strategy is used for performance
enhancement. Applications such as the mplayer video
player~\cite{Mplayer} can take advantage of predefined knowledge about
the possible target processor and enable optimized
methods to perform some of their tasks. Mplayer declares
this by displaying messages such as ``Using SSE optimized
IMDCT transform'' or ``Using MMX optimized resampler''.
Our procedure differs considerably from the mplayer's, as
it focuses on non-functional (dependability) enhancements. Furthermore, it
is a more general design methodology and makes use of
knowledge bases. Meta-object protocols, compiler
technology, and aspects could provide alternative way to
offer similar services.
\subsection{Choice of Fault-tolerance Design Patterns} \label{s:pattern}
The choice of which design pattern to use is known to
have a direct influence on a program's overall complexity
and performance. What is sometimes overlooked is the fact
that fault-tolerance design patterns have a strong
influence on a program's actual ability to tolerate
faults. For instance, a choice like the \textbf{redoing}
design pattern~\cite{SKS96}---i.e., repeat on failure---\emph{implies\/}
assumption $e_1:$ \{``The
physical environment shall exhibit transient faults''\},
while a design pattern such as \textbf{reconfiguration}---that is,
replace on failure---is
the natural choice after an assumption such as $e_2:$
\{``The physical environment shall exhibit permanent
faults''\}. Of course clashes are always possible, which
means in this case that there is a non-zero probability
of a Horning Assumption failure---that is, a case of the
\hbox{$\hbox{S}_H$}. Let us observe that:
\begin{enumerate}
\item A clash of assumption $e_1$
implies a livelock (endless repetition) as a result of redoing actions
in the face of permanent faults.
\item A clash of assumption $e_2$ implies an unnecessary
expenditure of resources as a result of applying
reconfiguration in the face of transient faults.
\end{enumerate}
The strategy we suggest to tackle this case
is to offer the designer the possibility to
postpone the binding of the actual
fault-tolerance design pattern and to condition it
to the actual behavior of the environment.
In what follows we describe a possible implementation
of this run-time strategy.
\begin{itemize}
\item First, we assume the software system
to be structured in such a way as to allow
an easy reconfiguration of its components.
Natural choices for this are service-oriented and/or
component-oriented architectures.
Furthermore we assume that the software architecture can be
adapted by changing a reflective meta-structure in the form
of a directed acyclic graph (DAG). A middleware supporting
this is e.g. ACCADA~\cite{GuD10+}.
\item Secondly, the designer draws a number of alternative
hypotheses regarding the faults to be experienced
in the target environments.
A possible choice could be for instance
$e_0$: ``No faults shall be experienced'' and then
$e_1$ and $e_2$ from above.
\item For each fault-tolerance assumption (in this case $e_1$
and $e_2$) a matching fault-tolerant design pattern is designed
and exported e.g. in the service or component registry.
The corresponding DAG snapshots are stored in
data structures $D_1$ and $D_2$.
\item Through e.g. publish/subscribe, the
supporting middleware component receives notifications
regarding the faults being detected by the main
components of the software system.
Such notifications are fed into an Alpha-count
filter~\cite{Bon96,BCDG00}, that is, a count-and-threshold
mechanism to discriminate between different types of
faults.
\item Depending on the assessment of the Alpha-count oracle,
either $D_1$ or $D_2$ are injected on the reflective DAG.
This has the effect or reshaping the software architecture
as in Fig.~\ref{f:d1tod2}.
Under the hypothesis of a correct oracle, such scheme avoids
clashes: always the most appropriate design
pattern is used in the face of certain classes of faults.
\end{itemize}
The above strategy is a second example of a way
to postpone the
choice among alternative design-time assumptions to the
right moment---in this case at run-time, when the
physical environment changes its characteristics or when
the software is moved to a new and different environment.
As a consequence, our strategy has the effect to help
avoiding \hbox{$\hbox{S}_H$}{} and to force the designer not to
neglect the problem of assumption failures.
\begin{figure}[t]
\centerline{\includegraphics[width=0.7\textwidth]{d1tod2.png}}
\caption{Transition from a redoing scheme $(D1)$ to
a reconfiguration scheme $(D_2)$ is obtained by replacing
component $c_3$, which tolerates transient faults by redoing its computation,
with a 2-version scheme where a primary component ($c_{3.1}$) is
taken over by a secondary one ($c_{3.2}$) in case of permanent faults.}
\label{f:d1tod2}
\end{figure}
We have developed a prototypical version of this strategy (see Fig.~\ref{f:alpha1})
and we are now designing a full fledged version based on the cited ACCADA
and on an Alpha-count framework built with Apache Axis2~\cite{Axis2}
and MUSE~\cite{Muse}.
\subsubsection{Comparison with existing strategies.}
Also in this case there exist strategies that postpone the
choice of the design pattern to execution time, though to the best of our
knowledge this has been done only with the design goal of achieving
performance improvements. A noteworthy example is FFTW,
a code generator for Fast Fourier Transforms that defines and assembles
(before compile time) blocks of C code that optimally solve FFT sub-problems
on a given machine~\cite{Fri04}.
Our strategy is clearly different in that
it focuses on dependability and makes use of a well-known count-and-threshold
mechanism.
\begin{figure}[t]
\hspace*{-0.05\textwidth}\includegraphics[width=1.1\textwidth]{alphacount2.png}
\caption{A scenario involving a watchdog (left-hand window)
and a watched task (right-hand). A permanent design fault
is repeatedly injected in the watched task.
As a consequence, the watchdog ``fires'' and an
alpha-count variable is updated. The value of
that variable increases until it overcomes a threshold (3.0) and correspondingly the
fault is labeled as ``permanent or intermittent.''}
\label{f:alpha1}
\end{figure}
\subsection{Assumptions Related to Dimensioning Replicated Resources}
As well known, a precise characterization of the amount of resources
necessary to deal with a certain situation is not always easy or even
possible to find out. In some cases, such amount is not to be considered
as a static value, fixed once and for all at design-time. Rather, it
should be modeled as a dynamic system, i.e. a variable changing over
time. When the situation to deal with is a threat to the quality
of a software service, then the common approach is to foresee
a certain amount of redundancy (time-, physical-, information-,
or design-redundancy). For instance, replication and voting can
be used to tolerate physical faults\footnote{Obviously
simple replication would not suffice to tolerate design faults,
in which case a design diversity scheme such as $N$-Version Programming
would be required.}. An important design problem is redundancy dimensioning.
Over-dimensioning redundancy or under-dimensioning it
would respectively lead to either a waste of resources or failures.
Ideally the replication and voting scheme should work with a
number of replicas that closely follows the evolution of
the disturbance. In other words, the system should be aware
of changes in certain physical variables or at least of the
effect they are producing to its internal state. Not doing so---that
is, choosing once and for all a certain degree of redundancy---means
forcing the designer to take one assumption regarding the
expected range of disturbances. It also means that the system
will have a predetermined, necessary ``motion'' that will not be
affected by changes, however drastic or sudden. In other words,
the system will be a Boulding's \emph{Thermostat}. In what follow we describe
a strategy that can be used to enhance the Boulding category of a
replication and voting scheme, thus avoiding a case of the \hbox{$\hbox{S}_B$}.
The strategy we propose is to isolate redundancy management at
architectural level, and to use an autonomic computing scheme
to adjust it automatically.
In what follows we describe a possible implementation
for this run-time strategy.
\begin{itemize}
\item First, we assume that the replication-and-voting service
is available through an interface similar to the one of the
Voting Farm~\cite{DeDL98a}.
Such service sets up a so-called
``restoring organ''~\cite{John89a} after the user supplied the
number of replicas and
the method to replicate.
\item Secondly, we assume that the number
of replicas is not the result of
a fixed assumption but rather an initial value possibly subjected to revisions.
Revisions are triggered by secure messages that ask to raise or
lower the current number of replicas.
\item Third, we assume a middleware component such as our
Reflective Switchboards~\cite{DF10a} to be available. Such middleware
deducts and publishes a measure of the current environmental disturbances.
In our prototypical implementation, this is done by computing,
after each voting, the
``distance-to-failure'' (dtof), defined as
\[
\hbox{dtof}(n,m) = \lceil \frac n2 \rceil - m,
\]
where $n$ is the current number of replicas and $m$ is the amount
of votes
that differ from the majority, if any such majority exists.
If no majority can be found dtof returns 0.
As can be easily seen,
dtof returns an integer in $[0, \lceil \frac n2 \rceil]$ that
represents how close we were to failure at the end of the last voting round.
The maximum distance is reached when there is full consensus among
the replicas. On the contrary the larger the dissent, the smaller
is the value returned by dtof, and the closer we are to
the failure of the voting scheme.
In other words, a large dissent (that is, small values
for dtof) is interpreted as
a symptom that the current amount of redundancy employed is
not large enough.
Figure~\ref{f:dtof}
depicts some examples when the number of replicas is 7.
\item When dtof is critically low, the Reflective Switchboards
request the replication system to increase the number of redundant replicas.
\item When dtof is high for a certain amount of consecutive runs---1000 runs
in our experiments---a request to lower the number of replicas is issued.
Figure~\ref{f:fromadidadr} shows how redundancy varies in
correspondence of simulated environmental changes.
\end{itemize}
Function dtof is just one possible example of how to estimate the
chance of an impending assumption failure when dimensioning
redundant resources.
Our experiments~\cite{DF10a} show that
even such a simplistic scheme allows most if not all dimensioning
assumption failures to be avoided. Despite heavy and diversified
fault injection, no clashes were observed during our experiments.
At the same time, as a side effect of assumption failure avoidance,
our autonomic scheme reduces the amount of redundant resources to be allocated and
managed. This can be seen for instance in
Fig.~\ref{f:redundyn} which plots in logarithmic scale the distribution of the
amount of redundancy employed by our scheme during one of our experiments.
\begin{figure}
\centerline{\includegraphics[width=0.8\textwidth]{dtof2.png}}
\caption{Distance-to-failure in a replication-and-voting scheme
with 7 replicas. In (a), consensus is reached, which corresponds
to the farthest ``distance'' to failure. From (b) to (d),
more and more votes dissent from the majority (red circles) and correspondingly
the distance shrinks. In (d), no majority can be found---thus,
failure is reached.}
\label{f:dtof}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=0.7\textwidth]{adapredun.png}}
\caption{During a simulated experiment, faults are injected,
and consequently distance-to-failure decreases. This triggers an autonomic
adaptation of the degree of redundancy.}
\label{f:fromadidadr}
\end{figure}
\begin{figure}[t]
\noindent\input histogram.tex
\caption{Histogram of the employed redundancy during an experiment
that lasted 65 million simulated
time steps. For each degree of
redundancy $r$ (in this case $r\in \{3,5,7,9\}$) the graph displays the
total amount of time steps the system adopted assumption
$a(r)$: \{``Degree of employed redundancy is $r$''\}.
A logarithmic scale is used
for time steps. Despite
fault injection, in the reported experiment the system spends
99.92798\% of its execution time
making use of the minimal degree of redundancy,
namely 3, without incurring in failures.}
\label{f:redundyn}
\end{figure}
The above strategy shows how \hbox{$\hbox{S}_H$}{} and \hbox{$\hbox{S}_B$}{} may be
avoided---in a special case---by creating context-aware,
autonomically changing Horning Assumptions. In other
words, rather than postponing the decision of the value
to bind our assumption to, here we embedded our software
system in a very simple autonomic architecture that dynamically revises
dimensioning assumptions. The resulting system complies to
Boulding's categories of ``Cells'' and ``Plants'', i.e.
open software systems with a self-maintaining structure~\cite{Bou56}.
\section{Related Technologies}\label{s:reltech}
As mentioned in the introduction, several
conceptual and
practical tools are available to deal to some extent
with problems related to assumption failures. Such tools may be
situated in one or more of the following ``time stages'': design-time,
verification-time, compile-time, deployment-time, and run-time.
In what follows we briefly discuss some families of those
tools pointing out their relations with the subjects
treated in this paper.
\begin{description}
\item[Verification and validation activities\mbox{,}] i.e.,
checking and certification of compliance to specifications,
are a fundamental tool to verify and prove the absence of
some context clashes.
In particular re-qualification is an activity prescribed
each time a system (not necessarily a software system)
is ``relocated'' (e.g. reused, or ported);
or in case of replacement of some of its parts; or when
a system is connected to another system.
We observe how, verification and validation being
in general off-line activities, assumptions
are matched against a reference context information
(the hypothized truth)
that might differ from the actual context---from ``real life'',
as it were.
Particularly interesting is the family of techniques known as
formal verification, which make use of formal (mathematical)
methods to assess a system's properties. Properties
are described through a formal specification.
Formal specification languages, such as the
Z notation~\cite{z:baum95,z:bish90}, can be used for
the non-ambiguous expression of software properties.
Compliant tools can then verify the validity of those
properties and detect cases of assumption failures.
Semantics~\cite{DBLP:journals/cacm/ShethVG06}
is another family of techniques that
aim at expressing and machine-verifying the meanings of
computing systems, their processing and environments.
\item[Unified Modeling Language] (UML) is the de-facto modeling
language for object-oriented software engineering.
A discussion of UML would lead
us astray, thus we shall just remark here how
UML provides means to describe
\begin{itemize}
\item the dependencies among the modeled software parts via component diagrams;
\item the mapping of software parts onto the target hardware and execution environment
via deployment diagrams;
\item assorted knowledge, in the form of annotations;
\item rules and constraints, as for instance in the contexts and properties of
the Object Constraint Language~\cite{OCL}.
\end{itemize}
UML and related tools situate themselves at design level though
can be used to generate implementation artifacts directly from
the models. By creating a stronger link between design-time
and other ``time stages'' such tools---when used correctly---make
it more difficult to incur in cases of the \hbox{$\hbox{S}_{{H\!I}}$}{} that are
due to model-vs.-code inconsistencies. We observe how the
produced artifacts are static entities that strictly follow
design-time rules; as such they are not able to self-adapt
so as to withstand faults or re-optimize in view of changed
conditions. In other words, those implementation artifacts
may suffer from the \hbox{$\hbox{S}_B$}.
\item[Design by Contract]~\cite{Mey92} is a design approach that systematically
deals with the mutual dependences of cooperating software components.
Depending on the context, any two software components may find
themselves in the role of a client and of a supplier of some service.
A well-defined ``contract'' formally specifies what are the obligations
and benefits of the two parties. This is expressed in terms of
pre-conditions, post-conditions, and invariants. Design by Contract
forces the designer to consider explicitly the mutual dependencies
and assumptions among correlated software components.
This facilitates assumption failures detection and---to some extent---treatment.
The concept of contracts has been recently successfully applied to security
of mobile applications~\cite{DMNS07,mw09.12}.
\item[Web Services standards] provide numerous examples
of specifications to expose, manage, and control
capabilities and features of web services architectures.
It is worth highlighting here a few of these standards:
\begin{description}
\item[WSDL] (Web Services Description Language) is an XML language
that allows a client to query and invoke the services exported by
any third-party web service on the Internet. This high degree
of flexibility exacerbates the problem of depending on third
party software components, i.e., software of unknown characteristics
and quality~\cite{Gre97}. The need to discipline this
potential chaos brought to a number of other specifications, such as
WS-Policy.
\item[WS-Policy] implements a sort of XML-based run-time version
of Design by Contract: using WS-Policy web service suppliers can
advertise their pre-conditions (expected requirements, e.g.
related to security), post-conditions (expected state evolutions),
and invariants (expected stable states).
\item[WSDM] (Web Services Distributed Management)
and its two complementary specifications MUWS
(Management Using Web Services) and MOWS (Management Of Web Services),
which respectively
expose manageability capabilities and define a monitoring
and control model for Web Services resources. This
allows for instance quality-of-service monitorings,
enforcing a service level agreement, or controlling a task.
\end{description}
\item[XML-based deployment descriptors] typical
of service-oriented and comp\-onent-oriented
middleware platforms such as J2EE or CORBA are
meant to reduce the complexity of deployment
especially in large-scale distributed systems.
Their main focus is clearly deployment-time.
Despite their widely recognized values,
some authors observe that they exhibit a
``semantic gap''~\cite{Sch06} between the design intent
and their verbose and visually dense syntax, which in practice
risks to conceal the very knowledge they are intended
to expose. This is probably not so relevant as the
exposed knowledge is meant to be reasoned upon by
machines.
\item[Introspection.]
The idea of introspection
is to gain access into the hidden software complexity, to inspect the
black-box structure of programs, and to interpret their
meaning through semantic processing, the same way the
Semantic Web promises to accomplish with the data scattered through
the Internet. Quoting~\cite{Intros}, ``introspection is a means
that, when applied correctly, can help crack the code of a software
and intercept the hidden and encapsulated meaning of the
internals of a program''. Introspection
is achieved e.g. by instrumenting software with data collectors producing
information available in a form allowing semantic processing,
such as RDF\cite{RDF}. This idea is being used in the Introspector
project, which aims at instrumenting the GNU programming
tool-chain so as to create a sort of semantic web of all
software derived from those tools. The ultimate goal is
very ambitious: ``To create a super large and extremely dense web of
information about the outside world
extracted automatically from computer language
programs''~\cite{Intros}. This would allow the design of software able
to reason about the dependability characteristics of other
software. Tools based on introspection include:
\begin{description}
\item[GASTA] (Gcc Abstract Syntax Tree Analysis)~\cite{gasta}, which uses introspection to
automatically annotate C code to analyze the presence of null pointer design faults),
\item[GCC-XML]~\cite{gccxml}, quite similar to GASTA, and
\item[XOGASTAN] (XML-Oriented Gcc Abstract Syntax Tree ANalyzer)~\cite{ADM04},
which uses the abstract syntax tree produced by the GNU compiler while processing a C file
and translates it into XML. Another of the XOGASTAN tools can then read the XML file and analyze it.
\end{description}
In its current form introspection is an off-line technique working at code level.
\item[Aspect orientation] logically
distinguishes a conventional language to encode the functional logics;
an aspect language to define
specific interconnections among a program's basic functional units; and
a so-called aspect weaver, that is a program that composes a
software system from both the functional and the aspect logics.
Multiple aspect logics can be defined to address different
systemic cross-cutting concerns, e.g. enhancing dependability, minimizing
energy expenditure, or increasing performance. This has two
consequences particularly important for our treatise: the most
obvious one is that aspect oriented languages realize pliable
software that can be more easily maintained and adapted. Secondly,
aspects encode knowledge that regard specific ``viewpoints'', and
encourage the designers doing so.
As such, aspect orientation offers a conceptual and practical
framework to deal with the three syndromes of software development.
\item[Model Driven Engineering] (MDE) is a relatively new paradigm that
combines a number of the above approaches into a set of
conceptual and practical tools that address several shortcomings
of traditional methods of software development. In particular,
MDE recognizes that ``models alone are insufficient to
develop complex systems''~\cite{Sch06}.
Contrarily to
other approaches, which develop general ``languages'' to express
software models in an abstract way, MDE employs so-called
domain-specific modeling languages, which make use of semantics to
precisely characterize the relationships between concepts and their
associated constraints. The ability to express
domain-specific constraints and to apply model checking
allows
several cases of assumption failures to be detected
early in the software life cycle.
Furthermore, MDE features
transformation engines and generators that synthesize
from the models
various types of artifacts, e.g. source code and
XML deployment descriptions.
MDE systematically combines several existing technologies
and promises to become soon one of the most important ``tools''
to tame the ever growing complexity of software.
As remarked by Schmidt~\cite{Sch06}, the elegance
and the potential power of MDE brought about many expectations;
this notwithstanding,
scientific studies about
the true potential of MDE are still missing~\cite{Bez03,Bez05,Sch06}.
\end{description}
\section{Lessons Learned and Vision}\label{s:vision}
In previous section we discussed very concisely a few families of
approaches that can be effectively used to solve
some of the problems we introduced in this paper. Here we first summarize lessons
learned while doing so, which then brings us to our vision on future
approaches and architectures to deal effectively with assumption failures.
First of all, we can summarize that a number of
powerful techniques and tools exist or are in the course
of being honed that can effectively help dealing
with assumption failures. What is also quite apparent is
that each of them only tackles specific aspects of the
problem and takes a privileged view to it.
Our position in this context is that
we are still lacking methodologies and architectures to tackle this
problem in its complex entirety. Fragmented views to this
very complex and entangled web are inherently ineffective
or at best sub-optimal. Missing one aspect means leaving a
backdoor open to the manifestations of the three syndromes introduced
in Sect.~\ref{s:hazards}.
In other words, a holistic approach is required. Taming
the complexity of software systems so as to reach true resilience
in the face of assumption failures
requires a unitary view to the whole of the ``time stages'' of
software development---what the General Systems Theory calls
a \emph{gestalt}~\cite{Bou56}. We believe
one such gestalt for software systems to be the concept of
assumption failure. As Boulding writes in the cited paper,
gestalts are ``of great value in directing research towards
the gaps which they reveal''---in the case at hand,
the gaps of each fragmented view offered by the
approaches mentioned in Sect.~\ref{s:reltech} to the
problems discussed in this paper\footnote{In
the cited paper Boulding applies this concept to the general system of disciplines
and theories: ``Each discipline corresponds to a certain segment of the empirical
world, and each develops theories which have particular applicability to its own
empirical segment. Physics, chemistry, biology, psychology, sociology, economics
and so on all carve out for themselves certain elements of the experience of man
and develop theories and patterns of activity (research) which yield satisfaction
in understanding, and which are appropriate to their special segments.''
Gestalts, that is meta-theories of systems, ``might be of value in directing the
attention of theorists toward gaps in theoretical models, and might even be of
value in pointing towards methods of filling them.''}.
In a sense, most if not all of those approaches may be regarded
as the result of an attempt to divide and
conquer the complexity of software development by abstracting and
specializing (that is, reducing the scope of) methods, tools,
and approaches. This specialization
ends up in the ultimate case of the
Hidden Intelligence syndrome. A better approach would probably
be considering the unity of the design intent and using
a holistic, ``cross layered'' approach to share
sensible knowledge unraveled in one layer and feed it back
into the others.
We envision a general systems theory of
software development in
which the model,
compile-, deployment-, and run-time layers
feed one another with deductions and control ``knobs''.
The strategies discussed in this paper could provide
the designer with useful tools to arrange such cross-layering
processes.
This would allow knowledge
slipping from one layer to be still caught in another,
and knowledge gathered
in one layer to be fed back into others. As an example,
the strategy discussed in Sect.~\ref{s:pattern}
could feed an MDE tool whose deductions could in turn be published
or reified into a context-aware middleware such as our Reflective
Switchboards~\cite{DF10a}.
One way to achieve this could be to arrange a web of
cooperating reactive agents serving different software design
concerns (e.g. model-specific, deployment-specific, verification-specific,
execution-specific)
responding to external stimuli and
autonomically adjusting their internal state. Thus a design assumption
failure caught by a run-time detector should trigger a request for
adaptation at model level, and vice-versa. We believe that
such a holistic approach would realize
a more complete, unitary vision of a system's behavior and properties
with respect to the sum of the detached and fragmented views
available so far.
\section{Conclusions}\label{s:end}
Software systems are characterized by predefined assumptions about
their intended platform, their internal state, and the environments they are
meant to be deployed in. They are often closed, ``blind'' systems built from
synchronous assumptions and designed so as to be plugged in immutable
hardware systems and environments whose changes, idiosyncrasies, or
fluctuations most of them deliberately ignore. We believe that this
approach to software development is not valid anymore. Software ought to
be designed and executed taking into account the inevitable occurrence
of potentially significant and sudden changes or failures in their
infrastructure and surrounding environments. By analyzing well-known
software failures we identified three main threats to effective
dependable software engineering, which we called the Hidden Intelligence
syndrome, the Horning syndrome, and the Boulding syndrome. In this paper
we expressed our thesis that services explicitly addressing those
threats and requirements are an important ingredient towards truly
resilient software architectures. For each of the above mentioned
syndromes we also provided exemplary treatment strategies, which form
the core of our current work in the adaptive-and-dependable software
systems task force of the PATS research group at the University of
Antwerp. The key idea is to provide the designer with the ability to
formulate dynamic assumptions (assumption variables) whose boundings get
postponed at a later, more appropriate, time: at compile time, when we
are dealing with hardware platform assumptions for a stationary code; at
deployment time, when the application can be assembled on that stage;
and at run-time, e.g. when a change in the physical environment calls
for adaptation to new environmental conditions. We believe that
an effective way to do this is by means of a web of
cooperating autonomic ``agents''
deducting and sharing knowledge, e.g. the type of faults being
experienced or the current values for properties regarding
the hardware platform and the execution environment.
We developed a number of these agents, e.g. Reflective
Switchboards~\cite{DF10a}, ACCADA~\cite{GuD10+}, and an Apache Axis2/MUSE
web service framework.
Our future steps include the design of a software architecture
for assumptions failure treatment based on the close cooperation
of those and other building blocks.
We conclude by observing how our research actually ``stands on the
shoulders of giants'', as its conclusions closely follow those in the
now classic 1956 paper of Kenneth Boulding on General Systems
Theory~\cite{Bou56}: indeed, current software engineering practices
often still produce systems belonging to Boulding's categories of
``Clockworks'' and ``Thermostats''. The root assumption of such systems
is their being closed-world, context-agnostic systems characterized by
predefined assumptions about their platform, their internal state, and
the environment they are meant to be deployed in, which makes them
fragile to change.
On the contrary, the unitary approach we envision,
based on the proposed role of gestalt for assumption failures,
would make it
possible to design and maintain actual open software systems with a
self-maintaining structure (known as ``Cells'' and ``Plants'' according
to Boulding's terminology) and pave the way to the design of
fully autonomically resilient software systems (Boulding's ``Beings'').
\subsection*{Acknowledments}
My gratitude goes to Jonas Buys, Ning Gui, and Hong Sun for providing
me with their helpful comments, suggestions, and valuable assistance, and
to Marc and Mieke Leeman for kindly providing me with the output from the
execution of the \texttt{lshw} command on several computers.
\bibliographystyle{splncs}
|
\section{Introduction}
In his seminal works in electromagnetic theory, O. Heaviside developed formal rules for dealing with the differentiation operator, much of which he arrived at intuitively. A mathematical basis for his operational calculus was done by mathematicians later with the aid of the Laplace transform.
A different explanation of Heaviside's calculus, based upon
the convolution ring of continuous functions, was presented by Mikusinski \cite{m}.
From this ring, which is commutative and without zero divizors, Mikusinski goes on to define the field of fractions. These fractions he calls operators.
In this note, we modify Mikusinski's approach by starting with the module structure of continuous functions over the ring of convergent formal series (say, in $t$). This module is {\em torsion free}. Imitating Mikusinski, we apply the fraction space construction and obtain this way a linear space over the field of convergent Laurent series. Elements of this space, which we call Mikusinski functions, constitute a very small part of
the
field of Mikusinski operators. Viewing the space of these functions as a module over the polynomial ring
(in $s=t^{-1}$)
and taking the factor module by "derivatives" of constant functions, we get a module that is canonically isomorphic
to the module ${\cal D}^\prime_{fin}$, the module of Schwartz distributions of finite order.
The goal of this note is to demonstrate that with such a representation of ${\cal D}^\prime_{fin}$ we have a natural, simple and rigorous theory for Heaviside's operational calculus.
Throughout, $I$ is a fixed interval of real axis containing $0$ and $t$ an indeterminate. The symbol $\bf{1}$ will stand for the unit function on $I$.
Let $C(I)$ be the space of all complex-valued continuous functions defined on the interval $I$ and let $\mathbb{C}\{t\}$ be
the ring of complex coefficient convergent formal
series in $t$.
(We remind that a formal series $\sum_{i\geq 0}b_it^i$ is said to be convergent if $\sum_{i\geq 0}|b_i|\varepsilon^i < +\infty$
for some positive real $\varepsilon$.) Let $\mathbb{C}(\{t\})$ denote the fraction field of $\mathbb{C}\{t\}$, the field of convergent Laurent series.
It is worth mentioning that $\mathbb{C}(\{t\})$ contains as a subring the polynomial ring $\mathbb{C}[s]$, where $s=t^{-1}$.
For every continuous function $u\in C(I)$, let $J(u)$ denote the function defined by
$$
J(u)(x)=\int_0^xu(\alpha)d\alpha,\ \ \ x\in I.
$$
\section{Mikusinski functions}
If
$g=\sum_{i\geq 0}b_it^i$ is a convergent formal series and $u$ is a continuous function on $I$, then the series
$$
\sum_{n\geq 0} b_nJ^n (u)
$$
converges uniformly on every compact subinterval of $I$ that contains $0$. We therefore can define
the product $gu$ by setting
$$
gu=\sum_{n\geq 0} b_nJ^n (u).
$$
This multiplication makes $C(I)$ a {\em module} over $\mathbb{C}\{t\}$.
(A module is a "linear space" over a ring!) It is easily seen that this module is torsion free. Indeed, suppose that $gu=0$, where $g\neq 0$. Write $g=t^ng_0$, where $g_0$ has nonzero
free coefficient. By the theorem on units (see \cite{r}), $g_0$ is invertible. Multiplying $gu=0$ by $g_0^{-1}$, we get
that $t^nu=0$. Because $J$ is an injective operator, it follows that $u=0$.
{\bf Definition}. We define the Mikusinski space $M(I)$
as the fraction space of the module $C(I)$.
By definition, $M(I)$ is a linear space over $\mathbb{C}(\{t\})$, and every Mikusinski function is represented by a ratio $ u/g$ with $u\in C(I)$ and $g\in \mathbb{C}\{t\},\ g\neq 0$.
Two such ratios $u_1/g_1$ and $u_2/g_2$ represent the same Mikusinski function if
$$
g_2u_1=g_1u_2.
$$
Because $C(I)$ is torsion free, the canonical map $u \mapsto u/1$ is injective. Therefore we can make the identification
$$
u=u/1.
$$
With this identification,
a Mikusinski function $w$ can be written in the form
$$w=s^nu,$$
where $u\in C(I)$.
Indeed, if $w=u/g$, then, writting $g=t^ng_0$ with $g_0$ as above, we have
$$\frac{u}{g}=s^n\frac{g_0^{-1}u}{1}=s^n(g_0^{-1}u).$$
We therefore have:
$$
C(I)\subseteq sC(I)\subseteq s^2C(I)\subseteq s^3C(I)\subseteq \cdots \ \ \tx{and}\ \ M(I)=\cup_{n\geq 0} s^nC(I).
$$
{\em Remark}. If one wants, one can define the Mikusinski space $M(I)$ as the inductive limit
of the sequence
$$
C(I)\stackrel{J}{\ra} C(I) \stackrel{J}{\ra}C(I) \stackrel{J}{\ra} C(I)\stackrel{J}{\ra} \cdots .
$$
Extend the operator $J$ to Mikusinski functions by setting
$$
J(w)=tw.
$$
This way, the integration operator becomes bijective.
The following theorem says that the space $M(I)$ is the smallest extension of $C(I)$
in which every function is an integral.
\begin{thm} Every continuous function is an $n$-fold integral of a Mikusinski function.
\end{thm}
{\em Proof}. This is obvious. Indeed, if $u\in C(I)$, then
$$
u=t^n(s^nu)=J^n(s^nu).
$$
$\quad\Box$
\section{Schwartz distributions (of finite order)}
Since multiplication by $t$ is the integration operator, a natural idea is to interpret multiplication by $s=t^{-1}$ as the differentiation operator. Postulating that constant functions must have zero derivatives,
we are led to the following definition.
{\bf Definition}. Define the generalized function space $G(I)$ to be
$$
G(I)\ =\ M(I)/N(I),
$$
where $N(I)=\{f{\bf 1}\ |\ f\in s\mathbb{C}[s]\}$.
Since $N(I)$ is a $\mathbb{C}[s]$-submodule of $M(I)$, $G(I)$ has a structure of
a module over $\mathbb{C}[s]$.
In particular, we can multiply generalized functions by $s$.
{\bf Definition}. Define the differential operator $D: G(I)\ra G(I)$ by setting
$$
D(\xi)=s\xi,\ \ \ \xi\in G(I).
$$
We need the following lemma.
\begin{lem} \
$
C(I) \cap N(I) = \{0\}.
$
\end{lem}
{\em Proof}. Assume there is $u\in C(I)$ that is not zero and belongs to $N(I)$. We
then have
$$
u = (a_0s^n + \cdots + a_{n-1}s){\bf 1}
$$
with $n\geq 1$, $a_i\in \mathbb{C}$ and $a_0\neq 0$. Multiplying both of the sides by $t^{n+1}$, we get
$$
t^{n+1}u = (a_0t+ \cdots + a_{n-1}t^n){\bf 1}.
$$
It follows that
$$
t(t^nu - (a_1+ \cdots + a_{n-1}t^{n-1}){\bf 1}) = a_0{\bf 1}.
$$
On the left, we have a continuous function having value 0 at 0. Hence, $a_0=0$, which is a contradiction.
$\quad\Box$
It follows from the previous lemma that the canonical map $C(I)\ra G(I)$ defined by
$$
u \mapsto u\ (\tx{mod}\ N(I))
$$
is injective. This permits us to identify $C(I)$ with its image under this map. In other words, for every
$u\in C(I)$, we can make the identification
$$
u= u\ (\tx{mod}\ N(I)).
$$
Notice that the derivative $D(u)$ of a classical continuously differentiable function $u$ coincides with the ordinary derivative $u^\prime$.
Indeed, by the Newton-Leibniz formula
$u=Ju^\prime + u(0){\bf 1}$,
and consequently
$$
su\ (\tx{mod}\ N(I)) =(u^\prime +su(0){\bf 1})\ (\tx{mod}\ N(I)) =u^\prime (\tx{mod}\ N(I))=u^\prime.
$$
The following theorem says that the space $G(I)$ is the smallest extension of $C(I)$ in which every function has a derivative.
\begin{thm} Any generalized function is an $n$-fold derivative of a continuous function.
\end{thm}
{\em Proof}. This is obvious. Indeed, for every continuous function $u$ and every nonnegative integer $n$, we have
$$
s^nu\ (\tx{mod}\ N(I))=D^n (u\ (\tx{mod}\ N(I)))= D^nu.
$$ $\quad\Box$
As an immediate consequence, we have the following corollary.
\begin{cor}
$G(I)$ is canonically isomorphic to ${\cal D}^\prime_{fin}(I)$, the space of Schwartz distributions of finite order.
\end{cor}
This corollary permits us to interpret generalized functions as Schwartz distributions of finite order.
{\em Remark}. We remind that, basically, distributions have been introduced by Schwartz in order to be able to differentiate all continuous functions
(see page 72 in Schwartz \cite{s}).
We close the section with the following important theorem.
\begin{thm} Let $f\in \mathbb{C}[s]$ be a polynomial. The equation
$$
f(D)\xi=\omega
$$
has a solution $\xi\in G(I)$ for every $\omega\in G(I)$.
\end{thm}
{\em Proof}. This is obvious. Indeed, $M(I)$ is a divisible $\mathbb{C}[s]$-module, and consequently $G(I)$ also is divisible.
To be more precise, if $\omega$ is represented by a Mikusinski function $w$, then $$\xi=\frac{1}{f}w \ \tx{mod} N(I)$$ is a solution.
$\quad\Box$
\section{Linear constant coefficient differential equations}
We already know how to find a particular solution of a nonhomogeneous linear constant coefficient differential equation.
So, we only need to consider the homogeneous case.
Given $g\in \mathbb{C}\{t\}$, define an entire analytic function $E(g)$ by setting $$E(g)=g{\bf 1}.$$
\begin{thm} Let $f\in \mathbb{C}[s]$ be a polynomial of degree $d$. The solutions of the linear differential equation
$$
f(D)\xi=0,\ \ \ \ \xi\in G(I)
$$
are given by the formula
$$
\xi=E(\frac{sr}{f}),
$$
where $r$ runs over the polynomials in $\mathbb{C}[s]$ of degree $\leq d-1$.
\end{thm}
{\em Proof}. It is easily seen
that all these functions are solutions. Indeed,
$$
f(D)\xi=f\frac{sr}{f}{\bf 1}=sr {\bf 1}=0.
$$
To show the converse, assume that $\xi$ is a solution of our equation and assume that a Mikusinski function $s^nx$ represents it. Then
$$
fs^nx\in N(I).
$$
This means that there exist complex numbers $a_0, a_1, \ldots , a_k$ such that
$$
fs^nx=s(a_0+a_1s+ \cdots + a_ks^k){\bf 1}.
$$
Multiplying both sides by $t^{d+n}$, from this we get
$$
(ft^d)x=(a_0t^{d+n-1}+\dots +a_{d+n-1}){\bf 1} + (a_{d+n}s+\cdots ){\bf 1}.
$$
Since the left side is an ordinary continuous function, we can see that $a_k=0$ for all $k\geq d+n$.
We therefore have
$$
s^nx=\frac{s(a_0+a_1s+ \cdots + a_{d+n-1}s^{d+n-1})}{f}{\bf 1}.
$$
By the Euclidean division theorem, there exists a polynomial $r$ of degree $\leq d-1$ such that
$$
\frac{a_0+a_1s+ \cdots + a_{d+n-1}s^{d+n-1}}{f} \equiv \frac{r}{f} \ \tx{mod}\ \mathbb{C}[s].
$$
So that
$$
\xi=E(\frac{sr}{f}).
$$
The proof is complete. $\quad\Box$
|
\section{Introduction}
For centuries, it has been known that the shape of an object influences the polarization state of reflected light.\footnote{Augustin-Jean Fresnel (1788-1827).} This principle underlies the Shape from Polarization (SfP) technique, which aims to recover the surface normals of an object from three polarized photos.
Classical approaches to SfP rely on specular reflections from an object (hereafter, specular SfP). In an effort to handle purely diffuse surfaces, Atkinson and Hancock introduced a landmark result, modifying the physical model to account for cases where all the light is diffusely reflected (hereafter, diffuse SfP)~\cite{atkinson2006recovery}. However, many surfaces exhibit properties that are neither diffuse nor specular, but somewhere in-between. A ``mixed reflection'' occurs: both diffusely and specularly reflected light return to the camera causing model mismatch.
Obtaining surface normals through polarization is mostly a laboratory problem, with several practical challenges. For example, one needs to know the refractive index of the material; the material must be either diffuse or specular; and ill-posed ambiguities exist for both zenith and azimuth angles. Recent work has used a coarse depth map to provide what may be a promising step toward ``in-the-wild'' uses of SfP~\cite{kadambi2015polarized} (hereafter, ``Polarized 3D''). While Polarized 3D has demonstrated compelling results, we believe our work offers complementary benefits.
At the heart of our work is an analysis of mixed reflections and their impact on existing techniques that use SfP. We find that, indeed, a mixed reflection perturbs the result to the point where correction is desirable. We therefore propose a physics-based technique to correct for mixed reflections using multiple viewpoints of an object, demonstrating the practical benefits of our approach through comparisons with previous work.
\paragraph{Scope:} Our contribution of extending SfP to handle mixed surfaces is a unified approach. Prior art has proposed a sequential approach, where the scene is first split into diffuse and specular components, following which the appropriate SfP algorithm can be used. For example, the work of Miyazaki et al.~\cite{miyazaki2003polarization} handles mixed surfaces by first using an algorithm for diffuse-specular separation, proposed by Tan and Ikeuchi~\cite{tan2005separating}, following which standard technique of SfP are used. Since the Tan and Ikeuchi technique is very general, i.e, it is not specific to polarization, we believe that the information from the Fresnel equations could be used to improve on previous work.
In this paper, we develop an approach that incorporates the SfP model to aid in separating the image into diffuse and specular components. We also show that our proposed approach allows simultaneous recovery of refractive index, while outperforming sequential approaches.
\begin{figure}[t]%
\centering
\includegraphics[width=1\columnwidth]{images/refl_cartoon/refl_cartoon}%
\caption{(a) Unpolarized light strikes a surface -- the specular reflection is partially polarized in accordance with the Fresnel equations. (b) The Fresnel equations were recently extended to the case of diffuse reflections by Atkinson and Hancock~\cite{atkinson2006recovery}, where it was assumed that the light emanating from the object is transmitted from an internal reflection. (c) Mixed reflections are the subject of this paper.}%
\label{fig:refl}%
\end{figure}
\section{Related Work}
In the context of related work, we believe our proposed technique is the first unified approach toward joint estimation of shape, diffuse-specular separation, and pixel-wise refractive index.
\paragraph{The Fresnel Equations} describe the behavior of electromagnetic radiation as it interacts with a surface. When light interacts with a surface, the first-order event that occurs is reflection and transmission at the boundary (Figure \ref{fig:refl}a).\footnote{``First-order'' refers to the behavior of light at the first transition between media. Light scattering through heterogeneous media, as in Figure \ref{fig:refl}b is a higher-order phenomena.} The Fresnel equations relate the angles of reflection (e.g., the zenith component of surface normal) with the refractive index of the medium as well as polarimetric properties.
\paragraph{Shape from Polarization (SfP)} is the term used in computer vision for a technique that estimates surface normals using the principles of the Fresnel equations. Classical SfP requires measurement of the polarimetric properties (through 3 polarized photos) and estimation of the refractive index to solve for the angle of reflection. We consider ourself with two primary branches of the SfP technique: first-order and higher-order. The first-order SfP techniques assume the reflection model akin in Figure \ref{fig:refl}a. Following this model has allowed shape estimation of
metals~\cite{morel2005polarization}, transparent objects~\cite{saito2001measurement,miyazaki2004transparent}, dark objects~\cite{miyazaki2012polarization}, and even ocean waves~\cite{zappa2008retrieval}. Higher-order SfP techniques rely on multiple interactions of light with a medium, as in the case of Figure~\ref{fig:refl}b. In such case, the Fresnel equations are applied differently,
allowing for shape recovery of diffuse, subscattering surfaces. This is described with compelling experimental support by Atkinson and Hancock~\cite{atkinson2006recovery}.
\paragraph{Diffuse-specular separation} refers to a broad class of computational and optical techniques to decompose an image into a specular-only image and diffuse-only image. The most general techniques use only image or color information, but these can be susceptible to artifacts. For example, Nishino et al. introduced a technique that uses view-independent effects to identify the diffuse reflection in an image~\cite{nishino2001determining}. Other strategies combine image-based measurements with color analysis~\cite{nishino2001determining,lin2001separation,lin2002diffuse,tan2004separating,tan2005separating}. In this paper we show that it is beneficial to leverage the behavior of polarization to perform this separation. Previous attempts have used polarization to separate diffuse and specular components in the context of active illumination. In particular, spherical gradient illumination has been used by Ma et al.~\cite{ma2007rapid} and Ghosh et al.~\cite{ghosh2011multiview} to achieve photorealistic geometric reconstructions. For passive conditions, Nayar et al. introduced a separation technique that uses polarization images and color cues~\cite{Nayar97}. While successful, the Nayar method, and a related method proposed by Zickler et al~\cite{zickler2006reflectance} are limited by smoothness assumptions. In crux, though it is possible to directly combine existing work---for example, the SfP paper in~\cite{miyazaki2003polarization} uses the separation method from~\cite{tan2005separating}---we show that joint incorporation of reflection separation with SfP physics results is a beneficial strategy for addressing the mixed reflection problem in SfP.
\paragraph{Extended topics in polarization} that are tangentially relevant, but outside the direct scope of this paper are described in brief. While our paper considers linear polarization effects,
work from~\cite{guarnera2012estimating} demonstrates shape reconstructions using circular polarization. Polarization information need not only be used for shape: prior art has considered problems like image dehazing~\cite{schechner2001instant}, illumination multiplexing~\cite{cula2007polarization}, panoramas~\cite{Schechner05}, underwater scattering~\cite{Treibitz09} or 3D displays~\cite{Lanman11}. In comparison to these related works, our paper is specific to the SfP problem. Future work could use, for example, the descattering model of~\cite{schechner2001instant}, to possibly obtain shape in scattering environments.
\begin{figure}[t]%
\centering
\includegraphics[width=0.75\columnwidth]{images/curves/curves.pdf}%
\caption{Notation used for intensity components. As the polarizer angle is rotated, the image intensity varies in accordance with Equation \ref{eq:formation}. Only the intensity quantities $I_{\max}$ and $I_{\min}$ can be directly measured with a camera. One of the aims of this work is to recover $I^{d}$ using the physical behavior of polarization, shape, and reflectance.}%
\label{fig:curves}%
\end{figure}
\section{Image Formation Model}
\newcommand{\mathbf{\phi}_{\text{pol}}}{\mathbf{\phi}_{\text{pol}}}
This section describes SfP in condensed form. Conceptually, SfP uses image-based measurements
to estimate the surface angles of azimuth, $\varphi$, and zenith, $\theta$. We will denote $\widehat \varphi$ and $\widehat \theta$ to represent
estimates of ground-truth quantities.
The measured irradiance at a single scene point
is expressed as
\begin{equation}
I\left( \mathbf{\phi}_{\text{pol}} \right) = \frac{{{I_{\max }} + {I_{\min }}}}{2} + \frac{{{I_{\max }} - {I_{\min }}}}{2}\cos \left( {2\left( \mathbf{\phi}_{\text{pol}} - \phi \right)} \right),
\label{eq:formation}
\end{equation}
where $\phi$ is the phase angle, and $I_{\max}$, $I_{\min}$ are the quantities shown in Fig. \ref{fig:curves}. Since a sinusoid has three unknowns, by sampling three different values of $\mathbf{\phi}_{\text{pol}}$ it is possible to estimate $\phi$, $I_{\max}$, and $I_{\min}$.\footnote{If it is difficult to rotate the polarizer precisely, the recent method from~\cite{polcalibrate} allows determination of the angles from image-based measurements.} From these measurements, as detailed below, separate mechanisms are used to obtain the azimuth or elevation angles.
\subsection{Challenges with estimating azimuth angle}
To obtain an estimate of the azimuthal angle, $\widehat \varphi$, early work in SfP has used a specular reflection model, as illustrated in Figure~\ref{fig:refl}a~\cite{wolff1991constraining,rahmann2001reconstruction,miyazaki2003polarization}. The maximum value of reflected light will occur when the light that reflects is perpendicularly polarized (since the Fresnel reflection coefficient for perpendicularly reflected light is greater). Then, since the maximum value of the cosine occurs at the origin, the azimuth angle is estimated as $\widehat \varphi = \phi$.
Atkinson and Hancock later introduced a compelling technique to recover object shape from a diffuse reflection model, as illustrated in Figure~\ref{fig:refl}b. For such a scenario, it was observed that the direction of light propagation is reversed: light is refracted from the surface to air~\cite{atkinson2006recovery}. Since the direction of propagation is flipped, the minimum irradiance is now of interest, resulting in a shift in the estimated azimuth angle of $\pi /2$ radians, i.e., $\widehat \varphi = \phi \pm \pi /2$.
Two key challenges occur with azimuthal estimation. First, since Equation \ref{eq:formation} includes a factor of 2 within the cosine, two azimuth angles, shifted apart by $\pi$ radians, cannot be distinguished in the polarized images.\footnote{Concretely, $\widehat \varphi$,and $\widehat \varphi + \pi$ return the same value for Equation \ref{eq:formation}} This first fundamental ambiguity is termed a \textbf{azimuthal ambiguity}, and applies to all SfP techniques. Second, for a general surface, not known \emph{a priori} to be diffuse or specular it is ambiguous as to whether the estimated angle should be shifted by $\pi /2$ radians or not. This second ambiguity, due to the surface reflectance, is termed as \textbf{azimuthal model mismatch}, and applies critically to our problem of mixed reflections.
\subsection{Challenges with estimating zenith angle}
Estimation of zenith angle relies on the degree of polarization, calculated as
\begin{equation}
\rho = \frac{I_{\max} - I_{\min}}{I_{\max} + I_{\min}}.
\label{eq:dop}
\end{equation}
As in the case of azimuthal estimation, the type of reflection influences the reflection model. First, consider the specular model in Figure \ref{fig:refl}a. Substituting the Fresnel equations (see~\cite{hecht2002optics}) into Equation \ref{eq:dop} allows the degree of polarization to be written as
\begin{equation}
\rho = \frac{2\sin\theta\tan\theta\sqrt{n^{2}-\sin^{2}\theta}}{n^{2}-2\sin^{2}\theta+\tan^{2}\theta}
\label{eq:rhospec}
\end{equation}
where $n$ denotes the refractive index and $\theta$ is the zenith angle. If some knowledge of $\rho$ and $n$ is obtained, then it is possible to solve Equation \ref{eq:rhospec} for an estimate of the zenith angle, $\widehat \theta$. This method is well-suited for highly specular objects and has been successfully used to estimate shape of metallic surfaces~\cite{morel2005polarization}.
For diffuse reflections, as illustrated in Figure~\ref{fig:refl}b, the Fresnel equations are once again combined with the degree of polarization. However, this is performed for the model where light is transmitted from the surface to air, such that the relation is now expressed as
\begin{equation}
\rho = \frac{{{{\left( {n - \frac{1}{n}} \right)}^2}{{\sin }^2}\theta }}{{2 + 2{n^2} - {{\left( {n + \frac{1}{n}} \right)}^2}{{\sin }^2}\theta + 4\cos \theta \sqrt {{n^2} - {{\sin }^2}\theta } }}.
\label{eq:rhodiff}
\end{equation}
This work addresses two problems with zenith estimation. First, it is difficult to obtain an estimate of refractive index, that is accurate, at each pixel. If an improper refractive index is used, an error in zenith estimation occurs, termed in previous work as \textbf{refractive distortion}~\cite{kadambi2015polarized}. Second, it can be hard to know whether to use the model for a specular surface (Equation~\ref{eq:rhospec}) or a diffuse surface (Equation~\ref{eq:rhodiff}). Only in ideal scenarios do surfaces conform to specular and diffuse models -- real-world surfaces exhibit mixed reflections. This second source of error is referred to as \textbf{zenith model mismatch} in this paper.
\section{Solving Model Mismatch}
To solve model mismatch error, consider the dichromatic reflection model, where the radiance from a single scene point is expressed as
\begin{equation}
I = I^d + I^s,
\end{equation}
where $I^d$ and $I^s$ refer to the radiant intensity of diffuse and specularly reflected light. The prior work in SfP calculates $\rho$ from $I_{\text{max}}$ and $I_{\text{min}}$, the maximum and minimum intensities observed when rotating the polarizer. Following the work of Nayar et al.~\cite{Nayar97}, it can be assumed that only the specular component causes appreciable variation, such that the measured degree of polarization for a mixed surface is expressed as
\begin{equation}
\rho = \frac{I_{\max}^{s} - I_{\min}^{s}}{I_{\max}^{s} + I_{\min}^{s} + I^{d}} = {\widetilde \rho} \frac{I^{s}}{I} = {\widetilde \rho} \frac{I - I^{d}}{I},
\label{eq:rhotilde}
\end{equation}
where $I_{\max}^{s}$ and $I_{\min}^{s}$ denote maximum and minimum irradiance observed from specularly reflected light (see Figure \ref{fig:curves}). Under this simplification, it is possible to substitute Equation \ref{eq:rhospec} into Equation \ref{eq:rhotilde} to express the diffuse intensity as
\begin{equation}
I^{d} = I(1- {\rho} \frac{n^{2}-2\sin^{2}\theta+\tan^{2}\theta}{2\sin\theta\tan\theta\sqrt{n^{2}-\sin^{2}\theta}}).
\label{eq:diff}
\end{equation}
This equation contains two unknowns: the intensity of diffuse reflection, $I^{d}$ and the refractive index $n$. Under a Lambertian approximation, the former quantity is constant across
different viewpoints. Additionally, since the refractive index is a physical property of the material, it is also constant across different viewpoints. The proposed strategy is to estimate the quantities on the right-hand-side of Equation~\ref{eq:diff} at different viewpoints, such that
\begin{equation}
{I^d} = f\left( {n,{\theta _i},{I_i},{ {\rho}_i}} \right) \triangleq I_i(1 - {\rho}_i \frac{{{n^2} - 2{{\sin }^2}\theta_i + {{\tan }^2}\theta_i }}{{2\sin \theta_i \tan \theta_i \sqrt {{n^2} - {{\sin }^2}\theta_i } }}),
\end{equation}
for the i-th view of $N$ total views. To recover $I^d$ a non-linear least squares problem can be solved of the form
\begin{equation}
\left\{ \widehat{I}^d, \widehat{n} \right\} = \argmin{I^d,n} \sum\limits_{i=1}^N (I_{d} - f(n, \theta_{i}, I_{i}, {\rho}_{i}))^2.
\end{equation}
In this paper a sequential quadratic program is used to perform the minimization. Please refer to the supplement for implementation details.
\begin{figure}[t]%
\centering
\includegraphics[width=0.97\columnwidth]{images/bunny/bunny.pdf}%
\caption{The proposed approach is able to separate reflectance for a variety of object textures (simulated example). Using the Mitsuba raytracer we render the Stanford bunny from three viewpoints, under three different material conditions (diffuse, glossy, and a spatially varying texture). The proposed technique is quantitatively compared with the previous work of~\cite{nishino2001determining}. By incorporating additional polarization, we demonstrate a significant reduction in error.}%
\label{fig:bunny}%
\vspace{-2mm}
\end{figure}
\section{Experimental Results}
Reflection models for SfP are not geared to handle mixed reflections. Existing solutions use a sequential approach: first, a robust algorithm is used to separate reflection components, following which SfP is performed. We provide a comparison to this sequential approach, using the multiview, reflection separation technique of Nishino et al.~\cite{nishino2001determining} as our point of comparison.
\paragraph{Implementation details:} All simulations were performed using the Mitsuba raytracer~\cite{jakob2010mitsuba}. The raytracer has been modified to acquire depth information and includes a Matlab script to simulate polarization measurements. The object remains static throughout all experiments as viewpoint diversity was acquired by moving the camera. Physical experiments
were performed with a Canon Rebel T3i camera with EF-S 18mm-55mm f/3.5-5.6 IS II SLR lens and a linear polarizer with quarter-wave plate, model Hoya CIR-PL. Three viewpoints were collected at 10 degree increments.
\paragraph{Diffuse-specular separation:} As shown in Figure \ref{fig:bunny}, the Stanford bunny is rendered with three different materials: clay, gloss, and a mixture of clay and gloss. Reflection separation is shown for the specular image component, for both the proposed technique and Nishino's method. Both techniques recover specular outliers, but the proposed technique recovers detail in regions that are of moderate specularity. As illustrated in the bar graph, the quantitative error is lower, for the proposed method, for all tested material configurations. Because the proposed
technique relies on viewpoint artifacts, classic artifacts like occlusions or lack of texture can lead to registration issues. This explains why our proposed method performs worse on the mixed material bunny (although the result is an improvement over Nishino's method).
\begin{figure}[t]%
\centering
\includegraphics[width=1.0\columnwidth]{images/sphere_sim/sphere_sim.pdf}%
\caption{A glossy sphere is neither diffuse nor specular (simulated rendering). (a) The ground truth surface normal, degree of polarization and zenith angle. (b) Without correction, the estimates of surface normal, degree of polarization and refractive index (see bar graph) are incorrect. (c) The proposed method recovers a more accurate normal map, degree of polarization and refractive index than (d) the previous method of Nishino et al.~\cite{nishino2001determining}.}%
\label{fig:spheresim}%
\vspace{-2mm}
\end{figure}
\paragraph{Surface normal recovery:} Figure~\ref{fig:spheresim} uses a rendered sphere to show that the recovered surface normals are not accurate using naive shape from polarization (Figure~\ref{fig:spheresim}b. Simple pre-processing with Nishino's method, as shown in Figure~\ref{fig:spheresim}d, does not allow for robust surface normal recovery. The Nishino method, as a general method that does not account for polarization information recovers a the degree of polarization that does not conform with the physical scene. This leads to a poor estimate of zenith angle, and ultimately, surface normals. In comparison, the proposed technique, shown in Figure~\ref{fig:spheresim}c, shows clear recovery of the surface normals, as well as the degree-of-polarization anzenith angle.
\paragraph{Refractive index recovery:} A benefit of the proposed technique is the ability to simultaneously recover per-pixel refractive index. The rendered sphere in Figure~\ref{fig:spheresim}
has a ground-truth refractive index of 1.5. The proposed technique estimates the refractive index as 1.49, for a mean absolute error of 0.01. Without applying the proposed correction for
mixed reflections, the error in refractive index estimation is 0.07. Interestingly, pre-processing with the Nishino method leads to a much greater error in estimating refractive index (0.18).
\begin{figure}[t]%
\centering
\includegraphics[width=0.99\columnwidth]{images/sphere_real/sphere_real.pdf}%
\caption{Validating the proposed technique with a physical experiment and comparisons to~\cite{nishino2001determining}. The uncorrect SfP result leaves something to be desired
as the normal map is noisy. The proposed correction algorithm reduces the MSE, while applying the technique from~\cite{nishino2001determining} increases the error.}%
\label{fig:spherereal}
\vspace{-2mm}%
\end{figure}
\paragraph{Physical experiment:} To validate our technique in the wild, a physical scene was set up in a similar fashion to the simulated examples. As illustrated in Figure~\ref{fig:spherereal},
a camera and polarizing filter are placed 50cm in front of a glossy sphere. Three viewpoints, at 10 degree increments, were captured. At each viewpoint, three polarized photos
were captured, for a total of nine photographs. The uncorrected surface normals, obtained from naive SfP, are poor. In particular, note the reflections of the ceiling lights in the normal map.
The proposed technique mitigates this issue, and reduces the MSE from 0.06 to 0.03. The Nishino method, while it does mitigate the dramatic specular reflections
from ceiling lights, results in a greater MSE. This is likely due to the significant holes in the normal map.
\section{Discussion}
In summary, we have proposed a new technique to separate reflection components from a scene using a combination of passive polarization and viewpoint. To our knowledge,
this is the first paper to do so. While there are alternate ways to address the reflection separation problem -- for instance, through the use of color channels -- we believe that
the proposed technique is a complementary approach that can be combined with previous methods.
\paragraph{Benefits:} The proposed technique may find direct application in improving the quality of SfP and related algorithms. Prior art in SfP has not analyzed in-depth the
impact of mixed reflections. This paper has shown that it may not be sufficient to sequentially apply an existing algorithm to first separate reflection components. Rather, the
physics of polarization that are used to obtain shape, can \emph{also} be used to separate reflection components. With the increased interest in multiview methods (e.g. KinectFusion~\cite{Izadi2011KinectFusion}), it seems logical to consider the inclusion of the proposed technique within such frameworks. In addition, the proposed technique has shown recovery of refractive index, which is a challenging problem in computational imaging often addressed with calibrated optical setups~\cite{wetzstein2011hand}. The ability to estimate refractive index is shown to greatly improve the accuracy of SfP, but may also find use in other applications like object detection.
\paragraph{Limitations:} We follow previous work in using the unpolarized world assumption -- the light incident on an object is initially unpolarized. In scenes with significant specular reflections -- like a house of mirrors -- the unpolarized world assumption is violated. It should be noted that prior art has empirically observed the validity of the unpolarized world assumption in realistic
scene conditions~\cite{kadambi2015polarized}. Although our paper also acquires refractive index, for the sole purpose of reflection separation, other strategies that use fewer images (e.g. Tan et al.~\cite{tan2005separating}) may be preferable. Specifically, the proposed technique uses three polarized images at a minimum of two viewpoints -- a minimum total of 6 images are required. However,
the intended application of this technique is to SfP, where it is expected to capture multiple images, and where it is desirable to estimate the refractive index.
\paragraph{Open challenges:} While the proposed technique forges a strong link betweeen shape, passive polarization, and reflectance, several open topics remain. For example,
would the technique improve if more viewpoints were captured; would circular polarization allow for more information to be gleaned; and could this method be combined with other frameworks (for example, color, viewpoint, and polarization)? In conclusion, we hope that this paper may improve the practicality of SfP, allowing surface normals to be estimated on surfaces
with mixed reflective properties.
\clearpage
\bibliographystyle{splncs}
|
\section{Introduction}
\label{sec:intro}
In animal cells, motility and morphology are strongly coupled and are largely due to the activity of the cell cytoskeleton. Research into these areas is broad and has many applications, from studying metastatic cancer cells to wound healing. In order to mimic aspects of these systems we model, both analytically and numerically, examples of active cytoskeletal material confined to droplets. An active material is defined as driven out-of-equilibrium by the internal energy of its constituent particles \cite{Marchetti2013}. We use the hydrodynamic model of an active polar fluid outlined in \cite{Kruse2004,Kruse2005,Furthauer2012} to model the behaviour of such a material at long length and time scales.
Over the past decade there have been a number of calculations of instabilities and non-equilibrium steady states in active liquid crystals; thin or 2D flat films \cite{Kruse2004,Voituriez2005,Voituriez2006,Kruse2006,Bois2011,Sarkar2015a}, thin cortical layers \cite{Zumdieck2005,Hawkins2011,Joanny2013,Khoromskaia2015}, confined in emulsion droplets or vesicles \cite{Callan-Jones2008a,Tjhung2012,Joanny2012a,Blanch-Mercader2013,Giomi2014,Whitfield2014,Marth2014,Tjhung2015}, and simplified models of animal and plant cells \cite{Hawkins2009,Woodhouse2012,Callan-Jones2013,Kumar2014,Turlier2014,Callan-Jones2016}. {In this paper we model deforming active droplets immersed in a passive fluid using linear perturbation theory. By making justified assumptions, we are able to predict non-equilibrium phase transitions in both of the systems we consider, and predict how the droplet deformation couples to these. These analytical calculations are presented for the three-dimensional case and also repeated for the two-dimensional analogue where we find qualitatively similar results. Numerical simulations use the two-dimensional Immersed Boundary method used in \cite{Whitfield2016} and are directly compared to the two-dimensional analytical calculation.
The models presented here are relevant to active systems \emph{in vitro} (constructed using techniques in \cite{Bendix2008, Sanchez2012, Keber2014}) as well mimicking aspects of cell dynamics. The two cases we consider correspond to two limits of active cytoskeletal behaviour (see figure \ref{fig:sketch}) that represent the minimum degrees of freedom required to observe interesting out-of-equilibrium dynamics. In both cases we consider a 1-component model used originally in \cite{Kruse2004}, which allows us to investigate the coupling with droplet shape dynamics analytically. The linear stability analyses are restricted by assumptions which enable an analytical understanding of the mechanisms involved in producing the observed behaviour in numerical simulations.
\begin{figure}[h!]
\hfill \includegraphics[width=\columnwidth]{fig1.pdf}
\caption{2D schematic of \textbf{(a)} Active fluid interface: active concentration $c$ on the droplet interface coupled to the internal concentration $\rho$. \textbf{(b)} Active polar droplet: constant density of active filaments with local average polarisation $\vect{p}$ (red arrows). Blue arrows indicate active contractile force dipoles.}
\label{fig:sketch}
\end{figure}
Firstly, we consider an isotropic layer of contractile active material confined to an interface between two fluids, which has physical similarities to the actomyosin cortex in cells. The stresses generated are confined to the plane of the interface giving rise to flows in the surrounding fluid and deformation of the interface itself. Interestingly, diffusion of the active particles through the bulk can result in { a change in which mode of the perturbation has lowest critical activity, from a single peak instability driving droplet motion to higher modes which produce symmetric deformation}. Furthermore, simulations show that advection through the bulk can stabilise such modes. This suggests that droplets with an active interface could spontaneously deform and possibly divide due to the feedback from the fluid flow.
Secondly, we consider a highly ordered active polar liquid crystal confined inside a fluid droplet. In this case the polarisation gradients direct the internal stresses giving rise to fluid flow. A polar anchoring condition at the interface means that the deformation of the droplet and polarisation field are strongly coupled. We find in this case there is a separation of swimming and stationary deforming modes, such that extensile activity destabilises the defect position and results in a swimming drop, whereas a contractile activity stabilises the centred defect position and gives rise to deformations of the interface.
\section{Active Fluid Interface}
\label{sec:model}
In this section we consider a fluid droplet coated by active particles on its interface that are isotropically ordered. Such systems have been found to self-organise in \emph{in vitro} experiments using reconstituted active cytoskeletal material contained in vesicles or droplets \cite{Tsai2011, Shah2014}. These experimental systems are a useful tool for understanding the more complex dynamics of cells. The model in this section makes predictions of interesting active phenomena including symmetry breaking, and droplet deformation, that are relevant to the field of cell mechanics.
\subsection{Model}
\label{sec:abmodel}
We consider a fluid droplet described by an interfacial surface $\Sigma$ separating the contained fluid domain $\Omega_0$ and external fluid domain $\Omega_1$ with viscosities $\eta_0$ and $\eta_1$ respectively. We define a concentration of active matter $c(\theta,\phi,t)$ on the interface $\Sigma$, which alters the droplet surface tension $\gamma$ such that $\gamma = \gamma_0 - \zeta_c \, c - Bc^2/2$. $\gamma_0$ is the bare surface tension, $\zeta_c$ is the activity ($\zeta_c<0$ for contractile) and $B$ is a passive repulsion force. This higher order repulsive term represents passive pressure, similar to that in \cite{Joanny2013}, which parametrises the compressibility of the active fluid on the interface. We denote the effective surface tension $\gamma'_0 = \gamma_0 - \zeta_c \, c_0 - Bc_0^2/2$, which is the value of $\gamma$ in the stationary state.
The force density on the droplet interface is then: $\vect{F} = \kappa \gamma \vect{\hat{n}} + \rndb{\nabla_s \gamma}\vect{\hat{t}}_i $, where $\vect{\hat{n}} = \vect{\hat{n}}(\theta,\phi,t)$ is the outward surface normal, $\vect{\hat{t}}_i = \vect{\hat{t}}_i(\theta,\phi,t)$ are the orthogonal surface tangent vectors, $\kappa = \nabla \cdot{\vect{\hat{n}}}$ is the local curvature, and $\nabla_s = (\vect{\hat{t}}_i \cdot \nabla)$ is the surface gradient. It is useful to define the effective activity $\tilde{\zeta}=\zeta_c+Bc_0$ which defines the scale of the force $\vect{F}$ for small deviations of the concentration $c$ from $c_0$. Thus, the interface has net contractility for $\tilde{\zeta}<0$.
The only forces acting on the system originate at the droplet surface $\Sigma$, with position $\vect{R} = R(\theta,\phi,t)\hat{\vect{e}}_r$ assuming this is single-valued with respect to the angular coordinates ($\theta$,$\phi$). Thus, the resulting force density in the fluid is $\vect{f}^{\rm ext}(r,\theta,\phi,t) = \vect{F} \delta\sqb{r - R(\theta,\phi,t)}$. We ignore inertia taking the low Reynolds' number limit, $Re = 0$, thus the incompressible fluid flow ($\nabla \cdot \vect{v} = 0$) is described by Stokes' equation $\eta_n \nabla^2\vect{v} + \vect{f}^{\rm ext} - \nabla P = 0$, where $n=0,1$ denotes the domain $\Omega_0$ or $\Omega_1$, $\vect{v}=\vect{v}(r,\theta,\phi,t)$ is the fluid velocity, $\vect{f}^{\rm ext}=\vect{f}^{\rm ext}(r,\theta,\phi,t)$ denotes any external force densities and $P=P(r,\theta,\phi,t)$ is the hydrostatic pressure. We take the limit of a zero-thickness interface and assume flow and stress continuity between the two fluids $\Omega_0$ and $\Omega_1$. This means the active particles act as an active surfactant, rather than a thin viscous layer (as in \cite{Bois2011, Hawkins2011, Joanny2013, Khoromskaia2015, Turlier2014, Callan-Jones2016}), which allows us to study the dynamics of deformation in a 3D viscous environment analytically.
The evolution of the surface concentration $c$ with respect to time $t$ is:
\begin{eqnarray}
\label{dcdt} \dot{c} = -\nabla_s \cdot (c \vect{v}_b) + D\nabla_s^2 c - k_{\rm off}c + k_{\rm on}\rho_b \,,
\end{eqnarray}
where $\dot{c}=\partial c/\partial t$, $\vect{v}_b = \vect{v}(r=R,\theta,\phi,t)$ is the interface flow velocity, $D$ is the diffusion constant for the active particles on $\Sigma$, and $k_{\rm on,off}$ are binding and unbinding rates of the particles to the interface. The concentration of unbound particles in the bulk of the drop is denoted $\rho = \rho(r,\theta,\phi,t)$. Binding occurs at the interface where we denote the concentration of unbound prticles $\rho_b = \rho(r=R,\theta,\phi,t)$. Note that $k_{\rm on}$ has units of velocity, as it contains the adsorption depth parameter. We assume that the active particles are insoluble in the external fluid, and so the evolution of the bulk concentration $\rho$ is given by:
\begin{eqnarray}
\label{drhodt} \dot{\rho} = -(\vect{v}\cdot\nabla)\rho + D_\rho\nabla^2 \rho
\end{eqnarray}
with the boundary condition $D_\rho(\vect{n}\cdot\nabla)\rho = k_{\rm on}\rho - k_{\rm off}c$ at $r=R$, to ensure conservation of mass. The parameter $D_\rho$ is the bulk diffusion constant of the active particles. Here we assume that the active particles only generate stresses at the interface, so the bulk concentration acts as a buffer to recycle the surface concentration.
\subsection{Linear Stability Analysis}
\label{sec:ablsa}
In this section we present the results of a linear perturbation to the stationary ground state of the droplet. The system is in a stationary (velocity $\vect{v} = 0$) steady state when the interface is spherical (fixed radius $R = R_0$) with a homogeneous concentration of active particles ($c = c_0$). Then the bulk concentration is $\rho_0 = k_{\rm off} c_0/k_{\rm on}$ inside the drop, and the hydrostatic pressure inside is $\vect{P} = P_{\rm ext} + (2\gamma_0 - \tilde{\zeta} c_0)/R_0$ where $P_{\rm ext}$ is the stationary state pressure in the external fluid. We perform a linear stability analysis by applying a small perturbation to the variables defined at the interface $R$ and $c$ of the form: $\tilde{g} = g_0 + \sum_{l=1}^{\infty}\sum_{m=-l}^{l}\delta g_{lm}(t)Y_l^m(\theta,\phi)$, where $Y_l^m$ are the spherical harmonic functions and $\delta g_{lm} \ll g_0$. To first order, the resulting flow is given by Lamb's solutions for flow around a sphere, which can be expressed as vector spherical harmonics \cite{Lamb1945}. Solving the Stokes equation with flow and stress continuity conditions at the droplet interface gives expressions for $\delta v^{(i)}_{lm}$ (as defined in \cite{Carrascal1991} and Supplementary Information appendix A) in terms of $\delta c_{lm}$ and $\delta R_{lm}$. The perturbation on the interface is also coupled to a perturbation of the internal concentration $\rho$ such that
\begin{eqnarray*}
\rho = \sqb{\frac{k_{\rm off} c_0}{k_{\rm on}} + \sum_{l=1}^{\infty}\sum_{m=-1}^{l}\delta\rho(r,t)Y_l^m} \, .
\end{eqnarray*}
We obtain analytical solutions for the stability by assuming a quasistatic solution for $\delta \rho$ (taking $\dot{\rho}=0$). This assumption corresponds to a fast relaxation of the bulk concentration $\rho$ compared to the timescale of evolution of the surface concentration $c$. At linear order, the solution for $\delta \rho$ simply satisfies the diffusion equation with a flux condition at the boundary:
\begin{eqnarray*}
\delta \rho = \frac{k_{\rm off} R_0 \delta c}{D_\rho l + k_{\rm on} R_0} \rndb{\frac{r}{R_0}}^l \,.
\end{eqnarray*}
This solution enables us to predict the effect of the feedback by diffusion through the bulk analytically. The full solutions to the coupled linear equations are solved exactly with Bessel functions as in \cite{Hawkins2011}, however these solutions do not permit an analytical calculation of the stability condition, hence we do not consider them here, but instead compare our approximate analytical solutions directly with the full dynamical simulations.
Finally, we evaluate the coupled system of dynamic equations for the concentration (equation \eqref{dcdt} in section \ref{sec:abmodel}) and radius $\dot{R} = \vect{v}_b.\hat{\vect{n}}$ (the normal velocity at the interface) to first order in the perturbations. We find instabilities by looking for positive eigenvalues of the stability matrix that relates $\dot{c}$ and $\dot{R}$ to $\delta c$ and $\delta R$ to first order in the perturbations (see Supplementary Information appendix A for further details of this calculation). From this analysis we find an instability threshold for the effective activity $\tilde{\zeta}<\alpha_I$ where
\begin{eqnarray}
\label{intZ} \alpha_I = -\frac{2\tilde{\eta}}{c_0}\rndb{2l+1}\rndb{\frac{D}{R_0} + \frac{D_\rho R_0 k_{\rm off}}{(l+1)\rndb{D_\rho l + k_{\rm on} R_0}}},
\end{eqnarray}
where $\tilde{\eta} = (\eta_0 + \eta_1)/2$ is the mean viscosity of the internal and external fluid. We see that $\alpha_I$ is independent of the effective surface tension $\gamma_0'$ which shows that the coupled droplet deformation does not contribute to the symmetry breaking threshold. However, the corresponding maximum eigenvalue of the stability matrix does weakly depend on the effective surface tension $\gamma_0'$ for $l>1$. This weak positive relation suggests that the instability should evolve more quickly in large surface tension drops when $l>1$. In this linear limit there is no contribution from the advection term in \eqref{drhodt} and the second term in \eqref{intZ} (proportional to the binding rates) always increases the threshold. This is because the binding terms allows the concentration on the interface to be recycled by unbinding and diffusing into the bulk of the drop.
{The stability analysis shows how the droplet will initially deform. This deformation is characterised at short times by the maximally unstable mode $l_{\rm max}$, which can be found exactly when binding is not included (see figure \ref{fig:lmax} and Supplementary Information appendix A). At linear order the instability is independent of the spherical harmonic parameter $m$. Generically, $l_{\rm max}$ predicts that as contractile activity is increased, the more concentration peaks will be initially formed on the droplet surface (figure \ref{fig:lmax}). The total droplet activity scales with droplet size, and so $l_{\rm max}$ is more sensitive to the activity parameter $\tilde{\zeta}$ in larger droplets. Thus it is easier to observe modes with small $l$ in smaller droplets, where the dynamics are less sensitive to small changes in the activity. Note that only the $l=1$ mode ($k=1$ in 2D) produces net propulsion of the droplet (i.e. $\int_\Sigma \dot{R}\hat{\vect{n}}{\rm d}S \neq 0 $), so the first unstable mode corresponds to front-back symmetry breaking of the droplet profile.
As shown in Supplementary Information appendix A, one can approximate the maximally unstable mode $l_{\rm max}$ analytically by solving $\dot{R} = 0$ for $\delta R_{lm}$. This approximation imposes that $R$ always assumes the steady state shape for a given fixed concentration perturbation $\delta c_{lm}$ (plotted in figure \ref{fig:lmax}). Physically, this assumes that the shape dynamics are much faster than the concentration dynamics, and so can be taken to be quasistatic. Interestingly, while this assumption does not represent the full coupled dynamics of $\delta c_{lm}$ and $\delta R_{lm}$, it does reproduce the critical activity threshold, and also approximates the mode structure well.
When binding is included ($k_{\rm off} \neq 0$) the dispersion relation changes, and as we see from \eqref{intZ} the active threshold is non-linear in $l$, and hence higher (non-swimming) modes can have lower activity thresholds than the $l=1$ (swimming) mode.}
\begin{figure}
\hfill\includegraphics[width=\columnwidth]{fig2.pdf}
\caption{Maximum mode number $l_{\rm max}$ plotted against activity in normalised units for increasing values of the droplet radius. Dashed lines show numerical solution and solid lines show analytical approximation using $\dot{R}=0$. Parameters used: $c_0 = 1$, $\gamma_0=1$, $D=0.05$, $\eta_0=\eta_1=1$ and $k_{\rm off}=0$. Insets show flow (blue arrows) and active concentration $c$ (colour gradient from purple (low) to yellow (high)) to linear order on the perturbed interface for a (i) $l=1$ mode and (ii) $l=2$ mode respectively. Deformation of the interface in (ii) is calculated {by solving $\dot{R}=0$ for $\delta R$ given the form of $\delta c$, and is exaggerated for visibility} using small $\gamma_0'$.}
\label{fig:lmax}
\end{figure}
Within the assumptions made here, the binding and unbinding dynamics always increase the activity threshold. We see that if the binding is fast $k_{\rm on} \gg D_\rho$, the critical activity takes the same form as the 1D model considered in \cite{Hawkins2011} where the active threshold is always minimal for $l=1$ and is proportional to the effective diffusion parameter $\tilde{D} = (Dk_{\rm on} + D_\rho k_{\rm off})/k_{\rm on}$. However, for fast bulk diffusion, geometrical effects become important. A single peak in the interfacial concentration gives rise to a concentration gradient in the bulk driving diffusion away from it. As the number of peaks on the interface increases the concentration gradients are more localised to the surface, and diffusion has a smaller effect. In this regime, the minimum critical activity can correspond to multi-peak modes ($l>1$) when the contribution from bulk diffusion is significant. This is analogous to the findings in \cite{Bois2011} for a one-dimensional active fluid consisting of two-components.
The droplet shape instability is enslaved to the concentration (as $\alpha_I$ is independent of $\gamma$), so we can estimate how the shape will deform due to certain concentration distributions on the interface by solving $\dot{R}=0$ for $\delta R$ (for $l>1$). Plotted in figure \ref{fig:lmax} is an example of these deformations and the associated flow to linear order. In order to calculate the resulting steady state dynamics we require numerical simulation.}
\subsection{Results and Comparison with Simulations}
\label{sec:absims}
We test these analytical results against the 2D simulations developed in \cite{Whitfield2016}. These use an Immersed Boundary method \cite{Peskin2002,Lai2008} to represent the active interface explicitly as a Lagrangian mesh which is coupled to the Cartesian mesh for the 2D fluid via a numerical Dirac delta function.
Repeating the stability analysis in 2D, we now take perturbations of the form $g = g_0 + \sum_{k=1}^{\infty}{\rm e}^{ik\theta}$. The calculation reveals that surface tension gradients do not deform the drop in 2D (as found in \cite{Yoshinaga2014}) however the concentration dynamics remain very similar. We compare our predictions in 2D to the results of the Immersed Boundary simulations in figure \ref{fig:abphase}. We run simulations varying the activity, binding rate (taking $k_{\rm off} = k_{\rm on}$) and diffusion parameters. At zero binding we observe two steady phases, a stationary state and a steady moving state \ref{fig:abphase}(a) separated by the threshold $\alpha_{\rm I,2D}$ which agrees well with the expected analytical result
\begin{eqnarray}
\label{2DintZ} \alpha_{\rm I,2D} = -4\frac{\tilde{\eta}}{c_0}\rndb{\frac{Dk}{R_0} + \frac{D_\rho R_0 k_{\rm off}}{\rndb{D_\rho k + k_{\rm on} R_0}}} \, .
\end{eqnarray}
{This moving steady state due to a surface tension gradient is also observed for the the self-propelled droplets studied in \cite{Yoshinaga2014,Ohta2009a}. The equations of motion we use (see Model section) are similar to those for the self-propelled droplets studied in \cite{Yoshinaga2014,Ohta2009a} and hence some of the same dynamical behaviour is observed. However, our model predicts new stable states and instabilities corresponding to pure deformation and division as discussed below. This arises due to the advection and diffusion of active particles through the bulk of the drop. Unlike in \cite{Yoshinaga2014,Ohta2009a} the model here conserves the active particles within the drop making it more relevant to cell cortex dynamics.}
\begin{figure}[h!]
\hfill\includegraphics[width=\columnwidth]{fig3.pdf}
\caption{Phase diagram of 2D simulation results for an active isotropic interface, each dot represents a single simulation run. Insets show steady state flow (blue arrows) and concentration fields (colour density, black to yellow) for the different phases. Low values of $k_{\rm off}$ transition from stationary (black squares) to motile (red circles) with a single peak in concentration (shown in (a)). Feedback from the internal concentration {produces intermediate oscillatory states} (magenta stars) and a stationary 2-peak state (blue triangles). Solid lines of increasing gradient show predicted activity threshold for modes $k=1,2$ (red, blue). Simulation parameters: $c_0 = 1$, $R_0=1$, $\gamma_0=1$, $D=0.05$, $D_\rho=0.5$, $\eta_0=\eta_1=1$. }
\label{fig:abphase}
\end{figure}
We next calculate the maximum mode number $k_{\rm max}$ (see Supplementary Information appendix A). In the regime where we predict $k_{\rm max}=2$, our simulations show initial formation of 2 peaks in droplet concentration. Without binding, these peaks are unstable and always coalesce to form one (as predicted for a flat active viscous layer in \cite{Bois2011}). In this case, the droplet swims persistently and steadily with the concentration peak at its rear. A decomposition of the Fourier modes of this steady state shows that the far field flow is puller like, i.e. its dipole moment is such that it pulls the surrounding fluid inward and pushes it outward along the axis perpendicular to its motion. The activity threshold predicted compares well to that in the simulations for small values of the binding. At larger binding rate, the interior dynamics is not completely diffusion dominated, and the critical activity is underestimated due to the approximation of $\dot{\rho} = 0$. As we increase $k_{\rm off}$ and $\tilde{\zeta}$ we see that eventually the droplet becomes immobile with 2 stable peaks in the concentration (see figure \ref{fig:abphase}). In the intermediate regime the droplet undergoes a `wandering' motion as the concentration profile oscillates between a single peak and two peaks. Equation \eqref{2DintZ} predicts a non-trivial $k$ dependence of the active threshold as binding terms become important. For the parameters used in figure \ref{fig:abphase}, this can be seen by the crossing of the lines for the $k=1$ and $k=2$ modes, meaning that the minimum critical activity is not necessarily for the lowest $k$ mode ($k=1$). Note this is very similar to the prediction in 3D in \eqref{intZ}.
The simulation results in figure \ref{fig:abphase} demonstrate that as the binding rate increases, advection of the concentration through the droplet bulk becomes more important. The advection can stabilise the two peaks at diametrically opposite points on the circle, resulting in a stationary droplet. However, we see that in 2D the drop does not deform, as the radial forces from the activity gradients are always cancelled by the hydrostatic pressure $P$. This is not the case for the full 3D system where we expect concentration gradients to deform the droplets as shown in figure \ref{fig:lmax}. Nonetheless, the 2D simulations show that advection can stabilise the 2 peak configuration, which in 3D would result in symmetric deformation and potentially division of the droplet. Such a 3D simulation is beyond the scope of this work, but would be useful for quantifying the full 3D morphology. Recent work has shown that non-adherent cells exhibit a swimming state similar to the motion described here, and so it would be of interest to test in future work whether the steady state shape in 3D for the model here resembles the `pear shape' observed in \cite{Ruprecht2015, Callan-Jones2016}.
\section{Active Polar Fluid Droplet}
In this section, we consider a droplet filled with an active polar liquid crystal of constant density everywhere. Realising this system experimentally in droplet systems requires high concentrations of active material so that the polar to isotropic phase transition is localised to the droplet centre. This has been achieved \emph{in vitro} for microtubule based active nematics but only in thin films thus far \cite{Sanchez2012,Keber2014}. In these systems the measured order parameter is approximately constant everywhere except in the vicinity of topological defects. Thus we consider the limit where the active fluid is strongly polarised and restrict the analysis to only the orientational degrees of freedom of the active liquid crystal, and do not consider the density or polarisation magnitude degrees of freedom.
\subsection{Model}
\label{sec:apfmodel}
We utilise the model of an active polar fluid developed by Kruse et al. in \cite{Kruse2004,Kruse2005,Furthauer2012} which has similarities to other continuum models of the cytoskeleton on surfaces (such as \cite{Lober2015, Ziebert2014}). We consider the case where the active fluid has strong local ordering and is far from the isotropic phase so that $\norm{\vect{p}}=1$ everywhere (except at defects where $\vect{p}$ is undefined). This approximation is commonly used to model active and passive liquid crystal systems analytically.
In the $Re=0$ limit the total stress in the active polar fluid, $\sigma_{i j}^{\mathrm{tot}} = \sigma_{ij}^{\mathrm{visc}} + \sigma_{ij}^{\mathrm{dist}} + \sigma_{ij}^{\mathrm{act}} $,
has viscous, distortion and active contributions respectively where:
\begin{align*}
\sigma_{ij}^{\mathrm{visc}} &= 2\eta_{n} u_{ij} = \eta_{0,1}\rndb{\partial_i v_j + \partial_j v_i} \, , \\
\sigma_{ij}^{\mathrm{dist}} &= \frac{\nu}{2} \rndb{p_i h_j + p_j h_i} + \frac{1}{2}\rndb{p_i h_j - p_j h_i} + \sigma_{ij}^{\mathrm{e}} \, , \\
\sigma_{ij}^{\mathrm{act}} &= - \zeta p_i p_j \, .
\end{align*}
The viscous stress is the response to flow assuming a Newtonian fluid. The distortion stress is that of a passive polar liquid crystal due to deviations in filament alignment, where the perpendicular part of the molecular field $h_i = -(\delta F/\delta p_j)(\delta_{ij}-p_ip_j)$ acts to minimise the free energy functional $F = \int_{\Omega+\Sigma} {\rm d}^3r f$ with respect to $\vect{p}$, given $\left| \vect{p} \right| = 1$. The Ericksen stress, $\sigma_{ij}^{\mathrm{e}} = f\delta_{ij} - (\partial f/(\partial(\partial_jp_n)))(\delta_{ij} - p_n p_k) \partial_i p_k $, is a generalisation of the hydrostatic pressure for complex fluids. Finally, the active stress represents the active dipolar force and thus is second order in $\vect{p}$.
The free energy functional $F$ gives the equilibrium properties of the system. Here for simplicity we use the one constant approximation of the Frank free energy:
\begin{eqnarray}
\label{FreeE} F = \int_\Omega {\rm d}^3r \frac{K}{2}(\partial_i p_j)^2 + \int_\Sigma {\rm d}S f_s \; ,
\end{eqnarray}
where $K$ is the elastic constant and $\norm{\vect{p}}=1$. Since we are modelling a finite droplet, the surface terms are important. We consider normal anchoring of the filaments to the interface, with surface distortion free energy density $f_s = W (\vect{p}\cdot\hat{\vect{n}} - 1)^2 $. {This form of the surface free energy includes the `spontaneous splay' term which is allowed in polar liquid crystals \cite{Pleiner1989}.}
The polarisation flux is
\begin{eqnarray}
\label{dpdt} \dot{\vect{p}} = -\rndb{\vect{v}\cdot\nabla}\vect{p} - \tens{\omega}\cdot\vect{p} - \nu\tens{u}\cdot\vect{p} + \frac{\vect{h}}{\Gamma}
\end{eqnarray}
where $\omega_{ij} = (\partial_iv_j - \partial_jv_i)/2$ and $\Gamma$ is the rotational viscosity.
\subsection{Linear Stability Analysis}
\label{sec:apflsa}
We contrast the model of an active interface to that of a droplet of active polar fluid of constant density. In this case, rather than the concentration of active particles, the important degree of freedom is the polarisation vector $\vect{p}$ denoting the average direction of the contractile filaments in the fluid.
We calculate the linear stability of the droplet in the limit of strong anchoring $W \rightarrow \infty$ in order to study the effects between the coupling of droplet morphology and polarisation. This equates to the boundary condition $\vect{p} = \hat{\vect{n}}$ at $\vect{r}=\vect{R}$. In the case of weak or no anchoring, instabilities can occur for both extensile ($\zeta>0$) and contractile ($\zeta<0$) active polar drops as shown analytically in \cite{Whitfield2015} and in simulations \cite{Tjhung2012}. The condition of fixed polarisation at the interface inhibits certain deformations of the polarisation field at low activities and so the preferred deformation modes are those which can couple to the droplet deformation. This was demonstrated in 2D simulations of active nematic drops in \cite{Giomi2014}. Here we explain this mechanism analytically in a 3D fluid drop by linear stability analysis. The polar nature of the anchoring produces a ``radial hedgehog'' topological defect at the droplet centre (or a radial defect with $+1$ winding number in 2D), giving a simple analytical description of the stationary state. Thus we are able to make analytical predictions about spontaneous symmetry breaking in these systems even in the general 3D case.
Unlike the case of an active interface, the active fluid here fills the drop, and hence active and passive stresses are generated in the bulk. The stationary steady state is given by the polarisation $\vect{p} = \hat{\vect{r}}$, $\vect{R} = R_0 \hat{\vect{r}}$, and $\vect{v} = 0$.
{To perform a general linear stability analysis, one would need to consider generic perturbations to both the polarisation field and interface and study the coupled equations for their evolution, this is not analytically tractable in this case. However, we can perform restricted perturbations that we expect to be representative of the dynamics in a particular limit. We consider the case where the polarisation field is enslaved everywhere to the shape of the boundary by the anchoring condition. This corresponds to the limit where bulk instabilities in the droplet are suppressed by its size (\emph{i.e.} small droplets). In larger droplets, (or equivalently for smaller $K$) the dynamics of the polarisation field becomes more independent of the anchoring condition, and we expect this approximation to break down. }
Due to the symmetry of the stationary state, we first need to consider the special case of the translational mode of perturbation, corresponding to the $l=1$ spherical harmonic mode. Without loss of generality we consider a perturbation along the $z$-direction ($m=0$). This mode implies a translation of the hedgehog defect away from the droplet centre. If we assume that the defect has some fixed finite core radius $R_{c}$ then we can treat the liquid crystal as contained between two boundary conditions, one at the defect $r=R_{c}$ and one at the droplet interface $r=R_0-\delta z \cos(\theta)$, where $\delta z$ is a small displacement of the defect position from the droplet centre along the $z$-direction. The calculation is done in the reference frame of the defect so that it coincides with the origin of our coordinate system. In the equilibrium case ($\zeta=0$), we can write a polarisation field to first order that minimises the bulk free energy in \eqref{FreeE} by solving $\vect{h} = 0$ for these boundary conditions:
\begin{eqnarray}
\label{pl1} \vect{p}_{l=1} = \vect{e}_r - \delta z \frac{r-R_c}{r(R_0-R_{c})}\sin(\theta) \vect{e}_\theta \, .
\end{eqnarray}
This method equates the defect to a small colloid with (polar) homeotropic anchoring, and in the strong anchoring case we expect the free energy minimum to correspond to the defect being positioned at the droplet centre as we observe in simulations, and is reported in \cite{Lubensky1998, Poulin1998}. Using the polarisation in equation \eqref{pl1} we can estimate what the bulk free energy increase will be for such a deformation (details in Supplementary Information Appendix B)
\begin{align}
\notag \Delta F_{\rm bulk} &= \frac{4 K \pi \delta z^2}{3 R_{0} (1-\epsilon)^2} \sqb{4- 3\epsilon - \epsilon^2 + 4\epsilon \log \rndb{\epsilon}} +\, O(\delta z^3) \\
\label{deltaF} &\approx \frac{16 K \pi \delta z^2}{3R_{0}}
\end{align}
where $\epsilon=R_{c0}/R_{0}$ is assumed small in the final approximation of the equation. This $\Delta F$ is positive for all $\epsilon$, suggesting that the free energy minimum corresponds to the defect being positioned at the droplet centre. Note that this polarisation field is only valid to first order in $\delta z$ and so higher order terms could affect the form of the quadratic term here.
We now introduce a small activity $\zeta$, such that equation \eqref{pl1} remains a valid approximation for the form of the polarisation field, then we see that this gives rise to active forces in the drop. We solve the force balance equations (omitting passive contributions, see Supplementary Information Appendix B) to find the active contribution to the flow. We then integrate to find the active contribution to the velocity of the defect core $\vect{v}_{c}$ and droplet $\vect{v}_{\rm drop}$. The relative velocity of the defect is then:
\begin{equation}
\Delta \vect{v} \equiv \vect{v}_c - \vect{v}_{\rm drop} \approx \zeta \delta z\frac{(2\eta_0+\eta_1)-\epsilon(\eta_0+\eta_1)}{2 \eta_0 (3 \eta_0 + 2 \eta_1)} \hat{\vect{e}}_z \; .
\end{equation}
We see that extensile activity ($\zeta>0$) always results in a relative defect velocity that is in the same direction as the initial defect displacement (along $\hat{\vect{e}}_z$), as shown by figure \ref{fig:lsavel}. This implies that extensile activity will destabilise the defect from the centre and give rise to motion of the droplet as a whole (which to linear order is also along $\hat{\vect{e}}_z$). Conversely, we expect contractile activity to stabilise the defect at the droplet centre, as the flows resulting from contractile activity ($\zeta<0$) act to restore the defect back to its stationary position at the droplet centre.
Thus, within the assumptions made above, one can predict that the active polar droplet will break translational symmetry spontaneously above some finite activity. This mode of symmetry breaking is independent of surface deformations at linear order, and so its critical activity threshold should not depend on the droplet surface tension. Hence the critical activity threshold will only depend on the increase in the passive free energy (equation \eqref{deltaF}), which goes to a finite value in the limit of a point defect and scales as the inverse of the droplet size. In general, the parameter $\epsilon$ is difficult to define, which is a consequence of the assumption of $\norm{\vect{p}}=1$, which breaks down around the defect. This can be avoided by using a Landau-De Gennes type free energy description for the passive part of the dynamics such that there is an polar-to-nematic phase transition at the centre of the droplet. However, such an approach is not analytically tractable, as it requires solving non-linear partial differential equations for the radial dependence of $\vect{p}$. Qualitatively though, the predictions here are consistent with what is observed in the simulations.
\begin{figure}[h!]
\hfill\includegraphics[width=\columnwidth]{fig4.pdf}
\caption{Active part of the flow field (blue arrows) to linear order in the perturbations for: (a) defect position (inner sphere) displaced in the vertical direction with $\zeta>0$ (extensile activity); (b) $l=2$ mode perturbation of the interface assuming strong anchoring of the polarisation field with $\zeta<0$ (contractile activity). The perturbations are made artificially large for visibility here.}
\label{fig:lsavel}
\end{figure}
For perturbation modes $l>1$ the flow at the origin will always be zero, and so one can assume that in the strong anchoring limit the defect will remain centred at the origin. We again require an assumption for the $r$-dependence of the polarisation perturbation. Taking $R_{c0} \rightarrow 0$, we can write a general form as $\delta \vect{p} \propto r^n$ for arbitrary $n\ge0$. Importantly, for all $n$, the active flows always give rise to an instability for $\zeta<0$ (contractile). Considering only active flows, the maximally unstable perturbation is for $n=0$. Thus, below we consider only the results of this mode, which allows us to consider the dynamics in the limit where the filament polarisation at the interface and in the droplet are strongly coupled. However it comes at the cost of reducing the quantitative power of our predictions, and is an important restriction to the dynamics considered. Note, in 2-dimensions, the assumption $n=0$ gives rise to an infinite passive contribution to the dynamics (proportional to $K$) and so we use $n=1$, which appears consistent with what is observed in simultions.
In the strong anchoring limit, the polarisation has to match the perturbed interface normal at $r=R$ to first order, such that
\begin{eqnarray}
\label{pl2} \vect{p} = \hat{\vect{r}} - \sum_{l=2}^\infty \sum_{m=-l}^{l} &\sqb{\frac{\delta R_{lm}(t)}{R_0} r(\nabla Y_l^m(\theta,\phi))} \, .
\end{eqnarray}
We calculate the resulting flows to first order in $\delta R$. Since $\vect{p}$ is enslaved to the deformation we then only need to consider the radius dynamics given by $\dot{R}$ (for details see Supplementary Information appendix B).
In this strong anchoring limit we find that the droplet is unstable if $\zeta<\alpha_P<0$, i.e. the activity threshold, $\alpha_P$, is always contractile. The threshold $\alpha_P$ increases linearly with $\gamma$ and $K$. Repeating the linear stability analysis calculation in 2D shows the same qualitative prediction, where this time we take $\delta p \propto r$ as this is the leading order contribution allowed. The analytical expressions for the activity threshold are given in Supplementary Information appendix B and a full discussion of the eigenvalues of the general stability matrix (for weak anchoring) can be found in \cite{Whitfield2015}.
The result of this analysis is somewhat surprising, in this strong anchoring limit we expect the $l=1$ mode to be unstable to extensile activity, whereas the higher modes of deformation are unstable for contractile activity. This suggests that, when our assumptions hold, we should see translational symmetry breaking with the defect moving to the droplet front for an extensile drop and symmetric modes of deformation for a contractile drop (see figure \ref{fig:lsavel}). This active threshold scales linearly with $K$ and $\gamma_0$, demonstrating the importance of the coupling of the morphology to the polarisation field. Contrast this to the case of the active interface where the shape does not affect the threshold for a phase transition.
This contractile instability can be understood physically by considering the splay in the drop due to perturbations in the interface curvature. High curvature couples to increased splay which couples to outward flow, further increasing the curvature of the interface and hence the splay. A sketch of this is given in figure \ref{fig:instsketch}.
\begin{figure}[h]
\centering
\hspace{30mm}\includegraphics[width=\columnwidth]{fig5.pdf}
\caption{Spatial change in splay induced by boundary pertubation. Dotted line indiciates $R_0$ and solid line the perturbed interface $R$. Increased splay in regions of higher curvature drive outward flows, coupling to further increase in boundary curvature. The black arrows indicate polarisation direction while the colour gradient indicates the splay magnitude $\norm{\nabla \cdot{\vect{p}}}$ relative to its value in the stationary state.}
\label{fig:instsketch}
\end{figure}
\subsection{Results and Comparison with Simulations}
\label{sec:apfsims}
In the 2D simulations (see figure \ref{fig:bulkphase}) we see symmetry breaking corresponding to the $k=1$ mode for extensile activity resulting in a steady motile state, as predicted by the stability analysis. This is characterised by the defect centre moving to the front of the drop and is independent of the boundary deformation (and hence $\gamma_0$). Due to the extensile nature of the activity this droplet is a pusher, pushing fluid out along its axis of motion and thus elongating parallel to its motion.
Conversely contractile activity stabilises the defect at the droplet centre and we observe a $k=2$ mode instability characterised by deformation of the droplet into a `dumbbell' shape. It is also observed that this phase behaviour breaks down as the value of $K/R_0^2$ is reduced. In this limit the distortions in the droplet bulk are not strongly coupled to those at the interface and so more complex distortions can occur without significant droplet deformation. Our analytical calculations do not predict this as we assume a form for the $r$-dependence of the polarisation such that it is strongly coupled to the curvature. This behaviour goes beyond the scope of the analytical work here as this corresponds to a transition to an `active turbulence' state, as numerically simulated in \cite{Whitfield2016}.
\begin{figure}[h!]
\hfill\includegraphics[width=\columnwidth]{fig6.pdf}
\caption{Active polar drop stability diagram. Stationary state (white, square dots), spontaneous symmetric deformation (blue, triangular dots) and spontaneous motility (red, round dots) are observed. Dashed line shows analytical prediction from linear stability analysis. Insets show the polarisation field $\vect{p}$ (black arrows) inside the droplet following symmetry breaking with defects labelled by blue dots. Note that due to the simulation method, the polarisation field in the simulations changes continuously from $\norm{p}=1$ inside the drop to $\norm{p}=0$ outside, hence the polarisation is defined everywhere in (i) and (ii). Parameters used: $K = 0.1$, $R_0=1$, $\eta_0=\eta_1=\Gamma=1$, $W = 50$ and $\nu=1.1$.}
\label{fig:bulkphase}
\end{figure}
Finally, we also observe rotational steady states in the simulations (for extensile activity when using $\nu=-1.1$) which can be characterised exactly by rotationally invariant distortions of the polarisation field \cite{Kruse2004,Kruse2005}, but these are not predicted for the parameter range used in figure \ref{fig:bulkphase}.
\section{Discussion}
\label{sec:discussion}
We have used analytical linear stability analysis and numerical simulation to characterise instabilities in active droplets and their resulting non-equilibrium steady states. Recent advances in experimental techniques mean that active gels of cytoskeletal material can be produced \emph{in vitro}. The predictions of our active interface model could be tested by adsorbing an isotropic actin gel onto the interface of an emulsion drop containing myosin and ATP \cite{Tsai2011,Shah2014}. We predict an activity threshold for spontaneous motion, and a further continuous transition to a stable symmetric state mediated by advection of motors through the droplet bulk. We predict that in 3D this symmetric configuration will be coupled to deformation of the drop, however this cannot be observed in the 2D model.
The active polar drop model we use only predicts some of the dynamics of a real active polar drop system as it ignores the density and ordering magnitude degrees of freedom. However, this model system gives us an insight into the intrinsic instabilities when droplet deformation and filament polarisation direction are strongly coupled. In particular, there is a contractile activity threshold that is linearly dependent on surface tension, above which the droplet spontaneously deforms into a characteristic dumbbell shape. We also see persistent motility in the case of extensile activity such that the droplet acts as a \emph{pusher}, compared to the \emph{puller} type motion exhibited in the active isotropic interface model. This is consistent with previous active droplet models that show contractile activity resulting in droplets which are \emph{pullers} and extensile activity resulting in \emph{pushers} \cite{Tjhung2012, Giomi2014, Marth2014, Khoromskaia2015, Tjhung2015}. An interesting future extension of this work would be to consider coupling between both of the active phases studied here within a single drop.
The finite active systems we study improve our understanding of how confinement and deformation affect steady state dynamics. Additionally, we see the importance of feedback, driven by advection through the droplet or the internal orientational order, resulting in more complex dynamics. These results should prove useful in characterising future experiments on \emph{in vitro} cytoskeletal networks and be useful in developing more complex models of multicomponent active systems in nature.
\section*{Acknowledgements}
We acknowledge the EPSRC for funding this work, grant reference EP- K503149-1.
|
\section{Main Results}
\label{sec:main}
In Section~\ref{sec:computation} below, we provide our main computational result: a poly-time approximation algorithm for BAVWM, which implies a poly-time truthful mechanism for revenue maximization that respects ex-post IR and ex-post budget constraints. In Section~\ref{sec:structure}, we detail the structure of the optimal mechanism in this setting, as well as our computationally efficient mechanism from Section~\ref{sec:computation}.
\subsection{Computational Results}\label{sec:computation}In this section, we provide a poly-time 3-approximation for BAVWM. We begin by writing a LP relaxation, allowing the designer to award fractions of items as long as the total fraction awarded doesn't exceed $1$. We split the fraction of item $j$ awarded to bidder $i$ into two parts, $\bar{x}_{ij}$ and $\hat{x}_{ij}$. Let $\bar{x}_{ij}$ denote the fraction of item $j$ assigned to agent $i$ before exceeding $b_i$. And let $\hat{x}_{ij}$ denote the fraction of item $j$ assigned after. In other words, if $x_{ij}$ is the fraction of item $j$ assigned to agent $i$, we have $\bar{x}_{ij} + \hat{x}_{ij} = x_{ij}$, $\sum_j \bar{x}_{ij} v_{ij} \leq b_i$, and $\sum_j \bar{x}_{ij} v_{ij} = b_i$ if for any $j$, $\hat{x}_{ij} > 0$. The idea is that assigning more of item $j$ to agent $i$ before exceeding his budget increases both terms in the ``goal'' above, but assigning more after exceeding the budget only affects the second term. The LP relaxation is as follows:\\
\noindent\textbf{Variables:}
\begin{itemize}
\item $\bar{x}_{ij}$, for all agents $i$ and items $j$, denoting the fraction of item $j$ assigned to agent $i$ contributing to both the budgeted-additive and additive terms in bidder $i$'s (virtual) welfare.
\item $\hat{x}_{ij}$, for all agents $i$ and items $j$, denoting the fraction of item $j$ assigned to agent $i$ contributing to just the additive term in bidder $i$'s (virtual) welfare.
\end{itemize}
\textbf{Constraints:}
\begin{enumerate}
\item $\sum_i (\bar{x}_{ij} + \hat{x}_{ij}) \leq 1$, for all $j$, guaranteeing that no item is allocated more than once.
\item $\sum_j \bar{x}_{ij}v_{ij} \leq b_i$, for all $i$, guaranteeing that contributions to the budgeted-additive term are not overcounted.
\end{enumerate}
\textbf{Maximizing:}
\begin{itemize}
\item $\sum_{ij} m_i \bar{x}_{ij}v_{ij} + \sum_{ij} w_{ij} (\bar{x}_{ij} + \hat{x}_{ij})$, the virtual welfare. Note that as each $m_i \geq 0$ and $v_{ij} \geq 0$, the optimal solution will never have $\hat{x}_{ij} > 0$ unless $\sum_j \bar{x}_{ij} = b_i$.
\end{itemize}
It is clear that any solution to BAVWM has a corresponding fractional solution to this LP. So the goal is to solve this LP and round the fractional solution to an integral one without too much loss. The idea is that the feasible region now looks pretty similar to that of the generalized assignment problem, asking for an assignment of jobs to machines such that the capacity of machine $i$ is at most $b_i$. We first prove the following rounding theorem, which is a near-direct application of Theorem~\ref{thm:ST}.
\begin{theorem}\label{thm:rounding}
The optimal fractional solution to the above LP can be rounded in polynomial time to an integral assignment such that:
\begin{enumerate}
\item $\sum_i (\bar{x}_{ij} + \hat{x}_{ij}) \leq 1$ for all $j$.
\item $\sum_j \bar{x}_{ij} v_{ij} \leq 2b_i$ for all $i$.
\item $\sum_{ij} m_i \bar{x}_{ij}v_{ij} + \sum_{ij} w_{ij} (\bar{x}_{ij} + \hat{x}_{ij}) \geq OPT$, where $OPT$ is the value of the LP.
\end{enumerate}
\end{theorem}
\begin{proof}
We show how to interpret our LP as an instantiation of a fractional LP for the generalized assignment problem, and then directly apply Theorem~\ref{thm:ST}. We use $p_{ij}$ to denote processing times, $c_{ij}$ to denote costs, and $T_i$ to denote capacities in the created generalized assignment problem instance.
\begin{itemize}
\item Machines:
\begin{enumerate}
\item A dummy machine, $0$.
\item For all bidders $i$, a hat machine $\hat{i}$ (corresponding to the hat variables in our LP).
\item For all bidders $i$, a bar machine $\bar{i}$ (corresponding to the bar variables in our LP).
\end{enumerate}
\item Jobs: A job $j$ for all items $j$.
\item Processing times and costs:
\begin{enumerate}
\item $p_{0j} = c_{0j} = 0$ for all $j$. $T_0 = 0$.
\item $\hat{p}_{ij} = 0$ for all $j$. $\hat{c}_{ij} = w_{ij}$ for all $j$. $\hat{T}_i = 0$.
\item $\bar{p}_{ij} = v_{ij}$. $\bar{c}_{ij} = m_i v_{ij} + w_{ij}$. $\bar{T}_i = b_i$.
\end{enumerate}
\end{itemize}
The fractional LP referenced in Theorem~\ref{thm:ST} on this instance would then be (note that the capacity constraints for machines $0$ and all $\hat{i}$ are vacuously satisfied, and that there do not exist any $i, j$ for which $p_{ij} > T_i$ by Observation~\ref{obs:2}):\\
\noindent\textbf{Variables:}
\begin{itemize}
\item $x_{0j}$, for all jobs $j$, denoting the fraction of job $j$ assigned to machine $0$.
\item $\bar{x}_{ij}$, for all machines $i$ and jobs $j$, denoting the fraction of job $j$ assigned to machine $\bar{i}$.
\item $\hat{x}_{ij}$, for all machines $i$ and jobs $j$, denoting the fraction of job $j$ assigned to machine $\hat{i}$.
\end{itemize}
\textbf{Constraints:}
\begin{enumerate}
\item $x_{0j} + \sum_i (\bar{x}_{ij} + \hat{x}_{ij}) = 1$, for all $j$, guaranteeing that every job is allocated exactly once.
\item $\sum_j \bar{x}_{ij}v_{ij} \leq b_i$, for all $i$, guaranteeing that the total processing time on machine $\bar{i}$ is at most $b_i$.
\end{enumerate}
\textbf{Maximizing:}
\begin{itemize}
\item $\sum_{ij} m_i \bar{x}_{ij}v_{ij} + \sum_{ij} w_{ij} (\bar{x}_{ij} + \hat{x}_{ij})$, the cost.
\end{itemize}
It's clear that this LP is exactly the same as our LP, just with an additional dummy bidder $0$ who collects all unallocated fractions of items. By Theorem~\ref{thm:ST}, the optimal fractional solution to this LP can be rounded in polynomial time to an integral solution whose total cost is at least as large, but where the capacity of machine $\bar{i}$ could be as large as $2b_i$, which is exactly an integral allocation of items to bidders with the desired properties.
\end{proof}
After applying Theorem~\ref{thm:rounding}, we now have an integral solution that is at least as good as the optimum, except our solution is infeasible. It's infeasible because it's ``getting credit'' for {(virtual) welfare in the budgeted-additive term} that is perhaps up to twice the budget (i.e. up to $2b_i$). An ``obvious'' fix to this problem might be to take this integral solution and {only take credit for budgeted-additive values up to $b_i$}, thereby making the solution feasible again. Unfortunately, because the objective is mixed sign, the resulting solution doesn't provide any approximation guarantee.\footnote{\mattnote{Consider, for example, the following instance: there is one buyer and two items. $v_{11} = v_{12} = 3$, $b_1 = 3$, $w_{11} = w_{12} = -2$. Then the allocation that awards both items and ``gets credit'' for up to $2b_i$ is believed to have virtual welfare $2$. However, the correctly computed virtual welfare of this allocation is actually $-1$, which clearly provides no meaningful approximation. Instead we must develop a procedure that, on this instance, would allocate just one of the items}.} Instead, we provide a simple procedure to select a feasible suballocation of this infeasible one that loses a factor of 3.
\begin{theorem}\label{thm:2}
Given an integral allocation $\vec{x}$ satisfying $\sum_i \bar{x}_{ij} + \hat{x}_{ij} \leq 1$ for all $j$, $\sum_j \bar{x}_{ij} v_{ij} \leq 2b_i$ for all $i$, and $\sum_{ij} m_i \bar{x}_{ij}v_{ij} + \sum_{ij} w_{ij} (\bar{x}_{ij} + \hat{x}_{ij}) = C$, one can find in poly-time an integral allocation $\vec{y}$ such that:
\begin{enumerate}
\item $\sum_i (\bar{y}_{ij} + \hat{y}_{ij}) \leq 1$ for all $j$.
\item $\sum_j \bar{y}_{ij} v_{ij} \leq b_i$ for all $i$.
\item $\sum_{ij} m_i \bar{y}_{ij}v_{ij} + \sum_{ij} w_{ij} (\bar{y}_{ij} + \hat{y}_{ij}) \geq C/3$.
\end{enumerate}
\end{theorem}
\begin{proof}
For each $i$, we wish to partition the set of items assigned to $i$ via $\bar{x}_{ij}$ (of the infeasible integral solution), $S$, into three disjoint sets $S_i^1,S_i^2,S_i^3$ such that $\sum_{j \in S_i^k} v_{ij} \leq b_i$ for all $k$. This is always possible: consider sorting the elements in decreasing order of $v_{ij}$ and greedily adding them one at a time to the $S_i^k$ with minimal weight so far. Assume for contradiction that some item $j^*$, when added, pushes some $S_i^k$ from below $b_i$ to above $b_i$. Then without $j^*$, each of $S_i^1,S_i^2,S_i^3$ must have had weight strictly larger than $b_i - v_{ij^*}$. As the total weight in all three (without $j^*$) is at most $2b_i - v_{ij^*}$, this means that $2b_i - v_{ij^*} > 3(b_i - v_{ij^*}) \Rightarrow v_{ij^*} > b_i/2$. But as we processed elements in decreasing order of $v_{ij}$, this would imply that $j^*$ was the third (or earlier) item processed, meaning that some set must have been empty, and $j^*$ couldn't have possibly pushed it over the limit (as $v_{ij} \leq b_i$ for all $j$). Therefore, at termination we must have $\sum_{j \in S_i^k} v_{ij} \leq b_i$ for all $k$. Now, define $k^* = \arg \max_{k}\{\sum_{j \in S_i^k}m_iv_{ij} + w_{ij}\}$. Let $\bar{y}_{ij} = 1$ iff $j \in S_i^{k^*}$, and $\hat{y}_{ij} = \hat{x}_{ij}$ for all $j$.
It's clear that $\sum_j \bar{y}_{ij}v_{ij} \leq b_i$ for all $i$. As $\bar{y}_{ij} \leq \bar{x}_{ij}$ for all $i,j$, it's also clear that $\sum_i \bar{y}_{ij} + \hat{y}_{ij} \leq 1$ for all $j$. Finally, by choice of $k^*$ it's also clear that $\sum_{ij} (m_iv_{ij}+w_{ij})\bar{y}_{ij} \geq \sum_{ij} (m_i v_{ij} + w_{ij})\bar{x}_{ij}/3$, and therefore $\sum_{ij} m_i \bar{y}_{ij}v_{ij} + \sum_{ij} w_{ij} (\bar{y}_{ij} + \hat{y}_{ij}) \geq C/3$, as desired.
\end{proof}
Combining Theorems~\ref{thm:rounding} and~\ref{thm:2} yields a feasible, integral allocation that is a 3-approximation by rounding the fractional solution output by our LP, and it is easy to see that the entire procedure runs in polynomial time.
\begin{theorem}\label{thm:algorithm}
There is a poly-time 3-approximation algorithm for Budgeted-Additive Virtual Welfare Maximization, which is a reformulation of GOOP($[n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue}$). Therefore, for all $\epsilon > 0$, there is a poly-time $(\epsilon,3)$-approximation algorithm for BMeD($[n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue}$). Specifically, if $\ell$ is the input length to an instance of BMeD($[n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue}$), the algorithm terminates in time $\text{poly}(\ell, 1/\epsilon)$ and succeeds with probability $1-\text{exp}(-\text{poly}(\ell, 1/\epsilon))$.
\end{theorem}
We conclude this section with a remark about the special case of a single (or small constant) number of items. Notice that BAVWM can be solved exactly by exhaustive search in time $\text{poly}(n^m)$. If $m$ is a small constant, exhaustive search may be computationally feasible, resulting in an exact algorithm (instead of a $3$-approximation).
\begin{remark}
Budgeted-Additive Virtual Welfare Maximization can be solved exactly in time $\text{poly}(n^m)$ by exhaustive search. Therefore, for all $\epsilon > 0$, there is an $(\epsilon,1)$-approximation algorithm for BMeD($[n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue}$). Specifically, if $\ell$ is the input length to an instance of BMeD($[n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue}$), the algorithm terminates in time $\text{poly}(\ell, n^m,1/\epsilon)$ and succeeds with probability $1-\text{exp}(-\text{poly}(\ell,1/\epsilon))$.
\end{remark}
Finally, we remark that the single-item case is \emph{especially} simpler than even the two item case. We refer the reader to~\cite{CaiDW12,CaiDW13b} for complete details, but essentially the sampling procedure that results in the $\epsilon$ error of Theorem~\ref{thm:CDW} can be replaced by an exact computation \emph{only} in the single item case (and not even in the two item case), and $\epsilon$ can be set to exactly $0$.
\begin{remark}
\label{remark:singleitemVWM}
Budgeted-Additive Virtual Welfare Maximization with $m = 1$ can be solved exactly in time $\text{poly}(n)$ by exhaustive search: there are only $n$ possible outcomes, corresponding to assigning the item to exactly one of the agents. Therefore, there is a $(0,1)$-approximation algorithm (i.e. an exact algorithm) for BMeD($[n+1],\mathbb{R}^{2}_+,\textsc{Revenue}$) (i.e. the single item case). Specifically, if $\ell$ is the input length to an instance of BMeD($[n+1],\mathbb{R}^{2}_+,\textsc{Revenue}$), the algorithm terminates in time $\text{poly}(\ell)$, and succeeds with probability $1$.
\end{remark}
\subsection{Structural Results}\label{sec:structure}\mattnote{
In this section, we discuss the structure of the optimal mechanism, and of the computationally efficient mechanism from Section~\ref{sec:computation}. We begin by characterizing the optimal mechanism by combining Theorem~\ref{thm:CDWstructure} with Observation~\ref{obs:1}.}
\begin{theorem}\label{thm:structure}\mattnote{ In any BMeD($[n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue}$) instance, the optimal mechanism can be implemented as a distribution over virtual welfare maximizers. Specifically, there exists a distribution $\Delta$ over mappings $(f_1^\delta,\ldots,f_n^\delta)$. Each mapping $f_i^\delta$ maps types $(\vec{v}_i,b_i) \in \mathbb{R}_+^{m+1}$ to a multiplier $m_i^\delta(\vec{v}_i,b_i) \in \mathbb{R}_+$ and a vector $\vec{w}^\delta(\vec{v},b_i) \in \mathbb{R}^m$. Define $\phi_i^\delta$ to be the mapping that takes as input types $(\vec{v}_i,b_i) \in \mathbb{R}_+^{m+1}$ and outputs a valuation function $\phi_i^\delta(\vec{v}_i,b_i)(\cdot)$ with $\phi_i^\delta(\vec{v}_i,b_i)(S) = m_i^\delta(\vec{v}_i,b_i) \cdot \min\{b_i,\sum_{j \in S} v_{ij}\} + \sum_{j \in S} w^\delta_{ij}(\vec{v}_i,b_i)$. The allocation rule of the optimal mechanism first samples $(f_1^\delta,\ldots,f_n^\delta)$ from $\Delta$, and on profile $(\vec{v},\vec{b})$, allocates the items according to $\arg \max_{S_1\sqcup \ldots \sqcup S_n \subseteq [m]} \{\sum_i \phi_i^\delta(\vec{v}_i,b_i)(S_i)\}$. Furthermore, if $m_i^\delta(\vec{v}_i,b_i) > 0$, bidder $i$ is charged $\min\{b_i,\sum_{j \in S_i} v_{ij}\}$. If $m_i^\delta(\vec{v}_i,b_i) = 0$, then bidder $i$ is charged $0$.}
\end{theorem}
\begin{proof}
\mattnote{The proof starts with an application of Theorem~\ref{thm:CDWstructure} to the problem BMeD($[n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue}$). By Observation~\ref{obs:1}, the joint optimization over allocations $x$ and price vectors $\vec{p}$ can be accomplished by transforming the optimization into one that depends only on the allocation. Once the allocation is found, optimization of the price vector follows as in Observation~\ref{obs:1}. }
\end{proof}
\mattnote{
We remark that the virtual types involved in Theorem~\ref{thm:structure} have valuation functions that are the sum of a budgeted-additive function, and an additive function (the latter may have negative item values). We also note that the budgeted-additive component depends in a very structured way on the input type $(\vec{v}_i,b_i)$. Specifically, $b_i$ is turned into a hard cap on the bidder's maximum valuation instead of a hard budget on her ability to pay, and the additive valuation $\vec{v}_i$ is kept the same, forming a budgeted-additive function that is scaled by a positive multiplier $m_i$. The multiplier $m_i$ and additional values $\vec{w}_i$ may show little structure with respect to the input types (or perhaps none at all). }
We also remark that the structure is especially simple in the case of a single item, because a budgeted-additive function for a single item is just a typical valuation function (where the bidder's value for the item is the minimum of her value and her budget). Specifically, the virtual type parameterized by $m_i^\delta(v_i,b_i)$ and $w_i(v_i,b_i)$ values the item at $m_i \min\{v_i,b_i\} + w_i(v_i,b_i)$. This observation leads to the following simplification:
\begin{remark}\label{remark:structure}
\mattnote{ In any BMeD($[n+1],\mathbb{R}^{2}_+,\textsc{Revenue}$) instance (i.e. the single item case), the optimal mechanism can be implemented as a distribution over virtual value maximizers. Specifically, there exists a distribution $\Delta$ over mappings $(f_1^\delta,\ldots,f_n^\delta)$. Each mapping $f_i^\delta$ maps types $(v_i,b_i) \in \mathbb{R}_+^{2}$ to an indicator bit $m_i^\delta(v_i,b_i) \in \{0,1\}$ and a virtual value $\phi_i^\delta(v_i,b_i)$. The allocation rule of the optimal mechanism first samples $(f_1^\delta,\ldots,f_n^\delta)$ from $\Delta$, and on profile $(\vec{v},\vec{b})$, allocates the item to any bidder $i^* \in \arg\max_{i} \{\phi_i^\delta(v_i,b_i)\}$ if her virtual value is non-negative, and doesn't allocate the item otherwise. Furthermore, if $m_{i^*}^\delta({v}_{i^*},b_{i^*}) =1$, bidder $i^*$ is charged $\min\{b_{i^*},v_{i^*}\}$. If $m_{i^*}^\delta({v}_{i^*},b_{i^*}) = 0$, then bidder $i^*$ is charged $0$.}
\end{remark}
\mattnote{We conclude with a statement regarding the format of our computationally efficient mechanisms from Section~\ref{sec:computation}. This is an instantiation of Algorithm~2 in~\cite{CaiDW13b}, which is used to prove Theorem~\ref{thm:CDW}. }
\begin{theorem}\label{thm:format} The mechanism providing the guarantee of Theorem~\ref{thm:algorithm} has the following format:
\\
\textbf{Phase One, Find the Mechanism:}
\begin{enumerate}
\item Write a linear program that optimizes revenue over the space of truthful, feasible interim forms (Section~\ref{sec:instantiation}).
\item Pick an $\epsilon > 0$. Using the algorithm developed in Section~\ref{sec:computation}, and the reduction of~\cite{CaiDW13b}, solve this linear program approximately.
\item This yields an interim form corresponding to a mechanism that is an $(\epsilon,3)$-approximation.
\item The linear program also outputs auxiliary information in the form of a distribution $\Delta$ over mappings $(f_1^\delta,\ldots,f_n^\delta)$ of the same format from Theorem~\ref{thm:structure}.
\end{enumerate}
\textbf{Phase Two, Run the Mechanism:}
\begin{enumerate}
\item Sample a mapping from $\Delta$ (provided in Phase One).
\item On profile $(\vec{v},b)$, run the approximation algorithm of Section~\ref{sec:computation} for Budgeted-Additive Virtual Welfare Maximization, with input budgets $b_i$, input values $v_{ij}$, input price multipliers $m_i^\delta(\vec{v}_i,b_i)$, and input virtual values $w_{ij}^\delta(\vec{v}_i,b_i)$. Select this allocation.
\item If $m_i^\delta(\vec{v}_i,b_i) > 0$, charge bidder $i$ the minimum of their budget and their value for the items they receive. Otherwise, charge them nothing.
\end{enumerate}
\end{theorem}
Note that this mechanism has basically the same structure as the optimal mechanism, except that on every profile it only approximately maximizes virtual welfare (and we also first have to find the mechanism, which is completely described by the distribution $\Delta$). In the special case of a single item, the structure can again be simplified.
\begin{remark}
In the special case of a single item, the following algorithm finds the optimal mechanism in polynomial time:
\\
\textbf{Phase One, Find the Mechanism:}
\begin{enumerate}
\item Write a linear program that optimizes revenue over the space of truthful, feasible interim forms (Section~\ref{sec:instantiation}).
\item Using the reduction of~\cite{CaiDW13b} and the observation in Remark \ref{remark:singleitemVWM} that Budgeted-Additive Virtual Welfare Maximization with $m = 1$ can be solved exactly, solve this linear program exactly.
This yields an interim form corresponding to the optimal mechanism.
\item The linear program also outputs auxiliary information in the form of a distribution $\Delta$ over mappings $(f_1^\delta,\ldots,f_n^\delta)$ of the same format from Remark~\ref{remark:structure}.
\end{enumerate}
\textbf{Phase Two, Run the Mechanism:}
\begin{enumerate}
\item Sample a mapping from $\Delta$ (provided in Phase One).
\item On profile $(\vec{v},b)$, award item $j$ to any bidder $i^* \in \arg\max_{i} \{\phi_i^\delta(v_i,b_i)\}$ if her virtual value is non-negative. Don't allocate item $j$ otherwise.
\item If $m_i^\delta(\vec{v}_i,b_i) =1$, charge bidder $i$ the minimum of their budget and their value for the items they receive. Otherwise, charge them nothing.
\end{enumerate}
\end{remark}
\section{Introduction}
Most of auction theory crucially depends on the assumption of quasi-linear utilities, that the utility is equal to valuation minus payments.
This assumption fails when bidders are budget constrained.\footnote{The terms financially constrained bidders or bidders with liquidity constraints are used synonymously.}
Auctions with budget constrained bidders are commonplace, and prominent examples of this are ad-auctions and auctions for government licensing such as the FCC spectrum auction.
An interesting example of budget constraint occurs in the auction for professional cricket players in the Indian Premier League: the league imposes a budget constraint on all the teams as a means of ensuring well balanced teams.
Another source of budget constraints is what \citet{CheGale1998} call the {\em moral hazard problem}: procurement is often delegated and budget constraints are imposed as a means of controlling the spend.
A budget represents the bidder's \emph{ability to pay}, in contrast to the valuation which represents his {\em willingness to pay}.
For this reason, budgets may be more tangible and easier to estimate than valuations.
It is therefore important to understand how budget constraints impact the design of auctions; this has been well established by now \citep{CheGale1998, PaiVohra, BenoitKrishna, LaffontRobert, Maskin, MalakhovVohra, CheGale2000, BhattacharyaGGM10}.
The theory of auctions in the presence of budget constraints on bidders lags far behind the theory of auctions without budgets.
For instance, consider the design of {\em optimal} (revenue maximizing) auctions that are Bayesian incentive compatible (BIC) and ex-post individually rational (IR).
While \citet{Myerson} gives a beautiful theory characterizing the optimal auction for any single parameter domain, no such characterization is known in the presence of private budgets (that could be correlated with the valuation).
As a way to deal with this difficulty, previous papers have considered special cases and auctions with a subset of the desired properties. (See \prettyref{sec:related} for details.) We adopt the Computer Science approach of approximation, while incorporating all the desired properties. {\em The main result of this paper is a 3-approximation to the optimal auction in the class of auctions that are
\begin{itemize}
\item BIC,
\item ex-post IR and
\item ex-post budget respecting, with private budgets that could be correlated with the valuations,
\end{itemize}
for multiple heterogenous items and additive valuations.} This is the first constant factor approximation for this class of auctions.
Moreover, the computational problem, even without any incentive constraints is already NP-Hard to approximate within a ratio of $16/15$ \citep{ChakrabartyGoel}. This too suggests that an approximation is necessary.
\subsection{Overview of Techniques}
{We prove our main result by making use of an algorithmic framework developed in~\cite{CaiDW13b}. The computational aspect of their framework provides a \emph{black-box reduction} from a wide class of Bayesian mechanism design problems to problems of purely algorithm design. More specifically, they show that any $\alpha$-approximation algorithm for a certain incentive-free algorithmic problem (induced by the mechanism design problem at hand) can be leveraged to find a BIC, IR mechanism that is also an $\alpha$-approximation (to the optimal BIC, IR mechanism) in polynomial time. Significant further details on their reduction and how to employ it can be found in Section~\ref{sec:CDW}. After applying their framework to our problem, there is still the issue of solving the algorithmic problem that pops out of the reduction. This turns out to be essentially a (virtual) welfare maximization problem (without budgets), but where bidder types are somewhat involved. The optimization involves a mixed sign objective (i.e. the objective is a sum of several terms which can be positive or negative). Such optimization problems are typically solvable \emph{exactly} in polynomial time or computationally hard to approximate within any finite factor, but rarely in between (due to the mixed signs in the objective). Interestingly, we obtain a 3-approximation for our mixed-sign objective problem despite the fact that it is NP-hard to optimize exactly. The design and analysis of our algorithm can be found in Section~\ref{sec:main}.}
{Cai et. al.'s framework also contains a structural result. We use it to show that the optimal auction in our setting is a \emph{distribution} over virtual welfare maximizers. By this, we mean that the optimal mechanism maintains a distribution over $n$ mappings, one mapping per bidder that maps types to virtual types, and, given a vector of reported types, it samples $n$ mappings from this distribution, uses them to map the reported types to virtual types, and proceeds to choose an allocation that optimizes virtual welfare. Note that by virtual types in the previous sentence we do \emph{not} mean the specific virtual types as computed by Myerson's virtual transformation, which aren't even defined for multi-dimensional types, but just \emph{some} virtual types that may or may not be the same as the true types. In particular, each mapping in the support of the mechanism's distribution will take as input a type (which is an additive function with non-negative item values plus a non-negative budget), and output a virtual type without a budget constraint and whose valuation function is the sum of a budgeted-additive function\footnote{{A function $v(\cdot)$ is budgeted-additive if there exists a $b$ such that $v(S) = \min\{b,\sum_{i \in S} v(\{i\})\}$ for all $S$. Note that this is different from an additive buyer with a budget, and that a budgeted-additive buyer indeed has quasi-linear utilities.}} with non-negative item values (which depends on the input type in a very structured way) plus an additive function with possibly negative item values (which may be unstructured with respect to the input type). We provide a formal statement of this structural claim in Section~\ref{sec:main} as well.}
Note that for the special case of a single item auction, this gives a particularly simple structure: the virtual types are now {just a single (possibly negative) real number, which could be interpreted as a virtual value.} The optimal auction simply maps reported types to virtual {\em values} and assigns the item to the bidder with the highest virtual value.
\input{related}
\subsection{Conclusions and Future Work}
{The goals of revenue-optimality, ex-post individual rationality, and ex-post budget feasibility seem to be at odds with one another. This is highlighted by the fact that, prior to our work, no known auctions even approximately satisfied all three conditions, even with just a single item and private budgets that are independent of values. We provide a computationally efficient 3-approximation for the significantly more general case of auctions for multiple heterogeneous goods and additive bidders with private budgets that can be correlated with their values. While this model is already quite general compared to the previous state-of-the-art, it is an important direction to see if our results can be extended to more complex classes of bidder valuations, or to more complex constraints on feasible allocations.}
In particular, well studied classes of valuations such as gross substitutes would be interesting next steps.
\section{Preliminaries}
We begin with formal definitions of the mechanism design problem we study.
We then outline the reduction of \citet{CaiDW13b} (\prettyref{sec:CDW})
and its implications (\prettyref{sec:instantiation}) for our problem.
Finally we state a related problem (\prettyref{sec:gap}), the Generalized Assignment Problem, which we use in the design of our algorithm.
\paragraph{Bidders} There are $n$ bidders, each with additive valuations over $m$ items and a hard budget constraint. Specifically, bidder $i$ has value $v_{ij}$ for item $j$, value $\sum_{j\in S} v_{ij}$ for set $S$, and hard budget $b_i$. We denote by $\vec{v}_i$ the vector of bidder $i$'s values for all $m$ items. We denote by $\mathcal{D}_i$ the joint distribution of $(\vec{v}_i, b_i)$. We denote by $\mathcal{D} = \times_i \mathcal{D}_i$ the joint distribution of all bidders' valuations and budgets.
\paragraph{Mechanisms} Our goal is to design \emph{Bayesian Incentive Compatible} (BIC) mechanisms that are \emph{ex-post Individually Rational} (IR) and that respect budgets ex-post. Formally, for a (randomized) mechanism $M$, we can denote by $x^M_{ij}(\vec{v},\vec{b},r)$ to be $1$ if bidder $i$ receives item $j$ when the profile of values/budgets reported to $M$ is $(\vec{v},\vec{b})$ and the random seed used by $M$ is $r$, or $0$ otherwise. Similarly, we denote by $q^M_{i}(\vec{v},\vec{b},r)$ to be the price paid by bidder $i$ in the same conditions. We can then define the interim allocation probability $\pi^M_{ij}(\vec{v}_i,b_i)$ to be the probability that bidder $i$ receives item $j$ when reporting $(\vec{v}_i,b_i)$ over the randomness of other agent's (valuation,budget)s $(\vec{v}_{-i},\vec{b}_{-i})$ being drawn from $\mathcal{D}_{-i}$, and any randomness in $M$. We can similarly define the interim price $p^M_i(\vec{v}_i,b_i)$ to be the expected payment made by bidder $i$ over the same randomness. Formally, $\pi^M_{ij}(\vec{v}_i,b_i) = \mathbb{E}_{(\vec{v}_{-i},\vec{b}_{-i})\leftarrow \mathcal{D}_{-i}, r}[x^M_{ij}(\vec{v_i};\vec{v}_{-i},b_i;\vec{b}_{-i},r)]$ and $p^M_{ij}(\vec{v}_i,b_i) = \mathbb{E}_{(\vec{v}_{-i},\vec{b}_{-i})\leftarrow \mathcal{D}_{-i}, r}[q^M_{ij}(\vec{v_i};\vec{v}_{-i},b_i;\vec{b}_{-i},r)]$. Formal definitions of BIC, IR, and ex-post budgets are below.
\begin{definition}\label{def:BIC}(Bayesian Incentive Compatible) A mechanism $M$ is BIC if for all bidders $i$, and types $(\vec{v}_i,b_i), (\vec{v}'_i, b'_i)$ the following holds:
$$\vec{v}_i \cdot \vec{\pi}^M_i(\vec{v}_i, b_i) - p^M_i(\vec{v}_i, b_i) \geq \vec{v}_i \cdot \vec{\pi}^M_i(\vec{v}'_i, b'_i) - p^M_i(\vec{v}'_i,b'_i).$$
\noindent A mechanism is said to be $\epsilon$-BIC if for all bidders $i$, and types $(\vec{v}_i,b_i), (\vec{v}'_i, b'_i)$ the following holds:
$$\vec{v}_i \cdot \vec{\pi}^M_i(\vec{v}_i, b_i) - p^M_i(\vec{v}_i, b_i) \geq \vec{v}_i \cdot \vec{\pi}^M_i(\vec{v}'_i, b'_i) - p^M_i(\vec{v}'_i,b'_i) - \epsilon.$$
\end{definition}
\begin{definition}\label{def:IIR}(Interim/Ex-Post Individually Rational) A mechanism $M$ is interim IR if for all bidders $i$, and types $(\vec{v}_i,b_i)$ the following holds:
$$\vec{v}_i \cdot \vec{\pi}^M_i(\vec{v}_i, b_i) \geq p^M_i(\vec{v}_i, b_i) .$$
Further, it is ex-post IR if for all bidders $i$, all profiles $(\vec{v},\vec{b})$ and random seeds $r$, we have:
$$ \vec{v}_i \cdot \vec{x}^M_i(\vec{v},\vec{b},r)\geq q^M_i(\vec{v},\vec{b},r).$$
\end{definition}
\begin{definition}(Ex-Post Budget Respecting) A mechanism $M$ respects budgets ex-post if for all type profiles $(\vec{v},\vec{b})$, all random seeds $r$, and all bidders $i$ we have:
$$q^M_i(\vec{v},\vec{b},r) \leq b_i.$$
\end{definition}
\begin{definition}(No Positive Transfers) A mechanism $M$ has no positive transfers if for all type profiles $(\vec{v},\vec{b})$, all random seeds $r$, and all bidders $i$ we have:
$$q^M_i(\vec{v},\vec{b},r) \geq 0.$$
\end{definition}
\subsection{Reduction from Mechanism to Algorithm Design}\label{sec:CDW} In recent work,~\cite{CaiDW13b} provide an algorithmic framework for mechanism design, showing how to design mechanisms by solving purely algorithmic problems. We use this reduction to reduce our mechanism design problem to an algorithm design problem and show a 3-approximation to this algorithmic problem. In the rest of this section, we state the general formulations of the mechanism design and the corresponding algorithm design problems considered by~\citet{CaiDW13b}. Then we give the precise statement of their reduction, and a structural characterization of the optimal mechanism obtained as a byproduct of their reduction. Finally we instantiate these to state the corresponding problems in our setting, and massage the resulting problems to simplify them.
\cite{CaiDW13b} call the mechanism design problems of study BMeD($\mathcal{F},\mathcal{V},\mathcal{O}$),\footnote{BMeD stands for \textbf{B}ayesian \textbf{Me}chanism \textbf{D}esign.} where feasibility constraints $\mathcal{F}$, possible valuations $\mathcal{V}$, and optimization objective $\mathcal{O}$ parameterize the problem. Formally, this problem is defined as:\\
\noindent\textbf{BMeD($\mathcal{F},\mathcal{V},\mathcal{O}$)}:\\
\textsc{Input}: For each bidder $i \in [n]$, a finite set $T_i \subseteq \mathcal{V}$, and a distribution $D_i$ over $T_i$, presented by explicitly listing all types in $T_i$ and their corresponding probability.\\
\textsc{Output}: A feasible (selects an allocation in $\mathcal{F}$ with probability $1$), BIC, (interim) IR mechanism for bidders drawn from $D = \times_i D_i$.\\
\textsc{Goal}: Find the mechanism that optimizes $\mathcal{O}$ in expectation, with respect to all BIC, IR mechanisms (when bidders with types drawn from $D$ play truthfully).\\
\textsc{Approximation}: An algorithm is said to be an $(\epsilon,\alpha)$-approximation if it finds an $\epsilon$-BIC mechanism whose expected value of $\mathcal{O}$ (when bidders drawn from $D$ report truthfully) is at least $\alpha\cdot \text{OPT} - \epsilon$.\\
In our problem, the feasible allocations are those that award each item to at most one bidder. So we could denote the set of feasible allocations as $[m+1]^n$ (with the convention that selecting the allocation $\vec{a}$ awards item $j$ to bidder $a_j$ if $a_j > 0$, or no one if $a_j = 0$). The possible bidder types are all additive functions over items (with non-negative multipliers), and non-negative budgets, which we could denote by $\mathbb{R}^{m+1}_+$. Our objective is revenue. To ensure that all feasible mechanisms are ex-post IR (note that their reduction only guarantees interim IR without extra work) and ex-post budget respecting, we will define the objective function $\textsc{Revenue}$ as follows. $\textsc{Revenue}$ takes as input a valuation profile $(\vec{v},\vec{b})$, an allocation $\vec{x}$ (where $x_{ij} = 1$ iff bidder $i$ is awarded item $j$), and a price vector $\vec{p}$. We define $\textsc{Revenue}(\vec{v},\vec{b},\vec{x},\vec{p}) = \sum_i p_i$, if $0 \leq p_i \leq \min\{b_i, \vec{v}_i\cdot \vec{x}_i\}$ for all $i$, or $\textsc{Revenue}(\vec{v},\vec{b},\vec{x},\vec{p}) = -\infty$ otherwise.
{There is a subtle issue with respect to \emph{why} we want to design mechanisms that respect budgets ex-post. Specifically, is it just because the designer wishes to offer this guarantee to the bidders, who have true quasi-linear preferences? If so, then this is exactly the setting we have described so far: the designer is constrained to select a mechanism that respects budgets ex-post, but bidders will still choose how to play as if they were quasi-linear. While this motivation is certainly mathematically interesting, it is also non-standard and perhaps unrealistic. Instead, the more common motivation is because bidders physically can't pay more than their budget, and would have utility $-\infty$ if asked to do so. In this case, the designer should actively \emph{exploit} this to extract higher revenue. For example, if for all $i$, bidder $i$ is awarded all the items and charged her budget with tiny probability $\epsilon /n$, then the designer needn't worry about bidders overreporting their budget (as otherwise they'd get utility $-\infty$ with probability $\epsilon /n$). Therefore, a mechanism can be made BIC in the latter case (while losing arbitrarily little revenue) iff for all $(\vec{v}_i, b_i), (\vec{v}'_i, b'_i)$ \textbf{with $b'_i \leq b_i$}, $\vec{v}_i \cdot \vec{\pi}^M_i(\vec{v}_i, b_i) - p^M_i(\vec{v}_i, b_i) \geq \vec{v}_i \cdot \vec{\pi}^M_i(\vec{v}'_i, b'_i) - p^M_i(\vec{v}'_i,b'_i).$ Note that this is a relaxed condition of the former setting, which requires the inequality to hold \textbf{for all $b_i, b'_i$}. Fortunately, the Cai et. al. framework applies in both settings, and the resulting structure and algorithmic problem are exactly the same. So all of our theorems, exactly as stated, hold in both of the described settings.}
Informally, the main result of~\cite{CaiDW13b} states that, for all $\mathcal{F}, \mathcal{V}, \mathcal{O}$, the problem BMeD($\mathcal{F},\mathcal{V},\mathcal{O}$) can be solved in polynomial time with black-box access to a poly-time algorithm for a purely algorithmic problem that they call GOOP($\mathcal{F},\mathcal{V},\mathcal{O}$).\footnote{GOOP stands for \textbf{G}eneralized \textbf{O}bjective \textbf{O}ptimization \textbf{P}roblem.} Below, $\mathcal{V}^\times$ denotes the closure of $\mathcal{V}$ under addition and (possibly negative) scalar multiplications (so for instance, $(\mathbb{R}^m_+)^\times = \mathbb{R}^m$).\\
\noindent\textbf{GOOP($\mathcal{F},\mathcal{V},\mathcal{O}$)}:\\
\textsc{Input}: A type $t_i \in \mathcal{V}$, multiplier $m_i \in \mathbb{R}$, {and virtual valuation function $g_i \in \mathcal{V}^\times$ for each $i \in [n]$.}\footnote{For other applications, the inputs $g_i(\cdot)$ are sometimes called instead cost functions.} \\
\textsc{Output}: An allocation $x \in \mathcal{F}$ and price vector $\vec{p} \in \mathbb{R}_+^n$.\\
\textsc{Goal}: Find $\arg \max_{x \in \mathcal{F},\vec{p}}\{\mathcal{O}(\vec{t},x,\vec{p}) + \sum_i m_i p_i + \sum_i g_i(x)\}$.\\
\textsc{Approximation}: $(x^*,\vec{p}^*)$ is said to be an $\alpha$-approximation if $\mathcal{O}(\vec{t},x^*,\vec{p}^*) + \sum_i m_i p_i^*+ \sum_i g_i(x^*) \geq \alpha \cdot \arg \max_{x\in \mathcal{F},\vec{p}}\{\mathcal{O}(\vec{t},x,\vec{p}) + \sum_i m_i p_i+ \sum_i g_i(x)\}$.\\
Further below we provide much more detail on the structure of the algorithmic focus of this paper, GOOP($[n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue}$), but we first conclude our discussion of the reduction we employ. The main result of~\cite{CaiDW13b} states that for all $\epsilon > 0$, an $(\epsilon,\alpha)$-approximation for BMeD($\mathcal{F},\mathcal{V},\mathcal{O}$) can be obtained from a poly-time $\alpha$-approximation for GOOP($\mathcal{F},\mathcal{V},\mathcal{O}$). The additive error (and failure probability in the theorem statement) is due to a sampling procedure in the execution of the reduction. We provide a full statement of their main result below.\footnote{The theorem statement is identical in content, but reworded for clarity and cleanliness.}
\begin{theorem}\label{thm:CDW}(Theorem~4 of~\cite{CaiDW13b}) For all $\mathcal{F},\mathcal{V},\mathcal{O}$, and $\epsilon > 0$, if there is a poly-time $\alpha$-approximation algorithm, $G$, for GOOP($\mathcal{F},\mathcal{V},\mathcal{O}$), there is a poly-time $(\epsilon,\alpha)$-approximation algorithm for BMeD($\mathcal{F},\mathcal{V},\mathcal{O}$) as well. Specifically, if $\ell$ denotes the input length of a BMeD($\mathcal{F},\mathcal{V},\mathcal{O}$) instance, the algorithm runs in time $\text{poly}(\ell, 1/\epsilon)$, makes $\text{poly}(\ell, 1/\epsilon)$ black box calls to $G$ on inputs of size $\text{poly}(\ell, 1/\epsilon)$, and succeeds with probability $1-\text{exp}(-\text{poly}(\ell, 1/\epsilon))$.
\end{theorem}
\cite{CaiDW13b} prove Theorem~\ref{thm:CDW} above by considering a linear program that optimizes over the space of interim forms that are both truthful (that satisfy the linear constraints in Definitions~\ref{def:BIC} and~\ref{def:IIR}), and feasible (those that correspond to an actual mechanism that selects an outcome $x \in \mathcal{F}$ on every profile with probability $1$).\footnote{In fact, they need to work with a generalization of interim forms, called {\em implicit forms}, to accommodate non-additive valuations. But we describe their proof for additive valuations for clarity of exposition, and because it is relevant for our setting.} Linear constraints enforcing that an interim form is BIC and interim IR can be written explicitly, but a computationally efficient separation oracle for the space of feasible interim forms is still required in order to solve the linear program. They show how to obtain such a separation oracle with black-box access to an algorithm that solves GOOP, and that this entire process preserves approximation as well.
\cite{CaiDW13b} further provide a structural characterization of the space of all feasible mechanisms (truthful or not), leading to a structured implementation of whatever interim form is output by the LP. Specifically, they show that the extreme points of the space of feasible interim forms correspond to mechanisms that associate a virtual valuation function $g_i(t_i)(\cdot)$ and price multiplier $m_i(t_i)$ to each type $t_i \in T_i$, and then selects on profile $(t_1,\ldots,t_n)$ the allocation and price vector that solves GOOP on input $t_1,\ldots,t_n$, $m_1(t_1),\ldots,m_n(t_n)$, $\sum_i g_i(t_i)(\cdot)$. They show further that solving the linear program explicitly finds a list of virtual valuation functions and multipliers whose resulting interim forms contain the optimal (truthful) interim form in their convex hull. Theorem~\ref{thm:CDWstructure} below captures the structural aspect of their result.
\begin{theorem}\label{thm:CDWstructure}(Implicit in~\cite{CaiDW13b}) For all BMeD instances, the optimal mechanism can be implemented as a distribution over generalized objective optimizers. Specifically, there exists a distribution $\Delta$ over mappings $(f^\delta_1,\ldots,f^\delta_n)$. Each mapping $f^\delta_i$ takes types $t_i$ in $T_i$ to price multipliers $m^\delta_i(t_i) \in \mathbb{R}$ and virtual valuation functions $g^\delta_i(t_i)(\cdot) \in \mathcal{V}^\times$. The optimal mechanism first samples $(f^\delta_1,\ldots,f^\delta_n)$ from $\Delta$, and on profile $\vec{t}$, selects the outcome and price vector $\arg\max_{x \in \mathcal{F},\vec{p}}\{\mathcal{O}(\vec{t},x,\vec{p})+\sum_i m^\delta_i(t_i)\cdot p_i + \sum_i g^\delta_i(t_i)(x)\}$.
\end{theorem}
In the section below, we provide further details surrounding instantiations of Theorems~\ref{thm:CDW} and~\ref{thm:CDWstructure} as they pertain to the problem at hand.
\subsection{Instantiations}\label{sec:instantiation}
The goal of this section is to provide more details of the instantiation of Theorems~\ref{thm:CDW} and~\ref{thm:CDWstructure} to our setting, but not to provide proofs (for which we refer the reader to~\cite{CaiDW13b}). We begin by describing the linear program that the reduction of~\cite{CaiDW13b} would try to solve for our setting. Below, $F([n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue})$ denotes the space of interim forms of all feasible (not necessarily truthful) mechanisms. Specifically, $(O,\vec{\pi},\vec{p}) \in F([n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue})$ if and only if there is a mechanism $M$ that awards each item at most once on every profile, is ex-post IR and ex-post budget respecting, awards bidder $i$ item $j$ when she reports type $t_i$ with probability exactly $\pi_{ij}(t_i)$ (w.r.t. all other bidders' types and the randomness in the mechanism) and chargers bidder $i$ price $p_i(t_i)$ in expectation (over all other bidders' types and the randomness in the mechanism), and whose expected revenue is exactly $O$. With this definition in mind, the linear program they solve is stated below.\\
\noindent\textbf{Variables:}
\begin{itemize}
\item $O$, denoting the expected revenue of the interim form found.
\item $\pi_{ij}(t_i)$ for all bidders $i$, items $j$, types $t_i$, denoting the probability that bidder $i$ receives item $j$ when reporting type $t_i$.
\item $p_i(t_i)$ for all bidders $i$ and types $t_i$, denoting the expected price paid by bidder $i$ when reporting type $t_i$.
\end{itemize}
\textbf{Constraints:}
\begin{enumerate}
\item $\sum_j \pi_{ij}(t_i) \cdot v_{ij}(t_i) - p_i (t_i) \geq \sum_j \pi_{ij}(t'_i) \cdot v_{ij}(t_i) - p_i(t'_i)$, for all bidders $i$ and types $t_i, t'_i$, guaranteeing that the interim form corresponds to a BIC mechanism.
\item $\sum_j \pi_{ij}(t_i) \cdot v_{ij}(t_i) - p_i (t_i) \geq 0$, for all bidders $i$ and types $t_i$, guaranteeing that the interim form corresponds to an interim IR mechanism.\footnote{Actually, this constraint is redundant as we will also enforce that the mechanism be ex-post IR to be considered feasible.}
\item $(O,\vec{\pi},\vec{p}) \in F([n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue})$, guaranteeing that the interim form corresponds to a feasible mechanism.
\end{enumerate}
\textbf{Maximizing:}
\begin{itemize}
\item $O$, the expected revenue.
\end{itemize}
The solution to this LP is the interim form of the optimal mechanism. The LP can be solved in polynomial time, so long as we have a poly-time separation oracle for the space $F([n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue})$. \cite{CaiDW13b} shows that this can be obtained via an algorithm for the related GOOP problem, which we instantiate in our setting below.
\paragraph{Budgeted-Additive Virtual Welfare Maximization} As discussed above, in order to find (approximately) optimal mechanisms for our setting, we need to study the purely algorithmic problem GOOP($[n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue}$), which we pose formally below.\\
\noindent\textbf{GOOP($[n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue}$)}:\\
\textsc{Input}: Values $v_{ij} \geq 0$ and virtual values $w_{ij} \in \mathbb{R}$ for all $i, j$. Budget $b_i\in \mathbb{R}_+$ and price multiplier $m_i\in \mathbb{R}$ for all $i$.\\
\textsc{Output}: An allocation $\vec{x} \in \{0,1\}^{mn}$ and prices $\vec{p}$ such that $\sum_i x_{ij} \leq 1$ for all $j$ (each item awarded at most once), $\sum_j x_{ij} v_{ij} \geq p_i$ (ex-post IR), $p_i \leq b_i$ (ex-post budget respecting), and $p_i \geq 0$ (no positive transfers).\\
\textsc{Goal}: Find $\arg \max_{\vec{x},\vec{p}}\{\sum_i (m_i+1)p_i + \sum_{ij} x_{ij} w_{ij}\}$ (virtual revenue plus virtual welfare). \\
\\
Note that in the above formulation, we have folded cases where $\textsc{Revenue}$ evaluates to $-\infty$ into feasibility constraints on the output. We make two quick further observations about the structure of GOOP($[n+1]^m,\mathbb{R}^{m+1}_+,\textsc{Revenue}$), and call the reformulation \textbf{B}udgeted-\textbf{A}dditive \textbf{V}irtual \textbf{W}elfare \textbf{M}aximization (BAVWM). Also, for cleanliness, we will replace the input price multipliers $m_i$ by $m_i - 1$ so that the term in the objective will be $\sum_i m_i p_i$. This is w.l.o.g. as each $m_i$ could be any real number.
\begin{observation}\label{obs:1}
If $m_i > 0$, the optimal choice for $p_i$ is always $\min\{b_i, \sum_j x_{ij} v_{ij}\}$. If $m_i \leq 0$, the best choice for $p_i$ is $0$.
\end{observation}
\begin{observation}\label{obs:2}
For all possible solutions $(\vec{x}, \vec{p})$, the quality of $(\vec{x}, \vec{p})$ for the input instance $(\vec{v},\vec{w},\vec{b},\vec{m})$ is the same as for the instance $(\vec{v}',\vec{w},\vec{b},\vec{m})$ where $v'_{ij} = \min\{v_{ij}, b_i\}$, for all $i,j $.
\end{observation}
In light of these, we may set all negative $m_i$ to $0$, and all $v_{ij}$ to $\min\{v_{ij},b_i\}$ without changing the problem, leading to the following reformulation.\\
\\
\noindent\textbf{Budgeted-Additive Virtual Welfare Maximization}:\\
\textsc{Input}: Budget $b_i$ for all agents. Values $v_{ij} \in [0,b_i]$ for all agents and items. Price multiplier $m_i \geq 0$ for all agents, and virtual value $w_{ij} \in \mathbb{R}$ for all agents and items.\\
\textsc{Output}: An allocation $\vec{x} \in \{0,1\}^{mn}$ such that $\sum_i x_{ij} \leq 1$ for all $j$ (each item awarded at most once).\\
\textsc{Goal}: Find $\arg \max_{\vec{x}}\{\sum_i (m_i \min\{b_i, \sum_j x_{ij} v_{ij}\} + \sum_{j} x_{ij} w_{ij})\}$.\\
Note that in the above formulation, we no longer need to optimize over the price vector, due to Observation~\ref{obs:1}. The problem can now be interpreted as just a welfare maximization problem, where bidder $i$'s valuation function is the sum of a budgeted-additive function (with non-negative item values) and an additive function (with possibly negative item values). Also, note that we can re-formulate the above problem to remove the multipliers $(m_i)_i$ from the input and the objective, by incorporating them in the $b_i$'s and the $v_{ij}$'s. We choose to leave them in so that it is more transparent how the inputs to BAVWM are related to the types reported by the bidders of the mechanism output by the~\citet{CaiDW13b} reduction.
\subsection{The Generalized Assignment Problem}\label{sec:gap}
Our main technical result will make use of a rounding algorithm for the \emph{Generalized Assignment Problem}. We give here a statement of the problem and a rounding theorem due to Shmoys and Tardos~\cite{ShmoysT93}.\\
\\
\noindent\textbf{Generalized Assignment Problem}:\\
\textsc{Input}: Processing times $p_{ij} \in \mathbb{R}_+$ and costs $c_{ij} \in \mathbb{R}$ for all machines $i$ and jobs $j$, capacities $T_i$ for all machines $i$.\footnote{Traditionally, some consider only costs $c_{ij} \in \mathbb{R}_+$, but the result we cite applies for negative costs as well.}\\
\textsc{Output}: An allocation $\vec{x} \in \{0,1\}^{mn}$ of jobs to machines such that $\sum_i x_{ij} = 1$ for all $j$ (each job is assigned) and $\sum_j x_{ij} p_{ij} \leq T_i$ (each machine processes at most its capacity).\\
\textsc{Goal}: Find $\arg \max_{\vec{x}}\{\sum_{i,j} x_{ij} c_{ij}\}$ (total cost).\footnote{Traditionally, it makes sense to minimize total cost. As costs are possibly negative, the use of max or min is irrelevant.}\\
Now, we provide an LP due to Shmoys and Tardos that outputs a fractional solution at least as good as OPT.\\
\noindent\textbf{Variables:}
\begin{itemize}
\item $x_{ij}$, for all machines $i$ and jobs $j$, denoting the fraction of job $j$ assigned to machine $i$.
\end{itemize}
\textbf{Constraints:}
\begin{enumerate}
\item $\sum_i x_{ij}= 1$, for all $j$, guaranteeing that every job is processed exactly once.
\item $\sum_j x_{ij} \leq T_i$, for all $i$, guaranteeing that no machine's capacity is violated.
\item $x_{ij} = 0$ if $p_{ij} > T_i$.
\end{enumerate}
\textbf{Maximizing:}
\begin{itemize}
\item $\sum_{i,j} x_{ij} c_{ij}$, the total cost.
\end{itemize}
\begin{theorem}\label{thm:ST}(\cite{ShmoysT93}) The optimal fractional solution to the above LP can be rounded in polynomial time to an integral solution such that:
\begin{enumerate}
\item $\sum_i x_{ij} = 1$, for all $j$.
\item $\sum_j x_{ij} \leq 2T_i$, for all $i$.
\item $\sum_j x_{ij} c_{ij} \geq \text{OPT}$.
\end{enumerate}
\end{theorem}
\subsection{Related Work}
\label{sec:related}
The result that comes closest to characterizing the optimal auction is that of \citet{PaiVohra}: they characterize the optimal budget respecting BIC auction for a single item. Their auction is implemented as an all-pay auction and is therefore not ex-post IR.
{They show that the optimal BIC, \emph{interim} IR mechanism that respects budgets ex-post takes on a form similar to Myerson's, but with additional \emph{pooling} to enforce that no bidder is asked to pay more than her budget, while also maintaining that no bidder has incentive to underreport their budget.} Earlier, \citet{LaffontRobert} and \citet{Maskin} considered the case where valuations are private information but budgets are common knowledge and identical.
\citet{MalakhovVohra} study the setting where there are two bidders, one has a known budget constraint while the other does not.
\citet{CheGale2000} characterize the optimal pricing scheme for a single item with a single bidder, with private valuation and budget that could be correlated with each other.
The limited special cases considered by these papers point to the difficulty of characterizing the optimal auction, which motivates the search for efficient approximations.
Another line of work ranks different auction formats by the revenue generated in the presence of budgets.
\citet{CheGale1998} compare first price, second price and all-pay auctions, while
\citet{BenoitKrishna} compare sequential and simultaneous auctions
In the computer science tradition, \cite{BhattacharyaGGM10} give a 4-approximation for multiple items with additive valuations, but they assume that the budgets are publicly known, and the auction is not ex-post IR.
\citet{ChawlaMM11} give a 2-approximation in a single parameter domain, but assume that the budgets are public.
They also consider private budgets, where budgets and values are independently distributed, in single parameter matroid domains, and MHR Distributions, and give a $3(1+e)$-approximation. Finally,~\citet{CaiDW12}, provide exactly optimal mechanisms for multiple items, additive valuations and private budgets, but their auctions are interim-IR.
Once again, all these auctions make additional assumptions when compared to us.
\citet{CaiDW13b} give a general reduction from mechanism design to algorithm design, which we use for our results.
For the special case of a single item auction with private budgets, we show that the algorithmic problem obtained through this reduction is quite easy to solve optimally, resulting in exactly optimal single-item auctions with budgets.
However, when there are multiple items the resulting algorithmic problem becomes NP-Hard \citep{ChakrabartyGoel}. We give a 3-approximation to this algorithmic problem which through the reduction gives a 3-approximately optimal multi-item auction with budgets.
Recently, \cite{Bhalgatetal} showed that (a weaker form of) the reduction of \citet{CaiDW13b} could be obtained using the simpler multiplicative weight update method instead of the ellipsoid algorithm used originally, and consider the variant of our setting where the items are \emph{divisible}. The algorithmic problem in this case is once again easy. \citet{DaskalakisW15} also use the reduction in \cite{CaiDW13b} to design an auction for a non-linear objective, namely the \emph{makespan} of an assigment of jobs to machines.
The auction design problem has also been considered in a \emph{worst-case} model, as opposed to a Bayesian model.
A standard framework is that of {\em competitive} auctions, where a bound is shown on the ratio of the revenue of an optimal auction to the revenue of the given auction on any instantiation of valuations and budgets.
\citet{Borgsetal2005} and \citet{Abrams2006} give constant competitive auctions for multi-unit auctions, under an assumption of \emph{bidder dominance}, that the contribution of a single bidder to the total revenue is sufficiently small.
\cite{DevanurHH13} give constant competitive auctions for single parameter downward-closed domains with a public, common budget constraint.
Since the worst-case setting is decidedly more difficult than the Bayesian setting, these results are not comparable to ours.
Another line of work considers the design of \emph{Pareto-optimal} auctions: \citet{DobzinskiLaviNisan} characterize single item auctions that are Pareto-optimal, with public budgets and show an impossibility of a similar auction for private budgets.
\citet{GoelMirrokniPaes-Leme} extend this auction to a more general poly-matroidal setting.
|
\section{Introduction}
It is well known that sets of rough objects (in various senses) are quasi or partially orderable. Specifically in classical or Pawlak rough sets \cite{ZPB}, the set of roughly equivalent sets has a Quasi Boolean order on it while the set of rough and crisp objects is Boolean ordered. In the classical semantic domain or classical meta level, associated with general rough sets, the collection of crisp and rough objects is quasi or partially orderable (or more generally a parthood relation is definable). Under minimal assumptions on the nature of these objects, many orders with rough ontology can be associated - these necessarily have to do with concepts of discernibility. There are situations where the following can happen:
\begin{itemize}
\item {Case-1: The order structure is simple and objects are labeled crisp or otherwise - this can happen for example in communication between agents. }
\item {Case-2: Some of the order structure is lost, but objects are labeled crisp or otherwise - this can happen for example in communication between agents. }
\item {Case-3: The parthood relation among crisp objects is known fully. But the parthood relation on the set of crisp and non crisp objects is not known in full.}
\item {Case-4: All/Some of the order structure is lost, but objects are labeled crisp or otherwise, but the agent had applied some non-rough method of arriving at the information - this again can happen for example in communication between agents.}
\item {Case-5: All objects are labeled and the parthood relation is known in full.}
\end{itemize}
Under minimal assumptions, it is shown that lot more information about possible models for the situation can be
deduced. In this order structures relatable to chains and antichains are chosen because of their value in representation of finite posets and certain other possible posets have been excluded.
The general problem that is investigated has the following form: \emph{given some information about number theoretic properties of crisp and non-crisp objects, then what can be said about the existence of models of possible rough ontology in the context?}
So the problem falls under the general class of inverse problems (\cite{AM240}), but with additional assumptions of a number theoretic nature. New application contexts can be found in studies in social sciences, psychology and human reasoning - in these contexts finite bounds on possible number of non-crisp objects and their distribution has significant effect (especially the latter). This approach also has the potential to bridge the gap between theoretical rough sets and algorithms used in practice. Two papers that approach the bridging aspect without any number-theoretic assumptions are \cite{NOV,YY15}.
In algorithms used in practice, the goals can be very different - but a dominant method is to \emph{coerce data to follow preconceived models}. Here the idea of \emph{preconceived} is dictated by factors like \emph{ease of computation} at the cost of representation. This is reflected for example in the book \cite{MoPiZi}. Similar methods of intrusion can be found in other branches of machine learning - decision trees can be used for prediction and knowledge representation and in general the methods adopted are to intrusively optimize over the following principles:
\begin{itemize}
\item {Decision trees are comprehensible when they have lesser number of nodes in the tree.}
\item {Decision tree algorithms work faster when the depth of the tree is smaller.}
\item {Accuracy of decisions increase with decrease in number of misclassifications.}
\end{itemize}
In this paper, the motivations are towards knowledge representation.
\subsection*{Background}
Let $S$ be any set and $l, u$ be lower and upper approximation operators on $\mathcal{S}\subseteq \wp (S)$ that satisfy monotonicity and $(\forall A\subseteq S)\, A\subseteq A^u$. An element $A\in\mathcal{S}$ will be said to be \emph{lower definite} (resp. \emph{upper definite}) if and only if $A^l = A$ (resp. $A^u = A$) and \emph{definite}, when it is both lower and upper definite.
In general rough sets, these ideas of definiteness are insufficient as it can happen that upper approximations of upper definites are not upper definite. The following variants are of natural interest (though surprisingly their characterizations in the rough context are not fully understood):
\begin{itemize}
\item {An upper definite object $A$ will be said to be \emph{strongly upper definite} if and only if \[A = A^u = A^{uu}. \] }
\item {An object $A$ will be said to be \emph{pre-strongly upper definite} if and only if \[A^u = A \,\& \, (\exists n) \, A^{u^{n}} = A^{u^{n+1}} .\] }
\item {An object $A$ will be said to be \emph{upper pre-definite} if and only if \[ (\exists n) \, A^{u^{n}} = A^{u^{n+1}} .\] }
\end{itemize}
Analogous concepts of lower definiteness and definiteness can be directly defined .
Possible concepts of rough objects considered in the literature include the following:
\begin{itemize}
\item {A \textsf{non definite subset} of $S$, that is $A$ is a rough object if and only if $A^l \neq A^u$. }
\item {\textsf{Any pair of definite subsets} of the form $(A , B)$ satisfying $A\subseteq B.$}
\item {\textsf{Any pair of subsets} of the form $(A^l ,A^u)$.}
\item {Sets in an \textsf{interval of the form} $(A^l, A^u)$.}
\item {Sets in an \textsf{interval of the form} $(A, B)$ satisfying $A\subseteq B$ and $A, B$ being definite subsets.}
\item {A \textsf{non-definite element in a RYS}, that is an $x$ satisfying $\neg \mathbf{P} x^u x^l $}
\item {An \textsf{interval of the form}, $(A, B)$ satisfying $A\subseteq B$ and $A, B$ being definite subsets.}
\end{itemize}
The idea of definite and rough objects can be varied substantially even when the approximations have been fixed and the above concepts are based on representation.
\textsf{Concepts of representation of objects necessarily relate to choice of semantic frameworks. In general, in most contexts, the order theoretic representations are of interest. In operator centric approaches, the problem is also about finding ideal representations. The central problem that is pursued in the present paper relates to combinatorial characterization and number theoretic properties of existence of models.}
Granular operator spaces, a set framework with operators introduced by the present author in \cite{AM6999}, will be used as all considerations will require quasi/partial orders in an essential way. The evolution of the operators need not be induced by a cover or a relation (corresponding to cover or relation based systems respectively), but these would be special cases. The generalization to some rough Y-systems \textsf{RYS} (see \cite{AM240} for definitions), will of course be possible as a result.
\begin{definition}
A \emph{Granular Operator Space} $S$ will be a structure of the form $S\,=\, \left\langle \underline{S}, \mathcal{G}, l , u\right\rangle$ with $\underline{S}$ being a set, $\mathcal{G}$ an \emph{admissible granulation}(defined below) over $S$ and $l, u$ being operators $:\wp(\underline{S})\longmapsto \wp(\underline{S})$ satisfying the following:
\begin{align*}
A^l \subseteq A\,\&\,A^{ll} = A^l \,\&\, A^{u} \subset A^{uu} \\
(A\subseteq B \longrightarrow A^l \subseteq B^l \,\&\,A^u \subseteq B^u)\\
\emptyset^l = \emptyset \,\&\,\emptyset^u = \emptyset \,\&\,S^{l}\subseteq S \,\&\, S^{u}\subseteq S.
\end{align*}
Here, \textsf{Admissible granulations} are granulations $\mathcal{G}$ that satisfy the following three conditions (Relative \textsf{RYS} \cite{AM240}, $\mathbf{P} = \subseteq$, $\mathbb{P} = \subset$) and $t$ is a term operation formed from set operations):
\begin{align*}
(\forall x \exists
y_{1},\ldots y_{r}\in \mathcal{G})\, t(y_{1},\,y_{2}, \ldots \,y_{r})=x^{l} \\
\tag{Weak RA, WRA} \mathrm{and}\: (\forall x)\,(\exists
y_{1},\,\ldots\,y_{r}\in \mathcal{G})\,t(y_{1},\,y_{2}, \ldots \,y_{r}) =
x^{u},\\
\tag{Lower Stability, LS}{(\forall y \in
\mathcal{G})(\forall {x\in \underline{S} })\, ( y\subseteq x\,\longrightarrow\, y \subseteq (x^{l})),}\\
\tag{Full Underlap, FU}{(\forall
x,\,y\in\mathcal{G})(\exists
z\in \underline{S} )\, x\subset z,\,y \subset z\,\&\,z^{l} = z^{u} = z,}
\end{align*}
\end{definition}
On $\wp(\underline{S})$, the relation $\sqsubset$ is defined by \[A \sqsubset B \text{ if and only if } A^l \subseteq B^l \,\&\, A^u \subseteq B^u.\] The rough equality relation on $\wp(\underline{S})$ is defined via $A\approx B \text{ if and only if } A\sqsubset B \, \&\,B \sqsubset A$.
Regarding the quotient $\underline{S}|\approx$ as a subset of $\wp(\underline{S})$, the order $\Subset$ will be defined as per \[\alpha \Subset \beta \text{ if and only if } \alpha^l \subseteq \beta^l \,\&\, \alpha^u \subseteq \beta^u.\] Here $\alpha^l$ is being interpreted as the lower approximation of any of the elements of $\alpha$ and so on. $\Subset$ will be referred to as the \emph{basic rough order}.
\begin{definition}
By a \emph{roughly consistent object} will be meant a set of subsets of $\underline{S}$ of the form $H = \{A ; (\forall B\in H)\,A^l =B^l, A^u = B^u \}$. The set of all roughly consistent objects is partially ordered by the inclusion relation. Relative this maximal roughly consistent objects will be referred to as \emph{rough objects}. By \emph{definite rough objects}, will be meant rough objects of the form $H$ that satisfy
\[(\forall A \in H) \, A^{ll} = A^l \,\&\, A^{uu} = A^{u}. \]
\end{definition}
However, this definition of rough objects will not necessarily be followed in this paper.
\begin{proposition}
$\Subset$ is a bounded partial order on $\underline{S}|\approx$.
\end{proposition}
\begin{proof}
Reflexivity is obvious. If $\alpha \Subset \beta$ and $\beta \Subset \alpha$, then it follows that $\alpha^l = \beta^l$ and $\alpha^u = \beta^u$ and so antisymmetry holds.
If $\alpha \Subset \beta$, $\beta \Subset \gamma$, then the transitivity of set inclusion induces transitivity of $\Subset$.
The poset is bounded by $0 = (\emptyset , \emptyset)$ and $1 = (S^l , S^u)$. Note that $1$ need not coincide with $(S, S)$.
\qed
\end{proof}
In quasi or partially ordered sets, sets of mutually incomparable elements are called \emph{antichains}. Some of the basic properties may be found in \cite{GG1998,koh}. The possibility of using antichains of rough objects for a possible semantics was mentioned in \cite{AM3690,AM9501} and has been developed subsequently in \cite{AM6999}. In the paper the developed semantics is applicable for a large class of operator based rough sets including specific cases of \textsf{RYS} \cite{AM240} and other less general approaches like \cite{CD3,CC5,YY9,IT2}. In \cite{CD3,CC5}, negation like operators are assumed in general and these are not definable operations relative the order related operations/relation.
\subsection{Concepts of Finite Posets}\label{wth}
Let $S$ be a finite poset with $\#(S) = n < \infty$. The following concepts and notations will be used in this paper:
\begin{itemize}
\item {If $\mathfrak{F}$ is a collection of subsets $\{X_i\}_{i\in J}$ of a set $X$, then a \emph{system of distinct representatives} \textsf{SDR} for $\mathfrak{F}$ is a set $\{x_i ; i\in J\}$ of distinct elements satisfying $(\forall i \in J) x_i \in X_i$.}
\item {For $a, b\in S$, $b$ covers $b$ will be denoted by $a \prec b$. $c(S)$ shall be the number of covering pairs in $S$.}
\item {A chain $C$ will be said to be \emph{saturated} if and only of if $a\prec_{|C} b$ (that is if $b$ covers a in the induced order $\leq_{|C}$ on $C$) implies $a\prec b$ }
\item {A \emph{chain cover} of a finite Poset $S$ is a collection $\mathcal{C}$ of chains in $S$ satisfying $\cup \mathcal{C} = S$. It is disjoint if the chains in the cover are pairwise disjoint.}
\item {$S$ has finite width $w$ if and only if it can be partitioned into $w$ number of chains, but not less.}
\item {The \emph{Hasse index} of $S$ is defined by $i(S) = \dfrac{c(S)}{\# (S)}$.}
\item {$S$ is \emph{graded} if there is a unique partition of $S$ into $\{A_i\}_{i=o}^{r}$ with $A_o$ being the set of minimal elements, and \[\forall x\in A_i (x\prec y \longrightarrow y\in A_{i+1}).\] $A_i$ are termed the \emph{levels} of $S$ and if $x\in A_i$, then \emph{rank} of $x$ is $rk(x) = i$.}
\item {A \emph{symmetric chain} $x_o \prec x_1 \prec \ldots x_r$ is a chain with $rk(x_i) = i$ for all $i$. A \emph{symmetric chain decomposition} of $S$ is a partition of $S$ into symmetric chains.}
\end{itemize}
The following theorems are well known:
\begin{theorem}
\begin{itemize}
\item {A collection of subsets $\mathfrak{F}$ of a finite set $S$ with $\#(\mathfrak{F}= r$ has an SDR if and only if for any $1 \leq k \leq r $,the union of any $k$ members of $\mathfrak{F}$ has size at least $k$, that is
\[(\forall{X_1,\ldots , X_k \in \mathfrak{F}}) \, k \leq \#(\cup X_i).\] }
\item {Every finite Poset $S$ has a disjoint chain cover of width $w = width(S)$.}
\end{itemize}
\end{theorem}
\section{Semantic Framework}\label{sf}
Since \textsf{objects} are assumed to be definable if at all by sets of \textsf{attributes}, all considerations can be in terms of attributes. For the considerations of the following sections on distribution of rough objects and on counting to be valid, a minimal set of assumptions are necessary. These are as follows:
\begin{align*}
\tag{FO1} S \text{ is a granular operator space}. \\
\tag{FO2} \mathbb{S} \subseteq \wp(\underline{S}) .\\
\tag{FO3} \# (\mathbb{S}) = n < \infty .\\
\tag{RO1} R \subset \mathbb{S} \text{- the set of rough objects in some sense}.\\
\tag{RO2} \# (R) = n-k < n .\\
\tag{CO1} C \subseteq \mathbb{S} \text{ is the set of crisp objects}.\\
\tag{CO2} \# (C) = k .\\
\tag{RC1} R\cap C = \emptyset .\\
\tag{RO3} \text{ there exists a map } \varphi : R \longmapsto C^{2} .\\
\tag{RC2} (\forall x\in R)(\exists a, b\in C) \varphi (x) = (a, b) \,\& \, a\subset b.
\end{align*}
Note that no further assumptions are made about the nature of $\varphi(x)$. It is not required that
\[\varphi (x) = (a, b) \, \&\, x^l = a \, \&\, x^u = b,\]
though this happens often.
The set of crisp objects is necessarily partially ordered. In specific cases, this order may be a lattice, distributive, relatively complemented or Boolean order. Naturally the combinatorial features associated with granular operator space depends on the nature of the partial order and results in situations that is way more involved than the situation encoded by the following simple proposition.
\begin{proposition}
Given a fixed value of $\# (\mathbb{S}) = n = \# ((\wp (S))$ and $\#(C)=k$, $R$ must be representable by a finite subset $K \subseteq C^2 \setminus C$.
\end{proposition}
The two most extreme cases of the ordering of the set $C$ of crisp objects correspond to $C$ forming a chain and $C\setminus \{0,1 \}$ forming an anti-chain.
\begin{definition}
For $a, b \in R$, let
\[\nu(a, b ) =\,\left\{
\begin{array}{ll}
0, & \text{ if } a^l = b^l \, \& \, a^u = b^u \\
1 & \text{ if } a^l \neq b^l \, \& \, a^u \neq b^u \\
\frac{1}{\pi} & \text{ if } a^l \neq b^l \, \& \, a^u = b^u \\
\frac{1}{e} & \text{ if } a^l = b^l \, \& \, a^u \neq b^u.
\end{array}
\right. \]
By the \emph{rough distribution index} of $R$ will be meant the sum
\[\iota(R,C) = \sum_{a, b\in R} \nu(a,b)\]
and the \emph{relative rough distribution index} of $R$ shall be
\[\iota^\ast (R, C) = \dfrac{\iota(R,C)}{(n-k)^2}\]
\end{definition}
\begin{theorem}
\[0 \leq \iota (R, C) \leq (n -k)^2\]
\end{theorem}
\begin{proof}
The lower bounds have been obtained on the assumption that the non crisp elements are mutually roughly equal.
\qed
\end{proof}
The measure gives an idea of the extent of distribution of non crisp objects over the distribution of the crisp objects and the relative measure is a bad approximation of the idea of seeking comparison across distributions of crisp objects.
\section{Example Contexts}
Often in the design, implementation and analysis of surveys (in the social sciences in particular), a number of intrusive assumptions on the sample are done and preconceived ideas about the population may influence survey design. Some assumptions that ensure that the sample is representative are obviously good, but as statistical methods are often abused \cite{MHRL} a minimal approach can help in preventing errors.
The idea of samples being representative translates into number of non crisp objects being at least above a certain number and below a certain number.
There are also situations (as when prior information is not available or ideas of representative samples are unclear) when such bounds may not be definable or of limited interest.
\section{Distribution of Objects in Chains: Case-0}
This a variant of the simplest case relative computations and requires three additional assumptions:
\begin{enumerate}
\item {$C$ forms a chain under inclusion order.}
\item {$\varphi$ is a bijection.}
\item {Pairs of the form $(x, x)$, with $x$ being a crisp object, also correspond to rough objects.}
\end{enumerate}
It should be noted that this interpretation is not compatible with the interval way of representing rough objects without additional tweaking. It is the pairs interpretation that is being targeted.
\begin{theorem}
Under the above two assumptions, the number of crisp objects is related to the total number of objects by the formula: \[k \stackrel{i}{=} \dfrac{(1+4n)^\frac{1}{2} -1 }{2}.\]
In the formula $\stackrel{i}{=}$ is to be read as \emph{if the right hand side (RHS) is an integer then the left hand side is the same as RHS.}
\end{theorem}
\begin{proof}
\begin{itemize}
\item {Clearly the number of rough objects is $n - k$ .}
\item {By the nature of the surjection $n - k$ maps to $k^2$ pairs of crisp objects.}
\item {So $n - k = k^2$.}
\item {So integral values of $\dfrac{(1+4n)^\frac{1}{2} -1 }{2}$ will work.}
\end{itemize}
\qed
\end{proof}
This result is associated with the distribution of odd square integers of the form $4n +1$ which in turn should necessarily be of the form $4(p^2 +p) +1$ (p being any integer). The requirement that these be perfect squares causes the distribution of crisp objects to be very sparse with increasing values of $n$. The number of rough objects between two successive crisp objects increases in a linear way, but this is a misleading aspect. These are illustrated in the graphs below (\textsf{Fig.1,Fig.2}).
\begin{figure*}[htb]
\centering
\includegraphics[width=11.7cm]{disrough+.png}
\caption{Rough Objects Between Crisp Objects: Special Chain Case}
\end{figure*}
\begin{figure*}[htb]
\centering
\includegraphics[width=11.7cm]{nem.png}
\caption{Values of n and k: Special Chain Case}
\end{figure*}
\section{Distribution of Objects Over Chains: Case-1}
The simplest case relative computations requires two additional assumptions:
\begin{enumerate}
\item {$C$ forms a chain under inclusion order.}
\item {$\varphi$ is an bijection onto $C^2 \setminus \Delta_C$ ($\Delta_C$ being the diagonal of $C$).}
\end{enumerate}
\begin{theorem}
Under the above two assumptions, the number of crisp objects is related to the total number of objects by the formula: \[n - k = k^2 - k\]
So, it is necessary that $n$ be a perfect square
\end{theorem}
\begin{proof}
\begin{itemize}
\item {Clearly the number of rough objects is $n - k$ .}
\item {By the nature of the surjection $n - k$ maps to $k^2 -k$ pairs of crisp objects (as the diagonal cannot represent rough objects).}
\item {So $n - k = k^2 - k$.}
\item {So $n= k^2$ is necessary.}
\end{itemize}
\qed
\end{proof}
\begin{theorem}
In the above context, the cardinality of Boolean algebras that are power sets and in which the rough objects form chains in the induced order correspond to integral solutions for $x$ in \[ 2^x = k^2 .\]
\end{theorem}
\begin{proof}
As the number of elements in a finite power set must be of the form $2^x$ for some positive integer $x$, the correspondence follows. If $2^x = k^2$, then $x = {2 \log_2 k}$
This translates to a very sparse distribution of such models. In fact for $n\leq 10^8$, the total number of models is $27$.
\qed
\end{proof}
\textsf{Fig. 3} gives an idea of the numbers that work:
\begin{figure*}[hbt]
\centering
\includegraphics[width=11.7cm]{existenceg+.jpg}
\caption{Existence of Power Rough Sets on Chain}
\end{figure*}
\section{Distribution of Objects Over Chains: Case-2}
The formalism of this case is motivated by the hope that existence of rough objects may be strongly regulated by essentially number-theoretic properties.
In the context of case-1, if the condition of $\varphi$ being a bijection into $C^2 \setminus \Delta_C$ is relaxed to $\varphi $ being an injection and it is assumed that \[\# (\varphi(R)) \leq \alpha (k^2 - k),\] for some rational $\alpha\in (0, 1]$ (the interpretation of $\alpha$ being that of a loose upper bound rather than an exact one), then the following theorems holds:
\begin{theorem}\label{fract}
Given fixed $n$, the possible values of $k$ correspond to integral solutions of the formula:
\[k = \dfrac{(\pi - 1) + \sqrt{(1-\pi)^2 +4n\pi}}{2\pi},\] subject to $k\leq \lfloor\sqrt{n}\rfloor$, $\# (\varphi(R)) = \pi (k^2 - k)$ and $0 < \pi \leq \alpha$.
\end{theorem}
\begin{proof}
\begin{itemize}
\item {When $n-k = \pi (k^2- k )$ then $\pi = \dfrac{(n - k)}{(k^2 - k)}$}
\item {So positive integral solutions of $k = \dfrac{(\pi - 1) + \sqrt{(1-\pi)^2 +4n\pi}}{2\pi}$ may be admissible.}
\item {The expression for $\alpha$ means that it can only take a finite set of values given $n$ as possible values of $k$ must be in the set $\{2, 3, \ldots , \lfloor\sqrt {\frac{n}{\alpha}}\rfloor\}$. The bounds for $k$ is not the best possible.}
\end{itemize}
\qed
\end{proof}
\begin{theorem}
In the proof of the above theorem (Thm. \ref{fract}), fixed values of $n$ and $\pi$ do not in general correspond to unique values of $k$ and unique models.
\end{theorem}
For fixed $n$ and possible values of $\pi$, the number of values of $k$ for which rough objects exist follows the
pattern described in \textsf{Fig. 4} below. The intended reading is \emph{For $\pi=0.5$ and $n=1000000$, the number of values of $k$ that work seems to be $1413$}.
\begin{figure*}[hbt]
\centering
\includegraphics[width=11.7cm]{alphrou9+.jpg}
\caption{Number of Possible Values of $k$}
\end{figure*}
If the bounds on $k$ are imposed on the graph in \textsf{Fig. 4} then \textsf{Fig. 5} is the result:
\begin{figure*}[hbt]
\centering
\includegraphics[width=11.7cm]{alphrouff+.jpg}
\caption{Trimmed Number of Possible Values of $k$}
\end{figure*}
\subsection*{Algorithms: Case-2}
An algorithm for computing admissible values of $alpha$ can be
\begin{enumerate}
\item {Fix the value of $n$.}
\item {Start from possible values of $k$ less than $\sqrt{n - 1}$.}
\item {Compute $\alpha$ for all of these values.}
\item {Suppose the computed values are $\alpha_1, \ldots \alpha_r$}
\item {Check the admissibility of solutions.}
\end{enumerate}
Another algorithm for converging to solutions is the following:
\begin{enumerate}
\item {Start from a sequence $\{\alpha_i\}$ of possible values in the interval $(0,1)$.}
\item {Check the admissibility and closeness to solutions}
\item {If a solution appears to be between $\alpha_i$ and $\alpha_{i+1}$, add an equally spaced subsequence between the two.}
\item {Check the admissibility and closeness to solutions.}
\item {Continue}
\item {Stop when solution is found}
\end{enumerate}
\begin{theorem}
Both of the above algorithms converge in a finite number of steps.
\end{theorem}
\begin{proof}
Convergence of the first algorithm is obvious.
Convergence of the second follows from the following construction:
\begin{itemize}
\item {Suppose the goal is to converge to an $\alpha \in (0,1) $.}
\item {Let $\alpha_o =0, \, \alpha_1 = 1 $ and for a fixed positive integer $n$ and $i= 1, \ldots, n$, let $\alpha_{1i} = \frac{i}{n}$ and $\alpha \in (\alpha_{1j}, \alpha_{1 j+1})$. }
\item {Form $n$ number of equally spaced partitions $\{\alpha_{2i}\}$ of $(\alpha_{1j}, \alpha_{1 j+1})$ and let $\alpha \in (\alpha_{2j}, \alpha_{2 j+1})$.}
\item {Clearly $(\forall \epsilon >0 \, \exists N\, \forall r>N)\, |\alpha - \alpha_{rj}| < \epsilon $}
\item {So the algorithm will succeed in finding the required $\alpha$. }
\end{itemize}
\qed
\end{proof}
\section{Bounded Distribution on Chains}
The idea of bounded distribution corresponds to the set $R$ being partitioned into disjoint subsets of size $\{r_i \}_{i=1}^{g}$ with $g= k^2 -k$ subject to the condition $\beta$\[a \leq r_i \leq b \leq n -k, \text{ with }a,\, b \text{ being constants.} \]
\begin{theorem}
If the crisp objects form a chain, then the total number of possible models $B$ is \[B \,= \sum_{\alpha\in \pi(r)|\beta}\prod_{i=1}^{k^2 -k} \alpha_i \text{ and }n_o a^{k^2 - k} \leq B \leq n_o b^{k^2 - k} ,\]
with the summation being over partitions $\alpha = \{\alpha_i \}$ of $r$ subject to the condition $\beta$ and $n_o$ being the number of admissible partitions under the conditions.
\end{theorem}
\begin{proof}
On a chain of length $k$, $k^2 -k$ spaces can be filled.
The next step is to determine the partitions $\pi(r)$ of $r$ into $k^2 -k$ distinct parts.
The condition $\beta$ eliminates many of these partitions resulting in the admissible set of partitions $\pi(r)|\beta$.
Each of the partitions $\alpha \in \pi(r)|\beta$ corresponds to $\prod_i \alpha_i $ number of possibilities.
So the result follows.
\qed
\end{proof}
\section{Distribution of Objects: Most General Set-Theoretic Context}
In general, in the context of the framework specified in the Sec.\ref{sf} it can be assumed that
\begin{itemize}
\item {$\# (\varphi(R)) = t \leq n-k$,}
\item {$t = \beta (k^2 -k)$ and,}
\item {$n - k = \alpha (k^2 - k)$. for some constants $t,\, \beta,\, \alpha$}
\end{itemize}
This can also be used when objects are neither crisp or non crisp. In practice, objects may be neither crisp nor clearly non crisp possibly when:
\begin{itemize}
\item {a consistent method of identifying crisp objects is not used or}
\item {some objects are merely labeled on the basis of poorly defined partials of features or}
\item {a sufficiently rich set of features that can provide for consistent identification.}
\end{itemize}
The following construction provides a way of integrating the order structure on the set of all crisp and non crisp objects into basic computational considerations:
\begin{definition}
The \emph{lower definable scope} $\mathbf{SL}(x)$ of an element $x\in R$ will be the set of maximal elements in $\downarrow (x) \cap C$, that is \[\mathbf{SL}(x) = \max(\downarrow(x) \cap C). \]
The \emph{upper definable scope} $\mathbf{SU}(x)$ of an element $x\in R$ will be the set of minimal elements in $\uparrow (x) \cap C$, that is \[\mathbf{SU}(x) = \min(\uparrow(x) \cap C). \]
\end{definition}
All representations of rough objects can be seen as the result of choice operations \[\psi_x : \mathbf{SL}(x) \times \mathbf{SU}(x) \longmapsto C^2 \setminus \Delta_C . \] Letting $\# (\mathbf{SL} (x)) = c(x)$ and $\# (\mathbf{SU} (x)) = v(x)$ formulas for possible values are obtainable. Finding a simplification without additional assumptions remains an open problem though.
\subsection*{Chain Covers}
Let $C^*$ be the set of crisp objects $C$ with the induced partial order, then by the theorem in Sec \ref{wth},
The order structure of the poset of crisp objects $C^*$ permits a disjoint chain cover. This permits a strategy for estimating the structure of possible models and counting the number of models.
\begin{itemize}
\item {Let $\{C_i \,:\, i=1, \ldots h \}$ be a disjoint chain cover of $C^*$. Chains starting from $a$ and ending at $b$ will be denoted by $[[a, b]]$.}
\item {Let $C_1$ be the chain $[[0, 1]]$ from the the smallest(empty) to the largest object.}
\item {If $C_1$ has no branching points, then without loss of generality, it can be assumed that $C_2 = [[c_{2l}, c_{2g}]]$ is another chain with least element $c_{2l}$ and greatest element $c_{2g}$ such that $0 \prec c_{21}$, possibly $c_{2g} \prec 1$ and certainly $c_{2g} < 1$. }
\item {If $c_{2g} < 1$, then the least element of at least two other chains ($[[c_{3l}, \, c_{3g}]]$ and $[[c_{4l}, \, c_{4g}]]$) must cover $c_{2g}$, that is $c_{2g}\prec c_{3l}$ and $c_{2g}\prec c_{4l}$.}
\item {This process can be extended till the whole poset is covered. }
\item {The first step for distributing the rough objects amongst these crisp objects consists in identifying the spaces distributed over maximal chains on the disjoint cover subject to avoiding over counting of parts of chains below branching points.}
\end{itemize}
The above motivates the following combinatorial problem for solving the general problem:
Let $H = [[c_l,c_g]] $ be a chain of crisp objects with $\#(H) = \alpha$ and let $c_o$ be a branching point on the chain with $\# ([[c_l , c_o]]) = \alpha_o$. Let \[S_C = \{(a, b) \,; \,a, b\in [[c_o, c_g]] \text{ or } c_l < a, b < c_o \}.\] In how many ways can a subset $R_f \subseteq R$ of rough objects be distributed over $S_C$ under $\# (R_f) = \pi$?
\begin{theorem}
If the number of possible ways of distributing $\pi$ rough objects over a chain of $\alpha$ crisp elements is $n(\pi,\alpha )$, then the number of models in the above problem is \[n (\pi,\alpha ) - n (\pi , \alpha_o).\]
\end{theorem}
\begin{proof}
This is because the places between crisp objects in $[[c_l,c_o ]] $ must be omitted. The exact expression of $n(\pi, \alpha)$ has already been described earlier.
\qed
\end{proof}
Using the above theorem it is possible to evaluate the models starting with splitting of $r$ into atmost $w$ partitions. \emph{Because of this it is not necessary to use principal order filters generated by crisp objects to arrive at direct counts of the number of possible cases and a representation schematics}.
\section{Interpretation and Directions}
The results proved in this research are relevant from multiple perspectives. In the perspective that does not bother with issues of contamination, the results mean that the number of rough models relative the number of other possible models of computational intelligence is low. This can be disputed as the signature of the model is restricted and categoricity does not hold.
In the perspective of the contamination problem, the axiomatic approach to granules , the results help in handling inverse problems in particular. From a minimum of information, it can also be deduced
\begin{itemize}
\item {whether a rough model is possible or}
\item {whether a rough model is not possible or}
\item {whether the given data is part of some minimal rough extensions}
\end{itemize}
The last possibility can be solved by keeping fixed the number of rough objects or otherwise. These problems apply for the contaminated approach too. It should be noted that extensions need to make sense in the first place.
The results are also expected to have many applications in probabilist approaches and variants.
All this is despite the paper being among the simplest in the literature on rough sets.
\bibliographystyle{splncs.bst}
|
\section*{Introduction}
Soon after the discovery of the Fe-based superconductors (Fe-SC)
great effort has been devoted to unveil their electron paring mechanism. Even after nearly a decade of intensive research,
the symmetry of the superconducting order parameter is still under dispute \cite{Hir2011,Hir2016}, nonetheless most
theories favor an unconventional $s^{\pm}$ symmetry with a sign change of the order parameter
between the hole and the electron Fermi sheets \cite{Maz2008,Kur2008,Wan2009,Kon2010}.
Among the members of the family of Fe-SC,
the binary compound FeSe has attracted considerable attention recently. This is mostly because the
crystal structure of FeSe is regarded as representative
of the entire family of Fe-SC. Further,
the superconducting transition temperature $T_{c}$ $\approx 8 $~K \cite{Hsu2008}
in bulk FeSe can be enhanced up to 37~K by application of
pressure \cite{Miz2008,Med2009,Mar2009,Ima2009} and even to 50 - 100~K
by growing it as a monolayer on a SrTiO$_{3}$ substrate \cite{Yan2012,Ge2015,Fan2015,Zha2016}.
Another unique feature of FeSe is that it undergoes a structural
phase transition from a tetragonal to an orthorhombic phase
at $T_{s} \approx$ 87~K \cite{Mc2009a}, which is not accompanied
or followed by a long-range magnetic order. At $T_{s}$, the $C_{4}$-rotational
symmetry of the underlying electronic system is also spontaneously broken.
The resulting electronic state with a $C_{2}$ symmetry is referred to as a
nematic phase. It is argued that the symmetry of the superconducting
order-parameter should give insight into the collective degree of freedom
that governs both superconductivity as well as nematicity in the Fe-SC \cite{Fer2014}.
However, the situation on the experimental front is far from being resolved.
Even in the case of single crystalline FeSe with relatively simple crystal structure, different experiments indicated different superconducting gap structures. While most experiments detected two superconducting gaps \cite{Kas2008,Dong2009,Kha2010,Pon2011,Lin2011,Abd2013,Kas2014,Hop2016}, no consensus has been reached concerning the magnitude of the superconducting gap as well as on the presence or absence of nodes within the structure. The residual linear component of the thermal conductivity $\kappa_{0}$/T in the $T\rightarrow$ 0 limit, which is particularly sensitive to nodal quasiparticles, revealed contradicting results \cite{Kas2014,Hop2016}. Further, surface sensitive scanning tunnelling spectroscopic (STS) measurements, performed on single crystalline \cite{Kas2014} and thin film samples \cite{Song2011}, detected ``V''-shaped spectra in the superconducting state indicating the presence of nodes. However, STS conducted on the twin boundaries displayed a full gap \cite{Wata2015}, suggesting nodeless superconductivity at the twin boundaries. In order to resolve this issue, it is necessary to perform both bulk and surface sensitive experiments on FeSe. Owing to the marked dependence of the superconducting properties even for FeSe samples grown by the same method \cite{Kas2014,Hop2016},
concerted investigations on identical single crystals are required to establish one of its most fundamental properties, $viz.$, the symmetry of the superconducting order parameter.
Here we report on specific heat $C(T)$ combined with low-temperature ($T \geq$ 0.35~K) scanning tunnelling microscopy (STM) measurements on a stoichiometric FeSe single crystal to establish its superconducting order parameter. As shown below, such a combination of techniques, bulk sensitive $C(T)$ and surface sensitive STM, allows us to unequivocally resolve the superconducting gap structure of FeSe to be nodeless.
\section*{Experimental Results}
For a general characterization of our single crystal, we measured the temperature dependence of resistivity $\rho(T)$ and magnetization $M(T)$, see Fig.~\ref{Fig1}. These measurements were carried out on the same single crystal which is shown in the inset of Fig.~\ref{Fig2}. In Fig.~\ref{Fig1}(a), $\rho(T)$ measurement along the $[100]_\mathrm{T}$ direction of the tetragonal structure is presented. The resistivities are $\rho_{300\mathrm{K}}$ = 0.51 m$\Omega$ cm at 300 K and $\rho_{15~\mathrm{K}}$ = 0.031 m$\Omega$ cm at 15 K. These values give a residual resistivity ratio (RRR) $\rho_{300\mathrm{K}}$/$\rho_{15\mathrm{K}}$=16.4. The RRR of our crystal is very similar to the FeSe crystal (sample B) investigated by Bourgeois-Hope $et~al.$ in Ref. \onlinecite{Hop2016}. A kink at $T_s$ = 87 K marks the structural transition temperature. In the inset, Fig.~\ref{Fig1}(b), the onset of superconducting transition can be seen at $T_{c}^{\mathrm{onset}}$=10.2 K. The sample achieves zero resistivity at $T_{c}$ = 8.5 K, which is the superconducting transition temperature of the bulk. In the inset Fig.~\ref{Fig1}(c), $M(T)$ measured in a field of 20 Oe, both in the zero-field-cooled (ZFC) and field-cooled protocol (FC) are shown. The sample displays a full diamagnetic shielding in the superconducting state with $4\pi\chi = -1$.\\
\begin{figure}
\centering
\includegraphics[clip,width=0.95\columnwidth]{Fig1.eps}
\caption{\textbf{Resistivity and magnetization.} (a) Resistivity $\rho (T)$ of FeSe single crystal presented in inset of Fig. 2. The temperature of the structural transition is marked by $T_s$. (b) The same $\rho (T)$ data in (a) zoomed in for $T < 40~K$ showing the superconducting transition. (c) Magnetization measured in a magnetic field of 20~Oe both in zero-field-cooled (ZFC) and field-cooled (FC) protocols.}
\label{Fig1}
\end{figure}
\begin{figure}
\centering
\includegraphics[clip,width=0.95\columnwidth]{Fig2.eps}
\caption{\textbf{Specific heat.} Specific heat divided by temperature, $C/T$ vs $T^{2}$, measured at magnetic fields of zero and 9 T. The solid line represents the normal-state specific heat $C_{n}$. Upper-left inset: a photograph of tetragonal FeSe single crystal used for specific-heat measurements. Lower-right inset: zero-field $C/T$ vs $T$ in an enlarged scale around $T_{c}$. Lines show how $T_{c}$ and $\Delta C/T_{c}$ were determined.}
\label{Fig2}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[clip,width=0.95\columnwidth]{Fig3.eps}
\caption{\textbf{Electronic part of the specific heat.} Zero-field electronic specific heat (with the normal-state specific heat being subtracted) divided by temperature. The solid line represents a fit by a smaller $s$-wave plus a larger extended $s$-wave models of the form $\delta C = x~\delta C_s + (1-x)~\delta C_{es}$ with $x=0.32$. The inset shows the entropy conservation required for a second-order phase transition.}
\label{Fig3}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[clip,width=1.6\columnwidth]{Fig4.eps}
\caption{\textbf{STM topography and spectroscopy.} (a) A topography of FeSe on an area of 40 $\times$ 40 nm$^2$ obtained at 0.35 K.
The white line mark one of the unidirectional electronic dimer of length $\sim$ 16~$a_{\textrm{Fe}}$, where $a_{\textrm{Fe}}$ is the distance of the Fe-Fe
atoms in the crystal structure.
The bias voltage and the tunnelling
currents were set at $V_b$ = 10 mV and $I_{sp}$ = 100 pA, respectively. The tunnelling conductance is acquired by the
standard lock-in technique with a small modulation voltage of 0.05 mV$_{\textrm{rms}}$. (b) Line scans along the blue and green lines marked in (a) displaying the heights of the impurities. (c) An average tunnelling spectrum measured within the area of 1 $\times$ 1 nm$^2$ [red square in (a)] at 0.35~K. The arrows indicate ``wing''-like features mentioned in the text. (d) A fit (solid line) of a Dynes gap function to the symmetrized data (open circles) in the $V_b$ range $\pm$3 mV for an $(s+es)$-wave model. For the fit, the thermal broadening as well as the broadening caused by a finite energy resolution was taken into account. Inset :The tunneling conductance at $|V_b|\rightarrow 0$. The arrows mark the voltage range at which d$I$/d$V \approx$ 0 (cf. Supplementary information, Fig.~S11).}
\label{Fig4}
\end{figure*}
The temperature ($T$) and magnetic field ($B$) dependence of the specific heat $C(T,B)$ was measured on the single crystal imaged in the inset of Fig.~\ref{Fig2}. The zero-field $C/T$ vs $T^{2}$ plot between 0.35 and 10~K presented in Fig.~\ref{Fig2} displays two anomalies, a $\lambda$-like transition at $T_{c} = 8.4(1)$~K and a broad shoulder below 2~K, better seen in Fig.~\ref{Fig3}. This is a typical behaviour of a two-gap superconductor such as MgB$_{2}$ \cite{Bouquet2001}, suggesting the presence of at least two superconducting gaps in FeSe. The $T_{c}$ was determined via local entropy conservation, i.e., the vertical line in the inset of Fig.~\ref{Fig2} segments equal areas in a $C/T$ vs $T$ plot. We describe the normal-state specific heat $C_{n}$ below 10~K by $C_{n}(T) = \gamma_{n}T + C_{lat}(T)$, where $\gamma_{n}T$ is the normal electronic contribution and $C_{lat}(T) = \beta_{3} T^{3} + \beta_{5} T^{5}$ represents the phonon contribution. The fit to $C/T$ is shown as a solid line in Fig.~\ref{Fig2}, which yields $\gamma_{n} = 6.5$~mJ/mol K$^2$, $\beta_{3} = 0.365$~mJ/mol~K$^4$, and $\beta_{5} =$ 1.94$\times 10^{-4}$~mJ/mol~K$^6$. The Debye temperature $\theta_{\texttt{D}}$ calculated from $\beta_{3}$ is 242~K. These parameters are comparable to those reported earlier \cite{Lin2011,Koz2014,Mc2009,Wang2016}. The normalized specific-heat jump at $T_{c}$, $\Delta C/\gamma_{n}T_{c}$, is estimated to be 1.55, which is slightly larger than the weak-coupling value 1.43 of Bardeen-Cooper-Schrieffer (BCS) theory \cite{Bar1957}.
The excess electronic specific heat contribution in the superconducting state given by $\delta C(T) = C(T, B = 0) - C_{n}(T)$ is plotted in Fig.~\ref{Fig3}. The inset illustrates the satisfaction of entropy conservation $\Delta S = \int_{0}^{T_{c}}(\delta C/T)dT$ justifying the validity of the parameters used to fit $C_{n}(T)$. In the $\delta C/T$ plot, the shoulder below 2~K arising due to the second superconducting gap is clearly visible. To our knowledge, such a shoulder feature has been only reported for pure, polycrystalline samples \cite{Mc2009} with compositions Fe$_{1.01}$Se and Fe$_{1.02}$Se. In order to further examine the superconducting order parameter, the data in Fig.~\ref{Fig3} were fitted to the one-band BCS equation \cite{Bar1957} given by
\begin{eqnarray}
\delta C =
2N(0)\beta k_{\textrm{B}}\frac{1}{4\pi}\int_{0}^{2\pi}d\theta \nonumber \\
\int_{0}^{\hbar\omega _{\textrm{D}}} [ -\frac{\partial f}{\partial E}\left ( E^{2}+\frac{1}{2}\beta \frac{d\Delta^{2}(T,\theta)}{d\beta} \right ) ]d\varepsilon-\gamma_{n}T,
\end{eqnarray}
where $N(0)$ is the density of states at the Fermi surface, $\beta =$ 1/$k_{\textrm{B}}T$, $E = [\varepsilon^{2}+\Delta^{2}(T, \theta)]^{1/2}$, $f = (1 + e^{\beta E})^{-1}$, and $\Delta(T, \theta) = \Delta_{es}^0(T)(1 + \alpha~\mathrm{cos} 4\theta)$ an extended $s$-wave where $\alpha$ and $\theta$ represent the gap anisotropy and polar angle, respectively \cite{Maier2009,Chubukov2009}. Note that in an angle-integrated measurement, the functional forms containing $\mathrm{cos} 4\theta$ and $\mathrm{cos} 2\theta$ result in the same parameters and hence, it is not possible to distinguish between an extended $s$-wave and a $d$-wave symmetry. We found that a single anisotropic $s$-wave model, either with or without accidental nodes, does not fit to the $\delta C/T$ data (see Supplementary information, Figs. S1-S3.).
Since $\delta C$ indicated a signature of two superconducting gaps, we also tried a phenomenological two-gap model \cite{Suh1959,Mos1959,Zeh2013}
by taking a sum of either two $s$-wave-gaps (see Supplementary information, Fig. S4) or an $s$-wave + an extended $s$-wave gaps ($s+es$), Fig.~\ref{Fig3} (see also Supplementary information, Fig. S5),
to describe the data.
In the fitting, more weight was given to the low-temperature data, i.e., the data below $T =$ 5 K.
The reason being, close to $T_{c}$ the thermal fluctuations become stronger, and may result in increasing deviation of the data from the applied models which are based on a mean-field framework.
We found that both models lead to satisfactory fits.
Hence, the exact superconducting gap structure of FeSe cannot be unambiguously determined from analyzing the specific-heat data alone. However, as discussed later, with the help of fitting several models also to the tunnelling spectra (see Supplementary information, Figs. S8-S11), we could select ($s+es$) model as a proper model to describe the $C(T)$-data, with the larger gap assigned to the extended $s$-wave. The goodness of fit for each model tried here is presented in the Supplementary information Table S1.
The gap values in the $T \rightarrow 0 $ limit estimated from the ($s+es$) model fitting are: a small $s$-wave gap of $\Delta_{s}(0) = 0.25(3)$ meV and a large anisotropic extended $s$-wave gap of $\Delta_{es}^0(0) = 1.67(3)$ meV with $\alpha = 0.34$. The value of $\alpha <$ 1 obtained here clearly rules out the presence of accidental nodes \cite{Hin2015}. \sr{Since the obtained isotropic gap value $\Delta_{s}(0)$ is very small, a possible anisotropy of this gap would be beyond the resolution of our experiments. Further, a recent heat capacity study of FeSe single crystals by Wang $et~al.$ \cite{Wang2016} reports a small residual value of the electronic specific heat originating from low-energy quasiparticle excitations indicating either line nodes or deep gap minima. However, their experiment did not show the shoulder in $C(T)$ which we observe below 2~K and interpret as the mark of a second, smaller superconducting gap. In our case, this shoulder limits the analysis of the functional form of $\delta C(T)$ as $T\rightarrow 0$. Therefore, if we consider the data only below 1~K, the presence of accidental nodes can not be ruled out based on the specific heat analysis.}
\begin{figure*}[t]
\centering
\includegraphics[clip,width=1.6\columnwidth]{Fig5.eps}
\caption{\textbf{Tunneling spectra at a twin boundary.} (a) A 25$\times$12 nm$^2$ topography of FeSe with a twin boundary (TB), which is zoomed from Fig. S6 (Supplementary information). (b) A line scan along the blue line depicted in (a). (c) Ten tunnelling spectra measured laterally at equidistant positions along the white dashed line in (a). The black arrow in (c) represents the direction of measurement shown in (a). Spectra $\#$2 and $\#$3 are measured on the TB. Curves are equally shifted vertically for clarity. The bias voltage and the tunnelling currents were set at $V_b$ = 10 mV and $I_{sp}$ = 100 pA, respectively.}
\label{Fig5}
\end{figure*}
In order to determine the superconducting gap
structure of FeSe microscopically, we performed STM measurements at 0.35~K.
The topographic images, $e. g.$ Fig.~\ref{Fig4}(a),
revealed atomically resolved clean Se-terminated surfaces indicating good sample quality, very similar to our previous report \cite{Ros2015}. The Se-Se
distance $a_{Se-Se} = 3.7(1)$ \AA~observed here is in line with the distance
of 3.7702(5)~\AA~found by x-ray diffraction on our crystals \cite{Koz2014}.
A few protrusions (see Fig.~\ref{Fig4}(b) for line scans across the impurities) on the top of the surface likely correspond to
Se-Se bound atoms left over from the top-most counter layer while cleaving, or to an impurity atom occupying the Fe-site underneath the topmost Se-layer.
Alternatively, recent density functional theory (DFT) based calculations suggested that Fe-site vacancies can perturb orbitals on neighbouring Se-sites, thereby producing atomic dumbbells \cite{Den2016}. All these defects act as impurities and induce additional dumbbell-like unidirectional depressions in the topography, known as ``electronic dimers'' \cite{Song}, marked by
the white line in Fig.~\ref{Fig4}(a). Interestingly, the unidirectional behaviour of these
electronic dimers can be unveiled from the fact that the orientation of the
dimers is independent of the orientation of the impurities but
rotates by 90$^\circ$ across the twin boundary (TB) obvious from the bright stripe in the topography, Supplementary information, Fig.~S6. This behaviour represents a broken $C_4$ symmetry in the impurity scattering. The length of the electronic dimers is $\approx$ 16~$a_{\textrm{Fe}}$, where $a_{\textrm{Fe}}$ is the distance of the Fe-Fe
atoms in the crystal structure, which is consistent with a previous report \cite{Song}. We note that all electronic dimers are oriented in the same direction in Fig.~\ref{Fig4}(a) indicating the entire image consists of a single crystallographic domain.\\
In Fig.~\ref{Fig4}(c), a tunnelling spectrum averaged over an area marked in Fig.~\ref{Fig4}(a) is presented. Since the spectra were measured approximately in the middle of Fig.~\ref{Fig4}(a), a possible TB is at least a minimum distance of 20 nm away.
The most prominent feature of the spectrum is that, as $V_b \rightarrow 0$, the spectrum attains a ``U''-shape. Here, ``U''-shape refers to the finite energy range within which the experimental tunnelling conductance is zero, as more clearly seen in Fig.~\ref{Fig4}(d). Due to the estimated small magnitude of the smaller gap (see above) a zero tunnelling conductance is only expected within a very narrow energy range. A tunnelling conductance of zero indicates the absence of quasiparticle excitations within the superconducting gap, thus providing strong evidence for a nodeless superconductivity in FeSe.
However, the spectrum contains additional hump-like features at energies $\approx -$10 mV and +5.4 mV. These represent either simply the bottom of an electron band and the top of a hole band, respectively, or more complex phenomena such as a density-wave type ordering \cite{Ros2015} or an electron-boson coupling \cite{Song2014}. In addition, there are ``wing''-like features contained in the coherence peaks at energies $V_b \approx \pm2.75$~mV, indicated by black arrows in Fig.~\ref{Fig4}(c). These features may be related to the fine details of the band structure, such as spin-orbit coupling \cite{Bor2015} induced band-hybridization \cite{Sch2013,Hin2015,Kom2015}, which is beyond the scope of this paper.
In an attempt to describe the tunnelling spectra, we used a Dynes gap function \cite{Dynes1978} to fit the data. Within the standard picture of the BCS model \cite{Bar1957}, the tunnelling conductance $\mathrm{d}I/\mathrm{d}V \propto f_D(E) $, where
\begin{equation}
\label{fDyne}
f_D=\int_{0}^{2\,\pi}d\,\theta\;\mathrm{Re}\,\frac{e\,V-\mathrm{i}\,\varGamma}{\sqrt{(e\,V-\mathrm{i}\,\varGamma)^2-[\Delta(\theta)]^2}}~.
\end{equation}
Here, the spectral broadening is given by the inverse quasiparticle lifetime $\varGamma$.
Since the experimental spectra did not indicate the presence of nodes, we first tried a single extended $s$-wave gap function $\Delta(\theta)=\Delta_{es}^0(1 + \alpha~\mathrm{cos} 4\theta)$ to fit the data (see Supplementary information, Fig. 11). Although the fitted curve followed the experimental data well in the applied bias voltage range $0.5~\mathrm{meV} < |V_b| < 0.75~$meV, below 0.5 meV, the fitted curve deviated from the experimental data as $V_b \rightarrow 0$. This suggested the presence of a small second gap as already inferred from the specific heat analysis. However, the coherence peaks corresponding to the smaller gap could not be resolved in our experiments due to its small magnitude, which is at the limits of our instrumental resolution. To take this into account, we
included an energy resolution of 0.16 meV in the fit procedure, which accounts for the spectral broadening caused by thermal effects (0.35 K) as well as a finite modulation voltage (0.05 mV$_\mathrm{rms}$).
In Fig.~\ref{Fig4}(d), the best fit to the experimental spectrum is presented. This fit corresponds to an ($s+es$) model with an $s$-wave gap of $\Delta_{s}=$0.6(1) meV and an extended $s$-wave gap $\Delta_{es}^0 = 1.35(2)$ meV and $\alpha = 0.30(1)$. These values are slightly different than those obtained from the specific heat analysis. Here we would like to emphasize that the models considered here should not be taken exhaustive, rather it should be understood as a minimum ansatz to describe the overall behaviour of the spectrum, which agrees semi-quantitatively with the specific-heat analysis. By considering the raw data alone and leaving the models aside, the multigap nature of the superconducting gap is derived from the specific heat measurements, whereas, the nodeless nature of the gap is concluded from the tunnelling spectroscopy measurements.\\
Following this indication towards nodeless superconductivity in FeSe, we now show --- using STM/STS --- that the gap structure appears to remain nodeless on different crystallographic twin domains as well as at the TB. In Fig.~\ref{Fig5}(a), an STM topographic image over an area of 25$\times$12 nm$^2$ containing a TB is presented. This image is a part of the topography of 40 $\times$40 nm$^2$ presented in Fig. S6 (Supplementary information). A height scan across the TB is shown in Fig. \ref{Fig5}(b). Several spectra were measured along the white line in Fig.~\ref{Fig5}(a) in such a way that the spectra were distributed on either side as well as on the TB. As can be seen in Fig.~\ref{Fig5}(c), the spectra retains an overall ``U''-shape across the TB, warranting the robustness of the nodeless gap structure in our single crystal. Alternatively, Watashige $et.al$ \cite{Wata2015} observed a lifting of nodes in the vicinity of a TB and interpreted this finding in terms of time reversal symmetry breaking caused by a $\pi/2$ rotation of the crystallographic domains. They also found that the influence of the TB on the superconducting gap structure extends up to a length scale of more than 50 nm. Since the spectra shown in Fig.~\ref{Fig5}(c) were measured only up to 9 nm away from the twin boundary, our results shown in Fig.~\ref{Fig5} do not directly contradict those of Ref. [\onlinecite{Wata2015}]. However, we did not find any signature of pair-breaking by observing a formation of bound states in the spectra taken on the TB, which was suggestive of a time reversal symmetry breaking. \sr{For the sake of confirming the U-shape of the tunneling spectra at small $V_b$ as a common feature of our sample, we performed STM/STS on a second crystal. In this case we conducted our measurements on an area of 100 $\times$100 nm$^2$ without any TB, see Supplementary information Fig. S7. As can be seen in Fig. S7(b), even the small gap could be resolved in some cases in the tunneling spectra. However, within a small range of $V_b $, the spectra retain a U-shape indicating the absence of low energy quasiparticle excitations owing to finite superconducting gap over the Fermi surface.}
\section*{Discussion}
In the framework of a single-band BCS theory, the zero-temperature upper critical field $H_{c2}(0)$ is proportional to $(\Delta/v_{\texttt{F}})^{2}$, $v_{\texttt{F}}$ being the Fermi velocity \cite{Shulga2002}. In a phenomenological two-gap model, $H_{c2}(0)$ is set by the larger gap, and the critical field for the smaller gap $H^{*}(0)$ can be determined by thermodynamic, e.g., specific-heat \cite{Bouquet2002} and thermal-conductivity \cite{Hop2016} measurements. Shubnikov-de Haas oscillation measurements have reported comparable values of $v_{\texttt{F}}$ for the different Fermi sheets in FeSe \cite{Tera2014}. By taking the maximum gap value $\Delta^\mathrm{max}_{es} = \Delta_{es}^0(1+\alpha)$ and $\Delta_{s}(0)$ obtained from the $C(T)$ analysis, we estimate, $H^{*}(0)/H_{c2}(0) = [\Delta_{s}(0)/\Delta^\mathrm{max}_{es}(0)]^{2} = (0.25/2.23)^{2} \approx 0.01$. This value is in good agreement with the data reported by Bourgeois-Hope $et$ $al.$~\cite{Hop2016}, thus further supporting the validity of the current analysis. As far as the absence of nodes is concerned, our results are also in good agreement with recent thermal conductivity \cite{Hop2016}, penetration depth \cite{Tek2016}, and microwave conductivity \cite{Li2016} measurements on single crystals of similar quality. \\
It is worthwhile to discuss the possible origin of the discrepancy between the STM results presented here in comparison to those in Refs. [\onlinecite{Kas2014,Wata2015}]. The nodes observed in FeSe are considered accidental, i.e., they are not imposed by symmetry \cite{Hir2011,Hir2016}. Theoretical investigations of multiorbital microscopic models have suggested that the nodes in the Fermi surface can be lifted by disorder \cite{Mis2009} or external strain \cite{Kan2014}. One of the ways to get a semi-quantitative estimation of the degree of disorder in a sample is to look at its RRR value.
However, we would like to point out that in the particular case of FeSe, the RRR calculated by taking the resistivity values from above and below $T_{s}$ contain additional contributions other than initial intrinsic disorder of the crystal which exists at room temperature. As observed by Kn\"oner $et~al.$ \cite{Kno2015}, cooling the samples through $T_{s}$ induces different twin states in the samples in question; which together with the finite in-plane anisotropy can produce different resistivity values below $T_{s}$. A similar observation was also made in Ref. \onlinecite{Hop2016}. Therefore, the crystals showing lower RRR likely contain more twins, and TB are considered accountable for lifting the nodes \cite{Wata2015}. Nonetheless, a very recent thermal conductivity \cite{Hop2016} measurement on samples grown by flux-vapour transport \cite{Boe2013} with the RRR values similar to those used in Refs. [\onlinecite{Kas2014,Wata2015}], exhibited two-gap nodeless superconductivity. It is rather intriguing that such negligibly small differences in the samples appear to be sufficient to influence the superconducting gap structure in FeSe. \\
Our observation of two superconducting gaps $\Delta_{s}$ and $\Delta_{es}$ with strongly different gap magnitudes, $i.e.$ $\Delta_{s} \ll \Delta_{es}$, indicates that superconductivity appears presumably in one band (producing a large gap $\Delta_{es}$) and may induce a second small gap $\Delta_{s}$ in another band due to a proximity effect \cite{Sch2013,Mc1968}. Nonetheless, both gaps open at the same temperature, but may have different temperature dependencies \cite{Hir2011}. So far in FeSe, only one hole Fermi sheet and one electron Fermi pocket are detected by angle-resolved photoemission spectroscopy (ARPES) \cite{Mal2014} and quantum oscillation experiments \cite{Tera2014,Wat2015a}. If this is correct, then an inter-band extended $s$-wave pairing with a sign reversal of the order parameter between different Fermi surface sheets might be the likely mechanism of superconductivity in FeSe \cite{Maz2008}. However, there are experimental indications for more than one electron pocket crossing the Fermi energy \cite{Wat2015, Huy2014}. In this case, a more exotic pairing mechanism such as band-hybridization induced odd frequency pairing can be expected \cite{Sch2013}.\\
In summary, we have investigated the superconducting gap structure of FeSe in a combined study of scanning tunnelling microscopy and specific heat measurements. \sr{The results indicate
multigap superconductivity in FeSe single crystals.} Our analysis suggests that the gap is of ($s+es$) type. The isotropic $s$-wave gap is much smaller than the anisotropic $s$-wave gap. Additionally, the tunnelling spectroscopy indicate at a superconducting gap which remains nodeless also on twin boundaries. These experimental results are expected to provide important ingredients for a unified theory of the superconducting paring mechanism for all FeSe-related superconductors.\\
\textit{Note~added}: In the revising stage of this manuscript we became aware of a new STM study on FeSe reported very recently \cite{Spr2016}. Our conclusions presented here are in excellent agreement with these complementary investigations in which the Bogoliubov quasipaticle scattering interference (BQPI) was used to determine the superconducting gap symmetry as extremely anisotropic, but nodeless with an OP changing sign between the hole and electron pockets. In addition, Ref. [\onlinecite{Spr2016}] also provides evidence for an orbital-selective Cooper pairing in FeSe.
\section*{Methods}
The single crystals were grown by chemical vapour transport\cite{Koz2014,Ros2016} of stoichiometric FeSe powder containing $\alpha$-Fe of less than 300 ppm. The ratio of FeSe to the transport additive AlCl$_3$ was taken as 50:1. Typically, a mixture of 1 g of FeSe powder and 20 mg of AlCl$_3$ was placed in a quartz ampoule of length 10 cm and diameter 2 cm prepared inside an argon-filled glove box. The ampoule containing the mixture was evacuated, sealed, and placed horizontally inside a two-zone furnace at temperatures from $T_{2}$=673~K and $T_{1}$=573~K. The crystal growth was carried out for 2 months. Finally, the ampoule was quenched in water. The product, which contained plate-shaped single crystals with edge lengths up to 400 $\mu$m perpendicular to the $c$ axis, was washed repeatedly in ethanol to remove remaining condensed gas phase, dried under vacuum and stored in the glove box. By extending the growth time to one year, larger single
crystals with dimensions up to $4 \times 2 \times 0.03$ mm$^3$ could be grown. The specific heat $C(T,B)$ was measured down to 0.5~K using a thermal-relaxation method in a physical property measurement system (Quantum Design) with the magnetic field $B$ applied parallel to the [001] direction of the single crystal. The scanning tunnelling microscopy/spectroscopy measurements were performed in an ultra-high vacuum ($p < 3 \times 10^{-9}$ Pa) cryogenic STM with a base temperature $T\approx 0.35$~K. The bias voltage and the tunnelling
currents were set at $V_b$ = 10 mV and $I_{sp}$ = 100 pA, respectively. The tunnelling conductance is acquired by the standard lock-in technique with a small modulation voltage of 0.05 mV$_{\textrm{rms}}$. Given the total energy resolution $\Delta E$ of the STM is limited by $\Delta E \approx \sqrt{(3.5k_{\mathrm{B}}T)^2+(2.5 eV_{\mathrm{mod}})^2}$ \cite{Lou2001}, the low temperature (0.35 K) and the small modulation voltage (0.05 mV$_{\textrm{rms}}$) used here allows us to resolve the fine structure of the superconducting gap. The FeSe single crystals were cleaved $in~situ$ at 20~K before being inserted into the STM-head.
|
\section{Introduction}
Understanding the spatial structure of ecological networks is vital in theoretical and experimental biology and
ecology \citep{May74,Smith74,Sole06,Smith82,Hof98,Now06,Weber14}. The stability of the dynamics of ecological networks
is influenced by a variety of factors such as the topology of the network as well as the weights of the links composing the network
(see \citep{Knebel13} for a recent example of Lotka-Volterra networks). A spatial setting, where individuals interact
locally and are mobile, strongly influences the
dynamics of the system and also has a large impact on the stability of an ecological network \citep{Rei08,Roman13}.
The introduction of space gives rise to rich, interaction network dependent, phenomena including pattern formation,
cluster coarsening, alliance formation and nested ecological niches.
These phenomena have yielded a huge theoretical interest recently, and
a variety of techniques \citep{Sza07,Fre09} have been employed to study these intriguing phenomena
for symmetric networks, ranging from simple cases with three \citep{Rei08,Rei07,Pel08,Rei08a,Ven10,Shi10,Wan10,
He10,Win10,He11,Rul11,Wan11,Nah11,Jia11,He12,Juu12,Jia12,Lam12,Ada12,Juu13,Rul13,Szs13,Sch13,Szs14,Gro15}
and four species \citep{Rom12,Sza04,Sza07b,Sza08,Int13,Gui13,Int15}
to complex situations with an
arbitrary number of species \citep{Roman13,Sza01,Sza01b,Sza05,Per07,Sza08a,Ave12a,Ave12b,Ave14,Mowlaei14,Sza07b,
Vuk13,Kan13,Ave14b,Che14}. Realistic ecologies, however, are endowed
with complex interaction networks that can not be captured fully by only considering symmetric networks.
As such it is important to develop theoretical approaches that allow to understand the dynamics of
general networks \citep{Sza07c,Lut12,Pro99,Van12,Knebel13,Dob14,Rul14,Var14,Sza15,Dal15,Szo15} and their effects on biodiversity,
correlations, and spatio-temporal patterns.
In the following we present an exact method that allows to reveal why species are partitioned in domains and what are the partition sets of the species labels in complex food networks.
This information is a necessary first step in order to develop a predictive theory for the fate of ecologies. Our approach
predicts the alliance formation between different species and fixes the emerging space-time patterns in many spatial
interaction networks. As a consequence, our approach also permits to shed light on the stability of ecological niches by breaking down this stability question into two distinctive parts: the stability of interactions among domains and within domains.
\section{Model and examples}
Fig.\ \ref{fig1} shows some typical two-dimensional space-time patterns that we aim at predicting with the approach discussed in the following.
Whereas snapshot (a) shows an example of a system with spirals where every wavefront contains only one species, panel (b) gives an
example of a coarsening process with two types of domains where inside every domain spirals are formed, thus yielding non-trivial
dynamics inside the coarsening domains. Example (c) reveals a coarsening process where every domain is formed by an alliance
of mutually neutral species. Finally, snapshot (d) shows a case of fuzzy spirals due to the mixing of different
species inside the waves.
\begin{figure} [!h]
\begin{center}
~~~~~~(a)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(b)~~~~~~~~~\\
\includegraphics[width=0.35\columnwidth]{figure1a}~~~
\includegraphics[width=0.35\columnwidth]{figure1b}~~~\\[0.2cm]
~~~~~~(c)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(d)~~~~~~~~~\\
\includegraphics[width=0.35\columnwidth]{figure1c}~~~
\includegraphics[width=0.35\columnwidth]{figure1d}~~~
\end{center}
\caption{(Color online) Four typical types of patterns observed in many species food networks (in all cases the rates where
chosen to be species independent): (a)
compact spirals in the $(5,3)$ model (with $\gamma = \delta =1$ and $\alpha = \beta =0$); (b) coarsening with spirals forming inside the domains in
the $(6,3)$ model (with $\gamma = \delta =1$ and $\alpha = \beta =0$); (c) coarsening of neutral species domains in the $(6,2)$ model
(with $\gamma = \delta =1$ and $\alpha = \beta =0$); (d) fuzzy spirals with mixing of species inside the spirals in the
$(9,3)$ model (with $\gamma = \delta = 0.5$ and $\alpha = \beta =0.25$).
Systems of $400 \times 400$ sites have been simulated.
\label{fig1}
}
\end{figure}
In order to focus our discussion we consider in the following systems of $N$ species in two space dimensions where
individuals of different species have mutual predator-prey
interactions, give birth to off-springs, and are mobile. For our method
the details of the implementation of these interactions do not matter, with the exception that the predation-prey interactions should
be binary, i.e. at any given event one predator consumes one prey. For the examples in Fig. \ref{fig1} as well as for the more complicated cases
discussed later we consider the following general interaction scheme ($X_i$ denotes a member of species $i$ and
${\emptyset}$ is an empty site):
\begin{eqnarray}
X_{i} + X_{j}& \xrightarrow{\delta_{ij}} & {\emptyset}+X_{i}\label{model1} \\
X_{i} + X_{j}& \xrightarrow{\alpha_{ij}} & X_{j}+X_{i} \label{model2} \\
X_{i} + {\emptyset}& \xrightarrow{\gamma_{i}} & X_{i}+X_{i} \label{model3} \\
X_{i} + {\emptyset}& \xrightarrow{\beta_{i}} & {\emptyset}+X_{i} \label{model4}
\end{eqnarray}
where every lattice site can be occupied by at most one individual, whereas reactions between different individuals only
occur when they are at neighboring sites. Our approach remains valid when considering in addition self-predation.
The model parameters $\delta_{ij}$, $\alpha_{ij}$, $\gamma_{i}$, and $\beta_{i}$
are the rates at which these reactions take place.
For the examples included in this paper, predation and swapping rates are always considered
to be independent of the prey, i.e. $\delta_{ij} = \delta_i$ and $\alpha_{ij} = \alpha_i$.
The separation of the predation and birth events is needed for the formation of spirals as those shown in Fig. \ref{fig1},
but is otherwise not relevant in order to establish the relationships between species.
We also allow for various ways of mobility (swapping of individuals
on neighboring sites (\ref{model3}) or diffusion by jumping to empty sites (\ref{model4})), in order to be as general as possible.
The reaction scheme (\ref{model1})-(\ref{model2}) contains many known situations as for example the cyclic $N$ species Lotka-Volterra and
May-Leonard models as well as more complicated situations like the $(N,r)$ models where each species preys on $r$ other species
in a cyclic way \citep{Roman13,Mowlaei14}. Whereas the examples shown in Fig. \ref{fig1} are those of $(N,r)$ models with
symmetric rates (hence we can drop the labels $i$, $j$ for the rates, see the figure caption), more general cases
will be discussed below.
The matrix approach discussed in the following is valid for rather general predator-prey networks, provided that
no interactions of higher order than second order are present. We assume that the species network gives rise
to a unique partition of the species into domains. While we have not encountered a case where this assumption is
not fulfilled, we can not exclude that there exist networks of species that exhibit in presence of noise two non-isomorphic partitions
of species into domains. For the cases used to illustrate our approach, we consider rates for which none
of the species go extinct before the quasi-steady state of the dynamics is reached, i.e.
the time scale on which the species are forming alliances and domains is shorter than the extinction time
for any of the species. Notably this requires that the diffusion and mobility rates are not too high so as to induce rapid extinction
of the species, but are still large enough to allow the formation of species alliances.
\section{Matrix approach}
The predation scheme (\ref{model1}) can be represented by a graph where one connects species $i$ and $j$ by a directed edge
if $i$ preys on $j$. The scheme can also be encoded in the $N \times N$ {\it adjacency matrix} $\bold{\underline{A}}$ where the element
$a_{ij} = 1$ if $i$ preys on $j$ and 0 otherwise. The {\it square of the adjacency matrix} $\bold{\underline{B}} = \bold{\underline{A}}^2$
contains information about preferred partnership formations based on the maximum protection a species can enjoy while being
close to some other species. Indeed, the element $b_{ij}$ counts the number of
directed paths of length 2 from vertex $i$ to vertex $j$, i.e. the number of paths of the form $i \longrightarrow k \longrightarrow j$
where $k$ is a vertex different from $i$ and $j$. Following the maxim that "the enemy of my enemy is my friend," species $j$
wants to be in close proximity to that species that attacks most of its predators. Mathematically, this preferred partner of
$j$ is obtained from the condition $\max_i b_{ij}$.
Of course, a given species does not want to undergo a close partnership with one of its predators. Adding this important information
yields the {\it alliance matrix} $\bold{\underline{S}}$ whose elements $s_{ij}$ provide information on the best possible allies $j$
for each species $i$:
\begin{equation}
s_{ij} = \delta_{b_{ji}, m_i}\delta_{0,a_{ji}}~,
\end{equation}
where $m_i = \max_{j} b_{ji}$. The first condition selects those species $j$ that are preferred partners of $i$ as they
provide the best protection by attacking the largest number of its predators, whereas the second condition makes sure that $j$ is not itself a predator of $i$, yielding
a matrix with elements $s_{ij} = 1$ for the best possible allies of species $i$ and zero otherwise.
In order to fully understand space-time patterns as those shown in Fig. \ref{fig1}, one needs to further distinguish between preferred allies
of species $i$ that are hunted by species $i$ and preferred allies towards which species $i$ has a neutral approach. This information is encoded by the two matrices
$\bold{\underline{P}}$ ({\it prey-allies}) and $\bold{\underline{N}}$ ({\it neutral allies}) with elements
\begin{eqnarray}
p_{ij} &=& s_{ij} \wedge a_{ij}=\delta_{s_{ij},a_{ij}} \\
n_{ij} &=& s_{ij }\wedge \neg \, a_{ij}=\delta_{s_{ij},(1-a_{ij})}
\end{eqnarray}
with the binary operators AND ($\wedge$) and NOT ($\neg$).
The two matrices $\bold{\underline{P}}$ and $\bold{\underline{N}}$ allow to understand (and predict) many of the space-time patterns that can emerge
in predator-prey systems governed by the reactions (\ref{model1})-(\ref{model2}). A matrix element $p_{ij}=1$ means that species $i$ seeks
the protection of species $j$, while at the same time preying on species $j$. As a result single species waves will form where a wave containing
individuals of species $i$ will follow one formed by species $j$. On the other hand $n_{ij}=1$ represents non-aggressive coexistence
where species $i$ is protected by a neutral species $j$ (i.e. $i$ does not prey on $j$). As a result $i$ will mix with $j$ in order
to enjoy the enhanced protection by its preferred ally. If $n_{ij}=n_{ji}=1$, then the two mutually neutral partners will form neutral domains
and protect each other.
Inspection of Fig. \ref{fig1}d reveals a case that is not yet covered by the matrices given above, namely that of 'fuzzy' waves where
three species mix to some degree. Consider three given species $i$, $j$, and $k$. The above
situation is encountered when species $i$ and $j$ are mutually neutral and species $j$ and $k$ are mutually neutral,
while at the same time species $i$ preys on species $k$. This means that the following conditions have to be fulfilled simultaneously:
$n_{ij}=1$, $n_{jk}=1$, and $a_{ik}=1$. As a result, $i$ chases species $k$ while both mix with the intermediate species $j$, giving fuzzy
looking waves as those seen in Fig. \ref{fig1}d. Defining the matrix elements of a matrix $\bold{\underline{F}}$ as
\begin{equation} \label{eq:f}
f_{ij}=\theta \left(\left( \sum_{k=1}^{N} \delta_{n_{ij},1} \delta_{n_{jk},1} \delta_{a_{ik},1} \right) - 1\right)~,
\end{equation}
with $\theta(x) = 1$ for $x \ge 0$ and zero otherwise,
we have that $f_{ij}=1$ if $j$ acts as a neutral intermediary between species $i$ and some other species, whereas $f_{ij}=0$ otherwise.
\section{Examples}
Let us first look at the four cyclic $(N,r)$ games of Fig. \ref{fig1} in order to see how to work with the different matrices.
For the $(5,3)$ game shown in Fig. \ref{fig1}a the relevant matrix is
\begin{equation}
\renewcommand{\arraystretch}{.45}
\bold{\underline{P}}=\left(
\begin{array}{ccccc}
0 & 1 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 1 \\
1 & 0 & 0 & 0 & 0 \\
\end{array}
\right)
\end{equation}
whereas $\bold{\underline{N}} = \bold{\underline{F}} = \bold{\underline{0}}$. We readily obtain from the prey-allies matrix
$\bold{\underline{P}}$ that species 1 wants to ally with its prey species 2, species 2 wants to ally
with its species 3, etc. This yields single species spirals where the species follow each other in the order
$1 \rightarrow 2 \rightarrow 3 \rightarrow 4 \rightarrow 5 \rightarrow 1$.
The $(6,3)$ in Fig. \ref{fig1}b is characterized by a similar looking $6 \times 6$ matrix
\begin{equation}
\renewcommand{\arraystretch}{.45}
\bold{\underline{P}}=\left(
\begin{array}{cccccc}
0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0 & 1 \\
1 & 0 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 & 0 \\
\end{array}
\right)
\end{equation}
where again $\bold{\underline{N}} = \bold{\underline{F}} = \bold{\underline{0}}$. This time, however, two family of spirals
(or teams) are formed, one with the ordering $1 \rightarrow 3 \rightarrow 5 \rightarrow 1$, the other with the ordering
$2 \rightarrow 4 \rightarrow 6 \rightarrow 2$. As every member in a team has both a prey and a predator in the other team,
these two family of spirals compete against each other, which yields a coarsening process with a three-species rock-paper-scissors
game within each domain.
The $(6,2)$ case of Fig. \ref{fig1}c is a case where the prey-allies matrix $\bold{\underline{P}}$ vanishes whereas the neutral allies matrix
$\bold{\underline{N}}$ is non-zero:
\begin{equation}
\renewcommand{\arraystretch}{.45}
\bold{\underline{N}}=\left(
\begin{array}{cccccc}
0 & 0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0 & 1 \\
1 & 0 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & 0 \\
\end{array}
\right)~.
\end{equation}
Inspection of $\bold{\underline{N}}$ reveals the formation of three teams composed of two mutually neutral species (1 and 4,
2 and 5, 3 and 6), resulting in a coarsening process with three different types of domains.
Finally, the $(9,3)$ game in Fig. \ref{fig1}d provides us with an example where both $\bold{\underline{N}}$ and $\bold{\underline{F}}$
are non-zero:
\begin{equation}
\renewcommand{\arraystretch}{.45}
\bold{\underline{N}}=\bold{\underline{F}}=
\left(
\begin{array}{ccccccccc}
0 & 0 & 0 & 0 & 0 & 1 & 0 & 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 & 0 & 0 & 0 & 1 \\
1 & 0 & 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 \\
0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 \\
\end{array}
\right)~,
\end{equation}
whereas $\bold{\underline{P}} = \bold{\underline{0}}$.
It follows from this that species 5, which is the preferred ally of species 1, acts as an intermediary between species 1 and 9. As species 9 preys
on species 1, this yields a wavefront of the type $9 \stackrel{\left[ 5 \right]}{\longrightarrow} 1$ where $\left[ \cdots \right]$ indicates
a neutral intermediary. As this happens in a cyclic way, one observes the emergence of fuzzy spirals, with the ordering
$1 \stackrel{\left[ 6 \right]}{\longrightarrow} 2 \stackrel{\left[ 7 \right]}{\longrightarrow} 3
\stackrel{\left[ 8 \right]}{\longrightarrow} 4 \stackrel{\left[ 9 \right]}{\longrightarrow} 5 \stackrel{\left[ 1 \right]}{\longrightarrow} 6
\stackrel{\left[ 2 \right]}{\longrightarrow} 7 \stackrel{\left[ 3 \right]}{\longrightarrow} 8 \stackrel{\left[ 4 \right]}{\longrightarrow}
9 \stackrel{\left[ 5 \right]}{\longrightarrow} 1$.
Let us now look at the two more involved examples shown in Figs. \ref{fig2} and \ref{fig3}
where the interaction scheme is no longer cyclic.
\begin{figure} [!h]
\begin{center}
~~~~~~~~~$t=100$~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~$t=900$~~~~~~~~~~\\
\includegraphics[width=0.35\columnwidth]{figure2a}~~~
\includegraphics[width=0.35\columnwidth]{figure2b}~~~
\end{center}
\caption{(Color online)
Six-species model defined by the adjacency matrix (\ref{A_example1}) and the rates (\ref{R_example1}).
Red: species 1, green: species 2, blue: species 3, yellow: species 4, cyan: species 5, pink: species 6.
The system size is $500 \times 500$.
Times are measured since preparing the system, where one time step corresponds to $500 \times 500$
proposed updates.
\label{fig2}
}
\end{figure}
The snapshots in Fig.\ \ref{fig2} result from a six-species model with an interaction scheme described by the matrix
\begin{equation} \label{A_example1}
\renewcommand{\arraystretch}{.45}
{\bold{\underline{A}}}= \left(
\begin{array}{cccccc}
0 & 1 & 1 & 1 & 0 & 0 \\
0 & 0 & 1 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 & 1 & 1 \\
1 & 0 & 0 & 0 & 1 & 0 \\
1 & 1 & 0 & 0 & 0 & 1 \\
1 & 0 & 1 & 0 & 0 & 0 \\
\end{array}
\right)
\end{equation}
and rates given by
\begin{equation} \label{R_example1}
\renewcommand{\arraystretch}{.45}
\left(
\begin{array}{cccc}
\delta_1 & \alpha_1 & \gamma_1 & \beta_1 \\
\delta_2 & \alpha_2 & \gamma_2 & \beta_2 \\
\delta_3 & \alpha_3 & \gamma_3 & \beta_3 \\
\delta_4 & \alpha_4 & \gamma_4 & \beta_4 \\
\delta_5 & \alpha_5 & \gamma_5 & \beta_5 \\
\delta_6 & \alpha_6 & \gamma_6 & \beta_6
\end{array}\right)=
\left(
\begin{array}{cccc}
0.9 & 0.08 & 0.8 & 0.16 \\
0.43 & 0.46 & 0.8 & 0.16 \\
0.9 & 0.08 & 0.8 & 0.16 \\
0.43 & 0.46 & 0.8 & 0.16 \\
0.9 & 0.08 & 0.8 & 0.16 \\
0.43 & 0.46 & 0.8 & 0.16
\end{array}\right).
\end{equation}
Following the construction explained above, we get the matrices
\begin{equation}
\renewcommand{\arraystretch}{.45}
{\bold{\underline{P}}}= \left(
\begin{array}{cccccc}
0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
1 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
\end{array}
\right)
~;~
{\bold{\underline{N}}}= \left(
\begin{array}{cccccc}
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 1 \\
0 & 0 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 & 0 \\
\end{array}
\right)
\end{equation}
which, together with the fact that ${\bold{\underline{F}}} = {\bold{\underline{0}}}$,
allow us to predict the alliances and space-time pattern appearing in this system. In the present case,
both the prey-allies and neutral allies matrices are non-zero, so that we can expect the formation of
both neutral alliances as well as of alliances where the partners remain in a predator-prey relationship.
Indeed, from the matrix ${\bold{\underline{P}}}$ we obtain that species 1, 3, and 5 form a three-species
cyclic game that results in the formation of compact wavefronts with the following ordering of
the species: $1 \rightarrow 3 \rightarrow 5 \rightarrow 1$. On the other hand, from the matrix ${\bold{\underline{N}}}$
follows that the species 2, 4, and 6 form a neutral alliance. As a result, two types of domains should form,
one composed of three neutral partners and one composed of three partners that undergo a cyclic rock-paper-scissors
game, resulting in a coarsening process with different internal dynamics inside the domains.
\begin{figure} [!h]
\begin{center}
~~~~~~~~~$t=200$~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~$t=600$~~~~~~~~~~\\
\includegraphics[width=0.35\columnwidth]{figure3a}~~~
\includegraphics[width=0.35\columnwidth]{figure3b}~~~
\end{center}
\caption{(Color online)
Five-species model defined by the adjacency matrix (\ref{A_example2}) and the rates (\ref{R_example2}).
Red: species 1, green: species 2, blue: species 3, yellow: species 4, cyan: species 5.
The system size is $400 \times 400$.
Times are measured since preparing the system, where one time step corresponds to $400 \times 400$
proposed updates.
\label{fig3}
}
\end{figure}
The five-species model shown in Fig. \ref{fig3} is defined by the adjacency matrix
\begin{equation} \label{A_example2}
\renewcommand{\arraystretch}{.45}
{\bold{\underline{A}}}= \left(
\begin{array}{ccccc}
0 & 1 & 1 & 1 & 1 \\
1 & 0 & 1 & 1 & 0 \\
1 & 0 & 0 & 1 & 0 \\
1 & 0 & 0 & 0 & 1 \\
1 & 1 & 0 & 0 & 0
\end{array}\right)~.
\end{equation}
For the time evolution shown in Fig. \ref{fig3} we used the rates
\begin{equation} \label{R_example2}
\renewcommand{\arraystretch}{.45}
\left(
\begin{array}{ccccc}
\delta_1 & \alpha_1 & \gamma_1 & \beta_1 \\
\delta_2 & \alpha_2 & \gamma_2 & \beta_2 \\
\delta_3 & \alpha_3 & \gamma_3 & \beta_3 \\
\delta_4 & \alpha_4 & \gamma_4 & \beta_4 \\
\delta_5 & \alpha_5 & \gamma_5 & \beta_5
\end{array}\right)=
\left(
\begin{array}{ccccc}
0.5 & 0.4 & 0.8 & 0.16 \\
0.7 & 0.24 & 0.8 & 0.16 \\
0.3 & 0.56 & 0.8 & 0.16 \\
0.8 & 0.16 & 0.8 & 0.16 \\
0.3 & 0.56 & 0.8 & 0.16
\end{array}\right).
\end{equation}
From ${\bold{\underline{A}}}$ we readily obtain
\begin{equation}
\renewcommand{\arraystretch}{.45}
{\bold{\underline{P}}}= \left(
\begin{array}{ccccc}
0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 \\
0 & 1 & 0 & 0 & 0
\end{array}\right)
~;~
{\bold{\underline{N}}}= \left(
\begin{array}{ccccc}
1 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 \\
0 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0
\end{array}\right)~,
\end{equation}
whereas ${\bold{\underline{F}}}={\bold{\underline{0}}}$ (i.e. no fuzzy wavefronts are formed). The matrices
immediately yield all the relevant information regarding the partnerships forming in this system. As
$n_{11} =1$, species 1 wants to self-segregate, which is a direct consequence of the fact that all the
other species are both its predators and preys. The neutral allies matrix ${\bold{\underline{N}}}$ also indicates that species 3 and 5
form a neutral alliance, in order to fight off their corresponding predators. Finally, the prey-allies matrix ${\bold{\underline{P}}}$
reveals that the three species 2, 4, and 5 tend to form an active three-species alliance with the following ordering of the wave fronts:
$2 \rightarrow 4 \rightarrow 5 \rightarrow 2$. Our method therefore predicts the formation of three types of domains (pure species 1 domains, domains
of the neutral alliance between species 3 and 5, as well as domains where species 2, 4, and 5 undergo a cyclic game), where, remarkably,
species 5 is involved in {\it two} different alliances. All these predictions are indeed verified when inspecting the snapshots in
Fig. \ref{fig3}.
Knowing the partnerships and the resulting space-time pattern is an important step in order to better understand
the fate of an ecology. This alone, though, is not enough to predict whether a given species will go extinct, as the
survival of a species also depends on the rates of its interactions with the other species. The reliability of our predictions at least for the $(N,r)$ model with symmetric rates is confirmed via exact agreement with the theoretical predictions in \citep{Mowlaei14} obtained using the complex Ginzburg Landau approach. Comprehensive numerical studies are required to fully understand the parameter regimes for which the presented ad-hoc methods are valid. We leave such calculations for future investigations.
\section{Summary}
We have presented a theoretical approach that allows to predict partnership formation and emerging space-time patterns in
many-species food networks. While our emphasis was on predator-prey systems, the proposed use of various matrices, derived from
the adjacency matrix, to understand alliances and spatio-temporal pattern formation can be straightforwardly expanded to
other systems, ranging from ecology to social sciences \citep{Smith82,Hof98,Now06,Hau02,Gok14},
governed by binary interactions between different 'species'.
\section*{Acknowledgements}
M.P. thanks the Galileo Galilei
Institute for Theoretical Physics for the hospitality and
the INFN for partial support during the completion of
this work.\\
{\it Funding:} This work is supported by the US National
Science Foundation through grant DMR-1205309.
|
\section{Introduction}
\label{sec:intro}
Consider the eigenvalue equation
\begin{equation}
\label{general}
-u''(x) + V(x) u(x) = \lambda u(x), \;\; x\in (a,b) \subseteq \R,
\end{equation}
associated with a one-dimensional Schr\"odinger operator $H = -d^2/dx^2+V$, where the potential $V:(a,b) \to \R$, and
the boundary condition if $(a,b) \neq \R$, are chosen such that the spectrum consists of a discrete sequence of eigenvalues $\{\lambda_k\}$.
One possible way of linking the behaviour of this sequence to properties of the potential $V$ is via a regularized trace formula for
the sum of the eigenvalues. The classical example is the formula attributed to Gelfand and Levitan,
which, if we take $(a,b) = (0,\pi)$ with Dirichlet boundary conditions on the endpoints, reads
\begin{equation}
\label{classtr}
\sum_{k=1}^\infty \left[\lambda_k - k^2 - \frac{1}{\pi}\int_0^\pi V(x)\dx{x} \right]
= \frac{1}{2\pi}\int_0^\pi V(x)\dx{x} - \frac{V(0)+V(\pi)}{4}
\end{equation}
(see, e.g., the book \cite{lesa}, also for other similar formulae). Since the values $k^2$ are in fact the eigenvalues of the
Dirichlet Laplacian, that is, the corresponding Schr\"odinger operator with zero potential, this is a comparison between the
eigenvalues of the operators $H$ and $H_0:=-d^2/dx^2$.
More recently it has also been shown that an analogous trace formula holds for the eigenvalues of~\eqref{general} on the
whole line $(a,b) = \R$~\cite{abp,puso}. The comparison case is now provided by the quantum harmonic oscillator
\begin{equation}
\label{qho}
-u''(x) + x^2 u(x) = \lambda u(x), \;\; x\in\R,
\end{equation}
whose eigenvalues are given by $\lambda_k^0 = 2k+1$ for $k \in \N$. Writing the potential in \eqref{general} as $V(x) = x^2 + q(x)$, that is,
as a perturbed harmonic oscillator, if the perturbation $q:\R \to \R$ is small enough in an appropriate sense,
then the eigenvalues of~\eqref{general}, which we denote by $\lambda_k$ for $k \in \N$, satisfy the trace formula
\begin{equation}
\label{photrace}
\dsum_{k=0}^{\infty}\left[ \lambda_{k}-\lambda_{k}^{0}-\frac{\ds 1}
{\ds \pi \sqrt{\lambda_{k}^{0}}}\dint_{\R} q(x) \dx{x}\right] = -\frac{\ds Z_0(1/2)}{\ds \pi}
\dint_{\R} q(x) \dx{x},
\end{equation}
where
\begin{equation}
\label{basespeczeta}
Z_0(s) =(1-2^{-s})\zeta(s) = \sum_{k=1}^{\infty} \frac{\ds 1}{\ds \left(\lambda_{k}^{0}\right)^{s}}
\end{equation}
is the spectral zeta function associated with \eqref{qho}, the second equality being valid for $\textrm{Re}\,{s}>1$,
and $\zeta(\,.\,)$ is the Riemann zeta function; see~\cite[Theorem~2]{abp} or~\cite[Eq.~(1.12)]{puso}.
We refer to \cite{sapo} for a wide-ranging general survey on the theory of regularized traces.
In a separate paper \cite{frke} we show that formula~\eqref{classtr} is in fact the limit as $n \to \infty$ of a sequence of inequalities
for the (finite) sums of the first $n$ eigenvalues given in terms of the Fourier coefficients of the potential, and that~\eqref{classtr}
can be proved by combining these inequalities with knowledge of the asymptotic behaviour of the eigenvalues and eigenfunctions~\cite{frke}.
In the present paper, which may be viewed as a continuation of~\cite{frke}, we show that a similar family of inequalities is valid
for the perturbed harmonic oscillator assuming that the perturbation $q$ is non-negative and of finite $L^1(\R)$-norm. More precisely, we
shall prove in Theorem~\ref{mainth1} below that there is a sequence of inequalities of the form
\begin{displaymath}
\dsum_{k=0}^{n}\left[ \lambda_{k}-\lambda_{k}^{0}-\frac{\ds 1}
{\ds \pi \sqrt{\lambda_{k}^{0}}}\dint_{\R} q(x) \dx{x}\right]\leq \frac{\ds \chi_{n}}{\ds \pi}
\dint_{\R} q(x) \dx{x}
\end{displaymath}
for all $n \in \N$ if $V(x)=x^2+q(x)$ with $0 \leq q \in L^1(\R)$, where the sequence $\chi_n$, which is given explicitly, depends
only on properties of the eigenfunctions and
eigenvalues of the quantum harmonic oscillator~\eqref{qho} and converges to $-Z_0(1/2)$ like $\bo(1/\sqrt{n})$ as $n\to\infty$.
A similar sequence of bounds will also be shown to hold for a certain class of negative or indefinite potentials (see Theorem~\ref{mainth2}),
and although the corresponding bounding sequence we obtain is larger than $\chi_n$, it is still explicit, and the order of convergence to
the known trace formula remains $\bo(1/\sqrt{n})$.
These results will be established via test function methods, using for this purpose the eigenfunctions of \eqref{qho} in a
suitable Rayleigh quotient expression for the eigenvalues of the perturbed harmonic oscillator, and then combining this with properties of
Hermite polynomials to analyze the resulting expression. We believe one of these properties, namely Lemma~\ref{lemma:hn}, which provides
an upper bound for the function $e^{-x^2}\left[H_{n+1}^{2}(x)-H_{n}(x)H_{n+2}(x)\right]$ to be new and interesting in its own right.
In fact, these results---and the corresponding proofs---differ from those in~\cite{frke} in that for them we do not use a decomposition of
the potential in terms of the eigenfunctions of the unperturbed problem. However, such an approach is also possible in this case and we
carry it out to obtain a different type of bound; see Theorem~\ref{genpot}. For this particular result we assume that
$V \in L^2(\R,e^{-x^2}\dm{x})$, that is, that the potential is no longer necessarily a perturbation of $x^2$, but rather more
generally merely square integrable with respect to the weighted $L^2$-measure most naturally associated with the problem~\eqref{qho}.
The resulting bounds (which are once again explicit) are expressed in terms of the Fourier-like coefficients of $V$ expanded as a sum of
Hermite polynomials. These are actually stronger than Theorems~\ref{mainth1} and~\ref{mainth2}, as the only inequality used now is
that which arises from the substitution of test functions in the Rayleigh quotient (see Remark~\ref{rem:genpot}(i)). However, now the
finite sums converging to the the left-hand side of the trace formula~\eqref{photrace} do not appear in a natural way; this will then be
derived as a simple corollary by writing the potential $V(x)$ as $x^2+q(x)$ and using the Fourier coefficients for $q$ instead.
We also generalize Theorems~\ref{mainth1} and~\ref{mainth2} to obtain bounds on sums of powers of the eigenvalues in Section~\ref{sec:power}.
\section{Schr\"odinger operators on the real line}
\label{sec:prelim}
Throughout this paper we will consider one-dimensional Schr\"odinger operators on the real line, that is, associated with the
equation~\eqref{general} for $x \in \R$, where the potential $V:\R \to \R$ is a locally measurable function on which we will impose
various (and varying) assumptions. We will always assume that $V(x) \to \infty$ as $|x| \to \infty$, so that the operator associated with the problem \eqref{general} considered as an operator on $L^2(\R)$ has discrete spectrum, and we will in general denote the associated eigenvalues by $\lambda_0 < \lambda_1 \leq \ldots \to \infty$.
As is well known, the eigenvalues of the quantum harmonic oscillator \eqref{qho}, which will play the role of our ``default'' problem,
are given by $\lambda_{k}^{0}=2k+1$ for $k\in\N$, with corresponding eigenfunctions $\psi_{k}(x) = e^{-x^2/2}H_{k}(x)$, which form an orthonormal basis of $L^2(\R)$. Here $H_{k}$ denotes the $k^{\rm th}$ Hermite polynomial (see, e.g., \cite[Ch.~5]{szeg}).
Of particular interest to us will be the perturbed harmonic oscillator
\begin{equation}
\label{pertharm}
-u''(x) + \left[ x^2 + q(x)\right] u(x) = \lambda u(x), \;\; x\in\R,
\end{equation}
which is easily seen to have discrete spectrum if $q \in L^p(\R)$ for some $p \in [1,\infty]$.
For a general potential $V:\R\to\R$, we can characterize the associated eigenvalues via classical variational methods. Denoting by
$\varphi \in H^1(\R) \cap L^2(\R,V(x)\dm{x})$ an arbitrary test function, we let
\begin{equation}
\label{rayleigh}
\mathcal{R}[V,\varphi]:= \frac{\dint_{\R} \left(\varphi'(x)\right)^2\dx{x}+\dint_{\R}V(x)\varphi^2(x)\dx{x}}
{\dint_{\R} \varphi^2(x)\dx{x}}
\end{equation}
be the Rayleigh quotient associated with (the Schr\"odinger operator with potential) $V$ at $\varphi$. A standard generalization of
the usual minimax formula for eigenvalues states that if $\varphi_0,\ldots,\varphi_n$ is a collection of $n+1$ such functions orthogonal in $L^2(\R)$, for any $n\in\N$, then
\begin{displaymath}
\sum_{k=0}^{n} \lambda_k \leq \sum_{k=0}^{n} \mathcal{R}[V,\varphi_k]
\end{displaymath}
(see, e.g., \cite{band}), with equality being achieved when the $\varphi_k$ are the first $n+1$ eigenfunctions. For us the most natural choice of test functions will be the eigenfunctions $\psi_k$ of the quantum harmonic operator.
\section{Bounds for the perturbed harmonic oscillator with a non-negative perturbation}
\label{sec:pho}
In this section we will state and prove our main theorem, obtaining the aforementioned finite version of the trace formula~\eqref{photrace} for the general perturbed harmonic oscillator \eqref{pertharm}.
\begin{theorem}
\label{mainth1}
Let $q$ be a non-negative potential defined on the real line having finite $L^{1}(\R)$ norm. Then the eigenvalues of \eqref{pertharm} satisfy the inequalities
\begin{equation}
\label{maineq}
\dsum_{k=0}^{n}\left[ \lambda_{k}-\lambda_{k}^{0}-\frac{\ds 1}
{\ds \pi \sqrt{\lambda_{k}^{0}}}\dint_{\R} q(x) \dx{x}\right]\leq \frac{\ds \chi_{n}}{\ds \pi}
\dint_{\R} q(x) \dx{x},\;\; n=0,1,\cdots,
\end{equation}
where
\begin{equation}
\label{chi}
\chi_{n} = \left\{
\begin{array}{ll}
\frac{\ds 2n+3}{\ds n+1}\frac{\ds \Gamma\left(\frac{\ds n}{\ds 2}+1\right)}
{\ds \Gamma\left(\frac{\ds n+1}{\ds 2}\right)}-\dsum_{k=0}^{n}\frac{\ds 1}
{\ds \sqrt{\lambda_{k}^{0}}},
& n \mbox{ odd,}\eqskip
(n+1)\frac{\ds \Gamma\left(\frac{\ds n+1}{\ds 2}\right)}{\ds \Gamma\left(\frac{\ds n}{\ds 2}+1\right)}
-\dsum_{k=0}^{n}\frac{\ds 1}{\ds \sqrt{\lambda_{k}^{0}}},
& n \mbox{ even.}
\end{array}
\right.
\end{equation}
Furthermore, $\chi_{n} = -Z_{0}(1/2) + \bo(1/\sqrt{n})$, where $Z_{0}(s) = (1-2^{-s})\zeta(s)$.
\end{theorem}
\begin{remark}
(i) It is essential for our method of proof that $q$ be non-negative. In Theorem~\ref{mainth2} below, we weaken this assumption
and obtain a slightly weaker set of inequalities which nevertheless still converge in the limit to the trace formula~\eqref{photrace}
with the same order of convergence $\bo(1/\sqrt{n})$. It is not clear if the inequalities \eqref{maineq} are true for arbitrary
$q \in L^1(\R)$; the trace formula \eqref{photrace} is itself currently only known to hold under stronger assumptions on $q$: in
\cite{abp} a certain rate of decay of $q$ at infinity is assumed, and in \cite{puso} it is assumed $q$ has compact support. We
remark however that having convergence of order $\bo(1/\sqrt{n})$ is most probably optimal, since this is the rate at which we
have convergence of the sequence whose limit defines $\zeta(1/2)$ (cf.~\eqref{zetaexp} and \eqref{z0}).
(ii) There do not exist corresponding lower bounds for \emph{finite} sums of eigenvalues: for any fixed $n \geq 0$ is it always
possible to find a function $0 \leq q \in L^1(\R)$ for which the left-hand side of \eqref{maineq} is arbitrarily large negative;
see Proposition~\ref{prop:lower} below. However, for a \emph{fixed} potential it is a natural question as to whether we can recover
a lower bound valid in the asymptotic limit. Indeed, it might be possible to extend our result to give a new proof of the trace
formula \eqref{photrace} for a different class of (non-negative) potentials $q$ from those considered
in~\cite{abp,puso}, namely $q \in L^1(\R)$. The idea would be to argue as in \cite{dikii} (or \cite{frke}), to show
that the degree of ``error'' which arises from using the eigenfunctions $\psi_k$ of the unperturbed problem as test
functions becomes asymptotically small as $k \to \infty$: denoting by $\varphi_k$ the eigenfunction associated with
$\lambda_k$ (corresponding to the potential $V(x)=x^2+q(x)$), we see that the trace formula holds whenever
\begin{equation}
\label{rayleigh-lim}
\lim_{n \to \infty} \dsum_{k=0}^{n} \left(\mathcal{R}[x^2+q(x),\varphi_k] - \mathcal{R}[x^2+q(x),\psi_k]\right) = 0,
\end{equation}
since by definition $\mathcal{R}[x^2+q(x),\varphi_k]=\lambda_k$. We can rewrite \eqref{rayleigh-lim} as a type of ``change of basis'' formula
\begin{displaymath}
\lim_{n \to \infty} \dsum_{k=0}^{n} \big(\langle \varphi_k, H\varphi_k \rangle - \langle \psi_k, H\psi_k\rangle\big)=0,
\end{displaymath}
where $H: D(H)\subset L^2(\R) \to L^2(\R)$ is the operator associated with the potential $x^2+q(x)$.
We \emph{expect} this to hold whenever the asymptotics for $\lambda_k$ and $\varphi_k$ are similar enough to those of
$\lambda_k^0$ and $\psi_k$, respectively, when $k \to \infty$. This is, however, likely to be a difficult problem,
and we shall not attempt an investigation of it here.
\end{remark}
For notational convenience, for $n\geq 0$ we define
\begin{equation}
\label{omega}
\omega_n:=\chi_n+\dsum_{k=0}^n \frac{1}{\sqrt{\lambda_k^0}}= \left\{
\begin{array}{ll}
\frac{\ds 2n+3}{\ds n+1}\frac{\ds \Gamma\left(\frac{\ds n}{\ds 2}+1\right)}
{\ds \Gamma\left(\frac{\ds n+1}{\ds 2}\right)},
& n \mbox{ odd,}\eqskip
(n+1)\frac{\ds \Gamma\left(\frac{\ds n+1}{\ds 2}\right)}{\ds \Gamma\left(\frac{\ds n}{\ds 2}+1\right)},
& n \mbox{ even,}
\end{array}
\right.
\end{equation}
and we also set $\omega_{-1}:= 0$.
\begin{proof}[Proof of Theorem~\ref{mainth1}]
Using the first $n+1$ eigenfunctions of the unperturbed harmonic oscillator \eqref{qho}, given by $\psi_k(x) =e^{-x^2/2} H_k(x)$,
$k=0,\ldots,n$, as test functions in the Rayleigh quotient \eqref{rayleigh} for $V(x)=x^ 2+q(x)$ yields
\begin{equation}
\label{basicsum}
\begin{array}{lll}
\dsum_{k=0}^{n} \lambda_{k} & \leq & \dsum_{k=0}^{n} \frac{\ds \dint_{\R}\left[\frac{\ds d}{\ds dx}
\left[e^{-x^2/2}H_{k}(x)\right]\right]^2
+
\left[x^2+q(x)\right]e^{-x^2}H_{k}^{2}(x)\dx{x}}{\ds \dint_{\R}e^{-x^2}H_{k}^{2}(x)\dx{x}}\eqskip
& = & \dsum_{k=0}^{n} \lambda_{k}^{0} + \dint_{\R} e^{-x^2} q(x)\dsum_{k=0}^{n}\frac{\ds 1}
{\ds 2^k k!\sqrt{\pi}}H_{k}^{2}(x)\dx{x}
\end{array}
\end{equation}
From basic properties of Hermite polynomials we have the identity
\begin{equation}
\label{sumH}
\dsum_{k=0}^{n}\frac{\ds 1}{\ds 2^k k!}H_{k}^{2}(x)=\frac{\ds 1}{\ds 2^{n+1}n!}\left[ H_{n+1}^{2}(x)
-H_{n}(x)H_{n+2}(x)\right].
\end{equation}
This arises in the context of Tur\'an's inequality for Hermite polynomials (cf. \cite[p. 404]{szeg1}), and can easily be derived directly
by induction in $n$---see also, for instance,~\cite[p.~106]{szeg}. By using the estimate of the function
\begin{equation}
\label{h}
h_{n}(x) := e^{-x^2}\left[H_{n+1}^{2}(x)-H_{n}(x)H_{n+2}(x)\right]
\end{equation}
given in Lemma~\ref{lemma:hn} below in \eqref{sumH} and inserting this into \eqref{basicsum}, we obtain
\begin{multline}
\label{comb}
\dsum_{k=0}^{n} \lambda_{k} \leq \dsum_{k=0}^{n} \lambda_{k}^{0} + \dint_{\R} e^{-x^2} q(x)\dsum_{k=0}^{n}\frac{\ds 1}
{\ds 2^k k!\sqrt{\pi}}H_{k}^{2}(x)\dx{x} \\
\leq \dsum_{k=0}^{n} \lambda_{k}^{0} + \frac{\omega_n}{\pi}\dint_{\R} q(x)\dx{x},
\end{multline}
which upon rearranging yields \eqref{maineq}.
We now give the (routine) proof that $\chi_n = -Z_0(1/2) + \bo(1/\sqrt{n})$ as $n \to \infty$. We first note that
\begin{equation}
\label{zetaexp}
\zeta(s) = \sum_{k=1}^n k^{-s}+ s\int_n^\infty \frac{\lfloor x\rfloor - x + \frac{1}{2}}{x^{s+1}} \dx{x} +
\frac{n^{1-s}}{s-1} - \frac{1}{2n^{s}},
\end{equation}
valid for $s>0$ (see \cite{titc}, Eq.~(3.5.3), pp.~49-50). Setting $s=1/2$ and passing to the limit as $n \to \infty$,
this means we can write
\begin{equation}
\label{z0}
-Z_0(1/2) = -\left(1-\frac{1}{\sqrt{2}}\right) \zeta(1/2)
= \left(1-\frac{1}{\sqrt{2}}\right)\lim_{n \to \infty} a_n,
\end{equation}
where for ease of notation we have set
\begin{equation}
\label{an}
a_n:= 2\sqrt{n} - \sum_{k=1}^n\frac{1}{\sqrt{k}}
\end{equation}
for $n \geq 1$. Now, recalling that $\lambda_k^0 = 2k+1$ for $k \in \N$, we have
\begin{equation}
\label{eulersum}
\chi_n = \omega_n - \sum_{k=0}^n \frac{1}{\sqrt{2k+1}} = \omega_n - \left(1-\frac{1}{\sqrt{2}}\right)\sum_{k=1}^{n} \frac{1}{\sqrt{k}} - \sum_{k=n+1}^{2n+1} \frac{1}{\sqrt{k}};
\end{equation}
we wish to show that this converges to $-Z_0(1/2)$ as $n\to\infty$. We first establish that
\begin{equation}
\label{omegaconv}
\omega_n = \sqrt{2n} + \bo\left(\frac{1}{\sqrt{n}}\right),
\end{equation}
using the following asymptotics for the quotient of two gamma functions (see~\cite{abrasteg}, formula $6.1.47$, for instance):
\begin{equation}
\label{gammaexp}
\frac{\Gamma\left(z+\frac{1}{2}\right)}{\Gamma(z)} = \sqrt{z} + \bo\left(\frac{\ds 1}{\ds \sqrt{z}}\right),
\end{equation}
for large $z$. For $n$ odd we obtain
\begin{displaymath}
\omega_n = \frac{2n+3}{n+1} \left[\sqrt{\frac{n+1}{2}} + \bo\left(\frac{1}{\sqrt{n}}\right)\right]
= \sqrt{2n} + \bo\left(\frac{1}{\sqrt{n}}\right).
\end{displaymath}
A similar calculation when $n$ is even gives
\begin{displaymath}
\omega_n = 2\, \frac{\Gamma\left(\frac{n+3}{2}\right)}{\Gamma\left(\frac{n}{2}+1\right)}
= 2\left[\sqrt{\frac{n}{2}+1} + \bo\left(\frac{1}{\sqrt{n}}\right)\right],
\end{displaymath}
proving \eqref{omegaconv}. Next, we observe that
\begin{displaymath}
\sum_{k=n+1}^{2n+1} \frac{1}{\sqrt{k}} = 2(\sqrt{2}-1)\sqrt{n} + \bo\left(\frac{1}{\sqrt{n}}\right)
\end{displaymath}
for large $n$, as can be seen, for example, by noting that
\begin{displaymath}
\dint_{n+1}^{2n+2} \frac{1}{\sqrt{x}}\dx{x} \leq \dsum_{k=n+1}^{2n+1}\frac{1}{\sqrt{k}} \leq
\dint_{n+1}^{2n+2} \frac{1}{\sqrt{x-1}}\dx{x}
\end{displaymath}
and evaluating the integrals. Substituting these two estimates into~\eqref{eulersum} yields
\begin{displaymath}
\begin{aligned}
\chi_n &= \sqrt{2n} - \left(1-\frac{1}{\sqrt{2}}\right)\dsum_{k=1}^{n} \frac{1}{\sqrt{k}} - 2\left(\sqrt{2}-1\right)\sqrt{n} +
\bo\left(\frac{1}{\sqrt{n}}\right)\\
&= \left(1-\frac{1}{\sqrt{2}}\right)a_n + \bo\left(\frac{1}{\sqrt{n}}\right).
\end{aligned}
\end{displaymath}
Letting $s$ equal $1/2$ in~\eqref{zetaexp} and using $-1< \lfloor x\rfloor - x\leq 0$ we obtain
\begin{displaymath}
-\frac{1}{\sqrt{n}}< \zeta(1/2) + a_{n} \leq 0,
\end{displaymath}
from which it follows that
\begin{displaymath}
\chi_n = -Z_{0}(1/2)+\bo\left(\frac{1}{\sqrt{n}}\right),
\end{displaymath}
as desired.
\end{proof}
\begin{lemma}
\label{lemma:hn}
The function $h_{n}$ defined by \eqref{h} is positive and satisfies
\begin{displaymath}
h_{n}(x)\leq \left\{\begin{array}{ll}
\frac{\ds 4^{n+1}}{\ds 2\pi}\frac{\ds 2n+3}{\ds n+1}\Gamma^{2}\left(\frac{\ds n}{\ds 2}+1\right), & n \mbox{ odd,}\eqskip\eqskip
\frac{\ds 4^{n+1}}{\ds 2\pi}(n+1) \Gamma^{2}\left(\frac{\ds n+1}{\ds 2}\right), & n \mbox{ even.}
\end{array}
\right.
\end{displaymath}
\end{lemma}
\begin{proof}
Positivity of $h_{n}$ is a direct consequence of \eqref{sumH}. Taking derivatives in $x$ and using the property $H_{n}'(x) = 2n H_{n-1}(x)$ yields
\begin{displaymath}
\begin{array}{lll}
h_{n}'(x) & = & e^{-x^2}\left\{ -2x\left[H_{n+1}^{2}(x)-H_{n}(x)H_{n+2}(x)\right]\right.\eqskip
& & \hspace*{0.5cm} \left.+ 2H_{n+1}(x)H_{n+1}'(x)-H_{n}'(x)H_{n+2}(x)-H_{n}(x)H_{n+2}'(x)\right\}\eqskip
& = & e^{-x^2}\left\{ 2H_{n+1}(x)\left[-x H_{n+1}(x)+2(n+1)H_{n}(x)\right] \right. \eqskip
& & \hspace*{0.5cm} +2x H_{n}(x)H_{n+2}(x)-2nH_{n-1}(x) H_{n+2}(x) \eqskip
& & \hspace*{1cm} \left.-2(n+2)H_{n}(x)H_{n+1}(x)\right\}\eqskip
& = & e^{-x^2}\left\{ 2H_{n+1}(x)\left[-x H_{n+1}(x)+n H_{n}(x)\right]\right.\eqskip
& & \hspace*{0.5cm} \left. +2 H_{n+2}\left[x H_{n}(x)-nH_{n-1}(x)\right]\right\}.
\end{array}
\end{displaymath}
Using the identity $H_{n+1}(x) = 2x H_{n}(x)-2nH_{n-1}(x)$ in the above expression yields
\begin{displaymath}
h_{n}'(x) = -2e^{-x^2}H_{n}(x)H_{n+1}(x),
\end{displaymath}
which integrated between zero and $x$ becomes
\begin{displaymath}
\begin{array}{lll}
h_{n}(x)-h_{n}(0) & = & -2 \dint_{0}^{x} e^{-t^2}H_{n}(t)H_{n+1}(t)\dx{t}\eqskip
& = & -\frac{\ds 1}{\ds 2(n+1)}\dint_{0}^{x} e^{-t^2}\frac{\ds d}{\ds dt} H_{n+1}^{2}(t)\dx{t}\eqskip
& = & -\frac{\ds 1}{\ds 2(n+1)} \left[ e^{-x^2}H_{n+1}^{2}(x)-H_{n+1}^{2}(0)+ 2\dint_{0}^{x}
t e^{-t^2}H_{n+1}^{2}(t)\dx{t}\right].
\end{array}
\end{displaymath}
Noting that the terms which depend on $x$ on the right-hand side above are non-positive, we obtain
\begin{equation}\label{hnbound}
h_{n}(x)-h_{n}(0) \leq \frac{\ds 1}{\ds 2(n+1)}H_{n+1}^{2}(0).
\end{equation}
For odd $n$, $h_{n}(0) = H_{n+1}^{2}(0)$ and the above becomes
\begin{displaymath}
\begin{array}{lll}
h_{n}(x) & \leq & \frac{\ds 2n+3}{\ds 2n+2} H_{n+1}^{2}(0)\eqskip
& = & \frac{\ds 2n+3}{\ds 2n+2} \Gamma^{2}(n+2)\Gamma^{-2}\left(\frac{\ds n+3}{\ds 2}\right)\eqskip
& = & \frac{\ds 4^{n+1}}{\ds 2\pi}\frac{\ds 2n+3}{\ds n+1}\Gamma^{2}\left(\frac{\ds n}{\ds 2}+1\right).
\end{array}
\end{displaymath}
For even values of $n$ the right-hand side of \eqref{hnbound} vanishes and we obtain
\begin{displaymath}
h_{n}(x) \leq h_{n}(0) = \frac{\ds 4^{n+1}}{\ds 2\pi}(n+1)\Gamma^{2}\left(\frac{\ds n+1}{\ds 2}\right).
\end{displaymath}
\end{proof}
We will now construct an example showing that no lower bound of the same form as in Theorem~\ref{mainth1} is possible.
\begin{proposition}
\label{prop:lower}
For any $n \geq 0$ and any $N>0$, there exists $0 \leq q \in L^1(\R)$ such that
\begin{displaymath}
\dsum_{k=0}^{n}\left[ \lambda_{k}-\lambda_{k}^{0}-\frac{\ds 1} {\ds \pi \sqrt{\lambda_{k}^{0}}}\dint_{\R}
q(x) \dx{x}\right]\leq -N.
\end{displaymath}
\end{proposition}
Before giving the proof, we note two points: firstly, that there \emph{exists} a potential for which the corresponding first $n$ eigenvalues are arbitrarily large negative is trivial; the key point here is that $q$ satisfies the same assumptions as in Theorem~\ref{mainth1}. Secondly, the sum here has to be regularized, since for any $q \geq 0$ we automatically have $\lambda_k \geq \lambda_k^0$ for all $k \geq 0$.
\begin{proof}
Fix $n \geq 0$ and $N>0$. If we use the $n+1$ functions $\psi_k(x)=e^{-x^2/2}H_k(x)$ for $k=1,3,\ldots,2n+1$, as test functions in the Rayleigh quotient, then for any $0 \leq q \in L^1(\R)$ we obtain after a certain amount of rearranging
\begin{multline}
\label{neg-test}
\dsum_{k=0}^{n}\left[ \lambda_{k}-\lambda_{k}^{0}-\frac{\ds 1} {\ds \pi \sqrt{\lambda_{k}^{0}}}\dint_{\R}
q(x) \dx{x}\right] \\ \leq C_n + \dsum_{k=0}^{n}\dint_{\R} e^{-x^2}q(x) \frac{H_{2k+1}^2(x)}{2^{k+1}(2k+1)!
\sqrt{\pi}}\dx{x}- \dsum_{k=0}^{n}\frac{\ds 1} {\ds \pi \sqrt{\lambda_{k}^{0}}}\dint_{\R} q(x) \dx{x},
\end{multline}
where the constant
\begin{displaymath}
C_n := \dsum_{k=0}^{n} \lambda_{2k+1}^0 - \dsum_{k=0}^{n} \lambda_k^0 \geq 0
\end{displaymath}
depends only on $n \geq 0$. We will show that we can find $q$ for which the first sum on the right-hand side of \eqref{neg-test} is arbitrarily small, while the second sum is arbitrarily large. The idea is to choose $q$ to have support in a very small neighbourhood of $0$ and use that all odd Hermite polynomials $H_{2k+1}$ satisfy $H_{2k+1}(0)=0$ (and hence are very small close to $0$). We start by fixing $K=K(n,N)>0$ large enough that
\begin{equation}
\label{k-choice}
C_n + 1 - K\dsum_{k=0}^{n}\frac{\ds 1} {\ds \pi \sqrt{\lambda_{k}^{0}}} < -N
\end{equation}
and for given $\delta>0$, to be specified later, we choose $q_\delta(x):=K\delta^{-1}\chi_\delta(x)$, where $\chi_\delta$ is the indicator function of the set $[-\delta/2,\delta/2]$. Then obviously $q_\delta \geq 0$ has $L^1$-norm equal to $K$ for any $\delta>0$. Since, as mentioned, $H_{2k+1}^2(0)=0$ for all $k=0,\ldots,n$, and $H_{2k+1}^2$ is obviously continuous, for any $\varepsilon>0$, there exists $\delta=\delta(\varepsilon, n) >0$ such that
\begin{displaymath}
0 \leq \frac{e^{-x^2} H_{2k+1}^2(x)}{2^{k+1}(2k+1)!\sqrt{\pi}} < \varepsilon
\end{displaymath}
for all $x \in [-\delta/2,\delta/2]$ and all $k=0,\ldots,n$. It follows that for this $\delta$, we have
\begin{displaymath}
\dsum_{k=0}^{n}\dint_{\R} e^{-x^2}q_\delta (x) \frac{H_{2k+1}^2(x)}{2^{k+1}(2k+1)!
\sqrt{\pi}}\dx{x} < \varepsilon(n+1) < 1,
\end{displaymath}
if we choose $\varepsilon < 1/(n+1)$. Inserting this estimate together with \eqref{k-choice} into \eqref{neg-test} yields the proposition.
\end{proof}
\section{Bounds for the perturbed harmonic oscillator with an integrable perturbation}
\label{sec:neg}
Here we generalize Theorem~\ref{mainth1} to allow for a class of perturbations $q$ which may now take on negative values.
Although the resulting estimate is not quite as tight as in Theorem~\ref{mainth1}, we still have convergence to the trace formula~\eqref{photrace} at the same rate as before.
\begin{theorem}
\label{mainth2}
Given the function $q \in L^1(\R)$, suppose that there exists a non-negative constant $q_m$ for which $q(x) + q_m e^{-x^2}$ is non-negative for almost all real values of $x$. Then the eigenvalues of the corresponding perturbed harmonic oscillator~\eqref{pertharm} satisfy the inequalities
\begin{equation}
\label{negeq}
\dsum_{k=0}^{n}\left[ \lambda_{k}-\lambda_{k}^{0}-\frac{\ds 1}
{\ds \pi \sqrt{\lambda_{k}^{0}}}\dint_{\R} q(x) \dx{x}\right]\leq \frac{\ds \chi_{n}}{\ds \pi}
\dint_{\R} q(x) \dx{x} + \varepsilon_{n} \frac{q_m}{\sqrt{\pi}}
\end{equation}
for $n = 0,1,\ldots$, where
\begin{equation}
\label{negasymp}
\varepsilon_{n} = \omega_n - \sqrt{2}\,\frac{\Gamma\left(n+\frac{3}{2}\right)}{\Gamma(n+1)} \geq 0
\end{equation}
and $\chi_n$ and $\omega_n$ are given by~\eqref{chi} and~\eqref{omega}, respectively. Moreover, $\varepsilon_{n}= \bo(1/\sqrt{n})$ as $n \to \infty$.
\end{theorem}
\begin{proof}
We suppose $q_m \geq 0$ is as in the statement of the theorem, and mimic the proof of Theorem~\ref{mainth1} to obtain
\begin{displaymath}
\begin{split}
\dsum_{k=0}^{n} \lambda_k &\leq \dsum_{k=0}^{n} \lambda_k^0 + \dint_\R e^{-x^2}q(x)\dsum_{k=0}^{n}
\frac{1}{2^k k!\sqrt{\pi}} H_k^2(x)\dx{x}\\
&= \dsum_{k=0}^{n} \lambda_k^0 + \dint_\R e^{-x^2}\left[q(x)+q_m e^{-x^2}\right]
\dsum_{k=0}^{n}\frac{1}{2^k k!\sqrt{\pi}} H_k^2(x)\dx{x}\\
&\qquad\qquad -\frac{q_m}{\sqrt{\pi}}\dsum_{k=0}^{n} \frac{1}{2^k k!} \dint_\R e^{-2x^2}H_k^2(x)\dx{x}.
\end{split}
\end{displaymath}
Since $q(x)+q_m e^{-x^2} \in L^1(\R)$ is positive by assumption, we may proceed as in the proof of Theorem~\ref{mainth1} to obtain
\begin{displaymath}
\begin{split}
\dint_\R e^{-x^2}\left[q(x)+q_m e^{-x^2}\right] \dsum_{k=0}^{n}\frac{1}{2^k k!\sqrt{\pi}} H_k^2(x)\dx{x}
&\leq \frac{\omega_n}{\pi}\dint_\R q(x)+q_m e^{-x^2}\dx{x}\\
&= \frac{\omega_n}{\pi} \dint_\R q(x)\dx{x} + \frac{q_m}{\sqrt{\pi}}\,\omega_n.
\end{split}
\end{displaymath}
Meanwhile, since
\begin{displaymath}
\dint_\R e^{-2x^2} H_k^2(x)\dx{x} = 2^{k-\frac{1}{2}} \Gamma\left(k+\frac{1}{2}\right),
\end{displaymath}
we have
\begin{displaymath}
\dsum_{k=0}^{n}\frac{1}{2^k k!}\dint_\R e^{-2x^2}H_k^2(x)\dx{x}
=\frac{1}{\sqrt{2}}\dsum_{k=0}^{n} \frac{\Gamma\left(k+\frac{1}{2}\right)}{k!}
=\sqrt{2} \,\frac{\Gamma\left(n+\frac{3}{2}\right)}{\Gamma(n+1)}.
\end{displaymath}
Combining the above expressions yields \eqref{negeq}. The asymptotic behaviour of $\varepsilon_{n}$ is an immediate consequence
of \eqref{omegaconv} together with the expansion~\eqref{gammaexp}.
Although $\varepsilon_n$ can be computed explicitly, to see that it is positive we use the following easier, indirect argument:
if for a given $q \in L^1(\R)$, \eqref{negeq} holds for some $q_m \geq 0$, then the above proof shows that it also holds for
all $c \geq q_m$. This is only possible if $\varepsilon_n \geq 0$ for all $n\geq 0$.
\end{proof}
\section{A bound for a general potential in terms of Hermite polynomials}
\label{sec:genpot}
Here we will consider the general problem \eqref{general}, supposing only that the potential $V: \R \to \R$ admits a series expansion in
terms of Hermite polynomials in the manner of an eigenfunction decomposition
\begin{displaymath}
V(x) = \dsum_{j=0}^{\infty} v_j H_j(x),
\end{displaymath}
where we now assume that $V(x) \in L^2(\R,e^{-x^2}\dm{x})$, or equivalently, since the $H_j$ form an orthonomal basis of $L^2(\R)$ with
respect to this measure, that the sequence $v_j$ is square summable. We will prove the following explicit estimate for the
$\lambda_k = \lambda_k(V)$ based on the Fourier-type coefficients $v_j$.
\begin{theorem}
\label{genpot}
Under the above conditions on the potential $V$, for every $n\in\N$, the $n$th eigenvalue of \eqref{general} with $(a,b) = \R$ satisfies
\begin{equation}
\label{gen}
\dsum_{k=0}^n \lambda_k \leq \dsum_{k=0}^n \frac{2^k (2k)!}{k!}\binom{n+1}{k+1}v_{2k} + \frac{1}{2}(n+1)^2.
\end{equation}
\end{theorem}
\begin{remark}
\label{rem:genpot}
(i) This theorem will be proved by using the eigenfunctions of the quantum harmonic oscillator as test functions in the Rayleigh
quotient, as was done in Theorem~\ref{mainth1}. The difference is that there we used an estimate for the sum of Hermite polynomials
resulting from the test functions (Lemma~\ref{lemma:hn}), whereas here we expand out the potential as a Fourier series in Hermite
polynomials and multiply this against our test functions, in the spirit of the arguments used in \cite{frke}.
Since the only inequality we use here is that which results from inserting the test functions into the Rayleigh quotient,
and there is no other estimate involved, it follows that the right-hand side of \eqref{gen} must necessarily be smaller than
the right-hand side of \eqref{comb} if $V$ is of the form $V(x)=x^2+q(x)$ for some $0 \leq q \in L^1(\R)$ (indeed,
it must be equal to the right-hand side of \eqref{basicsum}, i.e.~the middle expression in \eqref{comb}). However,
in practice the two estimates are fundamentally different in nature; for example, it is not easy to see any relation
between the right-hand side of \eqref{gen} and the trace formula \eqref{photrace}. See also Corollary~\ref{gen2} below.
(ii) As a trivial example to show that the above theorem is sharp, if $V(x)=x^2$, then the only two nonzero coefficients in the
Fourier expansion of $V$ are $v_2 = 1/4$, $v_0 = 1/2$, and it can easily be seen that~\eqref{gen} reduces to an equality.
\end{remark}
\begin{proof}
As mentioned, we will use the functions $\psi_k(x):= e^{-x^2/2} H_k(x)$ as test functions in the Rayleigh quotient. In order to
do so, we shall need some more fairly standard facts about integrals of Hermite polynomials $H_k$, which may be found in~\cite{szeg}, for instance: for $n,m \in \N$,
\begin{equation}
\label{double}
\ds \int_{\R}e^{-x^2}H_n(x)H_m(x) \dx{x} = \delta_{mn}\sqrt{\pi}2^n n!
\end{equation}
where $\delta_{jk}$ is the Kronecker delta; and, for $\alpha,\beta,\gamma,s\in\N$ with $\alpha+\beta+\gamma=2s$ even and
$s\geq \alpha,\beta,\gamma$, we have
\begin{equation}
\label{triple}
\ds \int_{\R}e^{-x^2}H_\alpha(x)H_\beta(x)H_\gamma(x) \dx{x} = \sqrt{\pi}\frac{2^s \alpha!\beta!\gamma!}{(s-\alpha)!
(s-\beta)!(s-\gamma)!};
\end{equation}
under any other conditions on $\alpha,\beta,\gamma$ and $s$, this integral is $0$. We also note that, combining a standard integration by parts, \eqref{double} and the formula $H_n'(x) = 2n H_{n-1}(x)$, we obtain easily that
\begin{equation}
\label{quadint}
\begin{split}
\ds \int_{\R}e^{-x^2}x^2 H_k^2(x) \dx{x} &= \frac{1}{2}\ds \int_{\R}e^{-x^2}H_k^2(x)\dx{x}
+2k^2\ds \int_{\R}e^{-x^2}H_{k-1}^2(x)\dx{x}\\
&=\sqrt{\pi}2^{k-1}k!+\sqrt{\pi}2^k k\, k!
\end{split}
\end{equation}
So, using the $\psi_k$ as test functions, as well the convergence of the $v_j$ to interchange integration and summation (noting that
the functions $V(x), H_k^2(x) \in L^2(\R,e^{-x^2}\dx{x})$, the latter being in $\textrm{span}\{H_0(x),H_2(x),\ldots,H_{2k}(x)\}$) together with \eqref{double},
\begin{displaymath}
\begin{split}
&\dsum_{k=0}^{n} \lambda_{k} \leq \dsum_{k=0}^{n} \frac{\ds \dint_{\R}\left[\frac{\ds d}{\ds dx}
\left[e^{-x^2/2}H_{k}(x)\right]\right]^2+
e^{-x^2}V(x) H_{k}^{2}(x)\dx{x}}{\ds \dint_{\R}e^{-x^2}H_{k}^{2}(x)\dx{x}} \\
& = \dsum_{k=0}^{n} \left(\lambda_{k}^{0} - \frac{\dint_{\R}e^{-x^2} x^2 H_{k}^{2}(x)\dx{x}}{\int_{\R}e^{-x^2}
H_k^2(x)\dx{x}}\right)
+ \dsum_{k=0}^{n}\dsum_{j=0}^{\infty}\frac{v_j \dint_{\R} e^{-x^2} (x) H_{j}(x) H_{k}^{2}(x)\dx{x}}{2^k k!\sqrt{\pi}}.
\end{split}
\end{displaymath}
Using \eqref{double} and \eqref{quadint},
\begin{displaymath}
\frac{\dint_{\R}e^{-x^2} x^2 H_{k}^{2}(x)\dx{x}}{\int_{\R}e^{-x^2}
H_k^2(x)\dx{x}} = k+\frac{1}{2},
\end{displaymath}
while \eqref{triple} with $\alpha=\beta=k$ and $\gamma=j$ implies $\int_{\R} e^{-x^2} (x) H_{j}(x) H_{k}^{2}(x)\dx{x} \neq 0$ if and only if $j$ is even and $j \leq 2k$, and under these conditions, writing $j =: 2m$ for $m=0,\ldots,k$,
\begin{displaymath}
\dint_{\R} e^{-x^2} (x) H_{2m}(x) H_{k}^{2}(x)\dx{x} = \sqrt{\pi}\frac{2^{k+m}(k!)^2(2m)!}{(m!)^2(k-m)!}
= \sqrt{\pi}\frac{2^{k+m}k!(2m)!}{m!}\binom{k}{m}.
\end{displaymath}
Combining the above yields
\begin{displaymath}
\begin{array}{lll}
\dsum_{k=0}^{n} \lambda_{k} & \leq & \dsum_{k=0}^{n} \left(\lambda_k^0-k-\frac{1}{2}\right)+\dsum_{k=0}^n
\dsum_{m=0}^k \frac{2^m (2m)!}{m!}\binom{k}{m} v_{2m}.
\end{array}
\end{displaymath}
To simplify this last sum, since $\binom{a}{b}=0$ for $b>a$, we may just as well sum $m$ from $0$ to $n$, giving the sum as
\begin{displaymath}
\dsum_{m=0}^n \frac{2^m(2m)!}{m!}v_{2m}\left(\sum_{k=0}^n \binom{k}{m}\right) = \dsum_{m=0}^n
\frac{2^m (2m)!}{m!}\binom{n+1}{m+1} v_{2m},
\end{displaymath}
using a standard formula for binomial coefficients. This establishes the theorem.
\end{proof}
We shall now assume explicitly that the potential $V$ is a perturbation of the harmonic potential and thus return to writing it as $V(x)=x^2+ q(x)$, where we will assume that $q$ is integrable. By adding the terms which are missing in the right-hand side of~\eqref{gen} in order to obtain a sequence which converges to the right-hand side of the trace formula~\eqref{photrace}, and expressing the coefficients in the left-hand side in terms of the Fourier coefficients of the function $q$, we obtain the following result.
\begin{corollary}
\label{gen2}
\begin{equation}
\label{gen2eq}
\begin{array}{ll}
\dsum_{k=0}^n \left[ \lambda_{k}-\lambda_{k}^{0}-\frac{\ds 1}
{\ds \pi \sqrt{\lambda_{k}^{0}}}\dint_{\R} q(x) \dx{x}\right] & \eqskip
& \hspace*{-3cm}\leq
\dsum_{k=0}^n \left[\frac{2^k (2k)!}{k!}\binom{n+1}{k+1}q_{2k}-\frac{\ds 1}
{\ds \pi \sqrt{\lambda_{k}^{0}}}\dint_{\R} q(x) \dx{x}\right].
\end{array}
\end{equation}
\end{corollary}
\begin{proof}From $V(x) = q(x) + x^2$ we obtain the relations
\[
\begin{array}{lll}
v_{j} & = & \left\{
\begin{array}{ll}
q_{0} +\frac{1}{2}, & j=0\eqskip
q_{2} +\frac{1}{4}, & j=2\eqskip
q_{k}, & j\neq 0,2.
\end{array}\right.
\end{array}.
\]
Replacing this in~\eqref{gen} and adding and subtracting the term
\[
-\lambda_{k}^{0}-\frac{\ds 1} {\ds \pi \sqrt{\lambda_{k}^{0}}}\dint_{\R} q(x) \dx{x}
\]
inside the summation on the left-hand side of~\eqref{gen}, we obtain, after some manipulations, the desired result.
\end{proof}
\begin{remark}
Clearly the integral term appearing inside both sums can be cancelled. However, in this way not only do we obtain
an expression where the left-hand side converges in the limit as $n$ goes to infinity (under additional assumptions
on $q$ as in \cite{abp,puso}), but since as noted in Remark~\ref{genpot}(i) the right-hand side of \eqref{gen2eq}
is necessarily smaller than the right-hand side of \eqref{maineq} (or \eqref{negeq}, depending on $q$), it follows
that it must converge to the right-hand side of the trace formula~\eqref{photrace} and at least as fast as
$\bo(1/\sqrt{n})$.
\end{remark}
\section{Power generalizations of Theorems~\ref{mainth1} and \ref{mainth2}}
\label{sec:power}
In this section we generalize the summation bounds obtained in Theorems~\ref{mainth1} and \ref{mainth2} to allow for the
summands (arranged in various ways) to be raised to a given negative power. We keep the notation and assumptions of
Sections~\ref{sec:pho} and \ref{sec:neg}, and start with the case where the perturbation $q$ is non-negative.
\begin{theorem}
\label{power1}
Under the assumptions and notation of Theorem~\ref{mainth1}, with $\omega_n$ as in \eqref{omega}, for all $n \geq 0$ and $s>0$,
\begin{equation}
\label{11}
\left(\frac{1}{n+1}\right)\dsum_{k=0}^n \left(\lambda_k - \lambda_k^0 \right)^{-s}
\geq \left[\frac{\omega_n}{(n+1)\pi}\dint_\R q(x)\dx{x}\right]^{-s}.
\end{equation}
\end{theorem}
Under certain additional assumptions on the potential, we can rearrange the order of the terms in the above bounds somewhat.
\begin{proposition}
\label{power1a}
If $\int_\R q(x)\dx{x} < 32\sqrt{\pi}$, then for all $n\geq 0$ and $s>0$,
\begin{equation}
\label{1a1}
\dsum_{k=0}^n {\lambda_k}^{-s} \geq \dsum_{k=0}^n \left[\lambda_k^0+\frac{\omega_k-\omega_{k-1}}{\pi}
\dint_\R q(x)\dx{x}\right]^{-s}.
\end{equation}
\end{proposition}
We next consider the situation covered by Theorem~\ref{mainth2}, where the perturbation $q$ may take on negative values, provided its negative part decays rapidly enough at infinity. For simplicity, we consider the special case where $q$ has zero mean.
\begin{theorem}
\label{powerzeromean}
Suppose in addition to the assumptions of Theorem~\ref{mainth2} that $\int_{\R} q(x)\dx{x} = 0$. Then for all $n \geq 0$ and $s>0$,
\begin{equation}
\label{zero1}
\dsum_{k=0}^{n} \lambda_k^{-s} \geq (s+1) \dsum_{k=0}^{n} (\lambda_k^0)^{-s} - sq_m\dsum_{k=0}^{n}
(\lambda_k^0)^{-s-1} (\varepsilon_k - \varepsilon_{k-1}),
\end{equation}
where $q_m \geq 0$ is defined in Theorem~\ref{mainth2}. Here $\varepsilon_n \geq 0$ is given by \eqref{negasymp} for $n \geq 0$ and we set $\varepsilon_{-1}:=0$.
\end{theorem}
These results will be proved by combining generic results on arbitrary increasing or decreasing sequences of real numbers
(see Lemma~\ref{negs} and what follows it) with the following particular properties of the $\omega_n$.
\begin{lemma}
\label{phoseqs}
The sequence $\{\omega_n\}_{n\in\N}$ is positive and strictly increasing, while $\{\tau_n\}_{n\in\N}$ given by $\tau_n:=\omega_{n+1}-\omega_n$ is positive and non-increasing.
\end{lemma}
\begin{proof}
The $\omega_n$ are obviously all positive. Using the formulae
\begin{displaymath}
\Gamma\left(\frac{z+1}{2}\right) = \frac{z! \sqrt{\pi}}{2^z \left(\frac{z}{2}\right)!}, \qquad \Gamma\left(\frac{z}{2}+1\right)
=\left(\frac{z}{2}\right)!
\end{displaymath}
for $z \in \N$ even, if we assume $n\geq 0$ is even and set
\begin{displaymath}
C_n:=(n+1)\frac{\Gamma\left(\frac{n+1}{2}\right)}{\Gamma\left(\frac{n}{2}+1\right)}= \frac{(n+1)!
\sqrt{\pi}}{2^n \left[\left(\frac{n}{2}\right)!\right]^2} > 0,
\end{displaymath}
then an elementary calculation shows that
\begin{displaymath}
\begin{aligned}
\omega_{n+1}-\omega_n &= \frac{C_n}{2(n+2)}\\
\omega_{n+2}-\omega_{n+1} &= \frac{C_n}{2(n+2)}\\
\omega_{n+3}-\omega_{n+2} &= \frac{(n+3)C_n}{2(n+2)(n+4)},
\end{aligned}
\end{displaymath}
from which we see that $\omega_n$ is increasing in $n$, while $\tau_n = \omega_{n+1}-\omega_n$ is positive and weakly decreasing.
\end{proof}
The following lemma appeared in \cite{frke}, but for the sake of completeness we state and prove it here as well. Here and throughout, we will use the notation $[y]_+$, $y\in\R$, to denote the expression taking on the value $y$ if $y\geq 0$ and zero otherwise; $[f(x)]_{g(x)\geq y}$ will represent $f(x)$ if $g(x) \geq y$ and zero otherwise.
\begin{lemma}
\label{negs}
Suppose the sequences $(a_k)_{k\in\N}$ and $(b_k)_{k\in\N}$ are positive, with $(b_k)_{k\in\N}$ non-decreasing in $k\geq 0$. Suppose also that the sequence $(c_k)_{k\in\N}$ satisfies
\begin{equation}
\label{auxbound}
\dsum_{k=0}^m a_k \leq \dsum_{k=0}^m c_k
\end{equation}
for all $m\geq 0$. Then for all $s>0$ and all $n\geq 0$ we have
\begin{equation}
\label{zeta}
\dsum_{k=0}^n (a_k)^{-s} \geq \dsum_{k=0}^n \left[ (s+1)(b_k)^{-s}-s(b_k)^{-s-1}c_k\right].
\end{equation}
If the sequence $(c_k)_{k\in\N}$ is itself positive and non-decreasing in $k\geq 0$, then the right-hand side of \eqref{zeta} is maximized when $b_k=c_k$ for all $0\leq k\leq n$, in which case \eqref{zeta} simplifies to
\begin{displaymath}
\dsum_{k=0}^n (a_k)^{-s} \geq \dsum_{k=0}^n (c_k)^{-s}.
\end{displaymath}
\end{lemma}
An examination of the proof shows that if we want \eqref{zeta} to hold for some fixed $n\geq 0$, then for the proof to work we
need \eqref{auxbound} to hold for all $0 \leq m \leq n$.
\begin{proof}
For $\lambda>0$, we use the identity, valid for all $s>0$,
\begin{equation}
\label{powerrep}
\lambda^{-s} = s(s+1)\dint_0^\infty \alpha^{-s-2}[\alpha-\lambda]_+\dx{\alpha}.
\end{equation}
Hence for $n\geq 0$, $s>0$ arbitrary,
\begin{displaymath}
\begin{split}
\sum_{k=0}^n \left(a_k^{-s}-b_k^{-s}\right) &= s(s+1)\dint_0^\infty \alpha^{-s-2}
\dsum_{k=0}^n \left([\alpha-a_k]_+-[\alpha-b_k]_+\right)\dx{\alpha}\\
&\geq s(s+1)\dint_0^\infty \alpha^{-s-2}\dsum_{k=0}^n [b_k-a_k]_{\alpha\geq b_k}\dx{\alpha}\\
&\geq s(s+1)\dint_0^\infty \alpha^{-s-2}\dsum_{k=0}^n [b_k-c_k]_{\alpha\geq b_k}\dx{\alpha}\\
&=\dsum_{k=0}^n s(s+1)(b_k-c_k)\dint_{a_k}^\infty \alpha^{-s-2}\dx{\alpha},
\end{split}
\end{displaymath}
which after simplification and rearrangement gives us \eqref{zeta}. For the maximizing property we consider each term on the right-hand side of \eqref{zeta} as a function of $b_k$
\begin{displaymath}
g_k(b_k):= (s+1)(b_k)^{-s}-s(b_k)^{-s-1}c_k.
\end{displaymath}
Differentiating in $b_k$ shows that $g_k$ reaches its unique maximum when $b_k=c_k$.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{power1a}]
Lemma~\ref{negs} may be applied directly to prove Proposition~\ref{power1a} in the obvious way; for \eqref{1a1}, it merely remains to be confirmed that the sequence
\begin{displaymath}
\left\{\lambda_k^0 + \frac{\omega_k-\omega_{k-1}}{\pi}\dint_\R q(x)\dx{x}\right\}_{k\in\N}
\end{displaymath}
is positive and non-decreasing. Now since $\lambda_{k+1}^0-\lambda_k^0=2$ for all $k\geq 0$, we need $\int_\R q(x)\dx{x}$ (which we assume to be nonzero and hence strictly positive) to be small enough that
\begin{displaymath}
\omega_{k+2}-2\omega_{k+1}+\omega_k \geq -\frac{2\pi}{\int_\R q(x)\dx{x}}
\end{displaymath}
for all $k\geq 0$. If $k$ is even, then the left-hand side is identically zero, as follows from the proof of Lemma~\ref{phoseqs}.
Otherwise, for $k+1$ odd, we have
\begin{displaymath}
\omega_{k+3}-2\omega_{k+2}+\omega_{k+1} = \frac{(k+3)C_k}{2(k+2)(k+4)} - \frac{C_k}{2(k+2)},
\end{displaymath}
which, using the definition of $C_k$, may be rearranged to give
\begin{displaymath}
-\frac{\sqrt{\pi}}{2(k+4)}\cdot\frac{k+1}{k+2}\cdot\frac{k-1}{k-2}\cdots\frac{3}{4}\cdot\frac{1}{2},
\end{displaymath}
which we see is negative and increasing in $k+1\geq 1$ odd. Thus $\omega_{k+3}-2\omega_{k+2}-\omega_{k+1}$ reaches its largest negative value, namely $-C_0/16 = -\sqrt{\pi}/16$, when $k=0$. The requirement on $q(x)$ is therefore that
\begin{displaymath}
-\frac{\sqrt{\pi}}{16} \geq -\frac{2\pi}{\int_\R q(x)\dx{x}},
\end{displaymath}
that is, we have shown the required sequence is increasing when $\int_\R q(x)\dx{x} \leq 32\sqrt{\pi}$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{power1}]
To prove \eqref{11} we use a similar idea to the one in Lemma~\ref{negs}, but since the right-hand side of \eqref{maineq}
is not a sequence, the method needs to be adapted slightly to this situation. Namely, starting with the representation \eqref{powerrep} of $\lambda =: \lambda_k - \lambda_k^0$,
\begin{equation}
\label{11proof}
\sum_{k=0}^n (\lambda_k-\lambda_k^0)^{-s} \geq s(s+1)\int_0^\infty \alpha^{-s-2} \sum_{k=0}^n
[\alpha - \lambda_k + \lambda_k^0]_{\alpha\geq M} \dx{\alpha}
\end{equation}
for all $M \in \R$; we make the choice $M:= \frac{\omega_n}{(n+1)\pi}\int_\R q(x)\dx{x}$. Using \eqref{maineq}, which, when rearranged, says that
\begin{equation}
\label{rearr}
\sum_{k=0}^n (\lambda_k-\lambda_k^0) \leq \frac{\omega_n}{\pi}\int_\R q(x)\dx{x},
\end{equation}
we have
\begin{displaymath}
\sum_{k=0}^n [\alpha - \lambda_k + \lambda_k^0]_{\alpha\geq \frac{\omega_n}{(n+1)\pi}\int_\R q(x)\dx{x}}
\geq (n+1)[\alpha-\frac{\omega_n}{(n+1)\pi}\int_\R q(x)\dx{x}]_+.
\end{displaymath}
Substituting this into \eqref{11proof} and applying \eqref{powerrep} yields \eqref{11}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{powerzeromean}]
This follows directly from Theorem~\ref{mainth2} and Lemma~\ref{negs}, where we take $a_k = \lambda_k$, $b_k = \lambda_k^0$ and $c_k = \lambda_k^0 + (\varepsilon_k - \varepsilon_{k-1})q_m$ (with $\varepsilon_{-1}:=0$).
\end{proof}
|
\section{Introduction}
Let $G$ be a graph with $n\ge 2$ vertices, chromatic number $\chi(G)$, fractional chromatic number $\chi_f(G)$ and independence number $\alpha(G)$. Let $A$ denote the adjacency matrix of $G$ and let $\mu_1 \ge \mu_2 \ge ... \ge \mu_n$ denote the eigenvalues of $A$. The inertia of $A$ is the ordered triple $(n^+, n^0, n^-)$ where $n^+, n^0, n^-$ are the numbers (counting multiplicities) of positive, zero and negative eigenvalues of $A$ respectively. Note that $\mathrm{rank}(A) = n^+ + n^-$ and $\mathrm{nullity}(A) = n^0$.
\section{An inertial lower bound}
\begin{theorem}
Let $G$ be any graph with inertia $(n^+, n^0, n^-)$. Then
\[
1 + \max\left(\frac{n^+}{n^-} , \frac{n^-}{n^+}\right) \le \chi(G).
\]
\end{theorem}
To better organize the proof of this theorem we formulate the following simple lemma. Recall that positive semidefinite matrices are Hermitian matrices with non-negative eigenvalues and $Z^*$ denotes the Hermitian transpose for matrices $Z\in\mathbb{C}^{n\times n}$.
\begin{lemma}\label{semirank}
Let $X,Y\in\mathbb{C}^{n\times n}$ be two positive semidefinite matrices satisfying $X \succeq Y$, that is, their difference $X - Y$ is positive semidefinite. Then
\[
\mathrm{rank}(X) \ge \mathrm{rank}(Y).
\]
\end{lemma}
\begin{proof}
Assume that $\mathrm{rank}(X) < \mathrm{rank}(Y)$. Then, there exists a non-trivial vector $v$ in the range of $Y$ that is orthogonal to the range of $X$. Consequently,
\[
v^*(X - Y)v = -v^* Y v < 0
\]
contradicting that $X - Y$ is positive semidefinite.
\end{proof}
\begin{proof}[Proof of Theorem 1]
Wocjan and Elphick \cite{wocjan13} proved that there exist $\chi$ unitary matrices $U_i$ such that:
\begin{equation}\label{eq:pawel}
\sum_{i=1}^{\chi-1} U_i A U_i^* = -A.
\end{equation}
Let $v_1, ... , v_n$ denote the eigenvectors of unit length corresponding to the eigenvalues $\mu_1 \ge ... \ge \mu_n$. Let $A = B - C$ where
\[
B = \sum_{i=1}^{n^+}\mu_iv_iv_i^*\quad\mathrm{and}\quad
C = \sum_{i=n-n^-+1}^n (-\mu_i)v_iv_i^*.
\]
Note that $B$ and $C$ are positive semidefinite and that $\mathrm{rank}(B) =n^+$ and $\mathrm{rank}(C) = n^-$. Let
\[
P^+ = \sum_{i=1}^{n^+} v_iv_i^* \mbox{ and } P^- = \sum_{i=n-n^-+1}^n v_i v_i^*
\]
denote the orthogonal projectors onto the subspaces spanned by the eigenvectors corresponding to the positive and negative eigenvalues respectively. Observe that $B = P^+AP^+$ and $C = -P^-AP^-$.
Equation (\ref{eq:pawel}) can be rewritten as:
\[
\sum_{i=1}^{\chi-1} U_iBU_i^* - \sum_{i=1}^{\chi-1} U_iCU_i^* = C - B.
\]
Multiplying both sides by $P^-$ from the left and the right we obtain:
\[
P^-\sum_{i=1}^{\chi-1} U_iBU_i^* P^- - P^-\sum_{i=1}^{\chi-1} U_iCU_i^* P^- = C.
\]
Using that
\[
P^-\sum_{i=1}^{\chi-1} U_iCU_i^* P^-
\]
is positive semidefinite, it follows that:
\[
P^-\sum_{i=1}^{\chi-1} U_iBU_i^* P^- \succeq C.
\]
Then using that the rank of a sum is less than or equal to the sum of the ranks of the summands, that the rank of a product is less than or equal to the minimum of the ranks of the factors, and Lemma~\ref{semirank}, we have that ($\chi - 1)n^+ \ge n^-.$ Similarly $(\chi - 1)n^- \ge n^+$ by multiplying equation (\ref{eq:pawel}) by $-1$ and multiplying both sides by $P^+$ from the left and right.
\end{proof}
\begin{corollary}\label{npmcorollary}
Let $G$ be any graph with inertia $(n^+, n^0, n^-)$. Then
\[
\max(n^+ , n^-) \le \frac{n(\chi - 1)}{\chi}.
\]
\end{corollary}
\begin{proof}
\[
\chi(G) \ge 1 + \frac{n^+}{n^-} = \frac{n^- + n^+}{n^-} = \frac{n - n^0}{n - n^+ - n^0} \ge \frac{n}{n - n^+}.
\]
Re-arranging and repeating with $\chi \ge 1 + n^-/n^+$ completes the proof.
\end{proof}
Griffith and Luhanga \cite{griffith11} conjectured that for connected, planar graphs $n^- \le 2n/3$. $K_4$ provides a counter-example, but it follows from this corollary that for planar graphs, $n^- \le 3n/4$.
\begin{corollary}
Let $\xi'(G)$ denote the normalized orthogonal rank of a graph. Then
\[
1 + \max\left(\frac{n^+}{n^-} , \frac{n^-}{n^+}\right) \le \xi'(G) \le \chi(G).
\]
\end{corollary}
\begin{proof}
Wocjan and Elphick \cite{wocjan13} proved that $\chi(G)$ can be replaced with $\xi'(G)$ in equality (\ref{eq:pawel}).
\end{proof}
\begin{corollary}
Let $A=(a_{k\ell})$ denote the adjacency matrix of a graph and let $W=(w_{k\ell})$ denote an arbitrary Hermitian matrix such that $a_{k\ell}=0$ implies $w_{k\ell}=0$ for all $k,\ell$.
\[
1 + \max\left(\frac{n^+(W)}{n^-(W)} , \frac{n^-(W)}{n^+(W)}\right) \le \chi(G).
\]
\end{corollary}
\begin{proof}
Wocjan and Elphick \cite{wocjan13} proved that $A$ can be replaced with $W$ in equality (\ref{eq:pawel}).
\end{proof}
\subsection{Extremal graphs}
We can compare this lower bound against the well known bound due to Hoffman \cite{hoffman70} that $1 + \mu_1/|\mu_n| \le \chi(G)$, and other bounds due to Nikiforov \cite{nikiforov07}, Kolotilina \cite{kolotilina10}, Wocjan and Elphick \cite{wocjan13}, Ando and Lin \cite{ando15} and Elphick and Wocjan \cite{elphick15}.
Theorem~1 is exact for example for bipartite graphs ($n^+ = n^-$), regular complete $q$-partite graphs ($n^+ = 1 , n^- = q - 1$) and $24$-cell $(n^+ = 5, n^- = 10)$, for which some of these other lower bounds are also exact. However Theorem 1 is also exact for example for the following singular and non-singular graphs, for which none of these other bounds is exact:
\begin{itemize}
\item barbell graphs on 2n vertices ($n^+ = 2 , n^- = 2n - 2$)
\item irregular complete $q$-partite graphs ($n^+ = 1, n^- = q - 1$)
\item various Circulant graphs such as Antiprism(9) ($n^+ = 5 , n^- = 10$)
\item Sextic(16,1) ($n^+ = 3 , n^- = 9$).
\end{itemize}
The full version of the Hoffman bound \cite{hoffman70} is that $\mu_1 + \mu_{n-\chi+2} + \ldots +\mu_n \le 0$, which is exact for all the itemised graphs.
\section{A conjectured stronger bound}
\begin{conjecture}\label{conjecture}
Let $G$ be any graph with inertia $(n^+, n^0, n^-)$. Then
\[
1 + \max\left(\frac{n^+}{n^-} , \frac{n^-}{n^+}\right) \le \chi_f(G).
\]
\end{conjecture}
We can prove this conjecture for non-singular graphs as follows.
\begin{proof}
A non-singular graph has $n^0 = 0$. Cvetkovi\'c \emph{et al.}~\cite{cvetkovic79} (page 88) proved that:
\[
\alpha (G) \le n^0 + \min(n^+ , n^-) = \min(n^+ , n^-).
\]
It is well known that $\chi_f(G) \ge n/\alpha(G)$, with equality for vertex-transitive graphs. Therefore
\[
\chi_f(G) \ge \frac{n}{\alpha} \ge \frac{n}{\min(n^+ , n^-)} = \frac{n^+ + n^-}{\min(n^+ , n^-)} = 1 + \max\left(\frac{n^+}{n^-} , \frac{n^-}{n^+}\right).
\]
\end{proof}
Conjecture~\ref{conjecture} is therefore true, for example, for all strongly regular graphs. Using standard notation, if $G$ is a strongly regular graph with spectrum $(k^1 , r^f , s^g)$ then:
\[
\chi_f \ge 1 + \max\left(\frac{n^+}{n^-} , \frac{n^-}{n^+}\right) = 1 + \max\left(\frac{1+f}{g} , \frac{g}{1+f}\right) = \max\left(\frac{n}{g} , \frac{n}{1+f}\right).
\]
\subsection{Extremal graphs}
Conjecture~\ref{conjecture} is exact for example for the Kneser graphs $K_{p,k}$, whose vertices correspond to the $k$-element subset of a set of $p$ elements, and where two vertices are joined if and only if the corresponding sets are disjoint. It is known that $\chi(K_{p,k}) = p - 2k + 2$ and that $\chi_f(K_{p,k}) = p/k$. The inertia of these graphs (see Godsil and Meagher \cite{godsil15} section 2.10) is as follows:
\[
n^+ = {p-1 \choose k} \mbox{ ; } n^0 = 0 \mbox{ ; } n^- = {p-1 \choose k-1}.
\]
Consequently
\[
1 + \max\left(\frac{n^+}{n^-} , \frac{n^-}{n^+}\right) = 1 + \frac{n^+}{n^-} = 1 + \frac{p - k}{k} = \frac{p}{k} = \chi_f.
\]
Conjecture~\ref{conjecture} is also exact for cycles. This is obvious for even cycles. For the odd cycle on $2n+1$ vertices:
\[
1 + \max\left(\frac{n^+}{n^-} , \frac{n^-}{n^+}\right) = 1 + \frac{n+1}{n} = 2 + \frac{1}{n} = \chi_f.
\]
\section{Other graph parameters and an example}
There are many graph parameters that lie between the clique number, $\omega(G)$, and the chromatic number. For example
\[
\omega(G) \le \chi_v(G) \le \theta(\overline{G}) \le \chi_f(G) \le \chi_c(G) \le \lceil \chi_c(G) \rceil = \chi(G),
\]
where $\chi_v(G)$ is the vector chromatic number, $\theta(\overline{G})$ is the Lov\'asz theta function of the complement of $G$ and $\chi_c(G)$ is the circular chromatic number. (These inequalities are sufficiently well known to be included in the Wikipedia entry on graph coloring and in \cite{brown09}.)
Bilu \cite{bilu06} proved that the Hoffman bound is a lower bound for the vector chromatic number
\[
1 + \frac{\mu_1}{|\mu_n|} \le \chi_v(G).
\]
We can demonstrate, however, using the self-complementary pentagon $(C_5)$, that
\[
1 + \max\left(\frac{n^+}{n^-} , \frac{n^-}{n^+}\right) \not\le \chi_v(G).
\]
The spectrum of $C_5$ is $\left(2, (\frac{-1+\sqrt{5}}{2})^2, (\frac{-1-\sqrt{5}}{2})^2\right)$ and Lov\'asz proved that $\theta(C_5) = \sqrt{5}$. Therefore
\begin{multline*}
\sqrt{5} = 1 + \frac{\mu_1}{|\mu_n|} = \chi_v(C_5) = \theta(\overline{C_5}) = \theta(C_5) = \sqrt{5} < \\ 2.5 = 1 + \frac{n^+}{n^-} = \chi_f(C_5) < \chi(C_5) = 3.
\end{multline*}
It is instructive to compare the various lower bounds for $\chi(G)$. Because $C_5$ is regular, the bounds due to Hoffman, Nikiforov and Kolotilina are all equal to each other at $\sqrt{5} = 2.236$. The generalisations of these bounds due to Wocjan and Elphick are also all equal to $\sqrt{5}$, as is a bound due to Lima \emph{et al.} \cite{lima11}. The bound due to Ando and Lin equals $2.1$. The bound in this paper equals $2.5$, because it uses the (integral) inertia of a graph instead of its eigenvalues. However, since $\chi$ is integral, all of these bounds imply $\chi \ge 3$.
\section{On the asymmetry between $n^+$ and $n^-$}
The inertial bound discussed above suggests symmetry between $n^+$ and $n^-$. In this section we explore asymmetry between $n^+$ and $n^-$, beginning with Nordhaus-Gaddum bounds for inertia.
In 1956, Nordhaus and Gaddum \cite{nordhaus56} proved that:
\[
2\sqrt{n} \le \chi(G) + \chi(\overline{G}) \le n + 1 \mbox{ and } n \le \chi(G) \cdot \chi(\overline{G}) \le \frac{(n + 1)^2}{4}.
\]
Similar bounds, now known as Nordhaus-Gaddum type inequalities, have been obtained for numerous graph parameters. For example the survey paper by Aouchiche and Hansen \cite{aouchiche12} reviews scores of Nordhaus-Gaddum type inequalities. The survey does not, however, refer to inertia.
\begin{theorem}
For any graph $G$,
\begin{align*}
1 &\le n^+(G) + n^+(\overline{G}) \le n + 1 \\
0 &\le n^0(G) + n^0(\overline{G}) \le n \\
n-1 &\le n^-(G) + n^-(\overline{G}).
\end{align*}
\end{theorem}
\begin{proof}
The lower bound for $n^+$ is trivial and is exact for complete graphs. The lower bound for $n^0$ is exact for most graphs since "almost all" graphs have a non-integral spectrum. There does not seem to be a straightforward upper bound for $n^-$. For example the Generalised Petersen (15,4) graph on 30 vertices has $n^-(G) + n^-(\overline{G}) = 37$.
The spectrum of $G$ begins:
\begin{multline*}
\mu_1 \ge \ldots \ge \mu_{n^+} > 0 = \mu_{n^+ + 1} = \ldots = \\
\mu_{n^+ + n^0} = 0 > \mu_{n^+ + n^0 + 1} \ge \ldots \ge \mu_{n^+ + n^0 + j} > -1
\end{multline*}
and continues
\[
-1 \ge \mu_{n^+ + n^0 + j + 1} \ge \ldots \ge \mu_n.
\]
One of the Courant-Weyl inequalities (see for example section 2.8 in \cite{brouwer10}) states that for Hermitian matrices $A$ and $B$
\[
\mbox{ if } i + j - n \ge 1 \mbox{ then } \mu_i(A) + \mu_j(B) \le \mu_{i+j-n}(A + B).
\]
Therefore
\[
\mu_i(G) + \mu_{n-i+2}(\overline{G}) \le \mu_2(K_n) = - 1, \mbox{ for } i \ge 2.
\]
Therefore, for $i \ge 2$, if $\mu_i > -1$ then $\mu_{n-i+2}(\overline{G}) \le -1 - \mu_i(G) < 0$. So
\begin{equation}\label{clive}
n^-(\overline{G}) \ge n^+(G) + n^0(G) + j -1.
\end{equation}
The $-1$ term in (\ref{clive}) follows from the constraint that $i \ge 2$. Consequently
\[
n^-(G) + n^-(\overline{G}) \ge n^-(G) + n^+(G) + n^0(G) + j - 1 \ge n - 1.
\]
Similarly, using (\ref{clive}) again
\begin{eqnarray*}
n^+(G) + n^+(\overline{G}) & = & n^+(G) + n - n^-(\overline{G}) - n^0(\overline{G}) \\
& \le & n^+(G) +n -(n^+(G) + n^0(G) + j -1) - n^0(\overline{G}) \\
& \le & n + 1.
\end{eqnarray*}
Finally, using the bounds above
\begin{eqnarray*}
n^0(G) + n^0(\overline{G}) & = & 2n - (n^+(G) + n^+(\overline{G})) -(n^-(G) + n^-(\overline{G})) \\
& \le & 2n - 1 - (n - 1) = n.
\end{eqnarray*}
\end{proof}
We can show that the upper bound for $n^+(G) + n^+(\overline{G})$ and the lower bound for $n^-(G) + n^-(\overline{G})$ are exact for strongly regular graphs as follows.
\begin{corollary}
Let $G$ be a strongly regular graph. Then $n^+(G) + n^+(\overline{G}) = n + 1$ and $n^-(G) + n^-(\overline{G}) = n - 1$.
\end{corollary}
\begin{proof}
Using standard notation let $G = SRG(n, k, \lambda, \mu)$ have spectrum $(k^1 , r^f , s^g)$, where $1$, $f$ and $g$ are multiplicities. It is well known that $\overline{G} = SRG(n, n-k-1 , n-2k+\mu-2, n-2k+\lambda)$. Clearly $1 + f(G) + g(G) = 1 + f(\overline{G}) + g(\overline{G}) = n$. It is well known that:
\begin{eqnarray*}
f & = & \frac{1}{2}\left((n - 1) - \frac{2k + (n - 1)(\lambda - \mu)}{\sqrt{(\lambda - \mu)^2 +4(k - \mu)}}\right) \\
g & = & \frac{1}{2}\left((n - 1) + \frac{2k + (n - 1)(\lambda - \mu)}{\sqrt{(\lambda - \mu)^2 +4(k - \mu)}}\right)
\end{eqnarray*}
A page of algebra demonstrates that $f(G) = g(\overline{G})$ and $g(G) = f(\overline{G})$. Therefore
\[
n^+(G) + n^+(\overline{G}) = 2 + f(G) +f(\overline{G}) = 2 + f(G) + g(G) = n + 1.
\]
It is then immediate that $n^-(G) + n^-(\overline{G}) = n - 1$
\end{proof}
It is clear that $n^- = n - 1$ for $K_n$. However $n^+ = n - 1$ only for $K_2$, because $\mu_1 \ge |\mu_n|$. The goal is therefore to find an upper bound for $n^+$ as a function of $n$, which is not an upper bound for $n^-$. Since the Nordhaus-Gaddum upper bound for $n^+$ is exact for SRGs, it is plausible that some SRGs will be extremal.
SRGs, other than the pentagon, have $\mu_2 \ge 1$ so it seems likely that SRGs that maximise $n^+$ have $\mu_2 = 1$. A couple of pages of simple algebra demonstrates that when $\mu_2 = 1$ then
\[
n^+ = n - \left(\frac{n - \lambda + 2\mu}{2 - \lambda + \mu}\right) \mbox{ and } k = 1 -\lambda + 2\mu.
\]
$n^+$ is maximised when $\lambda = 0$, in which case
\[
n^+ = n - \left(\frac{n + 2\mu}{2 + \mu}\right) \mbox{ and } k = 1 + 2\mu.
\]
We can also use the well known identity for all SRGs that
\[
n = 1 + k + \frac{k(k - 1 - \lambda)}{\mu}.
\]
This implies that
\[
n^+ = n - \left(\frac{8(n - 1)}{8 + n}\right),
\]
which suggests the following conjecture.
\begin{conjecture}\label{conjecture2}
For any graph
\[
n^+ \le n - \left\lfloor\frac{8(n - 1)}{8 + n}\right\rfloor.
\]
\end{conjecture}
This bound is exact for connected graphs such as the Petersen graph and SRG(16, 5, 0 2), and for disconnected graphs such as $2C_5$, without use of the floor function. We have tested this bound against all named graphs with up to 40 vertices in the Wolfram Mathematica database and found no counterexample. Conjecture~\ref{conjecture2} outperforms Corollary~\ref{npmcorollary} for some graphs, such as the Petersen graph.
For large $n$, this bound has $n^+ \approx n$. Such graphs do exist. For example the Taylor family of graphs are $SRG(q^3, \frac{1}{2}(q - 1)(q^2 +1), \frac{1}{4}(q - 1)^3 - 1, \frac{1}{4}(q - 1)(q^2 + 1))$, where $q$ is an odd prime. Nikiforov \cite{nikiforov15} discusses the inertia of these graphs and notes that for sufficiently large $n$, almost all of the eigenvalues are positive, since
\[
n^+ = 1 + (q - 1)(q^2 + 1) = q^3 + q - q^2 = n + n^{\frac{1}{3}} - n^{\frac{2}{3}} \approx n.
\]
\section{Conclusions}
It is worth noting that equation (\ref{eq:pawel}) can be used to prove (generalisations) of the bounds due to Hoffman, Nikiforov and Kolotilina and the bound due to Lima \emph{et al} (see \cite{wocjan13} and \cite{elphick15}) as well as this inertial bound, but we do not know how to prove the Ando and Lin bound using (\ref{eq:pawel}).
In 1976 van Nuffelen conjectured that $\chi(G) \le \mathrm{rank}(A)$, but a counter-example was found on 64 vertices by Alon and Seymour in 1989, which has rank 29 and chromatic number of 32. The spectrum of the counter-example is $56^1 , 4^7 , 0^{35} , -4^{21}$, so this graph provides an example of the Hoffman bound significantly outperforming Theorem 1.
The bound in this paper is the first spectral bound for the chromatic number of which we are aware, which uses only the numbers of positive and negative eigenvalues of a graph.
|
\section{Introduction}
There is accumulating evidence that quantum gravity may be
non-perturbatively renormalizable due to the existence of an
ultraviolet fixed point. This scenario of asymptotic safety
\cite{weinberg1979ultraviolet} has been found in four-dimensional
renormalization group investigations \cite{Reuter:1996cp}, based on
functional renormalization for the effective average action
\cite{Wetterich:1992yh,Reuter:1993kw}. Many extensions of the truncation
beyond the simplest Einstein-Hilbert truncation have confirmed the
presence of the ultraviolet fixed point
\cite{Dou:1997fg,Souma:1999at,Reuter:2001ag,Litim:2003vp,Codello:2006in,Machado:2007ea,Codello:2008vh,Fischer:2006fz,Benedetti:2009rx,Eichhorn:2010tb,Manrique:2010am,Donkin:2012ud,Christiansen:2012rx,Rechenberger:2012dt,Dietz:2012ic,Codello:2013fpa,Falls:2013bv,Benedetti:2013jk,Christiansen:2014raa,Christiansen:2015rva,Dietz:2015owa,Demmel:2015oqa,Falls:2015qga,Gies:2015tca,Gies:2016con},
for reviews see
\cite{Niedermaier:2006wt,Percacci:2007sz,Litim:2011cp,%
Reuter:2012id,Nagy:2012ef}.
Similar ideas have been explored in other quantum field theories, see,
e.g., \cite{Gies:2009hq,Braun:2010tt,Litim:2014uca}. Interactions
with matter have been included
\cite{Dou:1997fg,Folkerts:2011jz,Harst:2011zx,Dona:2013qba,Oda:2015sma,Meibohm:2015twa,Eichhorn:2016esv},
with emphasis on scalar matter in
refs. \cite{Narain:2009fy,Narain:2009gb,Percacci:2015wwa,Dona:2015tnf,Labus:2015ska}. Additional
arguments in favor of asymptotic safety arise from different
approaches to quantum gravity \cite{Hamber:2009mt, Ambjorn:2009ts}.
In the presence of an ultraviolet fixed point the flow of couplings
can be extended to the limit where the renormalization scale $k$ goes
to infinity. Within functional renormalization, observable quantities
are obtained in the opposite limit $k\to 0$. Furthermore, the gravitational
quantum field equations relevant for cosmology arise for
$k\to 0$. It is therefore important to find smooth trajectories from
the ultraviolet to the infrared limit as $k$ decreases to zero
\cite{Donkin:2012ud,Christiansen:2012rx, Nagy:2012rn}.
Gravity coupled to a scalar field offers the interesting perspective
that a realistic scale-symmetric theory of gravity and particle
physics can be formulated if the Planck mass is given by a scalar
field $\chi$ \cite{Fujii:1982ms,Wetterich:1987fm,Shaposhnikov:2008xi}. In
the absence of explicit mass scales the scale $k$ can only be compared
to $\chi$, with IR-limit $k/\chi\to 0$. If cosmological solutions for
$\chi$ approach an infrared fixed point with exact scale symmetry the
\textquotedblleft dilatation anomaly\textquotedblright\ close to the
fixed point can give rise to dynamical dark energy or quintessence
\cite{Wetterich:1987fm}. The crossover between an ultraviolet and
infrared fixed point can connect inflation and present dynamical dark
energy, both described by the same cosmon field
\cite{Wetterich:2014gaa}. Realistic cosmology is found in a picture where
the universe is not expanding during radiation and matter domination
\cite{Wetterich:2013aca} and the big bang singularity is absent
\cite{Wetterich:2014zta}.
In the functional renormalization group approach to quantum gravity
the system of a scalar field coupled to gravity was first studied in
ref. \cite{Narain:2009gb,Narain:2009fy}. The existence of a \textquotedblleft
global scaling solution\textquotedblright\ for all $k$ and $\chi$,
with general scalar potential and scalar-field dependent coefficient
of the curvature scalar, was investigated in
ref. \cite{Henz:2013oxa}. Recent advances towards a global scaling
solution have been made in \cite{Percacci:2015wwa,Borchardt:2015rxa}.
In the present work we derive for the first time candidates for
global scaling solutions in dilaton gravity. These solutions are
obtained from a qualitatively improved approximation to the effective
action in comparison to those used in previous works: Firstly, we include a $\chi$-dependent
coefficient of the scalar kinetic term, called the kinetial. This
closes a systematic derivative expansion in the second order of
derivatives. The second necessary improvement concerns the
computation of dynamical correlation functions on the basis of the
expansion scheme put forward in refs.
\cite{Christiansen:2012rx,Christiansen:2014raa,Christiansen:2015rva},
for gravity-matter systems see refs.
\cite{Meibohm:2015twa,Meibohm:2016mkp,Eichhorn:2016esv}.
This goes beyond the standard background
field approximation. With its relation to the constraints of
diffeomorphism symmetry in a gauge-fixed setting it is at the root of
background independence, for discussions see e.g. ref.
\cite{Litim:2002ce,Folkerts:2011jz,Christiansen:2015rva,Meibohm:2015twa}.
The expansion around a flat background as well as a vertex
construction allow us to compute the running of the kinetial as well as to
disentangle fluctuating and background fields.
\section{Setup}
We aim at finding global fixed point
solutions for the effective action of dilaton gravity
\begin{align}
\label{eq:action}
\Gamma =\int_x \sqrt{g} \left( V(\chi^2)-\frac12 F(
\chi^2)\,R+\frac{1}{2} K( \chi^2) g^{\mu\nu}
\partial_\mu\chi\partial_\nu\chi \right)\,,
\end{align}
where $\int_x = \int d^4 x$.
The three functions $V$, $F$ and $K$ depend on a scalar field $\chi$
and the renormalization scale $k$. For a scaling solution, the dimensionless functions
$V/k^4$, $F/k^2$ and $K$ only depend on the dimensionless ratio
$y=\chi^2/k^2$. For fixed $k=\mu$ the effective action \eq{eq:action}
constitutes a model of variable gravity, for which the cosmology is
discussed in detail in \cite{Wetterich:2013jsa}. In this work we extract
the scale dependence of the functions $V$, $F$ and $K$ from the
functional renormalization group. This translates directly to the
$\chi$-dependence of these functions and therefore to the field
equations relevant for cosmology.
We work with dimensionless functions and fields.
To derive the flow equations, we consider
$\Gamma[\bar g,\bar\chi; g,\chi]$.
Here,
$\bar g$, $\bar\chi$ are background
fields and $h=g-\bar g$, $\delta \chi = \chi - \bar \chi$ are the dynamical fluctuation fields.
While the occurrence of the background
metric is inherent to any gauge-fixed approach to quantum gravity, the
dependence on the dilaton background field is only introduced via the
regulator term, see refs. \cite{Litim:2002hj,Dietz:2015owa}.
The identification $\bar g = g$, $\bar \chi = \chi$ eliminates generalized gauge fixing terms and results in the gauge invariant effective action $\Gamma[g,\chi]$.
We are interested in the scaling solution for $\Gamma[g,\chi]$. The flows of $V$, $F$ and $K$ are extracted from the flow of the two point functions for the fluctuating fields $h$ and $\delta \chi$ \cite{Christiansen:2014raa,Christiansen:2015rva}.
We work in deDonder
gauge and neglect the ghost contributions, as they do
not couple directly to the dilaton field,
as well as some subleading terms in the $k$-dependence of the regulator.
We perform a systematic expansion in powers of $h$ to disentangle contributions from background and fluctuating fields. Accordingly, we compute the flow
equations for $V$ and $F$ via the momentum independent and dependent
part of the flow of the transverse-traceless graviton $2$-point
function, respectively. Moreover, we use the momentum-dependent part
of the scalar $2$-point function for the flow of $K$, expanding around
flat space \cite{Christiansen:2014raa}.
The full equations are too long to be displayed here.
\section{Large field limit}
\begin{figure*}[t]
\includegraphics[width=\columnwidth]{betaVF.pdf}
\includegraphics[width=0.95\columnwidth]{betaK.pdf}
\caption{Flowing couplings. The $\beta$-functions
$v_0,f_0,v_1,f_1,k_1$ are shown as functions of $\epsilon$, with other couplings kept fixed at values given by eq. \eqref{eqn:fpsol}. The simultaneous zero at $\epsilon=\epsilon_0=109.97$, as well as the
pole at $\epsilon=-6$ are clearly visible.}
\label{betafcts}
\end{figure*}
We first consider large $y$, which for finite field $\chi$ is
equivalent to sending the renormalization scale $k\to 0$ and thus
constitutes the infrared limit of dilaton gravity. For fixed $k$ this is the limit $\chi \to \infty$.
One possible IR-limit implies weak gravity with $F\sim y$. For $\chi\to\infty$ the
effective gravitational constant goes to zero and the gravitational
degrees of freedom decouple. Furthermore, if $K$ and $V/k^4$ approach
constants the asymptotic behavior in the scalar sector is a free
theory. This type of IR fixed point has a very simple physical content: a free massless scalar and a free graviton, with vanishing gravitational coupling. We investigate the scaling solutions connected to this fixed point.
For finite $y$ we expand $V$, $F$ and $K$ in inverse powers of
$y$. For a free scalar field with $F=\xi y$ and $K=K_0$, a rescaling
of $\chi$ multiplies $\xi$ and $K_0$ with the same factor. In
consequence, only the ratio $\epsilon=K_0/\xi$ appears in the flow
equations. More explicitly, for large $y$ we make the ansatz
\begin{align}\nonumber
V &= \sum_{j=0}^a \frac{v_j y^{-j}}{j ! \xi^j}\,,\qquad
F = \xi y + \sum_{j=0}^b \frac{f_j y^{-j}}{j ! \xi^j} ,\, \\[2ex]
K &= \xi \epsilon +\sum_{j=1}^c \frac{k_j y^{-j}}{j ! \xi^{j-1}}\,.
\label{largefieldexp} \end{align}
Switching to $x=1/(\xi y)$ and evaluating the flow at fixed $x$, one
finds to first order in powers of $x$ the set of flow equations
\begin{align}\nonumber
\partial_t V & = -4 V -2 x \frac{\partial V}{\partial x} + \frac{A_V}{B^3} + \frac{C_V}{B^4} x\,, \\[2ex]
\nonumber
\partial_t F & = -2 F -2 x \frac{\partial F}{\partial x} + \frac{A_F}{B^3} + \frac{C_F}{B^4} x\,,\\[2ex]
\partial_t K & = -2 x \frac{\partial K}{\partial x} + \frac{C_K}{B^4} x\,,
\label{eqn:betafcts}\end{align}
with $\partial_t = k\partial_k$ and
\begin{align}
\label{eqn:defB}
B=\epsilon+6= \frac{K_0}{\xi}+6.
\end{align}
The terms $-4V$ and $-2F$ reflect the dimensionality of $V$ and $F$,
while the terms $-2x \partial V /\partial x$, $-2x \partial F /\partial x$ and $-2x
\partial K /\partial x$ result from translating the flow to fixed $x$. The
coefficients $A_i$ and $C_i$ are computed from the one loop form of
the flow equation for the effective action. They read
\begin{align} \label{eqn:betafctscoeff} A_V=& \frac{1}{192 \pi ^2}\left(9 \epsilon
^3+82 \epsilon ^2+612 \epsilon +2760\right)\,, \\[2ex]\nonumber
C_V=& \frac{1}{2592 \pi^2}\left(1080 f_0 \epsilon ^3-3240 f_0
\epsilon ^2-62208 f_0 \epsilon \right.\\[2ex]\nonumber
&\left. +1080 k_1 \epsilon ^2-3240 k_1 \epsilon -62208 k_1+383 v_0
\epsilon ^4 \right. \\[2ex]\nonumber & \left.+5004 v_0 \epsilon
^3+51120 v_0 \epsilon ^2+278208 v_0 \epsilon +382320
v_0\right),\\[2ex]\nonumber A_F=& \frac{1}{3456 \pi ^2}\left(-253
\epsilon ^3-6094 \epsilon ^2-36240 \epsilon -51840\right),
\\[2ex]\nonumber C_F=& \frac{1}{5184 \pi ^2}\left(2148 f_0 \epsilon
^3-2916 f_0 \epsilon ^2-98712 f_0 \epsilon
+\right. \\[2ex]\nonumber & \left. 2310 k_1 \epsilon ^2-972 k_1
\epsilon -92880 k_1-345 v_0 \epsilon ^4\right. \\[2ex]\nonumber &
\left. -23498 v_0 \epsilon ^3-213492 v_0 \epsilon ^2-546552 v_0
\epsilon \right. \\[2ex]
&
\left. -263520 v_0\right)\,,\nonumber \\[2ex]\nonumber
C_K=& \frac{1}{36 \pi ^2}\left(-\epsilon ^4+90 \epsilon ^3+2079
\epsilon ^2+12636 \epsilon +26244\right)\,.
\end{align}
The fluctuation contributions do not enter the flow of the leading
terms $F=\xi / x$ and $K=K_0$ such that $\xi$ and $K_0$, and therefore
also $\epsilon$, are arbitrary couplings or \textquotedblleft
integration constants\textquotedblright . The appearance of a free
parameter $\epsilon$ corresponds to the undetermined $\xi$ in an
earlier calculation \cite{Henz:2013oxa} with field independent $K=K_0=1$.
We are interested in fixed point solutions where the left hand side of
\eq{eqn:betafcts} vanishes. The resulting system of
differential equations for the $x$-dependence is closed in every order
in the expansion in $x$. The fixed points for the $x$-independent
terms depend on $\epsilon$ and are given by
\begin{align}
v_0=\frac{A_V}{4 B^3}\,,\qquad \ f_0
=\frac{A_F}{2 B^3}\,.
\end{align}
Similarly, one has $k_1=C_K/(2 B^4)$. Inserting these values in $C_V$
and $C_F$ yields $v_1$ and $f_1$.
An interesting particular solution arises when one chooses integration constants such that $C_K=0$. This resembles the system investigated in \cite{Henz:2013oxa}.
The condition $C_k=0$ fixes $\epsilon$ to a certain $\epsilon_0$. For this
value the fixed point solution is given by
\begin{align}\begin{array}{l@{\quad}l@{\quad}l}
v_0=1.10 \cdot 10^{-3}, & f_0=-3.89\cdot 10^{-3},& \epsilon_0=109.97, \\[2ex]
v_1=2.32 \cdot 10^{-6}, & f_1=-2.81\cdot 10^{-6},& k_1=0.
\end{array}\label{eqn:fpsol}
\end{align}
It is the only real solution of this type which obeys the stability
condition $\epsilon \geq -6$.
The $\beta$-functions for the couplings $v_0,v_1,f_0,f_1,k_1$
correspond to the coefficients in the expansion of the r.h.s. of
eq. \eqref{eqn:betafcts} in powers of $x$, with
eq. \eqref{largefieldexp} inserted. The $\beta$-functions depend on
the couplings. We plot them in figure \ref{betafcts} for
$v_0,f_0,v_1,f_1$and $k_1$ given by eq. \eqref{eqn:fpsol}, as a
function of $\epsilon$ which is left free. They show a simultaneous
zero at $\epsilon_0=109.97$ as well as the pole at $\epsilon=-6.$ We
point out that our approximation is no longer valid for $B\to 0$.
\section{Global solution}
The expansion \eqref{largefieldexp} can be extended to rather high
powers of $x$, and one may perform Pad\'e approximations similar to
ref. \cite{Henz:2013oxa}. It is clear that such expansions will become
unreliable for small $\chi/k$ or large $x$, which we identify with the
ultraviolet limit. In order to connect with the region of large $x$ or small $y$ we need the full flow equations for the dimensionless functions $V/k^4$, $F/k^2$ and $K$. For simplicity, we keep the notation $\{ V,F,K \}$ for the dimensionless quantities. The system of flow equations has been
computed using algebraic algorithms. We have been able to solve numerically the fixed point equations $\partial_t V= \partial_t F=\partial_t K = 0$ for the full functions $V(y),\ F(y)$ and $K(y)$,
except for a range of very small $y$. Initial conditions for the
numerical solution are chosen at large $y$, where the expansion
\eqref{largefieldexp} is valid and can be solved analytically. More
precisely, we have taken the expansion \eqref{largefieldexp} for
$a=b=c=5$ and $y_\text{in}=10^5$. In figure \ref{globalsolall} we
display the numerical solutions for $\epsilon=\epsilon_0$ given by
eq. \eqref{eqn:fpsol}. Note that there is no point at which $\frac 1
2 F- V \approx 0$, meaning that the previously bothersome singularity
discussed in ref. \cite{Percacci:2015wwa} is not approached by the global
scaling solution.
We have also investigated solutions with values of $\epsilon$ different from $\epsilon_0$. We
were only able to obtain global solutions in the vicinity of
$\epsilon_0$, and only for a finite set of values for
$\epsilon$. While $V$ and $F$ are largely independent of $\epsilon$,
$K$ is rescaled by $\epsilon$ for large $y$, while $\epsilon$ becomes
less and less important for $y\to 0$.
For an estimate of the numerical error we compute the values of the
$\beta$-functions for our numerical solution, which should be
zero. They are normalized to the internal accuracy of the implicit
numerical differential equation solver employed, which was set to $8$
decimal digits. As long as this relative error is smaller than $1$, we can
assume the error to be negligible. This is the case for most parts of
the interval under consideration, with only small, local deviations
due to interpolation errors between the grid points of the numerical
solution. These are inevitable when taking derivatives to compute the
$\beta$-functions from the numerical solutions.
\begin{figure}[t]
\includegraphics[width=0.9\columnwidth]{numsollall_new.pdf}
\caption{Global numerical solution for the functions $V(y)$, $F(y)$
and $K(y)$ in a double-log plot. The matching with the Taylor
expansion was carried out at $y=10^5$.}
\label{globalsolall}
\end{figure}
\section{small field limit}
For an investigation of
the UV-behavior for $\chi/k\to 0$ we first consider the truncation
where $V$, $F$ and $K$ are $\chi$-independent. A scaling solution with
these properties closely resembles pure gravity in the
Einstein-Hilbert truncation \cite{Reuter:1996cp}, except for the presence of
a massless scalar field. In this approximation the coupling between
the scalar and the metric arises purely from the kinetic term. We
include the possibility of a $k$-dependent wave function
renormalization $Z$, or anomalous dimension $\eta=-\partial \log Z / \partial
\log k$. The wave function renormalization is defined by
$Z=K(y=0)$. At fixed $ y_R = Z y$ only the anomalous dimension $\eta$
(and not $Z$) appears in the flow equations. Here, $\eta$ is
determined by requiring that the $\eta$-dependent $\beta$-functions
maintain for $K( y_R=0)$ an arbitrary $k$-independent value. The
solution
\begin{align} \label{EHsolution} V=0.0006460\,,\quad F=0.002757\,,
\quad
\eta=1.6669\,,
\end{align}
can be interpreted as the Einstein-Hilbert approximation to our
system. It provides further evidence for the validity of the
Asymptotic Safety Scenario in a truncation extended by a running
kinetial.
In figure \ref{globalsolall} we have only plotted the range $y>10^{-7}$ for our numerical solution. For very small $y$ the
numerics become unstable. This is due to non-analytic behavior which
leads to diverging derivatives and therefore numerical
instabilities. Investigating the first derivative of the numerical
solutions, we find that they diverge as $y^{-\alpha}$ with
$\alpha=\frac 1 2$ for $y\to 0$. This explains why previous attempts
\cite{Narain:2009fy}, \cite{Percacci:2015wwa}, \cite{Henz:2013oxa} to expand in
integer powers of $y$ were not well suited. A best fit for the three
functions in the limit $y\to 0$, based on data generated from the
numerical solutions for $10^{-7}\leq y \leq 10^{-5}$, is given by
\begin{align}\nonumber
V(y) =& 6.3350\cdot 10^{-4} - 1.7137\cdot 10^{-4}\, y ^{1/2} \\[2ex]\nonumber
&+ 8.3172\cdot 10^{-3}\, y + 0.5626\, y^{3/2}\,, \\[2ex]\nonumber
F(y) =& 2.7077\cdot 10^{-3} - 3.7688\cdot 10^{-4} \, y ^{1/2} \\[2ex]\nonumber
& + 3.6550\cdot 10^{-3}\, y + 2.4944\, y^{3/2}\,, \\[2ex]\nonumber
K(y) =& 8.7018 + 455.269\, y ^{1/2} \\[2ex]
& +12940.3\, y + 223285\, y^{3/2}\,.
\label{eqn:bestfits}\end{align}
We observe that the constants approached by $V$ and $F$ in the limit
$y\to 0$ are very close to the Einstein-Hilbert truncation
\eqref{EHsolution}. This is no coincidence: For small $y$ our system
approaches the UV fixed point of asymptotically safe quantum
gravity. The slight numerical differences for $V(0)$ and $F(0)$ may
arise from the small $y$-dependence of $V$, $F$ and $K$ in equation
\eqref{eqn:bestfits}, which has been neglected in the truncation
leading to the values \eqref{EHsolution}. This may have a
stronger influence on the correct behavior of $K(y\to 0)$ and on the
value of the anomalous dimension. Extrapolating
eq. \eqref{eqn:bestfits} to $y\to 0$ would imply $\eta =0$. On the
other hand, it is not excluded that the correct matching of the
proposed scaling solution with the behavior for $y\to 0$ leads to
constraints on the allowed values of $\epsilon$. We are mainly
interested here in the generic IR-behavior which does not depend on
the precise details of the extreme limit $y\to 0$.
\section{Conformal invariants}
Conformal or Weyl transformations of the metric, $g_{\mu\nu} =
\omega^2(\chi)\tilde g_{\mu\nu}$, map the set of functions $\{V,F,K\}$ to a
new set $\{\tilde V,\tilde F,\tilde K\}$. By virtue of field
relativity the physical content of a model is specified by the two
invariants under this rescaling, namely \cite{Wetterich:2015ccd}
\begin{align}\hat V = \frac{V}{F^2}, \qquad \hat K = \frac
K F + \frac{6y}{F^2} \left(\frac{\partial F}{\partial y}\right)^2\,.
\end{align}
In figure \ref{VhatKhat} we show these invariants for the numerical
solutions with different values of $\epsilon$. While $\hat V$
shows very little $\epsilon$ dependency, the scaling $\hat K\sim
\epsilon$ is realized only for large $y$. The solutions with different
$\epsilon$ seem not to be equivalent.
\begin{figure}[t]
\includegraphics[width=0.9\columnwidth]{invariantseps.pdf}
\caption{Invariants $\hat V$ and $\hat K$ for the fixed point solution
for different values of $\epsilon$ in the vicinity of $\epsilon_0$.}
\label{VhatKhat}
\end{figure}
The invariants obey
\begin{align} \lim_{y\to\infty} \hat V= 0\,,\qquad\qquad \lim_{y\to\infty} \hat K= 0.
\end{align}
As both $V$ and $K$ go to a constant for $y\to \infty$, while $F$
grows with $\xi y$, we can immediately understand that the ratios
$V/F^2$ and $K/F$ vanish in this limit. Furthermore, we have $y (\partial
F / \partial y)^2 / F^2=1/y$. The potential $\hat V$ exhibits a maximum
located at
\begin{align} \label{eqn:maxVhat} y_\text{max}=0.02\,,\qquad \qquad
\hat V(y_\text{max})=87\,.
\end{align}
The vanishing of $\hat V$ for $y\to \infty$ implies that the effective
cosmological constant goes to zero asymptotically for cosmological
solutions where $\chi(t\to\infty)\to\infty$ \cite{Wetterich:1987fm}.
\section{Effective action in the Einstein frame}
Physical features of our system are most readily visible in the
Einstein frame, which is reached by a Weyl scaling leading to $\tilde
F=M^2/k^2$. Further using a rescaling of the scalar field to bring the
kinetic term to standard form yields in the
Einstein frame the effective action
\begin{align} \label{eqn:EAEinsteinframe} \Gamma =\int_x \sqrt{g}
\left( M^4 \hat V(\phi)-\frac12 M^2\,R+\frac{1}{2} g^{\mu\nu}
\partial_\mu\phi\partial_\nu\phi \right)\,.
\end{align}
From the kinetic term
\begin{align}
\label{XA}
\frac{M^2 \hat K}{2 k^2}\partial^\mu\chi\partial_\mu \chi = \frac 1 2 \partial^\mu
\phi \partial_\mu \phi\,,\quad \quad \hat K =\frac{16}{\alpha^2 y}\,,
\end{align}
one infers for constant $\alpha$ \begin{equation}\label{XB} \frac{1}{
y^2}=\exp\left(-\frac{\alpha \phi}{M}\right),\quad V=M^4\hat V =
\frac{v_0 M^4}{\xi^2}\exp\left(-\frac{\alpha \phi}{M}\right), \end{equation} with
modifications if $\alpha$ depends on $ y$. We observe that $ y$ is a
function of $\phi/M$, not involving $k$. All memory of $k$ has
disappeared in the Einstein frame. In figure \ref{fig:Vhatnormeps} we
plot the dimensionless potential in the Einstein frame $V/M^4 = \hat
V$ as a function of $\tilde \phi = \phi /M$. For the values of
$\epsilon$ for which numerical solutions could be established the
potential has a maximum for small values of $\tilde \phi$, as shown in
the inset.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\columnwidth]{Vnormeps.pdf}
\caption{Scalar potential $V / M^4$ in the Einstein frame for a standard normalization of
the scalar field. We show $\hat V$ as a function of $\tilde \phi = \phi/M$ for
different values of $\epsilon$ around $\epsilon_0$. Note the
maximum for small $\tilde \phi$ as well as the exponential tail for
large $\tilde \phi$. }
\label{fig:Vhatnormeps}
\end{figure}
\section{Discussion}
In this note we present for the first time candidates for a global scaling solution for dilaton quantum gravity.
We find several striking
features:\\[-2ex]
(i) For the scaling solution the effective action in the Einstein
frame shows no dependence on $k$. A realistic form of gravity results
directly for the scaling solution, without the need to deviate from
the fixed point. Variation of the effective action
\eqref{eqn:EAEinsteinframe} yields the quantum field equations and one
may discuss cosmological solutions. For fixed $\chi$ the limit $k\to
0$ corresponds to $\phi\to\infty$, but for arbitrarily small $k$ we
can always find values of $\chi$ such that $\phi$ is finite. We may
also consider the possibility that a separate physical cutoff sale
$\mu$ effectively stops the flow for $k\ll\mu$, leading to a deviation
from the scaling solution. A first approximation to the limit $k\to 0$
for this situation effectively replaces $k$ by $\mu$. There could be
additional terms realizing the physical cutoff, such as a mass term in the
potential \eq{eq:action}, $\Delta V = \mu^2\chi^2$.\\[-2ex]
(ii) The scaling solution is invariant under a simultaneous scaling of
fields and $k$. For any non-zero $k$ the effective action
\eq{eq:action} is not scale invariant if only the fields are
rescaled. Standard dilatation or scale symmetry can be recovered for
$k\to 0$. This is realized for our solution where $K$ and $F/y$ go to
constants for $y\to \infty$, while $V/y^2$ goes to zero. Scale
symmetry is spontaneously broken for cosmological solutions with
non-zero $\chi$. Therefore, one will find a massless Goldstone boson, the dilaton, in the limit $y\to\infty$,
$\phi\to\infty$. For finite
large $y$ this corresponds to the cosmon \cite{Wetterich:1987fm} with a
tiny field-dependent mass. Coupling matter fields in a scale invariant
way, e.g. with masses $\sim k\chi$, will lead to a massive particle
spectrum, with constant ratio particle mass / Planck mass.\\[-2ex]
(iii) For the scaling solutions that we have found the potential
asymptotically goes to a constant, $V(y\to\infty)=v_0$. With $F\sim y$
this results in $\hat V\sim y^{-2}$. The potential in the Einstein
frame decreases exponentially for $\phi\to\infty$, according to
equation \eqref{XB}. For cosmological solutions where $\phi\to\infty$
for $t\to\infty$, as realized if one starts on the right side of the
maximum in figure \ref{fig:Vhatnormeps}, the effective cosmological
constant goes to zero asymptotically. This is precisely the mechanism
proposed in the first paper on quintessence \cite{Wetterich:1987fm}. If
nature is characterized by a scaling solution for which $\hat V$
vanishes for $y\to\infty$, the cosmological constant problem is
solved, at least for asymptotic time. Our numerical solution for
$\epsilon=\epsilon_0$ yields $\alpha\approx0.4$. This is too small for
a realistic cosmological scaling solution for late cosmology, which
requires $\alpha \gtrsim 10$ \cite{Wetterich:2014gaa}.\\[-2ex]
(iv) The maximum of the potential for small $\tilde y$ shown in figure
\ref{fig:Vhatnormeps} offers, in principle, the possibility of an
inflationary stage. The end of inflation typically occurs
\cite{Wetterich:2014gaa, Wetterich:2013jsa} once $\hat K y$ drops below a
constant of order one. A glance at figures \ref{VhatKhat} and
\ref{fig:Vhatnormeps} yields typical values $\hat K y \gtrsim 10$ for
$\tilde y > \tilde y_\text{max}$ (the position of the maximum), with
$\hat K y \approx 10^2$ for $y\gtrsim 10^{-2}$. Inflation presumably
does not end for cosmologies resulting from the scaling solutions.\\[-2ex]
We conclude that the cosmology for the solutions found so far is not
yet realistic. Nevertheless, we emphasize that for a first time we can
directly connect cosmology to scaling solutions in quantum gravity,
without invoking any ad hoc association of $k$ with geometric
quantities. The reason is that no deviation from the scaling solution
is necessary in dilaton quantum gravity. The effective action in the
Einstein frame is independent of $k$, such that the limit $k\to 0$,
which is difficult in other settings, does not need to be performed
explicitly.
It is well possible that other scaling solutions exist beyond those found in this work. The fact that we have not found numerical
solutions for small $\epsilon$ may be related to numerical
instabilities rather than generic absence of such scaling
solutions. In the limit $B\to 0$ the limiting behavior for $y\to
\infty$ is expected to differ from the one discussed here -- the
limits $y\to\infty$ and $B\to 0$ do not commute. A possible scaling
behavior for $y\to \infty$ could be power-like, $V\sim y^{\delta_V},\
F\sim y^{\delta_F},\ K\sim y^{\delta_K}$. This results in $K +
6\delta_F^2 F/y = \hat K F \sim y^{\hat\delta_K}$, where $\hat
\delta_K=\max\left(\delta_K,\delta_F-1\right)$, a natural situation
arising for $\delta_K=\delta_F-1$. The asymptotic vanishing of the
cosmological constant occurs whenever $\delta_V < 2 \delta_F$. Such a
scaling behavior will be much closer to the properties of the infrared
fixed point discussed in ref. \cite{Wetterich:2014gaa}.
The type of scaling solution discussed here is special since we have
required that all three functions $V$, $F$ and $K$ depend only on
$y$. Scaling can also be realized if $y$ is replaced by a renormalized
variable $y_R=Z y$, see the discussion of the small field
limit. Furthermore, in view of possible field transformations we may
require only the invariants $\hat V$ and $y\hat K$ to be functions of
$y_R$. A possible explicit $k$-dependence of the individual functions
$V$, $F$, $K$ then merely reflects the $k$-dependence in the choice of
generalized renormalized fields. By an appropriate choice of a
renormalized metric we may always achieve standard forms as $F=1$ or
$F=y$. Furthermore, by a non-linear dependence of a renormalized
scalar field $\chi_R$ or $\chi$ we can realize standard forms either
of $K$ or of $V$ \cite{Wetterich:2013jsa}. Scaling solutions would then
merely require that the only remaining free function depends only on
the dimensionless ration $\chi_R / k$. Obviously, this condition is
much weaker than the simultaneous scaling of three functions $V$, $F$
and $K$. It will be an interesting question to find out if already a
simple truncation of dilaton quantum gravity admits scaling solutions
that lead to an acceptable cosmology.
\medskip\noindent {\em Acknowledgments.} We thank Andreas Rodigast
for collaboration on early stages of this project, as well as Aaron
Held for discussions. The authors acknowledge funding by the European
Research Council Advanced Grant (AdG), PE2, ERC-2011-ADG as well as
the German Academic Scholarship Foundation (Studienstiftung des
deutschen Volkes). \vfill
\bibliographystyle{bibstyle}
|
\section{Introduction to Hopf Formulas, HJ PDEs and Level Set Evolutions}
We briefly introduce Hamilton-Jacobi equations with initial data and
the Hopf formulas to represent the solution. We give some examples to
show the potential of our approach, including examples to perform level
set evolutions.
Given a continuous function $H:\Rn \rightarrow \R$ bounded from below
by an affine function, we consider the HJ PDE
\begin{align}
\label{1.1a}
\frac{\partial \varphi}{\partial t}(x,t) +
H(\nabla_x \varphi(x,t)) \;=\; 0 \quad \text{in } \Rn \times (0,+\infty),
\end{align}
where $\frac{\partial \varphi}{\partial t} $ and $\nabla_x \varphi$
respectively denote the partial derivative with respect to $t$ and the
gradient vector with respect to $x$ of the function $\varphi$. We are
also given some initial data
\begin{align}
\label{1.1b}
\varphi(x, 0) \;=\; J(x) & \quad \forall\, x \in \Rn,
\end{align}
where $J:\Rn \to \R$ is convex. For the sake of simplicity we only consider
functions $\varphi$ and $J$ that are finite everywhere. Results
presented in this paper can be generalized for $H:\Rn\to \R \cup \{
+\infty\}$ and $J:\Rn \to \R \cup \{ +\infty\}$ under suitable
assumptions. We also extend our results to an interesting class
of nonconvex initial data in section \ref{subsec.extension_future_work}.
We wish to compute the viscosity solution \cite{CL},\cite{CEL} for a
given $x \in \Rn$ and $t > 0$.
Using numerical approximations is essentially impossible for $n \geq
4$. The complexity of a finite difference equation is exponential in
$n$ because the number of grid points is also exponential in $n$.
This has been found to be impossible, even with the use of
sophisticated e.g. ENO, WENO, DG, methods
\cite{OS},\cite{HS},\cite{ZS}. High order accuracy is no cure for
this curse of dimensionality.
We propose and test a new approach, borrowing ideas from convex
optimization, which arise in the $\ell_1$ regularization convex
optimization \cite{GO},\cite{YOGD} used in compressive sensing
\cite{CRT},\cite{Do}. It has been shown experimentally that these
$\ell_1$ based methods converge quickly when we use Bregman and split
Bregman iterative methods. These are essentially the same as Augmented
Lagrangian methods \cite{He} and Alternating Direction Method of
Multipliers methods \cite{GM}. These and related first order and
splitting techniques have enjoyed a renaissance since they were
recently
used very successfully for these $\ell_1$ and related problems
\cite{YOGD},\cite{GO}.
One explanation for their rapid convergence is the ``error forgetting"
property discovered and analyzed in \cite{OY} for $\ell_1$
regularization.
We will solve the initial value problem \eqref{1.1a}-\eqref{1.1b}
without discretization, using the Hopf formula~\cite{Ho}
\begin{align}
\varphi(x,t) \;=\;& \left( J^* + t H \right)^*(x) \label{1.2}\\
\;=\;
&
-\min_{v\in R^{n}}
\left\{ J^*(v) + tH(v) - \langle x,v \rangle\right\}
\label{eq.Hopf_with_min}
\end{align}
where the Fenchel-Legendre transform $f^*:\Rn \to \R \cup \{
+\infty\}$
of a convex, proper, lower semicontinuous function $f:\Rn \to \R \cup
\{ +\infty\}$ is defined by
\cite{ekeland.76.book,hiriart-lemarechal.96.book-vol2,rockafellar.70.book}
\begin{equation}
J^*(v) = \sup_{x\in R^{n}}\left\{\langle v,x \rangle - J(x)\right\},
\label{1.3}
\end{equation}
where $\langle \cdot, \cdot\rangle$ denotes the $\ell_2(\Rn)$ inner
product. We also define for any $v \in \Rn$, $\|v\|_p =
\left(\sum_{i=1}^n |v_i|^p\right)^{\frac{1}{p}}$ for $1 \leq p < +
\infty$ and $\|v\|_\infty = \max_{i=1,\dots,n} |v_i|$.
Note that since $J:\Rn \to \R$ is convex we have that $J^*$ is
1-coercive \cite[Prop. 1.3.9,
p. 46]{hiriart-lemarechal.96.book-vol2},
i.e., $\lim_{\|x\| \to + \infty} \frac{J^*(x)}{\|x\|_2}\;=\;
+\infty$.
When the gradient $\nabla_x \varphi(x,t)$ exists, then it is precisely
the unique minimizer of (\ref{eq.Hopf_with_min}); in addition,
the gradient will also provide the optimal control considered in
this paper (see
Section \ref{sec.introduction_optimal_control}) For instance, this
holds for any $x\in \Rn$ and $t>0$ when $H : \Rn \to \R$ is convex,
$J:\Rn\to\R$ is strictly convex, differentiable and 1-coercive.
The Hopf formula only requires the convexity of $J$ and the continuity
of $H$, but we will often require that $H$ in (\ref{1.1a}) is also
convex.
We note that the case $H = \|\cdot\|_1$ corresponds to
$H(\nabla_x\varphi(x,t)) = \sum_{i=1}^n
\left|\frac{\partial}{\partial x_i}\varphi(x,t)\right|$, used, e.g. to
compute the Manhattan distance to the zero level set of $x\mapsto
J(x)$, \cite{Di}. This optimization is closely related to the $\ell_1$
type minimization \cite{CRT,Do},
\begin{equation}
\min_{v\in\Rn} \left\{\|v\|_1 + \frac{\lambda}{2} \|Av -
b\|_2^2\right\}, \nonumber
\end{equation}
where $A$ is an $m \times n$ matrix with real entries, $m < n$, and
$\lambda > 0 $ arising in compressive sensing.
Let $\Omega \subset \Rn$ be a closed convex set. We denote by
$\interior \Omega$ the interior of $\Omega$.
If $J(x) < 0$ for $ x \in \interior{\Omega}$,
$J(x) > 0$ for $x \in (\Rn \setminus \Omega)$ and $J(x) = 0$ for
$x \in (\Omega \setminus \interior{\Omega})$,
then the solution $\varphi$ of \eqref{1.1a}-\eqref{1.1b} has the
property that for any $t>0$ the set $\{x \in \Rn \;|\; \varphi(x,t) =
0\}$ is precisely the set of points in $\Rn \setminus \Omega$ for
which the Manhattan distance to $\Omega$ is equal to $t > 0$. This is
an example of the use of the level set method \cite{OSe}.
Similarly, if we take $H = \|\cdot\|_2$ then for any $t>0$ the
resulting zero level set of $x \mapsto \varphi(x,t)$ will be the
points in $\Rn \setminus \Omega$ whose Euclidean distance to $\Omega$
is equal to $t$. This fact will be useful later when we find the
projection from a point $x \in \Rn$ to a compact, convex set $\Omega$.
We present here two somewhat simple but illustrative examples to show the
potential power of our approach. Time results are presented in Section
\ref{sec.numerical_results} and show that we can compute solution of
some HJ-PDEs in fairly high dimensions at a rate below a millisecond
per evaluation on a standard laptop.
Let $H = \|\cdot\|_1$ and $J(x) = \frac{1}{2} \left(\sum_{i=1}^n
\frac{x_i^2}{a_i^2} - 1\right)$ with $a_i > 0$ for $i=1,\dots,n$.
Then, for a given $t > 0$, the set $\{x \in \Rn \;|\; \varphi(x,t) =
0\}$ will be precisely the set of points at Manhattan distance $t$
outside of the ellipsoid determined by $\{ x \in \Rn \;|\; J(x) \leq
0\}$. Following \cite[Prop. 1.3.1,
p. 42]{hiriart-lemarechal.96.book-vol2} it is easy to
see that $J^*(x) \;=\;
\frac{1}{2} \sum_{i=1}^n a_i^2 x_i + \frac{1}{2}$.
So:
\begin{align}
\varphi(x,t) \;=\;& -\frac{1}{2} - \min_{v \in R^{n}} \left\{\frac{1}{2}
\sum_{i=1}^n a_i^2 v_i^2 + t\sum_{i=1}^n |v_i| - \langle
x,v\rangle\right\}
\nonumber \\
\;=\;& -\frac{1}{2} - \min_{v \in R^{n}} \left\{\frac{1}{2}
\sum_{i=1}^n a_i^2 \left(v_i - \frac{x_i}{a_i^2}\right)^2 +
t \sum_{i=1}^n |v_i|\right\}
+ \frac{1}{2} \sum_{i=1}^n \frac{x_i^2}{a_i^2}. \nonumber
\end{align}
We note that the function to be minimized decouples into scalar
minimizations of the form
\begin{equation}
\nonumber
\min_{y\in\Rn} \left(\frac{1}{2} \|y-x\|_2^2 + \alpha\|y\|_1\right), \
\ \alpha > 0.
\end{equation}
The unique minimizer is the classical soft thresholding operator
\cite{lions.79.sna,figueiredo.98.asolimar,daubechies.04.cpam} defined
for any component~$i=1,\dots,n$ by
\begin{equation}
\label{eq.shrink1_def}
\left(\shrink_1(x,\alpha) \right)_i \;=\;
\begin{dcases}
x_i - \alpha & \mbox{ if } x_i > \alpha, \\
0 & \mbox{ if } |x_i| \leq \alpha, \\
x_i + \alpha & \mbox{ if } x_i < -\alpha.
\end{dcases}
\end{equation}
Therefore, for any $x\in\Rn$, any $t>0$ and any $i=1,\dots,n$ we have
\begin{displaymath}
\frac{\partial\varphi}{\partial x_i} (x,t)
\;=\;
\frac{1}{(a_i)^2} \ \left(\shrink_{1}(x,t)\right)_i,
\end{displaymath}
and
\begin{displaymath}
\varphi(x,t) = -\frac{1}{2} +
\sum_{i \in \{0,\dots,n\} \setminus B(t)}
\frac{1}{2}\left(\frac{|x_i|-t}{a_i}\right)^2.
\end{displaymath}
Here $B(t) \subseteq \{0,\dots,n\}$ consists of indices for which
$|x_i| \leq t$, and thus $\{0,\dots,n\} \setminus B(t)$ corresponds to
indices for which $|x_i| > t$, and the zero level set moves outwards
in this elegant fashion.
We note that in the above case we were able to compute the solution
analytically and the dimension $n$ played no significant role. Of
course this is rather a special problem, but this gives us some idea
of what to expect in more complicated cases, discussed in
section \ref{sec.algo_optimal_control}.
We will often need another shrink operator, i.e., when we solve the
optimization problem with~$\alpha > 0$ and $x \in \Rn$
\begin{displaymath}
\min_{v \in \Rn} \left\{\frac{1}{2} \|v - x\|_2^2 + \alpha \|v\|_2\right\}.
\end{displaymath}
Its unique minimizer is given by
\begin{equation}
\label{eq.shrink2_def}
\hbox{shrink}_2 (x,\alpha) \;=\;
\begin{cases}
\frac{x}{\|x\|_2} \max (\|x\|_2 - \alpha, 0) & \mbox{ if } x\neq 0,
\\
0 & \mbox{ if } x=0.
\end{cases}
\end{equation}
and thus its optimal value corresponds to the Huber function
(see \cite{winkler.03.book} for instance)
\begin{equation}
\nonumber
\min_{v \in \Rn} \left\{\frac{1}{2} \|v - x\|_2^2 + \alpha \|v\|_2\right\}
\;=\;
\begin{cases}
\frac{1}{2} \|x\|_2 & \hbox{ if } \|x\|_2 \leq \alpha, \\
\alpha \|x\|_2 - \frac{\alpha^2}{2} & \hbox{ if } \|x\|_2 > \alpha.
\end{cases}
\end{equation}
To move the unit sphere outwards with normal velocity 1,
we use the following formula
\begin{align}
\varphi(x,t) &= -\min_{v \in \Rn} \left(\frac{\|v\|_2^2}{2} - t\|v\|_2 -
\langle x,v\rangle\right) - \frac{1}{2} \nonumber \\
&= -\min_{v\in\Rn} \left\{\frac{1}{2} \|v-x\|_2^2 + t\|v\|_2\right\} +
\frac{1}{2} (\|x\|_2^2 - 1) \nonumber \\
&=
\begin{cases}
\frac{1}{2} (\|x\|_2 - t)^2 -\frac{1}{2} & \mbox{ if } \|x\|_2 > t\\
-\frac{1}{2} & \mbox{ if } \|x\|_2 \leq t \nonumber,
\end{cases}
\nonumber \\
\end{align}
and, unsurprisingly, the zero level set of $x \mapsto \varphi(x,t)$
is the set $x$ satisfying $\|x\|_2 = t+1$, for $t>0$.
These two examples will be generalized below so
that we can, with extreme speed, compute the signed distance, either
Euclidean, Manhattan or various generalizations, to the boundary of
the union of a finite collection of compact convex sets.
The remainder of this paper is organized as follows:
Section~\ref{sec.introduction_optimal_control} contains
an introduction to optimal control and its connection to HJ-PDE.
Section~\ref{sec.algo_optimal_control} gives the details of our
numerical methods. Section~\ref{sec.numerical_results} presents
numerical results with some details.
We draw some concluding remarks and give future plans in
Section~\ref{sec.conclusion}. The appendix links our approach to the
concepts of gauge and support functions in convex analysis.
\section{Introduction to Optimal Control}
\label{sec.introduction_optimal_control}
First, we give a short introduction to optimal control and its
connection to HJ PDE which is given in \eqref{eq.HJ_optimal_value}.
We also introduce positively homogeneous of degree one Hamiltonians
and describe their relationship to optimal control problems. We
explain how to recover the optimal control from the solution of the
HJ-PDE. An
appendix describes further connections between these Hamiltonians and
gauge in convex analysis. Second, we present some extensions of our
work.
\subsection{Optimal control and HJ-PDE}
We are largely following the discussion in \cite{Dol}, see also
\cite{Ev}, about optimal control and its link with HJ PDE. We briefly
present it formally and we specialize it to the cases considered in
this paper.
Suppose we are given a fixed terminal time $T \in \R$, an initial time
$t < T$ along with an initial $x\in\Rn$. We consider the Lipschitz
solution $x:[t, T] \to \Rn$ of the following ordinary differential
equation
\begin{equation}
\label{eq.ode-optimal-control}
\begin{cases}
\frac{d \mathrm{x}}{d s}(s)
\;=\;
f\left(\beta(s)\right) & \mbox{ in } (t, T),\\
\mathrm{x}(t) \;=\; x,
\end{cases}
\end{equation}
where $f: C \to \Rn$ is a given bounded Lipschitz function and $C$
some given compact set of~$\Rn$. The solution of
\eqref{eq.ode-optimal-control} is affected by the measurable function
$\beta:(-\infty, T] \to C$ which is called a control.
We note
$\setMeasurableFct = \{ \beta:(-\infty, T]\to C \; |\; \beta \mbox{ is
measurable}\}$.
We consider the functional for given initial time $t<T$, $x\in\Rn$ and
control $\beta$
$$
K(x,t ; \beta) \;=\;
\int_{t}^{T} L(\beta(s))ds
\,+\, J(\mathrm{x}(T)),
$$
where $\mathrm{x}$ is the solution of \eqref{eq.ode-optimal-control}.
We assume that the terminal cost $J:\Rn \to \R$ is convex. We also
assume that the running cost $L: \Rn \to \R \cup\{ + \infty\}$ is
proper, lower semicontinuous, convex, 1-coercive and $\dom{L}
\subseteq C$ where $\dom{L}$ denotes the domain of $L$ defined by
$\dom{L}\,=\, \{x\in\Rn\,|\, L(x) < +\infty \}$. The minimization
of $K$ among all possible controls defines the value function $v:\Rn
\times (-\infty,\, T] \to \R$ given for any $x\in\Rn$ and any $t < T$ by
\begin{equation}
\label{eq.value_function}
v(x,t) \;=\;
\inf_{\beta \in \setMeasurableFct} K(x,t; \beta).
\end{equation}
The value function \eqref{eq.value_function} satisfies the dynamic
programming principle for any $x \in \Rn$, any $t \geq T$ and any
$\tau \in (t,T)$
$$
v(x,t) \;=\;
\inf_{\beta \in \setMeasurableFct}
\left\{ \int_t^{\tau} L(\mathrm{\beta(s)})\,ds +
v\left(\mathrm{x}\left(\tau\right), \tau\right)
\right\}.
$$
The value function $v$ also satisfies the following
Hamilton-Jacobi-Bellman equation with terminal value
\begin{equation}
\nonumber
\begin{dcases}
\frac{\partial v}{\partial t}(x,t) +
\min_{c \in C}\{ \langle \nabla_x v(x,t), c \rangle + L(c)\} \;=\; 0
& \quad \text{in } \Rn \times (-\infty, T),
\\
v(x,T) \;=\; J(x) & \quad \forall x \in \Rn.
\end{dcases}
\end{equation}
Note that the control $\beta(t)$ at time $t \in (-\infty,T)$ in
\eqref{eq.ode-optimal-control} satisfies $\langle \nabla_x v(x,t),
\beta(t) \rangle + L(\beta(t))
\;=\; \min_{c\in C}\{\nabla_x v(x,t), c \rangle + L(c)\}$
whenever $v(\cdot,t)$ is differentiable.
Consider the function $\varphi: \Rn \times [0, +\infty) \to \R$
defined by $\varphi \left(x,t \right) \;=\; v(x,T-t)$. We have that
$\varphi$ is the viscosity solution of
\begin{equation}
\label{eq.HJ_optimal_value}
\begin{dcases}
\frac{\partial \varphi}{\partial t}(x,t) +
H(\nabla_x \varphi(x,t)) \;=\; 0 & \quad \text{in } \Rn \times (0,+\infty),
\\
\varphi(x,0) \;=\; J(x) & \quad \forall x \in \Rn.
\end{dcases}
\end{equation}
where the Hamiltonian $H: \Rn \to \R \cup \{ +\infty\}$ is defined by
\begin{equation}
\label{eq.def_Hamiltonian_optimal_control}
H(p) \;=\; \max_{c \in C} \left\{ \langle -f(c), p\rangle - L(c)\right\}.
\end{equation}
We note that the above HJ-PDE is the same as the one we consider
thoughout this paper.
In this paper we use the Hopf formula (\ref{1.2}) to solve
\eqref{eq.HJ_optimal_value}. We wish the Hamiltonian $H:\Rn\to\R$ to be not
only convex but also positively 1-homogeneous, i.e., for any $p\in \Rn$
and any $\alpha > 0$
$$
H(\alpha\, p) \;=\; \alpha \, H(p).
$$
We proceed as follows. Let us first introduce the
characteristic function
$\mathcal{I}_\Omega:\Rn\to \R\cup\{+\infty\}$ of the set $\Omega$
which is defined by
\begin{equation}
\label{eq.def_characteristic_function_set}
\characfct{\Omega}(x) \;= \;
\begin{cases}
0 & \mbox{ if } x \in \Omega,\\
+\infty & \mbox{ otherwise}.
\end{cases}
\end{equation}
We recall that $C$ is a compact convex set of $\Rn$. We take $f(c)
\,=\, -c$ for any $c \in C$ in \eqref{eq.ode-optimal-control} and
$$
L \;=\; \mathcal{I}_C.
$$
Then, \eqref{eq.def_Hamiltonian_optimal_control} gives the
Hamiltonian $H:\Rn\to \R$ defined by
\begin{equation}
\label{eq.H_as_support_function}
H(p) \;=\; \max_{c \,\in\, C}\, \langle c, p \rangle.
\end{equation}
Note that the right-hand side of \eqref{eq.H_as_support_function}
is called the support function of the
closed nonempty convex set~$C$ in convex analysis (see e.g.,
\cite[Def. 2.1.1, p. 208]{hiriart-lemarechal.96.book-vol1}). We check
that $H$ defined by \eqref{eq.H_as_support_function} satisfies our
requirement. Since $C$ is bounded, we can invoke \cite[Prop. 2.1.3, p.
208]{hiriart-lemarechal.96.book-vol1} which yields that the
Hamiltonian is indeed finite everywhere. Combining \cite[Def. 1.1.1,
p. 197]{hiriart-lemarechal.96.book-vol1}
and \cite[Prop. 2.1.2, p. 208]{hiriart-lemarechal.96.book-vol1}
we obtain that $H$ is positively 1-homogeneous and convex. Of course,
the Hamiltonian can also be expressed in terms of Fenchel-Legendre
transform; we have for any $p \in \Rn$
$$
H(p) \;=\; \max_{c \,\in\, C}\, \langle c, p \rangle \;=\;
\left(\characfct{C}\right)^*(p),
$$
where we recall that the Fenchel-Legendre is defined by \eqref{1.3}.
The nonnegativity of the Hamiltonian is related to the fact that $C$
contains the origin, i.e., $0 \in C$, and gauges. This connection is
described in the appendix.
Note that the controller $\beta(t)$ for $t \in (-\infty,T)$
in \eqref{eq.ode-optimal-control} is recovered for the solution
$\varphi$ of \eqref{eq.HJ_optimal_value} since we have
$$
\max_{c\in C}\langle c, \nabla \varphi(x,T-t) \rangle \;=\; \langle
\beta(t), \nabla_x \varphi(x,T-t) \rangle
$$
whenever $\varphi(\cdot,t)$ is differentiable. For any $p \in \Rn$
such that $\nabla H(p)$ exists we also have $H(p) = \langle p, \nabla
H(p) \rangle$. Thus we obtain that the control is given by $\beta(t) =
\nabla H(\nabla_x \varphi(x,T-t))$.
We present in Section\ref{sec.algo_optimal_control} our
efficient algorithm that computes not only $\varphi(x,t)$ but also
$\nabla_x\varphi(x,t)$. We emphasize that we do {\it not} need to use
some numerical approximations to compute the spatial gradient. In other
words our algorithm computes all the quantities we need to get the
optimal control without using any approximations.
It is sometimes convenient to use polar coordinates. Let us denote the
$(n-1)$-sphere by $\Snm1 \;=\; \{ x \in \Rn \;|\; \|x\|_2 =1 \}$.
The set $C$ can be described in terms of the Wulff shape~\cite{OM} by
the function $W:\Snm1 \to \R$. We set
\begin{equation}
\label{2.7}
C \;=\; \{ (R\, \theta) \in \Rn \;|\;
R \geq 0,\,
\theta \in \Snm1,\,
R \leq W\left( \theta \right)\}.
\end{equation}
The Hamiltonian $H$ is then naturally defined via $\gamma: \Snm1 \to
\R$ for any $R>0$ and any $\theta \in \Snm1$ by
\begin{equation}
\label{3.2}
H(R \theta) \; =\; R
\gamma\left(\theta \right),
\end{equation}
and where $\gamma$ is defined by
\begin{equation}
\label{eq.gamma}
\gamma(\theta)
\;=\;
\sup\left\{
\langle W(\theta')\,\theta', \theta\rangle
\;|\;
\theta' \in \Snm1
\right\}.
\end{equation}
The main examples are $H\;=\; \| \cdot\|_p$ for $p \in
[1,+\infty)$ and $H \;=\; \|\cdot \|_\infty$. Others include
the following two: $H = \sqrt{\langle \cdot, A \cdot\rangle} \,=\, \|
\cdot\|_A$ with $A$ a symmetric positive definite matrix, and $H$
defined as follows for any $p \in \Rn$
$$H(p) =
\begin{cases}
\frac{\langle p, A p \rangle }{\|p\|_2} & \mbox{ for } p\neq 0\\
0 & \mbox{ otherwise}.
\end{cases}
$$
In future work, we will also consider Hamiltonians defined as the
supremum of linear forms such as those that arise in linear
programming.
We will devise very fast, low memory, totally parallelizable and
apparently low time complexity methods for solving
\eqref{eq.HJ_optimal_value} with $H$ given by
\eqref{eq.H_as_support_function} in the next section.
\subsection{Some extensions and future work }
\label{subsec.extension_future_work}
In this section we show that we can solve the problem for a much more
general class of Hamiltonians and initial data which arise in optimal
control, including an interesting class of nonconvex initial data.
Let us first consider Hamiltonians that correspond to linear controls.
Instead of \eqref{eq.ode-optimal-control}, we consider the following
ordinary differential equation
$$
\frac{d \mathrm{x}}{d s}(s) \;=\; M\, \mathrm{x}(s) + N(s) \beta(s),
$$
where $M$ is a $n \times n$ matrices with real entries and $N(s)$ for
any $s \in (-\infty,T]$ is a $n \times m$ matrices
with real entries. We can make a change of variables
$$
z(s) \;=\; e ^{-s M } \mathrm{x}(s),
$$
and we have
$$
\frac{d z}{d s}(s) \;=\; e ^{-s M } N(s)\beta(s).
$$
The resulting Hamiltonian now depends on $t$ and
\eqref{eq.def_Hamiltonian_optimal_control} becomes:
$$
H(p,t) \;=\;
- \inf_{c \,\in\, \Rn} \left( \langle e ^{-t M} N(t) c, p\rangle +
L(c)\right).
$$
If $C$ is a closed convex set, and $L\; =\; \characfct{C}$ we have
$$
H(p,t) \;=\;
- \min_{c\,\in\, C}
\langle e ^{-t M } N(t) c, p\rangle,
$$
which leads to a positively 1-homogeneous convex Hamiltonians as a
function of $p$ for any fixed $t \geq 0$. If we
have convex initial data there is a simple
generalization of the Hopf formulas \cite{lions-rochet.86.ams}
\cite[Section 5.3.2, p. 215]{kurzhanski-varaiya.14.book}
$$
\varphi(z,t) \;=\;
- \min_{u \in \Rn} \left(
J^*(u) + \int_0^t H(u,s) ds - \langle z, u\rangle
\right).
$$
We intend to program and test this in our next paper.
\bigskip
We now review some well known results about the types of convex
initial value problems that yield
to max/min-plus algebra for optimal control problems (see e.g.
\cite{fleming.97.pisa,mceneaney.06.book,akian.06.book_chapter} for
instance).
Suppose we have $k$ different initial value problems $i=1,\dots,k$
\begin{equation}
\nonumber
\begin{dcases}
\frac{\partial \phi_i}{\partial t}(x,t) + H(\nabla_x \phi_i (x,t)) \;=\;
0 & \quad \text{in } \Rn \times (0,+\infty),
\\
\phi_i(x,0) \;=\; J_i(x) & \quad \forall x \in \Rn. \nonumber
\end{dcases}
\end{equation}
where all initial data $J_i:\Rn\to \R$ are convex,
the Hamiltonian $H:\Rn \to R$ is convex and 1-coercive.
Then, we may use the Hopf-Lax formula to get, for any $x \in \Rn$ and
any $t > 0$
$$
\phi_i(x,t) = \min_{z \in \Rn}
\left\{
J_i(z) + t H^*\left( \frac{x-z}{t}\right)
\right\},
$$
so
$$
\min_{i=1,\dots,k} \phi_i(x,t) =
\min_{z \in \Rn}
\left\{ \min_{i=1,\dots,k}
\left\{ J_i(z) + t H^*\left( \frac{x-z}{t}\right) \right\}
\right\}.
$$
So we can solve the initial value problem
\begin{equation}
\nonumber
\begin{dcases}
\frac{\partial \phi}{\partial t}(x,t) + H(\nabla_x \phi (x,t)) \;=\;
0 & \quad \text{in } \Rn \times (0,+\infty),
\\
\varphi_i(x,0) \;=\; \min_{i = 1,\dots,k} J_i(x) & \quad \forall x \in \Rn. \nonumber
\end{dcases}
\end{equation}
by simply taking the pointwise minimum over the $k$ solutions
$\phi_i(x,t)$, each of which has convex initial data. See
Section~\ref{sec.numerical_results} for numerical results. As an
important example, suppose
$$
H\; =\; \| \cdot\|_2,
$$
and each $J_i$ is a level set function for a convex
compact set $\Omega_i$ with nonempty interior and where the interiors
of each $\Omega_i$ may overlap with each other.
We have $J_i(x) < 0 $ inside $\Omega_i$ , $J_i(x) >0 $ outside $\Omega_i$, and
$J_i(x) = 0$ at the boundary of $\Omega_i$. Then $\min_{i=1,\dots,n} J_i$ is
also a level set function for the union of the $\Omega_i$.
Thus we can solve complicated level set motion involving merging
fronts and compute a closest point and the associated proximal points to
nonconvex sets of this type. See section~\ref{sec.algo_optimal_control}.
\bigskip
For completeness we add the following fact about the minimum of Hamiltonians.
Let $H_i:\Rn \to \R$, with $i=1,\dots,k$, be $k$ continuous
Hamiltonians bounded from below by a common affine function.
We consider for $i=1,\dots,k$
\begin{equation}
\nonumber
\begin{dcases}
\frac{\partial \phi_i}{\partial t}(x,t) + H_i(\nabla_x \phi (x,t)) \;=\;
0 & \quad \text{in } \Rn \times (0,+\infty),
\\
\varphi(x,0) \;=\; J(x) & \quad \forall x \in \Rn, \nonumber
\end{dcases}
\end{equation}
where $J:\Rn \to \R$ is convex.
Then,
\begin{align}
\nonumber
\min_{i=1,\dots,k}
\left(
-\phi_i(x,t)
\right)
\;=\; &
\nonumber
\min_{i=1,\dots,k} \left\{
\min_{u \in \Rn}
\left\{
J^*(u) + t H_i(u) - \langle u,x\rangle
\right\}
\right\},\\
\;=\; &
\nonumber
\min_{u\in \Rn} \left\{
J^*(u) + t \min_{i=1,\dots,k} \left\{H_i(u) - \langle u, x\rangle\right\}
\right\},
\end{align}
that is
\begin{align}
\max_{i=1,\dots,k} \phi_i(x,t) \;=\; -\min_{u \in \Rn}
\left\{
J^*(u) + t \min_{i=1,\dots,k} \left\{ H_i(u) - \langle u,x\rangle \right\}
\right\}.
\end{align}
So we find the solution to
\begin{equation}
\nonumber
\begin{dcases}
\frac{\partial \phi}{\partial t}(x,t) + \min_{i=1,\dots,k}
H_i(\nabla_x \phi (x,t)) \;=\;
0 & \quad \text{in } \Rn \times (0,+\infty),
\\
\varphi(x,0) \;=\; J(x) & \quad \forall x \in \Rn. \nonumber
\end{dcases}
\end{equation}
by solving $k$ different initial value problems and taking the
pointwise maximum. See Section~\ref{sec.numerical_results} for numerical results.
\bigskip
We end this section by showing that explicit formulas can be obtained
for the terminal value $\mathrm{x}(T)$ and the control $\beta(t)$ for
another class of running cost $L$. Suppose that $C$ is a convex
compact set containing the origin and take $f(c) = -c $ for any $c \in
C$. Assume also that $L: \Rn \to \R \cup \{+\infty\}$ is strictly
convex, differentiable when its subdifferential is nonempty and that
$\dom L$ has a non-empty interior with $\dom L \subseteq C$. Then the associated Hamiltonian $H$ is defined
by $H=L^*$. Then, using the results of \cite{darbon.13.cam}, we have
that the $(x,t) \mapsto \varphi(x,t)$ which solves
\eqref{eq.HJ_optimal_value} is given by the Hopf-Lax formula
$\varphi(x,t) \,=\, \min_{y\in\Rn}\left\{ J(y) +
tH^*(\frac{x-y}{t})\right\}$ where the minimizer is unique and
denoted by $\bar{y}(x,t)$. Note that the Hopf-Lax formula corresponds
to a convex optimization problem which allows us to compute
$\bar{y}(x,t)$. In addition, we can compute the gradient with respect
to $x$ since we have $\nabla_x \varphi(x,t) = \nabla
H^*\left(\frac{x-\bar{y}(x,t)}{t}\right) \in \partial J(y(x,t))$ for any given $x\in\Rn$ and
$t>0$. For any $t \in (-\infty, T)$ and fixed $x\in\Rn$ the control is
given by $\beta(t) = \nabla H(\nabla_x \varphi(x,T-xt))$ while the
terminal value satisfies $\mathrm{x}(T) = \bar{y}(x,T-t) = x - (T-t) \nabla
H(\nabla \varphi(x,(T-t)))$. Note that both the control and the
terminal value can be easily computed. More details about these facts
will be given in a forthcoming paper.
\section{Overcoming the Curse of Dimensionality for Convex Initial
Data and Convex Homogeneous Degree One Hamiltonians -- Optimal
Control}
\label{sec.algo_optimal_control}
We first present our approach for evaluating the solution of the
HJ-PDE and its gradient using the Hopf formula \cite{Ho},
Moreau's identity
\cite{moreau.65.proximite} and the split Bregman algorithm \cite{GO}.
We note that the split Bregman algorithm can be replaced by other
algorithms which converge rapidly for problems of this type. An
example might be the primal-dual hybrid gradient
method \cite{zhu-chan.08.cam,chambolle.11.jmiv}. Then, we show that our
approach can be adapted to compute a closest point on a closed set,
which is the union of disjoint closed convex sets with a non empty
interior, to a given point.
\subsection{Numerical optimization algorithm}
We present the steps needed to solve
\begin{equation}
\label{3.1}
\begin{dcases}
\frac{\partial \varphi}{\partial t}(x,t) +
H(\nabla_x \varphi(x,t)) \;=\; 0 & \quad \text{in } \Rn \times (0,+\infty),
\\
\varphi(x,0) \;=\; J(x) & \quad \forall x \in \Rn.
\end{dcases}
\end{equation}
We take $J:\Rn \to \R$ convex
and positively 1-homogeneous. We recall that solving \eqref{3.1},
i.e., computing
the viscosity solution, for a given $x \in \Rn,\, t > 0$ using
numerical approximations, is essentially impossible, for $n \geq 4$
due to the memory issue, and the complexity is exponential in~$n$.
An evaluation of the solution at $x\in\Rn$ and $t>0$ for the examples
we consider in this paper is of the order of $10^{-8}$ to $10^{-4}$
seconds on a standard laptop (see Section \ref{sec.numerical_results}).
The apparent time complexity seems to be polynomial in $n$ with
remarkably small constants.
We will use the Hopf formula \cite{Ho}:
\begin{equation}
\varphi(x,t) = -\min_v \left\{J^*(v) + tH(v)-\langle
x,v\rangle\right\}. \label{3.3}
\end{equation}
Note that the infimum is always finite and attained (i.e, it is a
minimum) since we have assumed that $J$ is finite everywhere on $\Rn$
and that $H$ is continuous and bounded from below by an affine function.
The Hopf formula \eqref{3.3}
requires only the continuity of $H$, but we will also require
the Hamiltonian $H$ be convex as well. We recall that the previous
section shows how to relax this
condition.
We will use the split Bregman iterative approach to solve this \cite{GO}
\begin{eqnarray}
v^{k+1} &=& \arg\min_{v \in \Rn} \{J^*(v) - \langle x,v\rangle +
\frac{\lambda}{2} \|d^k-
v-b^k\|_2^2\}, \label{3.4(a)} \\
d^{k+1} &=& \arg\min_{d \in \Rn} \left\{t H(d) + \frac{\lambda
}{2} \|d-v^{k+1}-b^k\|_2^2\right\}
\label{3.4(b)} \\
b^{k+1} &=& b^k + v^{k+1} - d^{k+1}. \label{3.4(c)}
\end{eqnarray}
For simplicity we consider $\lambda = 1$ and consider $v^0 = x$,
$d^0=x$ and $b^0 = 0$ in this paper. The algorithm still works for
any positive $\lambda$ and any finite values for $v^0$, $d^0$ and $b^0$.
The sequence $(v^k)_{k\in\mathbb{N}}$ and $(d^k)_{k\in\mathbb{N}}$ are
both converging to the same quantity which is a minimizer of~\eqref{3.3}.
We recall that when the minimizer of \eqref{3.3} is unique then
it is precisely the $\nabla_x \varphi(x,t)$; in other words,
both $(v^k)_{k\in\mathbb{N}}$ and $(v^k)_{k\in\mathbb{N}}$ converge
to $\nabla_x \varphi(x,t)$ under this uniqueness assumption. If the
minimizer is not unique then the sequences $(v^k)_{k\in\mathbb{N}}$
and $(v^k)_{k\in\mathbb{N}}$ converge to an element of the
subdifferential of $\partial (y\mapsto f(y,t))(x)$ (see below for the
definition of a subdifferential).
We need to solve \eqref{3.4(a)} and \eqref{3.4(b)}. Note that up to
some changes of variables, both optimization problems can be
reformulated as finding the unique minimizer of
\begin{equation}
\label{eq.general_opti_pb}
\Rn \ni w \mapsto
\alpha f(w) + \frac{1}{2} \|w - z\|_2^2,
\end{equation}
where $z \in \Rn$, $\alpha > 0$, and $f:\Rn \to \R \cup \{+\infty\}$
is a convex, proper, lower semi-continuous function. Its unique
minimizer $\bar{w}$ satisfies
the optimal condition
\begin{equation}
\nonumber
\alpha\, \partial f (\bar{w}) + \bar{w} - z \;\ni\; 0,
\end{equation}
where $\partial f (x)$ denotes the subdifferential (see
for instance \cite[p. 241]{hiriart-lemarechal.96.book-vol1},
\cite[Section 23]{rockafellar.70.book}) of $f$ at
$x \in \Rn$ and is defined by $\partial f (x) \;=\; \{ s \in \Rn\,|\,
\forall y \in \Rn,\, f(y)\, \geq\, f(x) + \langle s, y-x\rangle \}$.
We have
\begin{equation}
\nonumber
\bar{w} \;=\; \prox{\alpha\, \partial f} (z)
\;=\;
\arg\min_{w\in\Rn} \left\{ \alpha f(w) + \frac{1}{2} \|w - z\|_2^2
\right\},
\end{equation}
where $\prox{\partial f}$ denotes the \beginquote resolvent\endquote
operator of $f$ (see \cite[Def. 2, chp. 3, p. 144]{aubin.84.book},
\cite[p. 54]{brezis.73.book} for instance). It is also called the
proximal map of $f$ following the seminal paper of Moreau
\cite{moreau.65.proximite} (see also \cite[Def. 4.1.2, p.
318]{hiriart-lemarechal.96.book-vol2}, \cite[p.
339]{rockafellar.70.book}).
This mapping has been extensively studied in the context of
optimization (see for instance
\cite{combettes.07.siam-opt,eckstein-bertsekas.92.mpa,rockafellar.76.siam-co,teboulle.97.siam-op}).
\bigskip
Closed form formulas exist for the proximal of map for some specific
cases. For instance, we have seen in the introduction that
$\prox{\alpha \ \partial \|\cdot\|_i} \;=\shrink_i(\cdot,
\alpha)$ for $i=1,2$, where we recall that $\shrink_1$ and $\shrink_2$ are
defined by \eqref{eq.shrink1_def} and \eqref{eq.shrink2_def},
respectively. Another classical example consists of considering
a quadratic form $\frac{1}{2}\|\cdot\|_A^2 \;=\; \frac{1}{2} \langle
\cdot, A \cdot \rangle$, with $A$ a symmetric positive definite
matrix with real values, which yields $\prox{\alpha\, \partial
\left(\frac{1}{2}\|\cdot\|_A ^2\right)}\;=\; \left( I_n + \alpha \,
A\right)^{-1}$, where $I_n$ denotes the identity matrix of size $n$.
Assume $f$ is twice differentiable with a bounded Hessian, then the
proximal map can be efficiently computed using Newton's method.
Algorithms based on Newton's method require us to solve a linear
system that involves an $n \times n$ matrix. Note that typical high
dimensions for optimal control problems are about $n=10$. For
computational purposes, these order of values for $n$ are small.
We describe an efficient algorithm to compute the proximal map of
$\|\cdot\|_\infty $ in Section \ref{subsec.prox_infty} using
parametric programming \cite[Chap. 11, Section
11.M]{rockafellar.84.book}. An algorithm to compute the proximal map
for $\frac{1}{2}\| \cdot \|_1^2$ is described in Section
\ref{subsec.l1_linfty_squared}.
\bigskip
The proximal maps for $f$ and $f^*$ satisfy the celebrated Moreau
identity \cite{moreau.65.proximite} (see also \cite[Thm. 31.5, p.
338]{rockafellar.70.book}) which reads as follows: for any $w\in \Rn$
and any $\alpha > 0$
\begin{equation}
\label{eq.moreau-identity}
\prox{\alpha\, \partial f}(w) + \alpha \prox{\frac{1}{\alpha} \, \partial f^*}
\left( \frac{w}{\alpha}\right) \;=\; w.
\end{equation}
This shows that $\prox{\alpha \,\partial f}(w)$ can be easily computed
from $\prox{\frac{1}{\alpha} \, \partial
f^*}\left(\frac{w}{\alpha}\right)$. In other words, depending on the
nature and properties of the mappings $f$ and $f^*$, we choose the
one for which the proximal point is \beginquote easier\endquote
to compute. Section \ref{subsec.l1_linfty_squared} describes an
algorithm to compute the proximal map of $\frac{1}{2}\|\cdot
\|_\infty^2$ using only evaluations of $\prox{ \frac{\alpha}{2} \, \partial \|\cdot\|_1^2}$
using Moreau's identity \eqref{eq.moreau-identity}.
We shall see that Moreau's identity \eqref{eq.moreau-identity} can be
very useful to compute the proximal maps of convex and positively
1-homogeneous functions.
We consider problem \eqref{3.4(b)} that corresponds to compute the
proximal of a convex positively 1-homogeneous function $H:\Rn \to \R$.
(we use $H$ instead $f$ to emphasize that we are considering
positively 1-homogeneous functions and we set $\alpha=1$ to alleviate
notations.) We have that $H^*$ is the characteristic function of a
closed convex set $C \subseteq \Rn$, .i.e., the Wulff shape associated
to $H$~\cite{OM},
\begin{equation}
\nonumber
H^* \;=\; \characfct{C},
\end{equation}
and $H$ corresponds to the support function $C$, that is for any $p
\in\Rn$
\begin{equation}
\nonumber
H(p)\;=\; \sup_{ s \in C}\, \langle s, p \rangle.
\end{equation}
Following Moreau \cite[Example 3.d]{moreau.65.proximite},
the proximal point of $z \in \Rn$ relative to $H^*$ is
\begin{equation}
\nonumber
\prox{\partial H^*}(z)
\;=\;
\min_{w \in C} \left\{
\frac{1}{2}\|w-z\|_2^2 \right\}.
\end{equation}
In other words, $\prox{ \partial H^*}(z)$ corresponds to the
projection of $z$ on the closed convex set $C$ that we denote by
$\pi_C(z)$, that is for any $z \in \Rn$
\begin{equation}
\nonumber
\prox{\partial H^*}(z) \;=\; \pi_C(z).
\end{equation}
Thus, using the Moreau identity \eqref{eq.moreau-identity}, we see that
$\prox{\partial H}$ can be computed from the projection on
its associated Wulff shape and we have for any $z \in \Rn$
\begin{equation}
\label{eq.prox-projection}
\prox{\partial H} (z) \;=\;
z - \pi_C(z).
\end{equation}
In other words, computing the proximal map of $H$ can be performed by
computing the projection on its associated Wulff shape
$C$. This formula is not new, see e.g. \cite{chambolle.09.ijcv,OOTT}.
Let us consider an example. Consider Hamiltonians of the form
$H = \|\cdot\|_A = \sqrt{\langle \cdot, A \cdot\rangle}$ where $A$ is a
symmetric positive matrix. Here the Wulff shape is the ellipsoid $C \;=\;
\left\{ y\in \Rn\,|\, \langle x, A ^{-1} x \rangle \; \leq 1
\right\}$. We describe in Section \ref{subsec.projection_ellipsoid} an
efficient algorithm for computing the projection on an ellipsoid.
Thus, this allows us to compute efficiently the proximal map of norms
of the form $\| \cdot \|_A$ using \eqref{eq.prox-projection}.
\subsection{Projection on closed convex set with the Level Set Method}
\label{subsec.projection}
We now describe an algorithm based on the level set method \cite{OSe}
to compute the projection $\pi_\Omega$ on a compact convex set $\Omega
\subset \Rn$ with a nonempty interior. This problem appears to be of
great interest for its own sake.
Let $\psi:\Rn \times [0,+\infty)$ be the viscosity solution of the
eikonal equation
\begin{equation}
\label{eq.eikonal}
\begin{dcases}
\frac{\partial \psi}{\partial t}(y,s) +
\|\nabla_x \psi(y,s)\|_2 \;=\; 0 & \quad \text{in } \Rn \times (0,+\infty),
\\
\psi(y,0) \;=\; L(y) & \quad \forall y \in \Rn,
\end{dcases}
\end{equation}
where we recall that $\frac{\partial \psi}{\partial t}(y,t)$ and
$\nabla_x \psi(y,t)$ respectively denote the partial derivatives of
$\phi$ with respect to the time and space variable at $(y,s)$, and
where $L:\Rn \to \R$ satisfies for any $y \in \Rn$
\begin{equation}
\label{eq.L_level_set_fctx}
\begin{cases}
L(y) < 0 \mbox{ for any } y \in \mbox{int } \Omega, \\
L(y) > 0 \mbox{ for any } y \in (\Rn \setminus \Omega),\\
L(y) = 0 \mbox{ for any } y \in (\Omega \setminus \mbox{int } \Omega),\\
\end{cases}
\end{equation}
where $\mbox{int } \Omega$ denotes the interior of $\Omega$. Given
$s>0$, we consider the set
$$
\Gamma(s) \;=\;
\left\{
y \in \Rn \,|\, \psi(y, s) \;=\;0
\right\},
$$
which corresponds to all points that are at a (Euclidean)
distance $s$ from $\Gamma(0)$. Moreover, for a given point $y \in
\Gamma(s)$, the closest point to $y$ on $\Gamma(0)$ is exactly the
projection $\pi_\Omega(y)$ of $y$ on $\Omega$ and we have
\begin{equation}
\label{eq.projection_when_s_known}
\pi_{\Omega}(y)
\;=\;
y - s \frac{\nabla_x\psi(y,s)}{\|\nabla_x\psi(y,s)\|_2}.
\end{equation}
\bigskip
In this paper we will assume that $L$ is strictly convex, 1-coercive and
differentiable so that $\nabla_x \psi(s,y)$ exists for any $y \in \Rn$
and $s>0$. We note that if $\Omega$ is the finite union of sets of
this type then $\nabla_x \psi(s,y)$ may have isolated jumps. This presents
no serious difficulties.
Note that \eqref{eq.eikonal} takes the form of \eqref{3.1} with $H =
\|\cdot\|_2$ and $J = L$. We again use split Bregman to solve the
optimization given by the Hopf formula \eqref{1.2}. To avoid confusion
we respectively replace $J, v, d$ and $b$, by $L, w, e$ and $c$ in
\eqref{3.4(a)}-\eqref{3.4(c)}
\begin{eqnarray}
w^{k+1} &=& \arg\min_{w\in\Rn} \left\{L^*(w) - \langle z,w\rangle +
\frac{\lambda}{2} \|e^k-w -c^k\|_2^2\right\}, \label{3.9(a)}\\
e^{k+1} &=& \arg\min_{e\in\Rn} \left\{s\|e\|_2 +
\frac{\lambda}{2}\|e-w^{k+1}-c^k\|_2^2\right\}, \label{3.9(b)}\\
c^{k+1} &=& c^k + w^{k+1} - e^k. \label{3.9(c)}
\end{eqnarray}
An important observation here is that $e^{k+1}$ can be solved
explicitly in \eqref{3.9(b)} using the $\shrink_2$ operator defined by
\eqref{eq.shrink2_def}.
Note that the algorithm given by \eqref{3.4(a)}-\eqref{3.4(c)} allows
us to evaluate not only $\psi(y,s)$ but
also $\nabla_x\psi(y,s)$. Indeed
for any $s>0$ $\nabla_x \psi(y,s) \;=\;
\arg\min_{v\in\Rn} \{L^*(v) + s H(v)- \langle y, v\rangle\}$, the
minimizer being unique. Thus, the above algorithm \eqref{3.9(a)}-
\eqref{3.9(c)} generates sequences $(w^k)_{k \in \mathbb{N}}$ and
$(e ^k)_{k \in \mathbb{N}}$ that both converge to $\nabla_x
\psi(z,s)$.
The above considerations about the closest point and \eqref
{eq.projection_when_s_known} give us a numerical
procedure for computing $\pi_\Omega(y)$ for any $y \in (\Rn \setminus
\Omega)$.
Find the value $\bar{s}$ so that $\psi(y, \bar{s}) \;=\; 0$ where
$\psi$ solve \eqref{eq.eikonal}. Then, compute
\begin{equation}
\pi_{\Omega}(y)
\;=\;
y - \bar{s} \frac{\nabla_x\psi(y,\bar{s})}{\|\nabla_x\psi(y,\bar{s})\|_2},
\end{equation}
to obtain the projection $\pi_{\Omega}(z)$. We compute $\bar{s}$ using
Newton's method to find the $0$ of the function $(0, +\infty) \ni s
\mapsto \psi(z,s)$. Given an initial $s_0 > 0$ the Newton iteration
corresponds to computing for integers $l>0$
\begin{equation}
\label{eq.Newton_zero}
s_{l+1} \;=\;
s_{l} - \psi(z,s_{l}) \left(\frac{\partial \psi}{\partial t} (z,s_{l})
\right)^{-1}.
\end{equation}
From \eqref{eq.eikonal} we have $\frac{\partial \psi}{\partial t}
(z,s) = - \|\nabla_x \psi(z,s)\|_2$ for any $s > 0$. We can thus
compute \eqref{eq.Newton_zero}.
It remains to choose the initial data $L$ related to the set
$\Omega$. We would like it to be smooth so that the proximal point in
\eqref{3.9(a)} can be computed efficiently using Newton's
method. (If $L$ lacks differentiability the approach can be
easily modified using \cite[Chap 11, Section
11.M]{rockafellar.84.book} or using \cite{cheng-tsai.08.jcp}.) We
consider
$\Omega$ as Wulff shapes that are expressed, thanks to \eqref{2.7},
with the function $W:\Snm1\to \R$, that is
\begin{equation}
\nonumber
\Omega \;=\; \{ (R\, \theta) \in \Rn \;|\;
R \geq 0,\,
\theta \in \Snm1,\,
R \leq W\left( \theta \right)\}.
\end{equation}
As a simple example, if $W \equiv 1$, then we might try
$L = \|y\|_2 - 1$, the signed distance to $\Gamma(0)$. This does not
suit our purposes, because its Hessian is singular and its dual is the
indicator function of the $l_2(\Rn)$ unit ball. Instead we take
\begin{displaymath}
L(y) = \frac{1}{2} \left(\|y\|_2^2 - 1\right).
\end{displaymath}
Note $L^*(y) \; =\; \frac{1}{2} \left(\|y\|_2^2 + 1\right)$, both of
these are convex and $\mathcal{C}^2$ functions with non vanishing
gradients away from the origin, i.e. near $\Gamma(s) $. This gives us
a hint as how to proceed.
Recall that we need to get initial data which behaves as a level set function
should; i.e., as defined by \eqref{eq.L_level_set_fctx}. We also want
either $L$ or $L^*$ to be smooth enough, actually twice differentiable
with Lipschitz continuous Hessian, so that Newton's method can be
used.
We might take
$$
L( R\, \theta)\;=\; \frac{1}{2m}
\left(
\left(\frac{R}{W\left(\theta\right)}\right)^{2m} -1
\right),
$$
where $R \geq 0$, $\theta \in \Snm1$ and $m$ a positive integer. If we
consider the important case of $W(\theta) \,=\, \|\theta\|_A
^{-\frac{1}{2}} \,=\, \frac{1}{\sqrt{\langle \theta, A\, \theta
\rangle}}$, with $A$ a symmetric positive definite matrix, which
corresponds to $\Omega = \{x \in \Rn \,|\, \sqrt{\langle x, A\, x \rangle}
\,\leq\, 1\}$, then $m\geq2$ will have a smooth enough Hessian for $L$. In
fact, $m=2$
will lead us to a linear Hessian. There will be situations where using
$L^*$ is preferable because $L$ cannot be made smooth enough this way.
We can obtain $L^*$ using polar coordinates and taking
$1 \geq m > \frac{1}{2}$, that is for any $R\geq 0$ and any~$\theta \in \Snm1$
\begin{align}
\nonumber
L^*(R\, \theta)
\;=\;&
\displaystyle
\sup_{r\geq 0,\, v \in \mathcal{S}^{n-1}}
\left\{
R\, r\, \langle \theta, v\rangle + \frac{1}{2m} - \frac{1}{2m}
\left( \frac{r}{W(v)} \right)^{2m}
\right\}
\\
\nonumber
\;=\;&
\frac{1}{2m} \sup_{\|v\|_2=1}
\left\{
(2m-1)
\left( R \langle \theta, v\rangle W (v) \right)^{\frac{2m}{2m-1}} +1
\right\}
\\
\nonumber
\;=\;&
\frac{1}{2m}
\left\{
\left( R\, \gamma\left( \theta \right)\right)^{\frac{2m}{2m-1}}
(2m-1) +1
\right\}
\\
\nonumber
\;=\;&
\frac{1}{2m}
\left\{
\left( H\left( R\,\theta \right) \right)^{\frac{2m}{2m-1}}
(2m-1) +1
\right\}
\end{align}
where we recall that $\gamma$ is defined by \eqref{eq.gamma}.
As we shall see, $L^*$ is often preferable to $L$ yielding a
smooth Hessian for $ 1 \geq m > \frac{1}{2}$ with $m$ close to $\frac{1}{2}$.
Let us consider some important examples. If the set $\Omega$ is defined
by $\Omega = \{x\in\Rn \,|\, \|x\|_p \leq 1\}$ for $1 < p < \infty$,
then we can consider two cases
\begin{itemize}
\item[(a)] $2 \leq p < + \infty$
\item[(b)] $1 < p \leq 2$
\end{itemize}
For case $(a)$ we need only take
$$
L \;=\; \frac{1}{2m}
\left(
\|\cdot\|_p^{2m} -1
\right).
$$
If $m \geq 1$ it is easy to see that Hessian of $L$ is continuous and
bounded for $p\geq 2$. So we can use Newton's method for the
first choice above.
For case $(b)$, we construct the Fenchel-Legendre of the function
$$
L \;=\; \frac{1}{2m} \left( \|\cdot\|_p^{2m} - 1\right),
$$
but this time for $\frac{1}{2} < m \leq 1$. It is easy to
see that
$$
L^* \;=\; \frac{1}{2m}
\left(
(2m-1)\|\cdot\|_q^{\frac{2m}{2m-1}} + 1
\right),
$$
where $\frac{1}{p} + \frac{1}{q} = 1$ and if $\frac{1}{2} < m \leq 1$
the Hessian of $L^*$ is continuous.
Other interesting examples include the following regions defined by
$$
\Omega \;=\; \left\{
x \in \Rn \;|\; \langle x, Ax \rangle \leq \|x\|_2
\right\}.
$$
For $\Omega$ to be convex we require $A$ to be a positive definite
symmetric matrix with real entries and its maximal eigenvalue is
bounded by twice the minimal eigenvalue.
We can take
\begin{equation}
\label{eq.def_L_quad_over_norm}
L(x) \;= \;
\begin{cases}
\frac{1}{2m}
\left(
\left(
\frac{\langle x,\, A x
\rangle}{\|x\|_2} \right)^{2m}
-1 \right) & \mbox { if } x \neq 0,\\
0 & \mbox { if } x = 0.
\end{cases}
\end{equation}
for $m\geq 2$ and see that this has a smooth Hessian.
\section{Numerical Results}
\label{sec.numerical_results}
We will consider the following Hamiltonians
\begin{itemize}
\item $H = \|\cdot\|_p$ for $p = 1, 2, \infty$,
\item $H = \sqrt{\langle \cdot, A \cdot\rangle}$ with $A$ symmetric positive
definite matrix,
\end{itemize}
and the following initial data
\begin{itemize}
\item $J = \frac{1}{2}\|\cdot\|_p^2$ for $p = 1, 2, \infty$,
\item $J = \frac{1}{2}\langle \cdot, A\cdot \rangle$ with $A$ a
positive definite diagonal matrix.
\end{itemize}
It will be useful to consider the spectral decomposition of $A$, i.e.,
$A = P D P^\dagger$ where $D$ is a diagonal matrix, $P$ is an
orthogonal matrix and $P^\dagger$ denotes the transpose of $P$. The
identity matrix in $\Rn$ is denoted by $I_n$.
First, we present the algorithms to compute the proximal points for
the above Hamiltonians and initial data. We shall describe these
algorithms using the following generic formulation for the proximal map
$$
\prox{\alpha \,\partial f} (w)
\;=\;
\arg \min_{w \in \Rn}
\left\{
\frac{1}{2} \|w-z\|_2^2 +
\alpha f(w)
\right\}.
$$
Second, we present the time results on a standard laptop. Some time
results are also provided for a 16 cores computer which shows that our
approach scales very well. The latter is due to
our very low memory
requirement. Finally some plots that represent the
solution of some HJ PDE are presented.
\subsection{Some explicit formulas for simple specific cases}
We are able to obtain explicit formulas for the proximal map for some
specific cases of interest in this paper. For instance, as we have
seen, considering $f = \| \cdot \|_1$ gives for any $i=1,\dots,n$
$$
\left(\prox{\alpha \,\partial \|\cdot \|_1}(w)\right)_i
\;=\; \mbox{sign}(w_i) \max(|w_i|-\alpha, 0),
$$
where $\mbox{sign}(\beta) = 1$ if $\beta\geq 0$ and $-1$ otherwise.
The case $f=\|\cdot\|_2$ yields a similar formula
$$
\prox{\alpha \,\partial \|\cdot \|_2}(w)
\;=\;
\frac{\mu}{\alpha + \mu} x,
$$
with
$$
\mu = \max(\|w\|_2 - \alpha,0).
$$
The two above cases are computed in linear time with respect to the
dimension $n$.
Proximal maps for positive definite quadratic forms, i.e., $f(w) =
\frac{1}{2}\langle w, A w\rangle$ are also easy to compute since for
any $\alpha >0$ and any $z\in\Rn$
$$
\prox{ \alpha \,\partial f}(z) \;=\; ( I_n + \alpha A)^{-1} (z)
\;=\;
P (I_n + \alpha D)^{-1} P^\dagger z.
$$
where we recall that $A = P D P^\dagger$ with $D$ a diagonal matrix
and $P$ an orthogonal matrix. The time complexity is dominated by the
evaluation of the matrix-vector product involving $P$ and
$P^{\dagger}$.
\subsection{The case of $\|\cdot \|_\infty$ }
\label{subsec.prox_infty}
Let us now consider the case $f=\|\cdot\|_\infty$. Since
$\|\cdot\|_\infty$ is a norm, its Fenchel-Legendre transform is the
indicator function of its dual norm ball,
that is $(\|\cdot\|_\infty)^* \;=\;
\mathcal{I}_C$ with $C \;=\; \{ z \in \Rn \;|\;
\|z\|_1 \leq 1\}$. We use Moreau's identity
\eqref{eq.prox-projection} to compute $\prox{\alpha \|
\cdot\|_\infty}$; that is for any $\alpha>0$ and for any $z \in \Rn$
$$
\prox{\alpha \partial\,\| \cdot\|_\infty}(z)
\;=\; z - \alpha \pi_C(\frac{z}{\alpha}) = z - \pi_{\alpha C} (z)
$$
where we recall that $\pi_{\alpha C}$ denotes the projection operator onto
the closed convex set $\alpha C$. We use a simple variation of parametric
approaches that are well-known in graph-based optimization algorithm
(see \cite[Chap. 11, Section 11.M]{rockafellar.84.book} for instance).
Let us assume that $z \notin (\alpha C)$. The projection corresponds
to solve
$$
\pi_{\alpha C} (z) \;=\;
\begin{dcases}
\arg \min_{w\in\Rn} \frac{1}{2}\|w-z\|_2^2 \\
\mbox{s.t. } \|w\|_1 \leq \alpha.
\end{dcases}
$$
Now we use Lagrange duality (see
\cite[chap. VII]{hiriart-lemarechal.96.book-vol1} for instance). The
Lagrange dual function $g: [0,+\infty) \to \R$ is defined by
$$
g(\mu)
\;=\;
\min_{w\in\Rn}\left\{ \frac{1}{2} \| z -w \|_2^2 +
\mu \|w\|_1 - \mu \alpha \right\},
$$
that is
$$
g(\mu)
\;=\;
\| \mbox{shrink}_1(z, \mu)\|_1 - \mu \alpha.
$$
We denote by $\bar{\mu}$ the value that realizes the
maximum of $g$. Then, we obtain for any $z
\notin \alpha C$ that
\begin{equation}
\label{eq.projection_mu_linfty}
\pi_{\alpha C}(z) \;=\;
\mbox{shrink}_1(z, \bar{\mu}),
\end{equation}
where $\bar{\mu} $
satisfies
\begin{equation}
\label{eq.condition_mu_linfty}
\|\mbox{shrink}_1(z,\bar{\mu})\|_1 \;=\; \alpha.
\end{equation}
Computing the projection is thus reduced to computing the optimal
value of the Lagrange multiplier $\bar{\mu}$.
Consider the function $h : [0,\|z\|_1] \to \R$ defined by $ h(\mu)
\;=\; \|\mbox{shrink}_1(z, \mu)\|_1$. We have that $h$ is continuous,
piecewise affine, $h(0)=\|z\|_1$, $h(\|z\|_1) = 0$ and $h$ is
decreasing (recall that we assume $z \notin (\alpha C)$). Following
\cite[Chap. 11, Section 11.M]{rockafellar.84.book}, we call
breakpoints the values for which $h$ is
not differentiable. The set of breakpoints for $h$ is $B = \left\{0
\right\} \bigcup_{i=1,\dots,n} \left\{ |z_i|\right\}.$ We sort all
breakpoints in increasing order and we denote this sequence by $(l_i,
\dots, l_{m}) \in B^{m}$ with $l_i < l_{i+1}$ for $i=1,\dots,(m-1)$, where $m
\leq n$ is the number of breakpoints. This operation takes $O(n \log
n)$. Then, using a bitonic search, we can find $j$ such that
$\|\shrink_1(z, l_j)\|_1 \leq \|\shrink_1(z, \bar{\mu})\|_1 <
\|\shrink_1(z, l_{j+1})\|_1$ in $O(n \log n)$. Since $h$ is affine on
$[\|\shrink_1(z, l_j)\|_1, [\|\shrink_1(z, l_{j+1})\|_1]$ a simple
interpolation computed in constant time yields $\bar{\mu}$ that
satisfies \eqref{eq.condition_mu_linfty}. We then use
\eqref{eq.projection_mu_linfty} to compute the projection. The overall
time complexity is, therefore, $O(n \log n)$.
\subsection{The case $\|\cdot\|_A$ and projection
on a ellipsoid}
\label{subsec.projection_ellipsoid}
We follow the same approach as for $\|\cdot\|_\infty$. We consider $f
= \| \cdot \|_A = \langle \cdot, A \cdot \rangle$ which is a norm
since $A$ is assumed to be symmetric positive definite. The dual norm
is $\|\cdot\|$ \cite[Prop. 4.2, p. 19]{ekeland.76.book} is
$\| \cdot\|_{A ^{-1}}$.
Thus $\left(\| \cdot\|_{A}\right)^* = I_{\mathcal{E}_A}$ with
$\mathcal{E}_A$ defined by
$$
\mathcal{E}_A
\;=\; \left\{ y \in \Rn \;|\; \langle y, A ^{-1} y \rangle \leq 1\right\}.
$$
Using Moreau's identity \eqref{eq.prox-projection} we only need to
compute the
projection $\pi_{\mathcal{E}_{A}}(w)$ of $w\in \Rn$ on the ellipsoid
$\mathcal{E}_{A} $.
Note that we have $\pi_{\mathcal{E}_{A}} (w) = P\,
\pi_{\mathcal{E}_{D }} (P^\dagger w)$ where we recall that
$A=PDP^\dagger$ with $D$ and $P$ a diagonal and orthogonal matrix,
respectively. Thus we only describe the algorithm for the projection
on an ellipsoid involving positive definite diagonal matrices.
To simplify notation we take $d_i = D_{ii}$ for $i=1,\dots,n$.
We consider the ellipsoid $\mathcal{E_D}$ defined by
$$
\mathcal{E}_{D} \;=\;
\left\{
x \in \mathbb{R}^n\, |\, \sum_{i=1}^{n} \left(
\frac{x_i}{d_i}\right)^2 \leq 1
\right\}.
$$
Let $w \notin \mathcal{E}_D$.
We can easily show (see \cite[Exercise III.8]{hiriart.98.book} for
instance) that $\Pi_{\mathcal{E}_D}(w)$ satisfies for any
$i=1.\dots,n$
\begin{equation}
\label{eq.lagrange_multiplier_equation_ellipsoid}
\left(\Pi_{\mathcal{E}_D}(w)\right)_i \;=\;
\frac{d_i^2 w_i}{d_i^2 + \bar{\mu}},
\end{equation}
where the Lagrange multiplier $\bar{\mu} > 0$ is the unique solution
of $\sum_{i=1}^n \frac{d_i^2 w_i^2}{(d_i^2 + \mu)^2} = 1$. We find
such $\bar{\mu}$ by minimizing the function $[0,+\infty) \ni \mu
\mapsto \sum_{i=1}^{n} d_i ^2 w_i ^2 (d_i ^2 + \mu)^{-1} + \mu$ using
Newton's method which generates a sequence $(\mu_k)_{k \in \N}$
converging to $\bar{\mu}$. We set the initial value to $\mu_0 = 0$ and
we stop Newton's iterations for the first $k$ which statisfies
$|\mu_{k+1} - \mu_{k}| \leq 10^{-8}$. Once we have the value for $\bar{\mu}$ we
use \eqref{eq.lagrange_multiplier_equation_ellipsoid} to obtain the
approximate projection.
\subsection{The cases $\frac{1}{2}\|\cdot \|_1^2$ and
$\frac{1}{2}\|\cdot\|_\infty^2$}
\label{subsec.l1_linfty_squared}
First, we consider the case of $\frac{1}{2}\|\cdot \|_1^2$. We have
for any $\alpha >0$ and any $z\in \Rn$
$$
\prox{\frac{\alpha}{2}\, \partial \|\cdot \|_1^2} (z)
\;=\;
\prox{\alpha \|\cdot\|_1 \,\partial\|\cdot\|_1}(z).
$$
Thus, assuming there exists $\bar{\beta} \geq 0$ such that
\begin{equation}
\label{eq.optimality_beta_l1_2}
\bar{\beta} \;=\;
\alpha \left\| \shrink_1(z,\bar{\beta}) \right\|_1,
\end{equation}
we have for any $\alpha >0$ and any $z\in \Rn$
\begin{equation}
\label{eq.prox_l1_2_with_beta}
\prox{\frac{\alpha}{2} \,\partial \|\cdot \|_1^2} (z)
\;=\; \shrink_1(z, \bar{\beta}).
\end{equation}
The existence of $\bar{\beta}$ and an algorithm to compute
it follow.
Let us assume that $z\neq 0$ (otherwise, $\bar{\beta}=0$ works and the
solution is of course 0). Then, consider the function
$g:\left[0,\|z\|_1\right] \to \R$ defined by
$$
g(\beta) \;=\;
\alpha \|\shrink_1(z,\beta)\|_1 - \beta.
$$
It is continuous, and $g(0)= \alpha \|z\|_1$ while
$g\left(\|z\|_1\right) = -\|z\|_1$. The intermediate value theorem
tells us that there exists $\bar{\beta}$ such that $g(\bar{\beta}) =
0$, that is, satisfying \eqref{eq.optimality_beta_l1_2}.
The function $g$ is decreasing, piecewise affine and the breakpoints
of $g$ (i.e, the points where $g$ is not differentiable) are $B= \{0\}
\cup_{i=1,\dots,n} \{ |z_i|\}$. We now proceed similarly as for the
case $\|\cdot\|_\infty$. We note $(l_i,\dots, l_{m}) \in B^{m}$ the
breakpoints sorted in increasing order, i.e., such that $l_i <
l_{i+1}$ for $i=1,\dots,(m-1)$, where $m \leq n$ is the number of
breakpoints. We use a bitonic search to find the two consecutive
breakpoints $l_i$ and $l_{i+1}$, such that
$g(l_i) \;\geq\; 0 \;>\; g(l_{i+1})$. Since $g$ is affine on $[l_i,
l_{i+1}]$ a simple interpolation yields the value $\bar{\beta}$. We
then compute $\prox{\frac{\alpha}{2} \partial \|\cdot\|_1^2}(z)$
using \eqref{eq.prox_l1_2_with_beta}.
\medskip
We now consider the case $\frac{1}{2}\|\cdot\|_\infty^2$. We have
(for instance \cite[Prop. 4.2, p. 19]{ekeland.76.book})
$$
\left(\frac{1}{2}\| \cdot \|_\infty^2\right)^*
\;=\;
\frac{1}{2}\| \cdot \|_1^2
$$
Then Moreau's identity \eqref{eq.moreau-identity} yields for any
$\alpha >0 $ and for any $z \in \Rn$
$$
\prox{\frac{\alpha}{2}\,\partial \|\cdot\|_\infty^2}(z)
\;=\;
z - \prox{\frac{\alpha}{2}\,\partial \|\cdot\|_1^2}(z),
$$
which can be easily computed using the above algorithm for evaluating
$\prox{\frac{\alpha}{2}\,\partial \|\cdot\|_1^2}(z)$.
\subsection{Time results and illustrations}
We now give numerical results for several Hamiltonians and initial
data. We present time results on a standard laptop using a
single core which show that our approach allows us to evaluate very
rapidly HJ PDE solutions. We also present some time results on a
16 cores computer to show that are approach scales very well.
We also present some plots that depict the solution of some HJ PDEs.
We recall that we consider the following Hamiltonians
\begin{itemize}
\item $H = \|\cdot\|_p$ for $p = 1, 2, \infty$,
\item $H = \sqrt{\langle \cdot, D \cdot\rangle}$ with $D$ a diagonal positive
definite matrix,
\item $H = \sqrt{\langle \cdot, A \cdot\rangle}$ with $A$ symmetric positive
definite matrix,
\end{itemize}
and the following initial data
\begin{itemize}
\item $J = \frac{1}{2}\|\cdot\|_p^2$ for $p = 1, 2, \infty$.
\item $J = \frac{1}{2}\langle \cdot, D^{-1}\cdot \rangle$ with $D$ a
positive definite diagonal matrix,
\end{itemize}
where the matrix $D$ and $A$ are defined follows: $D$ is a diagonal
matrix of size $n\times n$ defined by $D_{ii} = 1 + \frac{i-1}{n-1}$
for $i=1,\dots,n$. The symmetric positive definite matrix $A$ of size
$n\times n$ is defined by $A_{ii} = 2$ for $i=1,\dots,n$
and $A_{ij} = 1 $ for $i,j=1,\dots,n$ with $i\neq j$.
All computations are performed using IEEE double precision
floating-points where denormalized number mode has been disabled.
The quanties $(x,t)$ are drawn uniformly in $[-10,10]^n \times
[0,10]$. We present the average time to evaluate a solution for
$1,000,000$ runs.
We set $\lambda =1$ in the split Bregman algorithm
\eqref{3.4(a)}-\eqref{3.4(c)}. We stop the iterations when the
following stopping criteria is met:
$\| v^k - v^{k-1}\|_2^2 \leq 10^{-8}$
and
$\| d^k - d^{k-1}\|_2^2 \leq 10^{-8}$ and
$\| d^k - v^{k}\|_2^2 \leq 10^{-8}$.
We first carry out the numerical experiments on an Intel Laptop Core
i5-5300U running at 2.3~GHz. The implementation here is {\it single}
threaded, i.e., only one core is used. Tables \ref{table.l2_2},
\ref{table.linfty_2}, \ref{table.l1_2} and \ref{table.diag_2} present
time results for several dimensions $n=4,8,12,16$ and with initial
data $J = \frac{1}{2}\|\cdot\|_2^2$, $J =
\frac{1}{2}\|\cdot\|_\infty^2$, $J = \frac{1}{2}\|\cdot\|_1^2$ and $J
= \frac{1}{2}\langle \cdot, D \rangle$ respectively. We see that it
takes about $10^{-8}$ to $10^{-4}$ seconds per evaluation of the
solution.
We now consider experiments that are carried out on a computer with 2
Intel Xeon E5-2690 processors running at 2.90GHz. Each processor has 8
cores. Table \ref{table.parallel} present the average time to compute
the solution with Hamiltonian $H=\|\cdot\|_\infty$ and initial data
$J=\|\cdot\|_1^2$ for
several dimensions and various number of used cores. We see that our
approach scales very well. This is due to the fact that our algorithm
requires little memory which easily fits in the L1 cache of each
processor. Therefore cores are not competing for resources. This
suggests that our approach is suitable for low-energy embedded
systems.
We now consider solutions of HJ PDEs in dimension $n=8$
on a 2-dimensional grid. We evaluate $\phi(x_1, x_2,0,0,0,0,0,0)$
with $x_i \in \cup_{k=0,\dots,99}\{-20 + k \frac{40}{99} \}$ for
$i=1,2$.
Figures \ref{fig.J_linfty_2__H__l2}, \ref{fig.J_l1_2__H__l1} and
\ref{fig.J_l1_2__H__DIAG_NORM} depict the solutions with initial data
$J=\frac{1}{2}\|\cdot\|_\infty^2$,
$J=\frac{1}{2}\|\cdot\|_1^2$ and $J=\frac{1}{2}\|\cdot\|_1^2$, and
with Hamiltonians $H= \|\cdot\|_2$, $H = \|\cdot\|_1$.
and $H=\sqrt{\langle \cdot, D \cdot\rangle}$ for various times,
respectively.
Figure~\ref{fig.min_initial_data} and
Figure~\ref{fig.min_hamiltonians} illustrate the max/min-plus
algebra results described in
Section~\ref{subsec.extension_future_work}. Figure~\ref{fig.min_initial_data} depicts the HJ solution for the initial data $J =
J= \min{}\left(\frac{1}{2}\|\cdot\|_2^2 - \langle b, \cdot\rangle,
\frac{1}{2}\|\cdot\|_2^2 + \langle b, \cdot\rangle\right)$ with
$b=(1,1,1,1,1,1,1,1)^\dagger$, and $H=\|\cdot\|_1$ for various
times. Figure~\ref{fig.min_hamiltonians} depicts the HJ solution
for various time with
$J= \frac{1}\|\cdot\|_2^2$ and
$H=\min{}\left(\|\cdot\|_1, \sqrt{\langle \cdot,
\frac{4}{3}D \cdot\rangle} \right) $.
\begin{table}[ht]
\begin{center}
\begin{tabular}{r|c|c|c|c|c}
n & $\|y\|_1$ & $\|y\|_2$ & $\|y\|_\infty$ & $\|y\|_D$ & $\|y\|_A$\\
\hline
4 & 6.36e-08 & 1.20e-07 & 2.69e-07 & 7.00e-07 & 8.83e-07 \\
8 & 6.98e-08 & 1.28e-07 & 4.89e-07 & 1.07e-06 & 1.57e-06 \\
12 & 8.72e-08 & 1.56e-07 & 7.09e-07 & 1.59e-06 & 2.23e-06 \\
16 & 9.24e-08 & 1.50e-07 & 9.92e-07 & 2.04e-06 & 2.95e-06 \\
\end{tabular}
\end{center}
\caption{Time results in seconds for the average time per call for
evaluting the solution of the HJ-PDE with the initial data $J =
\frac{1}{2}\|\cdot\|_2^2$, several Hamiltonians and
various dimensions~$n$.}
\label{table.l2_2}
\end{table}
\begin{table}[ht]
\begin{center}
\begin{tabular}{r|c|c|c|c|c}
n & $\|y\|_1$ & $\|y\|_2$ & $\|y\|_\infty$ & $\|y\|_D$ & $\|y\|_A$\\
\hline
4 & 1.79e-06 & 1.53e-06 & 1.84e-06 & 4.88e-06 & 7.77e-06 \\
8 & 3.77e-06 & 2.31e-06 & 3.50e-06 & 9.73e-06 & 1.92e-05 \\
12 & 6.31e-06 & 3.14e-06 & 5.54e-06 & 1.44e-05 & 2.91e-05 \\
16 & 9.61e-06 & 3.88e-06 & 8.22e-06 & 1.80e-05 & 4.04e-05 \\
\end{tabular}
\end{center}
\caption{Time results in seconds for the average time per call for
evaluting the solution of the HJ-PDE with the initial data $J =
\frac{1}{2}\|\cdot\|_\infty^2$, several Hamiltonians and various
dimensions~$n$.}
\label{table.linfty_2}
\end{table}
\begin{table}[ht]
\begin{center}
\begin{tabular}{r|c|c|c|c|c}
n & $\|y\|_1$ & $\|y\|_2$ & $\|y\|_\infty$ & $\|y\|_D$ & $\|y\|_A$\\
\hline
4 & 2.86e-06 & 4.42e-06 & 9.17e-06 & 1.79e-05 & 1.97e-05 \\
8 & 9.85e-06 & 1.63e-05 & 4.38e-05 & 9.37e-05 & 1.09e-04 \\
12 & 2.35e-05 & 3.84-05 & 1.19e-04 & 2.63e-04 & 3.24e-04 \\
16 & 4.35e-05 & 7.03e-05 & 2.46e-04 & 5.19e-04 & 6.92e-04 \\
\end{tabular}
\end{center}
\caption{Time results in seconds for the average time per call for
evaluting the solution of the HJ-PDE with the initial data $J =
\frac{1}{2}\|\cdot\|_1^2$, several Hamiltonians and
various dimensions $n$.}
\label{table.l1_2}
\end{table}
\begin{table}[ht]
\begin{center}
\begin{tabular}{r|c|c|c|c|c}
n & $\|y\|_1$ & $\|y\|_2$ & $\|y\|_\infty$ & $\|y\|_D$ & $\|y\|_A$\\
\hline
4 & 3.62e-07 & 5.19e-07 & 9.35e-07 & 2.79e-06 & 3.50e-06 \\
8 & 3.83e-07 & 5.25e-07 & 1.42e-06 & 4.40e-06 & 5.75e-06 \\
12 & 4.97e-07 & 6.62e-07 & 1.73e-06 & 5.70e-06 & 7.98-06 \\
16 & 5.92e-07 & 6.88e-07 & 2.27e-06 & 6.64e-06 & 1.04e-05 \\
\end{tabular}
\end{center}
\caption{Time results in seconds for the average time per call for
evaluting the solution of the HJ-PDE with the initial data $J =
\frac{1}{2}\langle\cdot, D^{-1}\cdot \rangle$, several Hamiltonians and
various dimensions $n$.}
\label{table.diag_2}
\end{table}
\begin{table}[h]
\begin{center}
\begin{tabular}{r|c|c|c|c|c}
n & 1 core & 4 cores & 8 cores & 16 cores\\
\hline
4 & 1.11e-05 & 2.81e-06 & 1.56e-06 & 8.36e-07 \\
8 & 4.77e-05 & 1.33e-05 & 6.81e-06 & 3.48e-06 \\
12 & 1.35e-04 & 3.90e-05 & 1.94e-05 & 9.90e-06 \\
16 & 3.24e-04 & 8.76e-05 & 4.40e-05 & 2.22e-05 \\
\end{tabular}
\end{center}
\caption{Time results in seconds for the average time per call for
evaluting the solution of the HJ-PDE with the initial data $J =
\frac{1}{2}\|\cdot\|_1^2$, and the Hamiltonian
$H=\|\cdot\|_\infty$, for various dimensions, and several cores.}
\label{table.parallel}
\end{table}
\begin{figure}[t
\begin{minipage}[b]{.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{J_linfty_2__H__l2__time_0}}
\vspace{-.1cm}
\centerline{\footnotesize{(a)}}\medskip
\end{minipage}
\hfill
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{J_linfty_2__H__l2__time_5}}
\vspace{-.1cm}
\centerline{\footnotesize{(b)}}\medskip
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{J_linfty_2__H__l2__time_10}}
\vspace{-.1cm}
\centerline{\footnotesize{(c)}}\medskip
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{J_linfty_2__H__l2__time_15}}
\vspace{-.1cm}
\centerline{\footnotesize{(d)}}\medskip
\end{minipage}
\caption{ Evaluation of the solution $\phi((x_1,x_2,
0,0,0,0,0,0)^\dagger,t)$ of
the HJ-PDE with initial data $J=\frac{1}{2}\|\cdot\|_\infty^2$
and Hamiltonian $H=\|\cdot\|_2$ for $(x_1,x_2) \,\in\, [-20,20]^2$ for
different times $t$.
Plots for $t=0, 5,10,15$ and respectively depicted in (a),
(b), (c) and (d). The level lines multiple of 5 are superimposed
on the plots.}
\label{fig.J_linfty_2__H__l2}
\end{figure}
\begin{figure}[ht
\begin{minipage}[b]{.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{J_l1_2__H__l1__time_0}}
\vspace{-.1cm}
\centerline{\footnotesize{(a)}}\medskip
\end{minipage}
\hfill
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{J_l1_2__H__l1__time_5}}
\vspace{-.1cm}
\centerline{\footnotesize{(b)}}\medskip
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{J_l1_2__H__l1__time_10}}
\vspace{-.1cm}
\centerline{\footnotesize{(c)}}\medskip
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{J_l1_2__H__l1__time_15}}
\vspace{-.1cm}
\centerline{\footnotesize{(d)}}\medskip
\end{minipage}
\caption{
Evaluation of the solution $\phi((x_1,x_2,0,0,0,0,0,0)^\dagger,t)$ of
the HJ-PDE with initial data $J=\frac{1}{2}\|\cdot\|_1^2$
and Hamiltonian $H=\|\cdot\|_1$ for $(x_1,x_2) \,\in\, [-20,20]^2$ for
different times $t$. Plots for $t=0, 5,10,15$ are
respectively depicted in (a), (b), (c) and (d). The level
lines multiple of 20 are superimposed on the plots.}
\label{fig.J_l1_2__H__l1}
\end{figure}
\begin{figure}[th
\begin{minipage}[b]{.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{J_l1_2__H__DIAG_NORM__time_0}}
\vspace{-.1cm}
\centerline{\footnotesize{(a)}}\medskip
\end{minipage}
\hfill
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{J_l1_2__H__DIAG_NORM__time_5}}
\vspace{-.1cm}
\centerline{\footnotesize{(b)}}\medskip
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{J_l1_2__H__DIAG_NORM__time_10}}
\vspace{-.1cm}
\centerline{\footnotesize{(c)}}\medskip
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{J_l1_2__H__DIAG_NORM__time_15}}
\vspace{-.1cm}
\centerline{\footnotesize{(d)}}\medskip
\end{minipage}
\caption{
Evaluation of the solution $\phi((x_1,x_2,0,0,0,0,0,0)^\dagger,t)$ of
the HJ-PDE with initial data $J=\frac{1}{2}\|\cdot\|_1^2$
and Hamiltonian $H=\|\cdot\|_D$, for $(x_1,x_2) \,\in\,
[-20,20]^2$ for different times $t$.
Plots for $t=0, 5,10,15$ are
respectively depicted in (a), (b), (c) and (d). The level
lines multiple of 20 are superimposed on the plots.
}
\label{fig.J_l1_2__H__DIAG_NORM}
\end{figure}
\begin{figure}[th
\begin{minipage}[b]{.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{min_l2_2mb_l2_2_pb__H__l1_time_0}}
\vspace{-.1cm}
\centerline{\footnotesize{(a)}}\medskip
\end{minipage}
\hfill
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{min_l2_2mb_l2_2_pb__H__l1_time_5}}
\vspace{-.1cm}
\centerline{\footnotesize{(b)}}\medskip
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{min_l2_2mb_l2_2_pb__H__l1_time_10}}
\vspace{-.1cm}
\centerline{\footnotesize{(c)}}\medskip
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{min_l2_2mb_l2_2_pb__H__l1_time_15}}
\vspace{-.1cm}
\centerline{\footnotesize{(d)}}\medskip
\end{minipage}
\caption{
Evaluation of the solution $\phi((x_1,x_2,0,0,0,0,0,0)^\dagger,t)$ of
the HJ-PDE with initial data $J=
\min{}\left(\frac{1}{2}\|\cdot\|_2^2 - \langle b, \cdot\rangle,
\frac{1}{2}\|\cdot\|_2^2 + \langle b, \cdot\rangle\right)$ with
$b=(1,1,1,1,1,1,1,1)^\dagger$ and Hamiltonian $H=\|\cdot\|_1$, for
$(x_1,x_2) \,\in\, [-20,20]^2$ for different times $t$.
Plots for $t=0, 5,10,15$ are
respectively depicted in (a), (b), (c) and (d). The level
lines multiple of 15 are superimposed on the plots.
}
\label{fig.min_initial_data}
\end{figure}
\begin{figure}[th
\begin{minipage}[b]{.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{l2_2___H__min_l1_D_time_2}}
\vspace{-.1cm}
\centerline{\footnotesize{(a)}}\medskip
\end{minipage}
\hfill
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{l2_2___H__min_l1_D_time_5}}
\vspace{-.1cm}
\centerline{\footnotesize{(b)}}\medskip
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{l2_2___H__min_l1_D_time_9}}
\vspace{-.1cm}
\centerline{\footnotesize{(c)}}\medskip
\end{minipage}
\begin{minipage}[b]{0.49\linewidth}
\centering
\centerline{\includegraphics[scale=0.35]{l2_2___H__min_l1_D_time_12}}
\vspace{-.1cm}
\centerline{\footnotesize{(d)}}\medskip
\end{minipage}
\caption{
Evaluation of the solution $\phi((x_1,x_2,0,0,0,0,0,0)^\dagger,t)$ of
the HJ-PDE with initial data $J=\frac{1}{2}\|\cdot\|_1^2$
and Hamiltonian $H=\min{}\left(\|\cdot\|_1, \sqrt{\langle \cdot,
\frac{4}{3}D \cdot\rangle} \right) $, for $(x_1,x_2) \,\in\,
[-20,20]^2$ for different times $t$.
Plots for $t=2, 5,9,12$ are
respectively depicted in (a), (b), (c) and (d). The level
lines multiple of 15 are superimposed on the plots.
}
\label{fig.min_hamiltonians}
\end{figure}
\newpage
\section{Conclusion}
\label{sec.conclusion}
We have designed algorithms which enable us to solve certain
Hamilton-Jacobi equations very rapidly. Our algorithms not only
evaluate the solution but also computes the gradient of the
solution. These include equations
arising in control theory leading to Hamiltonians which are convex and
positively homogeneous of degree $1$. We were motivated by ideas coming
from compressed sensing; we borrowed algorithms devised to solve
$\ell_1$ regularized problems which are known to rapidly converge.
We apparently extended this fast convergence to include convex
positively 1-homogeneous regularized problems.
There are no grids involved. Instead of complexity which is
exponential in the dimension of the problems, which is typical of grid
based methods, ours appears to be polynomial in the dimension with
very small constants. We can evaluate the solution on a laptop at
about $10^{-4} - 10^{-8}$ seconds per evaluation for fairly high
dimensions. Our algorithm requires very low
memory and is totally parallelizable which suggest that it is
suitable for low energy embedded systems.
We have chosen to restrict the presentation of the numerical
experiments to norm-based Hamiltonians and we
emphasize that our approach naturally extends to more elaborate
positively 1-homogeneous Hamiltonians (using the min/max algebra
results as we did for instance).
As an important step in this procedure we have also derived an equally
fast method to find a closest point
lying on $\Omega$, a finite union of compact convex sets $\Omega_i$,
such that $\Omega = \cup_{i}^{k}\Omega_i$ has a nonempty
interior, to a given point.
We can also solve certain so called fast marching
\cite{tsitsiklis.95.itac} and fast sweeping \cite{TCOZ} problems equally
rapidly in high dimensions. If we wish to
find $\psi:\Rn\to\R$ with, say $\psi=0$ on the boundary of a set
$\Omega$ defined above, satisfying
$$
\begin{cases}
\|\nabla_x \psi(x)\|_2 \;=\; 1 & \mbox{ in } \Rn,\\
\psi(x) \;=\; 0 & \mbox{ for any } x \in
(\Omega \setminus \interior{\Omega}),
\end{cases}
$$
then, we can solve for $u:\Rn\times
[0,+\infty) \to \R$
$$
\frac{\partial u}{\partial t} (x,t)
\,+\, \|\nabla_x u(x)\|_2 \;=\; 0 \quad \mbox{ in } \Rn \times
(0,+\infty),
$$
with initial data
$$
\begin{cases}
u(x,0) < 0 \mbox{ for any } x \in \mbox{int } \Omega, \\
u(x,0) > 0 \mbox{ for any } x \in (\Rn \setminus \Omega),\\
u(x,0) = 0 \mbox{ for any } x \in (\Omega \setminus \mbox{int } \Omega),
\end{cases}
$$
and locate the zero level set of $u(\cdot, t) = 0$ for any given
$t>0$. Indeed any $x \in \{y \in \Rn \;|\; u(y,t) \;=\;0\}$
satisfies $\psi(x) = t$.
Of course the same approach could be used for any
convex, positively 1-homogeneous Hamiltonian $H$ (instead of
$\|\cdot\|_2$), e.g., $H \,=\, \|\cdot\|_1$.
This will give us results related to computing the Manhattan distance.
We expect to extend our work as follows:
\begin{enumerate}
\item We will do experiments involving linear controls, allowing $x$
and $t>0$ dependence while the Hamiltonian $(p,x,t) \mapsto H(p,x,t)$ is
still convex and positively 1-homogeneous in $p$. The procedure
was described in section \ref{sec.introduction_optimal_control}.
\item We will extend our fast computation of the projection in several
ways. We will consider in detail the case of polyhedral regions
defined by the intersection of sets $\Omega_i \,=\, \{x\in\Rn\;|\;
\langle a_i, x \rangle - b_i \,\leq\,0\}$, $a_i,b_i \,\in\,\Rn$,
$\|a_i\|_2=1$, for $i=1,\dots,k$. This is of interest in linear
programming (LP) and related problems. We expect to develop
alternate approaches to several issues arising in LP, including
rapidly finding the existence and location of a feasible point.
\item We will consider nonconvex but positively 1-homogeneous
Hamiltonians. These arise in: differential games as well as in the problem
of finding a closest point on the boundary of a given compact convex
set $\Omega$, to an arbitrary point in the {\it interior} of $\Omega$.
\item As an example of a nonconvex Hamiltonians we consider the
following problems arising in differential games \cite{ES}. We need
to solve the following scalar problem for any $z\in\Rn$ and
any $\alpha >0$
\begin{displaymath}
\min_y \left\{\frac{1}{2} \|y-x\|_2^2 - \alpha\|y\|_1\right\}.
\end{displaymath}
It is easy to see that the minimizer is the stretch$_1$ operator which
we define for any $i=1,\dots,n$ as:
\begin{equation}
\label{eq.stretch_def}
\left(\hbox{stretch}_1 (x,\alpha)\right)_i \;=\;
\begin{cases}
x_i+\alpha & \hbox{if } \ x_i > 0, \\
0 & \hbox{if } x_i=0,\\
x_i-\alpha & \hbox{if } \ x_i < 0.
\end{cases}
\end{equation}
We note that the discontinuity in the minimizer will lead to a jump
in the derivatives $(x,t)\mapsto \frac{\partial\varphi}{\partial
x_i}(x,t)$, which is no surprise, given that this interface
associated with the equation
\begin{displaymath}
\frac{\partial\varphi}{\partial t}(x,t) - \sum_{i=1}^n \left|\frac{\partial
\varphi}{\partial x_i}(x,t)\right| \;=\; 0,
\end{displaymath}
and the previous initial data, will move inwards, and characteristics
will intersect. The solution $\varphi(x,t)$ will remain locally Lipschitz
continuous, even though a point inside the ellipsoid may be equally
close to two points on the boundary of the original ellipsoid in the
Manhattan metric.
So we are solving
\begin{align}
\varphi(x,t) &= -\frac{1}{2} - \min_{v\in \Rn}\left\{\frac{1}{2}
\sum_{i=1}^n a_i^2 v_i^2 - t \sum_{i=1}^n |v_i| +
\langle x,v\rangle\right\} \nonumber \\
&= -\frac{1}{2} + \frac{1}{2} \sum_{i=1}^n \frac{x_i^2}{a_i^2} -
\min_{v\in \Rn}
\left\{\frac{1}{2} \sum_i^n a_i^2 \left(v_i -
\frac{x_i}{a^2}\right)^2 - t \sum_{i=1}^n |v_i|\right\} \nonumber \\
&= -\frac{1}{2} + \frac{1}{2} \sum_{i=1}^n \frac{(|x_i| + t)^2}{a_i^2} \nonumber
\end{align}
The zero level set disappears when $t \geq \max_{i} a_i$ as it should.
For completeness, we also consider the nonconvex optimization problem
\begin{displaymath}
\min_{v\in \Rn} \left\{\frac{1}{2} \|v-x\|_2^2 - \alpha\,\|v\|_2\right\}.
\end{displaymath}
Its minimizer is given by the $\mbox{stretch}_2$ operator formally defined by
$$
\mbox{stretch}_2(x, \alpha) \;=\;
\begin{cases}
x + \alpha \frac{x}{\|x\|_2} & \mbox{ if } x\neq 0,\\
\alpha \theta & \mbox{ with } \| \theta \|_2=1 \mbox{ if } x=0.
\end{cases}
$$
This formula, although multivalued at $x=0$, is useful to solve the
following problem: move the unit
sphere inwards with normal velocity $1$. The solution comes from
finding the zero level set of
\begin{align}
\varphi(x,t) &= -\min_{v\in\Rn} \left\{\frac{|v|_2^2}{2} - t\|v\|_2 -
\langle x,v\rangle\right\} -\frac{1}{2}
\nonumber \\
&= -\min_{v\in \Rn} \left\{\frac{1}{2} \|v-x\|_2^2 - t\|v\|_2\right\}
+\frac{1}{2} (\|x\|_2^2 - 1) \nonumber \\
&= -\frac{1}{2} t^2 + t\|x\|_2 \left(1 + \frac{t}{\|x\|_2}\right) +
\frac{1}{2} (\|x\|_2^2 - 1) \nonumber \\
&= \frac{1}{2} (\|x\|_2 + t)^2 - \frac{1}{2} \nonumber
\end{align}
and, of course, the zero level set is the set of $x$ satisfying $\|x\|_2
= t-1$ if $t \leq 1$ and the zero level set vanishes for $t > 1$.
\end{enumerate}
|
\section{Introduction}
\label{int}
The interstellar medium of spiral galaxies contains large-scale
magnetic fields (of scale comparable with the galactic size)
which are believed to be excited by a
galactic dynamo, that is driven by the joint action of differential rotation
and mirror-asymmetric turbulent flows. As we
know from observations of several nearby spiral galaxies such as M~51,
as well as the Milky Way, this large-scale magnetic field is mainly azimuthal
in the disc,
with field strength of several $\mu$G, i.e. the magnetic energy is
close to equipartition with the kinetic energy of turbulent motions, see
e.g. Beck (2016).
In addition to the large-scale magnetic field, a small-scale component of comparable strength is present in the interstellar medium.
At least in the large isolated spiral galaxy NGC~6946, the large-scale magnetic field
is organized in magnetic arms located between the stellar arms. These features are more or less related to predictions of galactic dynamo theory
(for a review see e.g. \cite{Beck96}). There are some attempts (e.g. \cite{Arshakian09}) to follow magnetic field
evolution in the context of galactic evolution.
\begin{figure*}
\vspace*{2.5cm}
\resizebox{\hsize}{!}
{\includegraphics{n4513nuv.eps}
\includegraphics{ic5285sdss.eps}
\includegraphics{u5936sdss.eps}
\includegraphics{u9980sdss.eps}}
\hspace*{1cm}
{ \Large NGC~4513 $\mbox{ }$ \quad \quad \quad $\mbox{ }$ IC~5285 \quad \quad \quad $\mbox{ }$ \quad \quad \quad UGC~5936 $\mbox{ }$ \quad \quad \quad
UGC~9980 $\mbox{ }$}
\caption{Examples of
lenticular (they have no spiral arms but do have discs)
galaxies with outer rings: composite-coloured images from the SDSS survey and image from ultraviolet GALEX survey for NGC~4513. (The NGC~4513 image from
the SDSS survey shows the ring with a rather low contrast.)} \vspace*{2cm}
\end{figure*}
Observations and modelling of magnetic fields of barred galaxies,
considered as a particular
morphological class of galaxies, have been combined with interpretation of
the results obtained in the framework of galactic dynamos
(\cite{Metal98,Betal99,Metal01}). Previous experience with barred galaxies has shown
the investigation
of magnetic fields in galaxies of particular morphological classes to be fruitful (\cite{Betal02,Betal05,Metal07}), e.g. specific magnetic
structures near bars were identified and interpreted as being important in feeding
black holes at galactic centres.
Further studies of investigation of magnetic fields in particular morphological classes of galaxies should
include examples of classes which appear to be candidates for
future observations of polarized radio emission, which
is an important indicator of the presence of large-scale magnetic fields.
The aim of this paper is to discuss ring galaxies in this context;
note however that radio polarization detections of these galaxies are scarce.
Ring galaxies were defined as a class of disc galaxies by compiling
a list and atlas of galaxies in which the rings were separated from the bright central parts of the galaxies (\cite{vv60}). He stated
that the rings were "structural elements on the same level as bars, spiral arms, and discs". The typical
radii of the outer galactic rings range between 5 and 25 kpc (\cite{theys_spi}), and the width might be as large as a few
kpc. Though the rings are certainly stellar structures, they often contain gas and demonstrate some level of star formation.
For example, a recent catalogue of ring structures derived from The Spitzer Survey of Stellar Structure in Galaxies
(\cite{s4g}),
ARRAKIS (\cite{arrakis}), includes outer stellar rings detected in 50\%--60\%\ of all nearby S0 galaxies (\cite{lauri11,arrakis}); and among those,
half of all rings in S0s are seen in the ultraviolet -- in the NUV-band being mapped by the GALEX space
telescope -- so revealing recent star formation (\cite{kostuksil}). In early-type (S0) disc galaxies,
the star formation is mainly organized in large-scale rings (\cite{pogge_esk,salim}), and the neutral-hydrogen distributions
are also ring-like (\cite{vandriel}). The gas in the outer ultraviolet rings is often ionized by young stars (\cite{Silchenko14}).
van Driel \& van Woerden (1991), who had traced the rotation of the S0 discs far beyond the centre by observing neutral hydrogen in the rings,
noted that the rotation curves in the outer discs were flat and so the rotation was obviously differential.
In other words, the interstellar matter in ring galaxies has all the features that drive dynamo action in classical nearby
galaxies with magnetic field. The natural question here is: are the dynamo drivers in ring galaxies strong
enough to get self-excitation of magnetic fields?
Continuum radio emission, indicative of the presence of magnetic fields,
has been detected from a number of ring galaxies, e.g. Ghigo (1980), Jeske(1986), Appleton et al. (1999);
and there is a possible detection of polarized emission from
the polar ring galaxy NGC~660
(Wiegert et al 2015).
Estimates presented below
are optimistic enough to suggest that further
observational studies of these systems may be justified.
\begin{figure}
\begin{center}
\includegraphics[width=9cm]{fig2.eps}
\end{center}
\caption{Regular magnetic field generation in the ring for various
$d$ ($2d$ is the width of ring) and dynamo numbers $D$. The upper panel shows the magnetic field for $t=10^{10} \mbox{ yr}$ for $D=20,$ the
middle one the same for $D=50.$ The solid curves are for a ring with $d=1 \mbox{ kpc},$ the dashed curve, $d=1.5 \mbox{ kpc},$ and the dot-dashed curve is for $d=2 \mbox{ kpc}$. The lower panel shows the equipartition field
$B^*$ (which is determined by the kinetic energy of the turbulent gas)
versus radius for a representative case with $L=5$ kpc, $d=1$ kpc. }
\label{fig1}
\end{figure}
In the framework of our paper we have to recognize that the physical properties of the interstellar medium in
ring galaxies, and also the dynamo context of the problem,
are much less investigated than the corresponding topics for such galaxies
as the Milky Way or several nearby galaxies which are considered to be the site of typical examples of galactic dynamos (but see, e.g.
\cite{Cathy95}).
Their properties are quite diverse. Correspondingly, we are interested here in crude preliminary estimates rather than in detailed modelling.
Moreover, among the ring galaxies of quite diverse origins (see the next Section) we can select the cases which
are particularly promising for the dynamo development. Fortunately, rings are often found in the same plane as
the main
stellar disc component\footnote{There are examples of polar rings, i.e. rings inclined to the disc plane
(\cite{polarrings}).
However it is believed that after being accreted
with arbitrary inclined angular momentum, gaseous rings settle to the plane of the galactic disc in a few orbital
times, see
e.g. Steiman-Cameron \& Durisen (1988), Christodoulu \& Tohline (1993), Colley \& Sparke (1996).} This means that we can consider ring galaxies as
mainly flat systems and use various simplifications of the general dynamo equations previously developed for
spiral galaxies.
Additionally, the axisymmetry of the discs of ring galaxies simplifies dynamo studies.
\section{Interstellar gas in ring galaxies}
There are three distinct mechanisms that have been proposed for the
origin of the marked outer ring structures in disc galaxies.
The most exotic, and so in practice probably the rarest, is
the impact mechanism (\cite{freemanvauc,LyndTo76,theys_spi,fewmadore,AppStr,atha97}).
This mechanism involves the infall of a dwarf satellite or of a high-velocity massive
cloud impacting very close to the galactic
centre; this suggests that the initial configuration will be a strongly inclined, almost radial orbit of the satellite
or of the cloud around the host galaxy. Such collisions do not destroy the large-scale stellar disc
of the host galaxy but instead can provoke a compression wave running outward through the large-scale
gaseous (and stellar) disc of the host galaxy. The shock compresses the gas of the galaxy, initially
quasi-homogeneously
distributed over the whole disc, and
the compression becomes sufficient to initiate
star formation in a ring. Large spiral galaxies with high gas content appear to be the most favourable for this mechanism to operate.
\begin{figure}
\begin{center}
\includegraphics[width=9cm]{fig3.eps}
\end{center}
\caption{Dynamo generated magnetic fields for various values of $D$ and $L$ when the seed field is concentrated in the inner disc only. The ring half-width $d=1$ kpc in all cases. The upper panel
shows the magnetic field for $t=10^{10} \mbox{ yr}$ when the dynamo number
$D=20,$ the lower shows the same for $D=50.$ The solid curves
are for the gap extent $L=1 \mbox{ kpc},$ the dashed curves for $L=2 \mbox{\, kpc},$ and the dot-dashed curves are
for $L=3\mbox{\, kpc}.$}
\label{fig2}
\end{figure}
Another scenario, perhaps more frequent, invokes resonance (\cite{schommer_sullivan, atha82,buta95,butacombes}).
The resonance scenario requires violation of axisymmetry of the galactic disc and can occur in the
presence of a bar or at least of a slightly oval bulge (\cite{jung}). The triaxial gravitational
potential causes radial gas motions so that the gas can be acquired at the inner and outer Lindblad resonances
of the non-axisymmetric rigidly rotating pattern.
The third possibility for ring formation involves outer gas accretion (\cite{butacombes}).
We believe that
this scenario is probably the most common among early-type discs -- namely, among lenticular galaxies -- which possess outer rings
in more than a half of all cases (\cite{arrakis}). Interestingly, the frequency of strong bars diminishes among
lenticular galaxies compared to spirals (\cite{lauri11,lauri13}), while the frequency of outer rings rises
(\cite{arrakis}).
This implies a rather low probability that the outermost rings in lenticular galaxies are
due to resonance.
The high fraction of observed gas counter-rotation with respect to the stellar components, especially among
the S0s in rarified environments, strongly supports an origin through accretion (\cite{Katkov15}). Some examples
of the
outer star forming rings in visually unbarred
galaxies are shown in Fig.~1. This figure presents a particular class of ring galaxies -- the accreted
star forming
rings in axisymmetric S0s which are especially suitable to host a dynamo process.
The class is however far from being homogeneous. Fig.~1 illustrates that the ring galaxies of interest
can vary considerably in appearance.
This is why we
have to consider a variety
of possible dynamo effects which might contribute to magnetic field formation in such
scenarios and experiment with the dynamo numbers that estimate
the strengths of the effects. A particular challenge is that NGC 4513,
shown in the first panel of Fig.~1, has a counter-rotating ring.
IC 5285, UGC 5936 and UGC 9980 do not demonstrate
observable radial gas velocities, thus ruling out a collisional origin
for their rings
(\cite{uvrings,Silchenko14,Katkov15}).
In fact, to discriminate between the origin scenarios for an outer ring in a particular galaxy, kinematical
data are necessary, and 3D (two spatial coordinates on the plane of sky and one
along the wavelength axis, \cite{C95}) spectral data are especially useful (\cite{moisrev}). The collisional
rings and
OLR (Outer Lindblad Resonance) rings generate radial gas outflow while accreted gaseous rings rotate circularly,
though sometimes outside the main galactic plane. In addition, the optical spectral data provide the possibility
of determining the excitation mechanism for the ionized (warm) gas: the flux ratios of strong emission lines
discriminate easily between the mechanisms of shock excitation and of ionization by UV radiation of young stars
(\cite{bpt,vo87}).
To facilitate the work of the dynamo in magnetic-field generation, we need
\begin{itemize}
\item{long-lived (order of Gyrs) rings; so collisional rings, with their lifetimes of a few hundred Myr, are unlikely to be candidates;
}
\item {a supply of free electrons, so
an ionized state of the gas which can be provided and supported
by continuous star formation in the rings;
}
\item{regular circular gas rotation in the main galactic planes, presumably with a flat rotation curve, to
provide differential character of the gas rotation.
}
\end{itemize}
In some ways rings generated by accretion maybe be the most favourable
for our investigation, with estimated lifetimes possibly approaching
the Hubble time (Buta \& Combes 1996).
Based on the results of \cite{uvrings,Silchenko14,Katkov15} we
select a small sample with the properties mentioned above and present it in Fig.~1. The typical properties
of these galaxies are to be used in the further theoretical consideration of the dynamo magnetic-field generation
in ring galaxies. Some extension of the typical picture can be also considered,
by taking into account
the possible cases of the outer gas counter-rotation, as in NGC~4513 (\cite{Silchenko14}), or strong flaring
of the gas layer in the ring area, as in UGC~5936 (\cite{uvrings}).
Of this sample, continuum radio emission has been detected from IC~5285 (\cite{Co98}).
\section{Dynamos in ring galaxies}
From the viewpoint of galactic dynamos the main specific problem of ring galaxies
is to what extent
drivers of the galactic dynamo, i.e. differential rotation and mirror-asymmetric turbulence,
can produce a magnetic structure in the ring that is independent of that in the disc,
or is propagation of magnetic
field generated in the disc into the ring a more realistic process? Of course, the answer depends on properties
of the interstellar medium near the boundary between disc and ring. One possibility here is that rings are
overdense regions of warm gas superposed on
an exponential disk as in late-type disc galaxies. The other possibility is that there is a gap in density
distribution between disc
and ring. The latter option which is the case of lenticular galaxies where
warm ionized gas is often observed only in the rings (\cite{pogge_esk,salim}) is obviously favourable for formation
of independent magnetic structure in the ring and looks realistic especially for counter-rotating rings.
In the rings, the dynamo
drivers have in addition to overcome the turbulent diffusion of the large-scale magnetic field into the gap
between
the disc and the ring. We recall that in the standard disc dynamo the
drivers only have to compensate losses of
the large-scale magnetic field mainly through the upper and lower boundaries of the disc.
We start with an
order of magnitude estimate of the role of these losses.
We need to compare somehow the results obtained for ring galaxies with
those for spiral galaxies. We
proceed as follows. The conventional estimate for the intensity of dynamo action (e.g. \cite{Betal05}) is given in terms
of the dynamo number $D = 9 (\Omega h/v)^2$ where $\Omega$ is angular velocity, $h$ is disc thickness
and $v$ the rms turbulent velocity.
Usually $\Omega h \approx v$ and $D \approx 10$. This estimate is based on
the assumption that $r d \, \Omega/ d \, r = \Omega$,
which is an obvious idealisation for a flat (or outer Brandt)
rotation curve. For real rotation curves it seems reasonable to try
values $D$ of several dozen, i.e. significantly larger than the normal
estimates. This logic is directly
applicable for rings that co-rotate with the disc.
However a galaxy with a counter-rotating ring has unusually
strong shear in the region between the disc and ring.
We attempt to parametrize this situation as follows. We present the
rotation curve as a combination of two disjoint pieces of a Brandt
rotation curve with $\Omega$ of opposite signs, but
with absolute value matched smoothly through the gap (see Fig.~\ref{disc}),
and use the conventional estimate for $D$. In other words
we describe anomalous shear in region between the ring and the disc by the form of rotation curve. A problem
here is that we may overestimate the width of the region in which
the transition from the region co-rotating with
the disc to that co-rotating with the ring occurs.
Contemporary observations fail to resolve this
transition. In order to take this problem into account we consider
values of $D$ up to $D=200$, and demonstrate that the
results are more or less generically stable.
\subsection{Simple estimates}
The basic concept of the standard galactic dynamo is that the azimuthal magnetic field $B_\phi$ is obtained
from the radial
field $B_r$ by the action of differential rotation $\Omega$,
while the radial field is restored from the azimuthal by the
$\alpha$-effect which arises from the mirror asymmetry of interstellar turbulence.
Both effects are in competition with the destructive role of turbulent diffusion. This dynamo action can be represented in a very simple model by:
\begin{equation}
\frac{d B_{\varphi}}{d t}=r\frac{\partial \Omega}{\partial r} B_{r} - \eta(\pi^2/4) B_\varphi (1/h^2 +
1/d^2),
\label{eqBphi}
\end{equation}
\begin{equation}
\frac{d B_{r}}{d t}= -\frac{\alpha}{h}
B_{\varphi} - \eta(\pi^2/4) B_{r}(1/h^2 + 1/d^2).
\label{eqBr}
\end{equation}
Here $\eta$ is the turbulent diffusivity, $h$ is the half-thickness of the disc and $d$ is the half-width of the ring.
For the crude estimates of interest, we have
replaced all derivatives in the dynamo equations by corresponding algebraic terms and introduced numerical factors $\pi^2/4$ to fit numbers known from numerical experiments (e.g. \cite{Phillips01}). We estimate magnetic diffusivity in the framework of
conventional mixing length theory as $\eta = lv/3$ where $l$ is the basic scale of turbulence and $v$ is the r.m.s.
turbulent velocity, and the numerical factor 3 comes from the dimensionality of space. Another conventional estimate $\alpha = \Omega
l^2/h$ comes from the viewpoint that the mirror asymmetry of interstellar turbulence is caused by Coriolis force
action in a stratified medium. Note that this conventional estimate implies that, for the case of a counter-rotating ring, $\alpha$ changes sign in the gap between the disc and the counter-rotating ring.
This should be regarded as an order of magnitude estimate;
for our purposes the important factor is the dependence on $\Omega$.
\begin{figure}
\begin{center}
\includegraphics[width=9cm]{fig5.eps}
\includegraphics[width=9cm]{pitch.eps}
\end{center}
\caption{Upper panel. Regular magnetic field generation for various densities in the gap and the seed field in the inner disc. The
upper panel shows the magnetic field at $t=10^{10} \mbox{ yr}$ for $D=20,$ the
middle
panel gives the same for
$D=50.$ The solid curve shows $\rho_{\rm gap}=10^{-1}\rho_{0}$, the dashed curve $\rho_{\rm gap}=10^{-2} \rho_{0},$
and the dot-dashed curve $\rho_{\rm gap}=10^{-3} \rho_{0}$.
$R = 5$ kpc, $L=2$ kpc, $d=1$ kpc.
Without a ring the outer maximum in $B$ is absent, and $B$ continues to decline with radius. Lower panel. The dependence of pitch angle $p=\tan^{-1}(B_r/B_\phi)$ on radius.}
\label{fig2b}
\end{figure}
Dynamo self-excitation means that the linear dynamo equations above have exponential solutions $\sim \exp \gamma t, \gamma>0$. (In particular, it means that all magnetic field components grow with the same rate.) After
some algebra we obtain from Eqs.~(\ref{eqBphi}), (\ref{eqBr}) a dispersion relation
\begin{equation}
\gamma = \frac{\eta}{h^{2}}\left(-(\pi^2/4)(1+ h^2/d^2) + \sqrt D \right),
\label{estimgam}
\end{equation}
where $D=-9 (h^2 \Omega r\frac{\partial \Omega}{\partial r}/v^2)$ is a dimensionless quantity known as the dynamo number,
which measures the intensity of galactic dynamo action.
Dynamo self-excitation (positive $\gamma$) arises if $D$ exceeds
a critical level which for a thin galactic disc is $D_{\rm cr} \simeq \pi^4/16 \simeq 6$ (to be compared with $D_{\rm crit}=7.5$
found by \cite{Retal88} in a slightly different approach). For a flat rotation curve $|r
\frac{\partial \Omega}{\partial r}| \approx \Omega$ and for $h= 500 \mbox{ pc}$, $v=10 \mbox{ km s}^{-1}$ and
rotation velocity of $200 \mbox{ km s}^{-1}$ we find that at $r=10 \mbox{ kpc}$, $ h^2 \Omega r\frac{\partial
\Omega}{\partial r}/v^2 =1$, i.e. $D=9$.
Diffusive losses in the gap between the ring and disc increase $D_{\rm cr}$. It reaches $D_{\rm cr} =10$
for $d = 1.5 h$, giving $d =600 \mbox{ pc}$ for $h=400$ pc. On the other hand, the above estimate are crude
enough
that estimates $D = 10 \sim 50$ can be considered as realistic in a more
complete model.
Increase of the ring thickness $h$ or
differential rotation $r \partial \Omega/\partial r$ promote dynamo self-excitation. Indeed, rotation curves of ring
galaxies appear more complicated than given by flat rotation curves (\cite{Silchenko14})
and the rings can be quite thick (\cite{uvrings}). We
conclude that self-excitation of a large-scale dynamo driven magnetic field in the rings looks possible. For $h =d$
dynamo excitation needs $D = 30$, which seems quite possible for the interstellar medium.
\begin{figure}
\begin{center}
\includegraphics[width=9cm]{fig6.eps}
\end{center}
\caption{Regular magnetic field generation for a counter-rotating ring
at a nominal time of $10^{10}$ yr . The solid curve shows the magnetic field
for $|D|=20$, dashed curve for $|D|=50$. The upper panel shows $\alpha \propto \Omega$, middle panel
$\alpha(r)=\alpha_{0}>0$, lower panel $\alpha(r)=\alpha_{0}<0$. $R=5$ kpc, $L=2$ kpc.
We note that in the lower panel the dynamo action in the disc is suppressed by the choice of sign of $\alpha_{0}$.
}
\label{fig2c}
\end{figure}
\subsection{Towards detailed dynamo models for ring galaxies}
A natural development of the above estimates is to use what is known as the no-$z$
approximation~(\cite{Subramanian93,Moss95}). This approximation describes the magnetic field
components in the disc plane. The $z$-component of the field perpendicular to the disc plane is restored by the
solenoidality condition assuming that it is much less than the plane components ($B_{r}$ and $B_{\varphi}$).
\begin{figure}
\begin{center}
\includegraphics[width=9cm]{fig4.eps}
\end{center}
\caption{Magnetic field evolution for the model with the seed field concentrated in the inner disc only ($R= 5$ kpc, $L= 2 $ kpc). The solid curve shows $t=2 \cdot 10^{9} \mbox{ yr},$ the dashed curve $t=5 \cdot 10^{9} \mbox{ yr},$ and the
dot-dashed curve is for $t=10^{10} \mbox{ yr}.$}
\label{fig2a}
\end{figure}
In the framework of the no-$z$ model the dynamo equations read
$$\frac{\partial B_{r}}{\partial t}=-\frac{\alpha B_{\varphi}}{h}+$$
\begin{equation}\label{eq1}
+\eta \left(-(\pi^2/4)\frac{B_{r}}{h^{2}}+\frac{\partial}
{\partial r}\left(\frac{\partial}{r \partial r}(r B_{r})\right) +\frac{1}{r^{2}}\frac{\partial B_{r}}{\partial \varphi}\right),\end{equation}
$$\frac{\partial B_{\varphi}}{\partial t}=-r\frac{\partial \Omega}{\partial r} B_{r}-\Omega \frac{\partial B_{\varphi}}{\partial \varphi} $$
\begin{equation}\label{eq2}
+\eta
\left(-(\pi^2/4)\frac{B_{\varphi}}{h^{2}}+\frac{\partial}{\partial r}\left(\frac{\partial}{r \partial r}(r
B_{\varphi})\right) +\frac{1}{r^{2}}\frac{\partial B_{r}}{\partial \varphi}\right) \, .\end{equation}
The $\alpha$-coefficient includes saturation of the magnetic field growth when it reaches the equipartition value
$B^{*}=v\sqrt{4\pi\rho}$ where $\rho$ is the interstellar medium density. We use an algebraic quenching for dynamo action $\alpha \propto (1+\frac{B^{2}}
{B^{*2}})^{-1}$ and avoid for the time being more sophisticated parameterisations
~(e.g., \cite{Shukurov06,Sur07,Mikhailov13}). For $\Omega$ we use a Brandt (1960) rotation curve:
\begin{equation}
\Omega=\Omega_{\rm Br}(r)=\frac{\Omega_{0}}{\sqrt{1+(r/r_{\omega})^{2}}},
\label{Brandt}
\end{equation}
with parameters chosen to give an asymptotic rotational
velocity of about $200$ km s$^{-1}$ when $R_{\omega}=2\mbox{ kpc}.$
Now $\alpha\propto\Omega$.
We assume that $B^{*}=5 \mbox{ } \mu \mbox{G}$ in the main parts of the
galaxy.
For $\rho_{\rm gap}$ we
assume that it is an order of magnitude less than the density in the inner parts
and $B^*$ is then about 0.5 $\mu G$ there (Fig.~2, lower panel).
Usually we use zero boundary conditions at the central and outer boundaries of the galactic disc.
Additionally, we define $L$ to be the radial extent of the gap between
inner disc and outer ring, and $R$ the radius of the inner disc.
As a conservative assumption we take the rotation curve to be given
by Eq.~(\ref{Brandt}) everywhere, including in the gap.
The outer boundary of the disc in which we model dynamo action is fixed
at a radius of 10 kpc in all cases.
The magnetic field configuration in a ring galaxy depends on the lifetime of the ring and its origin, in
particular the nature of the seed magnetic field configuration. The main aim of this paper is to investigate whether a galactic
dynamo can generate a magnetic field in the ring.
In order to clarify this, we are interested initially in
long-lived rings and weak (much smaller than equipartition strength) seed magnetic fields.
This is why we choose as a basic example to present the field resulting from a weak seed field after $t=10^{10}$ yr.
We bear in mind that other possibilities should be considered later,
after confirming that dynamo generated magnetic
fields can be important for ring galaxies.
Generally the saturated state is attained in a shorter time, but at the
moment we wish mainly to demonstrate the eventual outcome of dynamo action.
In the framework of this research
we first perform some preliminary
experiments.
The modelling is quite conservative, in that
we use the Brandt rotation curve and ignore the fact that
differential rotation in ring galaxies is usually more pronounced,
and always consider an inner disc of maximum radius $R= 5$ kpc
independent of the width of the gap $L$.
Correspondingly,
an increase of $L$ moves the ring to a less dynamo-active radius.
\section{Results: no-$z$ model}
\label{res}
We have run our model for $D=9,\mbox{ }20, \mbox{ }50$ and
some other values that seem to be realistic for spiral galaxies
and their outer rings. For $D=9$ the initial magnetic field in the
ring decays. Fig.~2 shows $B=\sqrt{B_{r}^{2}+B_{\varphi}^{2}}$ for
$t=10^{10} \mbox{ yr}$ for $D=20$ (top panel) and $D=50$ (middle
panel) and the seed field present in the outer ring only
($B_{\varphi}(t=0)=B_{0}\sin \left(\pi \frac{r-r_{\rm
min}}{2d}\right)$, $B_r =0$ and $r_{\rm min}$ and $r_{\rm max}$
are the inner and the outer radii). We find that the intensity of
dynamo action ($D$) can be (under realistic assumptions) sufficient for magnetic field
self-excitation and that the field can reach a field strength that is close to
the equipartition value in a few Gyr.
Varying the half-widths of the ring from 1 to 2 kpc
(the magnetic field decays for smaller $d$) we find that the
magnetic configuration is
quite robust. However wider rings
and larger $D$ give more widely distributed and stronger magnetic field.
In order to isolate the effects of dynamo action in the ring we take here
a seed magnetic field for the dynamo that is present in the ring only.
We also can conclude that an initial magnetic field concentrated in
the ring can pass into the central parts.
\begin{figure}
\begin{center}
\includegraphics[width=9cm]{fig7.eps}
\end{center}
\caption{Regular magnetic field generated in an outer thin ring} in the nonaxisymmetric case (see discussion in the text).
\label{fig2d}
\end{figure}
The figure obtained for the large-scale magnetic field strength in the disc is in some cases as large as $30 \, \mu G$ which appears
rather large in the context of magnetic fields in spiral galaxies. This estimate
arises because we use $B_{\rm
eq} = 5 \, \mu G$ and $D=50$. Then, using the estimate $B = B_{\rm eq} \sqrt{D-D_0}$
obtained by Shukurov (2007), where $D_0 =7$ is the generation
threshold for the disc, we obtain the above field strength.
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{disc.eps}
\end{center}
\caption{ Radial dependence of the key disc properties for the
counter-rotating model. Upper panel, $B_{\rm eq}$ (solid, in units of $ B^*$).
Lower panel $\Omega$ (solid) and $\alpha$ (broken. }
\label{disc}
\end{figure}
Then we ran the model with the seed in the inner disc only (Figs.~3 and 4). We see that the outer ring becomes magnetized
after sufficient time, but it appears that magnetic field transport from the inner disc is at least as
important as in situ dynamo action in the outer ring.
Fig.~5 shows that in the framework of our model
the dynamo generated magnetic field is not very sensitive to the
gas density in the gap between the ring and the disc.
Without a ring, the outer maxima in $B_\phi$ are absent, and $B_\phi$ continues with radius.
The no-$z$ model gives a counter-intuitive result (Fig.~6) for the case of a counter-rotating ring
(such as observed in NGC~4513, although we do not attempt to model this galaxy explicitly). Then $\Omega=f(r)\Omega_{\rm Br}(r)$
where $f(r)=1$ in the inner parts, $f(r)=-1$ in the outer parts and in the gap it changes gradually. Qualitatively, $\Omega(r)$ is as shown in
Fig.~\ref{disc} , with allowance for differing positions of the gap.
Counter-rotation of the ring in respect to the inner disc means very large radial gradients in $\Omega$. A
natural expectation is to get effective dynamo action around the gap between the ring and the disc. In practice,
we face a problem how to parametrize in the framework of the model the quantity $\alpha$ for the counter rotating
ring. A straightforward parametrization $\alpha=\Omega l^{2}/h$ assumes that $\alpha$ vanishes in the
gap, suppressing dynamo action. (Here $l$ is the scale of interstellar turbulence: typically $l=100$ pc.) $\alpha$
is a quantity which is hard to predict, and we also tried the option that $\alpha$ does not change sign in
the gap (Fig.~6c). In the absence of counter-rotation, we tested the
case of $\alpha$ positive in the disc as well as in the
ring as well as the case of
$\alpha$ negative both in the disc and the ring
and found again that the
dynamo action in the gap region is quite moderate.
\begin{figure}
\begin{center}
a)\includegraphics[width=0.37\textwidth]{2D2_b.eps}\\
b)\includegraphics[width=0.45\textwidth]{B_2D2.eps}\\
\end{center}
\caption{a) Dependence of $B_\phi$ on radius in the disc plane,
b) $B_\phi$ contours for the model with $B_{\rm eq1}=0.1 B^*$ ; Here
$r_i=0.5R_{\rm gal}, r_o=0.7R_{\rm gal}$, with counter-rotation.} \label{2D2}
\label{2D2fig}
\end{figure}
We see from the above results that the magnetic field can pass through the gap between ring and disc.
If the seed is
non-zero in the disc only, it penetrates into the ring and vice versa.
A theoretical prediction for the field propagation (based on an estimate given in a much more general context in
\cite{Kolmogorov37}) is given by (\cite{Moss98,Mikhailov15}):
$$V=\sqrt{2 \gamma \eta},$$
where $\gamma$ is the growth rate of magnetic field. Taking
$\eta=0.33 \mbox{ kpc km s}^{-1},$ $\gamma=1.5
\mbox{ Gyr}^{-1},$ we obtain
$V=1 \mbox{ kpc Gyr}^{-1}$, quite close to the
numerical estimate that follows from Fig.~4.
We have assumed that the diffusivity, given by properties of the
turbulence is uniform through the inner disc and ring.
Of course, this may not be justified.
We recognize that these times significantly exceed the probable lifetimes
of the rings. Among other conservative assumptions in the model, we note
that the timescales depend inversely on the value of the diffusivity, which is
subject to substantial uncertainty. Also the choice of seed field and
the history of the galaxy before the formation of the ring can be important -- see the
comments in Sect.~\ref{2D}.
\begin{figure}
\begin{center}
a)\includegraphics[width=0.45\textwidth]{B_2D10.eps}\\
b)\includegraphics[width=0.37\textwidth]{2D10_b.eps}\\
c)\includegraphics[width=0.37\textwidth]{mod5_b.eps}
\end{center}
\caption{a) Dependence of $B_\phi$ on radius in the
disc plane; b) $B_\phi$ contours, for model with $B_{\rm eq1}=0.1B^*$
and enforced quadrupolar parity; c) Dependence of $|B_\phi|$ on radius
for the corresponding no-$z$ model. $r_i=0.5 R_{\rm gal}, r_o=0.7 R_{\rm gal}$, with
counter-rotation. $B_\phi$ is measured in units of the equipartition field $B^*$.} \label{2D10}
\end{figure}
The magnetic field in an isolated galactic disc is usually (almost) axisymmetric.
A ring however is a relatively thin,
azimuthally extended and
relatively isolated, body and it is a priori possible that generation
of a bisymmetric magnetic field could occur there.
We can note here that dominant bisymmetric magnetic configurations
are now considered to be rare in or completely absent from normal spiral galaxies
(see e.g. Beck 2016).
Nevertheless, it seems reasonable to examine briefly this option in thin rings, given that
dynamo theory suggests that thin rings
are believed to be more favourable for generation of nonaxisymmetric fields.
However, we found such a solution only for the case of a free outer boundary of the disc ($\partial B/\partial r=0$ at
$r=10$ kpc), $\Omega_{0}=67 \mbox{ km s}^{-1} \mbox{kpc}^{-1}$, $\alpha=2.68 \mbox{\, km s}^{-1}$ (corresponding to $R_\alpha\approx 4$, $R_\omega \approx 50$)
the radius of the inner disc is $5 \mbox{ kpc},$ the inner radius of the outer ring is $9 \mbox{ kpc}$, and its outer
radius is $10 \mbox{ kpc}$.
The initial field is zero if $r<9\mbox{ kpc}$ and bisymmetric in the ring.
With the standard dynamo parameters of Sect.~4, the bisymmetric field generated
only survives for about
$t=0.5 \mbox{ Gyr}$ and then disappears, to be replaced by an axisymmetric configuration. However with
slightly different dynamo parameters a stable
bisymmetric field can be found (Fig.~7).
We can conclude that, although possibly of some interest in a dynamo theory
context, nonaxisymmetric fields are not likely to be relevant to real ring galaxies.
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{t_2D2z.eps}
\end{center}
\caption{Dependence of global parity and energy on time for
a calculation beginning from a weak seed field of mixed parity.
} \label{evol_seed}
\end{figure}
\section{Counter-rotating discs -- beyond the no-z model}
\label{2D}
\subsection{Basic model}
\label{basic2D}
The rather surprising results obtained above for counter-rotating discs,
using the severely truncated no-$z$ model, suggests the need for verification
with an alternative model. Of course, such verification would be
helpful for case of a co-rotating disc as well, but in this case previous
comparisons show that the no-$z$ model provides a satisfactory
approximation to a conventional axisymmetric model with explicit $z$-dependence. Given that the no-$z$ models of Sect.~\ref{res}
suggest that fields are axisymmetric, we revisited the problem using an
axisymmetric embedded disc model, with cylindrical $r, z$ coordinates.
The code used is essentially that of Moss et al. (1998).
(This is an $\alpha^2\Omega$ code, whereas in the previous sections the
$\alpha\Omega$ approximation is used. However, in the regime of interest,
there is little difference between these approaches.)
In this Section we use a slightly different notation: $R_{\rm gal}$ is now the overall galactic radius, and we define the gap location and size by parameters $r_i$, $r_o$.
The reference rotation curve is again a Brandt curve (Eq.~(\ref{Brandt}))
with $r_\omega=0.2R_{\rm gal}$. In this $\alpha^2\Omega$ code,
the dynamo numbers are given
by $R_\omega=\Omega_0h^2/\eta_0, R_\alpha=\alpha_0 h/\eta_0$, where $\Omega_0$ is
the quantity defining the rotation curve (eq.~(\ref{Brandt}),
$\alpha_0$ is the corresponding value of $\alpha$, and $\eta_0$ is the value of the diffusivity
in the disc.
The diffusivity is independent of radius, and increases asymptotically to
$25\eta_D$ in the halo.
$\alpha$ varies sinusoidally in the disc ($\propto \sin\frac{z\pi}{h(r)}$),
and is zero in $|z|>h(r)$.
We continue to use a canonical value of $\eta_D=10^{26}$ cm$^2$s$^{-1}$ in the later
discussion, but always bearing in mind the possibility of some variation,
as this estimate is rather uncertain. Our probable underestimate of the angular velocity gradients, and
hence
effective dynamo number, might be considered to be compensated by the use of large values of $D$.
The "gap" is between $r=r_i$ and $r=r_o$,
and the equipartition field strength is uniform outside of the gap, i.e.
in $r\le r_i$ and $r\ge r_o$. Field strengths are measured in units of
this equipartition field (again defined by the kinetic energy
of the assumed turbulent gas motions). In the gap the equipartition field
drops quadratically to
the value $B_{\rm eq1}$ at $r=(r_i+r_o)/2$. In the initial investigation,
the dynamo parameters were
slightly different to those used in Sect.~\ref{res}: $R_\alpha=4$,
$R_\omega=50$. For the majority of models,
in $r\le r_i$, the disc thickness $h=0.04R_{\rm gal}$,
in $r\ge r_o$, $h=0.08R_{\rm gal}$, with a smooth interpolation in the gap.
Analogously, in $r\le r_i$, $\Omega=\Omega_B(r)$,
in $r\le r_o$, $\Omega=-\Omega_B(r)$, with a smooth transition.
Correspondingly, with $\alpha\propto \Omega$, $\alpha$ changes sign smoothly.
In the standard case, $r_i=0.5 R_{\rm gal}, \, r_o=0.7 R_{\rm gal} \ $.
Fig.~\ref{disc} shows the radial dependence of the key disc quantities.
We first computed a model with $B_{\rm eq1}=0.1 B^*$ (Sect.~3.2). Fig.~\ref{2D2}a
shows the dependence of $B_\phi$ on radius along the mid-plane $z=0$,
and Fig.~\ref{2D2}b the contours of $B_\phi$.
The notable feature of these plots is the departure from strict quadrupolar
parity -- in this case parity $P\approx -0.5$ in the steady state.
This feature could not be revealed by the no-$z$ model, which implicitly
assumes $P=+1$.
We have used the standard definition of dynamo field parity,
\begin{equation}
P=\frac{E_Q-E_D}{E_Q+E_D},
\end{equation}
where $E_Q, E_D$ are the energies of the even and odd parts respectively
of the field. Thus a strictly quadrupolar/dipolar field has $P+1/-1$.
In order to make a better comparison with the no-$z$ model, we recomputed
this model in the region $z\ge 0$, with quadrupolar parity enforced
by the boundary conditions on the plane $z=0$.
The contours of toroidal field and the radial variation of $B_\phi$
on $z=0$ are shown in the upper two panels of Fig.~\ref{2D10}.
The lower panel shows the variation of $|B|$ with radius in
the corresponding no-$z$ model.
Comparison of the lower two panels of this Figure shows a remarkable similarity
between the no-$z$ and $r,z$ models, and gives confidence in the modelling,
even for these much larger dynamo numbers.
(See also Phillips 2001).
Varying $B_{\rm eq1}$ between 1 and 0.01 produces very little variation in the eventual steady configuration,
with final parity around $P\approx -0.3$.
Moving the gap to $(r_i, r_o)=(0.6, 0.9)$ gives a final parity, $P\approx +0.4$,
and with (0.35, 0.7) gives $P\approx -0.28$.
A model with uniform disc thickness, $h/R_{\rm gal}=0.04$, no gap/counter-rotation,
$\alpha\propto \Omega$ and $R_\alpha=-1.0, R_\omega=20$ gives a strictly dipolar field,
$P=-1$. This result persists for considerably larger dynamo numbers, e.g. $R_\alpha=-4.0, R_\omega=50$. This means that the dynamo action in the disc of the model is strong enough to excite a dipolar
magnetic field (in which the azimuthal field changes sign at the central plane (antisymmetric)) while for a weaker dynamo action magnetic
fields in galactic discs have a quadrupolar
configuration (with azimuthal field that is symmetric with respect to the central plane). The counter-rotating ring and
associated magnetic field losses resulting from magnetic field transport into
the gap makes the parity of the configuration mixed, and the field configuration in the inner parts of the disc is
closer to
quadrupolar symmetry.
\begin{figure}
\begin{center}
(a)\includegraphics[width=0.45\textwidth]{b_teq70pt0.eps}\\
(b)\includegraphics[width=0.45\textwidth]{b_teq72pt0.eps}\\
(c)\includegraphics[width=0.45\textwidth]{b_teq74pt0.eps}\\
(d)\includegraphics[width=0.45\textwidth]{b_teq76pt0.eps}\\
(e)\includegraphics[width=0.45\textwidth]{b_teq78pt0.eps}
\end{center}
\caption{The radial distribution of $B_\phi$ in the disc plane (a) immediately
before the counter-rotation is turned on; (b), (c), (d), (e) at times
$0.9, 1.8, 2.7, 3.6$ Gyr afterwards.
}
\label{Bevol}
\end{figure}
Further models with $r_i=0.5 R_{\rm gal}, r_o=0.7 R_{\rm gal}$ with a) no counter-rotation,
$\alpha\propto \Omega, B_{\rm eq1}=0.1$, gave a strictly quadrupolar
$P=+1$; b) counter-rotation, $\alpha \propto |\Omega|$, $B_{\rm eq1}=0.1 B^*$,
resulted in $P=+0.98$.
In the counter-rotating case and $\alpha\propto \Omega$, as the dynamo numbers are reduced,
the contribution of the odd parity component also becomes smaller. For example,
with $R_\alpha=2, R_\omega=25$ in the steady state, the final parity is $P\approx 0.12$,
with $R_\alpha=1, R_\omega=25$ we get $P\approx 0.47$, and $R_\alpha=1, R_\omega=20$ also gives a mixed parity final state.
In the marginally
supercritical case $R_\alpha=1, R_\omega=15$, the steady field is purely
quadrupolar, $P=+1$; the field in the outer ring decreases in relative strength with reduction in dynamo numbers.
In summary, we can deduce that mixed parity fields (in which the
azimuthal field has no symmetry with respect to the central plane) are a robust feature of
the model with counter-rotation and $\alpha\propto \Omega$.
With $\alpha\propto |\Omega|$ the effect is much reduced.
The region with $\alpha < 0$ seems crucial to the effect, counter-rotation
with $\alpha>0$ everywhere produces only a small effect.
The computations described in this Section started from a weak seed field
of mixed parity distributed throughout the disc.
With our standard $\eta_D=10^{26}$ cm$^2$ s$^{-1}$, the
eventual steady state can take times in excess of $10^{10}$ years to become
established for the smaller dynamo numbers,
and departure from very near quadrupolar symmetry can then take a
substantial fraction of this time. Timescales can be substantially reduced by different choices of the field present at the onset of counter-rotation.
We emphasize again that these estimated
timescales are sensitive to the value of $\eta_D$. Prolonged
timescales are plausibly associated with the diffusion of field across the gap between
the inner disc and the counter-rotating ring.
\subsection{Timescales and evolution}
\label{timescale}
In Sect.~\ref{basic2D} we have presented the saturated steady magnetic configuration
for a particular model. The results presented using the no-$z$ approximation
suggest that evolution to the final configuration may be quite slow -- this
feature is associated with the diffusion of field across the gap as the
final configuration is determined.
If we start with an arbitrary seed field then, depending
on its strength, evolution to the final saturated configuration can take
times longer than the estimated lifetimes of rings, depending
on their believed mechanism of formation. For example, with rings
formed by accretion Buta \& Combes (1997) quote times that may approach the
Hubble time, whereas for rings resulting from impacts the models
of Athanassoula et al. (1996) suggest lifetimes of less than 1~Gyr.
We show in Fig.~\ref{evol_seed} the evolution of parity and energy for an example where the seed field is weak (magnitude much smaller than equipartition strength),
and of arbitrary parity.
After an initial brief fall in parity, a purely quadrupolar configuration
persists for about 4 Gyr. The nonlinear regime is then reached and the field
moves rapidly to the asymptotic state shown in Fig.~\ref{2D2fig}b.
However significant large-scale fields can be present in the ring regions
after $3-4$ Gyrs if the seed field is stronger.
Thus we now can consider two evolutionary
scenarios: either the counter-rotating ring is present from time $t=0$ with a
seed field of unknown strength and geometry, or a merger event occurs creating the ring after a standard
even parity field has been established.
To illustrate this, with a stronger seed field a significant large-scale field
of magnitude comparable with equipartition strength can
be established in the ring after $2-3$ Gyr, although the dynamo is far
from saturation, and a mixed parity field does not appear for another $0.5-1$ Gyr, with a longer time until the asymptotic state of Fig.~\ref{2D2}b is attained.
On the other hand, if the ring forms after an encounter or merger event
involving one or two pre-magnetised galaxies with near saturated dynamos,
then the effective seed field
for subsequent dynamo action may be quite strong and organised.
(We do not consider here the possibility that such an event generates turbulence
and small-scale magnetic fields of approximately equipartition strength,
which would provide a seed for subsequent dynamo action -- see, e.g., Moss et al. (2013).)
To gain some insight into this scenario, we allowed a standard, even parity, steady
global field to become established in a model with no counter-rotation, and
no gap, with reduced strength equipartion field. Counter-rotation was then
introduced, with $B_{\rm eq1}=0.1$, i.e. the model of Sect.~\ref{basic2D}.
In Fig.~\ref{Bevol} we show the dependence of $B_\phi$ in the plane $z=0$ on
radius at times $0, 0.9, 1.8, 2.7$ and $3.6$ Gyr after the introduction of counter-rotation.
A significant field is present in the ring ($r>7$ kpc) throughout.
We can conclude that mixed parity fields may be attainable within galactic
timescales, depending on the details and history of the model.
Other unexplored possibilities include that we underestimated angular velocity
gradients near the gap, and thus underestimated the strength of the dynamo action
in that region, or that the turbulent diffusivity is not constant
throughout the disc plane.
We have shown that large-scale fields may be present in galactic rings,
and that there is also the possibility that they may be of mixed parity.
With our current state of knowledge we cannot be more specific.
\section{Discussions and conclusions}
We have demonstrated that dynamo drivers in ring galaxies, i.e. differential rotation and mirror-asymmetry of
interstellar
turbulence, can be strong enough to excite magnetic fields both in the disc and in the rings. Our expectation
is that such rings do contain large-scale magnetic field (probably slightly weaker than in the inner discs) that
give rises to polarized radio emission.
We have found that the current concepts of galactic dynamo action tell us quite counter-intuitively that the strong angular velocity gradients associated
with
counter-rotation of galactic rings with respect to that of the disc does not lead to a substantial increase of
dynamo action
and magnetic field strength near the ring. In other words, observations of
magnetic field in galaxies with counter-rotating rings have the potential
to provide a
strong test for current galactic dynamo concepts.
Inter alia, this investigation demonstrates the limits of the no-$z$
model. It is known to be satisfactory (and very useful) for studying fields
that are believed with a high degree of confidence to have strictly even
parity with respect to the galactic equator, but is clearly unreliable
where this symmetry cannot be guaranteed, as in the models of Sect.~\ref{2D} above, where the stable fields are of mixed parity.
A potentially important feature of thin rings is the possibility
of generating non-axisymmetric fields. In our model, such fields can
more readily be generated
if magnetic field propagates freely through the outer boundary of the ring.
However dynamo numbers have to be chosen quite selectively to obtain nonaxisymmetric fields, and the result suggests that axisymmetric large-scale fields
are the normal case.
Potentially more importantly, a magnetic configuration that is asymmetric in respect to the
central galactic plane (a mixed parity solution) can arise near to the gap
between the disc and the counter-rotating ring.
We note that our paper does not cover all dynamo related problems that arise in
the context of ring galaxies studies.
In particular, it looks natural to investigate possible effects of dynamo action in resonant rings as well as
the dynamo driven evolution of a strong (i.e. comparable with the
equipartition field strength) and maybe small-scale seed field.
A further possibility is the effects of ongoing injections
of small-scale fields of approximately equipartition strength generated
by turbulent dynamo action, as in Moss et al. (2013).
Such extensions of the topic should however be addressed separately.
We further note that in order to reach steady-state magnetic configurations we
sometimes run our code for
times which are nominally several times longer than the anticipated
lifetimes of the rings. However, the eventual steady state configurations
may not be so significant -- see e.g. Fig.~\ref{Bevol}.
Timescales for our models, and their prior evolution, are quite uncertain. We have presented a scenario summarized in Fig.~\ref{Bevol} in which fields can be present in galactic
rings more-or-less throughout the lifetime of the rings.
Whether the rings live long enough for global mixed parity configurations to
become established is more uncertain, although Buta \& Combes (1996)
do suggest that ring lifetimes may be quite extended.
Observations of such fields
would anyway be challenging.
We reiterate that the timescale for
magnetic field evolution depends on the value of the
turbulent magnetic diffusivity which is not a directly observable
quantity and theoretical estimates contain substantial uncertainties.
We have taken a standard figure ($10^{26}$ cm$^2$ s$^{-1}$),
but need to keep in mind that we may underestimate the
speed of magnetic field transport in ring galaxies.
On the other hand, our results demonstrate that we can expect various long-lived
magnetic transients in ring galaxies,
and contemporary magnetic configurations there can be influenced
by features of seed magnetic fields (or the field present after the ring forms).
Chy\.zy \& Buta (2008) found a coherent magnetic spiral structure crossing the inner pseudo-ring of NGC~4736. It is
not exactly comparable with the outer rings discussed here,
but it appears to be an
interesting system.
We obtained an unexpected result for dynamo models with counter-rotating disc and ring. In spite of a very large
shear between disc and ring, the resulting dynamo action is quite moderate.
Even with our definition
of dynamo number $D$ the excitation threshold is found numerically to be
around $D=13$, instead of $D=7.5$ for spiral
galaxies. Thus the counter-rotating configuration inhibits the efficiency of dynamo action. The eigenfunction near marginal excitation is concentrated
in the inner regions, and so this increase in marginal dynamo number
is plausibly associated with its reduced radial scale,
and an opposition between the inner and outer dynamo drivers.
In general, we can conclude that magnetic field studies in ring galaxies have the potential to provide useful information
concerning the nature and evolution of such galaxies.
\begin{acknowledgements}
EM acknowledges support from the Dynasty Foundation and RFBR under grant 16-32-00056. DS acknowledges support from RFBR under grant 15-02-01407. RB
is grateful for support from DFG Research Unit FOR1254. Analysis of the observational data on ring galaxies is supported by the RNF grant N 14-22-0041,
and dynamo modelling of the counter-rotating ring by RNF grant 16-17-10097.
The authors thank Marita Krause, MPIfR's internal referee, for helpful comments,
and the external referee, Andrew Fletcher for suggesting a number of
improvements to the text.
\end{acknowledgements}
|
\section{Introduction}
Graphical approaches to clustering are appealing because they offer a natural way to compare samples, in the form of edge information.
Broadly, a good clustering should have similar nodes in the same cluster and dissimilar nodes in different clusters. However, which graph to use for clustering remains an open question \cite{Luxburg2007}. Previous work has considered edges to be the output of a similarity function\footnote{A similarity function is a function of two nodes' features, e.g. the RBF kernel} (e.g. spectral clustering), a Bernoulli random variable (e.g. stochastic block models), or some more general measure of similarity/dissimilarity (e.g. correlation clustering).
In reality, edge information can take a variety and multiplicity of forms. Edges in social graphs correspond to communication exchanges, mutual interests and types of relationships. In biology, protein-protein interaction networks involve complex underlying mechanisms and conditions under which an event may occur. And in economics, trades constitute numerous goods, prices and transaction types.
We are inspired by the complex interactions happening around us. Our relationships are more complicated than friend/not-friend, and our transactions are about more than the monetary value.
The motivation of this paper is to cluster with the additional information provided by multivariate edge features. This is partly supported by Thomas and Blitzstein's \yrcite{Thomas2011} recent results showing that converting to a binary graph makes recovering a partition more difficult. We are also interested in how to choose similarity functions which better capture the relationship between nodes, one of the challenges of spectral clustering \cite{Luxburg2007}. Choosing a scalar similarity function (e.g. the RBF kernel) may be overly restrictive and underutilize useful information. This is partly the cause of scale issues in spectral clustering \cite{Zelnik2004}. Our approach allows more complex similarity functions, such as the absolute vector difference.
We believe these results will be particularly useful for image segmentation, community discovery and entity resolution. These are all applications (a) with a large number of clusters and (b) where we have access to some labeled edges. With a large number of clusters, it is unlikely we have training samples from every class, let alone enough samples to train a multi-class supervised classifier. However, the small number of labeled edges will enable us to learn the typical cluster structure.
In this paper, we extend the planted partition model to general edge features. We also show how to partially recover a maximum likelihood estimator which is $\mathcal{O}(\log(n))$-close to the log likelihood of the true MLE by using an LP-rounding technique. Much of the analysis in planted partition models consider the probability of exactly recovering the partition. Depending on the cluster sizes and number of samples, this is often improbable. Our analysis addresses how good the result will be, regardless if it is exactly correct. Further, our theoretical results provide some insights on how to perform edge feature selection or, likewise, how to choose a similarity function for clustering. Experimental results show interesting clustering capabilities when leveraging edge feature vectors.
\section{Related Work}
Two areas of research are closely related to our work. Our graphical model is an extension of the stochastic block model from the mathematics and statistics literature. We also use some key results from correlation clustering in our algorithm and analysis.
\subsection{Stochastic Block Model}
The stochastic block model (SBM) was first studied by Holland et al. \yrcite{Holland1983} and Wang and Wong \yrcite{Wang1987} for understanding structure in networks. In its simplest form, every edge in the graph corresponds to a Bernoulli random variable, with probability depending on the two endpoints' clusters. In planted partition models\footnote{There is some inconsistency in the literature regarding the distinction between planted partition and stochastic block models. Occasionally the terms are used interchangeably} there are two Bernoulli probabilities $p$ and $q$ corresponding to if the endpoints are in the same or different clusters, respectively. These models are actually generalizations of the Erd\H{o}s-R\'enyi random graph, where $p=q$. Random graph models have a storied history and include famous studies such as the small-world experiment (popularized as ``six-degrees of separation'') by Milgram \yrcite{Milgram1967} and Zachary's Karate Club network \yrcite{Zachary1977}. For a more complete overview, we refer the interested reader to the review by Goldenberg et al. \yrcite{Goldenberg2010}.
More recently, planted partition models have gained popularity in the machine learning community for clustering. McSherry \yrcite{McSherry2001} and Condon \& Karp \yrcite{Condon2001} provided early spectral solutions to exactly recovering the correct partition, with probability depending on a subset of the parameters $p$, $q$, the number of samples $n$, the number of clusters $k$, and the smallest cluster size. Most results show recovering the partition is more difficult when $p$ and $q$ are close, $n$ is small, $k$ is large, and the smallest cluster size is small. Intuitively, if there are a high proportion of singleton clusters (i.e. ``dust''), mistaking at least one of them for noise is likely.
Some of the numerous alternative approaches to recovering the partition include variational EM \cite{Daudin2008, Airoldi2008, Park2010}, MCMC \cite{Park2010}, and variational Bayes EM \cite{Hofman2008, Aicher2013}. Some of these approaches may also be applicable to the model in this paper, though we found our approach simple to theoretically analyze.
The work most closely related to ours extends the stochastic block model edge weights to other parametric distributions. Motivated by observations that Bernoulli random variables often do not capture the degree complexity in social networks, Karrer \& Newman \yrcite{Karrer2011}, Mariadassou et al. \yrcite{Mariadassou2010} and Ball et al. \yrcite{Ball2011} each used Poisson distributed edge weights. This may also be a good choice because the Bernoulli degree distribution is asymptotically Poisson \cite{Xiaoran2007}. Aicher et al. considered an SBM with edge weights drawn from the exponential family distribution \yrcite{Aicher2013}. Like Thomas \& Blitzstein \yrcite{Thomas2011}, he also showed better results than thresholding to binary edges. Lastly, Balakrishnan et al. \yrcite{Balakrishnan2011} consider Normally distributed edge weights as a method of analyzing spectral clustering recovery with noise.
Other interesting extensions of the SBM include mixed membership (i.e.\ soft clustering) \cite{Airoldi2008}, hierarchical clustering \cite{Clauset2007, Balakrishnan2011} and cases where the number of clusters $k$ grows with the number of data points $n$ \cite{Rohe2011, Choi2012}. Combining our ideas on general edge features with these interesting extensions should be possible.
\subsection{Correlation Clustering}
Correlation clustering was introduced by Bansal et al. \yrcite{Bansal2004} in the computer science and machine learning literature. Given a complete graph with $\pm 1$ edge weights, the problem is to find a clustering that agrees as much as possible with this graph. There are two `symmetric' approaches to solving the problem. \textsc{MinimizeDisagreements} aims to minimize the number of mistakes (i.e.\ $+1$ inter-cluster and $-1$ intra-cluster edges), while \textsc{MaximizeAgreements} aims to maximize the number of correctly classified edges (i.e.\ $-1$ inter-cluster and $+1$ intra-cluster edges). While the solutions are identical at optimality, the algorithms and approximations are different.
The original results by Bansal et al. \yrcite{Bansal2004} showed a constant factor approximation for \textsc{MinimizeDisagreements}. The current state-of-the-art for binary edges is a 3-approximation \cite{Ailon2008}, which Pan et al. \yrcite{Pan2015} recently parallelized to cluster one billion samples in 5 seconds. Ailon et al. \yrcite{Ailon2008} also showed a linear-time 5-approximation on weighted probability graphs and a 2-approximation on weighted probability graphs obeying the triangle inequality. Demaine et al. \yrcite{Demaine2006} showed an $\mathcal{O}(\log(n))$-approximation for arbitrarily weighted graphs using the results of Leighton \& Rao \yrcite{Leighton1999}. Solving \textsc{MinimizeDisagreements} is equivalent to APX-hard minimum multi-cut \cite{Demaine2006, Charikar2003}.
For \textsc{MaximizeAgreements}, the original results by Bansal et al. \yrcite{Bansal2004} showed a PTAS on binary graphs. State-of-the-art results for non-negative weighted graphs are a 0.7664-approximation by Charikar et al. \yrcite{Charikar2003} and similar 0.7666-approximation by Swamy \yrcite{Swamy2004}. Both results are based on Goemans and Williamson \yrcite{Goemans1995} using multiple random hyperplane projections.
Later, we will use correlation clustering to partially recover the maximum likelihood estimator of our planted partition model. Kollios et al. \yrcite{Kollios2013} consider a similar problem of using correlation clustering on probabilistic graphs, although their algorithm does not actually solve for the MLE.
\section{Problem Statement}
Consider observing an undirected graph $G = (V, E)$ with $n=|V|$ vertices. Let $\psi : \left\{1,\dotsc,n\right\} \rightarrow \left\{1, \dotsc, k\right\}$ be a partition of the $n$ vertices into $k$ classes. We use the notation $\psi_{ij}=1$ if nodes $i$ and $j$ belong to the same partition, and $\psi_{ij}=0$ else. Edges $e_{ij} \in \mathbb{R}^d$ are $d$-dimensional feature vectors. Note we say that graph $G$ is `observed,' though the edges $E$ may also be the result of a symmetric similarity function $s$, where $e_{ij} = s(v_i, v_j)$.
We assume a planted partition model, $e_{ij} \sim P(e | \psi_{ij})$. From now on, we will use the shorthand $P_0( \cdot) = P(\cdot | \psi_{ij} = 0)$ and $P_1( \cdot) = P(\cdot | \psi_{ij} = 1)$. In the conventional planted partition model, $P_0$ and $P_1$ are Bernoulli distributions with parameter $q$ and $p$, respectively. However, in this work we make no assumptions about the probability density functions $P_0$ and $P_1$. We will make the key assumption that all stochastic block models make -- that the edges are independent and identically distributed, conditioned on $\psi$. Note if the edges $E$ are generated by a similarity function then it is unlikely the edges are actually independent, but we proceed with this assumption regardless.
In most planted partition models, the goal is to either partially or exactly recover $\psi$ after observing $G$. We aim to find the most likely partition, and bound our performance in terms of the likelihood. There is a subtle distinction between the two goals. Even if the maximum likelihood estimator is consistent, the non-asymptotic MLE may be different than the true partition $\psi$.
\section{Approximating the Maximum Likelihood Estimator}
Let $\theta : \left\{1,\dotsc,n\right\} \rightarrow \left\{1, \dotsc, k\right\}$ be a partition under consideration. In exact recovery, our goal would be to find a $\theta$ such that $\theta = \psi$. However, our goal is to find a partition $\hat \theta$ which is close to the likelihood of the maximum likelihood estimator $\theta_{MLE}$. Using the edge independence assumption, the likelihood $L$ is
\begin{equation}
L(\theta) = \prod_{i < j} P(e_{ij} | \theta_{ij}) \mathbbm{1}(\theta \in \Theta)
\end{equation}
where $\Theta$ is the space of all disjoint partitions.
The trick to finding an approximation $\hat \theta$ to the MLE $\theta_{MLE}$ is to reduce the problem to a correlation clustering instance. Consider forming a graph $G_O = (V, E_0)$ with binary edges defined by the sign of the log-odds $e_{0; ij} = sign\left(\log\left(P_1(e_{ij})/P_0(e_{ij})\right)\right)$. Let the cost of mislabeling each edge be the absolute log-odds $C_{ij} = \left|\log\left(P_1(e_{ij})/P_0(e_{ij})\right)\right|$. Then we can rewrite the log-likelihood $\ell$ as\footnote{We are playing fast and loose with the $\mathbbm{1}(\theta \in \Theta)$ terms here. $G_0$ is not required to be a valid partition and thus the $\mathbbm{1}(G_0 \in \Theta)$ term is not included in $\ell(G_0)$. However, $\theta$ is still required to be a valid partition.}
\begin{align}
\ell(\theta) &= \ell(G_0) - \sum_{\theta_{ij} \neq e_{0; ij}} \left|\log\left(\frac{P_1(e)}{P_0(e)}\right)\right| \nonumber \\
&= \ell(G_0) - \sum_{\theta_{ij} \neq e_{0; ij}} C_{ij} \label{eq:log-likelihood}
\end{align}
Maximizing $\ell(\theta)$ is equivalent to minimizing $\sum_{\theta_{ij} \neq e_{0; ij}} C_{ij}$, which is exactly \textsc{MinimizeDisagreements} where edges are labeled according to $E_0$ and have weighted costs $C\geq 0$. Intuitively, we consider the most likely graph $G_0$ (which is not a valid partition) and try to find the minimum number of weighted edge flips required to create a valid partition.
Unfortunately, we only have non-negativity bounds on the weights $C$. Thus we believe the only appropriate \textsc{MinimizeDisagreements} algorithm to solve Eq \ref{eq:log-likelihood} is the LP-rounding technique by Demaine et al. \yrcite{Demaine2006}.
\begin{theorem}
The above estimated clustering $\hat \theta$ is $c_1DIS\log(n)$-close to the log-likelihood of the true maximum likelihood estimator $\hat \theta_{MLE}$. This is an $\exp(-DIS(c_1\log(n)-1))$-approximation algorithm for the likelihood.
\end{theorem}
The constant $c_1 = 2+ 1/\log(n+1)$ is just slightly larger than 2. $DIS$ is a measure of disagreement between the graph $G_0$ and the optimal clustering, to be discussed shortly.
\begin{proof}
The results follow directly from Leighton \& Rao \yrcite{Leighton1999} and Demaine et al. \yrcite{Demaine2006}. Let $DIS$ be the optimal solution to \textsc{MinimizeDisagreements} on graph $G_0$ with weighs $C$. Then the log likelihood of the true MLE $\theta_{MLE}$ is
\begin{equation}
\ell(\theta_{MLE}) = \ell(G_0) - DIS
\end{equation}
Demaine et al. \yrcite{Demaine2006} showed an $c_1\log(n)$-approximation to \textsc{MinimizeDisagreements} on general weighted graphs. Thus the approximated MLE using this algorithm will yield
\begin{equation}
\ell(\hat \theta) \geq \ell(G_0) - c_1\log(n)DIS
\end{equation}
The approximation ratio result follows likewise.
\begin{align}
L(\hat \theta) &\geq L(G)\exp(-c_1\log(n)DIS) \nonumber \\
L(\theta_{MLE}) &= L(G)\exp(-DIS) \nonumber \\
\frac{L(\hat \theta)}{L(\theta_{MLE})} &\geq \exp(-DIS(c_1\log(n)-1)) \nonumber
\end{align}
\end{proof}
\subsection{Choosing Edge Features or Similarity Functions}
How to choose a similarity function remains a fundamental question in spectral clustering \cite{Luxburg2007}. A ``meaningful'' similarity function should have high similarity for samples belonging to the same cluster and low similarity for samples in different clusters, but how to judge that remains unclear. In practice, the radial basis function is commonly used and often provides favorable results. More precisely, we want to know which similarity functions make clustering easier and understand why they do.
This same question applies when doing edge feature selection. We want to choose features which are most informative for clustering and ignore the others. We can provide a more scientific answer to these questions by analyzing the $DIS$ coefficient.
\begin{theorem}
Let $n_0$ and $n_1$ be the number of inter and intra-cluster edges in $\psi$, respectively. Then
\begin{align}
\mathbb{E}[DIS] &= -n_1D_{KL}(P_1 || P_0)\Big|_{P_1 \leq P_0} \nonumber \\
& \qquad \qquad \qquad \qquad \quad - n_0D_{KL}(P_0 || P_1)\Big|_{P_0 \leq P_1}
\end{align}
where we use the notation $D(\cdot || \cdot )\Big|_{S}$ to denote the divergence evaluated only over the closed set $S$.
\end{theorem}
\begin{proof}
\begin{align}
\mathbb{E}[DIS] &= (n_1 + n_2)\mathbb{E}_{\psi_{ij} \neq e_{0; i,j}}[C] \nonumber \\
&= n_1\mathbb{E}_{\psi_{ij}=1, e_{0;i,j}=0}[C] + n_2\mathbb{E}_{\psi_{ij}=1, e_{0;i,j}=0}[C] \nonumber \\
&= n_1\int_{P_1(e) \leq P_0(e)} P_1(e)\log\left(\frac{P_0(e)}{P_1(e)}\right)de \nonumber \\
& \quad + n_2\int_{P_0(e) \leq P_1(e)} P_0(e)\log\left(\frac{P_1(e)}{P_0(e)}\right)de \nonumber \\
&= -n_1D_{KL}(P_1 || P_0)\Big|_{P_1 \leq P_0} \nonumber \\
& \qquad \qquad \qquad \qquad \quad- n_0D_{KL}(P_0 || P_1)\Big|_{P_0 \leq P_1}\nonumber
\end{align}
\end{proof}
Notice these restricted Kullback-Leibler divergences are always negative, and thus $\mathbb{E}[DIS] \geq 0$.
The intuition here is to choose edge features or similarity functions which are unlikely to create edges in the disagreement regions (i.e.\ edges which contribute to $DIS$). If $P_0$ and $P_1$ are completely divergent, then exactly recovering the partition is trivial because $G_0$ will be the set of disconnected cliques induced by $\psi$. Additionally, when mistakes are made, we want the KL divergence to be small (i.e.\ the mistake is not too `bad').
Along these lines, choosing higher dimensional edge features and similarity functions (e.g.\ the absolute vector difference instead of the Euclidean distance) makes clustering easier, by decreasing the disagreement region between $P_0$ and $P_1$. This confirms our earlier motivation that useful clustering information may be lost by only considering binary or scalar edge features and functions.
Considering only this approximation ratio when choosing a similarity function or edge features does not quite capture the complete picture. A trivial solution is select for $P_0 = P_1$ (a type of an Erd\H{o}s-R\'enyi random graph), which results in an approximation ratio of 1. Since every partition is equally likely in this scenario, finding an approximation to the MLE is trivial. However, exactly recovering $\psi$ is unlikely.
\subsubsection{Sparsity}
In many situations it is advantageous to induce sparsity into the graph $G$. Spectral clustering employs this trick to cluster large graphs, by only considering the most similar nodes when performing eigen decompositions. In the proposed approach, sparsity will also reduce the number of variables and constraints in the LP used to maximize Eq \ref{eq:log-likelihood}.
By the previous analysis, we want to choose a similarity function or edge features which achieve the desired sparsity while maintaining a small $DIS$ coefficient. In the \textsc{MinimizeDisagreements} problem, sparse edges will have cost $C_{ij} = 0$. This occurs when $P_0(e_{ij}) = P_1(e_{ij})$. Intuitively, the best edges to sparsify are the ones which we do not have strong evidence for whether they should be labeled positive or negative. For these edges, the probabilities $P_0$ and $P_1$ will be close. Unlike spectral clustering, which only considers the most similar edges, this sparsification considers the most similar \emph{and} the most dissimilar edges.
\section{Experiments}
\begin{figure*}[p]%
\centering
\subfigure[True Clusters]{\includegraphics[width=0.25\linewidth]{crossbones_true.eps}}\qquad
\subfigure[k-means]{\includegraphics[width=0.25\linewidth]{crossbones_kmeans.eps}}\quad
\subfigure[Spectral]{\includegraphics[width=0.25\linewidth]{crossbones_spectral.eps}}\qquad\\
\subfigure[This Paper]{\includegraphics[width=0.25\linewidth]{crossbones_mle.eps}}\qquad
\subfigure[Cluster Structure $\hat P_1(e)$]{\includegraphics[width=0.25\linewidth]{crossbones_P1.eps}}\quad
\subfigure[Non-Cluster Structure $\hat P_0(e)$]{\includegraphics[width=0.25\linewidth]{crossbones_P0.eps}}\qquad\\
\subfigure[True Clusters]{\includegraphics[width=0.25\linewidth]{grid_true.eps}}\qquad
\subfigure[k-means]{\includegraphics[width=0.25\linewidth]{grid_kmeans.eps}}\quad
\subfigure[Spectral]{\includegraphics[width=0.25\linewidth]{grid_spectral.eps}}\qquad\\
\subfigure[This Paper]{\includegraphics[width=0.25\linewidth]{grid_mle.eps}}\qquad
\subfigure[Cluster Structure $\hat P_1(e)$]{\includegraphics[width=0.25\linewidth]{grid_P1.eps}}\quad
\subfigure[Non-Cluster Structure $\hat P_0(e)$]{\includegraphics[width=0.25\linewidth]{grid_P0.eps}}
\caption{Results on synthetic 2D datasets. The true number of clusters $k$ is given as an input to k-means and spectral clustering, while our model consistently and naturally learns the correct number of clusters. The learned edge densities $P_0$ and $P_1$ are also shown.}
\label{fig:2D}
\end{figure*}
\begin{figure*}[t]%
\centering
\subfigure[True Clusters]{\includegraphics[width=0.4\linewidth]{skin_true.eps}}\qquad
\subfigure[k-means]{\includegraphics[width=0.4\linewidth]{skin_kmeans.eps}}\quad \\
\subfigure[Spectral]{\includegraphics[width=0.4\linewidth]{skin_spectral.eps}}\qquad
\subfigure[This Paper]{\includegraphics[width=0.4\linewidth]{skin_mle.eps}}\qquad
\caption{Results on the UCI Skin Segmentation dataset. Pink represents skin samples and blue represents non-skin samples. The axes correspond to RGB pixel values. The associated normalized mutual information scores are (b) 0.0042, (c) 0.1016 and (d) 0.6804.}
\label{fig:3D}
\end{figure*}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{motor_nmi.eps}
\cprotect\caption{Structured clustering with PCA (\verb!SC + PCA!) outperforms competitors on the 11 class, 48 dimensional UCI Sensorless Drive dataset. \verb!1D SC! is the same structured clustering model, except using the simpler Euclidean similarity function ($e_{ij} = ||v_i - v_j||_2$). Boxes correspond to the $25^{th}$ and $75^{th}$ percentile of 17 trials. Whiskers are the most extreme values.}
\label{fig:motor}
\end{figure}
We will experimentally demonstrate the performance of the proposed model and algorithm on several synthetic and real world datasets. Specifically, we show studying edge features enables learning the \emph{structure} of clusters. When compared to k-means and spectral clustering, the planted partition model with general edge features can correctly cluster some rather interesting examples which are not attainable with scalar similarity functions. And more importantly, it seems to outperform existing methods on real world datasets.
In practice, it is unlikely we have access to $P_0$ and $P_1$. By assuming a prior parametric distribution, previous approaches have inferred these distributions while simultaneously learning the clustering. To remain as general as possible, we do not make any prior assumptions on $P_0$ and $P_1$. Our focus here is different. In datasets where the number of classes is very large or when new previously unseen classes are introduced, it is unlikely we can perform traditional supervised classification. This is especially true in the entity resolution and record linkage domains, where clusters correspond to millions or billions of entities (e.g. people, businesses, items) and new entities are frequently introduced. Statistical networks and image segmentation exhibit this property.
In these problems, we frequently have access to labeled \emph{pairs}. Human adjudicators are often able to say ``These two samples describe the same person'' or conversely ``these two samples describe different people.'' This is the information we use to learn the cluster structure. Thus we assume we have access to some labeled pairs in order to learn $P_0$ and $P_1$. Standard dimensionality reduction techniques can be employed to perform analysis in a reasonable space. We show this is still more powerful than using conventional scalar similarity functions. For all of our experiments, we use kernel density estimation to estimate $P_0$ and $P_1$. To improve the dimensional scalability, we could also perform a single estimation of $P_1/P_0$ using direct density ratio estimation \cite{Kanamori2009}
To compare performance we evaluate against k-means and spectral clustering \cite{Shi2000}. Per the recommendations of Luxburg \yrcite{Luxburg2007}, we use the Gaussian similarity function, a mutual $k$-nearest neighbor graph where $k=20$ and the random walk graph Laplacian. Unless otherwise noted, the edge features used for our method are from the absolute vector difference function $e_{ij} = |x=v_i - v_j|$ and thus not independent. However, the results indicate that it may be an acceptable assumption.
Lastly, our model consistently and naturally learns the correct number of clusters $k$. We found it occasionally labeled outlier samples as singleton clusters, though this would have a very small impact on the normalized mutual information score. For k-means and spectral clustering we do provide $k$ as an input. There are certainly methods of estimating $k$ for these competitors (e.g. analyzing the spectral gap), although they are not intrinsic to the methods.
\subsection{Results on Synthetic Data}
We consider the two interesting synthetic examples shown in Figure \ref{fig:2D}. Traditional clustering algorithms such as k-means and spectral clustering are unable to correctly label these examples because the clusters occasionally cross each other. Our method is able to capture the unidirectional cluster structure, and thus correctly label the samples. This is not an occasional event, in fact we have yet to see our method fail on these examples.
For all the synthetic experiments, we estimated $P_0$ and $P_1$ using 5,000 labeled pairs and clustered 100 hold-out samples. We use the absolute vector difference as our similarity function, which is able to capture the distance \emph{and} direction, unlike the Gaussian similarity function. There may be other excellent choices for similarity function, this is the only one we have tried so far.
We have achieved comparable results on the classic Gaussian, two moons, concentric circles and swiss roll examples. There was little distinction between our method and spectral clustering on these problems, so they were omitted from this paper.
\subsection{Results on Real World Data}
The first real world data we consider is the UCI Skin Segmentation dataset\footnote{https://archive.ics.uci.edu/ml/datasets/Skin+Segmentation}, shown in Figure \ref{fig:3D}. Samples are RGB values and labeled according to whether they are skin or non-skin image pixels. Again, we estimated $P_0$ and $P_1$ using 5,000 labeled pairs and clustered 100 hold-out samples, and use the absolute vector difference similarity function.
Visually, this seems much easier than the previous synthetic examples. However, k-means and spectral clustering are still unsuccessful due to the data scale issue introduced by the oblong cluster nature. Feature whitening did not help the competitors, though we believe some extensions to the standard spectral clustering may be able to handle this type of data \cite{Zelnik2004}.
The second realistic example we consider is the UCI Sensorless Drive Diagnosis dataset\footnote{https://archive.ics.uci.edu/ml/datasets/Dataset+for\\+Sensorless+Drive+Diagnosis}. Features are derived from current and frequency measurements in defective electric motors, including the statistical mean, standard deviation, skewness and kurtosis of intrinsic mode function subsequences. In total, there are 48 features and 11 classes.
We repeat the same previous procedures, except we additionally perform principal component analysis on the training and hold-out edge features prior to estimating $P_0$ and $P_1$ (\verb!SC + PCA!). We also consider one dimensional features using the Euclidean distance similarity function (\verb!1D SC!). The results from 17 trials are shown in Figure \ref{fig:motor}. The strong performance on the PCA reduced edge features leads us to believe that even if the original vertices have high dimensional structure, the distinguishing edge features in clusters have a lower dimensional representation.
\section{Conclusions}
Overall, incorporating multivariate edge features and more powerful similarity functions improves performance in all the experiments we have conducted. And even when the edges are clearly not independently generated, our structured clustering model still outperforms competitors.
The key insight from our approach is that multidimensional edge features can be used to effectively learn structure in clusters. Relationships in real world data are more complex than a simple scalar similarity function, and our methods can benefit from capturing that additional complexity. Then we can use the learned cluster structure to both determine the correct number of clusters and to handle situations where we are given new, previously unseen clusters, by assuming similar structure.
Applications which may especially benefit from structured clustering usually (a) have some labeled edges to learn $P_0$ and $P_1$ and (b) have a large number of clusters which make training a supervised classifier impractical. For example, in community detection and entity resolution, we have many examples of communities or entities to learn $P_0$ and $P_1$, though we certainly do not have examples of every community and entity to perform classification. Intuitively, we expect communities and entities to exhibit some common behavior, and we can leverage this structure while clustering. In image segmentation we usually have many images with human labeled segments, but the segments (i.e.\ classes) in new images are likely of a different object. However, it is not unreasonable to assume the \emph{structure} of these new segments is similar to the previously seen segments.
We used the approximation algorithm by Demaine et al. \yrcite{Demaine2006} for \textsc{MinimizeDisagreements}. The solution for this sub-problem is not the main focus of this paper, and unfortunately this particular algorithm requires solving a large linear program which limited the scalability of our experiments.
Pan et al. recently clustered 1 billion samples in 5 seconds using a parallelizable, linear time algorithm for \textsc{MinimizeDisagreements}, but only with edge weight restrictions \yrcite{Pan2015}.
In future work, we intend to provide spectral solutions to this same problem, which we believe will help address scalability and provide better theoretical insights into the exact recoverability of the partition. A major goal of analyzing exact recoverability is understanding how to design smart, sparse and simple similarity functions.
Other interesting extensions include applying the same method to stochastic block models, which would require estimating a separate $P_0$ and $P_1$ for every pair of blocks. In record linkage problems the same technique could be used to cluster vertices with different feature types. For example, clustering \emph{across} multiple social networks is of particular interest for advertising and law enforcement.
We have independently provided similar analysis for \textsc{MaximizeAgreements} by extending the results of Swamy \yrcite{Swamy2004} and Charikar et al. \yrcite{Charikar2003} to graphs with negative edge weights, though the theoretical and experimental results were not as convincing as \textsc{MinimizeDisagreements}, so thus omitted from this paper.
|
\section{Introduction}\label{Sec1}
The great achievements in manipulating and probing ultracold atoms in a cavity offer a new platform for exploring many-body systems and their dynamics.
Up to now, the strong atom-cavity coupling has been experimentally demonstrated in several laboratories by using an atomic Bose-Einstein condensate (BEC)~\cite{Brennecke2007, Colombe2007, Goldwin2011, Baumann2011, Ritsch2013, Chen2015}.
The BEC-cavity experiments can be used to study the coupled atom-photon system with both large and small atom and photon numbers~\cite{Colombe2007, Brennecke2007, Goldwin2011, Chen2015}.
The BEC-cavity system provides an ideal platform for exploring exotic many-body quantum effects, such as quantum phase transition~\cite{Mekhov2007, Ritsch2013,Larson2008, Nagy2010, Baumann2010, Baumann2011, Vacanti2012, Bakhtiari2015}, many-body quantum entanglement~\cite{Haas2014, Sun2014, McConnell2015}, precision measurement~\cite{Zuppardo2015, Hamilton2015}, and many-body quantum dynamics~\cite{Alvermann2012, Garraway2012}.
On the other hand, as an important fundamental problem, the many-body Landau-Zener (LZ) tunneling with BEC had attracted much attention for a long time~\cite{Wu2000,Chen2011, Qian2013, Zhong2014}.
However, the many-body LZ tunneling in BEC-cavity system, in particular, how the cavity field affects the LZ dynamics of atoms, is still unclear.
Due to the atom-atom interaction, the many-body LZ tunneling is very different from the single-particle one.
The sequential LZ tunneling induced by interaction blockade have been found in Bose-Josephson system (BJS)~\cite{Lee2008, Cheinet2008}, Bose-Hubbard ladder~\cite{Deng2015} and spinor BEC within optical superlattices~\cite{Wagner2011}.
For an example, in a BJS~\cite{Gati2007, Lee2009, Gross2010, Lee2012}, due to the interplay between inter-mode bias and atom-atom interaction, the resonant tunneling and interaction blockade take place~\cite{Lee2008, Cheinet2008}.
There appears a population ladder indicating a series of interaction blockade.
The variance of the relative population would exhibit several peaks representing the single-atom resonant tunneling.
Due to the fixed Josephson coupling in such a BJS, the step slopes of the ladder and the resonant tunneling peaks are symmetric about the zero bias (which corresponds to the zero detuning).
In a BEC-cavity system, due to the quantization nature of the cavity field, the picture of the sequential LZ dynamics may have significant differences.
In comparison with the BJS, the Rabi frequency in a BEC-cavity system acts the role of the Josephson coupling.
Unlike the Josephson coupling, the Rabi frequency is no longer fixed and it is proportional to the square root of the cavity-photon number~\cite{Fink2008}.
In the case of few photons, as the photon number changes in the LZ process, the Rabi frequency dramatically changes when the detuning is tuned.
Furthermore, if there is no sufficient photons to excite all the atoms to the upper level, an incomplete sequential LZ process would occur.
Thus, in addition to the atom-atom interaction and the detuning, the photon number also plays an important role in the LZ process of the BEC-cavity system.
Simultaneously, during the time-evolution, the atoms may also have an influence on the cavity field.
Therefore, it is of great interest to investigate: (i) how the sequential LZ dynamics is affected by the cavity field? and (ii) how the cavity field changes during such a LZ process?
In this article, we study the LZ process of an ensemble of interacting Bose condensed atoms trapped in a cavity.
In the framework of second quantization, the BEC-cavity system can be described by a modified Tavis-Cummings model with an additional nonlinear term determined by atom-atom interaction~\cite{Garraway2012, Feng2015}.
If the initial cavity field is described by a Fock state, the BEC-cavity system only involves a single excitation number.
In the adiabatic sweeping of the detuning, the interplay between the detuning and the atom-atom interaction results in the asymmetric sequential LZ transitions, in which the atoms absorb or emit photons one by one and the sequential LZ transitions are asymmetric.
If the initial cavity field is described by a coherent state, the BEC-cavity system will involve multiple excitation numbers.
Different from the case of single excitation numbers, in the asymmetric sequential LZ process, the atoms may no longer absorb entirely one photon for every population step.
The amount of absorbed/emmited photons becomes less than one and it gradually decreases even in the adiabatic limit.
During the LZ process, the cavity field is no longer a coherent state and its intermittent collapse and revival appear.
This article is organized as follows.
In Sec.~\ref{Sec2}, we describe the model and give its Hamiltonian.
In Sec.~\ref{Sec3}, we investigate the asymmetric sequential LZ process involving single excitation numbers.
In Sec.~\ref{Sec4}, we consider the atoms initially coupled with a coherent cavity field, and then study the asymmetric sequential LZ process involving multiple excitation numbers. We derive an analytical formula for sequential ladders of the photon number and the relative atom number. The adiabaticity conditions for the sequential LZ transitions are also analytically estimated. An intriguing findings of fractional steps are observed when the average number is small compared with atom number. In addition, we analyze the effects of the cavity photon loss on the sequential LZ dynamics. The phenomena of collapse and revivals of the cavity field are also revealed.
Finally, in Sec.~\ref{Sec5}, we give a brief summary and discussion.
\section{Model}\label{Sec2}
We consider an ensemble of Bose condensed atoms occupying two different hyperfine levels and trapped in an optical cavity~\cite{Colombe2007, Brennecke2007}.
The sketch is shown in Fig.~\ref{Fig-schematic}.
The atoms are coupled with a cavity mode, in which one atom in the lower level may absorb a photon and jump to the upper level, or may be taken from the upper level into the lower level with a photon emitted.
\begin{figure}[htb]
\centering
\includegraphics[width=\columnwidth]{Fig1.eps}\caption{(Color online) Sketch of interacting Bose condensed atoms trapped in an optical cavity. Each atom is identically coupled to the cavity mode with coupling strength $g/\sqrt{N_a}$. Here, the frequencies of the cavity field and the atomic transition are denoted by $\omega$ and $\omega_0$, respectively. The detuning is $\delta=\omega-\omega_0$ and the effective atom-atom interaction is characterized by $E_c$.}
\label{Fig-schematic}
\end{figure}
Assuming that the coupling fields are spatially uniform and the atom-atom collision do not change the internal states. Each atom identically couples with a single-mode cavity cavity field~\cite{Lee2014}. Therefore, the system obeys a second quantized Hamiltonian within a single-mode approximation (in the unit of $\hbar=1$ throughout the paper)
\begin{eqnarray}\label{H1}
H &=& \omega \hat a^{\dagger} \hat a +\sum_{j=\uparrow,\downarrow}(E_{0j} \hat b^{\dagger}_{j} \hat b_{j} + \frac{1}{2}G_{jj} \hat b^{\dagger}_{j} \hat b^{\dagger}_{j} \hat b_{j} \hat b_{j})\nonumber \\
&+& G_{\uparrow \downarrow}\hat b^{\dagger}_{\downarrow} \hat b^{\dagger}_{\uparrow} \hat b_{\uparrow} \hat b_{\downarrow} + \frac{g}{\sqrt{N_a}}(\hat a \hat b^{\dagger}_{\uparrow} \hat b_{\downarrow}+\hat a^{\dagger} \hat b^{\dagger}_{\downarrow} \hat b_{\uparrow}),\\
&&\nonumber
\end{eqnarray}
where $\omega$ is the frequency of cavity mode, $N_a=N_{\uparrow}+N_{\downarrow}=\hat b^{\dagger}_{\uparrow} \hat b_{\uparrow} + \hat b^{\dagger}_{\downarrow}\hat b_{\downarrow}$ is the atom number, $E_{0j}$ is the hyperfine energy of the atoms in state $|j\rangle$, $G_{jj}$ is the atom-atom interaction energies of state $|j\rangle$, $G_{\uparrow\downarrow}$ is the atom-atom interaction energy between states $|\uparrow \rangle$ and $|\downarrow \rangle$, and $g$ is the homogenous coupling strength between the cavity mode and the Bose condensed atoms.
Regarding all the atoms as spin-1/2 particles, one can define the angular momentum operators as
\begin{eqnarray}\label{JxJyJz}
\hat J_{x} &=& \frac{\hat b^{\dagger}_{\uparrow} \hat b_{\downarrow}+ \hat b_{\uparrow} \hat b^{\dagger}_{\downarrow}}{2}, \\
\hat J_{y} &=& \frac{\hat b^{\dagger}_{\uparrow} \hat b_{\downarrow}- \hat b_{\uparrow} \hat b^{\dagger}_{\downarrow}}{2i}, \\
\hat J_{z} &=& \frac{\hat b^{\dagger}_{\uparrow} \hat b_{\uparrow}- \hat b^{\dagger}_{\downarrow} \hat b_{\downarrow}}{2}.
\end{eqnarray}
The system has two good quantum numbers: the atom number $N_a$ and the excitation number $N_e=\hat a^{\dagger} \hat a+N_{\uparrow}=\hat a^{\dagger} \hat a+\hat J_z+\frac{N_a}{2}$, satisfying the commutation $[N_{a},H]=0$ and $[N_{e},H]=0$.
Here, when the excitation number is much larger than the atom number, i.e., $N_e \gg N_a$, it corresponds to the large excitation number condition, which is close to the classical field limit.
While if the excitation number is comparable with or smaller than the atom number, i.e., $N_e \lesssim N_a$, it refers to the small excitation number condition.
The quantization of cavity mode would make a difference with a classical field and the photon number may affect the properties of the whole system.
We will carefully compare the results under large and small excitation conditions in the following.
Since $N_a$ and $N_e$ are conserved quantities, the constant terms $\mathcal{O}(N_a)$ and $\mathcal{O}(N^2_a)$ can be eliminated and the Hamiltonian becomes
\begin{equation}\label{H2}
H = \omega_0 \hat J_z + \omega \hat a^{\dagger} \hat a + \frac{E_c}{2}\hat J^2_{z} +\frac{g}{\sqrt{N_a}}(\hat a^{\dagger} \hat J_{-}+\hat a \hat J_{+}),
\end{equation}
where $\hat J_{\pm}=\hat J_x \pm i\hat J_y$ are the raising and lowering operators of atoms, the atomic transition frequency $\omega_0=E_{0\uparrow}-E_{0\downarrow}+\frac{1}{2}(N_a-1)(G_{\uparrow\uparrow}-G_{\downarrow\downarrow})$, and the effective atom-atom interaction $E_c = G_{\uparrow\uparrow}+G_{\downarrow\downarrow}-2G_{\uparrow\downarrow}$.
Further, shifting the zero point of the Hamiltonian~\eqref{H2}, one can get
\begin{equation}\label{H3}
H = -\frac{\delta}{2} \hat J_z + \frac{\delta}{2} \hat a^{\dagger} \hat a + \frac{E_c}{2}\hat J^2_{z} +\frac{g}{\sqrt{N_a}}(\hat a^{\dagger} \hat J_{-}+\hat a \hat J_{+}),
\end{equation}
with the detuning $\delta=\omega-\omega_0$. According to the last term, for fixed $g$ and $N_a$, the Rabi frequency is proportional to the square root of photon number (i.e., $\Omega\propto\sqrt{n}$), which is changeable with respect to the state of cavity field.
The state of the system can be expanded as
\begin{equation}\label{Psi}
|\psi\rangle=\sum_{n,m} C_{n,m} |n,m\rangle,
\end{equation}
where $n$ is the photon number of the Fock state, $m=(N_{\uparrow}-N_{\downarrow})/2$ is the relative atom number and $C_{n,m}$ are the coefficients of basis $|n,m\rangle$.
Due to the conservation of excitation number, when $N_e < N_a$, the number of bases can be reduced to $N_e+1$, the state of the system can be expanded as
\begin{equation}\label{Psi1}
|\psi\rangle=\sum_{l=0}^{N_e} C_{l}^{N_e} |N_e-l,-\frac{N_a}{2}+l\rangle.
\end{equation}
While for $N_e \ge N_a$, only $N_a+1$ bases are enough, and the state of the system can be expanded as
\begin{equation}\label{Psi2}
|\psi\rangle=\sum_{l=0}^{N_a} C_{l}^{N_e} |N_e-l,-\frac{N_a}{2}+l\rangle.
\end{equation}
Here, $C_l^{N_e}$ in Eq.~\eqref{Psi1} and~\eqref{Psi2} denotes the coefficient with $N_e-l$ photons and $l$ atoms populating in the upper level.
\section{Landau-Zener Process involving single excitation numbers}\label{Sec3}
Firstly, we consider the system with fixed single excitation numbers and study the static properties of the system in small and large excitation number conditions.
Starting from Hamiltonian~\eqref{H3}, we choose atom number $N_a=10$ and investigate the situations with excitation numbers $N_e=5$, $N_e=10$ and $N_e=1000$. Here, the system state $|\Psi\rangle$ can be expressed according to Eq.~\eqref{Psi1} or Eq.~\eqref{Psi2}. We vary $\delta$ from negative detuning to positive detuning and obtain the corresponding groundstates.
The detuning $\delta$ tends to populate the atoms into the upper level while the repulsive atom-atom interaction ($E_c>0$) tends to populate the atoms equally in both levels.
For far negative detuning, i.e., $\omega \ll \omega_0$, $\delta\rightarrow -\infty$, the groundstate is $|N_e,-\frac{N_a}{2}\rangle$, which corresponds to that all atoms are in the lower level and the photon number equals to the excitation number $N_e$.
For far positive detuning, i.e., $\omega \gg \omega_0$, $\delta\rightarrow \infty$, the form of groundstate is related to the excitation number. When $N_e\ge N_a$, the groundstate is $|N_e-N_a,\frac{N_a}{2}\rangle$, which is the state that all atoms occupy in the upper level and the photon number becomes $N_e-N_a$. When $N_e<N_a$, it becomes $|0,-\frac{N_a}{2}+N_e\rangle$, where only a portion of atoms occupy in the upper level and no photons remain.
In other regions of detuning, the groundstate depends sensitively on the parameters $g$, $\delta$ and $E_c$.
For better revealing the properties of the system, the expectations of photon number $\langle n \rangle=\langle \psi|\hat a^{\dagger} \hat a |\psi\rangle$ and the half relative atom number $\langle N_{\uparrow}-N_{\downarrow}\rangle/2=\langle \psi| \hat J_z |\psi\rangle$, and the variance of relative atom number $\text{Var} (N_{\uparrow}-N_{\downarrow})=4(\langle \hat J_{z}^2 \rangle-\langle \hat J_{z} \rangle^2)$ for the groundstate are calculated and shown in Fig.~\ref{Fig-single-excitation}.
\subsection{Asymmetric sequential population ladders}
\begin{figure*}[htb]
\centering
\includegraphics[scale=0.73]{Fig2.eps}\caption{(Color online) Static properties of the system involving single excitation numbers. The top, middle and bottom rows correspond to the cases of $N_e=5$, $N_e=10$ and $N_e=1000$. The left and right columns represent the expectations of half relative atom number $\langle N_{\uparrow}-N_{\downarrow}\rangle/2=\langle \hat J_z \rangle$, and variance of relative atom number $\text{Var} (N_{\uparrow}-N_{\downarrow})=4(\langle \hat J_{z}^2 \rangle-\langle \hat J_{z} \rangle^2)$ versus detuning, respectively. Here, $N_a=10$ and $g=1$.}
\label{Fig-single-excitation}
\end{figure*}
There appear a series of ladder-like steps for the half relative atom number.
These sequential population ladders are induced by the atom-atom interaction. The steps become steeper and the plateaus get more smooth when the atom-atom interaction becomes larger.
The height of every step equals exactly 1. For every step, one of the atoms in the lower level can absorb one photon and be excited to the upper level.
The appearance of sequential population ladders is similar to the interaction blockade observed in the Bose-Josephson junctions~\cite{Lee2008, Cheinet2008}.
For a two-component Bose-Einstein condensate linked by classical Raman fields, the competition between detuning and nonlinear interaction will result in a sequential population ladder for the relative atom number, where the slopes of the steps are exactly symmetric about $\delta=0$.
However, for the two-state Bose condensed atoms coupled with a quantized cavity mode, the number of the photons make the shape of the ladders significantly different.
When the excitation number is smaller than atom number $N_e<N_a$, the structure is totally asymmetric since there are not enough photons to excite all the atoms and the population ladder ceases to increase when there are no photons inside the cavity, see Fig.~\ref{Fig-single-excitation}~(a).
In the situation of $N_e=10$, the step slopes of the population ladders are also no longer symmetric about $\delta=0$.
For the same magnitude of detuning $|\delta|$, the step slope under negative detuning $-\delta$ is always less sharper than the one under positive detuning $+\delta$.
The asymmetry of the step slopes comes from the changing Rabi frequency during the sequential LZ processes. Adiabatically sweeping the detuning from negative to positive, the photon number decreases downstairs from $N_e$ to 0. Therefore, the Rabi frequency $\Omega$ is getting smaller. The sharpness is determined by the ratio between Rabi frequency and the atom-atom interaction $\Omega/E_c$, where the smaller $\Omega/E_c$ corresponds to the sharper step slope. Specially, when the photon number decreases to zero, the Rabi frequency vanishes, and no transitions would happen.
Since the photon number under negative detuning $-\delta$ is always larger than the one under positive detuning $+\delta$, the step slopes of the ladder for $\langle N_{\uparrow}-N_{\downarrow}\rangle/2$ is asymmetric about $\delta=0$. The first and the last step slopes have the biggest difference because they respectively correspond to the largest and smallest photon number, see the insets of Fig.~\ref{Fig-single-excitation}~(c). While for other pairs of step slopes, the difference is smaller.
As the excitation number increases, the photon number difference under $-\delta$ and $+\delta$ becomes less dramatic and the Rabi frequency $\Omega$ tends to be uniform.
For large excitation number $N_e=1000$, the step slopes of the ladder become almost symmetric about $\delta=0$, see the insets of Fig.~\ref{Fig-single-excitation}~(e). This is because when the photon number is extremely large compared with the atom number, the cavity field can be treated as a classical field with fixed complex number, and the sequential behaviors would return to the symmetric one in a BJS.
\subsection{Asymmetric variance of relative atom number}
For further investigation, we look in the variance of relative atom number $\text{Var} (N_{\uparrow}-N_{\downarrow})$.
There are a series of peaks for $\text{Var} (N_{\uparrow}-N_{\downarrow})$. The locations of the peaks correspond to the locations of the step slopes in the ladder of $\langle N_{\uparrow}-N_{\downarrow}\rangle/2$. The stronger atom-atom interaction $E_c$ induces sharper and narrower peaks, which is consistent with the sharpness of the step slopes.
When $N_e=5$, the peaks are completely asymmetric. There appear several peaks when $\delta<0$ and no peaks for $\delta>0$, see Fig.~\ref{Fig-single-excitation}~(b).
The asymmetry is also obvious for $N_e=10$, see Fig.~\ref{Fig-single-excitation}~(d). The total structure is asymmetric about $\delta=0$. The width and height of first and the last peaks are totally different. The last peak is much sharper than the first one, which is in accordance with the situation of the step slopes of population ladder.
The plateaus in the ladder of $\langle N_{\uparrow}-N_{\downarrow}\rangle/2$ indicate the interaction blockades. In the regions of the plateaus, the system is approximately in the state $|n,m\rangle$ with definite photon number and relative atom number. Therefore, the variance of the relative atom number are suppressed.
The resonance peaks are caused by the quasi-degeneracy between the two states $|N_e-k+1, -\frac{N_a}{2}+k-1\rangle$ and $|N_e-k, -\frac{N_a}{2}+k\rangle$ with $k=1,...,\text{min}(N_a,N_e)$ in the neighboring plateaus. In the intermediate region of the step slopes, the superposition of these two states leads to $\text{Var} (N_{\uparrow}-N_{\downarrow})$ increasing dramatically. One can treat the atom-photon coupling term as a perturbation~\cite{Lee2008, Cheinet2008} and obtain the locations of the peaks with $\delta/E_c= l+1/2$ for $l=\{-N_a/2, -N_a/2+1, ..., N_a/2-1\}$, which is in agreement with the numerical results.
More importantly, since the Rabi frequency becomes weaker when the photon number decreases, the relative atom number also shows asymmetric properties. When the Rabi frequency decreases, the perturbation term becomes smaller, and the two neighboring states $|N_e-k+1, -\frac{N_a}{2}+k-1\rangle$ and $|N_e-k, -\frac{N_a}{2}+k\rangle$ are closer to degeneracy. Therefore, the variance of the relative atom number would appear narrower and sharper peaks at the degeneracy location.
While for large excitation number, the asymmetry becomes much smaller. Under $N_e=1000$, the structure of the variance $\text{Var} (N_{\uparrow}-N_{\downarrow})$ become more symmetric about $\delta=0$, see Fig.~\ref{Fig-single-excitation}~(f).
The set of peaks that indicate the single-atom resonance transition can be used to design the single-atom device. For this system, varying the photon number in the cavity can adjust the sensitivity of the response and the atom-atom interaction need not be too strong, which can be promising for application in sensitive detection and high-precision metrology.
\section{Landau-Zener dynamics involving multiple excitation numbers}\label{Sec4}
\begin{figure*}[htb]
\centering
\includegraphics[scale=0.6]{Fig3.eps}\caption{(Color online) Dynamical evolution of the half relative atom number, variance of relative atom number and photon number. The atom number $N_a=4$ and the initial cavity field is in a coherent state with average photon number $|\alpha|^2=4$ and $|\alpha|^2=25$. The detuning is linearly swept according to $\delta(t)=\delta_0+\beta t$. The black solid lines are the adiabatic limit and other lines are numerical results with different sweeping rates $\beta$. In order to ensure the initial detuning is large enough, we choose $\delta_0=-3 E_c$ with $E_c=100$ for calculation.
}
\label{Fig-Dynamical-LZ}
\end{figure*}
In the following, we consider the ensemble of interacting Bose condensed atoms coupled with a cavity field in a coherent state. The coherent state of photons is natural and it is easy to be prepared in experiments.
We sweep the detuning to investigate the LZ dynamics of the system.
Initially, the atoms are all prepared in the lower level. The cavity field is a coherent state, which is the eigenstate of annihilation operator $\hat a$ and can be written as $|\alpha\rangle = e^{-\frac{|\alpha|^2}{2}}\sum_{n=0}^{\infty} \frac{\alpha^n}{\sqrt{n!}}|n\rangle$,
with $|\alpha|^2$ being the average photon number of the coherent state.
Therefore, the initial state of the whole system can be expressed as
\begin{equation}\label{Psi0}
|\Psi(0)\rangle=|\alpha, -\frac{N_a}{2}\rangle=\sum_{N_e=0}^{\infty} w_{N_e}|N_e,-\frac{N_a}{2}\rangle,
\end{equation}
which involves multiple excitation numbers with $w_{N_e}=e^{-\frac{|\alpha|^2}{2}} \frac{\alpha^{N_e}}{\sqrt{N_e!}}$.
During the sweeping, the time-evolution of the system state $\Psi(t)$ obeys the Schr\"{o}dinger equation,
\begin{equation}\label{SE}
i \frac{\partial |\Psi(t)\rangle}{\partial t}=H(t)|\Psi(t)\rangle,
\end{equation}
where $H(t)$ is the Hamiltonian~\eqref{H2} with time-dependent detuning $\delta(t)$.
The system state can be expanded as
\begin{equation}\label{Psi_ini}
|\Psi(t)\rangle=\sum_{N_e=0}^{\infty}w_{N_e}\left(\sum_{l}C_{l}^{N_e}(t)|N_e-l,-\frac{N_a}{2}+l\rangle\right),
\end{equation}
where $C_{l}^{N_e}(t)$ is the time-dependent coefficient of the basis in subspace with $N_e$ excitation number.
There are $N_e$ subspaces with different excitation numbers and the bases of each subspace depend on the excitation numbers.
Each subspace with different $N_e$ is orthogonal and decoupled with others. Therefore,
the dynamics in each subspace is independent and we can deal with the time-evolution problem individually and then sum up together according to the weight factor of each subspace $w_{N_e}$.
In addition, the average excitation number of the system is always conserved and equals the initial average photon number.
In our calculation, we fix the atom number $N_a=4$, coupling strength $g=1$ and atom-atom interaction $E_c=100$.
Meanwhile, the detuning $\delta$ is linearly swept according to $\delta(t)=\delta_0+\beta t$ with the initial detuning $\delta_0<0$ and the sweeping rate $\beta=\frac{2|\delta_0|}{\tau}$, where $\tau$ is the total sweeping time.
Here, we use $\beta$ to characterize the non-adiabaticity of the sweeping. The smaller sweeping rate $\beta$ refers to a slower sweeping of detuning, which means more probability staying in the instantaneous ground state when the LZ process occurs.
The properties of LZ dynamics with multiple excitation numbers may differ according to the distribution of the excitation numbers which is related to the average photon number of the initial coherent state.
Here, for initial coherent states, we choose the average photon number to be $|\alpha|^2=4$ and $|\alpha|^2=25$.
\subsection{Asymmetric sequential Landau-Zener dynamics}
The evolution of the photon number, the variance of relative atom number and the half relative atom number versus the detuning $\delta(t)$ for different sweeping rates $\beta$ are shown in Fig.~\ref{Fig-Dynamical-LZ}.
The three observables of the evolved state can be respectively calculated according to
\begin{equation}\label{PhoNum}
\langle n(t)\rangle=\langle \Psi(t)|\hat a^{\dagger} \hat a|\Psi(t)\rangle=\sum_{N_e,l} (N_e-l) |w_{N_e} C^{N_e}_{l}(t)|^2,
\end{equation}
\begin{eqnarray}\label{ExANum}
\frac{\langle N_{\uparrow}(t)-N_{\downarrow}(t)\rangle}{2}&=&\langle \Psi(t)|\hat J_{z}|\Psi(t)\rangle \nonumber\\
&=&\sum_{N_e,l} (-\frac{N_a}{2}+l) |w_{N_e} C^{N_e}_{l}(t)|^2,
\end{eqnarray}
and
\begin{eqnarray}\label{Variance}
&&\text{Var}(N_{\uparrow}(t)-N_{\downarrow}(t)) \nonumber\\
&=&4\left[\langle \Psi(t)|\hat J_{z}^2|\Psi(t)\rangle-\left(\langle \Psi(t)|\hat J_{z}|\Psi(t)\rangle\right)^2\right] \nonumber\\
&=&4\{\sum_{N_e,l}(-\frac{N_a}{2}+l)^2 |w_{N_e}C^{N_e}_{l}(t)|^2 \nonumber\\
&-&[\sum_{N_e,l}(-\frac{N_a}{2}+l) |w_{N_e} C^{N_e}_{l}(t)|^2]^2\}.
\end{eqnarray}
For $|\alpha|^2=4$, where the average photon number equals the atom number, the population ladder for half relative atom number coincides with the one in adiabatic limit when the sweeping rate is sufficiently small (e.g., $\beta=2$), see Fig.~\ref{Fig-Dynamical-LZ}~(a).
When the sweeping becomes fast (e.g., $\beta=10$), the population ladder oscillates and the heights of the steps become less obvious.
The sequential LZ transitions gradually disappear when the non-adiabatic effect is strong enough for extremely fast driving (e.g., $\beta=100$).
For $|\alpha|^2=25$, where the average photon number is much greater than the atom number, the sequential LZ dynamics versus the non-adiabatic effect is similar, see Fig.~\ref{Fig-Dynamical-LZ}~(b).
However, it does not require such slow driving to attain its adiabatic limit (e.g., $\beta=20$ is slow enough while for $|\alpha|^2=4$ is not).
That is because when the average photon number increases, the Rabi frequency as well as the gaps of the avoided energy level crossings become larger. The analytical analysis for adiabaticity condition is given in the next subsection.
The variances of the relative atom number for small and large $|\alpha|^2$ are very different, see Fig.~\ref{Fig-Dynamical-LZ}~(c) and (d).
For the adiabatic sweeping with $|\alpha|^2=25$, the variance $\text{Var}(N_{\uparrow}-N_{\downarrow})$ is similar to the one with large single excitation number shown in Fig.~\ref{Fig-single-excitation}~(f).
While for the adiabatic sweeping with $|\alpha|^2=4$, the variance $\text{Var}(N_{\uparrow}-N_{\downarrow})$ appears two peaks for the first and second LZ transitions and then shows a ladder-like structure with unequal step heights.
If the sequential LZ transitions for small $N_e$ are incomplete, the atomic state is a superposition state of different $m=(N_{\uparrow}-N_{\downarrow})/2$ and the corresponding variance $\text{Var}(N_{\uparrow}-N_{\downarrow})$ becomes large.
Thus, as the detuning increases, the systems is dominated by the incomplete sequential LZ transitions for small $N_e$, and the atomic number variance changes from peaks to ladder-like steps.
While for large $|\alpha|^2$, the atomic state in every subspace changes synchronously and the variance would appear a series of peaks due to the quasi-degeneracy.
When taking the non-adiabatic effects into account, the variances change dramatically.
For both $|\alpha|^2=4$ and $|\alpha|^2=25$, under a very slow sweeping, the variances approach to their
adiabatic limits.
Under intermediate sweeping, the atomic state becomes a superposition state of different
$m=(N_{\uparrow}-N_{\downarrow})/2$ after every LZ transitions.
As the LZ transitions occur sequentially, more different $m=(N_{\uparrow}-N_{\downarrow})/2$ components accumulate and induce the variance $\text{Var}(N_{\uparrow}-N_{\downarrow})$ to increase in ladder-like steps (e.g., the blue dash-dotted lines in Fig.~\ref{Fig-Dynamical-LZ}~(c) and (d)).
While for the fast sweeping, the system states hardly follow the instantaneous ground state and change slightly and the variances change more smoothly.
More interestingly, the sequential population ladders for small average photon number exhibit another kind of asymmetric property. The heights (but not the slopes) of the steps are obviously unequal and become fractional, see Fig.~\ref{Fig-Dynamical-LZ}~(a) and (e).
For $|\alpha|^2=4$, even though in the adiabatic limit, the height of each step is less than one and different with others. The relative atom population ladder increases upstairs with the height of the steps gradually becomes smaller and smaller.
The 4 photons on average could not be entirely absorbed by the 4 atoms even after 4 times of LZ transitions.
In contrast, for $|\alpha|^2=25$ in adiabatic limit, the population ladder of half relative atom number increases
upstairs with the each step's height equals nearly one, which is similar to the LZ process with single excitation numbers shown in Fig.~\ref{Fig-single-excitation}.
The coherent state with large average photon number can be approximately treated as a classical field.
When the average photon number is large enough, the initial atoms in the lower level could gradually absorb the photon one by one through every LZ transitions, , see Fig.~\ref{Fig-Dynamical-LZ}~(b) and (f).
It is worth to mention that, when $\delta(t)$ is swept adiabatically to 0, the evolved atomic state becomes a twin Fock state, in which the populations in upper and lower levels are exactly equal ($\langle N_{\uparrow}-N_{\downarrow}\rangle=0$) and the variance of the relative atom number $\text{Var}(N_{\uparrow}-N_{\downarrow})=0$.
The atomic twin Fock state is a highly entangled state that can be used to implementing quantum metrology~\cite{Lucke2011, Gerry2010}.
In our system, preparing the initial atomic state into $|-\frac{N_a}{2}\rangle$ and the cavity photon field into a coherent state with large $|\alpha|^2$, an atomic twin Fock state can be generated via an adiabatic linear sweeping from negative detuning to zero.
This offers a new scheme for quantum state engineering on highly entangled state preparation.
\subsection{Analytical analysis for sequential Landau-Zener dynamics}
\begin{figure}[htb]
\centering
\includegraphics[width=\columnwidth]{Fig4.eps}\caption{(Color online) The sequential LZ dynamics for photon number
in subspaces under different sweeping rate $\beta$. Here, $N_e$ are chosen from $1$ to $10$, $N_a$=4, $g=1$ and
$E_c=100$. The solid lines are obtained analytically according to Eq.~\eqref{Photon_LZ}. The dashed lines are
the numerical results.}
\label{Fig-LZ}
\end{figure}
To explore the sequential LZ dynamics, we derive an analytical formula for ladders of the photon number and the relative atom number.
In our model, the whole Hilbert space can be divided into several decoupled subspaces and we can investigate the dynamics individually in every subspace and finally sum up together.
For a subspace with excitation number $N_e$ and atom number $N_a$, it would respectively occur $N_a$ and $N_e$ times of LZ transitions for $N_e\ge N_a$ and $N_e< N_a$.
The $k$-th LZ transition happens between the states $|N_e-k+1, -\frac{N_a}{2}+k-1\rangle$ and $|N_e-k, -\frac{N_a}{2}+k\rangle$ with $k=1,...,\min(N_a,N_e)$.
Therefore, the sequential LZ transitions in a subspace can be treated as a sequence of conventional two-level LZ transitions.
In a conventional LZ transition of a sweeping rate $\beta$ with the minimum gap $\Delta$ for its avoided energy level crossing, and starting from its ground state at time $t\rightarrow-\infty$, the probability of finding the system remaining in the ground state at time $t\rightarrow+\infty$ is given by the LZ formula
~\cite{Landau1932, Zener1932},
\begin{equation}\label{LZ_formula}
P(\Delta,\beta)=1-\exp(-\frac{\pi\Delta^2}{2\beta}).
\end{equation}
By applying the conventional two-level LZ formula to each avoided energy level crossing, the probability of staying in the instant ground state at $k$-th LZ transition reads
\begin{equation}\label{P_LZ}
P(\Delta_k^{\!N_e},\beta)=1-\exp\left[-\frac{\pi\left(\Delta_k^{\!N_e}\right)^2}{2\beta}\right].
\end{equation}
where $\Delta_k$ denotes the minimum gap of the $k$-th LZ transitions, which corresponds to the energy gap at $\delta_k/E_c=-(N_a+1)/2+k$ with $k=1,2,...,\min(N_a,N_e)$.
The $k$-th minimum gap is proportional to the atom-photon coupling term of the Hamiltonian~\eqref{H3}, and it can be expressed in the form of
\begin{equation}\label{min_gap}
\Delta_k^{\!N_e}=2g\sqrt{\frac{k(N_a-k+1)(N_e-k+1)}{N_a}}.
\end{equation}
By applying the LZ formula~\eqref{P_LZ} one by one, we can obtain the final probability of staying in
the instant ground state after $k$ times of LZ transitions, which is given as
\begin{equation}\label{final_pro}
P_{N_e,k}(\beta)=\prod_{l=1}^{k} P(\Delta_l^{\!N_e},\beta).
\end{equation}
For every LZ transition, the probability of remaining in the ground state also corresponds to the probability of absorbing one photon.
From this point, we can figure out the heights of every population steps of the photon number during the sweeping.
The height of the $k$-th step equals exactly $P_{N_e,k}(\beta)$.
Therefore, the photon number of the plateaus in the ladder can be obtained analytically,
\begin{equation}\label{Photon_LZ}
\langle n \rangle_{\!N_e}=\left\{\begin{array}{lll}
N_e, \delta<\delta_1 \\
N_e-\sum_{m=1}^{\min(\!k,N_e\!)}P_{N_e,m}(\beta), \delta_k\le\delta<\delta_{k+1}\\
N_e-\sum_{m=1}^{\min(\!N_e,N_a\!)}P_{N_e,m}(\beta), \delta\ge\delta_{N_a}\\
\end{array}\right.
\end{equation}
Correspondingly, the half relative atom number can be expressed as
\begin{multline}\label{Jz_LZ}
\frac{\langle N_{\uparrow}-N_{\downarrow}\rangle_{\!N_e}}{2}\\
=\left\{\begin{array}{lll}
-\frac{N_a}{2}, \delta<\delta_1 \\
-\frac{N_a}{2}+\sum_{m=0}^{\min(\!k,N_e\!)}P_{N_e,m}(\beta), \delta_k\le\delta<\delta_{k+1}\\
-\frac{N_a}{2}+\sum_{m=0}^{\min(\!N_e,N_a\!)}P_{N_e,m}(\beta), \delta\ge\delta_{N_a}\\
\end{array}\right.
\end{multline}
where $\delta_k=\left[-(N_a+1)/2+k\right]E_c$ and the index $k=1,2,...,\min(N_a,N_e)$.
Given a sweeping rate $\beta$, one can figure out the population ladders analytically. In Fig.~\ref{Fig-LZ}, we plot the population ladders of photon number for $N_e=\{1, 2, ..., 10\}$ under sweeping rates $\beta=\{10,100\}$.
The dashed lines are the numerical results while the solid lines are obtained analytically according to Eq.~\eqref{Photon_LZ}. It is obvious that, the analytical population ladders are well consistent with the numerical ones.
Since the minimum gap $\Delta_{k}$ increases with excitation number $N_e$, it needs slower driving for smaller $N_e$. For fast sweeping $\beta=100$, the step height decreases obviously when $N_e$ changes from $10$ to $1$. While for slow sweeping $\beta=10$, the step heights for most $N_e$ approach to 1.
Therefore, for small $|\alpha|^2$, the dominated $N_e$ is small and requires slower sweeping to access the adiabaticity condition.
By summing up the results of all subspaces, the average photon number and the half relative atom number of the system are given as
\begin{equation}\label{sum_ph}
\langle n \rangle= \sum_{N_e=0}^{\infty} w_{N_e}^2 \langle n \rangle_{N_e},
\end{equation}
and
\begin{equation}\label{sum_at}
\frac{\langle N_{\uparrow}-N_{\downarrow}\rangle}{2}= \sum_{N_e=0}^{\infty} w_{N_e}^2 \frac{\langle N_{\uparrow}-N_{\downarrow}\rangle_{\!N_e}}{2}.
\end{equation}
In our system, the adiabatic condition can be estimated approximately according to
\begin{equation}\label{AdCondition}
\frac{\left[\min_k\left(\Delta_{k}^{|\alpha|^2-\min(\!|\alpha|^2,N_a\!)+1}\right)\right]^2}{2\beta} \gtrsim 2.
\end{equation}
For $|\alpha|^2=4$, $\beta\lesssim1$; for $|\alpha|^2=25$, $\beta\lesssim21$. This estimation can be confirmed by the comparison with numerical calculations, see Fig.~\ref{Fig-Dynamical-LZ}.
One of the interesting findings is the appearance of fractional steps in population ladders for small $|\alpha|^2$
even under adiabatic sweeping.
This peculiar phenomenon results from the property of coherent state with few average photon number. The initial coherent state is a superposition of infinite Fock state including the vacuum state with different weight coefficients $w_{N_e}$.
For small $N_e$, there is no sufficient photons to excite all the atoms to the upper level, and this incomplete sequential LZ process would contribute to the fractional steps if the weight coefficient $w_{N_e}$
is non-ignorable.
Below we give an analytical explanation for the fractional steps.
When the sweeping is adiabatic ($\beta\rightarrow 0$), all $P_{N_e,k}(\beta)\rightarrow 1$ and so that the average photon number in subspace of $N_e$ becomes
\begin{equation}\label{ph_ad}
\langle n \rangle_{\!N_e}=
\left\{\begin{array}{lll}
N_e, \delta<\delta_1 \\
N_e-\text{min}(k,N_e), \delta_k\le\delta<\delta_{k+1}\\
N_e-\text{min}(N_e,N_a), \delta\ge\delta_{N_a}\\
\end{array}\right.
\end{equation}
According to Eq.~\eqref{sum_ph}, the average photon number of the system can be obtained,
\begin{equation}\label{ph_coh_ad}
\langle n \rangle=
\left\{\begin{array}{lll}
\sum_{N_e=0}^{\infty} w_{N_e}^2 N_e, \delta<\delta_1 \\
\sum_{N_e=k}^{\infty} w_{N_e}^2(N_e-k), \delta_k\le\delta<\delta_{k+1}\\
\sum_{N_e=N_a}^{\infty} w_{N_e}^2(N_e-N_a), \delta\ge\delta_{N_a}\\
\end{array}\right.
\end{equation}
Using the condition that $\sum_{N_e=0}^{\infty} w_{N_e}^2 N_e=|\alpha|^2$, we can simplify Eq.~\eqref{ph_coh_ad} to
\begin{equation}\label{ph_coh_ad1}
\langle n \rangle=
\left\{\begin{array}{lll}
|\alpha|^2, \delta<\delta_1 \\
|\alpha|^2-k-\sum_{N_e=0}^{k-1} w_{N_e}^2(N_e-k), \delta_k\le\delta<\delta_{k+1}\\
|\alpha|^2-N_a-\sum_{N_e=0}^{N_a-1}w_{N_e}^2(N_e-N_a), \delta\ge\delta_{N_a}\\
\end{array}\right.
\end{equation}
From Eq.~\eqref{ph_coh_ad1}, we can further obtain the $k$-th step height of the photon number (as well as the relative atom number),
\begin{equation}\label{stepH}
h_{k}=1-\sum_{N_e=0}^{k-1}w_{N_e}^2.
\end{equation}
As a result, for small $|\alpha|^2$, the weight coefficients $w_{N_e}^2$ of small $N_e$ are dominated, which leads to the fractional steps.
For example, $|\alpha|^2=4$, $w_{0}^2=0.0183$, $w_{1}^2=0.0733$, $w_{2}^2=0.1465$, and $w_{3}^2=0.1954$ and we can get $h_1=0.9817$, $h_2=0.9084$, $h_3=0.7619$ and $h_4=0.5665$. The results are the same with the numerical findings.
While for large $|\alpha|^2$, the weight coefficients $w_{N_e}^2$ of small $N_e$ are nearly zero, and the height of every step equals 1.
Since the total excitation number of the system is conserved, the explanation is also valid to the relative atom number.
\subsection{Dissipative Sequential Landau-Zener dynamics}
\begin{figure*}[htb]
\centering
\includegraphics[scale=0.6]{Fig5.eps}\caption{(Color online) The sequential LZ dynamics under cavity phton losses. The influences of damping rates $\kappa$ on the photon number and the relative atom number under different sweeping rates $\beta$ are shown. Here, the parameters are chosen as $N_a=4$, $|\alpha^2|=4$, $g=1$, and $E_c=100$.}
\label{Fig-dissipation}
\end{figure*}
In cavity QED experiments, the system would suffer decoherence induced by the escape of photons out of the cavity or the decay of the atoms without emitting photons~\cite{Reiserer2015, Hartmann2008}.
These decoherence effects may affect the sequential LZ transitions.
In this subsection, we mainly discuss the effect of dissipation (cavity photon loss) on the sequential LZ dynamics.
During the sweeping, the cavity photons may escape from the cavity and the total excitation number would no longer be conserved.
Therefore, it is necessary to describe the system state by a reduced density matrix $\rho=|\tilde{\Psi}\rangle\langle\tilde{\Psi}|$, where the bases of $|\tilde{\Psi}\rangle$ include different total excitation numbers and differ from the ones in Eq.~\eqref{Psi_ini}. The sequential dynamics under cavity photon loss can be characterized by the Lindblad master equation~\cite{Keeling2008, Hartmann2008, Reiserer2015}
\begin{equation}\label{MEq}
\frac{\partial \rho}{\partial t}=-i\left[ H(t), \rho \right]+L_{\kappa}[\rho],
\end{equation}
with
\begin{equation}\label{Kappa}
L_{\kappa}[\rho]=\frac{\kappa}{2}\left( 2\hat a \rho \hat a^{\dagger} - \hat a^{\dagger} \hat a \rho -\rho \hat a^{\dagger} \hat a \right),
\end{equation}
where $\kappa$ is the cavity photon loss rate.
In our calculation, the reduced density matrix of the system is broken up into different density matrices $\rho_{N_e}$
in subspaces.
Then we solve the master equation~\eqref{MEq} independently and calculate the observable expectation by $\langle \hat O_{\!N_e}\rangle=Tr\left(\hat O \rho_{N_e}\right)$ and sum up together according to $\langle \hat O\rangle=w_{N_e}^2 \langle \hat O_{\!N_e}\rangle$. The time-evolution of photon number and half relative atom number under dissipation for $|\alpha|^2=4$ are shown in Fig.~\ref{Fig-dissipation}.
For $\kappa=0$, it returns to the ideal case, where the total excitation number is conserved.
For $\kappa>0$, the influences of the dissipation become obvious when the sweeping rates getting smaller.
Given $\beta=100$ and $\beta=10$, the influences of dissipation are very small and the sequential LZ dynamics
are almost unchanged, see Fig.~\ref{Fig-dissipation}~(a)-(d).
When the sweeping becomes slower, the total evolution time becomes longer and more photon losses accumulate.
For $\beta=2$, which approaches to the adiabatic limit for $|\alpha|^2=4$, although the cavity photon loss decreases the photon number, it only changes the step heights of the fractional ladders slightly, see Fig.~\ref{Fig-dissipation}~(e).
The cavity photon loss affects the ladder of the half relative atom number gradually, but the step heights shrink only a little during the whole processes, see Fig.~\ref{Fig-dissipation}~(f).
It is shown that, the cavity photon loss may also affect the sequential population ladders.
Obviously, the sequential population ladders will disappear if the lost photons is large compared with $|\alpha|^2$.
However, if the lost photons is small compared with $|\alpha|^2$, which may be realized in the strong coupling regime~\cite{Reiserer2015, Hartmann2008}, the fractional steps of half relative atom number can still be observed.
\subsection{Collapse and revivals of cavity field coherence}
\begin{figure*}[htb]
\centering
\includegraphics[scale=0.7]{Fig6.eps}\caption{(Color online) Collapse and Revivals of cavity field coherence. The evolution of cavity field coherence $\langle a(t) \rangle$ during LZ dynamics with different sweeping rate $\beta$ for (a) $|\alpha|^2=4$ and (b) $|\alpha|^2=25$. The final cavity field coherence $\langle a(\infty)\rangle$ versus the inverse of sweeping rate $1/\beta$ for (c) $|\alpha|^2=4$ and (d) $|\alpha|^2=25$. Here, $N_a=4$, $g=1$ and $E_c=100$.}
\label{Fig-Collapse-revivals}
\end{figure*}
We further investigate how the coherence of the cavity field varies during the LZ processes with different sweeping rates.
Since the evolution of each photon Fock state will accumulate a phase dependent on the photon number, the phase differences between different Fock state will also play an important role during the LZ dynamics.
In order to study the effect of this phase difference during the evolution, a common quantity to measure is the expectation of the photon filed annihilation operator~\cite{Keeling2008}
\begin{eqnarray}\label{Coh}
\langle a(t)\rangle&=&\langle \Psi(t)|\hat a|\Psi(t)\rangle \nonumber\\
&=&\sum_{N_e,l} \sqrt{N_e-l}w^{*}_{N_e-1}\! w_{N_e} C^{N_e-1*}_{l}\!(t) C^{N_e}_{l}\!(t),
\end{eqnarray}
which well describes the coherence of the cavity field.
The coherence of the photon filed $\langle a(t)\rangle$ versus the evolution time $t$ with different sweeping rates $\beta$ are shown in Fig.~\ref{Fig-Collapse-revivals}~(a) and (b).
The dynamical behaviors of $\langle a(t)\rangle$ are much different with the ones of photon number $\langle n(t) \rangle$.
When $|\alpha|^2=25$, for very fast sweeping $\beta=100$, the coherence of the cavity field $\langle a(t)\rangle$ changes slightly.
For modest sweeping $\beta=1$, $\langle a(t)\rangle$ oscillates in ladder shape.
For slower sweeping $\beta=0.1$, $\langle a(t)\rangle$ drops dramatically and remains zero when passing through all the LZ transitions.
However, for very slow sweeping $\beta=0.025$, $\langle a(t)\rangle$ drops rapidly to zero when the first LZ transition happens and begins to increase and oscillate after the second LZ transition. Finally, the coherence remains a certain value after all the LZ transitions.
This is some kind of collapse and revivals of the cavity field coherence due to the long time accumulated phase difference between different Fock state components of the coherent state.
The changes of the cavity field coherence for $|\alpha|^2=4$ is similar with the one of $|\alpha|^2=25$.
For fast sweeping, the coherence of the photon filed $\langle a(t)\rangle$ changes slightly with small oscillations.
For modest sweeping, $\langle a(t)\rangle$ oscillates in a downstairs ladder-like shape.
For very slow sweeping, the coherence of the cavity field $\langle a(t)\rangle$ first collapses and then revives. It drops quickly at the first LZ transition point and oscillates a lot afterwards until all the LZ transitions finished.
However, for $|\alpha|^2=4$, the coherence of the cavity field seldom decreases to zero and the oscillations are more dramatic compared with the ones of $|\alpha|^2=25$.
The coherence of the cavity field will tend to a steady value after all the LZ transitions and the final coherence $\langle a(\infty)\rangle$ depends on the sweeping rate.
Below the dependence of $\langle a(\infty)\rangle$ on the inverse sweeping rate $1/\beta$ is shown in Fig.~\ref{Fig-Collapse-revivals}~(c) and (d).
Obviously, the final coherence of the cavity field versus the inverse sweeping rate also exhibits the behavior of collapse and revivals.
For $|\alpha|^2=25$, with fast sweeping, the collapse occurs as the difference between different Fock state grows leading to a destructively interference. With slower sweeping, the coherence drops to death and remains zero for a wide range. For much slower sweeping, the phase differences grows back in phase and the coherence gradually begins to revive. However, the revival will oscillate and finally cease for extremely slow sweeping.
While for $|\alpha|^2=4$, the coherence changes more rapidly. The final coherence of the cavity field drops dramatically at first but revives quickly and then begins to oscillate in disorder around a modest range. It may be seen that, the cavity field coherence for small average photon number would not exhibit obvious collapse and revivals, which is different from the one for large average photon number.
\section{Summary and Discussions}\label{Sec5}
In summary, we have explored the novel phenomena of asymmetric sequential LZ dynamics in an ensemble of interacting two-level Bose condensed atoms trapped in an optical cavity.
The features of Bose condensed atoms in a cavity with small photon number are extremely different from the ones with large photon number.
For relatively small photon number, the interplay between the detuning and the atom-atom interaction leads to asymmetric sequential LZ transitions.
There appear asymmetric interaction blockade and single-atom resonance tunneling when the LZ process involves only a single excitation number, which is mainly due to the photon-number-dependent Rabi frequency.
The single-atom resonance may be used to design high-precision devices and sensitive detectors.
Instead of an initial Fock state, considering an initial coherent cavity field with small average photon number, an asymmetric population ladder with unequal fractional height steps is observed.
The asymmetric sequential LZ dynamics of the system are also studied.
We derive an analytical formula for sequential population ladders. The adiabaticity conditions for the sequential LZ transitions are analytically estimated. An intriguing findings of fractional steps are also explained. In addition, we analyze the effects of the cavity dissipation on the sequential LZ dynamics.
On the other hand, the state of the cavity field changes dramatically during the time-evolution process.
The behaviors of collapse and revivals of the cavity field coherence are revealed.
Further, for an initial coherent state with large average cavity-photon number, the sequential LZ dynamics in a cavity approach to the one in a continuous laser field.
Starting from all atoms in the lower energy level, the atoms can absorb photons one by one in the sequential LZ process if the detuning is swept sufficiently slow.
When the detuning is swept to zero, an atomic twin Fock state can be generated.
In addition, the photon state is no longer a coherent state and its coherence changes dramatically.
This may be associated to the generation of photon-deleted coherent states~\cite{Honer2011}, which is totally non-classical.
Similarly, if one sweeps the detuning oppositely, and the atoms are initially in the upper level, there would be the phenomenon that the atoms emit photons one by one during the sequential LZ process.
This may be related to the creation of the photon-added coherent states~\cite{Noriyuki2011}.
Different from most previous schemes which add or subtract only a single photon, our this scheme can add or subtract desired number of photons by dynamically controlling the detuning.
This indicates that it may provide a new tool for preparing non-classical photon states and also may be applied to demonstrate the bosonic commutation relations~\cite{Zavatta2004,Parigi2007,Zavatta2008}.
\section*{Acknowledgements}
This work is supported by the National Basic Research Program of China (NBRPC) under Grant No. 2012CB821305, the National Natural Science Foundation of China (NNSFC) under Grants No. 11374375, No. 11465008, and No. 11574405.
|
\section{Algebraic Toolkit}
\begin{appsec}\label{app:varieties}\textbf{Varieties of algebras.}
Fix a finitary signature~$\Gamma$, i.e. a set of
operation symbols with finite arities. A
\emph{$\Gamma$-algebra} is a
set $A$
equipped with an operation $\gamma^A: A^n\to A$ for each $n$-ary
$\gamma\in\Gamma$, and a \emph{morphism} of $\Gamma$-algebras
is a map preserving these operations. \emph{Quotients} and
\emph{subalgebras} of $\Gamma$-algebras are represented by
surjective resp. injective morphisms. A \emph{variety of algebras}
is
a class of $\Gamma$-algebras closure under quotients, subalgebras,
and products. Equivalently, a variety is class of $\Gamma$-algebras
specified by equations $s=t$ between
$\Gamma$-terms.
\end{appsec}
\begin{appsec}\label{app:orderedvarieties}\textbf{Varieties of ordered algebras.}
An \emph{ordered $\Gamma$-algebra} is a
poset $A$ equipped with a monotone operation $\gamma^A: A^n\to A$ for each
$n$-ary $\gamma\in\Gamma$, and a \emph{morphism} of
ordered $\Gamma$-algebras is a monotone map preserving these operations.
\emph{Quotients} of ordered algebras are
represented by
surjective morphisms, and \emph{subalgebras} by order-reflecting morphisms $m$
(i.e. $mx\leq my$ iff $x\leq y$). A \emph{variety of ordered
algebras}
is
a class of ordered $\Gamma$-algebras closed under quotients, subalgebras,
and products. Equivalently, a variety is class of ordered $\Gamma$-algebras
specified by inequations $s\leq t$ between
$\Gamma$-terms.
\end{appsec}
In the following, let $\mathscr{D}$ always denote a variety of algebras or ordered algebras.
\begin{appsec}\label{app:commvarieties}\textbf{Commutative varieties.}
A variety $\mathscr{D}$ of algebras or ordered algebras is \emph{commutative} if, for
any
$A\in\mathscr{D}$ and any
$n$-operation
symbol $\gamma\in\Gamma$, the corresponding operation
$\gamma^A: \under{A}^n\to \under{A}$ carries a morphism $\gamma^A: A^n\to A$
of $\mathscr{D}$. Equivalently, for any two
algebras $A,B\in \mathscr{D}$, the set $[A,B]$ of morphisms from $A$ to $B$ forms an
algebra of $\mathscr{D}$ under the pointwise $\Gamma$-operations, i.e.
$[A,B]$ carries a
subalgebra of $B^{\under{A}}$, the $\under{A}$-fold power of $B$.
\end{appsec}
\begin{appsec}\label{app:congruences}\textbf{Congruences and stable preorders.}
\begin{enumerate}[(1)]
\item A \emph{congruence} on a $\Gamma$-algebra $A$ is an
equivalence relation $\equiv$ on $A$ such that for all $n$-ary operations
$\gamma\in\Gamma$ and elements $a_1,\ldots,a_n,b_1,\ldots,b_n\in A$,
\[a_i\equiv b_i\,\, (i=1,\ldots,n) \quad\text{implies}\quad
\gamma^A(a_1,\ldots,a_n)\equiv\gamma^A(b_1,\ldots, b_n).\]
The set $A/\mathord{\equiv}$ of equivalence classes carries a
$\Gamma$-algebra structure defined by
\[ \gamma^{A/\mathord{\equiv}}([a_1],\ldots,[a_n]) :=
[\gamma^A(a_1,\ldots,a_n)],\]
and the projection map $\pi: A\twoheadrightarrow A/\mathord{\equiv}$, $a\mapsto [a]$, is a
surjective morphism of $\Gamma$-algebras.
\item Let $(A,\leq)$ be an ordered $\Gamma$-algebra. A \emph{stable preorder}
on $A$ is a preorder $\preceq$ on $A$ such that
(i) $a\leq b$ implies $a\preceq b$, and (ii) for all $n$-ary operations
$\gamma\in\Gamma$ and elements $a_1,\ldots,a_n,b_1,\ldots,b_n\in A$,
\[a_i\preceq b_i\,\, (i=1,\ldots,n) \quad\text{implies}\quad
\gamma^A(a_1,\ldots,a_n)\preceq\gamma^A(b_1,\ldots, b_n).\]
For any stable preorder, the equivalence relation $\equiv\,:=\,\preceq \cap
\succeq$ forms a congruence on $A$ in the sense of (1), and
$A/\mathord{\equiv}$ becomes an ordered $\Gamma$-algebra by setting
$[a]\leq[a']$ iff $a\leq a'$. We write $A/\mathord{\preceq}$ for this ordered
algebra. The projection map $\pi: A\twoheadrightarrow A/\preceq$, $a\mapsto [a]$,
is a
surjective morphism of ordered $\Gamma$-algebras.
\end{enumerate}
\end{appsec}
\begin{appsec}\label{app:jointlyinj}\textbf{Separating families.}
A family $\{f_i: A\to B_i\}_{i\in I}$ of morphisms in $\mathscr{D}$ is
\emph{separating} if the
morphism
$f: A\to \prod_i B_i$ with $f(a) =
(f_i(a))_{i\in I}$ is injective (resp. order-reflecting if $\mathscr{D}$ is a variety
of ordered algebras). Equivalently, for any
two elements $a,a'\in A$ with $a\neq a'$ (resp. $a\not\leq a'$), there
exists an $i\in I$ with $f_i(a)\neq f_i(a')$ (resp. $f_i(a)\not\leq
f_i(a')$).
Suppose that, for each $i\in I$, another separating family $\{\,g_{i,j}:
B_i\to C_{i,j}\,\}_{j\in J_i}$ is given. Then the combined family
$\{\,g_{i,j}\c f_i: A\to C_{i,j}\,\}_{i\in I, j\in J_i}$ is also separating.
\end{appsec}
\begin{appsec}\label{app:factsystems}\textbf{Factorization systems.}
Any variety $\mathscr{D}$ of algebras or ordered algebras has the factorization system
of surjective
and injective (resp. order-reflecting) morphisms. This means that (i) any
morphism $h: A\to B$ has a factorization $h=m\c e$ with $e$ surjective and $m$
injective (resp. order-reflecting), and (ii) the \emph{diagonal fill-in}
property holds: given a commutative square as displayed below with $e$ surjective and
$m$ injective (order-reflecting), there is a unique morphism $d$ making both
triangles commutative:
\[
\xymatrix{
D \ar@{->>}[r]^e \ar[d]_g & C \ar[d]^{h} \ar@{-->}[dl]_{d}\\
A \ar[r]_{m} & B_i
}
\]
The diagonal fill-in property generalizes to families of morphisms: suppose
that $e$, $g$,
$h_i$ and
$m_i$ ($i\in I$) are morphisms with $h_i\c e=m_i\c g$ for all $i$.
If $e$ is surjective and the family $\{\,m_i\,\}_{i\in I}$ is separating, then there exists a unique morphism
$d$ with $d\c e = g$ and $m_i\c d=h_i$ for all $i$:
\[
\xymatrix{
D \ar@{->>}[r]^e \ar[d]_g & C \ar[d]^{h_i} \ar@{-->}[dl]_{d}\\
A \ar[r]_{m_i} & B_i
}
\]
\end{appsec}
\begin{appsec}\label{app:tensorproducts}\textbf{Tensor products.}
Let $\mathscr{D}$ be \emph{commutative} variety of algebras or ordered algebras
\begin{enumerate}[(1)]
\item Let $A,B,C\in \mathscr{D}$. By a
\emph{bimorphism} from
$A$, $B$ to
$C$ is meant a function $f: \under{A}\times\under{B}\rightarrow\under{C}$ such that
the maps
$f(a,\mathord{-}): \under{B}\rightarrow \under{C}$ and $f(\mathord{-},b): \under{A}\rightarrow
\under{C}$ carry morphisms of $\mathscr{D}$ for every $a\in\under{A}$ and $b\in
\under{B}$. A \emph{tensor product} of $A$ and $B$ is a universal bimorphism
$t_{A,B}: \under{A}\times \under{B} \rightarrow \under{A\t B}$, i.e.
for any bimorphism $f: \under{A}\times \under{B} \rightarrow \under{C}$
there is a
unique $f': A\t B \rightarrow C$ in $\mathscr{D}$ with $f'\c t_{A,B} = f$. We write $a\t b$ for the element $t_{A,B}(a,b)\in\under{A\t B}$.
\item Tensor products
exist in any variety $\mathscr{D}$. Let us indicate how to construct them
in the case where $\mathscr{D}$ is a variety of ordered algebras. Given $A,B\in \mathscr{D}$,
form the free algebra
$\Psi(\under{A}\times\under{B})$ in $\mathscr{D}$ generated by the set
$\under{A}\times\under{B}$. (For simplicity, we assume that
$\under{A}\times\under{B}$ is a subset of $\Psi(\under{A}\times\under{B})$.)
Form the smallest stable preorder $\preceq$ on
$\Psi(\under{A}\times\under{B})$ containing all inequations of the form
\begin{align*}
(a,\gamma(b_1,\ldots,b_n)) &\preceq \gamma((a,b_1),\ldots,(a,b_n))\\
\gamma((a,b_1),\ldots,(a,b_n)) &\preceq (a,\gamma(b_1,\ldots,b_n))\\
(\gamma(a_1,\ldots,a_n),b) &\preceq (\gamma((a_1,b),\ldots,(a_n,b))\\
(\gamma((a_1,b),\ldots,(a_n,b))&\preceq (\gamma(a_1,\ldots,a_n),b)
\end{align*}
where $\gamma\in \Gamma$ is an $n$-ary operation symbol, $a,a_1,\ldots,a_n\in
A$ and $b,b_1,\ldots,b_n\in B$. Then the tensor product of $A$ and $B$ is
given by \[A\t B :=
\Psi(\under{A}\times\under{B})/\mathord{\preceq}\] and the universal bimorphism \[t_{A,B} :=
(\,\under{A}\times\under{B}\rightarrowtail \Psi(\under{A}\times\under{B}) \xrightarrow{\pi}
A\t B =
\Psi(\under{A}\times\under{B})/\mathord{\preceq}\,),\]
the composite of the inclusion map and the projection. In particular, $A\t B$
is generated by the elements $a\t b$ with $a\in\under{A}$ and $b\in\under{B}$.
The construction of
$A\t B$ for unordered algebras is analogous: just replace inequations and
stable
preorders by equations and
congruences.
\item For any $A,B,C\in \mathscr{D}$ there is a natural bijective correspondence between (i)
morphisms from $A\t B$ to $C$, (ii) bimorphisms from $A,B$ to $C$, and (iii)
morphisms
from $A$ to $[B,C]$. Indeed, the correspondence of (i) and (ii) follows from
the universal property of the tensor product. Further, any bimorphism $f:
\under{A}\times \under{B}\to \under{C}$ defines a morphism
\[ \lambda f: A\to [B,C],\quad (\lambda f)(a)(b) = f(a,b), \]
and the map $f\mapsto \lambda f$ gives the bijective correpondence between
(ii)
and
(iii).
\item Up to isomorphism, $\t$ is associative, commutative and has unit $\mathbf{1}_\mathscr{D}$.
More precisely, for any three objects
$A,B,C\in \mathscr{D}$ there are natural isomorphisms $\alpha_{A,B,C}$, $\sigma_{A,B}$,
$\rho_A$ and $\lambda_A$ making the
following squares commute:
\[
\xymatrix{
\under{(A\t B)\t C} \ar[r]^{\alpha_{A,B,C}} & \under{A\t (B\t C)}\\
(\under{A}\times \under{B})\times \under{C} \ar[u]^{t_{A\t B, C}
\c(t_{A,B}\times C)} \ar[r]_{\alpha'_{A,B,C}} & \under{A}
\times (\under{B}\times \under{C}) \ar[u]_{t_{A,B\t C} \c(A\times t_{B,C})}
}
\]
\[
\xymatrix{
\under{A\t B} \ar[r]^{\sigma_{A,B}} & \under{B\t A}\\
\under{A}\times \under{B} \ar[u]^{t_{A,B}} \ar[r]_{\sigma'_{A,B}} &
\under{B}\times
\under{A} \ar[u]_{t_{B,A}}
}\quad
\xymatrix{
\under{A}\t \under{\mathbf{1}_\mathscr{D}} \ar[r]^{\rho_A} & \under{A}\\
\under{A}\times \under{\mathbf{1}_\mathscr{D}} \ar[u]^{t_{A,\mathbf{1}_\mathscr{D}}} \ar[ur]_{\pi_A} &
}
\quad
\xymatrix{
\under{\mathbf{1}_\mathscr{D}\t A} \ar[r]^{\lambda_A} & \under{A}\\
\under{\mathbf{1}_\mathscr{D}}\times \under{A} \ar[ur]_{\pi_A'} \ar[u]^{t_{\mathbf{1}_\mathscr{D},A}} &
}
\]
where $\alpha'$ and $\sigma'$ are the canonical bijections, and $\pi_A$ and
$\pi_A'$ are the projection maps. We shall often omit indices and write
$t$ for $t_{A,B}$, $\alpha$ for $\alpha_{A,B,C}$, etc.
\item A \emph{$\mathscr{D}$-monoid} $(M,\iota,\mu)$ is
an object $M\in \mathscr{D}$ equipped with two morphisms $\iota: \mathbf{1}_\mathscr{D}\to
M$ and $\mu: M\t M\to M$ such that the following diagrams commute:
\[
\vcenter{
\[email protected]{
& (M\t M) \t M \ar[rr]^{\alpha} \ar[ld]_-{\mu \t M} & & M \t (M \t M)
\ar[rd]^{M \t \mu} \\
M \t M \ar[rrd]_{\mu} & & & & M \t M \ar[lld]^{\mu} \\
& & M
}
}\quad
\vcenter{
\xymatrix{
M \t \mathbf{1}_\mathscr{D} \ar[rd]^{\lambda} \ar[d]_{M \t \iota} \\
M \t M \ar[r]^{\mu} & M\\
\mathbf{1}_\mathscr{D} \t M \ar[ru]_{\rho} \ar[u]^{\iota \t M}
}
}
\]
A \emph{morphism} between $\mathscr{D}$-monoids $(M,\iota_M,\mu_M)$ and $(N,\iota_N,\mu_N)$ is
a morphism
$h: M\to N$ in $\mathscr{D}$ such that the following square commutes:
\[
\xymatrix@C+1em{
M\t M \ar[r]^>>>>>>>{\mu_M} \ar[d]_{h\t h} & M \ar[d]^h & \ar[dl]^{\iota_N} \ar[l]_{\iota_M}\mathbf{1}_\mathscr{D} \\
N\t N \ar[r]_>>>>>>>{\mu_N} & N &
}
\]
Due to $\t$ representing bimorphisms, the notion of $\mathscr{D}$-monoids and their morphisms given here is equivalent to the set-theoretic one of Definition \ref{def:dmonoid}.
\end{enumerate}
\end{appsec}
\section{Categorical Toolkit}
We assume familiarity with basic concepts from category theory, like
categories,
functors,
natural transformations, and (co-)limits (see e.g. \cite{macl}). However, we
recall here some definitions and facts concerning adjunctions.
\begin{appsec}\label{app:adjunctions}\textbf{Adjunctions.}
Let $U: \mathscr{A}\to \mathscr{X}$ be a functor between categories $\mathscr{A}$ and $\mathscr{X}$. Suppose that there
exists, for each
$X\in\mathscr{X}$, an object
$FX\in\mathscr{A}$ and a morphism $\eta_X: X\to U(FX)$ in $\mathscr{X}$ with the following universal
property: for any morphism $f: A\to UB$ in $\mathscr{X}$ with $B\in\mathscr{B}$, there is a
unique $\overline f: FX\to B$ in $\mathscr{B}$ (called the \emph{adjoint transpose} of $f$)
with $U(\overline f)\c \eta_X = f$. In this case the
object map $X\mapsto FX$ extends uniquely to a functor $F: \mathscr{X}\to\mathscr{A}$ such that
$\eta: \mathsf{Id}_\mathscr{X} \to UF$ becomes a natural transformation, and we say that the
functors $U$ and $F$ form an \emph{adjunction} (commonly denoted by
$F\dashv U:
\mathscr{A}\to\mathscr{X}$). The functor $U$ is the \emph{right adjoint}, $F$ the
\emph{left adjoint}, and the natural transformation $\eta$ the \emph{unit} of
the adjunction. $\eta$
induces
another natural transformation $\epsilon: FU \to \mathsf{Id}_\mathscr{A}$ with components
$\epsilon_A := \overline{\mathit{id}_{UA}}$, called the \emph{counit} of the adjunction. The
universal property gives rise to an isomorphism $\mathscr{A}(FX,A)\cong \mathscr{X}(X,UA)$
natural in $X\in \mathscr{X}$ and $A\in A$. An important fact about adjunctions is that
right adjoints preserve limits, and dually left adjoints preserve colimits
(and
thus, in particular, epimorphisms).
A typical source of adjunctions are free constructions in algebra. For any
variety
$\mathscr{D}$ of algebras or ordered algebras, the forgetful functor $\under{\mathord{-}}:
\mathscr{D}\to\mathbf{Set}$ (mapping an algebra to its underlying set) has the left adjoint
$\Psi: \mathbf{Set}\to\mathscr{D}$ that maps a set $X$ to the free algebra $\Psi X$ in $\mathscr{D}$
generated by $X$. The unit $\eta_X: X\to \under{\Psi X}$ is the inclusion of
generators. The freeness of $\Psi X$ amounts exactly to the universal property
of
$\eta_X$.
\end{appsec}
\begin{appsec}\label{app:compadjunctions}\textbf{Composition of adjunctions.} The \emph{composite} of two
adjunctions \[(F\dashv U: \mathscr{A}\to \mathscr{B},\eta,\epsilon)\quad\text{and}\quad(G\dashv
V: \mathscr{B}\to \mathscr{X},
\eta', \epsilon')\] is the adjunction $FG\dashv VU: \mathscr{A}\to \mathscr{X}$ with unit
$X\xrightarrow{\eta'_X} VGX \xrightarrow{V\eta_{GX}} VUFGX$ ($X\in \mathscr{X}$) and counit $FGVUA
\xrightarrow{F\epsilon'_{UA}} FUA \xrightarrow{\epsilon_A} A$ ($A\in \mathscr{A}$).
\end{appsec}
\begin{appsec}\label{app:yoneda}\textbf{Yoneda lemma (weak form).} Let $\mathscr{A}$ be a category and $A\in \mathscr{A}$. The \emph{hom-functor} $\mathscr{A}(A,\mathord{-}): \mathscr{A}\to\mathbf{Set}$ maps an object $B\in \mathscr{A}$ to the set $\mathscr{A}(A,B)$ of morphisms from $A$ to $B$, and a morphism $f: B\to B'$ to the function
\[ \mathscr{A}(A,f): \mathscr{A}(A,B)\to \mathscr{A}(A,B'),\quad g\mapsto f\c g. \]
The hom-functor determines objects of $\mathscr{A}$ up to isomorphism: any natural isomorphism $\theta:
\mathscr{A}(A,\mathord{-})\cong \mathscr{A}(A',\mathord{-})$ with $A,A'\in \mathscr{A}$ yields an isomorphism
$\theta_A(\mathit{id}_A): A'\cong A$.
\end{appsec}
\section{Monoidal Adjunctions}
Here we present some well-known facts about monoidal functors and adjunctions.
Since these facts appear scattered throughout the literature or are folklore
in category theory, we sketch the proofs for some statements for the
convenience of the reader. In the following, let $\mathscr{A}$, $\mathscr{B}$, $\mathscr{C}$, $\mathscr{D}$ be
commutative varieties of (ordered) algebras; we remark that all concepts
treated in this section can be introduced in a more general form for monoidal
categories, see e.g. \cite{macl}.
Recall from Definition \ref{def:monoidalfunc} the notion of a \emph{monoidal functor} and a \emph{monoidal natural transformation}. Recall also from \ref{app:tensorproducts} that $\mathscr{D}$-monoids $(M,1,\bullet)$ (see Definition \ref{def:dmonoid}) can be represented as triples $(M,\iota,\mu)$.
\begin{lemma}\label{lem:monlifting2}
Any monoidal functor $(G,\theta): \mathscr{A}\to \mathscr{B}$ lifts to a functor $\overline G:
\Mon{\mathscr{A}}\to
\Mon{\mathscr{B}}$ such that the following diagram commutes, where $U_\mathscr{A}$ and $U_\mathscr{B}$
are the forgetful functors:
\[
\xymatrix{
\Mon{\mathscr{A}} \ar[d]_{U_\mathscr{A}} \ar[r]^{\overline G} & \Mon{\mathscr{B}} \ar[d]^{U_\mathscr{B}} \\
\mathscr{A} \ar[r]_G & \mathscr{B}
}
\]
Explicitly, $\overline G$ maps an $\mathscr{A}$-monoid $(M,\iota,\mu)$ to the $\mathscr{B}$-monoid
\[
(GM,\ \mathbf{1}_\mathscr{B}\xrightarrow{\theta} G\mathbf{1}_\mathscr{A} \xrightarrow{G\iota} GM,\ GM\t GM \xrightarrow{\theta}
G(M\t M)\xrightarrow{G\mu}
GM),
\]
and an $\mathscr{A}$-monoid morphism $f$ to $Gf$. Moreover, any monoidal natural
transformation $\phi: (G,\theta)\to (H,\sigma)$ yields a natural
transformation $\overline\phi: \overline G\to \overline H$ with components
\[ \overline\phi_{(M,\iota,\mu)} = \phi_M: \overline G(M,\iota,\mu) \to \overline H(M,\iota,\mu). \]
\end{lemma}
\begin{proof}
\begin{enumerate}
\item It is straightforward to show that $(GM, G\iota \c \theta, G\mu \c
\theta_{M, M})$ is a $\mathscr{B}$-monoid. For example, associativity is
established by the commutative diagram below. Part $(1)$ commutes since
$G$ is monoidal; $(2)$ commutes since $(M,
\iota, \mu)$ is a monoid; $(3)$ and $(4)$ commute because $\theta_{A,
B}\colon
GA \t GB \to G(A \t B)$ is natural in $A$ and $B$.
\[
\xymatrix{
& (GM\t GM)\t GM \ar[rr]^{\alpha_{GM, GM, GM}} \ar@{}[rrdd]|{(1)}
\ar[d]_{\theta_{M, M}\t GM} & & GM \t (GM\t GM) \ar[d]^{GM\t \theta_{M,
M}}\\
& G(M\t M) \t GM \ar[d]_{\theta_{M\t M, M}} \ar@/_2pc/[ld]_-{G\mu \t GM} &
& GM \t G(M\t M) \ar[d]^{\theta_{M, M\t M}} \ar@/^2pc/[rd]^-{GM \t G\mu}
\\1
GM \t GM \ar@{}[r]|-{(3)} \ar@/_2pc/[rd]_{\theta_{M, M}} & G((M\t M)\t
M) \ar[d]_{G(\mu \t M)} \ar[rr]_{G\alpha_{M, M, M}}& \ar@{}[dd]|{(2)} &
G(M\t (M\t M)) \ar[d]^{G(M \t \mu)} \ar@{}[r]|-{(4)} & GM \t GM
\ar@/^2pc/[ld]^{\theta_{M, M}}\\
& G(M \t M) \ar[rd]_{G\mu} & & G(M\t M) \ar[ld]^{G\mu} \\
& & GM
}
\]
The unit laws
follow in a similar way.
Assume that $f\colon (M,\iota_M, \mu_M) \to (N, \iota_N, \mu_N)$ is an $\mathscr{A}$-monoid morphism.
Then the following diagram commutes:
\[
\xymatrix{
GM \t GM \ar[r]^{\theta_{M, M}} \ar[d]_{Gf \t Gf} & G(M\t M) \ar[r]^-{G\mu_M}
\ar[d]_{G(f\t f)} & GM \ar[d]^{Gf} \\
GN \t GN \ar[r]_{\theta_{N, N}} & G(N\t N) \ar[r]_-{G\mu_N} & GN
}
\]
where the right square commute because $f$ is a $\mathscr{D}$-monoid morphism and
the left
square commutes as $\theta_{M, M}$ is natural. Together with the
corresponding diagram for the preservation of the unit, this shows that
$Gf: \overline G(M,\iota_M,\mu_M)\to \overline G(N,\iota_M,\mu_N)$ is a $\mathscr{B}$-monoid
morphism.
\item To show that every monoidal natural transformation $\varphi$
lifts to a natural transformation from $\overline G$ to $\overline H$, it suffices to
show that $\varphi_M$ is a $\mathscr{B}$-monoid morphism for every $\mathscr{A}$-monoid
$(M, \iota,
\mu)$. The preservation of the multiplication follows from the following diagram:
\[
\xymatrix@C+2em{
GM \t GM \ar[r]^{\varphi_M \t \varphi_M} \ar[d]_{\theta_{M, M}}
& HM \t HM \ar[d]^{\sigma_{M, M}} \\
G(M \t M) \ar[r]^{\varphi_{M \t M}} \ar[d]_{G\mu} & H(M \t M) \ar[d]^{H\mu}
\\
GM \ar[r]_{\varphi_M} & HM
}
\]
where the upper square uses that $\phi$ is a monoidal natural
transformation and the lower square is the naturality of~$\varphi$. Similarly for the preservation of the unit.
\end{enumerate}
\end{proof}
\begin{lemma}\label{lem:moncompose}
For any two monoidal functors $(G,\theta)\colon \mathscr{A}\to \mathscr{B}$ and $(H,\sigma)\colon
\mathscr{B}\to\mathscr{C}$ the composite $HG$ becomes a monoidal functor via
\begin{align*}
(H\theta\c\sigma)_{A, B} & =
H(GA)\t H(GB) \xrightarrow{\sigma_{A, B}} H(GA\t GB) \xrightarrow{H\theta_{GA, GB}} HG(A\t B)
\\
G\theta \c \sigma & = \mathbf{1}\mathscr{C} \xrightarrow{\sigma} H\mathbf{1}_\mathscr{B} \xrightarrow{G\theta} HG\mathbf{1}_\mathscr{A}.
\end{align*}
\end{lemma}
\begin{proof}
The naturality of $(H\theta \c \sigma)_{A, B}$ follows from the naturality of
$\sigma_{A, B}$ and~$H\theta_{GA, GB}$. It remains to verify the diagrams in Definition \ref{def:monoidalfunc}. For example, for the left diagram we get
\[
\xymatrix@+4em@ru{
& (HGA \t HGB) \t HGC \ar[r]^{\alpha_{HGA, HGB, HGC}}
\ar[d]_{\sigma_{GA, GB} \t HGC} \ar@{}[rdd]|{(1)}
& HGA \t (HGB \t HGC) \ar[d]^{HGA \t \sigma_{HGB, HGC}} \\
& H(GA \t GB) \t HGC \ar[d]^{\sigma_{GA\t GB, GC}}
\ar[ld]_{H\theta_{A, B} \t HGC}
& HGA \t H(GB \t GC) \ar[d]^{\sigma_{GA, GB \t GC}}
\ar[rd]^{HA \t H\theta_{B, C}} \\
HG(A \t B) \t HGC \ar[rd]_{\theta_{GA\t GB, GC}}
\ar@{}[r]|{(3)}
& H((GA \t GB) \t GC) \ar[r]_{H\alpha_{GA, GB, GC}} \ar[d]_{H(\theta_{A,
B} \t GC)}\ar@{}[rdd]|{(2)}
& H(GA \t (GB \t GC)) \ar[d]^{H(GA \t \theta_{B, C}}
\ar@{}[r]|{(4)}
& HGA \t HG(B \t C) \ar[ld]^{\theta_{GA, G(B\t C)}}
\\
& H(G(A \t B) \t GC) \ar[d]_{H\theta_{A \t B, C}}
& H(GA \t G(B \t C)) \ar[d]_{H\theta_{A, B \t C}}
&
\\
& HG((A \t B) \t C) \ar[r]_{HG\alpha_{A, B, C}} & HG(A \t (B \t C))
}
\]
In $(1)$ we use that $H$ is monoidal, in $(2)$ that $G$ is monoidal, in (3)
the naturality of $\theta$, and in (4) the naturality of $\sigma$. Similarly
for the other two diagrams in Definition \ref{def:monoidalfunc}.
\end{proof}
\begin{expl}\label{ex:upsimonoidal}
\begin{enumerate}[(1)]
\item
For any commutative variety $\mathscr{D}$ of algebras or ordered algebras, the forgetful functor $\under{\mathord{-}}: \mathscr{D}\to \mathbf{Set}$ is a monoidal
functor w.r.t. the maps
\[ i: 1\rightarrowtail \under{\mathbf{1}_\mathscr{D}} = \under{\Psi 1}\quad\text{and}\quad t_{A,B}: \under{A}\times\under{B}\to \under{A\t B}\]
where $i$ is the inclusion of the generator.
Indeed, by Definition \ref{def:monoidalfunc} we need to verify that the following diagrams commute for all $A,B\in \mathscr{D}$:
\begin{center}
\begin{tabular}[x]{cc}
\parbox{0.6\textwidth}{
$
\xymatrix@[email protected]{
(\under{A}\times \under{B})\times \under{C} \ar[r]^{\alpha'} \ar[d]_{t \times \under{C}} & \under{A} \times
(\under{B}\times
\under{C}) \ar[d]^{\under{A}\times t}\\
\under{A\t B} \times \under{C} \ar[d]_{t} & \under{A} \times \under{B\t C}
\ar[d]^t \\
\under{(A\t B)\t C} \ar[r]_{\alpha}& \under{A\t (B\t C)}
}
$
}
&
\parbox{0.4\textwidth}{
$
\[email protected]{
\under{A}\times 1 \ar[r]^<<<<<{\under{A}\times i} \ar[d]_{\rho'} & \under{A}\times \under{\mathbf{1}_\mathscr{D}}
\ar[d]^t\\
\under{A} & \under{A\t \mathbf{1}_\mathscr{D}} \ar[l]_<<<<<<<{\rho}
}
$\\
$\[email protected]{
1\times\under{A} \ar[r]^<<<<<{i\times \under{A}} \ar[d]_{\lambda'} & \under{\mathbf{1}_\mathscr{D}}\times \under{A}
\ar[d]^t\\
\under{A} & \under{\mathbf{1}_\mathscr{D}\t A} \ar[l]_<<<<<<<{\lambda}
}
$
}
\end{tabular}
\end{center}
But this follows immediately from the definitions of $t$, $\alpha$, $\rho$, $\lambda$, see \ref{app:tensorproducts}. For example, both legs of the left hand diagram map an element $((a,b),c)\in (\under{A}\times\under{B})\times \under{C}$ to $a\t (b\t c)$. Similarly for the other two diagrams.
\item The left adjoint $\Psi: \mathbf{Set}\to\mathscr{D}$ to $\under{\mathord{-}}: \mathscr{D}\to \mathbf{Set}$ is also monoidal. Indeed, observe that for any two sets $X$ and
$Y$ we have the following bijections (natural in $A\in\mathscr{D}$), cf. \ref{app:tensorproducts}(3):
\begin{align*}
\mathscr{D}(\Psi(X\times Y), A) &\cong \mathbf{Set}(X\times Y, \under{A})\\
&\cong \mathbf{Set}(X,\under{A}^Y) \\
&\cong \mathbf{Set}(X, |[\Psi Y, A]|)\\
&\cong \mathscr{D}(\Psi X, [\Psi Y, A]) \\
&\cong \mathscr{D}(\Psi X\t \Psi Y, A).
\end{align*}
The Yoneda lemma, see \ref{app:yoneda}, gives a natural isomorphism $\Psi(X\times Y)\cong \Psi X\t \Psi Y$, mapping a pair $(x,y)\in X\times Y$ to
$x\t y\in \under{\Psi X\t \Psi Y}$. Its inverse $\theta_{X,Y}: \Psi X\t \Psi
Y \cong \Psi(X\times Y)$ together with the morphism $\theta = \mathit{id}:
\mathbf{1}_\mathscr{D}\to \Psi 1$, makes $\Psi$ a monoidal functor, i.e. the following diagrams commute for all sets $X,Y,Z$:
\begin{center}
\begin{tabular}[x]{cc}
\parbox{0.6\textwidth}{
$
\xymatrix@[email protected]{
(\Psi{X}\t \Psi{Y})\t \Psi{Z} \ar[r]^{\alpha} \ar[d]_{\theta \times \Psi{Z}} & \Psi{X} \t
(\Psi{Y}\t
\Psi{Z}) \ar[d]^{\Psi{X}\t \theta}\\
\Psi(X\times Y) \t \Psi{Z} \ar[d]_{\theta} & \Psi{X} \t \Psi(Y\times Z)
\ar[d]^\theta \\
\Psi((X\times Y)\times Z) \ar[r]_{\Psi\alpha'}& \Psi(X\times (Y\times Z))
}
$
}
&
\parbox{0.4\textwidth}{
$
\[email protected]{
\Psi{X}\t \mathbf{1}_\mathscr{D} \ar[r]^<<<<<{\Psi{X}\t \mathit{id}} \ar[d]_{\rho} & \Psi{X}\t \Psi 1
\ar[d]^\theta\\
\Psi{X} & \Psi(X\times 1) \ar[l]_<<<<<<<{\Psi\rho'}
}
$\\
$
\[email protected]{
\mathbf{1}_\mathscr{D}\t \Psi X \ar[r]^<<<<<{\mathit{id}\t \Psi X} \ar[d]_{\lambda} & \Psi 1\t \Psi X
\ar[d]^\theta\\
\Psi{X} & \Psi(1\times X) \ar[l]_<<<<<<<{\Psi\lambda'}
}
$
}
\end{tabular}
\end{center}
This follows directly from the definitions of $\theta$, $\alpha$, $\lambda$,
$\rho$. For example, both legs of the left-hand diagram map an element $(x\t
y)\t z)\in \under{(\Psi X\t \Psi Y)\t \Psi Z}$ (with $x\in X$, $y\in Y$, $z\in
Z$) to $(x,(y,z)) \in X\times (Y\times Z)\subseteq \Psi(X\times(Y\times Z))$. Since
the elements $(x\t y)\t z$ generate the algebra ${(\Psi X\t \Psi Y)\t \Psi
Z}$, see \ref{app:tensorproducts}, this shows that the diagram commutes.
Similarly for the other two diagrams.
\item The adjunction $\Psi\dashv \under{\mathord{-}}:\mathscr{D}\to\mathbf{Set}$ is monoidal (see
Definition \ref{def:monoidaladj}). To see this, we need to show that the unit
$\eta: \mathsf{Id}_\mathbf{Set}\to \under{\Psi}$ and the counit $\epsilon: \Psi\c
\under{\mathord{-}} \to \mathsf{Id}_\mathscr{D} $ are monoidal natural transformations. We only
prove
that $\epsilon$ is monoidal, since the proof for $\eta$ is similar. Note first
that by $\Psi\c \under{\mathord{-}}: \mathscr{D}\to\mathscr{D}$ is meant the composite monoidal
functor
in the sense of Lemma \ref{lem:moncomp}. Thus the associated morphisms are
\[ \Psi\under{A}\t \Psi\under{B} \xrightarrow{\theta_{\under{A},\under{B}}}
\Psi(\under{A}\times \under{B}) \xrightarrow{\Psi t_{A,B}} \Psi\under{A\t B} \]
and
\[ \mathbf{1}_\mathscr{D} \xrightarrow{\mathit{id}=\theta_1} \Psi 1 \xrightarrow{\Psi i} \Psi\under{\mathbf{1}_\mathscr{D}}. \]
To show that $\epsilon$ is a monoidal natural transformation, we need to
verify that the following
two
diagrams commute, cf. Definition \ref{def:monoidalfunc}.
\[
\xymatrix{
\Psi{\under{A}}\t\Psi\under{B} \ar[r]^>>>>{\epsilon_A\t \epsilon_B}
\ar[d]_{\Psi{\under{A},\under{B}}} & A\t B \ar@{=}[dd]\\
\Psi(\under{A}\times\under{B}) \ar[d]_{\theta t_{A,B}} & \\
\Psi\under{A\t B} \ar[r]_{\epsilon_{A\t B}} & A\t B
}\qquad\qquad
\xymatrix{
& \mathbf{1}_\mathscr{D} \ar[dl]_{\mathit{id}} \ar@{=}[ddr] &\\
\Psi 1 \ar[d]_{\Psi i} & &\\
\Psi\under{\mathbf{1}_\mathscr{D}} \ar[rr]_{\epsilon_{\mathbf{1}_\mathscr{D}}} && \mathbf{1}_\mathscr{D}
}
\]
But this follows immediately from the definitions of $\epsilon$, $\theta$ and
$i$. Both legs of the left diagram map an element $a\t b\in
\Psi\under{A}\t\Psi\under{B}$ (with $a\in \under{A}$ and $b\in\under{B}$) to
$a\t b$. Since $\Psi\under{A}\t\Psi\under{B}$ is generated by these elements,
this shows that the left diagram commutes. In the right diagram, both legs map
the generator of $\mathbf{1}_\mathscr{D}$ to itself, and thus both legs are the identity
morphism.
\end{enumerate}
\end{expl}
\section{Proofs}
\begin{notation}
Throughout this section, we assume the categorical setting introduced at the
beginning of Section 3, i.e. we have a concrete monoidal adjunction $F\dashv
U: \Mod{\S}\to\mathscr{D}$ with unit $\eta:\mathsf{Id} \to UF$. We denote the forgetful
functors by $\under{\mathord{-}}_\S: \Mod{\S}\to\mathbf{Set}$ and $\under{\mathord{-}}_\mathscr{D}: \mathscr{D}
\to\mathbf{Set}$, but usually drop the indices. For a morphism $f$ in $\mathscr{D}$ or
$\Mod{\S}$ we write
$f$ for $\under{f}$, i.e. there is no notational distinction between a
morphism and its underlying function. This is legimitate because the forgetful
functors are faithful, i.e. every function $\under{A}\to\under{B}$
carries at most one morphism from $A$ to $B$. The morphisms
witnessing that $U$ and $F$ are monoidal
functors are denoted by
\[ \theta^U: \mathbf{1}_\mathscr{D} \to U\S,\quad \theta^U_{A,B}: UA\t UB\to U(A\t B), \]
and
\[ \theta^F: \S\to F(\mathbf{1}_\mathscr{D}),\quad \theta^F_{X,Y}: FX\t FY\to F(X\t Y).\]
Here $\S$ is viewed as the free $\S$-module on one generator. Finally, we fix two $\mathscr{D}$-monoids $(M,1,\bullet)$ and $(N,1,\bullet)$. We usually write $xy$ for $x\bullet y$.
\end{notation}
\begin{rem}\label{rem:plusmorph}
Our assumption that $F\dashv U: \Mod{\S}\to \mathscr{D}$ is a concrete monoidal adjunction has the following consequences:
\begin{enumerate}[(1)]
\item An $\S$-module $A$ and the corresponding algebra $UA$ in $\mathscr{D}$ have the
same underlying set, i.e. $\under{A}_\S = \under{UA}_\mathscr{D}$. Likewise, an
$\S$-module morphism $f:A\to B$ and the
corresponding morphism
$Uf: UA\to UB$ in $\mathscr{D}$ have the same underlying function.
\item Every bimorphism from $A,B$ to $C$ in $\Mod{\S}$ is also a bimorphism from $UA,UB$ to $UC$ in $\mathscr{D}$. To see this, suppose that $f: \under{A}_\S\t\under{B}_\S\to \under{C}_\S$ is a bimorphism of $\Mod{\S}$, and let $f': A\t B\to C$ be the corresponding $\S$-linear map. Consider the diagram below:
\[
\xymatrix{
\under{UA\t UB}_\mathscr{D} \ar[r]^{\theta^U} & \under{U(A\t B)}_\mathscr{D} \ar[r]^<<<<<<{Uf'} \ar@{=}[d] & \under{UC}_\mathscr{D} \ar@{=}[d]\\
& \under{A\t B}_\S \ar[r]^{f'} & \under{C}_\S\\
\under{UA}_\mathscr{D}\times\under{UB}_\mathscr{D} \ar[uu]^{t_\mathscr{D}} \ar@{=}[r] & \under{A}_\S\times\under{B}_\S \ar[u]^{t_\S} \ar[ur]_{f} &
}
\]
Here $t_\mathscr{D}$ and $t_\S$ denote the universal bimorphisms.
The left part commutes because the monoidal functor $\under{\mathord{-}}_\S$ is the
composite of the monoidal functors $U$ and $\under{\mathord{-}}_\mathscr{D}$ (as $U$ is part
of a concrete monoidal adjunction); cf. Lemma \ref{lem:moncompose} and Example
\ref{ex:monoidalfunc}(1). The upper right square commutes because $f'$ and
$Uf'$ have the same underlying function, see (1). The triangle is the
definition of $f'$. It follows that $f: \under{UA}_\mathscr{D}\times
\under{UB}_\mathscr{D}\to\under{UC}_\mathscr{D}$ is a bimorphism from $UA,UB$ to $UC$ in $\mathscr{D}$,
being the composite of the $\mathscr{D}$-bimorphism $t_\mathscr{D}$ with the $\mathscr{D}$-morphism
$Uf'\c \theta^U$.
\item As a consequence of (2), for any $\S$-algebra $A$ the multiplication
$\under{A}\times\under{A}\xrightarrow{\o}\under{A}$ is a bimorphism from $UA, UA$ to
$UA$. Morever, the sum $\under{A}\times \under{A}\xrightarrow{+} \under{A}$
carries a morphism $+: A\times A\to A$ in $\Mod{\S}$ because the latter is a
commutative variety. Applying $U$, we obtain a morphism $+: UA\times UA\to UA$
in $\mathscr{D}$. Analogously for the scalar product $\lambda\o\mathord{-}:
A\to A$ $(\lambda\in S)$.
\end{enumerate}
\end{rem}
\begin{rem}
For any two morphisms $p: M\to \S$ and $q: N\to \S$ in $\mathscr{D}$ and elements $m\in
\under{M}$ and $n\in\under{N}$, we have
\begin{equation}\label{eq:tensorformula}(p\c (\mathord{-}\bullet m))\t (q\c
(\mathord{-}\bullet n)) = (p\t q)\c (\mathord{-}\bullet (m\t n))
\end{equation}
where $\mathord{-}\bullet m: M\to M$, $\mathord{-}\bullet n: N\to N$ and $\mathord{-}\bullet
(m\t n): M\t N\to M\t N$ are morphisms of $\mathscr{D}$ because $M$, $N$ and $M\t N$
are $\mathscr{D}$-monoids (i.e. their multiplication is a $\mathscr{D}$-bimorphism). Indeed, for
any $m'\in\under{M}$ and $n'\in \under{N}$ we have
\begin{align*}
[p\c (\mathord{-}\bullet m))\t (q\c
(\mathord{-}\bullet n))](m'\t n') &= p(m'\bullet m) \t q(n'\bullet n)\\
&= p\t q((m'\t n')\bullet (m\t n))\\
&= [(p\t q)\c (\mathord{-} \bullet (m\t n))](m'\t n').
\end{align*}
In the first step we use the definition of the tensor product of two
morphisms, see Remark \ref{rem:tensorproducts}. In the second step, we use the
definition of the multiplication in $M\t N$, see Remark
\ref{rem:tensormonoid}. The last step is obvious. Since the elements $m'\t n'$
generate $M\t N$, see \ref{app:tensorproducts}, this proves
\eqref{eq:tensorformula}.
\end{rem}
\begin{proof}[Lemma \ref{lem:mastn}]
We prove the case where $\mathscr{D}$ is a variety of ordered algebras, the unordered
case being analogous. Consider the preorder $\preceq$ on $M\t N$ defined
$x\preceq y$ iff $\pi(x)\leq \pi(y)$. By \ref{app:congruences} we only need to
show
that $\preceq$ is a stable preorder of the $\mathscr{D}$-monoid $M\t N$ (cf. Remark
\ref{rem:tensormonoid}); then, since $\pi$ is surjective,
\[\pi(x)\bullet
\pi(y) := \pi(x\bullet y)\quad (x,y\in \under{M\t N})\]
gives a well-defined
$\mathscr{D}$-monoid structure on $M\ast N$ making $\pi$ a $\mathscr{D}$-monoid morphism. In
particular for $x=m\t n$ and $y=m'\t n'$, the multiplication is given by
\begin{align*}
(m\ast n)\bullet (m'\ast n') &= \pi(m\t n)\bullet \pi(m'\t n')\\
& = \pi((m\t n)\bullet (m'\t n'))\\
& = \pi((mm')\t (nn'))\\
&= (mm')\ast (nn')
\end{align*}
Thus let us show that $\preceq$ is indeed stable. Clearly $x\leq y$ in $M\t N$
implies $x\preceq y$
because $\pi$ is monotone. Also, $\preceq$ is stable w.r.t. all operations
of
$\mathscr{D}$, since $\pi$ is morphism of $\mathscr{D}$. It remains to show that $x\preceq y$
implies $x\bullet z\preceq y\bullet z$ and $z\bullet
x\preceq z\bullet y$ (equivalently, $\pi(x)\leq
\pi(y)$
implies $\pi(x\bullet z)\leq \pi(y\bullet z)$ and $\pi(z\bullet x)\leq
\pi(z\bullet y)$) for all $x,y,z\in \under{M\t N}$. We may assume that
$z=m\t n$ for some $m\in\under{M}$ and $n\in \under{N}$; since $M\t N$ is
generated by
these elements, see \ref{app:tensorproducts}, and $\bullet$ is a $\mathscr{D}$-bimorphism, this implies the statement
for all $z$. So suppose that $\pi(x)\leq \pi(y)$. By \eqref{eq:s1s2} this
implies that
\[ \sigma\c (p\t q) (x) \leq \sigma\c (p\t q)(y)\quad\text{for all $p: M\to \S$ and $q: N\to \S$ in
$\mathscr{D}$}. \]
In particular we get, for all $p$ and $q$,
\[ \sigma\c (\,(p\c (\mathord{-}\bullet m))\t (q\c (\mathord{-}\bullet n))\,)(x)\leq \sigma\circ (\,(p\circ
(\mathord{-}\bullet m))\t (q\c (\mathord{-}\bullet n))\,)(y) \]
using that $\mathord{-}\bullet m: M\to M$ and $\mathord{-}\bullet n: N\to N$ are morphisms
of $\mathscr{D}$ since the multiplication of $M$ resp. $N$ is a bimorphism.
Equivalently, by \eqref{eq:tensorformula},
\[ \sigma\c (p\t q)\c (\mathord{-} \bullet (m\t n))(x)\leq \sigma\c (p\t q)\c (\mathord{-}
\bullet (m\t n))(y)\]
Thus, since $z=m\t n$,
\[ \sigma\c (p\t q)(x \bullet z)\leq \sigma\c (p\t q) (y
\bullet z).\]
for all $p$ and $q$.
By the definition of $\pi$, this means precisely that $\pi(x\bullet z)\leq \pi(y\bullet z)$. The proof of $\pi(z\bullet x)\leq \pi(z\bullet y)$ is symmetric.
\end{proof}
\begin{lemma}\label{lem:afg}
For any $\S$-algebra $A$ and any two $\mathscr{D}$-monoid morphisms $f: M \to \overline U A$
and
$g:
N\to \overline U A$,
the product
$M\times UA \times N$ in $\mathscr{D}$ carries a $\mathscr{D}$-monoid structure with unit $(1,0,1)$ and multiplication
\[ (m,a,n)(m',a',n') := (mm',\,f(m)\o a'+a\o g(n'),\,nn'). \]
Here the $0$ and the sum and product in the second component are taken in the $\S$-algebra $A$. Denoting this $\mathscr{D}$-monoid by $A^{f,g}$, the product projections $\pi_M: A^{f,g}\to M$ and $\pi_N: A^{f,g}\to N$ are $\mathscr{D}$-monoid morphisms.
\end{lemma}
\begin{proof}
That $(1,0,1)$ is the unit is clear since $f(1)=1$ and $g(1)=1$. For associativity, we
compute
\begin{align*}
&[(m,a,n)(m',a',n')](m'',a'',n'')\\
&= (mm',\,f(m)\o a'+a\o g(n'),\,nn')(m'',a'',n'') \\
&= ((mm')m'',\,f(mm')\o a'' + [f(m)\o a'+a\o g(n')]\o g(n''),\,(nn')n'')\\
&= (m(m'm''),\,f(m)\o [f(m')\o a''+a'\o g(n'')] + a\o g(n'n''),\,n(n'n''))\\
&= (m,a,n)(m'm'',\,f(m')\o a''+a'\o g(n''),\,n'n'')\\
&= (m,a,n)[(m',a',n')(m'',a'',n'')]
\end{align*}
It remains to verify that the multiplication of $A^{f,g}$ is a
$\mathscr{D}$-bimorphism. For simplicity, let us just prove that for any binary operation $\gamma$ in the signature of $\mathscr{D}$, the multiplication preserves $\gamma$ in the right component. Indeed, we have
\begin{align*}
& (m,a,n)[\gamma((m',a',n'),(m'',a'',n''))] & \\
&= (m,a,n)(\gamma(m',m''),\,\gamma(a',a''),\,\gamma(n',n'')) & (1)\\
&= (m\gamma(m',m''),\,f(m)\o\gamma(a',a'')+a\o g(\gamma(n',n'')),\,n\gamma(n',n'')) &(2)\\
&= (\gamma(mm',mm''),\,\gamma(f(m)\o a',f(m)\o a'')+
a\o \gamma(g(n'),g(n'')),\,\gamma(nn',nn''))& (3)\\
&= (\gamma(mm',mm''),\,\gamma(f(m)\o a',f(m)\o a'')+
\gamma(a\o g(n'),a\o g(n'')),\,\gamma(nn',nn'')) & (4)\\
&= (\gamma(mm',mm''),\,\gamma(f(m)\o a'+a\o g(n'),f(m)\o a''+a\o g(n'')),\,\gamma(nn',nn'')) &(5)\\
&= \gamma(\,(mm',f(m)\o a'+a\o g(n'), nn'),\,(mm'',
f(m)\o a''+a\o g(n''), nn'')\,) & (6)\\
&= \gamma((m,a,n)(m',a',n'), (m,a,n)(m'',a'',n'')) & (7)
\end{align*}
Explanation of the individual steps:
(1) Definition of the operation $\gamma$ in the product $M\times UA\times N$
in $\mathscr{D}$.
(2) Definition of the multiplication in $A^{f,g}$.
(3) In the first and third component we use that the multiplication of $M$ and
$N$ is a $\mathscr{D}$-bimorphism; in the second component we use that the
multiplication of $A$ is a $\mathscr{D}$-bimorphism on $UA$, see Remark
\ref{rem:plusmorph}, and moreover that $g$ is a $\mathscr{D}$-morphism.
(4) Again we use that the multiplication of $A$ is a bimorphism on $UA$.
(5) $+$ is a morphism of $\mathscr{D}$, see Remark \ref{rem:plusmorph}.
(6) Definition of the operation $\gamma$ in the product $M\times UA\times N$
in $\mathscr{D}$.
(7) Definition of the multiplication in $A^{f,g}$.\\
That $\pi_{M}$ is a $\mathscr{D}$-monoid morphism follows from the
computation
\begin{align*}
\pi_{M}((m,a,n)(m',a',n')) &= \pi_{M}(mm',
f(m)\o a'+a\o g(n'),nn')\\
&= mm'\\
&= \pi_{M}(m,a,n)\pi_{M}(m',a',n').
\end{align*}
Analogously for $\pi_{N}$.
\end{proof}
\begin{proof}[Lemma \ref{lem:schuetzwelldef}]
Consider the following morphisms in $\mathscr{D}$:
\begin{align*}
f &\equiv (\, M \xrightarrow{\rho} M\t \mathbf{1}_\mathscr{D} \xrightarrow{M\t \iota_N} M\t N \xrightarrow{\pi}
{M\ast N} \xrightarrow{\eta} U F({M\ast N})\, )\\
g&\equiv(\, N \xrightarrow{\lambda} \mathbf{1}_\mathscr{D}\t N \xrightarrow{\iota_M\t N} M\t N \xrightarrow{\pi}
{M\ast N} \xrightarrow{\eta} U F({M\ast N})\, )
\end{align*}
Note that $\pi$ is a $\mathscr{D}$-monoid morphism by Lemma \ref{lem:mastn},
$\overline\eta=\eta: M\ast N\to \overline U\,\overline F(M\ast N)$ is a $\mathscr{D}$-monoid morphism by
Lemma \ref{lem:monlifting}, and that $\rho$, $\lambda$, $M\t \iota_N$ and
$\iota_M\t N$ are $\mathscr{D}$-monoid morphisms follows easily from the definition of
the monoid structure on tensor products, see Remark \ref{rem:tensormonoid}.
Thus $f$ and $g$ and $\mathscr{D}$-monoid morphisms. Applying Lemma \ref{lem:afg} to
the $\S$-algebra $A=\overline F(M\ast N)$ and the morphisms $f$ and $g$ yields the
Sch\"utzenberger product $M\diamond N = [\overline F(M\ast N)]^{f,g}$.
\end{proof}
\begin{notation}
We denote the product projections in $\mathscr{D}$ by
\[
\xymatrix{
& M\diamond N \ar[dl]_{\pi_{M}} \ar[d]^{\pi_{MN}} \ar[dr]^{\pi_{N}} &\\
M & UF(M\ast N) & N
}
\]
For any $\mathscr{D}$-monoid morphism $f: \Psi\Sigma^* \to M\diamond N$, we put
\[ f_M := \pi_M\c f: \Psi\Sigma^*\to M,\quad f_N:= \pi_N\c f: \Psi\Sigma^*\to N,\]
and
\[ f_{MN} := \pi_{MN}\c f: \Psi\Sigma^*\to UF(M\ast N).\]
Recall that we put $m\ast n := \pi(m\t n)$ for $m\in\under{M}$ and $n\in\under{N}$, and that $\eta: M\ast N\to UF(M\ast N)$ denotes the universal map.
\end{notation}
The following lemma appears in \cite{reutenauer79} for the case $\mathscr{D}=\mathbf{Set}$:
\begin{lemma}\label{lem:reutformula}
For any $\mathscr{D}$-monoid morphism $f: \Psi\Sigma^* \to M\diamond N$ and
$u\in\Sigma^*$ we have
\[ f_{MN}(u) = \sum_{u=vaw} \eta(f_M(v)\ast 1)\o f_{MN}(a)\o \eta(1\ast
f_N(w)),\]
where the sum ranges over all factorizations $u=vaw$ with $a\in \Sigma$
and $v,w\in \Sigma^*$.
\end{lemma}
\begin{proof}
The proof is by induction on the length of $u$. For $u=\epsilon$, we have
$f(\epsilon)=(1,0,1)$ since $f$ is a $\mathscr{D}$-monoid morphism, and thus
$f_{MN}(\epsilon)=0$ (the empty sum). Now suppose that the formula holds for some $u\in\Sigma^*$, and
consider a word $ub$ with $b\in\Sigma$. Then
\begin{align*}
f_{MN}(ub) &= \pi_{MN}(f (ub)) & (1) \\
&= \pi_{MN}(f(u)f(b)) & (2)\\
&= \eta(f_M(u)\ast 1)\o f_{MN}(b) + f_{MN}(u)\o \eta(1\ast f_N(b)) & (3)\\
&= \eta(f_M(u)\ast 1)\o f_{MN}(b) & \\
&~~~+\left[\sum_{u=vaw} \eta(f_M(v)\ast 1)\o f_{MN}(a)\o \eta(1\ast f_N(w))\right]\o \eta(1\ast f_N(b)) & (4)\\
&= \eta(f_M(u)\ast 1)\o f_{MN}(b) & \\
&~~~+\sum_{u=vaw} \eta(f_M(v)\ast 1)\o f_{MN}(a)\o \eta(1\ast f_N(w))\o
\eta(1\ast f_N(b))& (5)\\
&= \eta(f_M(u)\ast 1)\o f_{MN}(b)\o \eta(1\ast f_N(\epsilon))\\
&~~~+\sum_{u=vaw} \eta(f_M(v)\ast 1)\o f_{MN}(a)\o \eta(1\ast f_N(wb)) & (6)\\
&= \sum_{ub=vaw'} \eta(f_M(v)\ast 1)\o f_{MN}(a)\o \eta(1\ast f_N(w')) & (7)\\
\end{align*}
Explanation of the individual steps:
(1) Definition of $f_{MN}$.
(2) $f$ is a $\mathscr{D}$-monoid morphism.
(3) Definition of the multiplication in $M\diamond N$.
(4) Induction hypothesis.
(5) Distributive law in the $\mathds{K}$-algebra $\overline F(M\ast N)$.
(6) Definition of the multiplication in $M\ast N$, $\overline\eta=\eta: M\ast N \to
\overline U\overline F(M\ast N)$ is a $\mathscr{D}$-monoid morphism (see \ref{lem:monlifting}, and
$f_N$ is a $\mathscr{D}$-monoid morphism.
(7) Any factorization $ub=vaw'$ has either $w'=wb$ for some $w$, or $a=b$ and
$w'=\epsilon$.
\end{proof}
\begin{notation}
\begin{enumerate}[(1)]
\item For any two morphisms $p: M\to \S$ and $q: N\to \S$ in $\mathscr{D}$, we denote
by
$\overline{p\ast q}:
F(M\ast N) \to \S$ the adjoint transpose of
$p\ast q: M\ast N \to \S$.
\item Recall from Definition \ref{def:recog} the morphism $L_\mathscr{D}:
\Psi\Sigma^*\to \S$ in $\mathscr{D}$ corresponding to a language $L: \Sigma^*\to S$.
Since $L$ and $L_\mathscr{D}$ agree on $\Sigma^*$, we
usually drop the index and write $L$ for $L_\mathscr{D}$.
\end{enumerate}
\end{notation}
\begin{proof}[Theorem \ref{thm:schurec}]
Let $g: \Psi\Sigma^*\to M$ and $h: \Psi\Sigma^*\to N$ be $\mathscr{D}$-monoid
morphisms recognizing $K$ resp. $L$. Thus there exist morphisms $p: M\to \S$
and $q: N\to \S$ in $\mathscr{D}$ with $K=p\c g$ and $L=q\c h$. Fix a letter
$a\in\Sigma$. We define a $\mathscr{D}$-monoid morphism $f: \Psi\Sigma^*\to M\diamond
N$ that recognizes the languages $K$, $L$ and $KaL$.
\begin{enumerate}[(1)]
\item Let $f: \Psi\Sigma^*\to M\diamond N$ be the unique $\mathscr{D}$-monoid
morphism
defined on letters $b\in \Sigma$ by
\[ f(b) =
\begin{cases}
(g(b),1,h(b)), & b=a;\\
(g(b),0,h(b)), & b\neq a.
\end{cases} \]
Then $f_M(b)=g(b)$ for all $b\in \Sigma$ and therefore $f_M = g$,
since any $\mathscr{D}$-monoid morphism with domain $\Psi\Sigma^*$ is determined by its
values on the generators $\Sigma$. Similarly, we have $f_N = h$.
Since $f_{MN}(a)=1$ and $f_{MN}(b)=0$ for $b\neq a$, Lemma
\ref{lem:reutformula} gives, for all $u\in
\Sigma^*$,
\begin{equation}\label{eq:e12}
f_{MN}(u) = \sum_{u=vaw} \eta(g(v)\ast h(w)).
\end{equation}
\item $f$ recognizes the language $K$ via the morphism $p\c \pi_M$, since
\[ K = p\c g = p\c f_M = (p\c \pi_M) \c f. \]
Analogously, $f$ recognizes $L$ via $q\c\pi_N$.
\item We show that $f$ recognizes the language $KaL$ via the morphism
\[s \ \equiv \ (\ M\diamond N \xrightarrow{\pi_{MN}}
UF(M\ast N)\xrightarrow{U(\overline{p\ast q})} \S\ ).\]
Indeed, for all $u\in\Sigma^*$ we have
\begin{align*}
s\c f(u) &= \overline{p\ast q}(f_{MN}(u)) &\text{(def. $s$ and
$f_{MN}$)} \\
&= \overline{p\ast q}\left(\sum_{u=vaw} \eta(g(v)\ast h(w))\right)
&\text{\eqref{eq:e12}} \\
&=\sum_{u=vaw} \overline{p\ast q}(\,\eta(g(v)\ast h(w))\,)
&\text{($\overline{p\ast q}\in \Mod{\S}$)} \\
&=\sum_{u=vaw} [p\ast q](g(v)\ast h(w))
&\text{(def. $\overline{p\ast q}$)} \\
&=\sum_{u=vaw} [\sigma\c (p\t q)](g(v)\t h(w))
&\text{(def. $p\ast q$)} \\
&=\sum_{u=vaw} \sigma(\,p(g(v))\t q(h(w))\,)
&\text{(def. $p\t q$)} \\
&=\sum_{u=vaw} \sigma(\,K(v)\t L(w)\,)
&\text{(def. $p$, $q$)} \\
&=\sum_{u=vaw} K(v)\o L(w)
&\text{(def. $\sigma$)} \\
&= (KaL)(u) & \text{(def. $KaL$)}
\end{align*}
\end{enumerate}
\end{proof}
From now on, we suppose that the Assumptions \ref{asm:dual} hold.
\begin{rem} The Assumptions \ref{asm:dual}(i) and (ii) have the following
consequences:
\begin{enumerate}[(1)]
\item $F$ preserves finite objects. Indeed, let $D$ be a finite object of
$\mathscr{D}$, and express $D$ as a
quotient $e: \Psi X\twoheadrightarrow D$ for some finite set $X$. Since the left adjojnt
$F$ preserves epimorphisms, see \ref{app:adjunctions}, the morphism $Fe:
\mathcal{P}_f X = F\Psi X \twoheadrightarrow FD$ is an
epimorphism in $\Mod{\S}$ and therefore surjective by Assumption
\ref{asm:dual}(ii). Since $\mathcal{P}_f X$ is finite, so is $FD$.
\item $U$ also preserves finite objects by Remark \ref{rem:plusmorph}.
\item Consequently the Schützenberger product of two finite
$\mathscr{D}$-monoids $M$
and $N$ is finite. To see this, recall from \ref{app:tensorproducts} that the
tensor
product $M\t
N$ is
generated by the finite set $\under{M}\times \under{N}$ and is thus
finite by Assumption \ref{asm:dual}(i). Therefore the quotient $M\ast N$ of
$M\t N$ is also finite.
Since $F$ and $U$ preserve finite objects by (1) and (2),
$UF(M\ast N)$ is finite. Thus $M\diamond N$ is finite, being carried
by the finite object $M\times
UF(M\ast N)\times N$.
\end{enumerate}
\end{rem}
\begin{defn}\label{def:langoperations}
\begin{enumerate}[(1)]
\item The set of languages over $\Sigma$ forms an $\S$-algebra w.r.t. to the
\emph{sum}, \emph{scalar product} and \emph{Cauchy product} of languages,
defined by
\[ (K+L)(u) = K(u)+L(u),\, (\lambda L)(u) = \lambda L(u),\, (KL)(u) =
\sum_{u=vw} K(v)\o L(w) \]
for languages $K, L: \Sigma^*\to S$, $\lambda \in S$, and $u\in\Sigma^*$.
Identifying a letter $a\in\Sigma$ with the language $a:\Sigma^*\to S$ that
sends $a$ to $1$ and all other words to $0$, the marked Cauchy product
$KaL$ is thus the Cauchy product of the languages $K$, $a$ and $L$.
\item The \emph{derivatives} of a language $L: \Sigma^*\to S$ are
the languages $a^{-1}L,\,La^{-1}\colon \Sigma^*\to S$ ($a\in\Sigma)$
with
\[a^{-1}L(u) = L(au) \quad\text{and}\quad La^{-1}(u) = L(ua).\]
\end{enumerate}
\end{defn}
\begin{lemma}\label{lem:recclosed}
Let $f: \Psi\Sigma^* \to M$ be a $\mathscr{D}$-monoid morphism. Then the set of
languages recognized by $f$ is closed under sum, scalar products, and
derivatives.
\end{lemma}
\begin{proof}
\begin{enumerate}[(1)]
\item \emph{Closure under derivatives.} Suppose that $L: \Sigma^*\to S$ is a
language recognized by $M$, i.e. there is a morphism $p: M\to \S$ in $\mathscr{D}$ with
$L=p\c
f$. We claim that $f$ recognizes the left derivative $a^{-1}L$ via the
morphism $p' := p\c (f(a)\bullet \mathord{-})$. (Note that $f(a)\bullet\mathord{-}: M\to
M$ is a morphism of $\mathscr{D}$ because $\bullet$ is a $\mathscr{D}$-bimorphism.) Indeed, we
have for all $u\in\Sigma^*$:
\[ p'(f(u)) = p(f(a)\bullet f(u)) = p(f(au)) = L(au) = (a^{-1}L)(u). \]
and thus $L=p'\c f$, as claimed. Analogously for right derivatives.
\item \emph{Closure under sums.}
Let $K$ and $L$ be two languages recognized by $f$, i.e. $K = p\c f$ and
$L=q\c f$ for morphisms $p, q: M\to \S$ in $\mathscr{D}$. Denote by $\overline p, \overline q: FM\to \S$
the adjoint transposes of $p$ and $q$ in $\Mod{\S}$ (w.r.t. the adjunction
$F\dashv U$). Since $\Mod{\S}$ forms a commutative variety, we have the morphism $\overline p +
\overline q: FM\to \S$ in $\Mod{\S}$, defined by $[\overline p+\overline q](x) = \overline p(x) + \overline
q(x)$. Then
$K+L$ is recognized by the morphism $U(\overline{p}+\overline{q})\c \eta: M\to \S$, since
for all $u\in\Sigma^*$,
\[ [\overline{p}+\overline{q}](\eta(f(u))) = \overline p(\eta(f(u))) +\overline q(\eta(f(u))) =
p(f(u))+q(f(u)) =
K(u)+L(u). \]
Thus $K+L = U(\overline p+\overline q)\c \eta\c f$, i.e. $f$ recognizes $K+L$.
\item \emph{Closure under scalar product.} Analogous to (2).
\end{enumerate}
\end{proof}
\begin{proof}[Theorem \ref{thm:universalprop}]
We first establish two preliminary
technical results (steps (1) and (2)).
\begin{enumerate}[(1)]
\item Consider the commutative diagram below, where $\overline t$ is the adjoint
transpose of the universal bimorphism $t$; note that $S^{(\mathord{-})} = F\Psi$
since
the adjunction $F\dashv U: \Mod{\S}\to \mathscr{D}$ is concrete.
\[
\xymatrix{
\under{\S^{(\under{M}\times\under{N})}} \ar@{->>}[r]^{F\overline t} & \under{F(M\t
N)}
\ar@{->>}[r]^{F\pi} & \under{F(M\ast N)}\\
\under{\Psi(\under{M}\times\under{N})} \ar[u]^\eta \ar@{->>}[r]^{\overline t} &
\under{M\t N} \ar[u]_\eta \ar@{->>}[r]^\pi & \under{M\ast N} \ar[u]_\eta \\
\under{M}\times \under{N} \ar@{>->}[u] \ar[ur]_t &&
}
\]
Since $F\dashv U$ is a concrete monoidal adjunction, precomposing the unit
$\eta:
\under{\Psi(\under{M}\times\under{N})} \to
\under{\S^{(\under{M}\times\under{N})}}$
with the unit $\under{M}\times\under{N}\rightarrowtail
\under{\Psi(\under{M}\times\under{N})}$ of the adjunction $\Psi\dashv
\under{\mathord{-}}_\mathscr{D}: \mathscr{D}\to \mathbf{Set}$ gives the unit
$\under{M}\times\under{N}\rightarrowtail \under{\S^{(\under{M}\times\under{N})}}$ of
the adjunction
$\S^{(\mathord{-})} \dashv\under{\mathord{-}}_\S: \Mod{\S}\to \mathbf{Set}$.
Thus the $\S$-linear map $F\pi\c F\overline t$ maps a generator $(m,n)$ of
$\S^{(\under{M}\times\under{N})}$ to $\eta(m\ast n)$. Since the left adjoint
$F$ preserves epimorphisms, $F\pi$ and $F\overline t$ are surjective by Assumption
\ref{asm:dual}(ii). It
follows that every element of the $\S$-module $F(M\ast N)$ can be expressed as
a linear
combination $\sum_{j=1}^n \lambda_j \eta(m_j\ast n_j)$ with $\lambda_j\in
S$, $m_j\in \under{M}$ and $n_j\in\under{N}$.
\item Let $p: M\to \S$ and $q: N\to \S$ be two morphisms in $\mathscr{D}$, and let $L$
be
the language recognized by $f$ via $U(\overline{p\ast q})\c \pi_{MN}$, i.e.
\begin{equation}\label{eq:lpq} L :=
U(\overline{p\ast q})\c \pi_{MN}\c f = U(\overline{p\ast q})\c f_{MN}.\end{equation}
By (1), each element $f_{MN}(a)\in\under{F(M\ast N)}$ with $a\in\Sigma$ can be
expressed as a linear combination
\begin{equation}\label{eq:lincomb} f_{MN}(a) = \sum_{j=1}^{n_a}
\lambda_j^a\eta(m^{a}_j\ast n^{a}_j). \end{equation}
with $\lambda_j^a\in S$, $m^{a}_j\in \under{M}$ and $n^a_j\in\under{N}$. For
$a\in\Sigma$ and $j=1,\ldots, n_a$, put
\begin{equation}\label{eq:l1l2} L^{a,j}_M := p \c (\mathord{-}\bullet m^{a}_j) \c
f_M \quad\text{and}\quad L^{a,j}_N := q \c (n^{a}_j\bullet \mathord{-}) \c
f_N,\end{equation}
where $\mathord{-}\bullet m^{a}_j: M\to M$ and $n^{a}_j\bullet \mathord{-}: N\to N$ are
morphisms in $\mathscr{D}$
because the monoid multiplication of $M$ resp. $N$ is a $\mathscr{D}$-bimorphism. Then
$L$ can be expressed as the following linear combination of
languages (cf. Definition \ref{def:langoperations}):
\begin{equation}\label{eq:recformula}
L = \sum_{a\in\Sigma}\sum_{j=1}^{n_a} \lambda_j^a(L_M^{a,j}aL_N^{a,j}).
\end{equation}
To prove this, we compute for all $u\in\Sigma^*$:
\begin{align*}
L(u) &= \overline{p\ast q}(f_{MN}(u)) & (1)\\
&= \overline{p\ast q}\left(\sum_a \sum_{u=vaw} \eta(f_M(v)\ast 1)\o f_{MN}(a)\o
\eta(1\ast f_N(w))\right) & (2)\\
&= \overline{p\ast q}\left(\sum_a \sum_{u=vaw} \eta(f_M(v)\ast
1)\o\left(\sum_{j=1}^{n_a}
\lambda_j^a\eta(m^{a}_j\ast n^{a}_j)\right)\o
\eta(1\ast f_N(w))\right) & (3)\\
&= \overline{p\ast q}\left(\sum_a \sum_{u=vaw}
\sum_{j=1}^{n_a} \lambda_j^a \eta(f_M(v)\ast
1)\o
\eta(m^{a}_j\ast n^{a}_j)\o
\eta(1\ast f_N(w))\right) & (4)\\
&= \overline{p\ast q}\left(\sum_a \sum_{u=vaw}
\sum_{j=1}^{n_a} \lambda_j^a \eta(\,(f_M(v)m^a_j)\ast (n^a_j
f_N(w)\,)\,)\right) & (5)\\
&= \overline{p\ast q}\left(\sum_a \sum_{j=1}^{n_a} \lambda_j^a \left(\sum_{u=vaw}
\eta(\,(f_M(v)m^a_j)\ast (n^a_j f_N(w)\,)\,)\right)\right) & (6)\\
&= \sum_a \sum_{j=1}^{n_a} \lambda_j^a \left(\sum_{u=vaw}
\overline{p\ast q}(\,\eta(\,(f_M(v)m^a_j)\ast (n^a_j f_N(w)\,)\,)\,)\right) & (7)\\
&= \sum_a \sum_{j=1}^{n_a} \lambda_j^a \left(\sum_{u=vaw}
[p\ast q](\,(f_M(v)m^a_j)\ast (n^a_j f_N(w)\,)\,)\right) & (8)\\
&= \sum_a \sum_{j=1}^{n_a} \lambda_j^a \left(\sum_{u=vaw}
[\sigma\c (p\t q)](\,(f_M(v)m^a_j)\t (n^a_j f_N(w)\,)\,)\right) & (9)\\
&= \sum_a \sum_{j=1}^{n_a} \lambda_j^a \left(\sum_{u=vaw}
p(f_M(v)m^a_j))\o q(n^a_j f_N(w))\right) & (10)\\
&=\sum_a \sum_{j=1}^{n_a} \lambda_j^a \left(\sum_{u=vaw} L_M^{a,j}(v)
L_N^{a,j}(w)\right) & (11) \\
&=\sum_a \sum_{j=1}^{n_a} \lambda_j^a (L_M^{a,j}aL_N^{a,j})(u) & (12)
\end{align*}
Explanation of the individual steps:
(1) Definition of $L$.
(2) Lemma \ref{lem:reutformula}.
(3) Apply equation \eqref{eq:lincomb}.
(4) The multiplication in the $\S$-algebra $\overline F(M\ast N)$ distributes over
sum and
scalar product.
(5) Use the definition of the multiplication in $M\ast N$, and the fact that
$\eta=\overline\eta: \overline U \overline F(M\ast N)$ is a $\mathscr{D}$-monoid morphism by Lemma
\ref{lem:monlifting}.
(6) Interchange sums, and use that the scalar product distributes over sums.
(7) $\overline{p\ast q}$ is an $\S$-linear map.
(8) Definition of $\overline{p\ast q}$.
(9) Definition of $p\ast q$.
(10) Definition of $\sigma$ and $p\t q$.
(11) Definition of $L_M^{a,j}$ and $L_N^{a,j}$.
(12) Definition of marked Cauchy product and scalar product of languages.
\item We are prepared to prove the theorem. Suppose that $e:
\Psi\Sigma^*\twoheadrightarrow P$ is a surjective $\mathscr{D}$-monoid morphism satisfying the
assumptions of the theorem. For any $p: M\to\S$ in $\mathscr{D}$, the language
$(\Sigma^*\rightarrowtail \Psi\Sigma^* \xrightarrow{f} M\ast N \xrightarrow{\pi_{M}} M \xrightarrow{p}\S)$
is recognized by $\pi_M\c f$ and hence, by
assumption, also by $e$. Consequently there exists a morphism $h^p: P\to \S$
in $\mathscr{D}$ with $h^p\c e = p\c \pi_M\c f$. Analogously, for each $q: N\to \S$
there exists a morphism $h^q: P\to \S$ in $\mathscr{D}$ with $h^q\c e = q\c \pi_N\c f$.
Moreover, for any pair of morphisms $p: M\to \S$ and $q: N\to \S$ in $\mathscr{D}$, the
language $L$ of \eqref{eq:lpq} is recognized by $e$. Indeed, by definition the
languages $L^{a,j}_M$ and $L^{a,j}_M$ of \eqref{eq:l1l2} are recognized by
$f_M=\pi_M\c f$
resp. $f_N = \pi_N\c f$. Thus, by the assumptions on $e$, the marked
product
$L^{a,j}_M a L^{a,j}_M$ is recognized by $e$ for every $a\in \Sigma$ and
$j=1,\ldots, n_a$.
Therefore, by Lemma
\ref{lem:recclosed},
$e$ recognizes $L$. That is, there exists a morphism $h^{p,q}: P\to \S$ in
$\mathscr{D}$ with $h^{p,q}\c e = L$ ($= U(\overline{p\ast q})\c \pi_{MN}\c f$).
To summarize, we have the following commutative diagrams for all $p: M\to \S$
and $q: N\to \S$:
\[
\xymatrix{
\Psi\Sigma^* \ar[d]_f \ar@{->>}[rr]^e && P \ar[d]^{h^p}\\
M\diamond N \ar[r]_{\pi_M} & M \ar[r]_{p} & \S
}\quad
\xymatrix{
\Psi\Sigma^* \ar[d]_f \ar@{->>}[rr]^e && P \ar[d]^{h^q}\\
M\diamond N \ar[r]_{\pi_N} & N \ar[r]_{q} & \S
}
\]
\[
\xymatrix{
\Psi\Sigma^* \ar[d]_f \ar@{->>}[rr]^e && P \ar[d]^{h^{p,q}}\\
M\diamond N \ar[r]_<<<<{\pi_{MN}} & UF(M\ast N) \ar[r]_<<<<<{U(\overline{p\ast q})}
& \S
}
\]
Since the family $\{\pi_M,\pi_N,\pi_{MN}\}$ of product projections is
separating, and using Assumption
\ref{asm:dual}(iii), the combined family
\begin{align*}
& \{\,M\diamond N\xrightarrow{\pi_M} M\xrightarrow{p} \S\,\}_{p: M\to \S}\\
&\cup \{\,M\diamond
N\xrightarrow{\pi_N} N\xrightarrow{q} \S\,\}_{q: N\to \S}\\
& \cup \{\,M\diamond
N\xrightarrow{\pi_{MN}} UF(M\ast N)\xrightarrow{U(\overline{p\ast q})} \S\,\}_{p: M\to \S,\,q: N\to \S}
\end{align*}
is separating, see \ref{app:jointlyinj}. Thus diagonal fill-in, see
\ref{app:factsystems}, gives a unique $h: P\to M\diamond N$ in $\mathscr{D}$ with $h\c
e = f$. Moreover, since both $e$ and $f$ are $\mathscr{D}$-monoid morphisms and $e$ is
surjective, $h$ is a $\mathscr{D}$-monoid morphism.
\end{enumerate}
\end{proof}
\begin{defn}\label{rem:recclosed}
A language is called \emph{recognizable} if it is recognized by some
finite $\mathscr{D}$-monoid. Denote by $\Rec{\Sigma}$ the set of $\mathscr{D}$-recognizable
languages over the alphabet $\Sigma$.
\end{defn}
\begin{rem}
Recognizable languages are precisely the regular languages $L: \Sigma^*\to S$,
i.e. languages accepted by some finite Moore automaton with output set $S$;
see \cite{uacm16}[Lemma G.1].
\end{rem}
In \cite{ammu14} we established (working with a subset of our present
Assumptions \ref{asm:dual}) the following results:
\begin{theorem}[Ad\'amek, Milius, Myers, Urbat
\cite{ammu14}]\label{thm:rec_properties}
$\Rec{\Sigma}$ forms an algebra in the variety $\mathscr{C}$ w.r.t to the
$\mathscr{C}$-algebraic operations on languages. Moreover, $\Rec{\Sigma}$ is closed
under derivatives, and the
maps $a^{-1}(\mathord{-})$ and $(\mathord{-})a^{-1}$ on $\Rec{\Sigma}$ are morphisms
of $\mathscr{C}$, i.e. the $\mathscr{C}$-algebraic operations preserve derivatives.
\end{theorem}
\begin{theorem}[Ad\'amek, Milius, Myers, Urbat
\cite{ammu14}]\label{thm:locvar_properties}
For any finite set $\mathcal{V}_\Sigma$ of recognizable languages over $\Sigma$, the
following statements are equivalent:
\begin{enumerate}[(i)]
\item $\mathcal{V}_\Sigma$ is closed under the $\mathscr{C}$-algebraic operations and
derivatives.
\item There exists a finite $\mathscr{D}$-monoid $P$ and a surjective $\mathscr{D}$-monoid morphism $e: \Psi\Sigma^*\twoheadrightarrow P$
such that $e$ recognizes precisely the
languages in $\mathcal{V}_\Sigma$.
\end{enumerate}
\end{theorem}
\begin{proof}[Theorem \ref{thm:schurec2}]
Let $\mathcal{V}_\Sigma$ be the closure of the set $\L_{M,N}(f)$ under $\mathscr{C}$-algebraic
operations and derivatives. We first show that $\mathcal{V}_\Sigma$ is finite. Indeed,
since $\L_{M,N}(f)$ is a finite set, and every recognizable language has
only finitely many derivatives by Lemma \ref{lem:recclosed}, the closure
$\mathcal{V}_\Sigma'$ of $\L_{M,N}(f)$ under derivatives is a finite set. By Theorem
\ref{thm:rec_properties}, the closure of $\mathcal{V}_\Sigma'$ under $\mathscr{C}$-algebraic
operations is again closed under derivatives, and thus equal to $\mathcal{V}_\Sigma$.
Since
$\mathscr{C}$ is a locally finite variety by Assumption \ref{asm:dual}(iv), this shows
that $\mathcal{V}_\Sigma$ is finite. Therefore, by
Theorem
\ref{thm:locvar_properties}, there exists a finite $\mathscr{D}$-monoid $P$ and a
surjective $\mathscr{D}$-monoid morphism
$e: \Psi\Sigma^*\twoheadrightarrow P$ such that $e$
recognizes
precisely the languages of $\mathcal{V}_\Sigma\supseteq \L_{M,N}(f)$. Theorem
\ref{thm:universalprop}
gives a $\mathscr{D}$-monoid morphism $h: P\to M\diamond N$ with $h\c e = f$.
Then every language recognized by $f$ (say via the morphism $s: M\diamond
N\to
\S$ in $\mathscr{D}$)
is also recognized by $e$ (via the morphism $s\c h$), and therefore lies in
$\mathcal{V}_\Sigma$.
\end{proof}
\begin{rem}
For the categories $\mathscr{C}/\mathscr{D}$ of Example \ref{ex:predualcats}) one can drop
the
closure under derivatives in Theorem \ref{thm:schurec2}: in each
case, the closure $\mathcal{V}_\Sigma$ of $\L_{M,N}(f)$ under $\mathscr{C}$-algebraic
operations is already closed under derivatives. To see this, note that by
Theorem
\ref{thm:rec_properties} it suffices to show that every derivative of a
language in
$\L_{M,N}(f)$ lies in $\mathcal{V}_\Sigma$. The latter is clear for the languages $K$ and
$L$ recognized by $\pi_M\c f$ resp. $\pi_N\c f$: by Lemma \ref{lem:recclosed}
their derivatives are even elements of $\L_{M,N}(f)$.
Now consider the languages of the form $KaL$ in $\L_{M,N}(f)$. One easily
verifies that
\[
b^{-1}(KaL) =
\begin{cases}
(b^{-1}K)aL,& b\neq a;\\
(a^{-1}K)aL + K(\epsilon)L, & b=a,
\end{cases}
\]
and analogously for right derivatives. In the case $\mathscr{D}=\Vect{\mathds{K}}$, this
shows that any derivative of $KaL$ is a linear combination (i.e. a
$\Vect{\mathds{K}}$-algebraic combination) of languages in
$\L_{M,N}(f)$, and thus lies in $\mathcal{V}_\Sigma$. For the other examples
($\mathscr{D}=\mathbf{Set},\mathbf{Pos},\mathbf{JSL}$ with $\S=\{0,1\}$), the above case $b=a$
states that
\[a^{-1}(KaL) = \begin{cases}
(a^{-1}K)aL & \epsilon\not\in L;\\
(a^{-1}K)aL \cup L & \epsilon\in L.
\end{cases}
\]
Thus every derivative of $KaL$ is a finite union of languages in $\L_{M,N}(f)$
and
therefore lies in $\mathcal{V}_\Sigma$, since the union is part of the $\mathscr{C}$-algebraic
operations
for $\mathscr{D}=\mathbf{Set},\mathbf{Pos},\mathbf{JSL}$.
\end{rem}
\section{Details for the Examples}
\textbf{Details for Example \ref{ex:monoidalfunc}.}
\begin{enumerate}[(1)]
\item See Example \ref{ex:upsimonoidal}.
\item This is a special case of Example \ref{ex:upsimonoidal} with
$\mathscr{D}=\mathbf{JSL}$
and $\Psi = \mathcal{P}_f$ (cf. Example \ref{ex:psi}). Explicitly, $\mathcal{P}_f$ is
monoidal
w.r.t. the isomorphism $\theta_{X,Y}: \mathcal{P}_f X\t \mathcal{P}_f Y \cong \mathcal{P}_f(X\times
Y)$
whose inverse $\theta_{X,Y}^{-1}$ maps $\{(x_1,y_1),\ldots,
(x_n,y_n)\}\subseteq X\times
Y$ to $(\vee_{j=1}^n x_j\t y_j) \in X\t Y$, and the
identity morphism $\theta_1 = \mathit{id}: \mathbf{1}_\mathbf{JSL} \to \mathcal{P}_f 1$.
\item Let us first verify that $\mathcal{D}_f$ is a left adjoint to $U$.
The unit is given by the monotone map $\eta_X: X\to U\mathcal{D}_f(X)$ with $\eta(x)=\mathord{\downarrow} \{x\}$.
Given a monotone map $h: X\to UA$ into a semilattice $A$, let $\overline
h: \mathcal{D}_f(X)\to A$ be the function that maps a
finitely generated down-set $S=\mathord{\downarrow} S_0$ of $X$ to the finite join
$\bigvee S = \bigvee S_0\in \under{A}$. Clearly $\overline h$ is a semilattice morphism and satisfied $U(\overline h)\c \eta_X = h$. Moreover, since every finitely generated down-set is a finite union of one-generated downsets $\mathord{\downarrow} \{x\}$, $\overline h$ is uniquely determined by this property.
To show that $U$ is monoidal, observe that for any two semilattices $A$ and
$B$, the universal bimorphism $t_{A,B}: \under{A}\times\under{B}\to \under{A\t
B}$ is monotone, since it preserves joins in each component (and is thus
monotone in each component). Thus we can view $t_{A,B}$ as a morphism
$t_{A,B}^*: UA\times UB \to U(A\t B)$ in $\mathbf{Pos}$, and in complete analogy to
Example \ref{ex:upsimonoidal}(1) one can show that $U$ is monoidal w.r.t. the
maps $t_{A,B}^*$ and the unit $\eta: 1\mapsto UF1 = U\mathbf{1}_\mathbf{JSL}$, where
$1=\mathbf{1}_\mathbf{Pos}$ is the one-element poset. Similarly,
the proof that $\mathcal{D}_f$ is monoidal is analogous to Example
\ref{ex:upsimonoidal}(2), with $\mathbf{Set}$ replaced by $\mathbf{Pos}$ and $\under{\mathord{-}}$,
$\Psi$ by $U$, $\mathcal{D}_f$. Explicitly, $\mathcal{D}_f$ is
monoidal
w.r.t. the isomorphism $\theta^*_{X,Y}: \mathcal{D}_f X \t \mathcal{D}_f Y \cong \mathcal{D}_f(X\times
Y)$
whose inverse $(\theta_{X,Y}^*)^{-1}$ maps a down-set $\{(x_1,y_1),\ldots,
(x_n,y_n)\}\subseteq X\times
Y$ to $(\vee_{j=1}^n x_j\t y_j) \in X\t Y$, and the
identity morphism $\theta_1 = \mathit{id}: \mathbf{1}_\mathbf{JSL} \to \mathcal{D}_f 1$.
\end{enumerate}
~\\
\noindent\textbf{Details for Example \ref{ex:monoidaladj}.}
The adjunction $\Psi\dashv \under{\mathord{-}}: \mathscr{D}\to\mathbf{Set}$ is monoidal by Example
\ref{ex:upsimonoidal}. The proof given in that example also works for $U\dashv
\mathcal{D}_f: \mathbf{JSL}\to \mathbf{Pos}$, replacing $\mathscr{D}$ by $\mathbf{JSL}$ and $\mathbf{Set}$ by $\mathbf{Pos}$. That
$\mathsf{Id}\dashv\mathsf{Id}: \mathscr{D}\to\mathscr{D}$ is a monoidal adjunction is trivial. \\
\noindent\textbf{Details for Example \ref{ex:monoidaladj}.}
(3), (4), (5) are trivially concrete monoidal adjunctions. So is $\mathcal{P}_f\dashv
\under{\mathord{-}}: \mathbf{JSL}\to \mathbf{Set}$, since the monoidal adjunction of $\mathbf{Set}$ is the
identity adjunction $\mathsf{Id}\dashv \mathsf{Id}: \mathbf{Set}\to\mathbf{Set}$. Concerning $\mathcal{D}_f\dashv U:
\mathbf{JSL}\to \mathbf{Pos}$, we clearly have $\under{\mathord{-}}_\mathbf{JSL} = \under{\mathord{-}}_\mathbf{Pos} \c U$
and $\mathcal{P}_f = \mathcal{D}_f \c \Psi_\mathbf{Pos}$. Since also the units and counits of these
three adjunctions compose accordingly, the composite of the adjunction $\mathcal{D}_f
\dashv U$ with
$\Psi_\mathbf{Pos} \dashv \under{\mathord{-}}_\mathscr{D}$ is the adjunction $\mathcal{P}_f \dashv
\under{\mathord{-}}_\mathbf{JSL}$.\\
\noindent\textbf{Details for Example \ref{ex:jointlym}.}
\begin{enumerate}[(1)]
\item For $\mathscr{D}=\mathbf{Set}$ or $\mathbf{Pos}$ and $\S=\{0,1\}$, the family $\{\,M\times N
\xrightarrow{p\times q} \{0,1\}\times \{0,1\} \xrightarrow{\sigma} \{0,1\}\,\}_{p,q}$ is
separating, where $\sigma(m,n)=m\o n$ is the multiplication of $S$. We prove this only for $\mathscr{D}=\mathbf{Pos}$, the argument
for $\mathscr{D}=\mathbf{Set}$ being analogous. Let $(m,n)$ and $(m',n')$ be two
elements of $M\times N$ with $(m,n)\not\leq(m',n')$, say
$m\not\leq m'$. Choose $p: M\to \{0,1\}$ to be the monotone map with
$p(x)=0$ iff $x\leq m'$, and $q: N\to\{0,1\}$ to be the constant map
on $1$. Then
\[\sigma\c (p\times q)(m,n) = p(m)\o q(n) = 1\o 1 = 1\] and \[\sigma\c
(p\times q)(m',n') = p(m')\o q(n') = 0\o 1 = 0,\] so $\sigma\c
(p\times q)(m,n) \not\leq \sigma\c(p\times q)(m',n')$. This shows
that
the family $(\sigma\c (p\times q))_{p,q}$ is separating.
\item Similarly, for $\mathscr{D}=\Vect{\mathds{K}}$ the family $\{\,M\t N
\xrightarrow{p\times q} \mathds{K}\t \mathds{K} \xrightarrow{\sigma} \mathds{K}\,\}_{p,q}$ is separating. To see this,
recall that if $M$ and $N$ are a vector spaces with bases $\{b_i\}_{i\in I}$
resp. $\{c_j\}_{j\in J}$, then
the
tensor product $M\t
N$ has the basis $\{b_i\t c_j\}_{i\in I, j\in J}$. It suffices to
show that, for any element $x\in
\under{M\t
N}$ with $x\neq 0$, there are linear maps $p: M\to \mathds{K}$ and $q: N\to \mathds{K}$ with
$\sigma\c
(p\t q)(x)\neq 0$. Suppose that $x=\sum_{i\in I, j\in J}
\lambda_{i,j} (b_{i}\t c_j)$ with $\lambda_{i,j}\in \mathds{K}$.
Since $x\neq 0$, there exist $i_0\in I$ and $j_0\in J$ with
$\lambda_{i_0,j_0}\neq 0$. Let $p: M\to \mathds{K}$ and $q:N\to\mathds{K}$ be the
linear maps defined by
\[ p(b_i) =
\begin{cases}
1,&i=i_0;\\
0,&i\neq i_0,
\end{cases}
\quad\text{and}\quad
p(c_j) =
\begin{cases}
1,&j=j_0;\\
0,&j\neq j_0.
\end{cases}
\]
Then
\begin{align*}
\sigma\c (p\t q)(x) = \sum_{i,j} \lambda_{i,j} [\sigma\c
(p\t q)(b_i\t c_j)]
= \sum_{i,j} \lambda_{i,j}
(p(b_i)\o q(c_j))
= \lambda_{i_0,j_0}
\neq 0.
\end{align*}
Here we use the linearity of $\sigma$ and $p\t q$ in the first step, the definition of $\sigma$ and $p\t q$ in the second step, and the definition of $p$ and $q$ in the last step.
\item If $\mathscr{D}=\mathbf{JSL}$ and $M$ and $N$ are finite idempotent semilattices, we can
describe $M\ast N$ as
follows. Consider the surjective semilattice morphism
\[ e\,\equiv\,(\, \mathcal{P}_f(\under{M}\times\under{N}) \xrightarrow{\overline t} M\t N
\xrightarrow{\pi} M\ast N \,), \]
where $\overline t$ is the adjoint transpose of the universal bimorphism $t:
\under{M}\times \under{N}\to \under{M\t N}$; cf. \ref{app:tensorproducts}. Let $\equiv\subseteq
\mathcal{P}_f(\under{M}\times\under{N})\times
\mathcal{P}_f(\under{M}\times\under{N})$ be the kernel of $e$, i.e. $X\equiv Y$
iff $\pi\c \overline t (X) = \pi\c \overline t (Y)$. By the definition of $\pi$, the latter means precisely that $\sigma\c
(p\t q)\c \overline t(X) =
\sigma\c
(p\t q)\c \overline t(Y)$ for all semilattice morphisms $p: M\to \{0,1\}$
and $q:
N\to \{0,1\}$. Since such morphisms correspond to ideals, see Example
\ref{ex:monrec}(3), we have $X\equiv Y$
iff, for all ideals $I\subseteq M$ and $J\subseteq N$,
\[ \exists (m,n)\in X: m\not\in I \wedge n\not\in
J\quad\Leftrightarrow\quad \exists (m',n')\in Y: m'\not\in I \wedge
n'\not\in J.\]
Since $\equiv$ is a semilattice congruence on
$\mathcal{P}_f(\under{M}\times\under{N})$, being the kernel of a semilattice
morphism, $X\equiv Y$ and $X\equiv Z$ implies $X=X\cup X \equiv Y\cup Z$.
Thus, for every $X\subseteq \under{M}\times\under{N}$ there exists a largest
set $[X]\subseteq \under{M}\times\under{N}$ with $X\equiv [X]$, viz.
the union of all $Y$ with $X\equiv Y$. It follows that $X\mapsto [X]$ defines
a
closure operator on $\mathcal{P}_f(\under{M}\times\under{N})$, and that every
equivalence class of $\equiv$ contains a unique closed subset of
$\under{M}\times\under{N}$ (viz. the union of all sets in the equivalence
class). One easily
verifies that $[X]$ consists of those elements $(x,y)\in
\under{M}\times\under{N}$ such that, for all ideals $I\subseteq M$ and $J\subseteq N$,
\[ x\not\in I\wedge y\not\in J\quad\Rightarrow\quad \exists(m,n)\in X: m\not\in I\wedge n\not\in J. \]
Since $\equiv$ is the kernel of $\pi\c \overline t$, we have
$M\ast N
\cong\mathcal{P}_f(\under{M}\times\under{N})/\mathord{\equiv}$. Identifying the
equivalence classes of $\equiv$ with the closed subsets of $M\times N$, the
join in the idempotent semiring $M\ast N$ is given by $[X]\vee [Y]=[X\cup Y]$, and the multiplication
(see Lemma \ref{lem:mastn}) by $[X][Y]=[XY]$.
\end{enumerate}
~\\
\noindent\textbf{Details for Example \ref{ex:predualcats}}
Clearly Assumption \ref{asm:dual}(i) holds for our examples
$\mathscr{D}=\mathbf{Set}$, $\mathbf{Pos}$, $\mathbf{JSL}$ and $\Vect{\mathds{K}}$ ($\mathds{K}$ a finite field). Also (ii) is
well-known in all these cases. For $\mathbf{Set}$, $\mathbf{Pos}$ and $\Vect{\mathds{K}}$, see e.g.
\cite[Example 7.40]{adherstr}. For $\mathbf{JSL}$, see \cite{hornkimuara}. Concerning
(iii), that $\mathscr{D}(M,\S)$
and $\mathscr{D}(N,\S)$ are separating is easy to
verify in all cases. Also, that $\{\,U(\overline{p\ast q}) \}_{p,q}$ forms a
separating family is trivial for $\mathscr{D}=\mathbf{JSL}$ and
$\Vect{\mathds{K}}$, since here $U=\mathsf{Id}$ and $U(\overline{p\ast q}) = p\ast q$, and the
morphisms $p\ast q$ are separating by definition.
It remains to consider the cases $\mathscr{D}=\mathbf{Set}$ and $\mathscr{D}=\mathbf{Pos}$. We only treat
$\mathbf{Pos}$, the argument for $\mathbf{Set}$ being the discrete special case. We
need to show
that the family of monotone maps
\[\{\,\overline{p\times q}: \mathcal{D}_f (M\times N)\to \{0,1\}\,\}_{p:
M\to\{0,1\},\,q:N\to\{0,1\}}\] is separating, where $\S=\{0,1\}$
(considered as a poset) is ordered by $0<1$.
Note that $\overline{p\times q}$ maps a finitely generated down-set $X\subseteq M\times
N$ to $1$ iff
there exists a pair $(m,n)\in X$ with $p(m)=1$ and $q(n)=1$.
Let $X\not\subseteq Y$ be two finitely generated down-sets of $M\times N$.
Choose an element $(m,n)\in
X\setminus
Y$, and define monotone maps $p: M\to \{0,1\}$ and $q: N\to \{0,1\}$ by
\[ \text{$p(x)=1$ iff $x\geq m$}\quad\text{resp.}\quad \text{$q(y)=1$ iff
$y\geq n$}.\]
Then we get
\[ \overline{p\times q}(X) = 1\not\leq 0 = \overline{p\times q}(Y), \]
i.e. $\overline{p\times q}$ separates $X$ and $Y$, as desired.
\section{Introduction}
Since the early days of automata theory, it has been known that regular
languages are precisely the languages
recognized by finite monoids. This observation is the origin of
algebraic language theory. One of the classical and ongoing challenges of this
theory is
the algebraic treatment of the concatenation of languages.
The most important
tool for this purpose is the \emph{Sch\"utzenberger product} $M\diamond N$
of
two monoids $M$ and $N$, introduced in \cite{sch65}. Its key
property is that it recognizes all marked products of languages
recognized by $M$ and $N$. Later,
Reutenauer \cite{reutenauer79} showed
that $M\diamond N$ is the ``smallest'' monoid with this property: any
language recognized by $M\diamond N$ is a
boolean combination of such marked products.
In the past decades, the original notion of language recognition by finite
monoids has
been refined to other algebraic structures, namely to ordered monoids
by Pin~\cite{pin95}, to idempotent semirings by Pol\'ak~\cite{polak01}, and to
associative algebras over a
field
by Reutenauer~\cite{reu80}. For all these structures,
a Sch\"utzenberger
product was introduced separately~\cite{pin03,klimapolak10,reu80}.
Moreover,
Reutenauer's characterization of the languages recognized by Sch\"utzenberger
products has been adapted to ordered monoids and idempotent semirings,
replacing
boolean combinations by positive boolean
combinations~\cite{pin03} and finite unions~\cite{klimapolak10}, respectively.
This paper presents a unifying approach to Sch\"utzenberger products,
covering the aforementioned constructions and results as special cases. Our
starting point is the observation that all the algebraic structures appearing
above (monoids, ordered monoids, idempotent semirings, and algebras over a
field $\mathds{K}$) are
\emph{monoids} interpreted in some variety $\mathscr{D}$ of algebras or
ordered
algebras, viz.\ $\mathscr{D}=$ sets, posets, semilattices, and $\mathds{K}$-vector spaces,
respectively.
Next, we note that these categories $\mathscr{D}$ are related to the
category $\Mod{\S}$ of modules over some semiring $\S$. Indeed, semilattices and
vector spaces are precisely modules over the two-element idempotent
semiring
$\S=\{0,1\}$ and the field $\S=\mathds{K}$, respectively. And every set
or poset freely
\emph{generates} a semilattice (i.e.\ a module over $\{0,1\}$), viz.
the semilattice
of finite subsets or finitely generated down-sets. Precisely speaking,
each of the above categories $\mathscr{D}$ admits a \emph{monoidal adjunction}
\begin{equation}\label{eq:monoidaladj}
\xymatrix@+1em{
\Mod{\S} \ar@<1.2ex>[r]^-{U}
\ar@{}[r]|-{\top} &
\mathscr{D}
\ar@<1.2ex>[l]^-{F}}
\end{equation}
for some semiring $\S$, where $U$ is a
forgetful functor and $F$ is a free construction.
In this paper we introduce the Sch\"utzenberger
product at the level of an
abstract monoidal adjunction \eqref{eq:monoidaladj}: for any two
$\mathscr{D}$-monoids $M$ and
$N$, we construct a $\mathscr{D}$-monoid $M\diamond N$ that recognizes all
marked products of languages recognized by $M$ and $N$ (Theorem
\ref{thm:schurec}), and prove that $M\diamond N$
is the ``smallest''
$\mathscr{D}$-monoid with this property (Theorem \ref{thm:universalprop}).
Further, we derive a characterization of the languages recognized
by $M\diamond N$ in
the spirit of Reutenauer's theorem \cite{reutenauer79}. To this end, we
consider another variety $\mathscr{C}$ that is
\emph{dual} to $\mathscr{D}$ on the level
of finite
algebras. For example, for $\mathscr{D}=$ sets we choose $\mathscr{C}=$ boolean algebras, since Stone's representation theorem gives a dual
equivalence between finite boolean algebras and finite sets.
We then prove that every language recognized by
$M\diamond N$ is a
``$\mathscr{C}$-algebraic combination'' of languages recognized by $M$
and $N$ and
their marked products (Theorem \ref{thm:schurec2}). The explicit
use of duality
makes our proof conceptually different
from the original ones.
By instantiating \eqref{eq:monoidaladj} to the proper adjunctions, we
recover the Sch\"utzenberger product for monoids, ordered monoids, idempotent
semirings and algebras over a field, and obtain a new Sch\"utzenberger
product
for algebras over a commutative semiring. Moreover, our Theorems
\ref{thm:schurec} and \ref{thm:schurec2} specialize to the corresponding
results
\cite{reutenauer79,pin03,klimapolak10} for (ordered) monoids and idempotent
semirings. In the
case of $\mathds{K}$-algebras,
Theorem \ref{thm:schurec2} appears to be a new result. Apart from that, we
believe that the main contribution of our paper is the identification of a categorical setting for language concatenation. We hope that the
generality and the conceptual nature of our approach can contribute to an improved understanding
of the various ad hoc constructions and separate results appearing in the
literature.
\paragraph{Related work.} In recent years, categorical approaches to
algebraic language theory have been a growing research topic. The present
paper is a natural continuation of \cite{amu15}, where
we showed that the construction of syntactic monoids works at the level
of
$\mathscr{D}$-monoids in any commutative variety $\mathscr{D}$, allowing for a uniform treatment of
syntactic (ordered) monoids, idempotent semirings and algebras over a
field. The systematic use of duality in algebraic language theory
originates in the work of Gehrke,
Grigorieff, and Pin \cite{ggp08}, who interpreted Eilenberg's variety theorem
in terms of Stone duality. In our papers
\cite{ammu14,ammu15,cu15} we extended their approach to an abstract
Stone-type
duality, leading to a uniform view of several Eilenberg-type theorems for
regular
languages. See also \cite{bcr15,uramoto16} for related
duality-based work. Recently, Boja\'nczyk \cite{boj15} proposed to use
\emph{monads}
instead of monoids to get a categorical grasp on languages beyond finite
words. By
combining this idea with our duality framework, we established in
\cite{camu16,uacm16} a variety theorem that covers
most Eilenberg-type correspondences known in the literature, e.g. for
languages
of finite words, infinite words, words on linear orderings, trees, and cost
functions.
\section{Preliminaries}
In this paper we study monoids and language recognition in algebraic categories.
The reader is assumed to be familiar with basic universal algebra and category
theory; see the Appendix for a toolkit. We call a
variety $\mathscr{D}$ of algebras or ordered algebras \emph{commutative} if, for any
two algebras $A,B\in \mathscr{D}$, the set $[A,B]$ of morphisms from $A$ to $B$ forms an
algebra of $\mathscr{D}$ with operations taken pointwise in $B$. Our applications involve
the commutative varieties $\mathbf{Set}$ (sets), $\mathbf{Pos}$ (posets, as
ordered algebras without any operation), $\mathbf{JSL}$ (join-semilattices with $0$),
$\Vect{\mathds{K}}$ (vector spaces over a field $\mathds{K}$) and
$\Mod{\S}$ (modules over a commutative semiring $\S$ with $0, 1$). Note that $\mathbf{JSL}$ and
$\Vect{\mathds{K}}$ are special cases of $\Mod{\S}$ for $\S=\{0,1\}$, the two-element
semiring with $1+1=1$, and $\S=\mathds{K}$, respectively.
\begin{notation}
Let $\mathscr{A}$, $\mathscr{B}$, $\mathscr{C}$, $\mathscr{D}$ always denote
commutative varieties of algebras or ordered algebras. We write
$\Psi=\Psi_\mathscr{D}\colon
\mathbf{Set}\rightarrow\mathscr{D}$ for the left adjoint to the forgetful functor
$\under{\mathord{-}}\colon\mathscr{D}\rightarrow\mathbf{Set}$; thus $\Psi X$ is the free algebra of $\mathscr{D}$
over~$X$. For simplicity, we assume that $X$ is a subset
of $\under{\Psi X}$ and the universal map $X\rightarrowtail \under{\Psi X}$ is the inclusion. Denote by $\mathbf{1}_\mathscr{D} = \Psi 1$ the free
one-generated algebra.
\end{notation}
\begin{expl}\label{ex:psi}
\begin{enumerate}[(1)]
\item For $\mathscr{D}=\mathbf{Set}$ or $\mathbf{Pos}$ we have $\Psi X = X$ (discretely ordered).
\item For $\mathscr{D} = \mathbf{JSL}$ we get $\Psi X
= (\mathcal{P}_f X,\cup)$, the semilattice of finite
subsets of $X$.
\item For $\mathscr{D}=\Mod{\S}$ we have $\Psi X=
\S^{(X)}$, the $\S$-module of all finite-support
functions $X\rightarrow \S$ with sum and scalar product defined pointwise.
\end{enumerate}
\end{expl}
\begin{defn} Let $A,B,C\in \mathscr{D}$. By a
\emph{bimorphism} from
$A$, $B$ to
$C$ is meant a function $f\colon \under{A}\times\under{B}\rightarrow\under{C}$ such that the maps
$f(a,\mathord{-})\colon \under{B}\rightarrow \under{C}$ and $f(\mathord{-},b)\colon \under{A}\rightarrow
\under{C}$ carry morphisms of $\mathscr{D}$ for every $a\in\under{A}$ and $b\in
\under{B}$. A \emph{tensor product} of $A$ and $B$ is a universal bimorphism
$t_{A,B}\colon \under{A}\times \under{B} \rightarrow \under{A\t B}$, in the sense that
for any bimorphism $f\colon \under{A}\times \under{B} \rightarrow \under{C}$
there is a
unique $f'\colon A\t B \rightarrow C$ in $\mathscr{D}$ with $f'\c t_{A,B} = f$. We denote by $a\t
b$ the element $t_{A,B}(a,b)\in\under{A\t B}$.
\end{defn}
\begin{expl}
In $\mathbf{Set}$ and $\mathbf{Pos}$ we have $A\t B= A\times B$.
In $\Mod{\S}$, $A\t B$ is the usual tensor product of $\S$-modules,
and $t_{A,B}$ is the universal $\S$-bilinear map.
\end{expl}
\begin{rem}\label{rem:tensorproducts}
\begin{enumerate}[(1)]
\item Tensor products exist in any commutative variety $\mathscr{D}$, see \cite{bn76}.
\item $\t$ is associative and has unit
$\mathbf{1}_\mathscr{D}$; that is, there are natural isomorphisms
\[\alpha_{A,B,C}\colon (A\t B)\t C \cong A\t (B\t C),\quad \rho_A\colon A\t \mathbf{1}_\mathscr{D}
\cong A,\quad \lambda_A\colon \mathbf{1}_\mathscr{D}\t A \cong A.\]
\item Given $f\colon A\to C$ and $g\colon B\to D$ in $\mathscr{D}$, denote by $f\t g\colon A\t B\to C\t
D$ the morphism induced by the bimorphism
$\under{A}\times\under{B}\xrightarrow{f\times g} \under{C}\times\under{D}
\xrightarrow{t_{C,D}}\under{C\t D}$.
\end{enumerate}
\end{rem}
\begin{defn}
\label{def:dmonoid}
A \emph{$\mathscr{D}$-monoid} is a triple
$(M, 1, \bullet)$ where $M$ is an object of $\mathscr{D}$ and $(\under{M}, 1,\bullet)$
is a monoid whose multiplication
$\under{M}\times\under{M}\xrightarrow{\bullet}\under{M}$ is a bimorphism of $\mathscr{D}$. A
\emph{morphism} $h\colon (M, 1_M,\bullet_M)\rightarrow (N, 1_N,\bullet_N)$ of
$\mathscr{D}$-monoids is a morphism $h\colon M\rightarrow N$ in $\mathscr{D}$ with $h(1_M)=1_N$ and
$h(m\bullet_M m') = h(m)\bullet_N h(m')$ for $m,m'\in\under{M}$.
We denote
the category of $\mathscr{D}$-monoids by $\Mon{\mathscr{D}}$.
\end{defn}
\begin{expl}\label{ex:dmon}
Monoids in $\mathscr{D}=\mathbf{Set}$, $\mathbf{Pos}$, $\mathbf{JSL}$ and $\Mod{\S}$ are precisely
monoids, ordered monoids, idempotent semirings, and associative algebras over
$\S$.
\end{expl}
\begin{proposition}[see \cite{ammu14}]
The free $\mathscr{D}$-monoid on a set $\Sigma$ is carried by $\Psi
\Sigma^*\in \mathscr{D}$, the free algebra in $\mathscr{D}$ on the set $\Sigma^*$ of finite
words
over $\Sigma$. Its multiplication extends the concatenation of words in
$\Sigma^*$, and
its
unit is the empty word $\epsilon$.
\end{proposition}
\begin{expl}
\begin{enumerate}[(1)]
\item In $\mathscr{D}=\mathbf{Set}$ or $\mathbf{Pos}$ we have $\Psi \Sigma^* = \Sigma^*$ (discretely
ordered).
\item In $\mathscr{D}=\mathbf{JSL}$ we have $\Psi\Sigma^* = \mathcal{P}_f
\Sigma^*$,
the idempotent semiring of all finite languages over $\Sigma$ w.r.t. union
and
concatenation of languages.
\item In $\mathscr{D}=\SMod{\S}$ we get $\Psi\Sigma^* = \S[\Sigma]$, the $\S$-algebra
of all polynomials $\Sigma_{i=1}^n c(w_i)w_i$ (equivalently, finite-support functions
$c\colon\Sigma^*\to \S$) w.r.t. the usual sum, scalar product and multiplication of polynomials.
\end{enumerate}
\end{expl}
\begin{rem}\label{rem:moncat}
Since the multiplication $\bullet\colon \under{M}\times\under{M}\to\under{M}$ of a
$\mathscr{D}$-monoid $(M,1,\bullet)$ forms a bimorphism, it
corresponds to a morphism $\mu_M\colon M\t M\to M$ in $\mathscr{D}$, mapping $m\t
m'\in \under{M\t M}$ to $m\bullet m'\in \under{M}$. Likewise, the
unit
$1\in\under{M}$ corresponds to the morphism $\iota_M\colon \mathbf{1}_\mathscr{D} \to M$
sending the generator of $\mathbf{1}_\mathscr{D}$ to $1$. We can thus represent
a
$\mathscr{D}$-monoid
$(M,1,\bullet)$ as the
triple $(M,\iota_M,\mu_M)$.
\end{rem}
\begin{rem}\label{rem:tensormonoid}
For any two $\mathscr{D}$-monoids $M$ and $N$,
the tensor
product
$M\t N$ in $\mathscr{D}$ carries a $\mathscr{D}$-monoid structure with unit
$\mathbf{1}_\mathscr{D} \xrightarrow{\cong} \mathbf{1}_\mathscr{D}\t \mathbf{1}_\mathscr{D} \xrightarrow{\iota_M\t \iota_N} M\t N$
and multiplication
$ (M\t N)\t (M\t N) \xrightarrow{\cong} (M\t M)\t
(N\t N) \xrightarrow{\mu_M\t \mu_N} M\t N$,
see e.g. \cite{porst08}. Equivalently, the unit of $M\t N$ is the
element
$1_M\t 1_N$, and the multiplication is determined by $(m\t
n)\bullet
(m'\t n') = (m\bullet_M m')\t (n\bullet_N n')$.
\end{rem}
\begin{defn}\label{def:monoidalfunc}
A \emph{monoidal functor} $(G,\theta)\colon\mathscr{C}\to\mathscr{D}$ is a functor $G\colon \mathscr{C}\to
\mathscr{D}$ with a morphism $\theta_1\colon \mathbf{1}_\mathscr{D} \to G\mathbf{1}_\mathscr{C}$ and morphisms
$\theta_{A,B}\colon GA\t GB \to G(A\t B)$ natural in $A,B\in \mathscr{C}$ such that the
following squares commute (omitting indices): \vspace{-0.4cm}
\begin{center}
\begin{tabular}[x]{cc}
\parbox{0.6\textwidth}{
$
\xymatrix@[email protected]{
(GA\t GB)\t GC \ar[r]^\alpha \ar[d]_{\theta\t GC} & GA \t
(GB\t
GC) \ar[d]^{GA\t \theta}\\
G(A\t B) \t GC \ar[d]_{\theta} & GA \t G(B\t C)
\ar[d]^\theta \\
G((A\t B)\t C) \ar[r]_{G\alpha}& G(A\t (B\t C))
}
$
}
&
\parbox{0.4\textwidth}{
$
\[email protected]{
GA\t \mathbf{1}_\mathscr{D} \ar[r]^-{GA\t \theta} \ar[d]_-{\rho} & GA\t G\mathbf{1}_\mathscr{C}
\ar[d]^-\theta\\
GA & G(A\t \mathbf{1}_\mathscr{C}) \ar[l]_-{G\rho}
}
$\\
$\[email protected]{
\mathbf{1}_\mathscr{D}\t GA \ar[r]^-{\theta\t GA} \ar[d]_-{\lambda} & G\mathbf{1}_\mathscr{C}\t GA
\ar[d]^-\theta\\
GA & G(\mathbf{1}_\mathscr{C}\t A) \ar[l]_-{G\lambda}
}
$
}
\end{tabular}
\end{center}
Given another monoidal functor $(G',\theta')\colon \mathscr{C}\to \mathscr{D}$, a natural
transformation $\phi\colon G\to G'$ is called \emph{monoidal} if the following
diagrams commute:
\[
\xymatrix@[email protected]{
GA\t GB \ar[r]^{\phi_A\t \phi_B} \ar[d]_{\theta} & G'A\t G'B
\ar[d]^{\theta'}\\
G(A\t B) \ar[r]_{\phi_{A\t B}} & G'(A\t B)
}\qquad
\[email protected]{
& \mathbf{1}_\mathscr{D} \ar[dl]_\theta \ar[dr]^{\theta'} & \\
G\mathbf{1}_\mathscr{C} \ar[rr]_{\phi_{\mathbf{1}_\mathscr{C}}} && G'\mathbf{1}_\mathscr{C}
}
\]
\end{defn}
\begin{expl}\label{ex:monoidalfunc}
\begin{enumerate}[(1)]
\item The functor
$\under{\mathord{-}}\colon \mathscr{D}\to\mathbf{Set}$ is monoidal w.r.t. the universal map
$1\rightarrowtail \under{\mathbf{1}_\mathscr{D}}=\under{\Psi 1}$ and the bimorphisms
$t_{A,B}\colon
\under{A}\times\under{B}\to \under{A\t B}$. Its left adjoint $\Psi\colon
\mathbf{Set}\to\mathscr{D}$
is also monoidal: there
is a natural \emph{isomorphism} $\theta_{X,Y}\colon
\Psi X \t \Psi Y \cong \Psi(X\times Y)$ with $\theta_{X,Y}^{-1}(x,y)=x\t
y$ for $(x,y)\in X\times Y$. Together with $\theta_1 = \mathit{id}\colon \mathbf{1}_\mathscr{D}\to
\Psi
1$, this makes $\Psi$ a monoidal functor.
\item In particular, the functors
$\under{\mathord{-}}\colon \mathbf{JSL}\to\mathbf{Set}$ and $\mathcal{P}_f\colon \mathbf{Set}\to\mathbf{JSL}$ (see
Example \ref{ex:psi}(2)) are
monoidal w.r.t. the morphisms chosen as in (1).
\item The forgetful functor $U\colon \mathbf{JSL}\to\mathbf{Pos}$ has a left adjoint
$\mathcal{D}_f\colon
\mathbf{Pos}\to\mathbf{JSL}$ constructed as
follows. For any
poset $A$ and
$X_0\subseteq A$ denote by
$\mathord{\mathord{\downarrow}} X_0 :=
\{\,a\in A:
a\leq x \text{ for some $x\in X_0$}\,\}$ the down-set generated by $X_0$.
Then $\mathcal{D}_f$ maps a poset $A$ to
$\mathcal{D}_f(A) \coloneqq \{\,X\subseteq A : X=\mathord{\downarrow} X_0 \text{ for some finite $X_0\subseteq
A$}\,\},
$ the semilattice (w.r.t. union) of finitely generated down-sets of $A$, and a
monotone map $h\colon A\to B$ to
the semilattice morphism
$\mathcal{D}_f(h) \colon \mathcal{D}_f(A)\to \mathcal{D}_f(B)$
with $\mathcal{D}_f(h)(X) = \mathord{\downarrow} h[X]$. Both $U$ and $\mathcal{D}_f$ carry monoidal
functors; the required morphisms, see Definition \ref{def:monoidalfunc}, are
chosen in analogy to $\under{\mathord{-}}$ and $\mathcal{P}_f$ in (2).
\item As a trivial example, the identity functor $\mathsf{Id}\colon \mathscr{D}\to\mathscr{D}$ is monoidal
w.r.t. the identity
morphisms $\mathit{id}\colon \mathbf{1}_\mathscr{D}\to \mathsf{Id}(\mathbf{1}_\mathscr{D})$ and $\mathit{id}\colon \mathsf{Id}(A)\t\mathsf{Id}(B)\to \mathsf{Id}(A\t
B)$.
\end{enumerate}
\end{expl}
The importance of monoidal functors is that they preserve monoid structures:
\begin{lemma}\label{lem:monlifting}
Let $(G,\theta)\colon \mathscr{C}\to \mathscr{D}$ be a monoidal functor. Then $G$ lifts to the
functor $\overline G\colon \Mon{\mathscr{C}}\to
\Mon{\mathscr{D}}$ mapping a $\mathscr{C}$-monoid $(M,\iota,\mu)$ to the $\mathscr{D}$-monoid
\vspace{-0.2cm}
\[ (GM,\ \mathbf{1}_\mathscr{D}\xrightarrow{\theta} G\mathbf{1}_\mathscr{C} \xrightarrow{G\iota} GM,\ GM\t GM \xrightarrow{\theta}
G(M\t
M)\xrightarrow{G\mu} GM), \]
and a $\mathscr{C}$-monoid morphism $h$ to $Gh$.
\end{lemma}
\begin{expl}
\begin{enumerate}[(1)]
\item $\mathcal{P}_f\colon \mathbf{Set}\to\mathbf{JSL}$ lifts to the functor
$\overline\mathcal{P}_f\colon \Mon{\mathbf{Set}}\to \Mon{\mathbf{JSL}}$ that maps a monoid $M$ to the semiring
$\overline\mathcal{P}_f M$ of finite subsets of $M$, with union as addition, and
multiplication
$XY=\{\,xy: x\in Y,\,y\in Y\,\}$.
\item $\mathcal{D}_f\colon \mathbf{Pos}\to\mathbf{JSL}$ lifts to $\overline\mathcal{D}_f\colon
\Mon{\mathbf{Pos}}\to\Mon{\mathbf{JSL}}$, mapping an ordered monoid $M$ to the semiring
$\overline\mathcal{D}_f(M)$ of finitely generated down-sets of $M$, with
union as addition, and
multiplication $XY = \mathord{\downarrow}\{\,xy: x\in X,\,y\in Y\,\}$.
\end{enumerate}
\end{expl}
\begin{lemma}\label{lem:moncomp}
Let $(G,\theta)\colon \mathscr{A}\to \mathscr{B}$ and $(H,\sigma)\colon
\mathscr{B}\to\mathscr{C}$ be monoidal functors. Then the composite $HG\colon \mathscr{A}\to \mathscr{C}$
is a
monoidal functor w.r.t. to
$H(\theta_1)\c \sigma_1\colon \mathbf{1}_\mathscr{C}\to HG(\mathbf{1}_\mathscr{A})$ and $H(\theta_{A,B})\c
\sigma_{GA,GB}\colon HGA\t HGB \to HG(A\t B)$.
\end{lemma}
\begin{defn}\label{def:monoidaladj}
A \emph{monoidal adjunction} between $\mathscr{C}$ and $\mathscr{D}$ is an adjunction
$F\dashv U\colon \mathscr{C}\to\mathscr{D}$ such that $U$ and $F$ are monoidal
functors and the unit $\eta\colon \mathsf{Id}_\mathscr{D}\to UF$ and counit $\epsilon\colon FU\to \mathsf{Id}_\mathscr{C}$
are monoidal
natural
transformations.
\end{defn}
\begin{expl}\label{ex:monoidaladj}
$\mathsf{Id}\dashv \mathsf{Id}\colon \mathscr{D}\to\mathscr{D}$, $\mathcal{D}_f\dashv U\colon \mathbf{JSL}\to\mathbf{Pos}$ and
$\Psi\dashv
\under{\mathord{-}}\colon\mathscr{D}\to\mathbf{Set}$ are monoidal adjunctions. We call the latter the
\emph{monoidal adjunction of $\mathscr{D}$}.
\end{expl}
\begin{rem}
If $(H\dashv V\colon \mathscr{C}\to \mathscr{B},\eta',\epsilon')$ and $(G\dashv
U\colon \mathscr{B}\to \mathscr{A},
\eta, \epsilon)$ are monoidal adjunctions, so is the composite adjunction
$(HG\dashv UV\colon \mathscr{C}\to \mathscr{A},\,U\eta' G\c \eta,\,\epsilon'\c H\epsilon V)$.
Here $HG$ and $UV$ are the composites of Lemma \ref{lem:moncomp}.
\end{rem}
\begin{defn}
A monoidal adjunction $F\dashv U\colon
\mathscr{C}\to\mathscr{D}$ is called a \emph{concrete monoidal adjunction} if its composite with
the
monoidal adjunction of $\mathscr{D}$ is the monoidal adjunction of $\mathscr{C}$.
\end{defn}
\section{Languages and Algebraic Recognition}
In this section we set the scene for our categorical approach to
Sch\"utzenberger products. For the rest of this paper let us fix a commutative
variety $\mathscr{D}$ of algebras or ordered algebras, a commutative semiring
$\S=(S,+,\o,0,1)$, and a concrete
monoidal adjunction $F\dashv U\colon \Mod{\S}\to \mathscr{D}$ with unit $\eta\colon \mathsf{Id}\to UF$.
Thus we have the
diagram of
functors below, where $\overline U$ and $\overline F$ are the lifted functors, see
Lemma \ref{lem:monlifting}, the
vertical functors are the forgetful functors, and $\Psi$ and $\S^{(\mathord{-})}$ are the left
adjoints to the forgetful functors of $\mathscr{D}$ and $\Mod{\S}$, see Example \ref{ex:psi}.
\vspace{-0.2cm}
\[
\[email protected]{
\Alg{\S} \ar[d] \ar@<0.5ex>[rr]^{\overline{U}} && \Mon{\mathscr{D}}
\ar@<0.5ex>[ll]^{\overline F} \ar[d] \\
\Mod{\S} \ar@<0.5ex>[rr]^U \ar@<-1.0ex>[dr]_{\under{\mathord{-}}} & & \mathscr{D}
\ar@<0.5ex>[ll]^F
\ar@<2.0ex>[dl]^{\under{\mathord{-}}} \\
& \mathbf{Set} \ar@<-1.0ex>[ur]^{\Psi} \ar[ul]_<<<<{\S^{(\mathord{-})}} &
}
\]
\begin{expl}\label{ex:adjunctions}
In our applications we will choose the concrete monoidal adjunctions listed
below. (The third and last column will be explained
later.)
\small
\begin{center}
\begin{tabular}{|c|c|c|c|c|l|c|}
\hline
&$\S$ & $\mathscr{C}$ & $\mathscr{D}$ & $\xymatrix{\Mod{\S}
\ar@<0.5ex>[r]^<<<<<{U} &
\mathscr{D}
\ar@<0.5ex>[l]^<<<<<{F}}$ & $\mathscr{D}$-monoids & $M\diamond N$
carried by\\
\hline
1&$\{0,1\}$ &$\mathbf{BA}$ & $\mathbf{Set}$ & $\xymatrix{\mathbf{JSL}
\ar@<0.5ex>[r]^<<<<<{\under{\mathord{-}}} &
\mathbf{Set}
\ar@<0.5ex>[l]^<<<<<{\mathcal{P}_f}}$ & monoids & $M\times \mathcal{P}_f(M\times
N)\times N$ \\
2&$\{0,1\}$ & $\mathbf{DL}$ & $\mathbf{Pos}$ & $\xymatrix{\mathbf{JSL} \ar@<0.5ex>[r]^<<<<<U & \mathbf{Pos}
\ar@<0.5ex>[l]^<<<<<{\mathcal{D}_f}}$ & ord. monoids & $M\times \mathcal{D}_f(M\times
N)\times N$ \\
3&$\{0,1\}$ & $\mathbf{JSL}$ & $\mathbf{JSL}$ & $\xymatrix{\mathbf{JSL}
\ar@<0.5ex>[r]^<<<<<{\mathsf{Id}} & \mathbf{JSL}
\ar@<0.5ex>[l]^<<<<<{\mathsf{Id}}}$ & id. semirings & $M\times (M\ast N)\times
N$\\
4&$\mathds{K}$ & $\Vect{\mathds{K}}$ & $\Vect{\mathds{K}}$ & $\xymatrix{\Vect{\mathds{K}}
\ar@<0.5ex>[r]^<<<<<{\mathsf{Id}} & \Vect{\mathds{K}}
\ar@<0.5ex>[l]^<<<<<{\mathsf{Id}}}$ & $\mathds{K}$-algebras & $M\times (M\t N)\t N$\\
5&$\S$ & ? & $\Mod{\S}$ & $\xymatrix{\Mod{\S}
\ar@<0.5ex>[r]^<<<<<{\mathsf{Id}} & \Mod{\S}
\ar@<0.5ex>[l]^<<<<<{\mathsf{Id}}}$ & $\S$-algebras & $M\times (M\ast N)\t N$\\
\hline
\end{tabular}
\end{center}
\normalsize
\end{expl}
\begin{notation}
We can view the semiring $\S$ as (i) an $\S$-algebra $\S_{\mathbf{Alg}}\in
\Alg{\S}$ with scalar product given by the multiplication of $\S$, (ii) a
$\mathscr{D}$-monoid $\S_{\mathbf{Mon}}\in \Mon{\mathscr{D}}$ (by applying $\overline U$ to
$\S_{\mathbf{Alg}}$), (iii) an $\S$-module $\S_{\mathbf{Mod}}\in \Mod{\S}$ (by
applying the forgetful functor to $\S_{\mathbf{Alg}}$) and (iv) an object
$\S_\mathscr{D}$ of $\mathscr{D}$ (by applying $U$ to $\S_{\mathbf{Mod}}$). The $\mathscr{D}$-monoid
$\S_\mathbf{Mon}$ is carried by the object $\S_\mathscr{D}$, and its multiplication is a
morphism of $\mathscr{D}$ that we denote by $\sigma\colon \S_\mathscr{D} \t \S_\mathscr{D} \to \S_\mathscr{D}$. For ease of notation we will usually drop the indices and simply write $\S$ for $\S_\mathscr{D}$, $\S_{\mathbf{Mod}}$, etc.
\end{notation}
\begin{defn}\label{def:recog}
\begin{enumerate}[(1)]
\item A \emph{language} (a.k.a. a
\emph{formal power series}) over a finite alphabet $\Sigma$ is a map $L\colon
\Sigma^*\to S$. Denote by $L_\mathscr{D}\colon \Psi\Sigma^*\to \S$ the adjoint
transpose
of
$L$ w.r.t. the adjunction $\Psi\dashv \under{\mathord{-}}\colon \mathscr{D}\to\mathbf{Set}$.
A
$\mathscr{D}$-monoid morphism $f\colon \Psi\Sigma^*\to M$ \emph{recognizes} $L$ if there
is a morphism $p\colon M\to \S$ in
$\mathscr{D}$ with $L_\mathscr{D} =p\c f$. In this case, we also say that $M$ \emph{recognizes}
$L$ (via
$f$ and $p$).
\item The \emph{marked Cauchy product} of two languages $K,L\colon \Sigma^*\to S$
w.r.t. a letter $a\in\Sigma$ is the language
$KaL\colon \Sigma^*\to S$ with $(KaL)(u) = \sum_{u=vaw} K(v)\o L(w)$.
\end{enumerate}
\end{defn}
For $\S=\{0,1\}$, a language $L\colon \Sigma^*\to \{0,1\}$ corresponds to a classical language $L\subseteq \Sigma^*$ by taking the preimage of $1$. Under this identification, we have
$KaL = \{\,vaw: v\in K,\,w\in L\,\}$. Our concept of language recognition by
$\mathscr{D}$-monoids originates in \cite{amu15} and specializes to several related
notions
from
the literature:
\begin{expl}\label{ex:monrec}
\begin{enumerate}[(1)]
\item $\mathscr{D} = \mathbf{Set}$ with $\S=\{0,1\}$: a map $p\colon M\to \{0,1\}$
corresponds
to a
subset $p^{-1}[1]\subseteq M$. Thus a monoid
morphism $f\colon \Sigma^*\rightarrow M$ recognizes the language $L\subseteq \Sigma^*$ iff $L$
is the
preimage under $f$ of some subset of $M$. This is the classical notion of
language recognition by a monoid, see e.g.\ \cite{pin15}.
\item $\mathscr{D} = \mathbf{Pos}$ with $\S_\mathbf{Pos}=\{0<1\}$: given an ordered
monoid $M$, a monotone map $p\colon M\rightarrow\{0,1\}$ defines an upper set
$p^{-1}[1]\subseteq M$. Hence a monoid morphism $f\colon \Sigma^*\rightarrow M$ recognizes
$L\subseteq \Sigma^*$ iff
$L$ is the preimage under $f$ of some upper set of $M$. This notion of
recognition is
due to Pin~\cite{pin95}.
\item $\mathscr{D}=\mathbf{JSL}$ with $\S_\mathbf{JSL}=\{0<1\}$: for
any idempotent semiring $M$, a semilattice morphism $p\colon M\rightarrow \{0,1\}$
defines an ideal $I=p^{-1}[0]$, i.e.
a nonempty down-set closed under joins. Hence a language $L\subseteq \Sigma$ is
recognized by a
semiring morphism $f\colon \mathcal{P}_f \Sigma^* \rightarrow M$ via $p$ iff $\overline L =
\Sigma^*\cap
f^{-1}[I]$. Here we identify $\Sigma^*$ with the set of all singleton languages
$\{w\}$, $w\in \Sigma^*$. This is the concept of language recognition by
idempotent semirings introduced
by Pol\'ak \cite{polak01}.
\item $\mathscr{D}=\Mod{\S}$: given an $\S$-algebra $M$, a formal power series $L\colon
\Sigma^*\to S$ is
recognized by
$f\colon \S[\Sigma]\to
M$
via $p\colon M\rightarrow \S$ iff $L_{\Mod{\S}}= p\c f$. This notion of recognition is due
to Reutenauer~\cite{reu80}. If $\S$ is a commutative ring, the
power series
recognizable
by $\S$-algebras of finite type (i.e. $\S$-algebras whose underlying
$\S$-module is finitely generated) are precisely rational power series.
\end{enumerate}
\end{expl}
\section{The Sch\"utzenberger Product}
We are ready to introduce the Sch\"utzenberger product for
$\mathscr{D}$-monoids. Fix two $\mathscr{D}$-monoids
$(M,1,\bullet)$ and $(N,1,\bullet)$, and write $xy$ for $x\bullet y$.
Our goal is to construct a $\mathscr{D}$-monoid $M\diamond N$ that recognizes
all
marked products of languages recognized by $M$ and $N$, and is the
``smallest'' such $\mathscr{D}$-monoid
(Theorems \ref{thm:schurec}, \ref{thm:universalprop},
\ref{thm:schurec2}).
\vspace{0.05cm}
\begin{construction} As a preliminary step, we define a
$\mathscr{D}$-monoid $M\ast N$ as fol-
\noindent\parbox{0.77\textwidth}{
lows. Call a family $\{\,f_i\colon A\to B_i\,\}_{i\in I}$ in
$\mathscr{D}$ \emph{separating} if the morphism $f \colon A\to \prod_i B_i$ with $f(a)
= (f_i(a))_{i\in I}$ is
injective (resp.
order-reflecting when $\mathscr{D}$ is a variety of ordered algebras). Any family
$\{f_i\}$ yields a separating family $\{\,f_i'\colon A'\to B_i\,\}_{i\in
I}$
by
factorizing
$f = m\c \pi$ with $\pi$ surjective and $m$ injective
(resp. order-reflecting), and setting $f_i' \coloneqq p_i\c m$, where $p_i$ is
\vspace{0.05cm}
}
\parbox{0.23\textwidth}{
$
\[email protected]@C-1.5em{
A \ar@/^4ex/[rr]^{f_i}\ar@{->>}[r]^\pi \ar[dr]_{f} & A'
\ar[r]^{f_i'}
\ar@{>->}[d]^-{m} & B_i\\
& \prod_i B_i \ar[ur]_{p_i} &}$
}
the projection. Now consider the family of all morphisms
$\sigma\c (p\t q)\colon M\t N\to \S$, where $p\colon M\to\S$ and $q\colon
N\to \S$ are arbitrary morphisms in $\mathscr{D}$. Applying the above construction to
this family $\{\,\sigma\c (p\t q)\,\}_{p,q}$ gives an algebra $M\ast N$ in
$\mathscr{D}$, a surjective morphism $\pi\colon M\t
N\twoheadrightarrow M\ast
N$, and a separating family $\{\,p\ast
q\colon M\ast N\to \S\,\}_{p,q}$, making the following diagram commute for all
$p$ and $q$:
\begin{equation}\label{eq:s1s2}
\begin{gathered}
\[email protected]{
& \S\otimes \S \ar[dr]^\sigma &\\
M\otimes N \ar[ur]^{p\otimes q} \ar@{->>}[r]_\pi & {M\ast
N}
\ar[r]_<<<<<{p\ast q} & \S\\
}
\end{gathered}
\end{equation}
\end{construction}
\begin{notation} For any $m\in \under{M}$ and $n\in\under{N}$, we write $m\ast
n$ for the element
$\pi(m\t n)\in\under{M\ast N}$.
\end{notation}
\begin{lemma}\label{lem:mastn}
There exists a (unique) $\mathscr{D}$-monoid structure on $M\ast N$ such that $\pi: M\t
N\twoheadrightarrow M\ast N$ is a
$\mathscr{D}$-monoid morphism. The multiplication is determined by $(m\ast n)\bullet
(m'\ast n') = (mm')\ast (nn')$, and the unit is $1\ast 1$.
\end{lemma}
\begin{expl}\label{ex:jointlym}
For $\mathscr{D}=\mathbf{Set}$, $\mathbf{Pos}$ or $\Vect{\mathds{K}}$, the family
$\{\,\sigma\c (p\t q)\,\}_{p,q}$ is
already
separating, and therefore $M\ast N = M\t N$ and $p\ast q = \sigma\c
(p\t q)$. For $\mathscr{D}=\mathbf{JSL}$ and in case $M$ and $N$ are \emph{finite} idempotent
semirings, we can
describe the idempotent semiring $M\ast N$ as
follows. For any subset $X\subseteq M\times N$, let $[X]\subseteq M\times N$ consist of
those elements $(m,n)\in M\times N$ such that, for all ideals $I\subseteq M$ and
$J\subseteq N$ with $m\not\in I$ and $n\not\in
J$, there exists some $(x,y)\in X$ with $x\not\in I$ and
$y\not\in J$. This gives us the closure operator $X\mapsto [X]$ on the power set
of $M\times N$ in~\cite{klimapolak10}. One can show that $M\ast N$ is isomorphic to the idempotent semiring
of all closed
subsets of $M\times N$, with sum and product defined by $[X]\vee[Y] = [X\cup
Y]$ and
$[X][Y]=[XY]$, where $XY=\{\,xy: x\in Y,\,y\in Y\,\}$.
\end{expl}
\begin{defn}
The \emph{Sch\"utzenberger product} of $M$ and $N$ is the $\mathscr{D}$-monoid
$M\diamond N$ carried
by the product $M\times UF(M\ast N)\times N$ in $\mathscr{D}$ and equipped with the
following monoid structure: representing elements $(m,a,n)\in
\under{M}\times\under{F(M\ast N)}\times \under{N}$ as upper triangular matrices
$
\begin{pmatrix}
m & a\\
0 & n
\end{pmatrix},
$
the multiplication and unit are given by
\[
\begin{pmatrix}
m & a\\
0 & n
\end{pmatrix}
\begin{pmatrix}
m' & a'\\
0 & n'
\end{pmatrix}
=
\begin{pmatrix}
mm'~~ & \eta(m\ast 1)\o a'+a\o \eta (1\ast n')\\
0 & nn'
\end{pmatrix}
\quad\text{and}\quad
\begin{pmatrix}
1 & 0\\
0 & 1
\end{pmatrix}.
\]
Here $\eta\colon M\ast N\to UF(M\ast N)$ is the universal map, and the sum,
product and $0$ in the upper right components are taken in the $\S$-algebra $\overline
F(M\ast N)$.
\end{defn}
\begin{lemma}\label{lem:schuetzwelldef}
$M\diamond N$ is a well-defined $\mathscr{D}$-monoid, and the product projections
$\pi_M\colon
M\diamond N\to M$ and $\pi_N\colon M\diamond N \to N$ are $\mathscr{D}$-monoid morphisms.
\end{lemma}
\begin{expl}
For the categories and adjunctions of Example \ref{ex:adjunctions}, we recover
four notions of Sch\"utzenberger products known in the literature, and obtain
a new Sch\"utzenberger product for $\S$-algebras:
\begin{enumerate}[(1)]
\item $\mathscr{D}=\mathbf{Set}$: given monoids $M$ and $N$, the monoid $M\diamond N$
is carried by the set $M\times \mathcal{P}_f (M\times
N)\times
N$, with multiplication and unit
\[
\begin{pmatrix}
m & X\\
0 & n
\end{pmatrix}
\begin{pmatrix}
m' & X'\\
0 & n'
\end{pmatrix}
=
\begin{pmatrix}
mm'~~ & mX'\cup Xn'\\
0 & nn'
\end{pmatrix}
\quad\text{and}\quad
\begin{pmatrix}
1 & \emptyset\\
0 & 1
\end{pmatrix},
\]
where $mX' = \{\,(my,z): (y,z)\in X'\,\}$ and $Xn' = \{\,(y,zn'):
(y,z)\in X\,\}$. This is
the original construction of Sch\"utzenberger \cite{sch65}.
\item $\mathscr{D}=\mathbf{Pos}$: for ordered monoids $M$ and $N$, the ordered monoid
$M\diamond N$ is carried by the poset $M\times \mathcal{D}_f(M\times
N)\times N$ with multiplication and unit
\[
\begin{pmatrix}
m & X\\
0 & n
\end{pmatrix}
\begin{pmatrix}
m' & X'\\
0 & n'
\end{pmatrix}
=
\begin{pmatrix}
mm'~~ & \mathord{\downarrow}(mX'\cup Xn')\\
0 & nn'
\end{pmatrix}
\quad\text{and}\quad
\begin{pmatrix}
1 & \emptyset\\
0 & 1
\end{pmatrix}.
\]
This construction is due to Pin \cite{pin03}.
\item $\mathscr{D}=\mathbf{JSL}$: given idempotent semirings $M$ and $N$, the idempotent
semiring $M\diamond N$ is carried by the semilattice $M\times (M\ast
N)\times N$. If $M$ and $N$ are finite, $M\ast N$ is the
idempotent semiring of closed subsets of
$M\times N$ by Example \ref{ex:jointlym}, and the multiplication and unit of
$M\diamond N$ are given by
\[
\begin{pmatrix}
m & X\\
0 & n
\end{pmatrix}
\begin{pmatrix}
m' & X'\\
0 & n'
\end{pmatrix}
=
\begin{pmatrix}
mm'~~ & [mX'\cup Xn']\\
0 & nn'
\end{pmatrix}
\quad\text{and}\quad
\begin{pmatrix}
1 & \emptyset\\
0 & 1
\end{pmatrix}.
\]
For the finite case, this construction is due to Kl\'ima and Pol\'ak
\cite{klimapolak10}.
\item $\mathscr{D}=\Vect{\mathds{K}}$: given $\mathds{K}$-algebras $M$ and $N$, the $\mathds{K}$-algebra
$M\diamond N$ is carried by the vector space $M\times (M\t N)\times
N$ with multiplication and unit
\[
\begin{pmatrix}
m & z\\
0 & n
\end{pmatrix}
\begin{pmatrix}
m' & z'\\
0 & n'
\end{pmatrix}
=
\begin{pmatrix}
mm'~~ & mz'+zn'\\
0 & nn'
\end{pmatrix}
\quad\text{and}\quad
\begin{pmatrix}
1~ & ~0\t 0\\
0~ & ~1
\end{pmatrix},
\]
where $mz' = (mm_0)\t n_0$ for $z'=m_0\t n_0$, and extending via bilinearity
for
arbitrary $z$; similarly for $zn'$.
This construction is due to Reutenauer \cite{reu80}.
\item $\mathscr{D}=\Mod{\S}$: given $\S$-algebras $M$ and $N$, the $\S$-algebra
$M\diamond N$ is carried by the $\S$-module $M\times (M\ast N)\times N$ with
multiplication and unit
\[
\begin{pmatrix}
m & z\\
0 & n
\end{pmatrix}
\begin{pmatrix}
m' & z'\\
0 & n'
\end{pmatrix}
=
\begin{pmatrix}
mm'~~ & mz'+zn'\\
0 & nn'
\end{pmatrix}
\quad\text{and}\quad
\begin{pmatrix}
1~ & ~0\ast 0\\
0~ & ~1
\end{pmatrix},
\]
where $mz' = (mm_0)\ast n_0$ for $z'=m_0\ast n_0$, and similarly for $zn'$.
This example specializes to (3) and (4) by taking $\S=\{0,1\}$ and $\S=\mathds{K}$, respectively,
but appears to be new construction for other semirings $\S$.
\end{enumerate}
\end{expl}
The following theorem gives the key property of $M\diamond N$.
\begin{theorem}\label{thm:schurec}
Let $K,L\colon \Sigma^*\to S$ be languages recognized by $M$ and $N$,
respectively. Then
$M\diamond N$
recognizes the languages $K$, $L$ and $K a
L$ for all $a\in\Sigma$.
\end{theorem}
Next, we aim to show that $M\diamond
N$ is the ``smallest''
$\mathscr{D}$-monoid satisfying the statement of the above theorem. This
requires
further assumptions
on
our setting.
\begin{notation}
Recall from \eqref{eq:s1s2} the morphism $p\ast q\colon M\ast N\to \S$. We denote
its adjoint transpose w.r.t. the adjunction $F\dashv U$ by $\overline{p\ast
q}\colon F(M\ast N)\to \S$.
\end{notation}
\begin{assumptions}\label{asm:dual}
From now on, suppose that:
\begin{enumerate}[(i)]
\item $\mathscr{D}$ is locally finite, i.e. every finitely generated algebra of $\mathscr{D}$ is
finite.
\item Epimorphisms in $\mathscr{D}$ and $\Mod{\S}$ are surjective.
\item $\mathscr{D}(M,\S)$, $\mathscr{D}(N,\S)$, and
$\{\,U(\overline{p\ast q})\colon UF(M\ast N)\to \S\,\}_{p\colon M\to
\S,\,q\colon N\to \S}$
are separating families of morphisms in $\mathscr{D}$.
\item There is a locally finite variety $\mathscr{C}$ of algebras such that
the full subcategories $\mathscr{C}_f$ and $\mathscr{D}_f$ on \emph{finite} algebras are dually
equivalent. We denote the equivalence functor by $E\colon
\mathscr{D}_f^{op}\simeq\mathscr{C}_f$.
\item The semiring $\S$ is finite, and $E(\S)\cong \mathbf{1}_\mathscr{C}$.
\end{enumerate}
\end{assumptions}
Let us indicate the intuition behind our assumptions.
First, (i) and (ii) imply that $M\diamond N$
is finite
if $M$
and $N$ are. This is important, as one is usually interested in
language
recognition by \emph{finite} $\mathscr{D}$-monoids. (iii) expresses that the semiring $\S$
has enough structure to separate elements of $M$, $N$ and $UF(M\ast N)$,
the three components of the Sch\"utzenberger product $M\diamond N$, by suitable
morphisms into $\S$. This
technical condition on $\S$ is the crucial
ingredient
for proving the ``smallness'' of $M\diamond N$ (Theorem
\ref{thm:universalprop}). Finally, the variety $\mathscr{C}$ in (iv) and (v)
will be used to determine, via duality, the algebraic operations to express
languages
recognized by $M\diamond N$ in terms of languages recognized by $M$ and $N$
(Theorem \ref{thm:schurec2}).
\begin{expl}\label{ex:predualcats} The categories and adjunctions
of Example \ref{ex:adjunctions}(1)-(4) satisfy our assumptions. Here we
briefly
sketch the dualities; see \cite{ammu14,ammu15} for details.
\begin{enumerate}[(1)]
\item For $\mathscr{D}=\mathbf{Set}$, choose $\mathscr{C}=\mathbf{BA}$ (boolean algebras). Stone duality
\cite{Johnstone1982} gives a dual equivalence $E\colon \mathbf{Set}_f^{op}\simeq \mathbf{BA}_f$
mapping a finite set to the boolean
algebra of all subsets.
\item For $\mathscr{D}=\mathbf{Pos}$, choose $\mathscr{C} = \mathbf{DL}$ (distributive
lattices with $0,1$). Birkhoff duality \cite{birkhoff37} gives a dual
equivalence $E\colon \mathbf{Pos}_f^{op}\simeq\mathbf{DL}_f$ mapping a finite poset to
the
lattice of all down-sets.
\item For $\mathscr{D}=\mathbf{JSL}$, choose $\mathscr{C} = \mathbf{JSL}$. The
dual equivalence $E\colon \mathbf{JSL}_f^{op}\simeq\mathbf{JSL}_f$ maps a finite semilattice
$(X,\vee)$ to its opposite
semilattice $(X,\wedge)$, see \cite{Johnstone1982}.
\item For $\mathscr{D}=\Vect{\mathds{K}}$, $\mathds{K}$ a \emph{finite} field, choose
$\mathscr{C}=\Vect{\mathds{K}}$. The dual equivalence $E\colon
\Vect{\mathds{K}}_f\simeq\Vect{\mathds{K}}_f^{op}$
maps
a space $X$ to its dual space $X^*=\hom(X,\mathds{K})$.
\end{enumerate}
\end{expl}
\begin{notation}\label{not:lmnf}
For any $\mathscr{D}$-monoid morphism $f\colon \Psi\Sigma^* \to M\diamond N$, put
\vspace{-0.2cm}
\[
\L_{M,N}(f) := \{\,K, L, KaL \mid \text{$a\in\Sigma$, $\pi_M\c f$ recognizes
$K$, $\pi_N\c
f$ recognizes $L$}\,\}
\]
\end{notation}
\begin{theorem}\label{thm:universalprop}
Let $f\colon \Psi\Sigma^*\to M\diamond N$ and $e\colon \Psi\Sigma^*\twoheadrightarrow
P$ be two $\mathscr{D}$-monoid morphisms. If $e$ is surjective and recognizes all
languages in $\L_{M,N}(f)$, then there exists a
unique
$\mathscr{D}$-monoid morphism $h\colon P\to M\diamond N$ with $h\c e = f$.
\end{theorem}
Using our duality framework, this theorem can be rephrased in terms of
language operations.
Recall that $E(\S)\cong\mathbf{1}_\mathscr{C}$ by Assumption \ref{asm:dual}(v). Putting $O_\mathscr{C}
\coloneqq
E(\mathbf{1}_\mathscr{D})$, we obtain a bijection
$i\colon S \cong \mathscr{D}(\mathbf{1}_\mathscr{D}, \S) \cong \mathscr{C}(E(\S),E(\mathbf{1}_\mathscr{D})) \cong \mathscr{C}(\mathbf{1}_\mathscr{C},
O_\mathscr{C}) \cong
\under{O_\mathscr{C}}$.
\begin{defn}
For any $n$-ary operation symbol $\gamma$ in the signature of $\mathscr{C}$ and
languages $L_1,\ldots, L_n\colon \Sigma^*\to S$, the language
$\overline\gamma(L_1,\ldots,L_n)\colon \Sigma^*\to S$ is given by
$\overline\gamma(L_1,\ldots,L_n)(u) \coloneqq i^{-1}(\,\gamma^{O_\mathscr{C}}(\,i(L_1u), \ldots, i(L_n u)\,)\,)$.
The operations $\overline\gamma$ are called the \emph{$\mathscr{C}$-algebraic operations} on
the set of languages
over $\Sigma$.
\end{defn}
\begin{expl} $O_\mathbf{BA}\cong\{0,1\}$ is the two-element boolean algebra, and the
$\mathbf{BA}$-algebraic operations are precisely the boolean operations (union,
intersection, complement, $\emptyset$, $\Sigma^*$) on languages. For example,
the operation symbol $\vee$ induces the language operation $(K\,\overline\vee\,L)(u) =
K(u)\vee L(v)$ corresponding to the union of languages. Similarly, for $\mathscr{C}=\mathbf{DL}$ we get
union, intersection, $\emptyset$, $\Sigma^*$, for $\mathscr{C}=\mathbf{JSL}$ we get union and
$\emptyset$, and for $\mathscr{C}=\Vect{\mathds{K}}$ we get sum, scalar product and $\emptyset$.
\end{expl}
All our constructions and results so far apply to arbitrary $\mathscr{D}$-monoids.
However, in the following theorem we need to restrict to \emph{finite}
$\mathscr{D}$-monoids. Recall that the \emph{derivatives} of a language
$L\colon\Sigma^*\to S$
are the languages $a^{-1}L,\,L{a^{-1}}\colon \Sigma^*\to S$ (where
$a\in\Sigma$)
defined by
$(a^{-1}L)(u) = L(au)$ and $(La^{-1})(u) = L(ua)$.
\begin{theorem}\label{thm:schurec2}
Let $M$ and $N$ be finite $\mathscr{D}$-monoids and $f\colon \Psi\Sigma^*\to M\diamond
N$ be a $\mathscr{D}$-monoid morphism. Then every
language
recognized by $f$ lies in the closure of $\L_{M,N}(f)$
under the $\mathscr{C}$-algebraic operations and derivatives.
\end{theorem}
Our proof uses the \emph{Local Variety Theorem} of \cite{ammu14}: for any
finite set
$\mathcal{V}$ of recognizable languages closed under $\mathscr{C}$-algebraic
operations and derivatives, there is a finite $\mathscr{D}$-monoid recognizing
precisely the languages of $\mathcal{V}$.
Coincidentally, for each of our categories of Example
\ref{ex:adjunctions}(1)-(4) it suffices to take the closure of
$\L_{M,N}(f)$
under $\mathscr{C}$-algebraic operations, as this set is already derivative-closed. For
example, for $\mathscr{C}=\Vect{\mathds{K}}$ we have
$a^{-1}(KaL)=(a^{-1}K)aL+K(\epsilon)L$, i.e. $a^{-1}(KaL)$ is a linear
combination of
languages in $\L_{M,N}(f)$ and thus lies in the closure of $\L_{M,N}(f)$
under $\Vect{\mathds{K}}$-operations. For $\mathscr{D}=\mathbf{Set}$, $\mathbf{Pos}$ and $\mathbf{JSL}$, Theorem
\ref{thm:schurec2}
then gives
\begin{corollary}[Reutenauer \cite{reutenauer79}, Pin \cite{pin03}, Kl\'ima and Pol\'ak \cite{klimapolak10}]
Let $M$ and $N$ be finite monoids [ordered monoids, idempotent semirings].
Then any language recognized by the Sch\"utzenberger product $M\diamond N$ is
a boolean combination [positive boolean combination, finite union] of
languages of the form $K$, $L$ and $KaL$, where $K$ is recognized by $M$, $L$
is recognized by $N$, and $a\in\Sigma$.
\end{corollary}
For $\mathscr{D}=\Vect{\mathds{K}}$, we obtain a new result for formal power series:
\begin{corollary}
Let $M$ and $N$ be finite algebras over a finite field $\mathds{K}$. Then any language
recognized by
$M\diamond N$ is a linear combination of power series of the form $K$,
$L$ and $KaL$, where $K$ is recognized by $M$, $L$ is recognized by $N$, and
$a\in\Sigma$.
\end{corollary}
\section{Conclusions and Future Work}
We presented a uniform approach to Sch\"utzenberger products
for various algebraic structures. Our categorical framework encompasses all
known instances of Sch\"utzenberger products in the
setting
of regular
languages. Two related constructions are the
Sch\"utzenberger products for \emph{$\omega$-semigroups} \cite{carton93}
(dealing with
$\infty$-languages), and for \emph{boolean spaces with
internal
monoids} \cite{gpr16} (dealing with non-regular
languages). Neither of
these structures are monoids in the categorical sense, and thus are not
covered by our present setting. The use of monads as in
\cite{boj15,camu16,uacm16} might pave the way to extending the scope of our
work.
Since our main focus in the present paper was to establish the categorical
setting, we restricted to \emph{binary} Sch\"utzenberger products $M\diamond
N$. For (ordered) monoids and semirings, a non-trivial $n$-ary generalization
of the Sch\"utzenberger product
is known \cite{straubing81,pin03,polak01}, and we aim to adapt our results to arbitrary $n$.
|
\section{Introduction and main results}
One way to create new projective varieties from a given variety $X \subset \P^{r}$ is to intersect with linear spaces. It is only natural to want to understand the collection of varieties obtained in this way.
There are different approaches to assigning ``moduli'' to the linear sections of $X$. Algebraically, we may assign a point in an appropriate Hilbert scheme $H$. This works locally as we perturb the linear section, but in order to produce a globally defined variation, we must assign the corresponding point in the quotient of the Hilbert scheme by the action of $\PGL$. The usual technical issues with GIT emerge from this approach, but ultimately one can construct an algebraic moduli map:
\begin{align*}
\mu_{m}: \G\left( m,r \right) \dashrightarrow H / \PGL
\end{align*}
which, to a general $m$-plane $\Lambda$ assigns the variety $[X \cap \Lambda]$.
Over the complex numbers, the Hodge structure on the middle cohomology of a linear section $X \cap \Lambda$ provides another way of attaching moduli to the section. Again, under ideal conditions, one gets a period map:
\begin{align*}
\pi_{m}: \G\left( m,r \right) \dashrightarrow {\mathcal D}/\Gamma
\end{align*}
In either case, whether algebraic or analytic, we call these maps {\sl moduli maps}. Three questions of particular interest are:
\begin{enumerate}
\item What is the dimension of the image of a moduli map?
\item When is a moduli map generically injective?
\item If a moduli map is generically finite, what is its degree?
\end{enumerate}
These questions, especially the first two, are certainly not new, and have held the attention of many authors. The first question is infinitesimal in nature -- one has to compute the rank of the differential of the moduli map. The second question is global in nature, and the third question is enumerative.
Although the local and global analyses of moduli maps would seem mostly disconnected, the seminal work of Donagi \cite{donagi} shows that the infinitesimal variation of Hodge structure (IVHS) of a variety can, in many situations, recover the variety itself. From the perspective of varying linear slices, we may recast Donagi's result as proving the generic injectivity of the moduli map for hyperplane sections of the $d$-uple Veronese varieties $X \subset \P^{r}$. In the complex analytic setting, such generic injectivity results are called {\sl weak Torelli} or {\sl generic Torelli} theorems.
In this paper, we prove a generic Torelli theorem for the family of $m$-plane slices of a generic hypersurface $X$. Numerous generic Torelli statements exist in the literature, including results for complete intersections in \cite{konno} and \cite{terasoma}, but the problem of generic Torelli for linear sections of a fixed hypersurface seems to have fallen through the cracks.
Question (1) is easy to answer in our setting -- a simple first-order calculation left to the reader shows that the images of all moduli maps are as large as possible. We note, however, that this question is easy only because of the nebulous requirement that $X$ be ``generic''. Once we try to qualify what ``generic'' means, question (1) becomes much more challenging. For example, after considerable effort, in \cite{joemazurpanda} it is shown (among other things) that if we let $X$ be an arbitrary {\sl smooth} hypersurface of degree $d$, then, provided $r$ is very very large, the moduli maps for linear slices has image as large as possible.
On the other extreme, an interesting result of Beauville \cite{beauville} characterizes the degree $d$ smooth hypersurfaces $X$ for which the moduli map of hyperplane sections is constant -- this only happens when $d=2$, or when $X$ is a Fermat hypersurface and $d-1$ is a power of the characteristic of the (algebraically closed) ground field.
We let $\Hyp(d,m)$ denote the moduli space of degree $d$ hypersurfaces in $\P^{m}$, modulo projective equivalence.
\begin{theorem}
\label{theorem:main}
Assume $d > 2r + 1$, and let $X \subset \P^{r}$ be a general hypersurface of degree $d$. Then the moduli maps $\mu_{m} : \G\left( m,r \right) \dashrightarrow \Hyp(d,m)$ are generically injective for all $m$.
\end{theorem}
The generic Torelli theorems mentioned earlier are proven using variants of Donagi's approach. Our technique for proving \autoref{theorem:main} is completely different; we deduce the result by a simple reduction to the case $m=1$ of lines.
In the case of lines, we study the moduli map $\mu_{1}$ by resolving its indeterminacy (at some special points), and then computing its degree at special points in the blown up domain. The points we consider are deep in the locus of indeterminacy, so the details of the resolution require special attention. We are able to carry out the resolution of indeterminacy because of an elementary, but fundamental {\sl versality} result (\autoref{corollary:versalkincident}) which essentially allows us to write the moduli map explicitly. This versality result is interesting in its own right, and should be true in a much larger context. Such a generalization will be the subject of future work.
Our analysis yields complete solutions to questions (2) and (3) in the case of lines. In order to answer question (2), we study the monodromy on the set of $k$-incident lines to $X$, where $k = d-(2r-2)$. These are lines which meet $X$ at $k$ points, i.e. {\sl as infrequently as possible}.
The set of $k$-incident lines to $X$ divides itself into groups according to the multiplicity vector $(m_{1}, \dots, m_{k})$, $\sum_{i=1}^{k}m_{i}=d$, describing the way the lines meet $X$. Clearly, monodromy preserves the multiplicity vector of a $k$-incident line.
We prove the following generalization of a monodromy result on flex lines of plane curves found in \cite{joe:galois}:
\begin{theorem}
\label{theorem:monodromymain}
Let $X$ be a general hypersurface of degree $d \geq 2r$, let $k = d - (2r-2)$, and let $(m_{1}, \dots, m_{k})$ be a multiplicity vector with at least two distinct entries.
Then the monodromy group of the set of $k$-incident lines to $X$ with multiplicity vector $(m_{1}, \dots, m_{k})$ is the full symmetric group.
\end{theorem}
The key to answering question (3) for the moduli map $\mu_{1}$ is to analyze the {\sl tri-incident} lines of $X$ -- these are lines meeting $X$ at exactly three points.
\begin{theorem}
\label{theorem:main2}
Let $X \subset \P^{r}$ be a general hypersurface of degree $d=2r+1$. Then
\begin{align*}
\deg \mu_{1} = 2\sum_{a \geq b >1}^{}n_{a,b,1} + 4n_{d-2,1,1}
\end{align*}
where $n_{i,j,k}$ is the number of tri-incident lines to $X$ with intersection multiplicities $\left( i,j,k \right)$.
\end{theorem}
\begin{remark}
\label{remark:periodmap}
\autoref{theorem:main}, along with Donagi's theorem on the generic injectivitiy of the period map $\Hyp(d,m) \dashrightarrow {\mathcal D}/\Gamma$, implies that the composite period map $\pi_{m}: \G\left( m,r \right) \dashrightarrow {\mathcal D}/\Gamma$ is injective for a generic degree $d$ hypersurface $X$ with $d > 2r+1$.
\end{remark}
\begin{remark}
\label{remark:notexhausted}
\autoref{theorem:main} does not exhaust all cases where $\dm \G\left( m,r \right) < \dm \Hyp(d,m)$. We also remark that a similar argument as found in the proof of \autoref{theorem:main} shows that knowledge of generic injectivity for $m$-planes implies the same for larger planes.
\end{remark}
\begin{remark}
\label{remark:genericityassumption}
Clearly, some sort of genericity condition on $X$ is required for an affirmative answer to question (2): Consider $X$ a Fermat hypersurface.
Interestingly, the example of the Fermat quintic in $\P^{2}$ shows that a genericity assumption stronger than ``smooth'' is required for a uniform answer to question (3). In \cite{cadmanlaza} , the degree of $\mu_{1}$ for a general quintic is $420$ (in agreement with \autoref{theorem:main2}), while for the Fermat quintic the degree drops to $150$.
\end{remark}
\subsection{Acknowledgements} I thank Francesco Cavazzani, Anand Deopurkar, Igor Dolgachev, Joe Harris, and Claire Voisin for useful conversations.
\subsection{Conventions} We work over an algebraically closed field of characteristic zero. If $F$ is a sheaf, we let $\Gamma (F)$ denote the space of global sections of $F$.
\section{Versality}
\subsection{Local models}
Throughout this section $X \subset \P^{r}$ will be a smooth hypersurface not containing any lines.
\begin{definition}
\label{definition:tautological}
We let $\pi : \P \to \G\left( 1,r \right)$ denote the tautological $\P^{1}$-bundle.
\end{definition}
\begin{definition}
\label{definition:incidencecorrespondence}
The {\sl point-line incidence correspondence for $X$} is the variety $\Sigma \subset \P$ defined by
\begin{align*}
\Sigma := \left\{ (x, [\ell]) \in X \times \G\left( 1,r \right) \mid x \in \ell \right\}
\end{align*}
We let $\varphi, \pi$ be the projections of $\Sigma$ to the first and second factors.
\end{definition}
\begin{lemma}
\label{lemma:sigmasmooth}
$\Sigma$ is a smooth variety.
\end{lemma}
\begin{proof}
The projection $\varphi : \Sigma \to X$ is a projective bundle, with fibers isomorphic to $\P^{r-1}$.
\end{proof}
Clearly, since $X$ is assumed to contain no lines, the map $\pi : \Sigma \to \G\left( 1,r \right)$ is a finite branched cover of degree $d$.
\begin{definition}
\label{definition:Wn}
We let $W_{n}$, $(n > 1)$, denote the pair $(Y_{n},\P^{1} \times \A^{n-1} \to (\A^{n-1},0))$, with $Y_{n} \subset \P^{1} \times \A^{n-1}$ defined by the equation
\begin{align*}
s^{n} + a_{n-2}s^{n-2}t^{2} + a_{n-3}s^{n-3}t^{3} + \dots + a_{0}t^{n} = 0,
\end{align*}
where $a_{i}$ are coordinates on $\A^{n-1}$, and $[s:t]$ are homogeneous coordinates on $\P^{1}$.
\end{definition}
\begin{remark}
\label{remark:dehomogenize}
To simplify notation, we will often write the affine equation $$z^{n} + a_{n-2}z^{n-2} + \dots a_{0}= 0$$ for the equation defining $W_{n}$. $W_{n}$ is often referred to as the ``mini-versal unfolding'' of the singularity $z^{n} = 0$.
\end{remark}
\begin{proposition}
\label{proposition:versality}
Let $B$ be a smooth variety, $b \in B$ a marked point. Suppose $Z \subset \P^{1} \times B$ is a divisor such that $Z_{b} \subset \P^{1} \times \left\{ b \right\}$ is a scheme isomorphic to $s^{n} = 0$. Then there is an \'etale neighborhood $U$ of $b$ and a map $\mu : \left( U,b \right) \to \left( \A^{n-1},0 \right)$ such that $Z_{U} \subset \P^{1}\times U$ is isomorphic to the pullback of $W_{n}$ under $\mu$.
\end{proposition}
\begin{proof}
Since $B$ is smooth, it is \'etale locally isomorphic to $\A^{m}.$ So we may reduce to the case $\left( B,b \right) = \left( \A^{m},0 \right)$.
After passing to an open set, the set $Z$ is defined by a single equation $$s^{n} + u_{n-1}s^{n-1}t + \dots + u_{0}t^{n}=0$$ where $u_{i} \in \fm_{0} \subset \Gamma (O_{\A^{m}})$. (Here $\fm_{0}$ is the maximal ideal of the origin.) We may ``complete the $n$-th power'' and find an isomorphic $Z' \subset \P^{1} \times \A^{m}$ defined by a similar equation, but with vanishing $u_{n-1}$ term.
The $n-1$ functions $u_{i} \in \fm_{0} \subset \Gamma \left( O_{\A^{m}} \right), i = 0, \dots n-2$ define a map $\left( \A^{m}, 0 \right) \to \left( \A^{n-1},0 \right)$ with the desired property.
\end{proof}
\begin{proposition}
\label{proposition:versalitymultiple}
Let $(B.b)$ be as in the previous proposition, but now suppose $Z \subset \P^{1} \times B$ is a divisor whose restriction to $Z_{b} \subset \P^{1} \times \left\{ b \right\}$ is of the form $n_{1}p_{1} + \dots + n_{i}p_{i}$ with each $n_{j} \geq 1$.
Then there exists an \'etale open neighborhood $U$ of $b$ such that $Z_{U}$ is a disjoint union $ \coprod_{j}Z_{j} $, and furthermore there exists a map $(U,b) \to \left( \prod_{j}\A^{n_{j}-1},0 \right)$ such that $Z_{j}$ is isomorphic to the pullback of $W_{n_{j}}$ under the composite $(U,b) \to \left(\prod_{j}\A^{n_{j}-1},0 \right) \to \left(\A^{n_{j}-1},0\right)$.
\end{proposition}
\begin{proof}
The argument is completely similar to the proof of \autoref{proposition:versality}.
\end{proof}
\begin{remark}
\label{remark:nonunique}
The maps to $W_{n}$ in these propositions are not unique, however it is easy to show that they induce a unique map on the tangent space of $B$ at $b$.
\end{remark}
\begin{definition}
\label{definition:versal}
Let $(B,b)$ be a smooth variety with marked point, and let $Z \subset \P^{1} \times B$ be a closed subset such that $Z_{b}$ is as in \autoref{proposition:versalitymultiple}. Then $Z \subset \P^{1} \times B \to B$ is {\sl versal at $b$} if any (and hence all) of the maps $(U,b) \to \left(\prod_{j}\A^{n_{j}-1},0 \right)$ is a local submersion, i.e. the map on tangent spaces is surjective.
\end{definition}
\begin{definition}
\label{definition:mini-versal}
Let $(B,b)$ be a smooth variety with marked point, and let $Z \subset \P^{1} \times B$ be a closed subset such that $Z_{b}$ is as in \autoref{proposition:versalitymultiple}. Then $Z \subset \P^{1} \times B$ is {\sl mini-versal at $b$} if any (and hence all) of the maps $(U,b) \to \left(\prod_{j}\A^{n_{j}-1},0 \right)$ is a local isomorphism.
\end{definition}
\subsection{Lines meeting $X$ infrequently}
\begin{notation} \label{notation1}
Let $\sum_{i}^{k}m_{i} = d$, $\sum_{i=1}^{j}n_{i}=d$, and let $\P^{N}$ denote the projective space of degree $d$ hypersurfaces in $\P^{r}$. We define the following three types of incidence correspondences:
\begin{enumerate}
\item \begin{align*}
I_{k} & \left( m_{1}, \dots, m_{k} \right) =\\
& \overline{\left\{ (X,\ell) \mid X \cap \ell = \sum_{i}m_{i}p_{i}, \, \, \text{for distinct $p_{i} \in \ell$} \right\}} \subset \P^{N} \times \G\left( 1,r \right)
\end{align*}
\item \begin{align*}
I'_{k} & \left( m_{1}, \dots, m_{k} \right) =\\
& \overline{\left\{ (X,\ell,p_{1}, \dots, p_{k}) \mid X \cap \ell = \sum_{i}m_{i}p_{i} \right\}} \subset \P^{N} \times \P \times_{\G\left( 1,r \right)} \dots \times_{\G\left( 1,r \right)}\P
\end{align*}
\item \begin{align*}
& I_{k,j}'' \left( m_{1},\dots, m_{k} \mid n_{1}, \dots, n_{j}\right) =\\
& \overline{\left\{ (X, \ell, p_{1}, \dots, p_{k}, \ell', q_{1}, \dots, q_{j}) \mid X\cap \ell = \sum_{i}m_{i}p_{i}, \, \, \, X \cap \ell' = \sum_{i}n_{i}q_{i} \right\}}
\end{align*}
\end{enumerate}
The closures are taken of the sets where the objects parametrized are all distinct (for points) and do not intersect (for lines).
\end{notation}
\subsubsection{Lemmas on correspondences}
\begin{lemma}
\label{lemma:Ikirreducible}
$I_{k}\left( m_{1}, \dots, m_{k} \right)$ is irreducible for all $k$.
\end{lemma}
\begin{proof}
Consider the correspondence $I_{k}'\left( m_{1}, \dots, m_{k} \right)$. The projection to the data $(\ell, p_{1}, \dots, p_{k})$ makes $I_{k}'$ into a projective space bundle over the parameter space of tuples $\left( \ell, p_{1}, \dots, p_{k} \right)$, and hence is irreducible.
$I_{k}'\left( m_{1}, \dots, m_{k} \right)$ dominates $I_{k}\left( m_{1}, \dots, m_{k} \right)$ by forgetting the data of the marked points $(p_{1}, \dots, p_{k})$. The lemma follows.
\end{proof}
\begin{lemma}
\label{lemma:2irreducible}
The varieties $I_{k,j}''\left( m_{1}, \dots, m_{k} \mid n_{1},\dots,n_{j} \right)$ are irreducible for all $k, j$, and all multiplicities $(m_{1}, \dots, m_{k}), ( n_{1}, \dots, n_{j} )$.
\end{lemma}
\begin{proof}
The general fiber of the projection
\begin{align*}
(X ,\ell, p_{1}, \dots, p_{k}, \ell', q_{1}, \dots, q_{j}) \mapsto (\ell, p_{1}, \dots p_{k}, \ell', q_{1}, \dots, q_{j})
\end{align*}
is a linear space in $\P^{N}$ with dimension independent of the pair of non-intersecting lines $\ell, \ell'$. This follows because the restriction map
\begin{align*}
\Gamma (O_{\P^{r}}(d)) \to \Gamma (O_{\ell \cup \ell'}(d))
\end{align*}
is surjective. The lemma follows at once.
\end{proof}
\begin{definition}
\label{definition:k-incident}
Let $X \subset \P^{r}$ be a hypersuface. A line $\ell$ is {\sl $k$-incident} to $X$ if the set-theoretic intersection $X\cap \ell$ consists of $k$ points.
\end{definition}
\begin{proposition}
\label{proposition:finitek-incident}
Assume $k =d - (2r-2)$ is positive. Then a general hypersurface $X \subset \P^{r}$ of degree $d$ has finitely many $k$-incident lines, and has no $k'$-incident lines for $k' < k$.
\end{proposition}
\begin{proof}
This is standard, and left to the reader.
\end{proof}
\subsubsection{First-order analysis near a $k$-incident line} Let $(X,\ell) \in I_{k}(m_{1},\dots,m_{k})$ be a general point. We will study the first order deformations of $X \cap \ell$, as we vary $\ell$.
Fix homogeneous coordinates $\left[ x_{0}, \dots , x_{r} \right]$ for $\P^{r}$ such that $\ell$ is given by $x_{0} = \dots = x_{r-2} = 0$. Let $X$ be given by the degree $d$ equation
\begin{align*}
F = P(x_{r-1}, x_{r}) + \sum_{j=0}^{d-2}x_{j}G_{j}
\end{align*}
where
\begin{align*}
P = \prod_{i=1}^{k}\left( x_{r-1} - \alpha_{i}x_{r} \right)^{m_{i}}.
\end{align*}
\begin{definition}
\label{definition:groupG}
Let $G \subset \PGL_{r+1}$ be the group of matrices of the form:
\begin{align*}
G = \begin{bmatrix}
1 & 0 & \dots & 0 & 0 \\
0 & 1 & \dots & 0 & 0 \\
\hdotsfor{5} \\
a_{r-1,0} & a_{r-1,1} & \dots & 1 & 0 \\
a_{r,0} & a_{r,1} & \dots & 0 & 1
\end{bmatrix}
\end{align*}
where $a_{i,j}$, $(i = r-1, r), (j = 0 , \dots , r-2)$ are any constants.
\end{definition}
\begin{remark}
\label{remark:opensetgrassman}
The orbit $G \cdot [\ell]$ is an open subset of $\G(1,r)$ isomorphic to $\A^{2r-2}$.
\end{remark}
\begin{definition}
\label{definition:familyG}
Let $Z \subset \ell \times G$ be the union
\begin{align*}
\bigcup_{g \in G} \left(\left(g \cdot X\right)\cap \ell,g\right).
\end{align*}
\end{definition}
\begin{lemma}
\label{lemma:ZisSigma}
The variety $Z \subset \ell \times G$ is isomorphic to the the variety $\Sigma \subset \P$ above the open set $G \subset \G\left( 1,r \right)$.
\end{lemma}
\begin{proof}
This is clear.
\end{proof}
Our goal here is to understand the first-order infinitesimal neighborhood of the the family $Z \subset \ell \times G \to G$ near the point $\left[ \ell \right] \in G$.
\begin{definition}
\label{definition:gj}
We let $g_{j}$ be the polynomial in $x_{r-1},x_{r}$ which is the restriction of $G_{j}$ to $\ell$.
\end{definition}
\begin{lemma}
\label{lemma:firstorderG}
Let $\left\{ c_{i,j} \right\}, (i=r-1,r), (j=0,\dots,r-2)$ be the elements $\partial_{a_{i,j}} \in T_{\left[ \ell \right]}G$.
The first-order family $Z_{\varepsilon} \subset \ell \times T_{\left[ \ell \right]}G \to T_{\left[ \ell \right]}G$ is isomorphic to the zero locus of:
\begin{align*}
P\left( x_{r-1},x_{r} \right) + \sum_{j=0}^{r-2}\left( c_{r-1,j}x_{r-1}+c_{r,j}x_{r} \right)g_{j}.
\end{align*}
\end{lemma}
\begin{proof}
The first-order family parametrized by the Lie algebra of the group $G$ is given by applying the infinitesimal substitutions
\begin{align*}
x_{0} &\mapsto x_{0} + \varepsilon(c_{r-1,0}x_{r-1}+ c_{r,0}x_{r})\\
x_{1} &\mapsto x_{1} + \varepsilon(c_{r-1,1}x_{r-1}+c_{r,1}x_{r})\\
&\vdots\\
x_{r-1} &\mapsto x_{r-1}\\
x_{r} &\mapsto x_{r}
\end{align*}
to the equation $F$ defining $X$ and then restricting to $\ell$. The constants $\left\{ c_{i,j} \right\}$ vary freely, and represent coordinates on the Lie algebra of $G$. The end result is the family given in the statement of the lemma.
\end{proof}
\begin{remark}
\label{remark:affine}
It will usually be more convenient to use the affine notation
\begin{align*}
p(z) + \sum_{j=0}^{r-2}\left( c_{r-1,j}z+c_{r,j} \right)g_{j}(z).
\end{align*}
\end{remark}
Using the infinitesimal analysis above, we will now prove some lemmas which will allow us to understand the versality of the point-line correspondence $\Sigma \to \G\left( 1,r \right)$ near a $k$-incident line $\ell$, where $k = d -(2r-2)$.
\subsection{Some lemmas on multiplication of polynomials} \label{multiplication}
\begin{notation}\label{notation2}
Throughout \autoref{multiplication}, we will keep the following notation.
\begin{enumerate}
\item Suppose $n_{1}, \dots n_{k}$ are positive integers such that $\sum_{i=1}^{k}n_{i}=d$ and $\sum_{i=1}^{k}(n_{i}-1)=2r-2$ for some positive $r$.
\item Let $P_{m}$ denote the vector space of polynomials in one variable $z$ with degree $\leq m$.
\item Finally, let $\alpha_{1}, \dots, \alpha_{k} \in \A^{1} = \Spec k[z]$ be distinct points.
\end{enumerate}
\end{notation}
\subsubsection{First Lemma}
\begin{lemma}
\label{lemma:surjection}
There exists an $r-1$-dimensional subspace $V_{r-1} \subset P_{d-1}$ such that the multiplication-then-restriction map
\begin{align*}
\rho: P_{1} \otimes V_{r-1} \to \bigoplus_{i=1}^{k}k[z]/(z-\alpha_{i})^{n_{i}-1}
\end{align*}
is an isomorphism.
\end{lemma}
\begin{proof}
We will explicitly construct a vector space $V_{r-1}$ with the desired property.
Our construction is slightly different depending on the parities of the numbers $n_{i}$. So, for the convenience of the reader, let us first assume all $n_{i}$ are odd.
Then for each $i$, we define the polynomial
\begin{align*}
p_{i} := \prod_{j \neq i}(z-\alpha_{j})^{n_{j}-1}.
\end{align*}
Next, for each $i$ consider the vector space
\begin{align*}
A_{i} := \Span\left\{p_{i} \cdot 1,p_{i} \cdot (z-\alpha_{i})^{2}, p_{i}\cdot (z-\alpha_{i})^{4}, \dots,p_{i} \cdot (z-\alpha_{i})^{n_{i}-3} \right\}.
\end{align*}
Then it is immediately verified that
\begin{align*}
\rho(P_{1} \otimes A_{i}) = (0, \dots, k[z]/(z-\alpha_{i})^{n_{i}-1}, 0 ,\dots).
\end{align*}
Note that $\dim A_{i} = (n_{i}-1)/2$. By letting $V_{r-1} = A_{1}+ \dots+ A_{k}$, the lemma follows in this case.
Now suppose $n_{1}$ is even. Since $\sum_{}^{}(n_{i}-1)$ is even, we may assume without loss of generality that $n_{2}$ is also even.
Then we define the polynomials
\begin{align*}
q_{1} = (z-\alpha_{2})^{n_{2}-2}\prod_{j \neq 1,2}(z-\alpha_{j})^{n_{j}-1}\\
q_{2} = (z-\alpha_{1})^{n_{1}-2}\prod_{j \neq 1,2}(z-\alpha_{j})^{n_{j}-1}.
\end{align*}
Next set
\begin{align*}
B_{1,2}:= \Span\left\{ q_{1}, q_{1}(z-\alpha_{1})^{2}, q_{1}(z-\alpha_{1})^{4} \dots q_{1}(z-\alpha_{1})^{n_{1}-2} = q_{2}(z-\alpha_{2})^{n_{2}-2}, \dots, q_{2}(z-\alpha_{2})^{2}, q_{2} \right\}.
\end{align*}
Then it is straightforward to verify that
\begin{align*}
\rho(P_{1} \otimes B_{1,2}) = (k[z]/(z-\alpha_{1})^{n_{1}-1},k[z]/(z-\alpha_{2})^{n_{2}-1},0,0, \dots ).
\end{align*}
Note that $\dim B_{1,2} = (n_{1}+ n_{2}-2)/2$. By pairing up the remaining even $n_{i}$ and creating the analogous vector spaces $B_{i,j}$, and then for odd $n_{i}$ using the spaces $A_{i}$, one can check that
\begin{align*}
V_{r-1}:= \sum_{n_{i}\text{\,\,odd}} A_{i} + B_{1,2} + \dots
\end{align*}
has the desired property, i.e. that $\rho$ is an isomorphism.
\end{proof}
\subsubsection{Second Lemma}
\begin{lemma}
\label{lemma:transversetothing}
Maintain the setup before the previous lemma. Then there exists an $r-1$-dimensional subspace $V_{r-1} \subset P_{d-1}$ such that the multiplication-then-restriction map
\begin{align*}
\rho': P_{1}\otimes V_{r-1} \to k[z]/(z-\alpha_{1})^{n_{1}} \oplus_{i=2}^{k}k[z]/(z-\alpha_{i})^{n_{i}-1}
\end{align*}
has the following properties:
\begin{enumerate}
\item $\rho'$ is injective.
\item The image of $\rho'$ does not contain the element
\begin{align*}
w = ((z-\alpha_{1})^{n_{1}-1},0,0,\dots).
\end{align*}
\end{enumerate}
\end{lemma}
\begin{proof}
The spaces $V_{r-1}$ constructed in the proof of \autoref{lemma:surjection} satisfy these two properties. We leave the verification to the reader.
\end{proof}
\begin{remark}
\label{remark:openproperty}
The conditions in these two lemmas are open -- a general $V_{r-1} \subset P_{d-1}$ will have the desired properties in each lemma.
\end{remark}
\begin{theorem}
\label{theorem:versal}
Let $\sum_{i=1}^{k}m_{k}=d$, and suppose $\sum_{i}^{}m_{i}-1 = 2r-2$. Suppose $\left( X,\ell \right)$ is a general point in $I_{k}\left( m_{1}, \dots , m_{k} \right)$. Then the point-line correspondence $\pi: \Sigma \to \G\left( 1,r \right)$ is mini-versal near the $k$-incident line $[\ell] \in \G\left( 1,r \right)$.
\end{theorem}
\begin{proof}
The map $\rho$ in \autoref{lemma:surjection} is the differential of (any of) the induced maps
\begin{align*}
\left( U, \left[ \ell \right] \right) \to W_{m_{1}} \times \dots W_{m_{k}}
\end{align*}
from \autoref{proposition:versalitymultiple} (Here $U$ is an \'etale neighborhood of $\ell$). The theorem follows immediately from \autoref{proposition:versalitymultiple}.
\end{proof}
\begin{corollary}
\label{corollary:versalkincident}
Let $X \subset \P^{r}$ be a general degree $d \geq 2r-1$ hypersurface, and let $k = d-(2r-2)$. Then
\begin{align*}
\left( \Sigma \subset \P \to \G\left( 1,r \right) \right)
\end{align*}
is versal at every $k$-incident line $\left[ \ell \right] \in \G\left( 1,r \right)$.
\end{corollary}
\begin{proof}
When $k = d-2(r-1)$, the irreducible variety $I_{k}\left( m_{1}, \dots, m_{k} \right)$ maps generically finitely onto $\P^{N}$. The locus consisting of pairs $(X,\ell) \in I_{k}(m_{1}, \dots , m_{k})$ for which the point-line correspondence fails to be versal at $\ell$ is a strict (closed) subset of $I_{k}\left( m_{1}, \dots ,m_{k} \right)$ by \autoref{theorem:versal}. Since this is true for all multiplicity data $\left( m_{1}, \dots, m_{k} \right)$, we conclude that there is a Zariski open subset of $\P^{N}$ for which the corollary holds.
\end{proof}
\section{Maximal monodromy} The purpose of this section is to establish a basic result about the monodromy of lines meeting a general hypersurface at the least possible number of points.
\begin{theorem}
\label{theorem:maximalmonodromytheorem}
Suppose $k = d - (2r-2) \geq 2 $, and suppose $\sum_{i=1}^{k}m_{i}=d$ is a partition with at least two distinct parts.
Then the Galois group of the generically finite projection
\begin{align*}
\varphi : I_{k}\left( m_{1}, \dots, m_{k} \right) \to
\P^{N} \end{align*}
is the full symmetric group.
\end{theorem}
Our approach to proving \autoref{theorem:maximalmonodromytheorem} is standard. Let $G$ denote the Galois group of $\varphi$. We break the proof up into two standard steps, in increasing difficulty: \begin{enumerate}
\item Show $G$ is $2$-transitive.
\item Show $G$ contains a simple transposition.
\end{enumerate}
These two properties imply that $G$ is the full symmetric group. Step $1$ is easy and quite standard, whereas Step $2$ requires some care, and will occupy most of our attention.
\subsection{$2$-transitivity}
\begin{lemma}
\label{corollary:2transitive}
The monodromy group $G$ is $2$-transitive.
\end{lemma}
\begin{proof}
It suffices to show that the variety $$I_{k}(m_{1}, \dots, m_{k}) \times_{\P^{N}} I_{k}\left( m_{1}, \dots, m_{k} \right) \setminus \Delta$$ is irreducible. (Here $\Delta$ is the diagonal.)
By \autoref{lemma:2irreducible} we know that $I_{k,k}''\left( m_{1}, \dots, m_{k} \mid m_{1}, \dots, m_{k} \right)$ is irreducible. Since this correspondence dominates $I_{k}(m_{1}, \dots, m_{k}) \times_{\P^{N}} I_{k}\left( m_{1}, \dots, m_{k} \right) \setminus \Delta$, we conclude the claimed irreducibility, and hence the corollary.
\end{proof}
\begin{remark}
\label{remark:incidentlines}
We do not have to concern ourselves with incident pairs of lines $\ell, \ell'$ -- the reader can check that a general $X \in \P^{N}$ will not have an intersecting pair of $k$-incident lines by an elementary dimension count.
\end{remark}
\subsection{In pursuit of a simple transposition}
\begin{notation}\label{notation3}
Until otherwise specified, we adopt the following notation:
\begin{enumerate}
\item $(X, \ell) \in I_{k-1}\left( m_{1}+m_{2}, m_{3}, \dots, m_{k} \right)$ is a general point, with $k = d-(2r-2)\ \geq 2$. Furthermore, we assume $m_{1} \neq m_{2}$.
\item We let $\A^{m_{1} + m_{2}- 1}$ denote the base of the mini-versal deformation $W_{m_{1}+m_{2}}$.
\item We let $Z_{m_{1},m_{2}} \subset \A^{m_{1}+m_{2}-1}$ be the locus parametrizing divisors of the form $m_{1}p + m_{2}q$.
\item $\pi : \Sigma \to \G\left( 1,r \right)$ denotes the point-line correspondence for $X$.
\end{enumerate}
\end{notation}
\begin{lemma}
\label{lemma:Zcurve}
$Z_{m_{1},m_{2}}$ is a uni-branched curve with a multiplicity two singularity at the origin $0 \in \A^{m_{1}+m_{2}-1}$.
\end{lemma}
\begin{proof}
The family of deformations of $z^{m_{1}+m_{2}} = 0$ parametrized by the set $Z_{m_{1},m_{2}}$ is the image of the map
\begin{align*}
t \mapsto (z-m_{2}t)^{m_{1}}(z+m_{1}t)^{m_{2}}.
\end{align*}
By expanding this polynomial in terms of $z$, we obtain an expression
\begin{align*}
z^{m_{1}+m_{2}} + c_{2}t^{2}z^{m_{1}+m_{2}-2} + c_{3}t^{3}z^{m_{1}+m_{2}-3} + \dots + c_{m_{1}+m_{2}}t^{m_{1}+m_{2}}
\end{align*}
where, under the assumption that $m_{1} \neq m_{2}$, all coefficients $c_{j}$ are nonzero. The lemma follows.
\end{proof}
\begin{corollary}
\label{corollary:multiplicitytwointersect}
Let $D \subset \A^{m_{1}+m_{2}-1}$ be a divisor given by the vanishing of an equation $f(a_{2}, \dots, a_{m_{1}+m_{2}})=0$, where $f$ is such that $0 \in D$, and such that $\frac{\partial}{\partial a_{2}}f(0) \neq 0$. Further suppose $D$ does not contain $Z_{m_{1},m_{2}}$.
Then $$\length (D \cap Z_{m_{1}, m_{2}}) = 2.$$
\end{corollary}
\begin{proof}
This easily follows from the parametrization of $Z_{m_{1},m_{2}}$ found in the proof of \autoref{lemma:Zcurve}.
\end{proof}
\begin{proposition}
\label{proposition:transversefamily}
Let $(X,\ell) \in I_{k-1}(m_{1}+m_{2}, m_{3}, \dots, m_{k})$ be a general point, and let $\pi: \Sigma \to \G\left( 1,r \right)$ be the point-line correspondence.
Denote by $$\nu: (U, [\ell]) \to (\A^{m_{1}+m_{2}-1} \times \A^{m_{3}-1} \times \dots \A^{m_{k}-1},0)$$ a local map to the mini-versal deformation space where $U \subset \G\left( 1,r \right)$ is a sufficiently small \'etale neighborhood of $\left[ \ell \right]$. Then
\begin{enumerate}
\item The map $\nu$ is an immersion near $[\ell]$.
\item The image $\nu(U)$ is transverse to $Z_{m_{1},m_{2}} \subset (\A^{m_{1}+m_{2}-1}, 0, \dots, 0)$ in the sense of \autoref{corollary:multiplicitytwointersect}.
\end{enumerate}
\end{proposition}
\begin{proof}
The two statements are proved at the level of tangent spaces. They follow, respectively, by applying points $1$ and $2$ in \autoref{lemma:transversetothing}.
\end{proof}
\subsubsection{``No unintended consequences'' lemma} Now suppose
\begin{align*}
(X,\ell,\ell') \in I_{k-1,k}(m_{1}+m_{2}, m_{3}, \dots, m_{k} \mid m_{1}, \dots, m_{k})
\end{align*}
is a general point. The next lemma and corollary show that the condition of possessing the $k-1$-incident line $\ell$ has no effect on the $k$-incident line $\ell'$ of $X$. Hence, there are in general ``no unintended consequences elsewhere'' for having a $k-1$-incident line.
\begin{lemma}
\label{lemma:nocommunication}
The point-line correspondence
\begin{align*}
\pi : \Sigma \to \G\left( 1,r \right)
\end{align*}
is mini-versal near $[\ell'] \in \G\left( 1,r \right)$.
\end{lemma}
\begin{proof}
After using the $\PGL_{r+1}$-action, we may assume
\begin{align*}
\ell = [x_{0}:x_{1}]\\
\ell' = [x_{2}:x_{3}]
\end{align*}
and therefore, that the homogeneous ideal of the union $\ell \cup \ell'$ is
\begin{align*}
I_{\ell \cup \ell'} = (x_{0}x_{2}, x_{1}x_{2}, x_{0}x_{3}, x_{1}x_{3}, x_{4}, x_{5}, \dots x_{r}).
\end{align*}
Therefore, we may write the defining equation of $X$ in the form
\begin{align*}
F = p(x_{0},x_{1}) + q(x_{2},x_{3}) + \sum_{i=0,1 ; j=2,3}^{}x_{i}x_{j}A_{i,j} + \sum_{i=4}^{r}x_{i}G_{i}
\end{align*}
where $A_{i,j} \in \Gamma (O_{\P^{r}}(d-2))$, and $G_{i} \in \Gamma (O_{\P^{r}}(d-1))$.
We let $a_{i,j} = A_{i,j}|_{\ell'}$, and $g_{i} = G_{i}|_{\ell'}$.
As in the proof of \autoref{lemma:firstorderG}, we apply the infinitesimal substitutions
\begin{align*}
x_{0}& \mapsto x_{0}+\varepsilon \ell_{0}\\
x_{1}& \mapsto x_{1} + \varepsilon \ell_{1}\\
x_{2}& \mapsto x_{2}\\
x_{3}& \mapsto x_{3}\\
x_{4}& \mapsto x_{4} + \varepsilon \ell_{4}\\
\vdots& \\
x_{r} &\mapsto x_{r} + \varepsilon \ell_{r}
\end{align*}
and then mod out by the ideal $I_{\ell'} = (x_{0}, x_{1}, x_{4}, \dots, x_{r})$. (Here $\ell_{i}$ are linear forms in the variables/$x_{2}$ and $x_{3}$.)
The result is the infinitesimal family
\begin{align*}
f_{\varepsilon} = q(x_{2},x_{3}) + \varepsilon \ell_{0}(x_{2}a_{0,2} + x_{3}a_{0,3}) + \varepsilon \ell_{1}(x_{2}a_{1,2} + x_{3}a_{1,3}) + \sum_{i = 4}^{r}\varepsilon \ell_{i} g_{i}.
\end{align*}
Set $g_{0} := x_{2}a_{0,2}+x_{3}a_{0,3}$ and $g_{1} := x_{2}a_{1,2} + x_{3}a_{1,3}$. Then, general choices of the four polynomials $a_{i,j}$ yield general choices of $g_{0}$ and $g_{1}$.
These observations tell us that the first-order analysis of $\pi: \Sigma \to \G\left( 1,r \right)$ near $[\ell']$ is exactly the same as in the case for a general point $(X,\ell') \in I_{k}(m_{1}, \dots m_{k})$. In particular, since $g_{0}, g_{1}, g_{4}, \dots, g_{r}$ are taken to be generic, $\pi$ is mini-versal near $[\ell']$.
\end{proof}
\begin{corollary}
\label{corollary:nostrangecomm}
Let $(X,\ell) \in I_{k-1}(m_{1}+m_{2}, m_{3}, \dots, m_{k})$ be a general point. Then the point-line correspondence
\begin{align*}
\pi : \Sigma \to \G\left( 1,r \right)
\end{align*}
is mini-versal at all $k$-incident lines $[\ell'] \in \G\left( 1,r \right)$.
\end{corollary}
\begin{proof}
Indeed, the variety $I''_{k-1, k}(m_{1}+m_{2}, m_{3}, \dots, m_{k} \mid m_{1}, m_{2}, \dots, m_{k})$ is irreducible, and generically finite over $I_{k-1}(m_{1}+m_{2}, m_{3}, \dots, m_{k})$. The previous lemma shows that there is a Zariski open subset of $I_{k-1,k}$ with $\ell'$ versal. The corollary follows.
\end{proof}
\begin{lemma}
\label{lemma:onlyonespecialline}
Let $(X,\ell) \in I_{k-1}(n_{1}, \dots, n_{k-1})$ be a general point. Then $X$ does not possess any other $k-1$-incident lines.
\end{lemma}
\begin{proof}
This follows from a simple dimension count -- we leave it to the reader.
\end{proof}
\begin{theorem}
\label{proposition:simpletransposition}
The monodromy group $G$ contains a simple transposition.
\end{theorem}
\begin{proof}
Let $(X, \ell) \in I_{k-1}(m_{1}+m_{2}, \dots, m_{k})$ be a general point.
All $k$-incident lines of multiplicity $(m_{1}, \dots, m_{k})$ are simple, i.e. occur with multiplicity one, according to \autoref{corollary:nostrangecomm}. The line $[\ell]$ occurs as a multiplicity $2$ $k$-incident line of type $(m_{1}, \dots, m_{k})$, according to \autoref{corollary:multiplicitytwointersect}.
Therefore, above the point $[X] \in \P^{N}$, the correspondence $I_{k}(m_{1}, \dots, m_{k})$ has exactly one nonreduced point of length $2$, and all other points are reduced.
Since the curve $Z_{m_{1},m_{2}} \subset \A^{m_{1}+m_{2}-1}$ is irreducible, the two nearby $k$-incident lines (for a nearby hypersurface $X'$) limiting to $[\ell]$ experience monodromy.
Hence we obtain a simple transposition in the monodromy group $G$.
\end{proof}
\subsubsection{Moduli of lines meeting $X$ infrequently}
\begin{definition}
\label{definition:cross-ratios-map}
Let $k \geq 4$. We define $i_{(m_{1}, \dots, m_{k})}: I_{k}(m_{1}, \dots , m_{k}) \dashrightarrow \m_{0,k}/S_{k}$ to be the map sending $(X,\ell)$ to $\left[ \ell \cap_{set} X \subset \ell \right]$. Here , subscript ``set'' means ``set theoretic''.
\end{definition}
\begin{corollary}
\label{corollary:distinct-cross-ratios}
Assume $k = d- (2r-2) \geq 4$, and let $X$ be a general degree $d$ hypersurface. Then any two distinct $(X,\ell), (X,\ell') \in I_{k}(m_{1}, \dots, m_{k})$ have distinct images under $i_{(m_{1},\dots,m_{k})}$.
\end{corollary}
\begin{proof}
Let $J \subset I_{k}(m_{1}, \dots ,m_{k}) \times_{\P^{N}}I_{k}(m_{1}, \dots ,m_{k}) $ denote the closure of the locus of pairs $\left( (X,\ell), (X,\ell') \right)$ having equal images under $i_{(m_{1},\dots,m_{k})}$, and such that $\ell \cap \ell'$ is empty.
By \autoref{corollary:2transitive}, The fiber product $I_{k}(m_{1}, \dots ,m_{k}) \times_{\P^{N}}I_{k}(m_{1}, \dots ,m_{k}) \setminus \Delta $ is irreducible, where $\Delta$ denotes the diagonal.
It is easy to exhibit two lines $\ell, \ell' \subset \P^{r}$ and a degree $d$ hypersurface $X$ such that $\ell$ and $\ell'$ are $k$-incident with multiplicity $(m_{1}, \dots, m_{k})$ and such that $i_{(m_{1}, \dots, m_{k})}(\ell) \neq i_{(m_{1}, \dots, m_{k})}(\ell')$.
Therefore, $J$, being a closed set, must have dimension strictly smaller than this fiber product's dimension. But the latter is the dimension of $\P^{N}$, i.e. the locus $J$ projects to a strict subset of $\P^{N}$. This proves the corollary.
\end{proof}
\section{Extending $\mu_{1}$ across a $k$-incident line}
\begin{notation}\label{notation4}
Throughout this section:
\begin{enumerate}
\item $X \subset \P^{r}$ is a general hypersurface of degree $d$.
\item We set $k := d - (2r-2)$, and assume $k \geq 1$.
\item We let $\pi : \Sigma \to \G\left( 1,r \right)$ be the point-line correspondence for $X$.
\item We continue to let
\begin{align*}
\mu_{1}: \G\left( 1,r \right) \dashrightarrow \Hyp(d,1)
\end{align*}
denote the moduli map.
\item Finally, we let $\ell$ be a $k$-incident line of $X$ with multiplicity $\left( m_{1}, \dots, m_{k} \right)$.
\end{enumerate}
\end{notation}
\begin{remark}
\label{remark:assumefamiliarity}
We will assume that the reader has some familiarity with the moduli space $\barm_{0,d}$. We will review some important background as appropriate.
\end{remark}
Our goal is to understand the resolution of indeterminacy of $\mu_{1}$ near $\ell$.
The moduli space $\Hyp(d,1)$ is none other than the space $\m_{0,d}/S_{d}$. Unfortunately the latter moduli space is not fine, so we will first change settings to circumvent this technical inconvenience.
\begin{definition}
\label{definition:orderedSigma}
We define
\begin{align*}
\Sigma^{\left[ d \right]} := \overline{\Sigma \times_{\pi} \dots \times_{\pi}\Sigma \setminus \text{Diagonals}}
\end{align*}
We let $s: \Sigma^{[d]} \to \G\left( 1,r \right)$ denote the structural degree $d!$ map.
\end{definition}
\begin{proposition}
\label{proposition:smoothSigmad}
The scheme $\Sigma^{[d]}$ is smooth at any point $\hat{[\ell]} \in s^{-1}([\ell])$.
\end{proposition}
\begin{proof}
This is a consequence of \autoref{corollary:versalkincident}. Indeed, \'etale locally around every $[\ell] \in \G\left( 1,r \right)$, we may assume $\pi : \Sigma \to \G\left( 1,r \right)$ is isomorphic to products of maps of the form
\begin{align*}
\left\{ (z,a_{0}, \dots ,a_{n-2}) \mid z^{n} + a_{n-2}z^{n-2} + \dots a_{0}=0 \right\} \to (a_{0}, \dots ,a_{n-2})
\end{align*}
If $r_{1}, \dots r_{n}$ denote the roots of $z^{n} + a_{n-2}z^{n-2} + \dots + a_{0}$, then the variety $$\Spec k\left[ r_{1}, \dots r_{n}\right]/(r_{1}+\dots+ r_{n}=0)$$ maps to $\Spec k\left[ a_{n-2}, \dots , a_{0}, \right]$ by the elementary symmetric expressions in the $r_{i}$.
This, in turn, means that the scheme $\Sigma^{\left[ d \right]}$, above the point $\left[ \ell \right] \in \G\left( 1,r \right)$, is \'etale locally isomorphic to a product of affine spaces $\Spec k[r_{1}, \dots, r_{n}]/(r_{1} + \dots + r_{n})$, and in particular smooth.
\end{proof}
\begin{lemma}
\label{lemma:numberpreimages}
The set $s^{-1}([\ell])$ has $\frac{d!}{m_{1}!m_{2}! \dots m_{k}!}$ elements.
\end{lemma}
\begin{proof}
The group $S_{d}$ acts transitively on the set $s^{-1}([\ell])$. From the local description of $s$ over $[\ell]$ explained in the proof of the previous proposition, we see that the stabilizer of a point in $s^{-1}([\ell])$ is a copy of $S_{m_{1}} \times \dots \times S_{m_{k}} \subset S_{d}$.
\end{proof}
\begin{definition}
\label{definition:Sigmasmoothd}
We let $\pi^{\left[ d \right]} : \P \to \Sigma^{\left[ d \right]}$ denote the tautological $\P^{1}$-bundle, pulled back from $\G\left( 1,r \right)$. $\pi^{\left[ d \right]}$ has $d$ tautological sections $\sigma_{i}$, $i = 1, \dots d$.
We let $U \subset \Sigma^{\left[ d \right]}$ denote the open set over which none of the sections $\sigma_{i}$ intersect one another.
We obtain an induced moduli map which we write as:
\begin{align*}
\mu_{1}^{\left[ d \right]} : U \to \barm_{0,d}.
\end{align*}
Finally, we let $\hat{[\ell]}$ be any point in $\Sigma^{[d]}$ lying over $[\ell]$.
\end{definition}
\subsection{Resolution near $\hat{[\ell]}$}
Now that we have altered our setting, we begin the study of the resolution of indeterminacy of $\mu_{1}^{\left[ d \right]}$, viewed as a rational map
\begin{align*}
\mu_{1}^{\left[ d \right]}: \Sigma^{\left[ d \right]} \dashrightarrow \barm_{0,d}
\end{align*}
near the point $\hat{[\ell]}$.
\begin{definition}
\label{definition:resolutionZ}
We define $$Z \subset \Sigma^{\left[ d \right]} \times \barm_{0,d}$$ to be the closure of the graph of $m_{1}^{\left[ d \right]}$. Furthermore, we let $\alpha : Z \to \barm_{0,d}$ denote the projection to the second factor and we let $\beta: Z \to \Sigma^{\left[ d \right]}$ be the natural (birational) projection.
\end{definition}
\subsection{Blowup description of $\barm_{0,d}$} Recall that the moduli space $\barm_{0,d}$ can be obtained as a blow up
$$\kappa: \barm_{0,d} \to (\P^{1})^{d-3}$$ which, to a general point $(\P^{1}, s_{1}, \dots, s_{d}) \in \barm_{0,d}$ assigns the coordinates of $(s_{4}, \dots, s_{d})$ once we use the $\PGL_{2}$-action to send $s_{1}$ to $0$, $s_{2}$ to $1$ and $s_{3}$ to $\infty$.
\begin{notation}
\label{notation:simplex} Moving forward, we will use the following objects and notation:
\begin{enumerate}
\item Let $n \geq 1$ be any integer. We let $\Delta_{n}$ denote the scheme $$\Spec k\left[ r_{1}, \dots, r_{n} \right] / (r_{1}+ \dots +r_{n} = 0).$$
\item We fix $\alpha_{1}, \dots, \alpha_{k} \in \A^{1}$, and corresponding multiplicities $m_{1}, \dots, m_{k}$ such that $\sum_{i=1}^{k}m_{i} = d$.
\item We assume $k \geq 3$.
\item We define the {\sl tautological family} $F_{n}$ over $\Delta_{n}$ anchored at the point $\alpha_{i}$ to be the scheme
\begin{align*}
\prod_{j=1}^{n}\left( z - \alpha_{i} - r_{j} \right) = 0.
\end{align*}
Note: $F_{n}$ is a closed subscheme of $\Delta_{n} \times \A^{1}$.
\item Let $\varphi_{n}(\alpha_{i}): \Delta_{n} \to (\A^{1})^{n}$ be defined by the formula:
\begin{align*}
(r_{1}, \dots, r_{n}) \mapsto (\alpha_{i}+r_{1}, \dots, \alpha_{i}+r_{n}).
\end{align*}
\item We let
\begin{align*}
\varphi_{m_{1}, \dots, m_{k}}\left( \alpha_{1}, \dots, \alpha_{k} \right): \Delta_{m_{1}} \times \dots \times \Delta_{m_{k}} \to (\A^{1})^{d} \subset (\P^{1})^{d}
\end{align*}
denote the map $\varphi_{m_{1}}(\alpha_{1}) \times \dots \times \varphi_{m_{k}}(\alpha_{k})$.
\item Let $x \in (\P^{1})^{m}$ be any point. We let $D_{x}$ denote the vector space of first order deformations of $x$ induced by the diagonal $\PGL_{2}$-action on $(\P^{1})^{m}$.
\end{enumerate}
\end{notation}
\begin{remark}
\label{remark:ofcourse}
Of course, $\Delta_{n} \simeq \A^{n-1}$, but we want to emphasize the particular presentation of the coordinate ring.
\end{remark}
\begin{proposition}
\label{proposition:localisomorphism}
Under the assumptions above, the map
\begin{align*}
\varphi_{m_{1}, \dots, m_{k}}\left( \alpha_{1}, \dots, \alpha_{k} \right): \Delta_{m_{1}} \times \dots \Delta_{m_{k}} \to (\P^{1})^{d}
\end{align*}
has the following properties
\begin{enumerate}
\item The induced map on tangent spaces is injective at $0 \in \Delta_{m_{1}} \times \dots \times \Delta_{m_{k}}$.
\item The tangent space $T_{0}\Delta_{m_{1}} \times \dots \Delta_{m_{k}}$ and the space $D_{\varphi_{m_{1}, \dots, m_{k}}(\alpha_{1}, \dots, \alpha_{k})(0)}$ are linearly independent vector subspaces of the tangent space of $(\P^{1})^{d}$ at $\varphi_{m_{1}, \dots, m_{k}}(\alpha_{1}, \dots, \alpha_{k})(0)$.
\end{enumerate}
\end{proposition}
\begin{proof}
The first point is basically by definitions.
An element $v \in D_{\varphi_{m_{1}, \dots, m_{k}}(\alpha_{1}, \dots, \alpha_{k})(0)} $ is determined by its effect at the three distinct points $\alpha_{1}, \alpha_{2}$ and $\alpha_{3}$ in $\P^{1}$. Assume the first order deformation $v$ has the following effect:
\begin{align*}
\alpha_{1} \mapsto \alpha_{1} + \varepsilon v_{1}\\
\alpha_{2} \mapsto \alpha_{2} + \varepsilon v_{2}\\
\alpha_{3} \mapsto \alpha_{3} + \varepsilon v_{3}.
\end{align*}
Due to the defining equation $(r_{1}+ \dots + r_{n} = 0)$ for $\Delta_{n}$, In order for $v$ to be in $T_{0}\Delta_{m_{1}} \times \dots \times \Delta_{m_{k}}$, we must have:
\begin{align*}
m_{1}v_{1} = 0\\
m_{2}v_{2} = 0\\
m_{3}v_{3} = 0.
\end{align*}
But then all $v_{i}$ are zero, and hence $v$ is. This is what we wanted to show.
\end{proof}
\begin{corollary}
\label{corollary:localisomorphismontoimage}
The induced map
\begin{align*}
\bar{\varphi}: \Delta_{m_{1}} \times \dots \times \Delta_{m_{k}} \to (\P^{1})^{d} / \PGL_{2} \simeq (\P^{1})^{d-3}
\end{align*}
is a local immersion near the point $0 \in \Delta_{m_{1}} \times \dots \times \Delta_{m_{k}}$.
\end{corollary}
\begin{proof}
This follows from \autoref{proposition:localisomorphism}.
\end{proof}
\begin{corollary}
\label{corollary:mainresolutioncorollary}
There exists an \'etale neighborhood $V \subset \Sigma^{[d]}$ containing $\hat{[\ell]}$ such that
\begin{align*}
\alpha: \beta^{-1}(V) \to \barm_{0,d}
\end{align*}
is an isomorphism onto its image. In particular, $\alpha$ is unramified at all points of $\beta^{-1}(\hat{[\ell]})$.
\end{corollary}
\begin{proof}
This follows from \autoref{corollary:localisomorphismontoimage}, from the fact that $\Sigma^{[d]}$ is \'etale locally isomorphic to $\Delta_{m_{1}} \times \dots \times \Delta_{m_{k}}$ near $\hat{[\ell]}$, and the description of $\barm_{0,d}$ as a blowup $\kappa: \barm_{0,d} \to (\P^{1})^{d-3}$.
\end{proof}
\subsection{Stable reduction} For the reader's convenience, we briefly review the procedure of stable reduction for families of $d$-pointed rational curves.
\begin{notation} \label{notation5} The following notation will appear frequently in this section:
\begin{enumerate}
\item Let $B$ be a smooth curve, and $b \in B$ a specified point.
\item Let $\sigma_{1}, \dots, \sigma_{d}$ be distinct sections of the projection $\P^{1} \times B \to B$. We assume that away from $b$, the sections $\sigma_{i}$ are mutually disjoint. We let $\P^{1}_{b}$ denote $\P^{1} \times \left\{ b \right\}$.
\item Let
\begin{align*}
D := \left( \bigcup_{i}\sigma_{i} \right) \cap \P^{1}_{b}.
\end{align*}
and write $D = p_{1} + \dots p_{n} + a_{1}q_{1} + \dots a_{k-n}q_{k-n}$, with $a_{i}>1$ and all points $p_{j}, q_{j}$ distinct.
\item We denote stable genus $0$ curves in the usual way as $(P,s_{1}, \dots, s_{d})$. The nodes and marked points of $(P,s_{1}, \dots, s_{d})$ are called {\sl special} points.
\end{enumerate}
\end{notation}
\begin{proposition}
\label{proposition:stablereduction}
Assume $k \geq 3$. Then the stable replacement $\left( P, s_{1}, \dots , s_{d} \right) \in \barm_{0,d}$ of the pair $D \subset \P^{1}_{b}$ is a union of
\begin{enumerate}
\item $\P^{1}_{b}$ and
\item $T_{i}, i = 1, \dots k-n$, possibly-nodal rational tails with $T_{i} \cap \P^{1}_{b} = q_{i}.$
\end{enumerate}
The marked points $\left\{ s_{1}, \dots, s_{d} \right\}$ are distributed as follows: on $\P^{1}_{b}$ the points $p_{1}, \dots, p_{n}$ are marked, while each $T_{i}$ contains $a_{i}$ marked points.
\end{proposition}
\begin{proof}
We recall the stable reduction procedure, from which the proposition is immediately verified.
\begin{enumerate}
\item We view $C := \bigcup \sigma_{i} \subset \P^{1} \times B$ as a curve on a surface, with singularities $q_{i} \in \P^{1}_{b}$.
\item We resolve the singularities $q_{i} \in C$ by blowing up repeatedly until all branches of $C$ through $q_{i}$ are separated.
\item Since each blow up in the resolution process happens at smooth centers, the new fiber $P$ of the blown up surface over the point $b \in B$ is the union of the original $\P^{1}_{b}$ and several rational tails attached to $\P^{1}_{b}$ at the points $q_{i}$.
\item We contract any components of $P$ which have $\leq 2$ special points.
\end{enumerate}
The proposition follows from this description of the stable reduction process.
\end{proof}
\begin{remark}
\label{remark:stablereductionprocess}
The process of stable reduction described in the proof of \autoref{proposition:stablereduction} will be referred to frequently in the remainder of this section. We emphasize that when $k \geq 3$, the original fiber $\P^{1}_{b}$ {\sl naturally} persists as a component of the stable reduction $P$.
\end{remark}
\begin{definition}
\label{definition:chain}
A {\sl chain} is a nodal curve $T = P_{1} \cup \dots \cup P_{j}$ where each $P_{i} \equiv \P^{1}$, and $P_{i}$ intersects $P_{i+1}$ at one node, for all $i = 1, \dots j-1$, with no further intersections between components.
\end{definition}
\begin{proposition}
\label{proposition:limitedorigin}
Assume $(P,s_{1}, \dots , s_{d}) \in \barm_{0,d}$ is a stable curve which is the nodal union of:
\begin{enumerate}
\item a $\P^{1}$ with $n$ marked points $p_{1}, \dots, p_{n}$, and
\item $k-n$ chains $T_{i}$, $i = 1, \dots k-n$, with each $T_{i}$ intersecting $\P^{1}$ once at a point $q_{i}$, every component of $T_{i}$ having exactly three special points, and each $T_{i}$ containing $a_{i}>1$ marked points.
\end{enumerate}
Suppose $(P, s_{1}, \dots, s_{d})$ is the stable replacement for $D \subset \P^{1}_{b}$ in \autoref{notation5}.
Then
\begin{enumerate}
\item $D \subset \P^{1}_{b}$ is supported on $\leq k$ points.
\item If $D$ is supported on $k$ points, and if we assume $k \geq 4$, then $D$ must equal $p_{1}+ \dots + p_{n} + a_{1}q_{1} + \dots a_{k-n}q_{k-n}$.
\end{enumerate}
\end{proposition}
\begin{proof}
These two points follow from the stable reduction procedure described in the proof of \autoref{proposition:stablereduction}, and from the observations in \autoref{remark:stablereductionprocess}.
\end{proof}
\subsection{Dandelions and the proof of \autoref{theorem:main}}
\begin{definition}
\label{definition:k-dandelion}
Any point $[(P,s_{1},\dots,s_{d})] \in \barm_{0,d}$ having dual graph
\tikzset{main node/.style={thick,circle,draw},}
\begin{tikzpicture}
\node[main node, label={[xshift=-.1cm]\small $k-1$}] (1) {f};
\node[main node, label={\small $1$}] (2) [right = .7cm of 1] {s};
\node[main node, label={\small $1$}] (3) [right = .7cm of 2] {s};
\node[main node, label={\small $2$}] (4) [right = .7cm of 3] {s};
\path (1) edge (2);
\path (2) -- node[auto=false]{\ldots} (3);
\path (3) edge (4);
\end{tikzpicture}
is called a {\sl $k$-dandelion}. There are a total of $d-k$ vertices labelled ``s''.
(Here, the numbers above a vertex indicate the number of marked points on the corresponding component of the stable curve.)
The {\sl stem} of the dandelion is the curve which is the union of components labelled ``s''. The {\sl flower} of the $k$-dandelion is the component corresponding to the vertex labelled ``f''.
\end{definition}
\begin{proposition}
\label{proposition:wheredandelions}
Let $\left[P \right] \in \alpha(Z)$ be a $k$-dandelion with $k = d-(2r-2)\geq 4$. Then the the map $\alpha$ is unramified at every preimage of $\left[ P \right]$. Furthermore, every preimage of $\left[ P \right]$ projects, under $\beta$ to a point in $\Sigma^{[d]}$ lying over a $k$-incident line $[\ell] \in \G\left( 1,r \right)$ with multiplicity $(d-k+1,1, \dots, 1)$.
\end{proposition}
\begin{proof}
The hypersurface $X$ does not have any $k'$-incident lines with $k' < k$. Therefore, the proposition follows from point $2$ of \autoref{proposition:limitedorigin}.
\end{proof}
\begin{proposition}
\label{proposition:imagedandelion}
Suppose $k = d-(2r-2) \geq 3$. Then there exists a $k$-dandelion in the image of $\alpha$.
\end{proposition}
\begin{proof}
Consider the one-parameter family of polynomials $(z-t)(z-t^{2})\dots(z-t^{d-k})(z+t+t^{2}+\dots+t^{d-k})$ viewed as a map $\A^{1} \to \Delta_{d-k+1}$ given by $(r_{1}, \dots ,r_{d-k+1}) = (t,t^{2}, \dots, t^{d-k}, -t-t^{2}- \dots- t^{d-k})$. If $\ell$ is a $k$-incident line of $X$ with multiplicities $(d-k+1,1,\dots,1)$ then $\Delta_{d-k+1}$ is a local chart for any point $\hat{[\ell]} \in \Sigma^{[d]}$ lying above $\ell$, and therefore the one-parameter family above is realized, \'etale locally, as a curve on $\Sigma^{[d]}$. Applying stable reduction to this one parameter family yields a $k$-dandelion, as the reader can easily verify.
\end{proof}
We have all ingredients to prove the $m=1$ case of \autoref{theorem:main}.
\begin{theorem}
\label{theorem:genericinjectivity}
Suppose $d \geq 2r+2$, and $X \subset \P^{r}$ is a general degree $d$ hypersurface. Then the moduli map
\begin{align*}
\mu_{1} : \G\left( 1,r \right) \dashrightarrow \Hyp(d,1)
\end{align*}
is generically injective.
\end{theorem}
\begin{proof}
By \autoref{proposition:imagedandelion}, we know that there exists at least one $k$-dandelion $[P] \in \alpha(Z)$.
If $\left[ P \right] \in \alpha(Z)$ is a $k$-dandelion, we have shown: The preimage $\alpha^{-1}([P])$ originates from a unique $k$-incident line $\ell$, read off from the cross-ratios of the flower (\autoref{corollary:distinct-cross-ratios}), and the marking on the flower and stem of $[P]$ determines a point $\hat{[\ell]}$ and a point $z \in \beta^{-1}(\hat{[\ell]})$, respectively, showing that $\alpha^{-1}([P]) = z$.
We conclude from \autoref{proposition:wheredandelions} that the preimage of $[P]$ is the reduced point $z$, and from this the theorem follows.
\end{proof}
We now give the proof of \autoref{theorem:main}.
\begin{proof}[Proof of \autoref{theorem:main}]
By \autoref{theorem:genericinjectivity}, we know that the map $\mu_{1} : \G\left( 1,r \right) \dashrightarrow \Hyp(d,1)$ is generically injective.
Let $\Lambda \subset \P^{r}$ be a general $m$-plane. The injectivity of $\mu_{1}$ implies that a general line $\ell \subset \Lambda$ will be general in the sense that there will be no other line $\ell'$ with $[X \cap \ell \subset \ell] \simeq [X \cap \ell' \subset \ell']$.
Therefore, if $\Lambda'$ is another $m$-plane with $[X \cap \Lambda \subset \Lambda] \simeq [X \cap \Lambda' \subset \Lambda']$, it must be that $\ell \subset \Lambda \cap \Lambda'$. Since this is true for a general $\ell \subset \Lambda$, we conclude that $\Lambda = \Lambda'$, which proves the theorem.
\end{proof}
\subsection{Straight trees and the proof of \autoref{theorem:main2}}
\begin{notation}
During this section, we assume $d = 2r+1$ and $X \subset \P^{r}$ a general hypersurface of degree $d$.
\end{notation}
\begin{definition}
\label{definition:straighttree}
A {\sl straight tree} is any stable curve $(P,s_{1}, \dots, s_{d})$ having dual graph of the form
\tikzset{main node/.style={thick,circle,draw},}
\begin{tikzpicture}
\node[main node, label={\small $2$}] (1) {};
\node[main node, label={\small $1$}] (2) [right = .7cm of 1] {};
\node[main node, label={\small $1$}] (3) [right = .7cm of 2] {};
\node[main node, label={\small $1$}] (4) [right = .7cm of 3] {};
\node[main node, label={\small $2$}] (5) [right = .7cm of 4] {};
\path (1) edge (2);
\path (2) edge (3);
\path (3) -- node[auto=false]{\ldots} (4);
\path (4) edge (5);
\end{tikzpicture}
(There are a total of $d-2$ vertices.)
\end{definition}
\begin{lemma}
\label{lemma:numberstraighttrees}
Assume $d \geq 4$. There are $d!/8$ distinct straight trees.
\end{lemma}
\begin{proof}
Indeed, we may exchange the two points on each extremal component of a straght tree, or we may flip the entire tree over, without altering the moduli of the straight tree. Therefore, we must divide $d!$ by $2 \times 2 \times 2$.
\end{proof}
\begin{lemma}
\label{lemma:wherestraighttrees}
Suppose $z \in Z$ is such that $\alpha(z) \in \barm_{0,d}$ is a straight tree. Then $\hat{[\ell]} := \beta(z) \in \Sigma^{[d]}$ lies over a $3$-incident line $\ell$. Furthermore, the multiplicity of $\ell$ is $(1,a,b)$, where $a+b = d-1$.
\end{lemma}
\begin{proof}
$X$ possesses no $2$-incident lines. The lemma follows from \autoref{proposition:limitedorigin}, and from the observation in \autoref{remark:stablereductionprocess}.
\end{proof}
\begin{lemma}
\label{lemma:existsstraighttrees}
Let $a,b$ be positive integers such that $a+b=d-1$. Then for every $3$-incident line $\ell$ of multiplicity $(1,a,b)$, and every point $\hat{[\ell]} \in \Sigma^{[d]}$ lying over $\ell$, there are precisely $a!b!/4$ points $z \in \beta^{-1}(\hat{[\ell]})$ such that $\alpha(z)$ is a straight tree.
\end{lemma}
\begin{proof}
Let $p, q$ be the points of multiplicity $a$ and $b$ on $\ell$. Then the variety $\beta^{-1}(\hat{[\ell]})$ is the collection of stable curves obtained by attaching all rational tails with $a$ marked points to the point $p$, and similarly with $q$. The lemma follows, after recalling that the two extremal points in a straight tree may be exchanged without changing the moduli of the tree -- this explains the division by $4$.
\end{proof}
\begin{lemma}
\label{lemma:11d-2}
For every $3$-incident line $\ell$ of multiplicity $(1,1,d-2)$, and every point $\hat{[\ell]} \in \Sigma^{[d]}$ lying over $\ell$, there are precisely $(d-2)!/2$ points $z \in \beta^{-1}(\hat{[\ell]})$ such that $\alpha(z)$ is a straight tree.
\end{lemma}
\begin{proof}
The proof is completely analogous to the proof of the previous lemma. The only difference is that we are only attaching rational tails to {\sl one} point on $\ell$, hence we must divide by two only once.
\end{proof}
\begin{theorem}
\label{theorem:enumerativtheoremmain}
For any multiplicity $(a,b,c)$, let $n_{a,b,c}$ denote the number of $3$-incident lines to $X$ having the given multiplicity.
Then the degree of the map $\alpha: Z \to \barm_{0,d}$ is
\begin{align*}
\deg \alpha = 2 \sum_{a \geq b > 1}^{}n_{(a,b,1)} + 4n_{d-2,1,1}.
\end{align*}
Furthermore, $\deg \alpha = \deg \mu_{1}$.
\end{theorem}
\begin{proof}
The last statement is clear.
By symmetry considerations, every straight tree $[P] \in \barm_{0,d}$ has the same number of preimages in $Z$, all preimages arise from $3$-incident lines, by \autoref{lemma:wherestraighttrees}, and all preimages are unramified points of $\alpha$, by \autoref{corollary:localisomorphismontoimage}.
Above a $3$-incident line $\ell$ with multiplicity $(a,b,1)$, where $a,b>1$, there are $\frac{d!}{a!b!}$ points in $s^{-1}([\ell])$, by \autoref{lemma:numberpreimages}. Each contributes $a!b!/4$ straight trees by \autoref{lemma:existsstraighttrees}. Therefore, we obtain $\frac{d!}{a!b!}\frac{a!b!}{4} = \frac{d!}{4}$ total straight trees arising from $3$-incident lines with multiplicities $(a,b,1)$, where $a,b > 1$. Since there are $\frac{d!}{8}$ straight trees total (\autoref{lemma:numberstraighttrees}), we see that each such line contributes $2$ to the degree of $\alpha$.
The same argument, using \autoref{lemma:11d-2} explains the coefficient $4$ for lines of type $(d-2,1,1)$.
\end{proof}
\section{Further questions} There are many remaining questions worth understanding.
For instance, there is the question of extending \autoref{theorem:monodromymain} to the {\sl equi-multiplicity} setting, i.e. where the multiplicity vector is $(m,m,\dots,m)$ for some integer $m$. Recall the situation for plane curves: The monodromy groups of flexes of a plane cubic, and of bitangents of a plane quartic are not the respective full symmetric groups. Does this special phenomenon persist in higher dimensions? The first open case is to determine the monodromy group of $5$-tangent lines of a general quintic surface in $\P^{3}$.
\begin{remark}
\label{remark:error}
There is one instance in the literature worth noting. In \cite{dsouza}, it is argued that the monodromy group of the set of $4$-incident lines of type $(2,2,2,2)$ to an octic surface $X \subset \P^{3}$ is the full symmetric group. Unfortunately, it appears (to the author) that there is a small gap in the proof -- in the proof of Lemma 2.8, a particular variety $I_{8}$ is claimed to be locally irreducible, {\sl because it is irreducible}. This is false as it stands.
However, the author believes the result is still true, and that D'Souza's argument can be made to work.
\end{remark}
We have settled the enumerative question $(3)$ raised in the introduction in the case $m=1$ of lines. However, there are many more instances where the preconditions of question $(3)$ are met. We give the following interesting open example:
\begin{example}
\label{example:quarticthreefold}
Let $X_{4} \subset \P^{4}$ be a general quartic threefold. What is the degree of the rational map
\begin{align*}
\mu_{2}: \G\left( 2,4 \right) \dashrightarrow \m_{3}?
\end{align*}
\end{example}
\begin{remark}
\label{remark:approach}
One possible approach, similar to that pursued in this paper, would be to count the number of ``complete quadrilaterals'' on $X_{4}$ -- these are plane curves which are the union of four general lines. The stable reduction procedure would also have to be analyzed.
\end{remark}
Other enumerative puzzles emerge.
\begin{example}
\label{example:planesextic}
Let $X_{6} \subset \P^{2}$ be a general plane sextic. Then the moduli map $\mu_{1}: \P^{2*} \dashrightarrow \m_{0,6}/S_{6}$ is generically injective. Since the target is three dimensional, we expect the double-point locus of $\mu_{1}$ to be a curve $Y \subset \P^{2*}$. Is this the case?
If so, the assignment $X \mapsto Y$ is a contra-invariant of ternary sextics -- what are the order and degree of this contra-invariant?
\end{example}
\begin{example}
\label{example:Fermat}
Cadman and Laza \cite{cadmanlaza} show that the degree of $\mu_{1}$ for a plane Fermat quintic is $150$. What is the degree in higher dimension, e.g. for the Fermat septimic surface in $\P^{3}$? We suspect that the answer is $7^{3} \times 24$, the size of the automorphism group of the Fermat septimic.
\end{example}
Finally, and perhaps most importantly, there remains the question of improving \autoref{theorem:versal}. Indeed, it should be the case that the point-line correspondence is versal at every $k$-incident line with $k \geq d - (2r-2)$ for a general hypersurface $X$. In fact, there should be an analogous versality result for higher dimensional plane sections of $X$ -- this is the subject of future work.
|
\section{INTRODUCTION}
\label{sec:intro}
With the ever increasing access to global travel and movement of goods, the demand for a non-ionising, non-destructive and non-invasive device which is capable of detecting a range of materials is growing. Despite the various surveillance methods currently employed, many may present health hazards such as exposure to ionising radiation, or their effectiveness can be impeded by creative concealment of regulated or undesirable goods.
We therefore propose a non-invasive, tunable detector capable of imaging conductive objects concealed by conductive enclosures such as cargo holds or thick shieldings. The device relies on a Radio-Frequency Optical Atomic Magnetometer (RF-OAM) operated in the Magnetic Induction Tomography (MIT) modality\cite{cam, oamsens}. MIT has already proved its usefulness in a number of industrial applications such as eddy current testing, but has since also demonstrated applications in medicine, for example in the detection of cerebral stroke \cite{stroke} and of heart arrythmias \cite{scirep}.
\section{Magnetic Induction Tomography with Optical Atomic Magnetometers}
MIT relies on the detection of a secondary magnetic field ($\mathbf{B_{2}}$) generated by eddy currents excited in a conductive target by a primary AC magnetic field ($\mathbf{B_{1}}$). The sketch in Fig.~\ref{fig:mitter} schematically represents this process. The secondary field contains information about the conductivity, permittivity and permeability of the object of interest \cite{mit}.
\begin{figure}[htbp]
\includegraphics[width=0.25\textwidth]{mitter.pdf}
\centering
\caption{MIT principle. $\mathbf{B_{1}}$, the primary magnetic field, is generated by a coil driven by an AC current source. It induces eddy currents (in white) into a conductive object, which generate a secondary field $\mathbf{B_{2}}$, whose amplitude and phase lag with respect to $\mathbf{B_{1}}$ depend on the dielectric properties of the supporting medium. A total magnetic field $\mathbf{B_{Tot}=B_{1}+B_{2}}$ is detected by a dedicated sensor.}
\label{fig:mitter}
\end{figure}
In conventional MIT systems, coils are used to produce eddy currents and also to pick-up the signal generated by $\mathbf{B_{2}}$. In our setup, based on magnetic induction imaging performed with an RF-OAM, the pick-up coil is replaced by an atomic magnetometer sensing unit \cite{cam}. A similar approach has recently been proposed by Wickenbrock \textit{et al}\cite{arnearxiv}.
The use of OAMs reduces or eliminates the impact of limitations and drawbacks of conventional sensors\cite{oamsens}, mainly in terms of sensitivity, available bandwidth and removal of capacitive coupling between induction and pick-up coils, which could alter the detection of the sample's properties. Alkali-metal vapor magnetometers have been shown to reach sensitivities which outperform SQUID magnetometers \cite{squidsens}. Moreover, the potential for miniaturisation for OAMs, thanks to micro-fabricated vapor cells, provides the potential for increased spatial resolution, reduced footprint and increased portability.
Other practical advantages include the operation of OAMs at room temperature and, in MIT modality with a phase-sensitive detection system, without the need for magnetic shielding. Finally, OAMs do not require calibration.
\section{Imaging Concealed Objects}
Our RF-OAM MIT setup has been extensively described elsewhere \cite{cam}; hence, we report here only the main details and a diagram of the setup (Fig.~\ref{fig:setup}). The RF-OAM is based on a rubidium vapor cell, where ${}^{87}$Rb is optically pumped and probed by laser radiation at 780 nm.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\textwidth]{layout.pdf}
\caption{RF-OAM based MIT experimental setup. Key: OI is an optical isolator, PD is a balanced photodiode, AOM is an acousto-optical modulator, IC the induction coil generating the primary field and driving the OAM, HC is a Helmholtz coil pair, LIA is a lock-in amplifier and DAQ is the data acquisition device connected to a laptop. HWP and QWP refer to half and quarter waveplates, respectively. The sensing cell filled with a natural mixture of $^{87}$Rb and $^{85}$Rb is in the location of intersection of the pump and probe beams, between the Helmholtz coils. DAVLL is the dichroic atomic vapor laser lock used for laser's frequency stabilization.}
\label{fig:setup}
\end{figure}
The pump and probe laser beams intersect orthogonally at the center of the vapor cell. A smaller overlap potentially increases the spatial resolution of the MIT system since the only atomic spins involved in the operation are enclosed in that volume. In other words, the OAM would be sensitive only to local variations of $\mathbf{B_{Tot}}$. Nevertheless, this reduces the signal-to-noise ratio in given conditions. Therefore, a practical trade-off between spatial resolution and signal level must be found in view of desired applications.
The inductive coil (IC) sketched in Fig.~\ref{fig:setup} has two simultaneous functions. The first one is to coherently drive atomic precession in the RF-OAM. The second function is to induce eddy currents in the object of interest, as in conventional MIT configuration. The RF-OAM is tuned to resonance by adjusting the DC magnetic field generated by the Helmholtz coils pair, through suitable changes in the supplied current.
The eddy currents induced in the object of interest circulate mainly on the surface of the material, with the skin depth $\delta$ controlling their penetration. This is dependent on the inherent properties of the material itself and the frequency supplied to the induction coil. In the case of conductive and non-magnetic materials, it is given by \cite{sdepth}:
\begin{equation}
\delta=\sqrt{\dfrac{2}{\omega \mu_{0}\sigma}}~,\label{skinnyd}
\end{equation}
\noindent where $\sigma$ is the DC conductivity of the material, $\mu_{0}$ is the permeability of free space and $\omega$ the angular frequency of the AC field probing the object. The corresponding dependency for copper and aluminium is shown in Fig.~\ref{skind}.
\begin{figure}[h]
\includegraphics[width=0.5\linewidth]{skindepth.pdf}
\centering
\caption{Penetration depth through a conductive object as a function of frequency supplied to induction coil for copper ($\sigma$=59.8$ \times$10$^{6}$ S/m) and aluminium ($\sigma$=37.7$\times$10$^{6}$ S/m) samples.}
\label{skind}
\end{figure}
The choice of operation frequency depends on both the materials and the desired penetration (Eq.~\ref{skinnyd}). Thanks to the bandwidth and flexibility provided by the RF-OAM, the frequency can be chosen in order to image objects hidden by conductive shields. Once the induction coil frequency has been selected, the RF-OAM is tuned accordingly.
The signal from the probe beam at the photodiode is due to the total magnetic field acting upon the OAM unit, $\mathbf{B_{Tot}}$; that is to say it contains information about the driving field due to the induction coil and the signal due to eddy currents circulating in the target itself. A lock-in amplifier (LIA) is used to extract the information encoded the in OAM output, from which the magnetic field's properties and hence the dielectric characteristics of the object of interest are inferred. In particular, amplitude (radius) and retardation (phase) of the secondary field's contribution are measured.
\section{Imaging of Concealed Conductive Objects}
Here, the capability of our RF-OAM to detect concealed objects \cite{cam}, is investigated in detail. A copper square with sides 25 mm and thickness of 1 mm was shielded by a larger aluminium square with sides 37 mm and thickness also of 1 mm.
The test objects were placed on a Perspex mount positioned directly above the center of the atomic sensor. The mount, attached to a translation stage, was moved by micrometer screws. At each position of the object, $10^{3}$ data points obtained from the LIA output were averaged and recorded. After averaging, data were stored in a 2D array by means of a DAQ connected to a laptop computer. Data were then processed using a nearest neighbour filter with a radius r=2, and displayed in color-coded maps.
\begin{figure}[h]
\centering
\includegraphics[width=1\linewidth]{cu_square.pdf}
\caption{MIT images of the unshielded copper square. \textbf{a)} and \textbf{b)}: radius image of the copper square and corresponding contour plot. \textbf{c)} and \textbf{d)}: phase image and corresponding contour plot of the same sample. Images were taken with a 1 kHz driving field, which corresponds to $\delta_{Cu}=2.1$ mm in copper.}
\label{fig:noshield}
\end{figure}
The unshielded copper square images in Fig.~\ref{fig:noshield} accurately reproduce the size and shape of the Cu sample. The rounded shape of the image may be due in part to the nature of the circular pattern of eddy currents. The nearest neighbour filter also emphasizes this effect.
The copper square was then shielded from the sensor using an aluminium square, with a thin insulating layer between them to prevent eddy currents being passed between the two materials, unlike the previous investigation \cite{cam}. The images were taken with a driving frequency of 500 Hz which corresponds to a skin depth of $\delta_{Al}=3.7$ mm in aluminium and $\delta_{Cu}=2.9$ mm in copper. The images were generated in a single measurement with no need for subtraction of the background. More importantly, no subtraction of the Al contribution was necessary, given the complete penetration of the AC fields through the concealing barrier.
Results are shown in Fig.~\ref{shieldedsquare} for the radius and phase data respectively. By inspection, it is clear that the edges of the radius maps (panels a) and b)) are more noisy and hence less defined than those obtained in the phase maps (panels c) and d)). This deformation is particularly pronounced along the x axis. This has been attributed\cite{cam} to the presence of possible edge effects and mutual induction between the sample and the Al barrier.
\begin{figure}[h]
\centering
\includegraphics[width=1\linewidth]{shieldedsq.pdf}
\caption{MIT images of copper square shielded from induction coil with a larger aluminium square, with a non-conductive layer between the two metals. \textbf{a)} and \textbf{b)}: radius image and contour map of the shielded square. \textbf{c)} and \textbf{d)}: corresponding phase images. Images were taken with a 500 Hz driving field, which corresponds to $\delta_{Al}=3.7$ mm and $\delta_{Cu}=2.9$ mm.}
\label{shieldedsquare}
\end{figure}
Moreover, both maps appear smaller than those obtained in absence of the Al screen. This, in turn, can be explained by edge effects produced at the boundary of the two object, which may be particularly severe when the size of the samples are similar.
\section{Conclusion}
Our RF-OAM based MIT system has demonstrated the possibility of a real-time scanning device for the imaging and detection of conductive objects, with the ability to penetrate conductive shieldings.
The sensitivity of our RF-OAM allows the imaging of conductive targets concealed by a conductive screen, without the need for magnetic shielding, or background subtraction. Furthermore, our system, thanks to the flexibility of the RF-OAM, can be easily tuned to image a wide variety of materials.
The current setup has a reduced complexity compared with the initial proof-of-principle demonstration \cite{wickenbrock2014}, reducing cost and size, thus improving everyday practical capabilities such as the potential for a hand-held scanner in an unshielded environment. This is ideal for applications in security and surveillance where a non-ionising, non-invasive easy-to-use scanner is required to detect a wide range of concealed objects.
\section{Acknowledgements}
S.~H. is supported by DSTL - Defence and Security PhD - Sensing and Navigation using Quantum 2.0 technology. L.~M. is supported by Innovate UK as part of the AMMIT project. C.~D. is supported by the EPSRC Centre
for Doctoral Training in Delivering Quantum Technologies.
|
\section{Introduction}
The body-centered tetragonal $AM_2X_2$ ternary compounds ($A$ = rare or alkaline earth, $M$ = transition metal, $X$ = Si, Ge, P, As, Sb) with the ${\rm ThCr_2Si_2}$ structure\cite{Just1996} have generated tremendous interest in the scientific community due to their novel electronic and magnetic properties. Prominent among these is the iron-arsenide family of parent compounds $A{\rm Fe_2As_2}$ ($A$ = Ca, Sr, Ba, Eu).\cite{Johnston2010, Stewart2011, Scalapino2012, Dagotto2013, Fernandes2014, Hosono2015, Dai2015, Inosov2016, Si2016} These materials are metallic and show nearly contiguous antiferromagnetic (AFM) spin-density wave and structural transitions at temperatures~$T$ up to \mbox{$\sim 200$~K}. The suppression of these transitions by external pressure or chemical doping leads to superconductivity with bulk superconducting transition temperatures $T_{\rm c}$ up to 56~K\@. It is believed that the FeAs layer as the conducting sheet in this structure plays a crucial role in the occurrence of superconductivity.\cite{Johnston2010, Stewart2011, Scalapino2012, Dagotto2013, Fernandes2014, Hosono2015, Dai2015, Inosov2016, Si2016} Hence, it is important to investigate other related materials with similar compositions and structures in the search for new superconductors and other novel phenomena. For example, SrNi$_2$As$_2$ ($T_{\rm c} = 0.62$~K, Ref.~\onlinecite{Bauer2008}) and BaNi$_2$As$_2$ ($T_{\rm c} = 0.7$~K, Ref.~\onlinecite{Ronning2008}) were both found to be superconductors.
On the other hand, SrCo$_2$As$_2$ (Refs.~\onlinecite{Pandey2013, Jayasekara2013, Wiecki2015}) and BaCo$_2$As$_2$ (Refs.~\onlinecite{Sefat2009, Anand2014}) are correlated metals with no structural, superconducting or long-range magnetic ordering transitions. From inelastic neutron scattering measurements, SrCo$_2$As$_2$ is found to exhibit strong AFM correlations at the same stripe wavevector as do the superconducting iron arsenides, which raises the interesting question of why SrCo$_2$As$_2$ is not a high-$T_{\rm c}$ superconductor.\cite{Jayasekara2013} The reason has been suggested from NMR measurements to be that SrCo$_2$As$_2$ exhibits strong {\it ferromagnetic} (FM) spin correlations/fluctuations in addition to the AFM correlations and these compete with the AFM correlations that are the presumptive glue for superconductivity in these systems. Subsequent NMR studies indicated that the large range of $T_{\rm c}$ observed within the FeAs-based systems may also arise from the competition between FM and AFM correlations.\cite{Wiecki2015b}
Recently significant attention has focussed on Mn arsenides. Our studies of the properties of the parent and doped BaMn$_2$As$_2$ systems were originally motivated by their potential to be ThCr$_2$Si$_2$-type high-$T_{\rm c}$ superconductors analogous to the cuprates. The semiconductor BaMn$_2$As$_2$ shows G-type (checkerboard-type) local-moment collinear AFM order below its high N\'eel temperature $T_{\rm N} = 625$~K with the ordered moments aligned along the tetragonal $c$~axis.\cite{Singh2009, Singh2009b, Johnston2011} Thus magnetoelastic coupling does not cause a distortion of the crystal structure below $T_{\rm N}$, contrary to the orthorhombic distortion associated with AFM ordering in the $A{\rm Fe_2As_2}$ compounds due to the collinear Fe ordered moments aligned in the $ab$~plane. An optical gap of 48~meV was inferred for BaMn$_2$As$_2$ from the optical conductivity,\cite{Antel2012} consistent with results from the electrical resistivity $\rho$ versus temperature~$T$ measurements in the $ab$~plane.\cite{Singh2009} Furthermore, this optical study\cite{Antel2012} found that BaMn$_2$As$_2$ is much more two-dimensional in its electronic properties than are the $A{\rm Fe_2As_2}$ parent compounds.\cite{Singh2009} A neutron scattering study of isostructural ${\rm BaMn_2Bi_2}$ found the same G-type AFM structure as in ${\rm BaMn_2As_2}$ but with a lower $T_{\rm N} = 387$~K.\cite{Calder2014}
Only 1.6\% K substitution for Ba transforms BaMn$_2$As$_2$ into a local-moment AFM metal.\cite{Pandey2012, Yeninas2013} Higher doping levels lead to the onset of FM at $\approx 16$\% K-doping\cite{Bao2012} and half-metal FM behavior below the Curie temperature $T_{\rm C} \sim 100$~K at 40\% K~doping (Refs.~\onlinecite{Pandey2013b, Ueland2015}) and at 60\% Rb doping.\cite{Pandey2015} The FM is thus thought to be associated with FM ordering of the itinerant doped-hole spins and coexists with the G-type AFM order of the local Mn moments with $T_{\rm N}>300$~K.\cite{Ueland2015,Lamsal2013}
Unlike BaMn$_2$As$_2$ with the tetragonal ${\rm ThCr_2Si_2}$ structure, the compounds \sma\ and \cma\ both crystallize in the trigonal \cas-type structure\cite{Mewis1978, Brechtel1978} containing a corrugated honeycomb Mn sublattice which can be viewed as a triangular lattice bilayer. The possibility of geometrically-frustrated triangular-lattice exchange connectivity exists and such compounds often show novel physical behaviors associated with the geometric frustration.\cite{Ramirez1994, Moessner2006, Balents2010} Single crystals of \sma\ were grown previously using Sn flux.\cite{Wang2011} These authors' in-plane electrical resistivity $\rho(T)$ measurements indicated that the ground state is insulating with activation energies of 0.29--0.64~eV depending on the $T$ range, and their magnetic susceptibility $\chi(T)$ measurements indicated an AFM transition at $T_{\rm N} = 125$~K\@.\cite{Wang2011}
Two neutron powder diffraction studies\cite{Ratcliff2009, Bridges2009} of the related \cas-type ${\rm CaMn_2Sb_2}$ revealed AFM ordering below $T_{\rm N}= 88$~K and 85~K, respectively, with an AFM propagation vector ${\bf k} = (0,0,0)$, i.e., the crystal and AFM unit cells are the same. In the former paper the AFM structure was deduced to be collinear, with the ordered moments aligned in the $ab$~plane with a low-$T$ ordered moment of 2.8(1)~$\mu_{\rm B}$/Mn, where $\mu_{\rm B}$ is the Bohr magneton. In the latter paper, a model was favored with the ordered moments canted at $\pm 25^\circ$ with respect to the $ab$ plane with an ordered moment of 3.38(6)~$\mu_{\rm B}$/Mn.
Herein, we report the growth, crystal structure, $\rho(T)$, magnetization as a function of magnetic field $M(H)$, $\chi(T)$ and heat capacity $C_{{\rm p}}(T)$ measurements of \cma\ and \sma\ single crystals. These studies were initiated because of the above-noted possibility that the Mn spin lattice might exhibit novel magnetic behaviors associated with the presence of geometric frustration within the triangular-lattice Mn layers. If the strongest AFM interactions are indeed within a triangular lattice layer, this should lead to a noncollinear AFM structure below $T_{\rm N}$. Instead, in a companion neutron diffraction study to the present work, the AFM structure of \sma\ was found to be collinear with the ordered Mn moments aligned in the $ab$ plane with magnitude 3.6~$\mu_{\rm B}$/Mn.\cite{Das2016} This magnetic structure is the same as one of the two AFM structures proposed for ${\rm CaMn_2Sb_2}$ (Ref.~\onlinecite{Ratcliff2009}) discussed above.
We discovered that the $\chi(T)$ of \sma\ and \cma\ and the $C_{\rm p}(T)$ of \sma\ above their respective N\'eel temperatures $T_{\rm N}$ of 120 and 62~K exhibit behaviors characteristic of strong dynamic short-range AFM spin correlations up to at least 900~K, likely arising from quasi-two-dimensional connectivity of strong AFM Mn--Mn exchange interactions within the corrugated honeycomb Mn spin sublattice. This result is interesting because such strong AFM spin correlations up to high temperatures and the suppression of $T_{\rm N}$ to much lower temperatures than expected from molecular field theory, due to AFM fluctuations associated with the low dimensionality of the exchange interaction connectivity, may give rise to novel physical properties upon doping the compounds into the metallic state.
\section{Experimental Details}
Single crystals of \sma{} and \cma{} were grown using Sn flux. High-purity elements Sr (99.95\%) from Sigma Aldrich, and Ca (99.95\%), Co (99.998\%), As (99.9999\%) and Sn (99.999\%) from Alfa Aesar were taken in the ratio (Sr,Ca):Mn:As:Sn = 1:2:2:20 and placed in an alumina crucible that was subsequently placed in a silica tube that was evacuated, partially refilled with high-purity argon ($\approx$1/4 atm pressure) and then sealed. After preheating at 600$^{\circ}$C for 5 h, the assembly was heated to 1150~$^{\circ}$C at the rate of 50~$^{\circ}$C/h and held at this temperature for 20 h for homogenization. Then the furnace was slowly cooled at the rate of 5~$^{\circ}$C/h to 700~$^{\circ}$C\@. At this temperature the molten Sn flux was decanted using a centrifuge. Shiny hexagonal-shape single crystals of maximum dimensions $4\times 3\times 1\ {\rm mm}^3$ were obtained.
Semiquantitative chemical analyses of the single crystals were performed using a JEOL scanning electron microscope (SEM) equipped with an EDX (energy-dispersive x-ray analysis) detector, where a counting time of 120 s was used. A room-temperature powder x-ray diffraction (XRD) pattern was recorded on crushed single crystals using a Rigaku Geigerflex powder diffractometer with Cu K$\alpha$ radiation at diffraction angles 2$\theta$ from 10$^{\circ}$ to 110$^{\circ}$ with a 0.02$^{\circ}$ step width. The data were analysed by Rietveld refinement using FullProf software.\cite{fullprof}
$M(T)$ measurements for ${\rm 1.8~K \leq}~T~{\rm \leq 300~K}$ and $M(H)$ measurements for $H \leq 5.5$~T were carried out using a Quantum Design, Inc., Magnetic Properties Measurement System (MPMS). The high-temperature $M(T)$ for ${\rm 300~K \leq T \leq 900~K}$ was measured using the vibrating sample magnetometer (VSM) option of a Quantum Design, Inc., Physical Properties Measurement System (PPMS)\@. In this paper we exclusively use Gaussian cgs units for $M$, $\chi$ and $H$ (see Sec.~3.5.1 of Ref.~\onlinecite{Johnston2010}). In this system of units, the Tesla~(T) is a unit of convenience for~$H$ defined as 1~T~=~$10^4$~Oe, where Oe is the conventional cgs unit for~$H$.
$C{\rm_p}(T)$ data were obtained using a relaxation method with the heat capacity option of the PPMS\@. Four-probe $\rho(T)$ data were obtained with an ac current amplitude $I=1\,\rm{\mu A}$ at a frequency of 37.7~Hz using the ac transport option of the PPMS\@. Electrical contacts to a crystal were made by soldering 0.05~mm diameter Pt wire to a crystal using indium solder.
\section{Experimental Results}
\subsection{\label{Sec:Struct} Crystal Structure}
\begin{figure}
\includegraphics[width=3.5in]{Fig_01a.eps}
\includegraphics[width=3.5in] {Fig_01b.eps}
\caption{\label{Fig:xrd}(Colour online) Powder x-ray diffraction patterns (open circles) of (a) \sma{} and (b) \cma{} at room temperature. The solid line represents the Rietveld refinement fit calculated for the \cas{}-type trigonal structure with space group $P\bar{3}m1$ together with the Sn impurity phase.}
\end{figure}
\begin{table}
\caption{\label{Table:crystal structure parameter} Crystallographic and Rietveld refinement parameters
obtained from powder XRD of \sma{} and \cma{} crystals. The structures are trigonal ${\rm CaAl_2Si_2}$-type with space group $P\bar{3}m1$. The atomic coordinates of \scma{} are Sr/Ca: $1a$ (0, 0, 0); Mn: $2d$ (1/3, 2/3, $z{\rm_{Mn}}$); and As: $2d$ (1/3, 2/3, $z{\rm_{As}}$). The shortest Mn--Mn interatomic distances in \scma{} [see Fig.~\ref{Fig:crystal_structure}(b)] are also listed.}
\begin{ruledtabular}
\begin{tabular}{lcc}
& \sma{} & \cma{}\\
\hline
Lattice parameters \\
$a$~(\AA) & 4.2962(1) & 4.2376(1) \\
$c$~(\AA) & 7.2997(2) & 7.0331(2) \\
$c/a$ & 1.6991(1) & 1.6596(1) \\
$V_{\rm cell}~(\rm{\AA}^3)$ & 116.682(6) & 109.372(6) \\
\hline
Atomic coordinates \\
$z{\rm_{Mn}}$ & 0.6231(1) & 0.6248(4) \\
$z{\rm_{As}}$ & 0.2667(2) & 0.2537(3) \\
\hline
Refinement quality \\
$\chi^2$ & 3.05 & 4.03 \\
$R_{\rm p}$ (\%) & 10.3 & 12.7 \\
$R_{\rm wp}$ (\%) & 13.6 & 16.4 \\
\hline
Shortest Mn--Mn\\
distances (\AA)\\
$d_1$ & 3.06306(8) & 3.0112(2)\\
$d_2$ & 4.29620(5) & 4.23760(5)\\
$d_3$ & 5.27633(7) & 5.1985(2)\\
$d_{z1}$ & 6.0357(2) & 5.8171(4)\\
$d_{z2}$ & 7.2997(2) & 7.0331(2)\\
\end{tabular}
\end{ruledtabular}
\end{table}
The crystal symmetry of several \scma{} crystals was checked by x-ray Laue back scattering which showed trigonal symmetry with well-defined diffraction spots which clearly indicated the good quality of the crystals. In this paper we use the hexagonal setting for the trigonal unit cell. The data also revealed that the \scma\ platelike crystals grow with the plate surface parallel to the hexagonal $ab$~plane. SEM imaging and EDX analyses were performed to check the chemical composition and surface morphology of the crystals. The average elemental ratio of the samples was in agreement with the expected 1:2:2 stoichiometry of the compounds to within the errors. The amount of Sn incorporated into the crystal structure from the Sn flux is zero to within the experimental error. The present analyses did not show any other elements.
The phase purity of our \scma{} crystals was confirmed by powder XRD\@. Their XRD patterns at 300 K along with the results of Rietveld refinements are shown in Figs.~\ref{Fig:xrd}(a) and~\ref{Fig:xrd}(b), respectively. One sees the presence of adventitious elemental Sn flux, so two-phase Rietveld refinements were carried out. The refinement results confirm that the crystals have the trigonal \cas-type structure with space group $P\bar{3}m1$. The refinement and crystal parameters obtained are listed in Table~\ref{Table:crystal structure parameter}. The crystal parameters are in good agreement with previously reported values.\cite{Mewis1978, Brechtel1978, Wang2011}
|
\section{Introduction}\label{section:intro}
The motivation behind this work is to investigate the characteristics of quantum
computation when viewed as randomized algorithms. It is known that quantum
amplitude amplification, the key technique underlying Grover's unordered
search algorithm, is able to reduce and even eliminate error of one-sided quantum
black-box algorithms for search problems~\cite{BHMT}. We explored that direction further for
two-sided error algorithms for decision problems based on
the key observation that quantum algorithms appear to be
better at distinguishing between two given probability distributions compared to
classical randomized algorithms.
Suppose we are given a biased coin whose
distribution is either
$\mu_1 = \langle 1/3, 2/3 \rangle$ or $\mu_2 = \langle 2/3, 1/3 \rangle$.
A classical problem of probabilistic classification is to determine the
distribution of the coin by tossing it several times.
Various techniques exist like Bayesian classification and
maximum likelihood estimation, all of which aim to minimize some kind of error
that is inherent in such a probabilistic inference. But it is not believed to be
possible to confidently classify a distribution without any error. This is
true even if $\mu_1 = \langle 0,1\rangle$ instead.
However, such classification is possible when the distributions come
from a {\em quantum system}, our definition of a quantum source of
random samples.
We define a quantum system (QS) as a combination of a quantum circuit
$C$, an input to the circuit $\ket{\psi}$ and
a two-outcome projective measurement operator $\mathcal{P}=\langle P_E, I -
P_E \rangle$ (two outcomes will be {\em always labeled as $E$ and $F$} for
convenience) and denote it by $\langle \ket{\psi}, C, \mathcal{P} \rangle$. If we are given an actual
instance of a QS and we {\em
apply the circuit on the input followed by measurement using the projective
operator}, we will obtain a
sample in $\{E,F\}$ from the output probability distribution $\langle p_E, 1-p_E
\rangle$ where $p_E$
denotes the probability of observing outcome $E$ when
$C\ket{\psi}$ is measured using $\mathcal{P}$.
The quantum version of the above question of classifying between $\mu_1$ and
$\mu_2$ becomes this:
given an
instance $\mathcal{Q}$ which can be either a quantum system $\mathcal{Q}_1$ with output distribution $\mu_1$ or
QS $\mathcal{Q}_2$ with output distribution $\mu_2$, can we confidently
figure out if $\mathcal{Q}$ is $\mathcal{Q}_1$ or $\mathcal{Q}_2$ (in other words, determine the actual distribution
of $\mathcal{Q}$) by using $\mathcal{Q}$ in a black-box manner? Assume that both $\mathcal{Q}_1$ and$\mathcal{Q}_2$ involve the same
number of qubits and the same set of outcomes ($E$ and $F$). This is analogous to asking if two or more distinct
distributions (over same support of two elements) can be distinguished
without any probability of error. Even though classical techniques cannot
identify the exact distribution from the sample distribution without any
error,
we show that it is possible to do so for
distributions of quantum systems.
\begin{theorem}\label{theorem:application}
Given a quantum system $\mathcal{Q} = \langle \ket{\psi}, C, \mathcal{P} \rangle$ whose output distribution can either be
$\pair{\delta}{1-\delta}$ or $\pair{\epsilon}{1-\epsilon}$ for some $0 \le \delta
< \epsilon \le 1$, there is a quantum circuit $C'$ which can determine
the output distribution of $\mathcal{Q}$ without any probability of error. $C'$ takes
$\ket{\psi}$ as input, makes repeated calls to $C$, $C^\dagger$ and
employs gates that depend upon operators of $\mathcal{P}$ and $\ket{\psi}$.
\end{theorem}
The core technique is once again, {\em quantum amplitude amplification}.
It can be thought of as a quantum analog of repeated
trials used in randomized algorithms for reducing mis-classification error.
It is the workhorse
behind Grover's famous quantum unordered search algorithm~\cite{Grover1996} and was later
shown to be also applicable to the Deutsch-Jozsa
problem~\cite{BeraDJ2015}.
It appears that quantum algorithm designers simply
cannot wave it enough; it is applicable to almost any search
problem to yield a surprising improvement, usually quadratic, over classical algorithms.
Since its inception, amplitude amplification have been used, either directly
or in the form of Grover's search algorithm for a vast range of problems like minimum of an unordered
array~\cite{Durr96}, minimum spanning tree~\cite{Durr2004} and even
clustering~\cite{Aimeur2007}.
Nevertheless, we feel that the technique still has a long way to go, especially,
when used in a non-blackbox manner.
The most generalized and popular version of this technique was given by
Brassard et al.
\begin{theorem}[Exact amplitude amplification~\cite{BHMT}]\label{theorem:bhmt}
Consider a Boolean function $\Phi: X \to \{0,1\}$ that partitions a set $X$
between its {\em good} (those which $\Phi$ evaluates to 1) and {\em bad} (those
which evaluate to 0) elements. Consider also a quantum algorithm that uses
no measurements and uses oracle gates for computing $\Phi$
such that $C\ket{0}$ is quantum superposition of the elements of $X$ and let
$a > 0$ denote the success probability that a good element is observed if
$C\ket{0}$ is measured (in the standard basis).
There exists a quantum circuit (that depends upon $a$) which finds a
good solution with certainty using at most $\Theta(1/\sqrt{a})$ applications of $C$ and
$C^\dagger$.
\end{theorem}
This theorem is highly versatile as it is. However, for our applications we
require further generalizations. For example, we are interested in not only
one-sided, but also two-sided error algorithms. We also want to apply it to
algorithms which are measured not necessarily in the standard basis. Lastly,
we want algorithms which act on non-$\ket{0}$ input states, specifically, input
states that correspond to the input $\Phi$, suitably encoded -- this is similar to classical Boolean circuits without
oracle gates.
Lastly, for the results of this paper we stick to only
decision versions of the above theorem (though our results could be extended to
circuits that output some solution).
The following theorem is our version of Theorem~\ref{theorem:bhmt} with the constraint that the probability $a$ is fixed for every
possible $\Phi$ (condition of {\em exactness}).
\begin{theorem}[Decision version of generalized exact amplitude
amplification]\label{theorem:aa-two-sided}
Consider a Boolean function $\Phi: X \to \{0,1\}$ that partitions a set $X$
between its {\em good} (those which $\Phi$ evaluates to 1) and {\em bad} (the
rest of $X$) elements.
Suppose $C$ is a quantum algorithm (or circuit) that uses no measurement and
decides $\Phi$ with two-sided exact error $(\delta,\epsilon)$ for some $\delta <
\epsilon$. That is, the probability of error when $C$ is given a good $x \in
X$ is
{\em exactly} $\epsilon$ and when $x$ is bad is {\em exactly} $\delta$.
Here success and error is determined upon measurement of the output state of
$C$ by any projective measurement with two outcomes.
There exists a quantum circuit $C'$ that
calls $C$ and $C^\dagger$, uses the same input as that of $C$ (maybe with
ancill\ae), is measured using an
extension of the measurement operator for $C$ and decides $\Phi$ with certainty,
\end{theorem}
The primary contribution of this paper are a few interesting applications of
amplitude amplification. If we have two quantum
systems which differ only in their circuit, then we can essentially use their
output distribution, after suitably amplifying the systems, to distinguish
between those circuits. We show how this can be used to detect faults in
quantum circuits.
On the other hand, if we have two systems that differ only in
their input states, then we get a way to amplify their probability of
acceptance. This is exactly at the core of our proof that quantum classes
equivalent to exact two-sided and exact one-sided error classes can be
``derandomized'', in the sense that their errors can be completely eliminated.
One of the major, and still open, questions of {\em Complexity
Theory} is how \P compares to \ensuremath{\mathbf{RP}}\xspace and
\ensuremath{\mathbf{BPP}}\xspace, one-sided and two-sided bounded error polynomial-time complexity classes. The current best results
are the obvious inclusions $\P \subseteq \ensuremath{\mathbf{RP}}\xspace \subseteq \ensuremath{\mathbf{BPP}}\xspace$, though there are
some evidences of their equivalence.
Same question for their quantum analogs is in an equally indeterminate state,
i.e., $\ensuremath{\mathbf{EQP}}\xspace \subseteq \ensuremath{\mathbf{RQP}}\xspace \subseteq \ensuremath{\mathbf{BQP}}\xspace$; these are quantum analogs of \P,
$\ensuremath{\mathbf{RP}}\xspace$ and $\ensuremath{\mathbf{BPP}}\xspace$, respectively. There is not even much evidence
that $\ensuremath{\mathbf{EQP}}\xspace = \ensuremath{\mathbf{BQP}}\xspace$. One
approach towards settling this question is studying restricted versions of these
classes. Our results show that their
exact error versions, $\ensuremath{\mathbf{ERQP}}\xspace$ and $\ensuremath{\mathbf{EBQP}}\xspace$, are identical to $\ensuremath{\mathbf{EQP}}\xspace$ as long as
the two(one)-sided errors are fixed
for all instances~\footnote{The same question for classical classes was asked here:
\url{http://cstheory.stackexchange.com/questions/20027/in-what-class-are-randomized-algorithms-that-err-with-exactly-25-chance}.}.
\vspace*{1em}\noindent{\bf Organization:} The rest of the paper is organized as
follows.
We discuss quantum distinguishability of quantum systems in
Section~\ref{section:distinguish-qs}. The proof of our main theorem on
distinguishability is given in Section~\ref{section:proof}. This theorem, even
though quite general, is not suitable enough to amplify a collection of
quantum systems in a uniform manner; in Section~\ref{section:uniform} we
discuss a uniform version of our main theorem.
Section~\ref{section:distinguish-circuits} contains one of the
applications about detection of faults in quantum circuits and
in Section~\ref{section:exact-error-classes} we show that $\ensuremath{\mathbf{EBQP}}\xspace=\ensuremath{\mathbf{ERQP}}\xspace=\ensuremath{\mathbf{EQP}}\xspace$ and
prove Theorem~\ref{theorem:aa-two-sided} for regular circuits and those with
oracle gates.
\section{Distinguishing quantum systems}\label{section:distinguish-qs}
We will use $\mu_p$ to denote a distribution $\pair{p}{1-p}$ over outcomes
$\pair{E}{F}$ and $\mu(\mathcal{Q})$ to denote output distribution of a quantum system $\mathcal{Q}$.
As explained earlier, the main problem we are interested in involves a given instance of a quantum system $\mathcal{Q}$ which
can be either $\mathcal{Q}_\delta$ with output
distribution $\mu_\delta = \pair{\delta}{1-\delta}$ or $\mathcal{Q}_\epsilon$ with output
distribution $\mu_\epsilon = \pair{\epsilon}{1-\epsilon}$. We want to construct
a quantum algorithm, rather a circuit, that can ``call $\mathcal{Q}$ as a subroutine''
and determine if $\mathcal{Q}=\mathcal{Q}_\delta$ or $\mathcal{Q}=\mathcal{Q}_\epsilon$.
We can even extend this
to multiple quantum systems $\S = \{\mathcal{Q}_1, \mathcal{Q}_2, \ldots \}$ where output
distribution of any $\mathcal{Q}_i$ is either $\mu_\delta$ or $\mu_\epsilon$. We use the
notation $QD(\mathcal{Q}_1, \mathcal{Q}_2, \ldots)$ or even shorter $QD(\S)$ to refer to the {\em
quantum distinguishability} problem among quantum systems of $\S$.
Our goal is to design a quantum circuit in which we can ``embed any given $\mathcal{Q}$'' as a
black-box.
This motivated us to define a notion of black-box extension for quantum
systems, similar to quantum algorithms with subroutines or quantum circuits with
black-box operators, allowing only trivial extensions to inputs states and
projection
operators. We refer to these as $\mathcal{B}$-transforms ($\mathcal{B}$
standing for ``black-box''). A general illustration is given in Figure~\ref{fig:b_transform}.
\begin{figure*}[!h]
\centering \resizebox{0.8\linewidth}{!}{\input b-transform.pdf_t}
\caption{Schematic for $\mathcal{B}$-transform}
\label{fig:b_transform}
\end{figure*}
\begin{definition}[$\mathcal{B}$-transform]\label{defn:bb-transform}
A (non-uniform) $\mathcal{B}^n_{\delta,\epsilon}$-transform for $n$-qubit systems is
a (non-uniform) procedure for extending an $n$-qubit QS $\mathcal{Q}_1
= \big\langle \ket{\psi}, C, \mathcal{P} \big\rangle$ to a (possibly larger) QS $\mathcal{Q}_2 = \big\langle \ket{\psi'}, C', \mathcal{P}' \big\rangle$
whose components are black-box extensions of
the components of $\mathcal{Q}_1$
\begin{itemize}
\item The input in $\mathcal{Q}_2$ is an extension of the input in $\mathcal{Q}_1$
supplemented by ancill\ae\ qubits initialized to a fixed
state (wlog. in state $\ket{\mathbf{0}}$), i.e.,
$\ket{\psi'}=\ket{\psi} \otimes \ket{00\cdots 00}$
\item The projection operator in $\mathcal{Q}_2$ is an extension of the projection
operator in $\mathcal{Q}_1$ to include measurement of the ancill\ae\ in a basis
independent of $\mathcal{Q}_1$, i.e., $\mathcal{P}' = \mathcal{P} \otimes \mathcal{P}_a$.
\item The number of ancill\ae\ and the operator $\mathcal{P}_a$ are independent of $\mathcal{Q}_1$ and depend upon $\delta, \epsilon$.
\item The circuit in $\mathcal{Q}_2$ calls $C$ and $C^\dagger$ and uses
additional gates that depend upon $\delta$ and $\epsilon$.
\item $C'$ may also use gates that depend upon $\mathcal{P}_E$ and
$\ket{\psi}$.
\end{itemize}
\end{definition}
We call the transformations that satisfy the final condition as ``non-uniform'' since the transformed circuit
could be using gates that depend upon the input states and measurement operators
of the respective quantum system.
Note that the non-uniformity is not with respect to $n$, the number of qubits of
the quantum system, but with respect to the gates of the transformed
circuit. It will be clear from the proof of
Theorem~\ref{theorem:main-separable} that the
transformations that will be used in this paper are anyway uniform in $n$. In any
case, we will always drop $n$ from the superscript of $\mathcal{B}^n_{\delta,\epsilon}$.
We will revisit the notion of non-uniformity in Section~\ref{section:uniform}.
We want transformed quantum circuits that solve the quantum distinguishing problem without any error which
motives the next definition.
\begin{definition}\label{defn:solution}
For a set of quantum systems $\S = \{\mathcal{Q}_1, \mathcal{Q}_2, \ldots \}$ with output distributions
either $\mu_p$ or $\mu_q$ (for $p < q$), a
$\mathcal{B}$-transform $\mathcal{B}$ is said to solve $QD(\S)$ with error
$(\delta,\epsilon)$, in other words $\mathcal{B}$ is a
$(\delta,\epsilon)$-solution of $QD(\S)$, if the following holds for some
$\delta < \epsilon$ and all $\mathcal{Q}
\in \S$.
\begin{itemize}
\item If $\mu(\mathcal{Q}) = \mu_p$, then outcome of $\mathcal{B}(\mathcal{Q})$ is $E$ with
probability $\delta$.
\item If $\mu(\mathcal{Q}) = \mu_q$, then outcome of $\mathcal{B}(\mathcal{Q})$ is $E$ with
probability $\epsilon$.
\end{itemize}
$QD(\S)$ is said to have a {\em perfect solution} if $\mathcal{B}$ is a
$(0,1)$-solution of $QD(\S)$.
\end{definition}
It can be seen that the identity $\mathcal{B}$-transform is a trivial solution of the
above $QD(\S)$ with error $(p,q)$.
The last part of the above definition is based on the fact that if $\mathcal{B}$ is a
$(0,1)$-solution of $QD(\S)$, then the outcome of $\mathcal{Q}' = \mathcal{B}(\mathcal{Q})$ can be used to
correctly infer the output distribution of any given instance $\mathcal{Q} \in \S$.
Let $\mathcal{Q}' = \mathcal{B}(\mathcal{Q})$ -- which is essentially an extension of the input of $\mathcal{Q}$
with some ancill\ae, an extension of its measurement operator and a circuit that
can call the circuits of $\mathcal{Q}$ (and its inverse) in a black-box manner.
If the output distribution of $\mathcal{Q}$ is $\mu_p$, then the outcome of
$\mathcal{Q}'$ is never $E$ and otherwise (i.e., if the output
distribution of $\mathcal{Q}$ is $\mu_q$) the outcome of $\mathcal{Q}'$ is always $E$ without any error.
The main theorem of our work is stated next.
\begin{theorem}\label{theorem:main-separable}
Let $\S = \{\mathcal{Q}_1, \mathcal{Q}_2, \ldots \}$ be a collection of quantum systems such
that output distribution of any $\mathcal{Q}_i \in \S$ is either $\mu_\delta$
or $\mu_\epsilon$ for some $\delta < \epsilon$.
Then $\S$ is perfectly-solvable via some $\mathcal{B}$-transition
$\mathcal{B}_{\delta,\epsilon}$, i.e., any $\mathcal{Q}_i \in \S$ can be transformed by $\mathcal{B}_{\delta,\epsilon}$ to some
$\mathcal{Q}'_i$ such that:
\begin{itemize}
\item if output distribution of $\mathcal{Q}_i$ is $\mu_\delta$, then outcome of
$\mathcal{Q}'_i$ is never $E$ and
\item if output distribution of $\mathcal{Q}_i$ is $\mu_\epsilon$, then outcome of
$\mathcal{Q}'_i$ is always $E$.
\end{itemize}
\end{theorem}
The proof of this theorem is presented in the next section.
Note that, unlike Theorem~\ref{theorem:bhmt} which only applies to one-sided error
algorithms, we prove that two-sided error algorithms can also be ``amplified to
certainty''.
A straight-forward application of
this is to exactly distinguish between two QS with known output
distributions, such as Theorem~\ref{theorem:application}
(Section~\ref{section:intro}).
\begin{proof}[Proof of Theorem~\ref{theorem:application}]
Consider the transformation $\mathcal{B}_{\delta,\epsilon}^n$ from
Theorem~\ref{theorem:main-separable}. Given an $n$-qubit $\mathcal{Q} = \langle \ket{\psi}, C, \mathcal{P}
\rangle$, construct the transformed QS $\mathcal{B}^n_{\delta,\epsilon}(\mathcal{Q})=\big\langle \ket{\psi}\otimes \ket{\mathbf{00\ldots
0}}, C', \mathcal{P} \otimes \mathcal{P}_a \big\rangle$.
By Theorem~\ref{theorem:main-separable}, the output state of the transformed
circuit $C'$, when given
$\ket{\psi}$ (along with a few ancill\ae\ in a fixed state), upon measurement by
a simple extension of $\mathcal{P}$, has outcome either $E$ or $F$, depending upon
whether $\mu(\mathcal{Q})=\mu_\delta$ or $\mu(\mathcal{Q})=\mu_\epsilon$.
\end{proof}
\section{Proof of Theorem~\ref{theorem:main-separable}}
\label{section:proof}
We first state and prove our main technical tool -- the {\em Separability
Lemma} which essentially amplifies amplitudes of one-sided error algorithms. The
Lemma can be proven using already known techniques of amplitude amplifications
(e.g., see \cite[Sec 2.1]{BHMT}).
We give an alternative recursive construction that is optimized
towards amplifying fixed probabilities.
We use the following notation for the sake of brevity. Given a collection of
quantum systems $\{\mathcal{Q}_1, \mathcal{Q}_2, \ldots \}$ (such collections will be always
denoted by $\S$), we say that $\S$ is
$(\delta,\epsilon)$-separable (for some $\delta <\epsilon$) if output
distribution of any $\mathcal{Q}_i$ in $\S$ is either $\mu_\delta$ or $\mu_\epsilon$.
\begin{lemma}\label{lemma:fully-separable}[Separability]
For $\delta < \epsilon < 1$ and a collection of quantum systems $\S_1$ which
is $(\delta,\epsilon)$-separable, there is a $\mathcal{B}$-transform $\mathcal{B}_\epsilon$ which converts
$\S_1$ to a $(\delta',1)$-separable collection of quantum systems (for some $\delta \le \delta' < 1$). Additionally, $\delta=\delta'=0$ if and only if $\delta=0$.
\end{lemma}
Given an instance $\mathcal{Q} = \langle \ket{\psi}, C, \mathcal{P} \rangle$ of some $\mathcal{Q}_i \in \S_1$, Lemma~\ref{lemma:fully-separable} gives us a way to
determine whether the distribution of $\mathcal{Q}$ is $\pair{0}{1}$ or
$\pair{\epsilon}{1-\epsilon}$ by first transforming $\mathcal{Q}$ to $\mathcal{B}(\mathcal{Q}) = \mathcal{Q}' =
\langle \ket{\psi'}, C', \mathcal{P}' \rangle$ and then measuring the output of $C'$
on $\ket{\psi'}$ (which is a simple extension of the original input state)
using measurement operator $\mathcal{P}'$ (which is also a simple extension of the
original measurement operator).
\subsection{Grover iterator}
\label{subsection:grover}
As is usual in all analysis of amplitude amplification, the main operator to
study is the Grover iterator~\cite{Grover1996,BHMT}. Suppose we have a circuit $C$ acting
on an input state $\ket{\psi}$ and supposed to be measured using a two-output
projective measurement operator $\mathcal{P}=\langle P_E, I-P_E \rangle$. We consider a generalized version,
similar to the one studied by H{\o}yer~\cite{Hoyer2000}: $G(C,\ket{\psi}, \mathcal{P}, \theta,\alpha)
= C S_{\ket{\psi}} C^\dagger S_\mathcal{P} C$ using these additional gates:
$S_{\ket{\psi}} = I - (1-e^{\imath\theta})\ketbra{\psi}{\psi}$ and $S_\mathcal{P} = I -
(1-e^{\imath\alpha})P_E$.
Let $\ket{\psi'}=C\ket{\psi}$ denote the output state, $\ket{\psi_E} = P_E
\ket{\psi'}$
and $p$ denote $\braket{\psi_E}{\psi_E}$ -- the probability of measuring outcome
$E$ for this output state.
It is easy to see that $C S_{\ket{\psi}} C^\dagger
= I - (1-e^{\imath\theta})\ketbra{\psi'}{\psi'}$ and
$S_\mathcal{P} C \ket{\psi} = \big(I - (1-e^{\imath\alpha})P_E\big)\ket{\psi'}$.
One can then compute $\ket{\psi''}=G \ket{\psi}$ as
$\big( e^{\imath\theta} +
(1-e^{\imath\alpha})(1-e^{\imath\theta})p \big) \ket{\psi'} -
(1-e^{\imath\alpha})\ket{\psi_E}$
and $P_E \ket{\psi''} = \big( e^{\imath\theta} + e^{\imath\alpha} - 1 +
(1-e^{\imath\alpha})(1-e^{\imath\theta})p \big) \ket{\psi_E}$.
We get the following lemma summarizing the relative increase in probability after
one application of our Grover iterator.
We will use $p'(\theta,\alpha,p)$ to denote the new probability of
measuring outcome $E$ on the output state
after applying $G$ on input $\ket{\psi}$.
\begin{lemma}\label{lemma:grover-iterator}
Given a quantum system $\mathcal{Q}_1 = \langle \ket{\psi}, C, \mathcal{P} \rangle$ and
$\alpha,\theta \in [0,\pi]$, let $G$ be
the circuit for the Grover iterator $G(C,\ket{\psi}, \mathcal{P}, \theta,\alpha) =
C S_{\ket{\psi}} C^\dagger S_\mathcal{P} C$. If $p$ denotes the probability of observing
outcome $E$ for $\mathcal{Q}_1$ and $p'$ denotes the same probability for the QS $\langle
\ket{\psi}, G, \mathcal{P} \rangle$, then $p' = p\Delta$ where $\Delta = \left|\big( e^{\imath\theta} + e^{\imath\alpha} - 1 +
(1-e^{\imath\alpha})(1-e^{\imath\theta})p \big)\right|^2$.
\end{lemma}
First, $p=0$ if and only if $p'=0$ which means amplification has no effect on
impossible outcomes.
On the other hand, if $p > 0$, $p'$ is maximized when $\theta=\alpha$;
it can be shown that
$\Delta = \big( (1-2p)\cos\theta - 2(1-p) \big)^2 +
\sin^2\theta$ in that case.
We will use $\Delta^*_p$ to denote the maximum value of $\Delta$ for any $p$ and
using optimal $\theta$ and $\alpha$. The corresponding {\em optimal Grover iterator} will be
denoted as $G^*_p(C,\ket{\psi}, \mathcal{P})$; note that $G^*$ increases the probability from
$p$ to $p' = p \Delta^*_p$. Table~\ref{table:delta} summarizes the optimum value
of $p'$ and the relative increase for different possible values of initial
probability $p$.
Details of the relevant calculations are given in Appendix.
The following definition and corollary essentially describes the optimum
$\mathcal{B}$-transform.
\begin{definition}[Optimal $\mathcal{B}$-transform]\label{defn:opt-bb-transform}
$\mathcal{B}_{p}^* : \big\langle \ket{\psi}, C, \mathcal{P} \big\rangle \longrightarrow
\big\langle \ket{\psi}, G^*_p\big(C, \ket{\psi}, \mathcal{P} \big), \mathcal{P}
\big\rangle$
\end{definition}
\begin{corollary}\label{cor:opt-bb-transform}
If the output distribution of a QS $\mathcal{Q}$ is
$\mu_\epsilon$, then the output distribution of $\mathcal{B}^*_\epsilon(\mathcal{Q})$ is $\langle \epsilon \Delta^*_\epsilon,
1-\epsilon\Delta^*_\epsilon \rangle$. On the other hand, if the output distribution is
$\mu_\delta$ (for some $\delta < \epsilon$), then the output distribution of $\mathcal{B}^*_\epsilon(\mathcal{Q})$
is $\langle \delta', 1-\delta' \rangle$ for some $\delta' \ge \delta$ which
can be computed using $\delta$ and $\epsilon$. Furthermore, $\delta =
\delta'$ if and only if $\delta = 0$ (in which case, $\delta' = 0$).
\end{corollary}
\begin{table}
\begin{tabular}{|c|c|c|c|}
\hline
\parbox[c][.7cm]{2.5cm}{\centering Range of initial probability $p$} & Optimum $\alpha=\theta$ & Relative increase
$\frac{p'}{p}=\Delta^*_p$ & \parbox[c][1cm]{3cm}{\centering Amplified probability $p' = p \Delta^*_p$}\\
\hline
$p = 0.5$ & $\pi/2$ & 2 & 1 \\
\hline
$0.25 \le p \le 0.5$ & $\arccos \left( 1-\frac{1}{2p} \right)$ &
$\frac{1}{p}$ & 1\\
\hline
$p \le 0.25$ & $\pi$ & $(3-4p)^2 \ge 4$ & $p(3-4p)^2 \ge 4p$ \\
\hline
\end{tabular}
\caption{Optimum Grover iterator for different values of initial
probability\label{table:delta}}
\end{table}
In the next few subsections, we prove Separability
Lemma for different values of $\epsilon$.
\subsection{$\mathcal{B}_\epsilon$ for $\epsilon \in [1/4,1/2]$}\label{subsection:bb-half}
This is the simplest of all cases, to $\mathcal{B}$-transform
$(\delta,\epsilon)$-separable $\S_1$ to a $(\delta',1)$-separable one, for any
$1/4 \le \epsilon \le 1/2$ and for some
$\delta \le \delta'$. We can clearly use $\mathcal{B}_{\epsilon} = \mathcal{B}^*_{\epsilon}$ defined in
Definition~\ref{defn:opt-bb-transform}.
Separability Lemma immediately follows from Corollary~\ref{cor:opt-bb-transform} and
Table~\ref{table:delta}.
\subsection{$\mathcal{B}_{\epsilon}$ for $\epsilon > \frac{1}{2}$}\label{subsection:bb-half+}
We use the idea proposed by Brassard et al.~\cite{BHMT} to first convert
$\S_1$ to a $(\delta',\frac{1}{2})$-separable $\S_2$; let $\mathcal{B}^+_\epsilon$ denote
this transformation which is illustrated in Equation~\ref{eqn:bb-half+}.
This involves an
additional qubit in state $\ket{0}$ and an additional projective operator
$\mathcal{P}_\epsilon = \langle P_\epsilon^0, I - P_\epsilon^0 \rangle$,
where, $$P_\epsilon^0 = \tfrac{1}{2\epsilon}\ketbra{0}{0} +
\sqrt{1-\tfrac{1}{2\epsilon}}\sqrt{\tfrac{1}{2\epsilon}}\ketbra{1}{0} +
\sqrt{1-\tfrac{1}{2\epsilon}}\sqrt{\tfrac{1}{2\epsilon}}\ketbra{0}{1} +
\left(1-\tfrac{1}{2\epsilon}\right) \ketbra{1}{1}$$
Then we convert $\S_2$ to a $(\delta'',1)$-separable $\S_3$ by using
$\mathcal{B}_{\frac{1}{2}}$ (see Subsection~\ref{subsection:bb-half}). Combining both of these, we
propose the following transformation for $\mathcal{B}_\epsilon$. Here $\mathcal{P}'$ denotes
$\mathcal{P} \otimes \mathcal{P}_\epsilon$.
\begin{align}
\big\langle \ket{\psi}, C, \mathcal{P} \big\rangle &
\stackrel{\mathcal{B}^+_\epsilon}{\longrightarrow} \big \langle \ket{\psi} \otimes \ket{0}, C \otimes I, \mathcal{P}' \big\rangle
\stackrel{\mathcal{B}_{\frac{1}{2}}}{\longrightarrow} \big \langle \ket{\psi} \otimes \ket{0}, G^*_{1/2}\big(C \otimes I,
\ket{\psi}\otimes \ket{0}, \mathcal{P}' \big), \mathcal{P}' \big\rangle
\label{eqn:bb-half+}
\end{align}
\begin{proof}[Proof of Separability Lemma:]
The transformation from $\S_2$ to $\S_3$ was shown to be correct in
Subsection~\ref{subsection:bb-half}. Correctness of
$\mathcal{B}^+_\epsilon$ follows from the fact that the probability of measuring
outcome $0$ on the state $\ket{0}$ is $\frac{1}{2\epsilon}$ (since $\frac{1}{2} <
\epsilon \le 1$, $\frac{1}{2} \le \frac{1}{2\epsilon} < 1$). Let $p$ denote the probability of measuring outcome $E$
for some $\mathcal{Q} = \langle \ket{\psi}, C, \mathcal{P} \big\rangle \in \S_1$ and let
$p'$ denote the same
probability for the QS $\langle \ket{\psi} \otimes \ket{0}, C \otimes I, \mathcal{P}
\otimes \mathcal{P}_\epsilon \rangle$ of $\S_2$. Observe that, if $p=0$, then
$p'=0$; furthermore, if $p = \epsilon > \frac{1}{2}$, then $p'=\epsilon
\frac{1}{2\epsilon} = \frac{1}{2}$. Of course, the transformation does not
depend upon $\delta$.
\end{proof}
\subsection{$\mathcal{B}_{\epsilon}$ for $\epsilon < \frac{1}{4}$}\label{subsection:bb-small}
To transform $(\delta,\epsilon)$-separable $\S_1$ to $(\delta',1)$-separable one,
we first repeatedly apply the optimum Grover iterator enough number of times to
amplify $\epsilon$ beyond $\frac{1}{4}$ and then apply a suitable
$\mathcal{B}_{\epsilon_k}$ from Subsection~\ref{subsection:bb-half}.
Suppose $\epsilon < 1/4$.
Let $\epsilon_1 = \epsilon \Delta^*_\epsilon$, $\epsilon_2
= \epsilon_1 \Delta^*_{\epsilon_1}$, $\epsilon_3 = \epsilon_2
\Delta^*_{\epsilon_2}, \cdots$. Let $k$ be the smallest integer such that
$\epsilon_k \ge 1/4$; clearly, $\epsilon_1, \ldots, \epsilon_{k-1} < 1/4$ and
$\epsilon_k \in [1/4,1/2]$.
We define $\mathcal{B}_\epsilon$ as the $k$ transformations $\mathcal{B}_{\epsilon}^*,
\mathcal{B}^*_{\epsilon_1}, \mathcal{B}^*_{\epsilon_2}, \ldots \mathcal{B}^*_{\epsilon_{k-1}}$ applied successively and then
followed by $\mathcal{B}_{\epsilon_k}$.
\begin{align*}
\mathcal{B}_\epsilon: \langle \ket{\psi}, C, \mathcal{P} \rangle & \substack{~\mathcal{B}^*_\epsilon~\\\longrightarrow} \langle
\ket{\psi}, C_1, \mathcal{P} \rangle & \mbox{output
dist.}=\pair{\epsilon_1}{1-\epsilon_1} \mbox{ \& } C_1 = G^*_\epsilon(C,
\ket{\psi},\mathcal{P})\\
& \substack{~\mathcal{B}^*_{\epsilon_1}~\\\longrightarrow} \langle \ket{\psi}, C_2,
\mathcal{P} \rangle & \mbox{output dist.}=\pair{\epsilon_2}{1-\epsilon_2}
\mbox{ \& } C_2 = G^*_{\epsilon_1}(C_1,
\ket{\psi},\mathcal{P})\\
& \substack{~\mathcal{B}^*_{\epsilon_2}~\\\longrightarrow} \cdots & \dots \\
& \substack{~\mathcal{B}^*_{\epsilon_{k-1}}~\\\longrightarrow} \langle \ket{\psi}, C_k,
\mathcal{P} \rangle & \mbox{output dist.}=\pair{\epsilon_k}{1-\epsilon_k}
\mbox{ \& } C_{k} = G^*_{\epsilon_{k-1}}(C_{k-1},
\ket{\psi},\mathcal{P})\\
& ~\substack{\mathcal{B}_{\epsilon_k}\\\longrightarrow} \langle
\ket{\psi'}, C_{k+1}, \mathcal{P}' \rangle &
\end{align*}
\begin{proof}[Proof of Separability Lemma:]
Satisfiability Lemma is easily proved by observing that $\epsilon_k \in
[1/4,1/2]$ and so, applying $\mathcal{B}_{\epsilon_k}$ (from
Subsection~\ref{subsection:bb-half}) at the last step ensures that the final QS
has output distribution $\pair{1}{0}$.
It is also easy to check that these output distributions remain unchanged if and only
if $\delta = 0$.
\end{proof}
\subsection{Performance Evaluation}
Even though we propose a recursive approach to reduce error-probability of exact
error quantum systems, we show that our approach is essentially same as the
existing iterative approaches for amplitude amplification in terms of the number
of calls to $C$ and $C^\dagger$.
Take any quantum system $QS =
\langle \ket{\psi}, C, \mathcal{P} \rangle$. The existing approaches~\cite{BHMT,Hoyer2000} repeatedly apply the iterative Grover
operator
$\mathcal{Q} = (C S_{\ket{\psi}} C^\dagger S_\mathcal{P})$ (generalized to act on input encoded as the initial state and output
state to be measured by any projective operator) on $C \ket{\psi}$. Here $S_{\ket{\psi}}$ and $S_\mathcal{P}$
modify the phase of certain states by $\theta=\alpha=\pi$ as specified in
Subsection~\ref{subsection:grover}.
Let $\epsilon$ denote the probability of
observing outcome $E$; let $\beta \in [0,\pi/2]$ be such that $\sin^2 \beta =
\epsilon$. Then, the probability of observing $E$ on repeated applications of
$\mathcal{Q}$, say $b$ times, on $C \ket{\psi}$ (i.e., on the output state of $\mathcal{Q}^b C
\ket{\psi}$) can be shown to be $\sin^2 \big( (2b+1)\beta \big)$.
As shown in Table~\ref{table:delta}, suitably choice of phases in
$S_{\ket{\psi}}$ and $S_\mathcal{P}$ can amplify any $\epsilon \in [0.25,1]$ to 1
using a $\mathcal{B}$-transform that effectively corresponds to one application of
$\mathcal{Q}$ on $C\ket{\psi}$. So, if $\epsilon \ge 0.25$, our recursive method and the
iterative approach are identical.
So, we will now analyze $\mathcal{B}_{\epsilon}$ for $\epsilon < 0.25$, in fact, for $\epsilon
\ll 0.25$. Let $k$ be the number of $\mathcal{B}^*$-transforms required.
Recall from Subsection~\ref{subsection:bb-small} that $\mathcal{B}_\epsilon$ keeps the input
and the projective operator unchanged and converts $C$ to some $C_{k+1}$ via
intermediate circuits $C_1, C_2, \ldots, C_k$ where
$C_{j+1} = G^*_{\epsilon_j}(C_j,
\ket{\psi}, \mathcal{P})$ for $\epsilon < \epsilon_1 < \ldots < \epsilon_k \in
[1/4,1/2]$. The $S_{\ket{\psi}}$ and $S_\mathcal{P}$ operators in those $G^*$ are defined
using phases $\theta=\alpha=\pi$ as per Table~\ref{table:delta}.
\begin{restatable}{lemma}{ckinduction}\label{lemma:c_k_induction}
For any $j \in [1,k]$, $C_j = \mathcal{Q}^{\frac{3^j-1}{2}} C$.
\end{restatable}
This lemma can be easily proved by induction on $k$ (see Appendix).
It shows that the final circuit obtained by our recursive approach is identical
to that obtained by apply a fixed $\mathcal{Q}$ a certain number of times.
Therefore, $\epsilon_{k} = \sin^2 \left( 3^k \beta
\right)$ which must be at least $1/4$. This stipulates that $k \ge \log_3
\frac{\pi}{6\beta}$. The total number of calls to $C$ and $C^\dagger$ made by
our recursive algorithm to amplify $\epsilon < 0.25$ to some $\epsilon_k > 0.25$
can then be easily shown to be $1+\frac{\pi}{3\beta}$ (rather, the next higher
integer) -- which is exactly the same as that in $\mathcal{Q}^{(3^k-1)/2}C$.
\subsection{Proof of Theorem~\ref{theorem:main-separable}}
We are now ready to prove Theorem~\ref{theorem:main-separable} using
Separability Lemma.
We will use the following notation. If $\mathcal{B}$ is a transformation for a set of
quantum systems $\S$, then the set of {\em transformed quantum systems} after
applying $\mathcal{B}$ will be denoted by $\mathcal{B}(\S)$.
\begin{proof}
The given $\S$ in the theorem is $(\delta,\epsilon)$-separable. Our required
$\mathcal{B}_{\delta,\epsilon}$ will be composed of a series of $\mathcal{B}$-transforms:
$\mathcal{B}_\epsilon$, $\mathcal{B}_2$ and $\mathcal{B}_\delta$.
$\mathcal{B}_\epsilon$ is chosen such so as to solve $QD(\S)$ with error $(\delta',1)$ for
some $\delta < \delta'$. This step can skipped ($\mathcal{B}_\epsilon$ can be set to
identity) if $\epsilon = 1$; on the other hand, if $\epsilon < 1$,
we can use $\mathcal{B}_\epsilon$ from Lemma~\ref{lemma:fully-separable}, which
implies that $\mathcal{B}_\epsilon(\S)$ is
$(\delta',1)$-separable for some $\delta'$ (that depends on $\delta$ and
$\epsilon$). Let $\S_1$ denote $\mathcal{B}_\epsilon(\S)$.
$\mathcal{B}_2$ is the following transform: $\big\langle \ket{\psi}, C, (P_1, P_2)
\big\rangle \longrightarrow \big\langle \ket{\psi}, C, (P_2, P_1) \big\rangle$.
Let $\S_2 = \mathcal{B}_2(\S_1)$. Any $QS \in \S_1$ with $\mu(QS)=\mu_{\delta'}$ is
transformed to $QS' \in \S_2$ with $\mu(QS') = 1-\delta'$ and similarly, if
$\mu(QS) = \mu_1$, then $\mu(QS') = \mu_0$. Therefore, $\S_2$ is $(0,1-\delta')$-separable.
By property of $\mathcal{B}_\epsilon$, $\delta=\delta'=0$ if and only if $\delta=0$ and in
that case, we have obtained $(0,1)$-separable $\S_2$. On the other hand,
if $\delta > 0$, then $\delta' > 0$. Let $\delta''$ denote
$1-\delta'$.
Since $\S_2$ is $(0,\delta'')$-separable, apply
Lemma~\ref{lemma:fully-separable} again to get $\mathcal{B}_{\delta}$ such that
$\S' = \mathcal{B}_{\delta}(\S_2)$ is $(0,1)$-separable.
Our required transform $\mathcal{B}$ is a sequential application of $\mathcal{B}_\epsilon$ followed
by $\mathcal{B}_2$ followed by $\mathcal{B}_{\delta}$. As explained above,
$\mathcal{B}_\delta(\mathcal{B}_2(\mathcal{B}_\epsilon(\cdot)))$ is a $(0,1)$-solution of $QD(\S)$.
\end{proof}
\section{Uniform version of
Theorem~\ref{theorem:main-separable}}\label{section:uniform}
The non-uniformity in Definition~\ref{defn:bb-transform} is not very helpful if
we wish to obtain a true black-box extension of a quantum system $\mathcal{Q} = \big\langle
\ket{\psi},C,\mathcal{P} \big\rangle$.
Note that the extension to the input qubits and the
extension to the projective measurement operator is anyway independent of $\mathcal{Q}$
and $n$, the gates in $C'$ are uniform in $n$,
and furthermore, the transformed circuit $C'$ is allowed to call the original
circuit $C$ (and its inverse $C^\dagger$) in a black-box manner; however, some
of the gates in $C'$ may additionally depend upon $\ket{\psi}$ and operators of
$\mathcal{P}$. It would be really good to obtain a more uniform conversion which
necessitates the following definition.
\begin{definition}[Uniform $\mathcal{B}$-transform]
A $\mathcal{B}$-transform for converting multiple QS $\{\mathcal{Q}_1, \mathcal{Q}_2, \ldots \}$ is said
to be {\em uniform} if the circuit of $\mathcal{B}(\mathcal{Q}_i)$ is identical for all source $\mathcal{Q}_i$ except for the
calls to $C$ and $C^\dagger$ corresponding to $\mathcal{Q}_i$.
\end{definition}
\subsection{Uniform Grover iterator}
We want to study some
sufficient conditions for the $\mathcal{B}$-transforms to be uniform by constructing a
uniform version of Grover iterator.
Since Grover iterator uses $\S_\mathcal{P}$, it is crucial to have identical measurement operators for
all quantum systems. This is, however, not such a major requirement since it is always
possible to change measurement operators by extending a quantum circuit with
suitable operators.
Except the gates $S_{\ket{\psi}} = I - (1-e^{\imath\theta})\ketbra{\psi}{\psi}$ which depend upon the
corresponding input to the circuit ($\ket{\psi}$), none of the other gates
used in $\mathcal{B}$-transforms that are involved in the
proof of Theorem~\ref{theorem:main-separable} depend upon the input
state (see Section~\ref{section:proof}).
However, a $\mathcal{B}$-transform may still become uniform if all the inputs in $\S_1$,
and hence all such $S_{\ket{\psi}}$ gates, will be identical.
Now consider a second option -- all measurement operators are identical and all the input states are not identical but
they form an orthonormal set. We show that it is still possible to
apply $S_{\ket{\psi}}$ in a uniform manner. Recall that this gate changes the
phase of any state depending upon whether it is $\ket{\psi}$ or not and the main
difficulty appears to be the fact that the input state cannot be copied and
stored for a later application of the conditional phase gate. So our main
idea is to convert $\ket{\psi}$ to some state in the standard basis since
it is possible to copy and store states in the standard basis using the {\em quantum
fanout gate}~\cite{Durr1999}. This gate copies a standard basis state to
another register: $F_m \ket{x_1 \ldots x_m}\ket{b_1 \ldots b_m} = \ket{x_1 \ldots
x_m}\ket{(x_1 \oplus b_1) \ldots (x_m \oplus b_m)}$ for $x_1 \ldots x_m \in
\{0,1\}^m$ and $b_1 \ldots b_m \in \{0,1\}^m$ shows the operation for
``copying'' $m$-qubits.
\begin{figure}[!htb]
\begin{subfigure}[b]{1.2in}
{\raisebox{8mm}{\resizebox{1.2in}{!}{\input S-psi-basis.pdf_t}}}%
\caption{$S_{\ket{\psi}}$\label{fig:s-psi-uniform-a}}
\end{subfigure}%
~
\begin{subfigure}[b]{3in}
{\resizebox{3in}{!}{\input S-psi-basis-implement.pdf_t}}%
\caption{Uniform $S_{\ket{\psi}}$\label{fig:s-psi-uniform-b}} %
\end{subfigure}%
~
\begin{subfigure}[b]{1.2in}
{\raisebox{1cm}{\resizebox{1.2in}{!}{\input sv.pdf_t}}}%
\caption{$S_\theta$ operator \label{fig:s-psi-uniform-c}}
\end{subfigure}
\caption{\label{fig:s-psi-uniform} Applying operator $S_{\ket{\psi}}$ in a
uniform manner. Figure~\ref{fig:s-psi-uniform-a} shows the non-uniform
operator and Figure~\ref{fig:s-psi-uniform-b} shows its uniform version
(dotted box on the left shows initialization and dotted box on the right
shows $S_{\ket{\psi}}$ being applied uniformly).
Figure~\ref{fig:s-psi-uniform-c} shows the $S_\theta$ operator from
Figure~\ref{fig:s-psi-uniform-b}.}
\end{figure}
See Figure~\ref{fig:s-psi-uniform} for a uniform version of $\S_{\ket{\psi}}$.
Figure~\ref{fig:s-psi-uniform-a} shows $S_{\ket{\psi}}$ as a part of an arbitrary quantum
circuit, say $C$ that takes as input an $m$-qubit state $\ket{\psi}$ (and some
ancill\ae) and $\S_{\ket{\psi}}$, on $m$-qubits, is one of its gates
Since we are now considering the case that $C$ is applied only on orthogonal
input states (suppose denoted by $\ket{\psi_1}, \ket{\psi_2},
\cdots$), therefore, there exists a one to one mapping between these states
and a subset of the $m$-qubit standard basis states $\ket{1}, \ket{2}, \cdots$.
Let $U$ denote the unitary operator for the mapping, i.e., $U\ket{\psi_v} =
\ket{v}$.
Figure~\ref{fig:s-psi-uniform-b} illustrates a circuit $C'$ that applies $S_{\ket{\psi}}$ without
requiring a gate that explicitly depends upon $\ket{\psi}$. Apart from the two
registers of $C$ (the input $\ket{\psi}$ and ancill\ae\ qubits), $C'$ also uses
$m$ additional ancill\ae\ qubits in state $\ket{0}$. Other than the standard gates
($T$ stands for the unbounded fanout Toffoli and $X$ is the quantum NOT gate),
$C'$ uses three additional gates: $F_m$, $P_\theta$ and $S_\theta$. The $F_m$ gate is
the quantum fanout gate. $P_\theta$ changes phase
of $\ket{1}$ by $e^{\imath\theta}$: $P_\theta = I - (1-e^{\imath\theta})\ketbra{0}{0}$.
The $S_\theta$ gate uses
an additional reusable ancill\ae\ $\ket{0}$ and changes the phase by
$e^{\imath\theta}$ only for the state $\ket{0^m}$ (illustrated in
Figure~\ref{fig:s-psi-uniform-c}).
The state of the first two registers after the left dotted box in
Figure~\ref{fig:s-psi-uniform-b} is simply $\ket{0^m}\ket{\psi} \to
\ket{v}\ket{\psi}$ where $\ket{v}$ is the standard basis vector $U\ket{\psi}$.
We will next analyze the operator for the right dotted
box, say denoted by $U_R$. $S_\theta$ can be written as $I -
(1-e^{\imath\theta})\ketbra{0^m}{0^m}$ and the $F_m$ operator essentially behaves
like $F_m \ket{b_1 \ldots b_m} \to \ket{(v_1 \oplus b_1), \ldots (v_m
\oplus b_m)}$.
The following calculation (for only the qubits involved) shows that the operator for the right dotted box is
identical with $S_{\ket{\psi}}$.
\begin{align*}
U_R = & (I \otimes U^\dagger) F_m (I \otimes S_\theta) F_m (I \otimes U)
= (I \otimes U^\dagger) F_m \big(I \otimes (I -
(1-e^{\imath\theta})\ketbra{0^m}{0^m}) \big) F_m (I \otimes U)\\
= & (I \otimes U^\dagger) \big(I \otimes (I -
(1-e^{\imath\theta})\ketbra{v}{v}) \big) (I \otimes U)
= I \otimes (I - (1-e^{\imath\theta})\ketbra{\psi}{\psi}) = I \otimes
S_{\ket{\psi}}
\end{align*}
The results of this subsection can be summarized in the following lemma.
\begin{lemma}\label{lemma:uniform-bb-transform}
The $\mathcal{B}$-transform in Theorem~\ref{theorem:main-separable} can be made uniform if
all projection operators in the quantum systems of $\S$ are identical and
all input states in $\S$ are either identical or form an orthonormal set of
states.
\end{lemma}
\section{Distinguishing two circuits}\label{section:distinguish-circuits}
Suppose we are given a quantum circuit $C$ (as black-box) and two
different operators $C_1$ and
$C_2$, all acting on the same Hilbert space, and we are told that the operator for $C$ is either $C_1$ or $C_2$. We have to
determine $C$ corresponds to which one. We assume that we also have access to its inverse operator $C^\dagger$.
The analogous problem for deterministic (classical) functions is
trivial. Two distinct functions must differ at some input which can be
determined from their function descriptions (the problem is NP-hard
but we are not concerned about feasibility, not efficiency, for this discussion). The output of
$C$ on this input will identify whether $C$ is $C_1$ or $C_2$.
However, if $C$ is a randomized circuit or
algorithm, then except for a few trivial cases, the output of $C$
generates a sample distribution over the output of $C_1$ and $C_2$; the
question of determining the correct distribution of $\mathcal{C}$ without any error is believed
to be hard, if not impossible.
However, it is possible to give a positive answer to the same question for
quantum circuits.
Select a suitable $\ket{\phi}$ and compute the two possible output states
${\ket{\psi_1} = C_1\ket{\phi}}$, ${\ket{\psi_2} = C_2\ket{\phi}}$. Choose
projective operators $\mathcal{P}=\langle I - \ketbra{\psi_1}{\psi_1},
\ketbra{\psi_1}{\psi_1} \rangle$ with respective outcomes $E$ and $F$.
Consider these two quantum systems: $\langle \ket{\phi},
C_1, \mathcal{P} \big\rangle$ and $\langle \ket{\phi}, C_2, \mathcal{P} \big\rangle$.
The output distribution of the first QS is $\pair{0}{1}$ and that of the second
is $\pair{\epsilon}{1-\epsilon}$ where $\epsilon =
1-|\braket{\psi_1}{\psi_2}|^2 > 0$.
Now, Theorem \ref{theorem:main-separable} can be applied on the QS $\langle
\ket{\phi}, C, \mathcal{P} \rangle$ which essentially gives us a circuit $C'$ (that
calls $C$ and $C^\dagger$) along with suitably extended input and
measurement operators,
with the property that if the outcome of the QS is $E$, then $C$ is surely $C_1$ and
otherwise $C_2$.
It is perfectly okay to use any $\ket{\phi}$ as the input state; however, since
the size of $C'$ depends inversely upon $\epsilon$ so it makes sense to have the
largest possible $\epsilon$. A recent result~\cite{2015BeraATPG} can be used to
determine the optimum initial state (details of this is presented in the
Appendix).
\paragraph*{Single-fault detection}
Fault detection is a major step in the workflow of circuit
fabrication. It is common in research and industry to assume that practically
most faults appear according to a few known fault models.
A standard approach to detecting if a circuit is faulty is to generate a set of
test patterns (inputs) such that the output of a fault-free circuit would be
different from that of a faulty-circuit. This method is known as ATPG (automatic
test-pattern generation) and is well-studied for classical
circuits and very
recently, seeing use even for quantum circuits~\cite{paler2011tomographic}.
ATPG is computationally difficult being NP-hard~\cite{IbarraSahniATPG}, and even harder
for quantum circuits because the measurement output of these circuits is
probabilistic, and hence even a single test pattern will generate a distribution
over possible outcomes.
However, the technique described earlier in this section can come to our rescue
in the special case of only one fault model, i.e., given a circuit $C$ as a
black-box unit, we wish to determine if $C$ is fault-free (i.e., $C=C_1$) or $C$
is faulty (with fault model $C_2$). We can reliably answer this question without
any chance of error using the approach described above.
\section{Exact Error Algorithms}\label{section:exact-error-classes}
Usual probabilistic classes like \ensuremath{\mathbf{RP}}\xspace and \ensuremath{\mathbf{BPP}}\xspace are defined in terms of errors that are upper bounded
by constants. They are rarely defined in terms of exact error, primarily due to
the lack of robustness in definition that accompanies this concept. There is no known
technique to show that the class of problems with one-sided error exactly same as $0.3$
remains unchanged if the error is instead $0.301$.
Consider, for example, the simplified class \ensuremath{\mathbf{ERP}}\xspace ($\P \subseteq \ensuremath{\mathbf{ERP}}\xspace \subseteq
\ensuremath{\mathbf{RP}}\xspace$) whose problems have randomized algorithms similar to those
for \ensuremath{\mathbf{RP}}\xspace, but with an additional requirement that the error is same for
all ``no'' instances (of any length). We similarly define \ensuremath{\mathbf{EBPP}}\xspace as the class of
problems with exact two-sided error polymomial-time algorithms.
Based on what we know, $\P \not= \ensuremath{\mathbf{ERP}}\xspace \not= \ensuremath{\mathbf{EBPP}}\xspace$.
However, we were able to prove that the
quantum analogs of these classes have identical complexity
using our generalization of quantum amplitude amplification.
\begin{definition}
$\ensuremath{\mathbf{EBQP}}\xspace_{\delta,\epsilon}$ is the class of languages $L$ for which there
exists a uniform family of polynomial-size quantum circuits $\{C_n\}$, a uniform family of
states for $a_n$ ancill\ae\ qubits
$\ket{A_n}$ and a uniform family of two-outcome projective measurement operators
$\{\mathcal{P}_n\}$ such that $C_n$ and $\mathcal{P}_n$ act on a space of $n+a_n$ qubits and
the following hold for any $x \in \{0,1\}^n$, $\forall n$:
\begin{itemize}
\item if $x \not\in L$, then the output distribution of $\big\langle \ket{x}
\otimes \ket{A_n}, C_n, \mathcal{P}_n \big\rangle$ is $\mu_\delta$ (i.e., when
the output state of $C_{n}$ on input state $\ket{x} \otimes \ket{A_n}$ is measured
using $\mathcal{P}_n$, outcome $E$ is observed with probability $\delta$) and
\item if $x \in L$, then the output distribution of $\big\langle \ket{x}
\otimes \ket{A_n}, C_n, \mathcal{P}_n \big\rangle$ is $\mu_\epsilon$ (i.e.,
outcome $E$ is observed with probability $\epsilon$ upon similar
measurement as the above case).
\end{itemize}
\noindent$\ensuremath{\mathbf{ERQP}}\xspace_\epsilon$ is simply $\ensuremath{\mathbf{EBQP}}\xspace_{0,\epsilon}$.
Define $\displaystyle\ensuremath{\mathbf{EBQP}}\xspace = \bigcup_{\epsilon>\delta\ge 0}
\ensuremath{\mathbf{EBQP}}\xspace_{\delta,\epsilon}$ and $\displaystyle\ensuremath{\mathbf{ERQP}}\xspace =
\bigcup_{\epsilon>0} \ensuremath{\mathbf{ERQP}}\xspace_\epsilon$.
\end{definition}
Note that, unlike the usual definitions of probabilistic classes, for these
classes it is not even clear if the different classes
$\ensuremath{\mathbf{EBQP}}\xspace_{\delta,\epsilon}$ for different $\delta$ and $\epsilon$ are identical.
However, the following lemma is obvious from these definitions.
\begin{lemma}
$\ensuremath{\mathbf{EQP}}\xspace = \ensuremath{\mathbf{EBQP}}\xspace_{0,1} = \ensuremath{\mathbf{ERQP}}\xspace_1$ and
$\ensuremath{\mathbf{EQP}}\xspace \subseteq \ensuremath{\mathbf{ERQP}}\xspace \subseteq \ensuremath{\mathbf{EBQP}}\xspace$.
\end{lemma}
The main result of this section is a simple application of
Theorem~\ref{theorem:main-separable} and Lemma~\ref{lemma:uniform-bb-transform}.
\begin{theorem}\label{theorem:eqp-ebqp}
$\ensuremath{\mathbf{EQP}}\xspace = \ensuremath{\mathbf{ERQP}}\xspace = \ensuremath{\mathbf{EBQP}}\xspace$.
\end{theorem}
\begin{proof}
We essentially need to show that $\ensuremath{\mathbf{EBQP}}\xspace \subseteq \ensuremath{\mathbf{EQP}}\xspace$. To prove this we will
show that for any $L$, if $L \in
\ensuremath{\mathbf{EBQP}}\xspace_{\delta,\epsilon}$ (for any $\epsilon > \delta \ge 0$), then $L \in
\ensuremath{\mathbf{EBQP}}\xspace_{0,1}$.
Fix an arbitrary $n$.
For any binary string $x$ of length $n$, define the quantum system $\mathcal{Q}_x = \big\langle \ket{x} \otimes \ket{A_n}, C_n, \mathcal{P}_n \big\rangle$
where $\ket{A_n}$, $C_n$ and $\mathcal{P}_n$ are obtained from the definition of
$\ensuremath{\mathbf{EBQP}}\xspace_{\delta,\epsilon}$ and the fact that $L \in \ensuremath{\mathbf{EBQP}}\xspace_{\delta,\epsilon}$.
Now consider these sets of quantum systems $\S_n = \{\mathcal{Q}_x ~:~ x \in \{0,1\}^n \}$ for all $n
> 0$. Clearly, there are two possible output distributions of any $\S_n$, namely,
$\mu_\delta$ and $\mu_\epsilon$. Since the input states in $\S_n$ are
orthonormal and the projection operators therein are identical, we can therefore apply
Theorem~\ref{theorem:main-separable} and Lemma~\ref{lemma:uniform-bb-transform}
to obtain a uniform transformation $\mathcal{B}_{\delta,\epsilon}$ which
perfectly solves the problem of $QD(\S_n)$. Let $\mathcal{B}_{\delta,\epsilon}(\mathcal{Q}_x) =
\mathcal{Q}'_x = \big\langle \ket{x} \otimes\ket{A_n} \otimes \ket{00\ldots 0}, C'_n, \mathcal{P}'_n
\big\rangle$ which gives us (i) a circuit $C'_n$ which calls $C_n$ (and
$C_n^\dagger$) (ii) a two-outcome projective measurement operator $\mathcal{P}'_n$ and
a (iii) set of ancill\ae\ qubits in state $\ket{\mathbf{00\ldots 0}}$
such that the following holds for the outcome of $C'_n$ on $\ket{x} \otimes
\ket{A_n} \otimes \ket{00\ldots 0}$ when measured using $\mathcal{P}'_n$.
\begin{itemize}
\item If $x \not\in L$, then the output distribution of $\mathcal{Q}'_x$ is $\mu_0$,
i.e., the outcome is never $E$.
\item If $x \in L$, then the output distribution of $\mathcal{Q}'_x$ is $\mu_1$,
i.e., the outcome is always $E$.
\end{itemize}
Therefore, we get a uniform family of circuits $\{C'_n\}$, a uniform family of
ancill\ae\ qubits $\ket{A_n} \otimes \ket{\mathbf{00\ldots 0}}$ and a uniform family of two-outcome
projective measurement operator $\{\mathcal{P}'_n\}$ such that the outcome of
$C'_{|x|}$ on any $\ket{x}$, with additional ancill\ae\ qubits in a uniformly
generated state, when measured by $\mathcal{P}'_{|x|}$ indicates whether $x \in L$
without any probability of error.
Since $C_n'$ uses constantly many calls to $C_n$ and $C_n^\dagger$ along with other
gates (the constant depends only on
$\delta$ and $\epsilon$), this shows that $L \in \ensuremath{\mathbf{EBQP}}\xspace_{0,1}$.
\end{proof}
\begin{figure*}[!h]
\centering \resizebox{0.65\linewidth}{!}{\input thm2.pdf_t} \qquad %
\resizebox{0.25\linewidth}{!}{\input thm1-s0.pdf_t}
\caption{Circuit for $\mathcal{C}'$ (left) and $S_0$ gate in $\mathcal{C}'$ (right)}
\label{fig:erqp-half}
\end{figure*}
We illustrate an application of the above theorem to obtain an error-free
circuit for an $\ensuremath{\mathbf{ERQP}}\xspace_{1/2}$ language $L$ (see Appendix for an explicit proof). Consider circuit $C$ in
Figure~\ref{fig:erqp-half}(a) which can identify if $x \in L$ with one-sided
error $0.5$. As is typical in quantum circuits, in this example only one of the output qubits of
the circuit is measured in the standard basis ($P_E = \ketbra{0}{0} \otimes I$
and $\mathcal{P} = \langle P_E, 1-P_E \rangle$); therefore, if $x \not\in L$, then
the output qubit is never observed in state $\ket{0}$ and if $x \in L$, then the
output qubit is observed in states $\ket{0}$ or $\ket{1}$ with equal
probability. The circuit $C'$ shown in Figure~\ref{fig:erqp-half}(b) shows how to
remove the probability of error; the same output qubit is measured in the
standard basis for outcome and some additional qubits in state $\ket{0}$ are
used as ancill\ae. Apart from calling $C$ and $C^\dagger$, $C'$ uses the
n-qubit Fanout gate $F_n$, a
conditional phase gate $S_0$
\footnote{$S_0
\ket{0 0 \ldots 0} = \imath \ket{00\ldots 0}$ and for other states $S_0
\ket{x_1 x_2 \ldots x_k} = \ket{x_1 \ldots x_k}$ (illustrated in
Figure~\ref{fig:erqp-half}(c)).}
which changes phase of $\ket{00\ldots 0}$ by $\imath$, and $P$
does the same to $\ket{1}$.
\subsection{Exact amplitude amplification
(Theorem~\ref{theorem:aa-two-sided})}
\begin{proof}[Proof of Theorem~\ref{theorem:aa-two-sided}]
Let $\mathcal{P}$
denote the two-outcome projective measurement operator used in the original
two-sided exact error circuit $C$. $C$ can be of two types depending on how
it accesses its input. Any input $x \in X$ can be accessed
either through the input state $\ket{x}$ (along with ancill\ae\ initialized to $\ket{00\ldots}$, wlog.) or through an oracle gate $U_x:
\ket{x,b} \to \ket{x, b\oplus \Phi(x)}$ (for $b \in \{0,1\}$). If $C$ is of the
former type, then Theorem~\ref{theorem:aa-two-sided} is essentially same as
Theorem~\ref{theorem:eqp-ebqp}.
Next we focus on
circuits with oracle
gates. Let $C^{U_x}$ denote this circuit
when given
$U_x$ as the oracle gate corresponding to an input $x \in X$. The input state to
$C^{U_x}$ can be taken to be $\ket{00\ldots 0}$, wlog.
The proof follows by applying Theorem~\ref{theorem:main-separable} on this
collection of quantum systems: $\left\{ \big\langle
\ket{00\ldots}, C^{U_x}, \mathcal{P} \big\rangle ~:~ x \in X \right\}$.
Observe that this collection satisfies the conditions of
Lemma~\ref{lemma:uniform-bb-transform}. So, the corresponding $\mathcal{B}$-transform
is uniform which implies that all the transformed circuits
for these quantum systems are identical, except for the calls to
$C$ and $C^\dagger$. Therefore, we can choose this transformed oracle circuit as
our required $C'$ of Theorem \ref{theorem:aa-two-sided}.
\end{proof}
\section{Conclusion}
Is there a classical method that can accurately decide the distribution of a random
variable $X$ among two given distributions based on multiple samples of $X$?
Probably no.
On the
other hand, if the random variables come from a quantum source, we show that
quantum circuits exist that can do the same without any probability of error.
A quantum circuit, along with an input state and a measurement operator, can be
consider as a quantum source of samples drawn over the distribution of
the measurement outcomes.
The underlying technique is a generalization of quantum amplitude
amplification to two-sided error and for circuits without oracle gates. We used
our amplification technique to distinguish between two circuits, when used as a
black box, which has application in fault detection of quantum circuits. We also
defined a restricted version of quantum one-sided and two-sided bounded error
classes and used generalized amplification to show that those complexity
classes collapse to (error-free) quantum polynomial time complexity
class.
It would be interesting to investigate if this approach can
be used for ATPG with more than one fault models and for amplifying
standard bounded-error classes $\ensuremath{\mathbf{BQP}}\xspace$ and $\ensuremath{\mathbf{RQP}}\xspace$.
\include{sampling2-arxiv-bbl}
|
\section{Introduction}
In modern rational mechanics,
when describing a material body,
the general principles such as field equations or jump conditions
are separated from the so-called constitutive equations or material laws
(cf. \cite[pp. 1--2]{TrNo2004}).
Whereas the former are common among all representatives
of a major class of materials such as solids or fluids,
the latter are meant to uniquely characterize each particular material.
Typically, a material law describes the response of a material to various stimuli applied
and is modelled by an algebraic or an operator equation
(see, e.g., \cite[Chapter C3]{TrNo2004}).
In macroscopic theories such as the classical theory of (thermo)elasticity
or its modern generalizations and unifications such as the one proposed
by Green and Naghdi \cite{GrNa1995.I, GrNa1995.II, GrNa1995.III}, etc.,
the microscopic structure of the material is ignored
and the material law is obtained in the form of a stress-strain relation,
which can be measured experimentally.
Recently, Rajagopal \cite{Ra2003} made a strong case
for using implicit material laws, which are physically more sound
as they let the stress, i.e., the force,
induce the strain and not vice versa.
In anelastic bodies, e.g., the thermoelastic ones,
the entropy and the resulting irreversibility play a very important role.
In the theory of thermoelasticity,
a relation between the heat flux and the temperature gradient is postulated
such that the second theorem of thermodynamics,
usually in form of Clausius-Duhem inequality, is satisfied \cite[Sections 96 and 96 bis]{TrNo2004}.
Alternatively, an entropy balance equation can be used
as proposed by Green and Naghdi \cite{GrNa1995.II, GrNa1995.III}.
When being applied to various materials with microstructure,
the macroscopic theory of (thermo)elasticity
does not provide an adequate mechanical and thermodynamical description.
They include but are not limitted to
porous elastic media, micropolar elastic solids, materials with microstructure
and nonsimple elastic solids,
which were first introduced in the works of
Truesdell and Toupin \cite{TruTou1960}, Green and Rivlin \cite{GrRi1964}, Mindlin \cite{Mi1964} and Toupin \cite{Tou1964}.
Their common feature is that they incorprorate extra field variables
such as microstresses or microrotations, hyperstresses or volume fractions, etc.
For further details, we refer the reader
to comprehensive monographs by Ciarletta and Ie\c{s}an \cite{CiaIes1993}
and Ie\c{s}an \cite{Ies2004}.
We now briefly summarize the linear thermoelasticity model
for nonsimple materials without energy dissipation introduced by Quintanilla \cite{Quin1}.
To this end, consider a rigid body occupying in a reference configuration a bounded domain $\Omega$ of $\mathbb{R}^{d}$
with the Lipschitz boundary $\Gamma := \partial \Omega$.
For a period of time $t \geq 0$ and a material point $\mathbf{x} = (x_{i}) \in \Omega$,
let $\mathbf{u} = (u_{i})$, $(u_{iL})$ and $T$, being functions $\mathbf{x}$ and $t$,
denote the displacement vector,
the homogeneous deformation tensor of the `particle' with its center of mass located at $\mathbf{x}$
and the relative temperature measured with respect to a constant reference temperature $T_{0}$,
which is assumed to be attained at some time $t_{0} \geq 0$.
Further, we define the thermal displacement
\begin{equation}
\tau(\mathbf{x}, t) := \int_{t_{0}}^{t} T(\mathbf{x}, s) \mathrm{d}s. \notag
\end{equation}
With $\mathbf{t} = (t_{Kj})$ denoting the first Piola \& Kirchhoff stress tensor,
the linear balance of momentum reads as
\begin{equation}
\rho \ddot{u}_{i} = t_{Ki, K} + \rho f_{i}, \notag
\end{equation}
where $\rho$ stands for the material density and $\mathbf{f} = (f_{i})$ is the volumetric force.
Here and the sequel,
we employ the Einstein's summation convention
as well as the standard notation for temporal and spatial
derivatives of scalar and tensor fields (cf. \cite[Chapter 1]{Ies2004}).
In contrast to the classical linear theory of thermoelasticity,
which utilizes the (generalized) Hooke's law
to postulate a linear relation between the elastic part of $\mathbf{t}$
and the infinitesimal Cauchy strain tensor
\begin{equation}
\boldsymbol{\varepsilon} = \frac{1}{2} (\nabla \mathbf{u}^{T} + (\nabla \mathbf{u})\big), \notag
\end{equation}
the hyperstresses need to be accounted for.
The latter can be shown to incorporate higher order derivatives of $\mathbf{u}$
(cf. \cite[Chapter 7.1]{Ies2004}).
Following Quintanilla \cite[Sections 1 and 2]{Quin1},
the linear equations of thermoelasticity without energy dissipation for a simple material with a center of symmetry read as
\begin{align}
\rho \ddot{u}_{i} &= \big(A_{iJRs} u_{s, R} - \beta_{Ji} T - (C_{iJKSRl} u_{l, RS} + M_{iJKR} \tau_{, R})_{, K}\big)_{, J} + \rho f_{i},
\label{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_PDE_1} \\
a \ddot{\tau} &=
-\beta_{Ki} \dot{u}_{i, K} - M_{jLKI} u_{j, LKI} + K_{IJ} \tau_{, IJ} + \rho T_{0}^{-1} R
\label{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_PDE_2}
\end{align}
in $\Omega \times (0, \infty)$, where $R$ stands for the volumetric heat sources.
We refer the reader to \cite[Section 2]{Quin1} for an explanation of material constants
in Equations (\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_PDE_1})--(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_PDE_2})
as well as a discussion on their symmetry and positive definiteness properties.
The homogeneous Dirichlet-Dirchlet boundary conditions for
Equations (\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_PDE_1})--(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_PDE_2})
on $\Gamma$ are given by
\begin{equation}
u_{i} = 0, \quad u_{i, A} = 0, \quad \tau = 0 \text{ in } \Gamma \times (0, \infty),
\label{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_BC}
\end{equation}
whereas the initial conditions are stated as
\begin{equation}
u_{i}(\cdot, 0) = u_{i}^{0}, \quad
\dot{u}_{i}(\cdot, 0) = u_{i}^{1}, \quad
\tau(\cdot, 0) = \tau^{0}, \quad
T(\cdot, 0) = T^{0} \text{ in } \Omega.
\label{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_IC}
\end{equation}
In \cite{Quin1},
Quintanilla proposed a logarithmically convex energy-like function
to prove the uniqueness for Equations
(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_PDE_1})--(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_IC}).
He further employed the operator semigroup theory to obtain the existence of solutions.
Fern\'{a}ndez Sare et al. \cite{FeSaMuRiQui2010}
considered a one-dimensional counterpart of
Equations (\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_PDE_1})--(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_IC})
formulated in terms of Green \& Naghdi's type I thermoelasticity (cf. \cite{GrNa1995.II})
under various sets of boundary conditions.
Based on the operator semigroup theory and Gearhart \& Pr\"uss' theorem,
the well-posedness and exponential stability of solution were shown.
Further, the spectral analyticity criterion was employed to show the lack of analyticity for the underlying semigroup.
Finally, the authors proved the impossibility
of the solutions to localize in time.
Similar results have later been obtained
by Maga\~{n}a and Quintanilla \cite{MaQui2014} also for the case of the more comprehensive type III thermoelasticity.
Pata and Quintanilla \cite{PaQui2010}
considered a 3D version of Equation (\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_PDE_1}) with $T \equiv 0$
for the case vanishing anti-plane share deformations (i.e., $u_{1} = u_{2} = 0$).
The Equation for $u := u_{3}$ reduces then to
\begin{equation}
\rho \ddot{u} = \mu(0) \triangle u
- l_{2}(0) \triangle^{2} u
+ \int_{0}^{\infty} \big(\mu'(s) \triangle u(t - s) - l_{2}'(s) \triangle^{2} u(t - s)\big) \mathrm{d}s = 0
\label{EQUATION_PA_QUI_EQ1}
\end{equation}
together with the boundary condition
\begin{equation}
u = \triangle u = 0 \label{EQUATION_PA_QUI_EQ2}
\end{equation}
and the initial condition
\begin{equation}
u(\cdot, -s) = g(\cdot, s) \text{ for } s \in [0, \infty). \label{EQUATION_PA_QUI_EQ3}
\end{equation}
First, an abstract version of the initial-boundary value problem
(\ref{EQUATION_PA_QUI_EQ1})--(\ref{EQUATION_PA_QUI_EQ3}) was studied using the operator semigroup theory.
Based on a modification of Gearhart \& Pr\"uss' theorem,
a condition for the exponential stability was derived.
Gawinecki and \L{}azuka \cite{GaLu2006} considered a Cauchy problem
for the genuinely nonlinear version of
Equations (\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_PDE_1})--(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_PDE_2})
within Green \& Naghdi's type I thermoelasticity for homogeneous isotropic media.
Under appropriate conditions on the nonlinearity,
a global classical solutions was obtained based on $L^{p}$-$L^{q}$-estimates.
Aouadi and Moulahi \cite{AouMou2015} studied an initial-boundary value problem
for the anti-plane share component of a nonsimple thermoelastic body
with a control distributed over an open subset $\omega$ of $\Omega$
\begin{align}
u_{tt} - c^{2} \triangle u + \alpha \triangle^{2} u + c^{2} \gamma (-\triangle)^{1/2} \theta &= \chi_{\omega} u_{1},
\label{EQUATION_AOUADI_MOULAHI_PDE_1} \\
\theta_{t} - \triangle \theta - \gamma (-\triangle)^{1/2} u_{t} &= \chi_{\omega} u_{2}
\label{EQUATION_AOUADI_MOULAHI_PDE_2}
\end{align}
subject to the homogeneous Dirichlet-Dirichlet boundary conditions
and the standard initial conditions. Here, $\chi_{\omega}$ stands for the indicator function of $\omega$.
For the uncontrolled case, i.e., $u_{1} = u_{2} \equiv 0$,
a well-posedness and exponential stability result for
Equations (\ref{EQUATION_AOUADI_MOULAHI_PDE_1})--(\ref{EQUATION_AOUADI_MOULAHI_PDE_2}) was proved.
Further, the authors showed
the system is approximately controllable
over the standard state space
by $L^{2}\big(0, T; L^{2}(\Omega)\big)$-controls at any time $T > 0$
if $\omega$ is nonempty.
In the present paper, we first consider Equations
(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_PDE_1})--(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_IC})
restated within the type III thermoelasticity
(cf. \cite[Section 2]{MaQui2014} for the 1D system)
with a (macroscopic) Kelvin-Voigt damping for $\mathbf{u}$.
Hence, our equations read as
\begin{align}
\label{EQUATION_NONSIMPLE_THERMOELASTICITY_LINEAR_KELVIN_VOIGT_DAMPING_1}
\rho \ddot{u}_i &= \big(A_{iJKj} u_{j,K} - \beta_{Ji} \dot{\tau} -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} - B_{iJKj}\dot{u}_{j})_{,K}\big)_{,J} \\
\label{EQUATION_NONSIMPLE_THERMOELASTICITY_LINEAR_KELVIN_VOIGT_DAMPING_2}
a \ddot{\tau} &=
-\beta_{Ki} \dot{u}_{i,K} + m_{IJ}\dot{\tau}_{,IJ} + M_{jLKI} u_{j,LKI} + K_{IJ}\tau_{,IJ}.
\end{align}
together with the boundary and initial conditions in Equations
(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_BC})--(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_IC}).
Throughout the paper,
the following (natural) positivity and positive definitess conditions are postulated:
\begin{itemize}
\item[I.] $a,\rho > 0$.
\item[II.] There exists $\alpha > 0$ such that
\begin{align}
\notag
C_{iIJKLj} u_{i,JI} u_{j,LK} + M_{iJKL} u_{i,KJ} \tau_{, L} & \\
\label{DPositDefined1}
+ M_{jLKI} u_{j,LK}\tau_{,I} + K_{IJ} \tau_{,I}\tau_{,J} & \ge
\alpha(\tau_{,R}\tau_{,R} + u_{k,ST}u_{k,ST}), \\
\label{DPositDefined2} A_{iKLj} u_{i,K} u_{j,L} & \ge \alpha
u_{k,R}u_{k,R}.
\end{align}
\end{itemize}
In Section \ref{SECTION_WELL_POSEDNESS}, we show
Equations (\ref{EQUATION_NONSIMPLE_THERMOELASTICITY_LINEAR_KELVIN_VOIGT_DAMPING_1})--(\ref{EQUATION_NONSIMPLE_THERMOELASTICITY_LINEAR_KELVIN_VOIGT_DAMPING_2}),
(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_BC})--(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_IC}) are well-posed.
Note that no positive definiteness conditions on $(B_{iJKj})$ or $(m_{IJ})$ are imposed.
By a standard perturbation argument (e.g., \cite[Chapter 3.1, Theorem 1.1]{Pazy1},
the problem remains well-posed if the Kelvin-Voigt damping
is replaced or complemented with a frictional damping.
In Section \ref{SECTION_LACK_OF_EXPONENTIAL_STABILITY}, we prove that,
in contrast to the 1D situation (cf. \cite[Section 2]{MaQui2014}),
if $B_{iJKj} \equiv 0$,
the positive definiteness of $(m_{IJ})$ is not sufficient
to exponentially stabilize the system.
Motivated by the necessity of an additional damping for $\mathbf{u}$,
we let $B_{iJKj} \equiv 0$, but add a linear frictional damping
to Equation (\ref{EQUATION_NONSIMPLE_THERMOELASTICITY_LINEAR_KELVIN_VOIGT_DAMPING_1}) thus obtaining
\begin{equation}
\label{EQUATION_NONSIMPLE_THERMOELASTICITY_LINEAR_FRICTIONAL_DAMPING}
\rho \ddot{u}_i = \big(A_{iJKj} u_{j,K} - \beta_{Ji} \dot{\tau} -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} - B_{iJKj}\dot{u}_{j})_{,K}\big)_{,J} - E_{ij} \dot{u}_{j}.
\end{equation}
Further, in Section \ref{SECTION_EXPONENTIAL_STABILITY},
under a positive definiteness assumption on $(E_{ij})$,
we use the Lyapunov's method to show the exponential stability of Equations
(\ref{EQUATION_NONSIMPLE_THERMOELASTICITY_LINEAR_FRICTIONAL_DAMPING}), (\ref{EQUATION_NONSIMPLE_THERMOELASTICITY_LINEAR_KELVIN_VOIGT_DAMPING_2}),
(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_BC})--(\ref{EQUATION_LINEAR_NONSIMPLE_THERMOELASTICITY_GENERAL_IC}).
In Section \ref{SECTION_HYPERBOLIZED_SYSTEM}, a hyperbolization of (\ref{EQUATION_NONSIMPLE_THERMOELASTICITY_LINEAR_KELVIN_VOIGT_DAMPING_2})
together a semilinear frictional damping in Equation (\ref{EQUATION_NONSIMPLE_THERMOELASTICITY_LINEAR_KELVIN_VOIGT_DAMPING_1}) are considered.
The resulting semilinear system reads then as
\begin{align}
\rho \ddot{u}_i &= \left( A_{iJKj} u_{j,K} - \beta_{Ji} \dot{\tau} -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} )_{,K} \right)_{,J} -
E(|\dot{u}|) \dot{u}_i,
\label{EQUATION_NONSIMPLE_THERMOELASTICITY_SEMILINEAR_DAMPED_HYPERBOLIZED_1} \\
a \ddot{\tau} &=
-\beta_{Ki} \dot{u}_{i,K} + m_{IJ}q_{I,J} + M_{jLKI} u_{j,LKI} +
K_{IJ}\tau_{,IJ},
\label{EQUATION_NONSIMPLE_THERMOELASTICITY_SEMILINEAR_DAMPED_HYPERBOLIZED_2} \\
\kappa \dot{q}_i &= \dot{\tau}_{,i} - q_{i},
\label{EQUATION_NONSIMPLE_THERMOELASTICITY_SEMILINEAR_DAMPED_HYPERBOLIZED_3}
\end{align}
with a small parameter $\kappa > 0$.
Here, $(m_{IJ} q_{I, J})$ plays the role of the heat flux.
If the function $E$ is constant,
similar to the case $\kappa = 0$,
Equations (\ref{EQUATION_NONSIMPLE_THERMOELASTICITY_SEMILINEAR_DAMPED_HYPERBOLIZED_1})--(\ref{EQUATION_NONSIMPLE_THERMOELASTICITY_SEMILINEAR_DAMPED_HYPERBOLIZED_3})
are exponentially stable.
In the nonlinear case,
we exploit a technique due to Lasiecka and Tataru \cite{Lasiecka1}
to prove the uniform stability of the nonlinear system
dependent on the behavior of $E$ at $0$ and infinity.
Here, we use a generalization of a technique dating back to Haraux \cite{Haraux1},
which is given in the appendix \ref{SECTION_APPENDIX}.
\section{Linear sytem with a Kelvin-Voigt damping for the elastic part: Well-posedness}
\label{SECTION_WELL_POSEDNESS}
In this section, we consider the equations of type III thermoelasticity for a nonsimple material
with a Kelvin-Voigt damping for $\mathbf{u}$.
By a standard perturbation argument \cite[Chapter 3.1, Theorem 1.1]{Pazy1},
the results remain valid also when a frictional damping, i.e., a term like $E_{ij} \dot{u}_{j}$,
is considered instead of or in addition to $B_{iJKj} \dot{u}_{j, KJ}$.
The equations read then as
\begin{align}
\label{SystemTwoFrictionSecondCase1}
\rho \ddot{u}_i &= \big(A_{iJKj} u_{j,K} - \beta_{Ji} \dot{\tau} -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} - B_{iJKj}\dot{u}_{j})_{,K}\big)_{,J} \\
\label{SystemTwoFrictionSecondCase2}
a \ddot{\tau} &=
-\beta_{Ki} \dot{u}_{i,K} + m_{IJ}\dot{\tau}_{,IJ} + M_{jLKI} u_{j,LKI} + K_{IJ}\tau_{,IJ}
\end{align}
in $\Omega \times (0, \infty)$ together with the boundary conditions
\begin{equation}
\label{WPBoundaryConditions}
u_i = 0, \quad u_{i,J} = 0, \quad \tau = 0 \text{ in } \Gamma \times (0, \infty)
\end{equation}
and the initial conditions
\begin{equation}
\label{WPInitialConditions}
u_i(\cdot, 0) = u_i^0, \quad \dot{u}_i(\cdot,0) = \dot{u}_i^0, \quad
\tau (\cdot, 0) = \tau^0, \quad \dot{\tau}(\cdot, 0) = \dot{\tau}^0 \text{ in } \Omega,
\end{equation}
where $\Gamma := \partial \Omega$ is assumed Lipschitzian.
In addition to conditions I, II, we assume
\begin{align}
B_{iJKj} \xi_{j,K} \xi_{i,J} \ge 0, \quad
m_{IJ} \xi_{I} \xi_{J} \ge 0.
\label{EQUATION_POSITIVE_SEMIDEFINITENESS_DAMPING}
\end{align}
Letting $v = \dot{u}$ and $\theta = \dot{\tau}$, we rewrite the system
(\ref{SystemTwoFrictionSecondCase1})--(\ref{SystemTwoFrictionSecondCase2}) as follows:
\begin{align}
\label{WPSystem1}
\dot{u}_i &= v_i, \\
\label{WPSystem2}
\dot{v}_i &= \frac{1}{\rho} \big(A_{iJKj} u_{j,K} - \beta_{Ji}
\theta - (C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} - B_{iJKj}v_{j})_{,K} \big)_{,J}, \\
\label{WPSystem3}
\dot{\tau} &= \theta, \\
\label{WPSystem4}
\dot{\theta} &= \frac{1}{a} \big(-\beta_{Ki} v_{i,K} +
(m_{IJ}\theta_{,J} + M_{jLKI} u_{j,LK} + K_{IJ}\tau_{,J})_{,I}\big).
\end{align}
We assume the evolution is taking place on the Hilbert space
\begin{equation}
\mathcal{H} = \left\{U \,|\, U = (u, v, \tau, \theta)^{T} \in
\left(H^2_{0}(\Omega)\right)^{d} \times \left( L^2(\Omega) \right)^{d} \times
H_0^1(\Omega) \times L^2(\Omega)\right\}, \notag
\end{equation}
where
\begin{equation}
H^{s}_{0}(\Omega) :=
\mathrm{clos}\big(C_{0}^{\infty}(\Omega), \|\cdot\|_{H^{s}}\big) \text{ for } s \in \mathbb{N}, \notag
\end{equation}
equipped with the scalar product
\begin{align}
\langle U, U^{\ast} \rangle_{\mathcal{H}} &= \int_{\Omega} \big(\rho v_{i} v_{i}^* +
a \theta \theta^* + A_{iKLj} u_{i,K} u^*_{j,L} + C_{iIJKLj} u_{i,JI}
u^*_{j,LK} + K_{IJ} \tau_{,J} \tau^*_{,I} \big) \mathrm{d}x \notag \\
&+ \int_{\Omega} \big(M_{iJKL} u_{i,KJ} \tau^*_{,L} + M_{jLKI}
u^*_{j,KL} \tau_{,I} \big) \mathrm{d}x, \label{EQUATION_SCALAR_PRODUCT_WEIGHTED}
\end{align}
for $U = (u, v, \tau, \theta)^{T}$, $U^{\ast} = (u^{\ast}, v^{\ast}, \tau^{\ast}, \theta^{\ast})^{T} \in \spH$.
Here and in the following, for the sake of simplicity,
we drop the physical convention to bold the vector and tensor fields.
It is easy to verify that this scalar product is equivalent with the usual product
induced by the product topology.
Consider the linear operator $\opA: D(\opA) \subset \mathcal{H} \to
\mathcal{H}$ defined by
\begin{equation}
\label{WPOpADefine}
\opA U =
{\small
\begin{pmatrix}
v_i,
\\
\frac{1}{\rho} \big( A_{iJKj} u_{j,K} - \beta_{Ji} \theta -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} - B_{iJKj}v_{j} )_{,K}
\big)_{,J},
\\
\theta,
\\
\frac{1}{a} \left( -\beta_{Ki} v_{i,K} + (m_{IJ}\theta_{,J} +
M_{jLKI} u_{j,LK} + K_{IJ}\tau_{,J})_{,I} \right)
\end{pmatrix}},
\end{equation}
where
\begin{align}
\begin{split}
D(\opA) = \big\{ U \in \mathcal{H} \,|\, \phantom{-}&v_i\in H^2_0(\Omega),\
\theta \in H^1_0(\Omega), \\
\phantom{-}&(A_{iJKj} u_{j,K} - \beta_{Ji} \theta)_{,J} \\
-&(C_{iJKLIj} u_{j,IL} M_{iJKL} \tau_{,L} - B_{iJKj}v_{j} )_{,K} \big)_{,J} \in L^2(\Omega), \\
\phantom{-}&(m_{IJ}\theta_{,J} + M_{jLKI} u_{j,LK} +
K_{IJ}\tau_{,J})_{,I} \in L^2(\Omega)\big\}
\end{split}
\label{WPOperatorDomain}
\end{align}
Thus, the abstract form of Equations (\ref{WPSystem1})--(\ref{WPSystem4}) is given by
\begin{equation}
\label{WPCauchyProblem}
\dot{U}(t) = \opA U(t) \text{ for } t > 0, \quad U(0) = U_0.
\end{equation}
with $U_{0} := (u^{0}_{i}, u^{1}_{i}, \tau^{0}, \tau^{1})^{T}$.
We first prove the following auxiliary lemma.
\begin{lemma}\label{WPMainLemma}
The operator $\opA$ defined in Equations
(\ref{WPOpADefine})--(\ref{WPOperatorDomain}) is the infinitesimal generator of
a $C_0$-semigroup of contractions on $\mathcal{H}$.
\end{lemma}
\begin{proof}
The proof of Lemma \ref{WPMainLemma} resembles the one of \cite[Theorem 2]{Po2014}.
{\it Denseness: }
Utilizing the fact that
\begin{equation}
\left(C_{0}^{\infty}(\Omega) \right)^d \times
\left(C_{0}^{\infty}(\Omega)\right)^d \times C_{0}^{\infty}(\Omega)
\times C_{0}^{\infty}(\Omega) \notag
\end{equation}
is a subset of the domain of $\opA$, we conclude that $D(\opA)$ is
dense in $\mathcal{H}$.
{\it Disipativity: }
A straightforward calculation involving the
Green's formula yields
\begin{equation}
\langle \opA U , U \rangle_{\mathcal{H}} = - \int_{\Omega} \big(B_{iJKj}
v_{i,K} v_{i,J} + m_{IJ} \theta_{,I} \theta_{,J} \big) \mathrm{d}x.
\end{equation}
Using Equation (\ref{EQUATION_POSITIVE_SEMIDEFINITENESS_DAMPING}), we get
\begin{equation}
\langle \opA U , U \rangle_{\mathcal{H}} \le -\alpha \int_{\Omega}
\big(v_{i,K}v_{i,K} + \theta_{,I}\theta_{,I} \big) \mathrm{d}x \le 0.
\end{equation}
{\it Maximality: }
Now we prove that $0\in \rho(\opA)$ with $\rho(\opA)$ standing for
the resolvent set of $\opA$. For $F = (F^1, F^2, F^3, F^4)^{T} \in \mathcal{H}$,
consider the operator equation $\opA U = F$ or, expicitly,
\begin{align*}
v_i &= F^1_i,
\\
\frac{1}{\rho} \left( A_{iJKj} u_{j,K} - \beta_{Ji} \theta -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} - B_{iJKj}v_{j} )_{,K}
\right)_{,J} &= F^2_i,
\\
\theta &= F^3,
\\
\frac{1}{a} \left( -\beta_{Ki} v_{i,K} + (m_{IJ}\theta_{,J} +
M_{jLKI} u_{j,LK} + K_{IJ}\tau_{,J})_{,I} \right) &= F^4.
\end{align*}
Eliminating $v_i$ and $\theta$, we obtain
\begin{align}
\label{WPLaxMilgramProblem1}
\frac{1}{\rho} \left( A_{iJKj} u_{j,K} - \beta_{Ji} F^3 -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} - B_{iJKj}F^1_{j} )_{,K}
\right)_{,J} &= F^2_i,
\\
\label{WPLaxMilgramProblem2} \frac{1}{a} \left( -\beta_{Ki}
F^1_{i,K} + (m_{IJ}F^3_{,J} + M_{jLKI} u_{j,LK} +
K_{IJ}\tau_{,J})_{,I} \right) &= F^4.
\end{align}
To solve this system, we exploit the lemma of Lax \& Milgram.
We consider the Hilbert space
\begin{equation}
\mathcal{V} = \left(H_0^2(\Omega)\right)^d \times H_0^1(\Omega) \notag
\end{equation}
equipped with the standard inner product associated with the product
topology and introduce the bilinear form $\frak{a}: \mathcal{V}
\times \mathcal{V} \to \mathbb{R}$ via
\begin{align*}
\frak{a}(V,V^*) &= \int_{\Omega} (A_{iJKj} u_{s,R} u^*_{i,J} +
C_{iIJKLj}u_{j,LK} u^*_{i,IJ}) \ \mathrm{d}x \\
&+\int_{\Omega} (M_{iJKL} \tau_{,L} u^*_{i,JK} + M_{jLKI} u_{j,KL}
\tau^*_{,I} + K_{IJ} \tau_{,I} \tau^*_{,J})\ \mathrm{d}x,
\end{align*}
where $V=(u, \tau)^{T}$ and $V^* = (u^*, \tau^*)^{T}$.
After multiplying Equations (\ref{WPLaxMilgramProblem1})--(\ref{WPLaxMilgramProblem2})
in the inner product $(L^2(\Omega))^n$ and $L^2(\Omega)$ with $\rho W^1$ and $a W^2$, respectively,
summing up the resulting equations and integrating by parts,
we obtain a weak formulation of Equations (\ref{WPLaxMilgramProblem1})--(\ref{WPLaxMilgramProblem2}) in the form:
\begin{align}
\begin{split}
&\text{Determine $V\in\mathcal{V}$ such that } \\
&\frak{a}(V,W) = -\langle G^1, W^1\rangle_{(L^2(\Omega))^n}
- \langle G^2, W^2\rangle_{L^2(\Omega)}
- m_{IJ}\langle F^3_{,J}, W^2_{,I}\rangle_{L^2(\Omega)}.
\end{split}
\label{WPWeakProblem}
\end{align}
Here, $G^1_i = \rho F^2_i + \beta_{Ji} F^3_{, J} - B_{iJKj} F^1_{j,KJ},\
G^2 = aF^4 + \beta_{Ki}F^1_{i,K}$.
The bilinear form $\frak{a}$ is continuous and coercive on $\mathcal{V}$ due to the conditions in
(\ref{DPositDefined1})--(\ref{DPositDefined2}).
Clearly, the linear functional
\begin{equation}
(W^1,W^2) \mapsto \langle G^1, W^1\rangle_{(L^2(\Omega))^n}
+ \langle G^2, W^2\rangle_{L^2(\Omega)}
+ m_{IJ}\langle F^3_{,J}, W^2_{,I}\rangle_{L^2(\Omega)}
\notag
\end{equation}
is continuous on $\mathcal{V}$.
Applying now the Lemma of Lax \& Milgram,
we deduce existence of a unique solution $V = (u, \tau)^{T} \in\mathcal{V}$
to Equations (\ref{WPLaxMilgramProblem1})--(\ref{WPLaxMilgramProblem2}).
Recalling the definition of $v_{i}$ and $\theta$,
we further get $v_i = F^1_i \in H^2_0(\Omega)$
and $\theta = F^3 \in H^1_0(\Omega)$.
Hence, the conditions in Equations (\ref{WPOperatorDomain}) are satisfied.
Therefore, we have $(u,v,\tau,\theta)^{T} \in D(\opA)$ implying that $(u,v,\tau,\theta)^{T}$ is a strong solution.
Finally we show the continuous dependence of the solution on $F$.
It follows from the Lemma of Lax \& Milgram that there exists $c_0 > 0$ such that
\begin{equation}
\label{WPAdditionalIneq1}
\| V \|_{\mathcal{V}} \le c_0 \| \widehat{F} \|_{\mathcal{V}^*},
\end{equation}
where $\widehat{F} = (F^2, F^4)^{T}$.
Since $\mathcal{V}$ is continuously embedded into $\left(L^2(\Omega)\right)^{d+1}$, we have
\begin{equation}\label{WPAdditionalIneq2}
\|\widehat{F}\|_{\mathcal{V}^*} \le c_1 \|\widehat{F}\|_{\left(L^2(\Omega)\right)^{d+1}}.
\end{equation}
Suppose $\|F_{n}\|_{\mathcal{H}} \to 0$ as $n \to \infty$.
By definition,
$\|F^1_{n}\|_{\left(H^2_0(\Omega) \right)^d} \to 0$ and $\|F^3_{n}\|_{H^1_0(\Omega)} \to 0$ as $n \to \infty$
and, therefore, $\big\|(v_i)_{n}\big\|_{L^2(\Omega)} \to 0$ and $\|\theta_{n}\|_{L^2(\Omega)} \to 0$ as $n \to \infty$.
Further, $\|F^2_{n}\|_{\left(L^2(\Omega)\right)^d}, \|F^4_{n}\|_{L^2(\Omega)} \to 0$ as $n \to \infty$
and it follows from Equations (\ref{WPAdditionalIneq1})--(\ref{WPAdditionalIneq2}) that
$\|V\|_{\mathcal{V}} \to 0$.
Therefore, the solution $U_{n}$ of equation
$\opA U_{n} = F_{n}$ tends to $0$ in $\mathcal{H}$ as $n \to \infty$.
Now, after we have shown $\opA$ is $m-$dissipative,
the claim of the present theorem follows from the Lumer \& Phillips theorem
(cf. \cite[Theorem 4.3, p. 14]{Pazy1}).
\end{proof}
Now, by virtue of \cite[Theorem 1.3, p. 102]{Pazy1},
Equations (\ref{SystemTwoFrictionSecondCase1})--(\ref{WPInitialConditions})
as well as their abstract formulation (\ref{WPCauchyProblem}) are well-posed.
\begin{theorem}
Let $U_0 \in \mathcal{H}$.
There exists then a unique mild solution $U \in C^0\big([0, \infty), \spH\big)$
to Equation (\ref{WPCauchyProblem}).
If $U_0\in D(\opA)$, the mild solution is even a
classical one satisfying $U \in C^1\big([0, \infty), \spH) \cap C^{0}\big([0, \infty), D(\mathcal{A})\big)$.
\end{theorem}
Applying \cite[Chapter 3.1, Theorem 1.1]{Pazy1}, we further obtain
\begin{corollary}
Replacing Equation (\ref{SystemTwoFrictionSecondCase1}) with
\begin{equation}
\rho \ddot{u}_i = \big(A_{iJKj} u_{j,K} - \beta_{Ji} \dot{\tau} -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} - B_{iJKj}\dot{u}_{j})_{,K}\big)_{,J} +
E_{ij} \dot{u}_{j}
\label{SystemTwoFrictionSecondCase1_FRICTIONAL_DAMPING}
\end{equation}
for an arbitrary matrix $(E_{ij}) \in \mathbb{R}^{d \times d}$
in Equations (\ref{SystemTwoFrictionSecondCase1_FRICTIONAL_DAMPING}),
(\ref{SystemTwoFrictionSecondCase2})--(\ref{WPInitialConditions}),
the resulting abstract Cauchy problem is well-posed on $\mathcal{H}$.
\end{corollary}
\section{Linear system with an undamped elastic part: Lack of exponential stability}
\label{SECTION_LACK_OF_EXPONENTIAL_STABILITY}
In this section, we consider Equations
(\ref{SystemTwoFrictionSecondCase1})--(\ref{SystemTwoFrictionSecondCase2}) for the case $B_{iJKj} \equiv 0$, i.e.,
\begin{align}
\label{SystemOneFriction1}
\rho \ddot{u}_i &= \big(A_{iJKj} u_{j,K} - \beta_{Ji} \dot{\tau} -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L}, \\
\label{SystemOneFriction2}
a \ddot{\tau} &=
-\beta_{Ki} \dot{u}_{i,K} + m_{IJ}\dot{\tau}_{,IJ} + M_{jLKI} u_{j,LKI} + K_{IJ}\tau_{,IJ}.
\end{align}
Equations (\ref{SystemOneFriction1})--(\ref{SystemOneFriction2})
are precisely the type III thermoelasticity for nonsimple materials.
Under a suitable choice of natural boundary conditions,
i.e., the ones appearing in the Green's formula,
and a feasible selection of the coefficient tensors
$(A_{iJKj})$, $(\beta_{Ji})$, $(C_{iJKLIj})$, $(m_{IJ})$, $(K_{IJ})$
as well as restricting $\Omega$ to the rectangular configuration
\begin{equation}
\Omega = (0, \pi) \times (0, \pi), \notag
\end{equation}
we show that the system
(\ref{SystemOneFriction1})--(\ref{SystemOneFriction2}) is lacking exponential stability.
Selecting the relatively open in $\Gamma$ disjunctive sets
\begin{align*}
\Gamma_{1} := (0, \pi) \times \{0, \pi\} \text{ and }
\Gamma_{2} := \{0, \pi\} \times (0, \pi),
\end{align*}
we have $\Gamma = \overline{\Gamma}_1 \cup \overline{\Gamma}_2$.
The coefficient tensors
$(A_{iJKj}), (\beta_{Ji}), (C_{iJKLIj}), (m_{IJ}), (K_{IJ})$
are chosen in \cite[p. 5]{Quin1} to model an isotropic material with a center of symmetry.
Further, we select
\begin{equation}
(m_{ij}) :=
\left(
\begin{array}{cc}
\delta & 0 \\
0 & \delta \\
\end{array}\right) \text{ for some } \delta > 0. \notag
\end{equation}
The balance equations and material laws (cf. \cite{Quin1}) read then as
\begin{align*}
\tau_{JI,J} &= \sigma_{RJI,RJ} + \rho \ddot{u}_{I},
\\
(\phi_{I})_{,I} &= \beta \dot{e}_{LL} - D_{IJ} \dot{\tau}_{,IJ} + a\ddot{\tau},
\\
\tau_{IJ} &= \lambda \delta_{IJ} e_{LL} + 2\mu e_{IJ} - \beta \delta_{IJ} \dot{\tau},
\\
\sigma_{IJK} &= \frac12 a_1 (\delta_{JK} \kappa_{LLI} + 2\delta_{IJ} \kappa_{KLL} + \delta_{IK} \kappa_{LLJ}) + a_2(
\delta_{JI} \kappa_{ILL} + \delta_{IK} \kappa_{JLL})
\\
&+ 2a_3 \delta_{IJ} \kappa_{LLK} + 2a_4\kappa_{IJK} + a_5(\kappa_{KJI} + \kappa_{KIJ}) + m_1\delta_{IJ} \tau_{,K}
\\
&+ m_2(\delta_{JK} \tau_{,I} + \delta_{IK} \tau_{,J}),
\\
\phi_{I} &= m_1\kappa_{LLI} + m_2(\kappa_{ILL} + \kappa_{LIL}) +
\delta_{IJ} \tau_{,J},
\\
e_{IJ} &= \frac12 (u_{I,J} + u_{J,I}),\ \kappa_{IJK} = u_{K,IJ}.
\end{align*}
After plugging the constitutive equations into the balance equations,
we arrive at the system
\begin{align}
\notag
\rho \ddot{u}_{1} &= (\lambda+2\mu)u_{1,11} + \mu u_{1,22}
+ (\lambda + \mu) u_{2,12} - \beta \dot{\tau}_{,1} - b_1 u_{1,1111}
- b_2 u_{1,1122}
\\ \label{NESElemSyst1}
&- b_3 u_{1,2222} - b_4 u_{2,1112} - b_5
u_{2,1222} -(m_1+2m_2)(\tau_{,111} + \tau_{,122}),
\\
\notag
\rho \ddot{u}_{2} &= (\lambda+2\mu)u_{2,22} + \mu u_{2,11} +
(\lambda + \mu) u_{1,12} - \beta \dot{\tau}_{,2} - b_1 u_{2,2222} -
b_2 u_{2,1122}
\\
\label{NESElemSyst2} &- b_3 u_{2,1111} - b_4 u_{1,1112} - b_5
u_{1,1222} -(m_1+2m_2)(\tau_{,112} + \tau_{,222}),
\\
\notag
a\ddot{\tau} &= (m_1 + 2m_2)(u_{1,111}+u_{1,122} + u_{2,112}
+ u_{2,222}) + \tau_{,1} + \tau_{,2}
\\
\label{NESElemSyst3} &- \beta (\dot{u}_{1,1} + \dot{u}_{2,2}) +
\delta (\dot{\tau}_{,11} + \dot{\tau}_{,22}),
\end{align}
where
\begin{align*}
b_1 &= 2(a_1+a_2+a_3+a_4+a_5), & b_2 &= 2(a_1+a_2+2a_3+2a_4+a_5), \\
b_3 &= 2(a_3+a_4), & b_4 &= b_5=2(a_1+a_2+a_5).
\end{align*}
Throughout this section, we employ the following set of boundary conditions,
which naturally arise from partially integrating of the system (\ref{NESElemSyst1})--(\ref{NESElemSyst3}):
\begin{align} \notag
\sigma_{111} |_{\Gamma} =
a_1(u_{1,11}+u_{1,22}+u_{1,12}+u_{2,12}) +2a_2(u_{1,11}+u_{2,21}) &
\\
\label{NESBoundaryConditions1}
+2a_3(u_{1,11} + u_{1,22}) + 2(a_4+a_5)u_{1,11} +
(m_1+2m_2)\tau_{,1} |_{\Gamma} & = 0,
\\
\notag
\sigma_{221}|_{\Gamma} =
a_1(u_{1,11}+u_{2,12}) +2a_3(u_{1,11} + u_{1,22}) +2a_4 u_{1,22} &
\\ \label{NESBoundaryConditions2}
+ 2a_5 u_{2,12} + m_1\tau_{,1} |_{\Gamma} & = 0,
\\ \notag
\sigma_{112}|_{\Gamma} = \frac12
a_1(u_{1,11}+u_{1,22}) + a_2(u_{1,11} + u_{2,21}) +2a_4 u_{2,12} & \\
\label{NESBoundaryConditions3} + a_5 ( u_{1,22} + u_{2,12}) +
m_2\tau_{,1} |_{\Gamma} & = 0,
\\
\label{NESBoundaryConditions4}
( \sigma_{112,1}, \sigma_{222,2} ) \cdot (\nu_1, \nu_2) |_{\Gamma}
&= 0,
\\
\label{NESBoundaryConditions5} (\tau_{,1}, \tau_{,2})^T \cdot
(\nu_1, \nu_2)^T |_{\Gamma_2} = 0, \tau |_{\Gamma_1} & = 0,
\end{align}
where $\nu = (\nu_1, \nu_2)^T$ is the unit exterior normal vector to $\Gamma = \partial \Omega$.
To eliminate the trivial kernel, the ground space is then selected as the Hilbert space
\begin{equation}
\spH_* = \left( H^2(\Omega)/ \{1\} \right)^2 \times (L^2(\Omega)/
\{1\})^2 \times \left(H^1(\Omega)/ \{1\}\right) \times
\left(L^2(\Omega)/ \{1\}\right)
\notag
\end{equation}
i.e., each component is taken as the orthogonal complement
of the one-dimensional subspace of constant functions.
Without loss of generality, we equip this new space $\spH_*$ with the canonical unweighted inner product,
which is equivalent with our original weighted definition in Equation (\ref{EQUATION_SCALAR_PRODUCT_WEIGHTED}).
The operator $\opA$ is defined similar to (\ref{WPOpADefine}).
The boundary conditions in Equations
(\ref{NESBoundaryConditions1})--(\ref{NESBoundaryConditions5}) are
incorporated into the weak definition of the domain $D(\opA)$
similar to Equation (\ref{WPOperatorDomain}).
\begin{theorem}
Assume $a_3 + a_4 \ge 0$.
Then the system (\ref{NESElemSyst1})--(\ref{NESElemSyst3}) with boundary conditions
(\ref{NESBoundaryConditions1})--(\ref{NESBoundaryConditions5}) is not exponentially stable.
\end{theorem}
\begin{proof}
We prove there exists a sequence $(\lambda_n)_n \subset \mathbb{R}$ with
\begin{equation}
\lim_{n\to\infty} |\lambda_n| = \infty \notag
\end{equation}
as well as the sequences $(U_n)_n \subset D(\opA)$ and
$(F_n)_n\subset \spH_0$ such that
\begin{equation}
(\ii \lambda_n - \opA)U_n = F_n
\text{ is uniformly bounded w.r.t. }
n\in\mathbb{N} \text{ and } \lim_{n\to\infty} \| U_n \|_{\spH_*} = \infty. \notag
\end{equation}
The lack of exponential stability will then follow from the well-known
Gearhart \& Pr\"uss' theorem (see e.g. \cite[Theorem 1.3.2, p. 4]{Liu1}).
For $\lambda\in\mathbb{R}$, the solution
$U = (u_1, u_2, v_1, v_2, \tau, \theta)^T$ of the resolvent equation $\protect{(\lambda \ii - \opA)U = F}$
satisfies
\begin{align*}
\lambda_n u_1 \ii - v_1 &= 0,
\\
\lambda_n u_2 \ii - v_2 &= 0,
\\
\lambda_n v_1 \ii - \frac{1}{\rho} \big((\lambda+2\mu)u_{1,11} +
\mu u_{1,22} + (\lambda + \mu) u_{2,12} - \beta \dot{\tau}_{,1} -
b_1 u_{1,1111} &
\\
- b_2 u_{1,1122} - b_3 u_{1,2222} - b_4 u_{2,1112} - b_5 u_{2,1222}&
\\
-(m_1+2m_2)(\tau_{,111} + \tau_{,122})\big) &= f_3,
\\
\lambda_n v_2 \ii - \frac{1}{\rho} \big((\lambda + 2\mu)u_{2,22} +
\mu u_{2,11} + (\lambda + \mu) u_{1,12} - \beta \dot{\tau}_{,2} - b_1 u_{2,2222} &
\\
- b_2 u_{2,1122} - b_3 u_{2,1111} - b_4
u_{1,1112} - b_5 u_{1,1222} &
\\
-(m_1+2m_2)(\tau_{,112} + \tau_{,222})\big) &= f_4,
\\
\lambda_n \tau \ii - \theta &= 0 ,
\\
\lambda_n \theta \ii - \frac{1}{a}\big((m_1 + 2m_2)(u_{1,111}+u_{1,122} + u_{2,112} +
u_{2,222}) & \\
+ \tau_{,1} + \tau_{,2} - \beta (\dot{u}_{1,1} +
\dot{u}_{2,2}) + \delta (\dot{\tau}_{,11} + \dot{\tau}_{,22}) \big) &= f_6,
\end{align*}
where $f_3,f_4,f_6$ will be selected as
\begin{equation}
f_3 = \sin (nx) \sin (ny), \quad
f_4 = \cos (nx) \cos (ny), \quad
f_6 = \cos (nx) \sin (ny) \notag
\end{equation}
for $n \in \mathbb{N}$.
Eliminating $v_1, v_2, \theta$, we obtain for $u_1, u_2, \tau$ the following algebraic system
\begin{align}
\notag
-\lambda^2_n u_1 - \frac{1}{\rho} \big((\lambda+2\mu) u_{1,11} +
\mu u_{1,22} + (\lambda + \mu) u_{2,12} &
\\
\notag
-\beta \lambda_n \ii \tau_{,1} - b_1 u_{1,1111} - b_2 u_{1,1122} - b_3 u_{1,2222}
\\
\label{NESReducedSystem1}
- b_4 u_{2,1112} - b_5 u_{2,1222} -(m_1+2m_2)(\tau_{,111} + \tau_{,122})\big) &= f_3,
\\
\notag
-\lambda^2_n u_2 - \frac{1}{\rho} \big((\lambda + 2\mu) u_{2,22} +
\mu u_{2,11} + (\lambda + \mu) u_{1,12} &
\\
\notag
-\beta \lambda_n \ii \tau_{,2} - b_1 u_{2,2222} - b_2 u_{2,1122} - b_3 u_{2,1111}
\\
\label{NESReducedSystem2}
- b_4 u_{1,1112} - b_5 u_{1,1222} -(m_1+2m_2)(\tau_{,112} + \tau_{,222})\big) &= f_4,
\\
\notag
-\lambda^2_n \theta \ii - \frac{1}{a} \big((m_1 + 2m_2)(u_{1,111}+u_{1,122} + u_{2,112} + u_{2,222}) &
\\
\label{NESReducedSystem3}
+ \tau_{,1} + \tau_{,2} - \beta \lambda_n \ii (u_{1,1} +
u_{2,2}) + \delta \lambda_n \ii (\tau_{,11} + \tau_{,22})\big) &= f_6.
\end{align}
To solve Equations (\ref{NESReducedSystem1})--(\ref{NESReducedSystem3}),
we employ the ansatz
\begin{equation}
u_1 = A\sin (nx) \sin (ny), \quad
u_2 = B\cos (nx) \cos (ny), \quad
\tau = C\cos (nx) \sin (ny), \notag
\end{equation}
where $A, B, C$ will actually depend on $n$.
It should be pointed out that this choice is compatible with the
boundary conditions in Equations (\ref{NESBoundaryConditions1})--(\ref{NESBoundaryConditions5}).
Thus, system (\ref{NESReducedSystem1})--(\ref{NESReducedSystem3}) is
equivalent with finding $A, B, C$ such that
\begin{align}
\notag
\big(-\lambda_n^2 \rho + (\lambda+ 3\mu) n^2 + 2b_2n^4\big) A - \big((\lambda +\mu) n^2 + 2b_4 n^4\big) B &
\\
\label{NESFinalSystem1}
+ \big(-\beta \lambda_n n \ii + 2 (m_1+2m+2) n^3\big) C &= \rho,
\\
\notag
\big(-\lambda_n^2 \rho + (\lambda+ 3\mu) n^2 + 2b_2n^4\big) B - \big((\lambda +\mu) n^2 + 2b_4 n^4\big) A &
\\
\label{NESFinalSystem2}
+ \big(\beta \lambda_n n \ii - 2(m_1 + 2m+2) n^3\big) C &= \rho, \\
\label{NESFinalSystem3}
\big(2(m_1 + 2m_2) n^3 + \beta \lambda_n n \ii\big)(A - B)
+ (-\lambda_n^2 a + 2n^2 +2\lambda_n \delta n^2 \ii) C &= a.
\end{align}
Let
\begin{equation}
\lambda_n = \sqrt{\big(8(a_3+a_4) n^4 + 2\mu n^2 - 1\big)/\rho}. \notag
\end{equation}
It is easy to verify that the linear algebraic system
(\ref{NESFinalSystem1})--(\ref{NESFinalSystem3}) has a unique solution.
Summing up Equations (\ref{NESFinalSystem1}) and
(\ref{NESFinalSystem2}), we get
\begin{equation}
(A + B) \big(-\lambda_n^2 \rho + 2\mu n^2 + 8(a_3 + a_4) n^4\big) = 2\rho, \notag
\end{equation}
where $a_3 + a_4 \ge 0$ follows from the positive definiteness of $C_{iIJKLj}$.
Equivalently,
\begin{equation}
A + B = 2 \rho. \notag
\end{equation}
and, therefore, by virtue of Young's inequality,
\begin{align*}
\|U\|^2_{\spH_*} &\ge
\|v_1\|^2_{L^2(\Omega)} + \|v_2\|^2_{L^2(\Omega)}
= \lambda_n^2 \left(\|u_1\|^2_{L^2(\Omega)} + \|u_2\|^2_{L^2(\Omega)}\right)
\\
&= \lambda_n^2 \frac{\pi^2}{2} (A^2 + B^2)
\ge \frac{\lambda_n^2 \pi^2}{2} \left(\frac{A+B}{2}\right)^2
= \frac{\lambda_n^2 \pi^2 \rho^2}{2} \to \infty,
\end{align*}
whereas we estimate
\begin{equation}
\|F\|_{\spH^{\ast}}^{2} = \frac{3 \pi^{2}}{4} \text{ for any } n \in \mathbb{N}, \notag
\end{equation}
i.e., $(F_n)_{n}$ remains uniformly bounded w.r.t. $n \in \mathbb{N}$.
Hence, the claim follows.
\end{proof}
\section{Linear system with a frictional damping for the elastic part: Exponential stability}
\label{SECTION_EXPONENTIAL_STABILITY}
In Section \ref{SECTION_LACK_OF_EXPONENTIAL_STABILITY},
if no additional damping for the elastic component is present,
Equations (\ref{SystemTwoFrictionSecondCase1})--(\ref{SystemTwoFrictionSecondCase2})
were shown, in general, not to exhibit an exponential decay rate.
This justifies the necessity of introducing a damping mechanism for the elastic variable.
In the following, we consider a frictional damping
which leads to a system of partial differential equations reading as
\begin{align}
\label{SystemTwoFrictionFirstCase1}
\rho \ddot{u}_i &= \big(A_{iJKj} u_{j,K} - \beta_{Ji} \dot{\tau} -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} )_{,K}\big)_{,J} -
E_{ij} \dot{u}_j,
\\
\label{SystemTwoFrictionFirstCase2}
a \ddot{\tau} &=
-\beta_{Ki} \dot{u}_{i,K} + m_{IJ}\dot{\tau}_{,IJ} + M_{jLKI} u_{j,LKI} +
K_{IJ}\tau_{,IJ}
\end{align}
together with the boundary conditions (\ref{WPBoundaryConditions})
and initial conditions (\ref{WPInitialConditions}).
For this system, we assume
\begin{equation}
E_{ij} \xi_{i} \xi_{j} \geq \alpha \xi_{i} \xi_{i} \text{ and }
m_{IJ} \xi_{I} \xi_{J} \geq \alpha \xi_{I} \xi_{I} \text{ for some } \alpha > 0.
\label{EQUATION_ASSUMPTION_POSITIVE_DEFINITENESS_OF_DAMPING}
\end{equation}
The natural first-order energy associated with the mild solution to
Equations (\ref{SystemTwoFrictionFirstCase1})--(\ref{SystemTwoFrictionFirstCase2}),
(\ref{WPBoundaryConditions})--(\ref{WPInitialConditions}) reads as
\begin{align}
\label{ESFCEnergyDefinition}
\begin{split}
\mathcal{E}(t) &= \frac{1}{2} \int_{\Omega} (\rho \dot{u}_{i}^2 + a
\dot{\tau}^2 + A_{iKLj} u_{i,K} u_{j,L} + C_{iIJKLj} u_{i,JI}
u_{j,LK}) \, \mathrm{d}x
\\
&+ \frac12 \int_{\Omega}
(K_{IJ} \tau_{,I} \tau_{,J} + 2 M_{iJKL} u_{i,KJ} \tau_{,L}) \, \mathrm{d}x,
\end{split}
\end{align}
where, as before, the Einstein's summation convention is applied to terms like $\dot{u}_{i}^{2} = \dot{u}_{i} \dot{u}_{i}$, etc.
Due to the assumptions in Equation (\ref{EQUATION_ASSUMPTION_POSITIVE_DEFINITENESS_OF_DAMPING}),
there exists a number $c_{\mathcal{E}} > 0$ such that
\begin{equation}
\label{ESFCEstimationForEnergy}
\mathcal{E}(t) \ge c_{\mathcal{E}} \int_{\Omega} (\dot{u}_{i}^2 +
\dot{\tau}^2 + u^2_{i,J} + u^2_{i,JK} +
\tau^2_{,I})(t,x) \, \mathrm{d}x.
\end{equation}
\begin{theorem}
\label{ESFCTheorem}
Let $U_0 \in \spH$.
There exist then positive constants $C$ and $c_0$ such that
\begin{equation}
\mathcal{E}(t) \le C \mathcal{E}(0) e^{-c_0 t}
\text{ holds true for every } t \geq 0. \notag
\end{equation}
\end{theorem}
\begin{proof}
Without loss of generality, let $U_0 \in D(\opA)$.
Indeed, if this is not the case,
due to the dense embedding of $D(\opA) \hookrightarrow \spH$,
an appropriate approximating sequence from $D(\opA)$ can be selected.
Further, let $(u, \dot{u}, \tau, \dot{\tau})^{T}$ denote the classical solution to
(\ref{SystemTwoFrictionFirstCase1})--(\ref{SystemTwoFrictionFirstCase2}),
(\ref{WPBoundaryConditions})--(\ref{WPInitialConditions}) for the initial data $U_0$.
We construct a Lyapunov functional $\mathcal{F}$.
Computing $\p \mathcal{E}(t)$, we get
\begin{align*}
\p \mathcal{E}(t) &= \int_{\Omega} \left(-\frac{1}{\rho} E_{ij}
\dot{u}_i \dot{u}_j + \frac{1}{a} m_{IJ} \dot{\tau}_{,IJ}
\dot{\tau}\right) \, \mathrm{d}x
\\
&= -\int_{\Omega} \left(\frac{1}{\rho} E_{ij} \dot{u}_i \dot{u}_j
+ \frac{1}{a} m_{IJ} \dot{\tau}_{,I} \dot{\tau}_{,J}\right) \
\mathrm{d}x\le -\alpha \int_{\Omega} \left(\dot{u}_i^2 +
\dot{\tau}^2_{,I}\right) \, \mathrm{d}x.
\end{align*}
Letting
\begin{equation}
\mathcal{F}_1(t) = \rho \int_{\Omega} \dot{u}_i u_i \, \mathrm{d}x \text{ and }
\mathcal{F}_2(t) = a \int_{\Omega} \dot{\tau} \tau \, \mathrm{d}x,
\end{equation}
after a partial integration, we arrive at
\begin{align*}
\p \mathcal{F}_1(t) &= \int_{\Omega} \left(-A_{iKLj} u_{i,K} u_{j,L} -
\beta_{Ji} \dot{\tau}_{,J} u_i - C_{iIJKLj} u_{i,JI}
u_{j,LK}\right) \, \mathrm{d}x
\\
&+ \int_{\Omega }\left(-
M_{iJKL} \tau_{, L} u_{i, KJ} - E_{ij} \dot{u}_i u_j + \rho
\dot{u}^2\right) \, \mathrm{d}x,
\\
\p \mathcal{F}_2(t) = & \int_{\Omega}\left( \beta_{Ki} \dot{u}_{i}
\tau_{,K} - m_{IJ} \dot{\tau}_{,I} \tau_{,J} - M_{iJKL} \tau_{, L}
u_{i, KJ} - K_{IJ} \tau_{,I}\tau_{,J} + a \dot{\tau}^2\right) \, \mathrm{d}x.
\end{align*}
Let $R_0 > \max\{\beta_{Ji}, m_{IJ}, E_{ij}, a, \rho \}$.
Utilizing the generalized Young's inequality and the first Poincar\'{e}'s inequality, we obtain
\begin{align*}
\p & (\mathcal{F}_1 + \mathcal{F}_2)(t) = \int_{\Omega}\left(
-A_{iKLj} u_{i,K} u_{j,L} - C_{iIJKLj} u_{i,JI} u_{j, LK} - 2M_{iJKL}
\tau_{, L} u_{i, KJ}\right) \, \mathrm{d}x
\\
& + \int_{\Omega} \left(\beta_{Ki} \dot{u}_{i} \tau_{,K} - K_{IJ} \tau_{,I}\tau_{,J} - \beta_{Ji} \dot{\tau}_{,J}
u_{i} - m_{IJ} \dot{\tau}_{,I} \tau_{,J} - E_{ij} \dot{u}_i u_j +
\rho \dot{u}^2 + a \dot{\tau}^2\right) \, \mathrm{d}x
\\
\le & -\alpha \int_{\Omega} \left( u^2_{i,J} + u^2_{i,KL} +
\tau^2_{,I}\right) \ \mathrm{d}x + \frac{R_0}{2} \int_{\Omega}
\left(\frac{1}{\varepsilon} \dot{u}^2_i + \varepsilon
\tau^2_{,K}\right) + \left(\frac{1}{\varepsilon} \dot{\tau}^2_{,J} +
\varepsilon u_i^2 \right) \, \mathrm{d}x
\\
& +\frac{R_0}{2}\int_{\Omega} \left(\frac{1}{\varepsilon} \dot{\tau}^2_{,I} + \varepsilon
\tau_{,J}^2 \right) + \left(\frac{1}{\varepsilon} \dot{u}^2_{i} +
\varepsilon u_{j}^2 \right) + \left(\dot{u}^2 + \dot{\tau}^2\right) \, \mathrm{d}x
\\
\le & -\alpha \int_{\Omega}\left( u^2_{i,J} + u^2_{i,KL} +
\tau^2_{,I}\right) \ \mathrm{d}x + \frac{R_0n}{2} \int_{\Omega}
\left(\frac{1}{\varepsilon} \dot{u}^2 + \varepsilon \tau^2_{,K} +
\frac{1}{\varepsilon} \dot{\tau}^2_{,I}\right) \, \mathrm{d}x
\\
&+ \frac{R_0n}{2} \int_{\Omega}
\left(\varepsilon c_{F} u_{i,J}^2 + \frac{1}{\varepsilon}
\dot{\tau}^2_{,I} + \varepsilon \tau_{,J}^2 + \frac{1}{\varepsilon}
\dot{u}^2 + \varepsilon c_{F} u_{j,I}^2 + \dot{u}^2 +
c_{F}\dot{\tau}_{,I}^2\right) \, \mathrm{d}x.
\end{align*}
Now, selecting $\varepsilon>0$ such that
\begin{equation}
\frac{\alpha}{2} > \frac{R_0 n}{2} \max \{ 2\varepsilon, 2\varepsilon c_{F}\}, \notag
\end{equation}
we find
\begin{align*}
\p (\mathcal{F}_1 + \mathcal{F}_2)(t) &\le -\frac{\alpha}{2}
\int_{\Omega}\left( u^2_{i,J} + u^2_{i,KL} +
\tau^2_{,I}\right) \ \mathrm{d}x
\\
&+ \frac{R_0n}{2} \int_{\Omega}
\left(\frac{2}{\varepsilon}+1\right) \dot{u}^2 +
\left(\frac{2}{\varepsilon}+c_{F}\right) \dot{\tau}^2_{,I} \, \mathrm{d}x.
\end{align*}
Next, we define
\begin{equation}
\mathcal{F}(t) = \mathcal{F}_1(t) + \mathcal{F}_2(t) + N \mathcal{E}(t) \notag
\end{equation}
for some $N> \frac12 + \frac{R_0n}{2\alpha}\left(\frac{2}{\varepsilon} + \max\{1,c_{F}\} \right)$ to be fixed later.
Then
\begin{equation}
\p \mathcal{F}(t) \le -\frac{\alpha}{2} \int_{\Omega}
\left(u^2_{i,J} + u^2_{i,KL} + \tau^2_{,I} + \dot{u}^2 +
\dot{\tau}^2_{,I}\right) \, \mathrm{d}x. \notag
\end{equation}
Using the first Poincar\'{e}'s inequality and Equation (\ref{ESFCEstimationForEnergy}), we obtain
\begin{equation}
\p \mathcal{F}(t) \le -\widehat{C} \mathcal{E}(t). \notag
\end{equation}
Taking into account
\begin{equation}
\big|(\mathcal{F}_1 + \mathcal{F}_2)(t)\big| \le \widetilde{C} \mathcal{E}(t), \notag
\end{equation}
we conclude that
\begin{equation}
(N - \widetilde{C}) \mathcal{E}(t) \le \mathcal{F}(t) \le (N+\widetilde{C}) \mathcal{E}(t). \notag
\end{equation}
If necessary, $N$ is increased to make $N-\widetilde{C}$ positive.
Gronwall's inequality now yields
\begin{equation}
\mathcal{E}(t) \le \frac{1}{N-\widetilde{C}} \mathcal{F}(t) \le
\frac{1}{N-\widetilde{C}} \mathcal{E}(0)
e^{-\widehat{C}/(N+\widetilde{C})} t = C \mathcal{E}(0) e^{-c_0t}. \notag
\end{equation}
This completes the proof.
\end{proof}
\section{Hyperbolized system with a frictional damping in the elastic part: Global existence and exponential stability}
\label{SECTION_HYPERBOLIZED_SYSTEM}
In this last section, we want to study the impact
of a nonlinear frictional damping on `the' equations of thermoelasticity for nonsimple materials.
In contrast to Section \ref{SECTION_EXPONENTIAL_STABILITY},
we replace Equation (\ref{SystemTwoFrictionSecondCase2}) with a Cattaneo-like hyperbolic relaxation
(see, e.g., \cite{TaZh1998}).
Thus, the resulting system becomes purely hyperbolic
and we can use the well-known technique due to Lasiecka and Tataru
\cite{Lasiecka1} to utilize an observability inequality for the linear system
to obtaine a uniform decay for the nonlinear one.
Letting $\mathbf{q} = (q_{i})$ such that $(m_{IJ} q_{I, J})$ represents the heat flux,
for a (small) relaxation parameter $\kappa > 0$,
we consider the following semilinear initial-boundary value problem
\begin{align}
\label{NLSystemWithNonLinDamping1}
\rho \ddot{u}_i &= \left( A_{iJKj} u_{j,K} - \beta_{Ji} \dot{\tau} -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} )_{,K} \right)_{,J} - E(|\dot{u}|) \dot{u}_i,
\\
\label{NLSystemWithNonLinDamping2}
a \ddot{\tau} &=
-\beta_{Ki} \dot{u}_{i,K} + m_{IJ}q_{I,J} + M_{jLKI} u_{j,LKI} + K_{IJ}\tau_{,IJ},
\\
\label{NLSystemWithNonLinDamping3}
\kappa \dot{q}_i &= \dot{\tau}_{,i} - q_{i}
\end{align}
in $\Omega \times (0, \infty)$ subject to the boundary conditions
\begin{align}
\label{NLSystemWithNonLinDampingBoundCond}
u_i = 0, \quad u_{i,J} = 0, \quad \tau = 0 \text{ in } \Gamma \times (0, \infty)
\end{align}
and the initial conditions
\begin{align}
\label{NLSystemWithNonLinDampingInitCond}
\begin{split}
u_i(\cdot, 0) &= u_i^0, \quad \dot{u}_i(\cdot, 0) = \dot{u}_i^0, \\
\tau (\cdot, 0) &= \tau^0, \quad \dot{\tau}(\cdot, 0) = \dot{\tau}^0, \quad q_i(\cdot,0) = q^0_i
\text{ in } \Omega.
\end{split}
\end{align}
Note that if $\kappa = 0$ and $E(\cdot)$ is linear,
Equations (\ref{NLSystemWithNonLinDamping1})--(\ref{NLSystemWithNonLinDamping3})
reduce to (\ref{SystemTwoFrictionSecondCase1})--(\ref{SystemTwoFrictionSecondCase2}).
\subsection{Preliminaries}
We assume the function $E \colon [0, \infty) \to \mathbb{R}$ satisfies the following conditions
\begin{itemize}
\item $E(s) > 0$ for $s > 0$,
\item $s \mapsto E(s)s$ is continuous and monotonically increasing,
\item $\lim\limits_{s \searrow 0} E(s)s = 0$,
\item There exist $m,M$ such that
\begin{equation} \label{NLAssumptionForE4}
0 < m \le E(s) \le M \text{ for } s > 1,
\end{equation}
\item The function $g_i$ given by $g_i(v) = E(|v|)v_i$ is uniformly Lipschitz-continuous, i.e.,
\begin{equation}\label{NLAssumptionForE5}
|g_i(\phi) - g_i(\psi) | \le L |\phi - \psi|
\end{equation}
for all $\phi,\psi\in \mathbb{R}^d$.
\end{itemize}
For Equations (\ref{NLSystemWithNonLinDamping1})--(\ref{NLSystemWithNonLinDamping3}),
we assume the evolution is taking place on the Hilbert space
\begin{align*}
\mathcal{H}_{\kappa} &= \Big\{U = (u, v, \tau, \theta, q)^{T} \in
\big(H^2_{0}(\Omega)\big)^d \times \big(L^2(\Omega)\big)^d \times
H_0^1(\Omega) \times L^2(\Omega) \times (L^2(\Omega))^d\Big\}
\\
&= \spH \times (L^2(\Omega))^d
\end{align*}
equipped with the scalar product
\begin{align*
\langle U , U^* \rangle_{\mathcal{H}_\kappa} = \big\langle
(u,v,\tau,\theta)^T , (u^*,v^*,\tau^*,\theta^*)^T\big\rangle_{\mathcal{H}}
+ \int_{\Omega} \kappa m_{ij} q_i^* q_j \, \mathrm{d}x
\end{align*}
for $U = (u, v, \tau, \theta, q), U^* = (u^*, v^*, \tau^*, \theta^*, q^*)\in \spH_{\kappa}$.
(See Equation (\ref{EQUATION_SCALAR_PRODUCT_WEIGHTED}).)
As before, one can easily prove the inner product
is equivalent with the usual product on the Hilbert space $\mathcal{H}_{\kappa}$.
We rewrite the problem
(\ref{NLSystemWithNonLinDamping1})--(\ref{NLSystemWithNonLinDampingInitCond}) in the abstract form:
\begin{equation}
\label{NLOperatorCauchyProblem}
\dot{U}(t) + \mathcal{K}\big(U(t)\big) = 0 \text{ for } t > 0, \quad U(0) = U_0.
\end{equation}
Here, $\mathcal{K} = - \mathcal{L} + \mathcal{N}$ is a nonlinear operator
with the domain $D(\mathcal{K}) := D(\mathcal{L})$, where
\begin{equation}\label{NLDefineL}
\mathcal{L} U =
\left(
\begin{array}{c}
v_i \\
\frac{1}{\rho} \left( A_{iJKj} u_{j,K} - \beta_{Ji} \theta -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} )_{,K} \right)_{,J} \\
\theta \\
\frac{1}{a}\left( -\beta_{Ki} v_{i,K} + m_{IJ}q_{I,J} + M_{jLKI} u_{j,LKI} + K_{IJ}\tau_{,IJ}\right) \\
\frac{1}{\kappa} \theta_{,i} \\
\end{array}
\right)
\end{equation}
for $u \in D(\mathcal{L})$ and
\begin{equation}
\label{NLDefineN}
\mathcal{N}(U) =
\Big(0,
\frac{1}{\rho} E(|v|) v_i,
0,
0,
\frac{1}{\kappa} q_{i}\Big)^{T} \text{ for } u \in D(\mathcal{N})
\end{equation}
as well as
\begin{align}
D(\mathcal{L}) &= \Big\{U \in \mathcal{H} \,|\, \phantom{-}v_i \in H^1_0(\Omega), \, \theta \in H^1_0(\Omega), \label{NLOperatorLDomain} \\
&\phantom{= \phantom{-}\Big\{U \in \mathcal{H} \,|\,}
\big(A_{iJKj} u_{j,K} - \beta_{Ji} \theta \notag \\
&\phantom{= \Big\{U \in \mathcal{H} \,|\,} - (C_{iJKLIj} u_{j,IL}
+ M_{iJKL} \tau_{,L} )_{,K} \big)_{,J} \in L^2(\Omega) \notag \\
&\phantom{= \phantom{-}\Big\{U \in \mathcal{H} \,|\,}
(m_{JI}q_{J} + M_{jLKI} u_{j,LK} + K_{IJ}\tau_{,J})_{,I} \in L^2(\Omega)\Big\}, \notag \\
D(\mathcal{N}) &= \mathcal{H}_{\kappa}. \label{NLOperatorNDomain}
\end{align}
\begin{definition}
Let $U_{0} \in \mathcal{H}$
and assume $\mathcal{L}$ generates a $C_{0}$-semigroup $\big(S(t)\big)_{t \geq 0}$ on $\mathcal{H}_{\kappa}$.
A function
$U \in C^{0}\big([0, \infty), \mathcal{H}_{\kappa}\big)$
satisfying the integral equation
\begin{equation}
U(t) = S(t) U_{0} + \int_{0}^{t} S(t - s) \mathcal{N}\big(U(s)\big) \mathrm{d}s \notag
\end{equation}
is referred to as a mild solution to Equation (\ref{NLOperatorCauchyProblem}).
If $U$ additionally satisfies
\begin{equation}
U \in H^{1}_{\mathrm{loc}}(0, \infty; \mathcal{H}_{\kappa}) \cap
L^{2}_{\mathrm{loc}}\big(0, \infty; D(\mathcal{L})\big), \notag
\end{equation}
we call $U$ a strong solution.
\end{definition}
By $\nlE(t)$ we denote the corresponding energy
\begin{align*
\nlE(t) &= \frac12 \int_{\Omega} \left(\rho |\dot{u}|^2 + a
\dot{\tau}^2 + A_{iKLj} u_{i,K} u_{j,L} + C_{iIJKLj} u_{i,JI}
u_{j,LK} + K_{IJ}
\tau_{,I} \tau_{,J}\right) \, \mathrm{d}x \\
&+ \frac12 \int_{\Omega}\left( 2 M_{iJKL} u_{i,KJ} \tau_{,L} +
\kappa |q|^2\right) \, \mathrm{d}x
\end{align*}
associated with a mild solution $U$ to nonlinear Equation (\ref{NLOperatorCauchyProblem}).
We observe
\begin{align}
\langle \mathcal{L}V , V \rangle_{\mathcal{H}_\kappa} &= 0
\text{ for any } V \in D(\mathcal{L}) \text{ and }
\label{NLDissipOfNonLinSyst1} \\
\p \nlE(t) &= -\int_{\Omega} E(|v|) |v|^2 + m_{ij} q_i q_j \, \mathrm{d}x
\text{ a.e. in } (0, \infty)
\label{NLDissipOfNonLinSyst2}
\end{align}
for any strong solution $U$ to Equation (\ref{NLOperatorCauchyProblem}).
\begin{remark}
By virtue of Stone's theorem
(cf. \cite{St1932}, \cite[Theorem 3.8.6, p. 105]{TuWei2009}),
Equation (\ref{NLDissipOfNonLinSyst1}) is equivalent with $\mathcal{A}$ being skew-adjoint.
\end{remark}
A straighforward adaption of the proof of Lemma \ref{WPMainLemma} yields:
\begin{lemma}
\label{NLMainLemma}
The operator $\mathcal{L}$ defined in Equation (\ref{NLDefineL}) is the infinitesimal
generator of a $C_0-$semigroup of contractions on $\mathcal{H}_\kappa$.
\end{lemma}
For the nonlinear Cauchy problem (\ref{NLOperatorCauchyProblem}),
we have the following existence and uniqueness theorem.
\begin{theorem}
For $U_0 \in \spH_\kappa$,
there exist unique mild solution $U$ to Equation (\ref{NLOperatorCauchyProblem}).
If $U_{0} \in D(\opK)$, the mild solution is strong.
\end{theorem}
\begin{proof}
By virtue of Equation (\ref{NLAssumptionForE5}),
\begin{align*}
\big\|\mathcal{N}(U^{1}) - \mathcal{N}(U^{2})\big\|_{\mathcal{H}_{\kappa}} &=
\Big(\rho \int_{\Omega} \big|E(|v^{1}|) v^{1} - E(|v^{2}|) v^{2}\big|^{2} \, \mathrm{d}x\Big)^{1/2} \\
&\leq L \sqrt{\rho} \|v^{1} - v^{2}\|_{L^{2}}
\leq L \big\|U^{1} - U^{2}\|_{\mathcal{H}_{\kappa}} \notag
\end{align*}
for $U^{i} = (u^{i}, v^{i}, \tau^{i}, \theta^{i}, q^{i})^{T}$, $i = 1, 2$,
i.e., the mapping $\mathcal{N}$ is globally Lipschitzian on $\mathcal{H}_{\kappa}$.
Hence, the first claim is a direct consequence of \cite[Theorem 1.2, p. 184]{Pazy1}.
Since $\mathcal{H}_{\kappa}$, being a Hilbert space, is reflexive,
the second claim readily follows from \cite[Theorem 1.4, p. 189]{Pazy1}.
\end{proof}
By utilizing Lemma \ref{NLMainLemma}
and constructing a Lyapunov's functional similar to that one in Section \ref{SECTION_EXPONENTIAL_STABILITY} (see also \cite{Po2014}),
we further obtain the following linear stability theorem.
\begin{theorem}
\label{THEOREM_LINEAR_HYPERBOLIZED_SYSTEM_EXPONENTIAL_STABILITY}
For
\begin{equation*}
\label{NLDefineD}
\mathcal{D} \colon \mathcal{H}_{\kappa} \to \mathcal{H}_{\kappa}, \quad
\mathcal{D} U := \Big(
0,
\frac{1}{\rho} E_{ij} v_j,
0,
0,
\frac{1}{\kappa} q_{i}
\Big)^{T} \text{ for } U \in \mathcal{H}_{\kappa},
\end{equation*}
the unique (mild or classical) solution to the Cauchy problem
\begin{equation}
\dot{U}(t) = \opL U(t) - \mathcal{D}U(t) \text{ for } t > 0, \quad
U(0) = U_0
\end{equation}
is exponentially stable.
\end{theorem}
\subsection{Uniform stability}
Recalling Equation (\ref{NLDissipOfNonLinSyst1})
and using Theorem \ref{THEOREM_LINEAR_HYPERBOLIZED_SYSTEM_EXPONENTIAL_STABILITY},
we apply Theorem \ref{APPTheorem} in Appendix \ref{SECTION_APPENDIX} to obtain
the following observability result for the linear part of Equation (\ref{NLOperatorCauchyProblem}).
\begin{theorem}
\label{ObservInequality}
There exist a time period $T_{0} > 0$ and a positive constant $C$ such that
for any $T \geq T_{0}$
every mild solution $(u, \dot{u}, \tau, \dot{\tau}, q)^{T}$ to the Cauchy problem
\begin{equation}
\dot{U}(t) = \opL U(t) \text{ for } t > 0, \quad
U(0) = U_{0} \notag
\end{equation}
satisfies
\begin{equation}
\mathcal{E} (0) \le C \int_0^T
\int_{\Omega} \big(|\dot{u}|^2 + |q|^2\big) \, \mathrm{d}x \mathrm{d}t. \notag
\end{equation}
\end{theorem}
In the following, we adapt the techniques introduced by Lasiecka and Tataru \cite{Lasiecka1}.
The technical difficulties in our case are due to the big system size
and mixed-order structure of Equations (\ref{NLSystemWithNonLinDamping1})--(\ref{NLSystemWithNonLinDamping3}).
\begin{lemma}
\label{NLLemaEstimationForEnergy}
For any $T > 0$ and $K > 0$,
there exists a positive constant $C$ such that
\begin{equation}
\nlE(T) \le C \int_0^T \int_{\Omega}\left( E^2(|\dot{u}|)|\dot{u}|^2
+ |\dot{u}|^2 + |q|^2 \right) \, \mathrm{d}x \mathrm{d}t \notag
\end{equation}
for any strong solution $\Psi = (u_i, \dot{u}_i, \tau, \dot{\tau}, q)^{T}$
to Equation (\ref{NLOperatorCauchyProblem}) additionally satisfying $\nlE(0) \le K$.
\end{lemma}
\begin{proof}
Suppose the converse is true which means there exist $T > 0$ and $K > 0$ together with
a sequence of strong solutions
$\big((u^{(m)}, \dot{u}^{(m)}, \tau^{(m)}, \dot{\tau}^{(m)}, q^{(m)})\big)_{m}^{T}$ such that $\nlE_m(0)\le K$
to the Cauchy problem (\ref{NLOperatorCauchyProblem}),
whereas
\begin{equation}
\frac{\nlE_m(T)}{\int_0^T \int_{\Omega}
E^2(|\dot{u}^{(m)}|)|\dot{u}^{(m)}|^2 + |\dot{u}^{(m)}|^2 +
|q^{(m)}|^2 \, \mathrm{d}x \mathrm{d}t} \to \infty \text{ as } m \to \infty. \notag
\end{equation}
Denote
\begin{equation}
\Pi_m(\cdot) :=
\int_0^T \int_{\Omega} E^2(|\dot{u}^{(m)}|)|\dot{u}^{(m)}|^2 + |\dot{u}^{(m)}|^2 +
|q^{(m)}|^2 \, \mathrm{d}x \mathrm{d}t. \notag
\end{equation}
Using Equation (\ref{NLDissipOfNonLinSyst2}), we then get
\begin{equation*}
\frac{K}{\Pi_m} \ge \frac{\nlE_m(0)}{\Pi_m} \ge
\frac{\nlE_m(T)}{\Pi_m} \to \infty \text{ as } m \to \infty,
\end{equation*}
and, thus, $\Pi_m \to 0$, i.e.,
\begin{equation*}
\int_{0}^{T} \|\dot{u}^{(m)}\|_{L^2}^{2} \mathrm{d}t \to 0 \text{ and }
\int_{0}^{T} \|q^{(m)}\|_{L^2}^{2} \mathrm{d}t \to 0 \text{ as } m \to \infty.
\end{equation*}
Letting
\begin{equation}
\nu_m = \sqrt{\nlE_m(0)}, \quad
\bar{u}^{(m)} = \frac{u^{(m)}}{\nu_m}, \quad
\bar{\tau}^{(m)} = \frac{\tau^{(m)}}{\nu_m}, \quad
\bar{q}^{(m)} = \frac{q^{(m)}}{\nu_m}, \notag
\end{equation}
we observe $\big(\bar{u}^{(m)}, \bar{\tau}^{(m)}, \bar{q}^{(m)}\big)^{T}$ is a strong solution to
\begin{align}
\label{NLSystemNormalized1}
\rho \ddot{u}_i &= \big( A_{iJKj} u_{j,K} - \beta_{Ji} \dot{\tau} -
(C_{iJKLIj} u_{j,IL} + M_{iJKL} \tau_{,L} )_{,K}\big)_{,J} -
E(|\dot{u}|\nu_m) \dot{u}_i
\\
\label{NLSystemNormalized2}
a \ddot{\tau} &=
-\beta_{Ki} \dot{u}_{i,K} + m_{IJ}q_{I,J} + M_{jLKI} u_{j,LKI} +
K_{IJ}\tau_{,IJ},
\\
\label{NLSystemNormalized3}
\kappa \dot{q}_i &= \dot{\tau}_{,i} - q_{i}
\end{align}
in $\Omega \times (0, \infty)$ subject to the boundary conditions
\begin{align*}
u_i = 0, \quad u_{i,J} = 0, \quad \tau = 0 \text{ on } \Gamma \times (0, \infty)
\end{align*}
and the initial conditions
\begin{align}
u_i = \frac{u_i^0}{\nu_{m}}, \quad
\dot{u}_i = \frac{\dot{u}_i^0}{\nu_{m}}, \quad
\tau = \frac{\tau^0}{\nu_{m}}, \quad
\dot{\tau} = \frac{\dot{\tau}^0}{\nu_{m}}, \quad
q_i = \frac{q^0_i}{\nu_{m}} \text{ in } \Omega \times \{0\}. \notag
\end{align}
By $\normE_m(t)$ denote the energy of the solution to system
(\ref{NLSystemNormalized1})--(\ref{NLSystemNormalized3}).
Note that
\begin{equation}
\label{NLInitEnergyOfNonLinSys}
\normE_m(0) = \frac{\nlE_m(0)}{\nu_m} = 1.
\end{equation}
Now,
\begin{align*}
\int_0^T \int_{\Omega} \big(E^2(|\dot{u}^{(m)}|)|\dot{\bar{u}}^{(m)}|^2 +
|\dot{\bar{u}}^{(m)}|^2 + |\bar{q}^{(m)}|^2 \big) \mathrm{d}x \mathrm{d}t =
\frac{\Pi_m}{\nu^2_m} = \frac{\Pi_m}{\nlE_m(0)} \to 0
\end{align*}
as $m \to \infty$. Therefore,
\begin{equation}
\label{NLTauToZero}
\int_{0}^{T} \|\dot{\bar{u}}^{(m)}\|_{L^2}^{2} \mathrm{d}t \to 0 \text{ and }
\int_{0}^{T} \|\bar{q}^{(m)}\|_{L^2}^{2} \mathrm{d}t
\to 0 \text{ as } m \to \infty
\end{equation}
and also
\begin{equation}\label{NLDampingToZero}
\int_{0}^{T} \big\|E\big(|\dot{u}^{(m)}|\big) |\dot{\bar{u}}^{(m)}|\big\|_{L^2}^{2} \mathrm{d}t \to 0
\text{ as } m \to \infty.
\end{equation}
Now, consider the following linear system
\begin{align}
\label{NLCorrespLinSystem1}
\rho \ddot{v}_i &= \big(A_{iJKj} v_{j,K} - \beta_{Ji} \dot{\omega} -
(C_{iJKLIj} v_{j,IL} + M_{iJKL} \omega_{,L} )_{,K}\big)_{,J},
\\
\label{NLCorrespLinSystem2}
a \ddot{\omega} &=
-\beta_{Ki} \dot{v}_{i,K} + m_{IJ}r_{I,J} + M_{jLKI} v_{j,LKI} +
K_{IJ}\omega_{,IJ},
\\
\label{NLCorrespLinSystem3}
\kappa \dot{r}_i &= \dot{\omega}_{,i} - r_{i}
\end{align}
subject to the boundary conditions
\begin{equation}
\label{NLCorrespLinSystemBoundCond}
v_i = 0, \quad \ v_{i,J} = 0, \quad \omega = 0 \text{ on } \Gamma \times (0, \infty)
\end{equation}
and the initial conditions
\begin{align}
v_i = \frac{u_i^0}{\nu_{m}}, \quad
\dot{v}_i = \frac{\dot{u}_i^0}{\nu_{m}}, \quad
\omega = \frac{\tau^0}{\nu_{m}}, \quad
\dot{\omega} = \frac{\dot{\tau}^0}{\nu_{m}}, \quad
r_i = \frac{q^0_i}{\nu_{m}} \text{ in } \Omega \times \{0\}.
\label{NLCorrespLinSystemInitCond}
\end{align}
Letting
$\big(\bar{v}^{(m)}, \bar{\omega}^{(m)}, \bar{r}^{(m)}\big)^{T}$
denote the strong solution to Equations
(\ref{NLCorrespLinSystem1})--(\ref{NLCorrespLinSystemInitCond}) and
$\lE_m(t)$ be the corresponding energy,
the initial conditions (\ref{NLCorrespLinSystemInitCond}) imply
\begin{equation}
\label{NLEqualityOfEnergies}
\lE_m(0) = \normE_m(0). \notag
\end{equation}
Exploring Equations (\ref{NLSystemNormalized1})--(\ref{NLSystemNormalized3})
and (\ref{NLCorrespLinSystem1})--(\ref{NLCorrespLinSystemInitCond}),
we deduce that
$w^{(m)}_i = \bar{u}^{(m)}_i - \bar{v}^{(m)}_i$,
$\chi^{(m)} = \bar{\tau}^{(m)} - \bar{\omega}^{(m)}$,
$p^{(m)}_i = \bar{q}^{(m)}_i - \bar{r}^{(m)}_i$
strongly solves the `incremental' system
\begin{align}
\label{NLMixedSystem1}
\rho \ddot{w}_i &= \big(A_{iJKj} w_{j,K} - \beta_{Ji} \dot{\chi} -
(C_{iJKLIj} w_{j,IL} + M_{iJKL} \chi_{,L} )_{,K}\big)_{,J}
\\
\notag &- E(|\dot{\bar{u}}^{(m)}\nu_m|) \dot{\bar{u}}_i^{(m)}
\\
\label{NLMixedSystem2}
a \ddot{\chi} &=
-\beta_{Ki} \dot{w}_{i,K} + m_{IJ}p_{I,J} + M_{jLKI} w_{j,LKI} +
K_{IJ}\chi_{,IJ},
\\
\label{NLMixedSystem3}
\kappa \dot{p}_i &= \dot{\chi}_{,i} - p_{i},
\end{align}
together with the boundary conditions
\begin{equation}
\label{NLMixedSystemBoundVal}
w_i = 0, \quad w_{i,J} = 0, \quad \chi = 0 \text{ on } \Gamma \times (0, \infty)
\end{equation}
and the initial conditions
\begin{equation}
\label{NLMixedSystemInitVal}
w_i(\cdot, 0) = \dot{w}_i(\cdot, 0) = 0, \quad
\chi(\cdot, 0) = \dot{\chi}(\cdot,0) = 0, \quad
p_i(\cdot, 0) = 0 \text{ in } \Omega.
\end{equation}
With $\mE_m(t)$ denoting the associated energy,
Equation (\ref{NLMixedSystemInitVal}) implies $\mE_m(0) = 0$,
for which we easily verify
\begin{equation}
\p \mE_m(t)
= \int_{\Omega} \Big(-E(|\dot{u}^{(m)}|)
\dot{\bar{u}}_i^{(m)} \dot{w}_i^{(m)} - m_{ij} \bar{q}^{(m)}_j p_i\Big) \, \mathrm{d}x.
\notag
\end{equation}
Therefore,
\begin{equation}
\begin{split}
\mE_m(t) &= \mE_m(t) - \mE_m(0) \\
&= -\int_0^t \int_{\Omega} \left(
E(|\dot{u}^{(m)}|) \dot{\bar{u}}_i^{(m)} \dot{w}_i^{(m)} + m_{IJ}
\bar{q}^{(m)}_j p_i \right) \, \mathrm{d}x \mathrm{d}t.
\end{split}
\label{EQUATION_DEFINITION_ENERRGY_DIFF}
\end{equation}
Note that $\int_{0}^{T} \|\dot{w}^{(m)}\|_{L^2}^{2} \mathrm{d}t$ and
$\int_{0}^{T} \|p^{(m)}\|_{L^2}^{2} \mathrm{d}t$ are bounded w.r.t. $m \in \mathbb{N}$.
Indeed, a straighforward computation yields
\begin{equation}
\p \lE_m(t) = 0. \notag
\end{equation}
Thus, $\lE_m(t) = \lE_m(0)$ and, therefore,
\begin{equation}
\frac{\rho}{2} \int_{\Omega} |\dot{\bar{v}}^{(m)}|^2 \, \mathrm{d}x +
\frac{\kappa\alpha}{2} \int_{\Omega} |\bar{r}^{(m)}|^2 \, \mathrm{d}x
\le \lE_m(t) = \lE_m(0) = \normE_m(0) = 1, \notag
\end{equation}
whence
\begin{equation}
\int_0^T \int_{\Omega} \Big(\frac{\rho}{2}
|\dot{\bar{v}}^{(m)}|^2 + \frac{\kappa\alpha}{2} |\bar{r}^{(m)}|^2\Big) \mathrm{d}x \mathrm{d}t \le T.
\notag
\end{equation}
Estimating
\begin{align*}
\Big(\int_{0}^{T} \|\dot{w}^{(m)}_i\|_{L^2}^{2} \mathrm{d}t\Big)^{1/2} &\le
\Big(\int_{0}^{T} \|\dot{\bar{u}}^{(m)}_i\|_{L^2}^{2} \mathrm{d}t\Big)^{1/2} +
\Big(\int_{0}^{T} \|\dot{\bar{v}}^{(m)}_i\|_{L^2}^{2} \mathrm{d}t\Big)^{1/2}
\text{ and } \\
\Big(\int_{0}^{T} \|p^{(m)}_i\|_{L^2}^{2} \mathrm{d}t\Big)^{1/2} &\le
\Big(\int_{0}^{T} \|\bar{q}^{(m)}_i\|_{L^2}^{2} \mathrm{d}t\Big)^{1/2} +
\Big(\int_{0}^{T} \|\bar{r}^{(m)}_i\|_{L^2}^{2} \mathrm{d}t\Big)^{1/2}
\end{align*}
and recalling Equation (\ref{NLTauToZero}), we arrive at the boundness of
$\int_{0}^{T} \|\dot{w}^{(m)}\|_{L^2}^{2} \mathrm{d}t$ and
$\int_{0}^{T} \|p^{(m)} \|_{L^2}^{2} \mathrm{d}t$ w.r.t. $m \in \mathbb{N}$.
Using this fact and applying Cauchy \& Schwarz' inequality to Equation (\ref{EQUATION_DEFINITION_ENERRGY_DIFF}),
we further find a positive constant $C$ such that
\begin{align*}
\mE_m(t) &\le \Big(\int_0^t\int_{\Omega} E^2(|\dot{u}^{(m)}|)
|\dot{\bar{u}}_i^{(m)}|^2 \, \mathrm{d}x\mathrm{d}t\Big)^{1/2}
\Big(\int_0^t\int_{\Omega}
|\dot{w}_i^{(m)}|^2 \, \mathrm{d}x \mathrm{d}t\Big)^{1/2}
\\
&\phantom{\leq} + \max_{i,j} |m_{ij}| \Big(\int_0^t\int_{\Omega} |\bar{q}^{(m)}_j|^2
\, \mathrm{d}x\mathrm{d}t\Big)^{1/2}
\Big(\int_0^t\int_{\Omega} |p_i|^2 \ \mathrm{d}x\mathrm{d}t
\Big)^{1/2}
\\
&\le C \Big(\int_0^T\int_{\Omega} E^2(|\dot{u}^{(m)}|)
|\dot{\bar{u}}_i^{(m)}|^2 \, \mathrm{d}x \mathrm{d}t\Big)^{1/2}
+ C
\Big(\int_0^T\int_{\Omega} |\bar{q}^{(m)}_j|^2 \, \mathrm{d}x \mathrm{d}t\Big)^{1/2}
\\
&\le C \left(\Big(\int_{0}^{T} \|E(|\dot{u}^{(m)}|) |\dot{\bar{u}}^{(m)}|\|_{L^2}^{2} \mathrm{d}t\Big)^{1/2}
+ \Big(\int_{0}^{T}\|\bar{q}^{(m)}\|_{L^2}^{2} \mathrm{d}t\Big)^{1/2}\right).
\end{align*}
Hence,
\begin{equation}
\label{NLMixedEnergyToZero}
\max_{t \in [0, T]} \mE_m(t) \le
C \bigg(\Big(\int_{0}^{T} \|E(|\dot{u}^{(m)}|) |\dot{\bar{u}}^{(m)}|\|_{L^2}^{2} \mathrm{d}t\Big)^{1/2}
+\Big(\int_{0}^{T}\|\bar{q}^{(m)}\|_{L^2}^{2} \mathrm{d}t\Big)^{1/2}\bigg).
\end{equation}
Further, using the trivial inequality
$a^2 = (a-b + b)^2 \le 2((a-b)^2 + b^2)$, we get
\begin{align*}
\int_{\Omega} \big(|\dot{\bar{v}}^{(m)}|^2 + |\bar{r}^{(m)}|^2 \big) \mathrm{d}x
&\le 2\int_{\Omega} \big(|\dot{\bar{u}}^{(m)} -
\dot{\bar{v}}^{(m)}|^2 + |\bar{q}^{(m)} - \bar{r}^{(m)}|^2 +
|\dot{\bar{u}}^{(m)}|^2 + |\bar{q}^{(m)}|^2\big) \mathrm{d}x
\\
&= 2 \int_{\Omega} \big(|\dot{w}^{(m)}|^2 + |p^{(m)}|^2 +
|\dot{\bar{u}}^{(m)}|^2 + |\bar{q}^{(m)}|^2\big) \mathrm{d}x
\\
&\le C_1 \mE_m(t) + 2
\int_{\Omega} \big(|\dot{\bar{u}}^{(m)}|^2 + |\bar{q}^{(m)}|^2\big) \mathrm{d}x
\\
&\le C_1 \max_{t \in [0, T]} \mE_m(t) +
2 \int_{\Omega} \big(|\dot{\bar{u}}^{(m)}|^2 + |\bar{q}^{(m)}|^2\big) \mathrm{d}x.
\end{align*}
Integrating the latter inequality w.r.t. $t$ and
using Equations (\ref{NLTauToZero}), (\ref{NLDampingToZero}) and (\ref{NLMixedEnergyToZero}), we obtain
\begin{equation}
\int_0^T \int_{\Omega} \big(|\dot{\bar{v}}^{(m)}|^2 + |\bar{r}^{(m)}|^2\big)
\mathrm{d}x \mathrm{d}t \to 0 \text{ as } m \to \infty. \notag
\end{equation}
On the other hand, recalling Equations
(\ref{NLInitEnergyOfNonLinSys}), (\ref{NLEqualityOfEnergies}) and Theorem \ref{ObservInequality},
we find
\begin{equation}
1 = \normE_m(0) = \lE_m(0) \le c \int_{\Omega}
\big(|\dot{\bar{v}}^{(m)}|^2 + |\bar{r}^{(m)}|^2\big) \mathrm{d}x \mathrm{d}t \to 0
\text{ as } m \to \infty, \notag
\end{equation}
which is a contradiction to our original assumption.
\end{proof}
To proceed further, let us introduce some notations.
Thanks to our assumptions on the function $E$,
according to \cite{Lasiecka1}, there exists a real-valued function $h \colon [0, \infty) \to [0, \infty)$,
being concave, strictly increasing and satisfying $h(0) = 0$ and
\begin{equation}
\label{NLDefinOfH}
h(s^2E(s)) \ge s^2 + E^2(s)s^2 \text{ for } s \in [0, 1].
\end{equation}
Further, we define a function $r$ by means of
\begin{equation}
\label{NLDefinOfR}
r(s) = h\left(\frac{s}{|Q|}\right) \text{ for } s \ge 0,
\end{equation}
where $Q = \Omega \times (0, T)$ and $|Q|$ is the standard Borel measure of $Q$.
Further, let
\begin{equation}
\label{NLDefinOfP}
p(s) = (cI+r)^{-1} \left(Ms\right) \text{ for } s \ge 0,
\end{equation}
where $M$ and $c$ are some positive constants to be defined later.
Finally, let
\begin{equation}
\label{NLDefinOfQ}
q(s) = s - (I+p)^{-1} \left(s\right) \text{ for } s \ge 0.
\end{equation}
We quote the following lemma due to Lasiecka and Tataru \cite{Lasiecka1} we use below.
\begin{lemma}
\label{LemmaLasiecka}
Let the functions $p, q$ be defined as above.
For any number sequence $(s_n)_n \subset (0, \infty)$ with $s_{m+1} + p(s_{m+1}) \le s_m$,
we have
\begin{equation}
s_m \le S(m) \text{ for every } m\in \mathbb{N}, \notag
\end{equation}
where $S(t)$ is a solution of the scalar Cauchy problem
\begin{equation}
\label{NLDefineS}
\dot{S}(t) + q\big(S(t)\big) = 0 \text{ for } t > 0, \quad
S(0) = s_0.
\end{equation}
Moreover, if $p$ satisfies $p(s) > 0$ for $s > 0$, then
\begin{equation}
\lim_{t \to \infty} S(t) = 0. \notag
\end{equation}
\end{lemma}
Now, we are in position to prove the uniform stability result for the Cauchy problem (\ref{NLOperatorCauchyProblem}).
\begin{theorem}
\label{NLTheoremOnStability}
For a sufficiently large number $T > 0$,
we have
\begin{equation}
\nlE(t) \le S\left(\frac{t}{T}-1\right) \text{ for } t > T. \notag
\end{equation}
Moreover,
$S(t) \to 0$ as $t \to \infty$, where $S(t)$ solves the ODE (\ref{NLDefineS}).
\end{theorem}
\begin{proof}
Consider the sets
\begin{align*}
Q_1 &= \big\{(x, t) \in \Omega \times (0,T) \,|\, \big|\dot{u}(x, t)\big| > 1 \big\},
\\
Q_2 &= \big\{(x, t) \in \Omega \times (0,T) \,|\, \big|\dot{u}(x, t)\big| \le 1 \big\}.
\end{align*}
Due to the additivity property of Lebesgue integral, we have
\begin{align*}
\int_0^T \int_{\Omega} \big(E^2(|\dot{u}|)|\dot{u}|^2 + |\dot{u}|^2 \big) \mathrm{d}x \mathrm{d}t
&= \int_{Q_1} \big(E^2(|\dot{u}|)|\dot{u}|^2 + |\dot{u}|^2\big) \mathrm{d}x \mathrm{d}t
\\
&+\int_{Q_2} \big(E^2(|\dot{u}|)|\dot{u}|^2 + |\dot{u}|^2\big) \mathrm{d}x \mathrm{d}t.
\end{align*}
On the set with $|\dot{u}| > 1$, Equation (\ref{NLAssumptionForE4}) implies
\begin{align*}
E^2(|\dot{u}|)|\dot{u}|^2 \le M E(|\dot{u}|)|\dot{u}|^2 \text{ and }
|\dot{u}|^2 \le \frac{1}{m} E(|\dot{u}|)|\dot{u}|^2
\end{align*}
with $m, M$ from Equation (\ref{NLAssumptionForE4}).
Therefore,
\begin{align}
\notag
\int_{Q_1} \big(E^2(|\dot{u}|)|\dot{u}|^2 + |\dot{u}|^2\big) \mathrm{d}x \mathrm{d}t
&\le \left(M + \frac{1}{m}\right) \int_{Q_1} E(|\dot{u}|)|\dot{u}|^2 \, \mathrm{d}x \mathrm{d}t
\\
\label{NLIneqForQ1}
&\le c_1 \int_{Q} E(|\dot{u}|)|\dot{u}|^2 \, \mathrm{d}x \mathrm{d}t,
\end{align}
where $c_1 = M + m^{-1}$.
Next, consider the integral over $Q_2$.
Using Equation (\ref{NLDefinOfH}), we estimate
\begin{align*}
\int_{Q_2} E^2(|\dot{u}|)|\dot{u}|^2 + |\dot{u}|^2 \, \mathrm{d}x \mathrm{d}t
\le \int_{Q_2} h(E(|\dot{u}|)|\dot{u}|^2) \, \mathrm{d}x \mathrm{d}t.
\end{align*}
Now, Jensen's inequality yields
\begin{align*}
\int_{Q_2} h(E(|\dot{u}|)|\dot{u}|^2)\ \mathrm{d}x \mathrm{d}t \le |Q| h\left(
\int_{Q_2} \frac{E(|\dot{u}|)|\dot{u}|^2}{|Q|} \, \mathrm{d}x \mathrm{d}t\right),
\end{align*}
and, recalling the definition of function $r$ in Equation (\ref{NLDefinOfR}), we get
\begin{align}
\notag
\int_{Q_2} \big(E^2(|\dot{u}|)|\dot{u}|^2 + |\dot{u}|^2 \big) \mathrm{d}x \mathrm{d}t
&\le |Q| r\left(\int_{Q_2} E(|\dot{u}|)|\dot{u}|^2 \, \mathrm{d}x \mathrm{d}t\right) \\
\label{NLIneqForQ2}
&\le |Q| r\left(\int_{Q} E(|\dot{u}|)|\dot{u}|^2 \, \mathrm{d}x \mathrm{d}t\right).
\end{align}
Combining Equations (\ref{NLIneqForQ1}) and (\ref{NLIneqForQ2}), we obtain
\begin{align*}
\int_0^T \int_{\Omega} \big(E^2(|\dot{u}|)|\dot{u}|^2 + |\dot{u}|^2\big) \mathrm{d}x \mathrm{d}t \le
\left(c_1 \mathrm{I} + |Q|r\right) \left(\int_{Q} E(|\dot{u}|)|\dot{u}|^2 \,
\mathrm{d}x \mathrm{d}t\right),
\end{align*}
where $\mathrm{I}$ is the identity function, i.e.,
\begin{align*}
\left(c_1 \mathrm{I} + |Q|r\right)(s) = c_1 s + |Q| r(s) \text{ for } s \ge 0.
\end{align*}
Further, exploiting Lemma \ref{NLLemaEstimationForEnergy} and the monotonicity of $r$, we get
\begin{align*}
\nlE(T) &\le C \left(\left( c_1 \mathrm{I} + |Q|r\right) \left(\int_{Q} E(|\dot{u}|)|\dot{u}|^2 \, \mathrm{d}x \mathrm{d}t\right)
+ \int_Q |q|^2 \ \mathrm{d}x \mathrm{d}t\right)
\\
&\le C \left(c_2 \int_{Q} \big(E(|\dot{u}|)|\dot{u}|^2 + |q|^2 \big) \mathrm{d}x \mathrm{d}t +
|Q|r\left(\int_{Q} E(|\dot{u}|)|\dot{u}|^2 \, \mathrm{d}x \mathrm{d}t\right)\right)
\\
&\le C \left(c_2 \int_{Q} \big(E(|\dot{u}|)|\dot{u}|^2\ + |q|^2\big) \mathrm{d}x \mathrm{d}t +
|Q|r\left(\int_{Q} \big(E(|\dot{u}|)|\dot{u}|^2 + |q|^2\big) \mathrm{d}x \mathrm{d}t\right)\right)
\\
&= C \left((c_2 \mathrm{I} + |Q|r) \left(\int_{Q} \big(E(|\dot{u}|)|\dot{u}|^2 + |q|^2\big) \mathrm{d}x \mathrm{d}t\right)\right),
\end{align*}
where $c_2 = \max\{c_1, 1\}$.
Thus, we obtain
\begin{align*}
\frac{1}{C|Q|} \nlE(T) \le \left(\frac{c_2}{|Q|} \mathrm{I} + r\right)
\left(\int_{Q} \big(E(|\dot{u}|)|\dot{u}|^2 + |q|^2\big) \mathrm{d}x \mathrm{d}t\right),
\end{align*}
or, equivalently,
\begin{align*}
M \nlE(T) \le (c \mathrm{I} + r)(\nlE(0) - \nlE(T)).
\end{align*}
Now, recalling Equation (\ref{NLDefinOfP}), we get
\begin{align*}
p(\nlE(T)) \le \nlE(0) - \nlE(T),
\end{align*}
or, equivalently,
\begin{align*}
p(\nlE(T)) + \nlE(T) \le \nlE(0).
\end{align*}
It can easily be seen that, replacing $0, T$ with $mT, (m+1)T$ above, we obtain
\begin{align*}
p\big(\nlE((m + 1)T)) + \nlE((m + 1)T\big) \le \nlE(mT),
\end{align*}
Note that the constants $C, c, c_1, c_2, |Q|$ and the functions $h,r,p,q$ remain the same.
As before, applying Lemma \ref{LemmaLasiecka} with $s_m=\nlE(mT)$,
we conclude
\begin{equation}
E(mT) \le S(m) \text{ for } m \in \mathbb{N} \text{ and }
\lim_{t\to\infty} S(t) = 0. \notag
\end{equation}
It can further be easily shown that $q(s) \ge 0$ for $s\ge 0$ and, thus,
$S(\cdot)$ is monotonically decreasing.
Therefore, for $t = mT + \tau$ with $\tau \in [0, T]$, we have
\begin{align*}
E(t) \le E(mT) \le S(m) = S\left(\frac{t-\tau}{T}\right) \le S\left(\frac{t}{T} - 1\right),
\end{align*}
which finishes the proof.
\end{proof}
\section*{Funding}
This work has been partially funded by the Young Scholar Fund at the University of Konstanz, Germany.
MP has been supported by the ERC-CZ Project LL1202 `MOdelling REvisited + MOdel REduction' at Charles University in Prague, Czech Republic and
the Deutsche Forschungsgemeinschaft (DFG) through CRC 1173 at Karlsruhe Institute of Technology, Germany.
|
\section{The ALICE Collaboration}
\begingroup
\small
\begin{flushleft}
J.~Adam\Irefn{org39}\And
D.~Adamov\'{a}\Irefn{org85}\And
M.M.~Aggarwal\Irefn{org89}\And
G.~Aglieri Rinella\Irefn{org35}\And
M.~Agnello\Irefn{org111}\And
N.~Agrawal\Irefn{org48}\And
Z.~Ahammed\Irefn{org134}\And
S.~Ahmad\Irefn{org19}\And
S.U.~Ahn\Irefn{org69}\And
S.~Aiola\Irefn{org138}\And
A.~Akindinov\Irefn{org59}\And
S.N.~Alam\Irefn{org134}\And
D.S.D.~Albuquerque\Irefn{org122}\And
D.~Aleksandrov\Irefn{org81}\And
B.~Alessandro\Irefn{org111}\And
D.~Alexandre\Irefn{org102}\And
R.~Alfaro Molina\Irefn{org65}\And
A.~Alici\Irefn{org12}\textsuperscript{,}\Irefn{org105}\And
A.~Alkin\Irefn{org3}\And
J.R.M.~Almaraz\Irefn{org120}\And
J.~Alme\Irefn{org18}\textsuperscript{,}\Irefn{org37}\And
T.~Alt\Irefn{org42}\And
S.~Altinpinar\Irefn{org18}\And
I.~Altsybeev\Irefn{org133}\And
C.~Alves Garcia Prado\Irefn{org121}\And
C.~Andrei\Irefn{org79}\And
A.~Andronic\Irefn{org98}\And
V.~Anguelov\Irefn{org95}\And
T.~Anti\v{c}i\'{c}\Irefn{org99}\And
F.~Antinori\Irefn{org108}\And
P.~Antonioli\Irefn{org105}\And
L.~Aphecetche\Irefn{org114}\And
H.~Appelsh\"{a}user\Irefn{org54}\And
S.~Arcelli\Irefn{org27}\And
R.~Arnaldi\Irefn{org111}\And
O.W.~Arnold\Irefn{org94}\textsuperscript{,}\Irefn{org36}\And
I.C.~Arsene\Irefn{org22}\And
M.~Arslandok\Irefn{org54}\And
B.~Audurier\Irefn{org114}\And
A.~Augustinus\Irefn{org35}\And
R.~Averbeck\Irefn{org98}\And
M.D.~Azmi\Irefn{org19}\And
A.~Badal\`{a}\Irefn{org107}\And
Y.W.~Baek\Irefn{org68}\And
S.~Bagnasco\Irefn{org111}\And
R.~Bailhache\Irefn{org54}\And
R.~Bala\Irefn{org92}\And
S.~Balasubramanian\Irefn{org138}\And
A.~Baldisseri\Irefn{org15}\And
R.C.~Baral\Irefn{org62}\And
A.M.~Barbano\Irefn{org26}\And
R.~Barbera\Irefn{org28}\And
F.~Barile\Irefn{org32}\And
G.G.~Barnaf\"{o}ldi\Irefn{org137}\And
L.S.~Barnby\Irefn{org102}\textsuperscript{,}\Irefn{org35}\And
V.~Barret\Irefn{org71}\And
P.~Bartalini\Irefn{org7}\And
K.~Barth\Irefn{org35}\And
J.~Bartke\Irefn{org118}\Aref{0}\And
E.~Bartsch\Irefn{org54}\And
M.~Basile\Irefn{org27}\And
N.~Bastid\Irefn{org71}\And
S.~Basu\Irefn{org134}\And
B.~Bathen\Irefn{org55}\And
G.~Batigne\Irefn{org114}\And
A.~Batista Camejo\Irefn{org71}\And
B.~Batyunya\Irefn{org67}\And
P.C.~Batzing\Irefn{org22}\And
I.G.~Bearden\Irefn{org82}\And
H.~Beck\Irefn{org54}\textsuperscript{,}\Irefn{org95}\And
C.~Bedda\Irefn{org111}\And
N.K.~Behera\Irefn{org49}\textsuperscript{,}\Irefn{org51}\And
I.~Belikov\Irefn{org56}\And
F.~Bellini\Irefn{org27}\And
H.~Bello Martinez\Irefn{org2}\And
R.~Bellwied\Irefn{org123}\And
R.~Belmont\Irefn{org136}\And
E.~Belmont-Moreno\Irefn{org65}\And
L.G.E.~Beltran\Irefn{org120}\And
V.~Belyaev\Irefn{org76}\And
G.~Bencedi\Irefn{org137}\And
S.~Beole\Irefn{org26}\And
I.~Berceanu\Irefn{org79}\And
A.~Bercuci\Irefn{org79}\And
Y.~Berdnikov\Irefn{org87}\And
D.~Berenyi\Irefn{org137}\And
R.A.~Bertens\Irefn{org58}\And
D.~Berzano\Irefn{org35}\And
L.~Betev\Irefn{org35}\And
A.~Bhasin\Irefn{org92}\And
I.R.~Bhat\Irefn{org92}\And
A.K.~Bhati\Irefn{org89}\And
B.~Bhattacharjee\Irefn{org44}\And
J.~Bhom\Irefn{org129}\textsuperscript{,}\Irefn{org118}\And
L.~Bianchi\Irefn{org123}\And
N.~Bianchi\Irefn{org73}\And
C.~Bianchin\Irefn{org136}\And
J.~Biel\v{c}\'{\i}k\Irefn{org39}\And
J.~Biel\v{c}\'{\i}kov\'{a}\Irefn{org85}\And
A.~Bilandzic\Irefn{org82}\textsuperscript{,}\Irefn{org36}\textsuperscript{,}\Irefn{org94}\And
G.~Biro\Irefn{org137}\And
R.~Biswas\Irefn{org4}\And
S.~Biswas\Irefn{org4}\textsuperscript{,}\Irefn{org80}\And
S.~Bjelogrlic\Irefn{org58}\And
J.T.~Blair\Irefn{org119}\And
D.~Blau\Irefn{org81}\And
C.~Blume\Irefn{org54}\And
F.~Bock\Irefn{org75}\textsuperscript{,}\Irefn{org95}\And
A.~Bogdanov\Irefn{org76}\And
H.~B{\o}ggild\Irefn{org82}\And
L.~Boldizs\'{a}r\Irefn{org137}\And
M.~Bombara\Irefn{org40}\And
M.~Bonora\Irefn{org35}\And
J.~Book\Irefn{org54}\And
H.~Borel\Irefn{org15}\And
A.~Borissov\Irefn{org97}\And
M.~Borri\Irefn{org84}\textsuperscript{,}\Irefn{org125}\And
F.~Boss\'u\Irefn{org66}\And
E.~Botta\Irefn{org26}\And
C.~Bourjau\Irefn{org82}\And
P.~Braun-Munzinger\Irefn{org98}\And
M.~Bregant\Irefn{org121}\And
T.~Breitner\Irefn{org53}\And
T.A.~Broker\Irefn{org54}\And
T.A.~Browning\Irefn{org96}\And
M.~Broz\Irefn{org39}\And
E.J.~Brucken\Irefn{org46}\And
E.~Bruna\Irefn{org111}\And
G.E.~Bruno\Irefn{org32}\And
D.~Budnikov\Irefn{org100}\And
H.~Buesching\Irefn{org54}\And
S.~Bufalino\Irefn{org35}\textsuperscript{,}\Irefn{org26}\And
P.~Buncic\Irefn{org35}\And
O.~Busch\Irefn{org129}\And
Z.~Buthelezi\Irefn{org66}\And
J.B.~Butt\Irefn{org16}\And
J.T.~Buxton\Irefn{org20}\And
J.~Cabala\Irefn{org116}\And
D.~Caffarri\Irefn{org35}\And
X.~Cai\Irefn{org7}\And
H.~Caines\Irefn{org138}\And
L.~Calero Diaz\Irefn{org73}\And
A.~Caliva\Irefn{org58}\And
E.~Calvo Villar\Irefn{org103}\And
P.~Camerini\Irefn{org25}\And
F.~Carena\Irefn{org35}\And
W.~Carena\Irefn{org35}\And
F.~Carnesecchi\Irefn{org27}\And
J.~Castillo Castellanos\Irefn{org15}\And
A.J.~Castro\Irefn{org126}\And
E.A.R.~Casula\Irefn{org24}\And
C.~Ceballos Sanchez\Irefn{org9}\And
J.~Cepila\Irefn{org39}\And
P.~Cerello\Irefn{org111}\And
J.~Cerkala\Irefn{org116}\And
B.~Chang\Irefn{org124}\And
S.~Chapeland\Irefn{org35}\And
M.~Chartier\Irefn{org125}\And
J.L.~Charvet\Irefn{org15}\And
S.~Chattopadhyay\Irefn{org134}\And
S.~Chattopadhyay\Irefn{org101}\And
A.~Chauvin\Irefn{org94}\textsuperscript{,}\Irefn{org36}\And
V.~Chelnokov\Irefn{org3}\And
M.~Cherney\Irefn{org88}\And
C.~Cheshkov\Irefn{org131}\And
B.~Cheynis\Irefn{org131}\And
V.~Chibante Barroso\Irefn{org35}\And
D.D.~Chinellato\Irefn{org122}\And
S.~Cho\Irefn{org51}\And
P.~Chochula\Irefn{org35}\And
K.~Choi\Irefn{org97}\And
M.~Chojnacki\Irefn{org82}\And
S.~Choudhury\Irefn{org134}\And
P.~Christakoglou\Irefn{org83}\And
C.H.~Christensen\Irefn{org82}\And
P.~Christiansen\Irefn{org33}\And
T.~Chujo\Irefn{org129}\And
S.U.~Chung\Irefn{org97}\And
C.~Cicalo\Irefn{org106}\And
L.~Cifarelli\Irefn{org12}\textsuperscript{,}\Irefn{org27}\And
F.~Cindolo\Irefn{org105}\And
J.~Cleymans\Irefn{org91}\And
F.~Colamaria\Irefn{org32}\And
D.~Colella\Irefn{org60}\textsuperscript{,}\Irefn{org35}\And
A.~Collu\Irefn{org75}\And
M.~Colocci\Irefn{org27}\And
G.~Conesa Balbastre\Irefn{org72}\And
Z.~Conesa del Valle\Irefn{org52}\And
M.E.~Connors\Aref{idp1793936}\textsuperscript{,}\Irefn{org138}\And
J.G.~Contreras\Irefn{org39}\And
T.M.~Cormier\Irefn{org86}\And
Y.~Corrales Morales\Irefn{org111}\textsuperscript{,}\Irefn{org26}\And
I.~Cort\'{e}s Maldonado\Irefn{org2}\And
P.~Cortese\Irefn{org31}\And
M.R.~Cosentino\Irefn{org121}\And
F.~Costa\Irefn{org35}\And
J.~Crkovska\Irefn{org52}\And
P.~Crochet\Irefn{org71}\And
R.~Cruz Albino\Irefn{org11}\And
E.~Cuautle\Irefn{org64}\And
L.~Cunqueiro\Irefn{org55}\textsuperscript{,}\Irefn{org35}\And
T.~Dahms\Irefn{org94}\textsuperscript{,}\Irefn{org36}\And
A.~Dainese\Irefn{org108}\And
M.C.~Danisch\Irefn{org95}\And
A.~Danu\Irefn{org63}\And
D.~Das\Irefn{org101}\And
I.~Das\Irefn{org101}\And
S.~Das\Irefn{org4}\And
A.~Dash\Irefn{org80}\And
S.~Dash\Irefn{org48}\And
S.~De\Irefn{org121}\And
A.~De Caro\Irefn{org12}\textsuperscript{,}\Irefn{org30}\And
G.~de Cataldo\Irefn{org104}\And
C.~de Conti\Irefn{org121}\And
J.~de Cuveland\Irefn{org42}\And
A.~De Falco\Irefn{org24}\And
D.~De Gruttola\Irefn{org12}\textsuperscript{,}\Irefn{org30}\And
N.~De Marco\Irefn{org111}\And
S.~De Pasquale\Irefn{org30}\And
R.D.~De Souza\Irefn{org122}\And
A.~Deisting\Irefn{org95}\textsuperscript{,}\Irefn{org98}\And
A.~Deloff\Irefn{org78}\And
E.~D\'{e}nes\Irefn{org137}\Aref{0}\And
C.~Deplano\Irefn{org83}\And
P.~Dhankher\Irefn{org48}\And
D.~Di Bari\Irefn{org32}\And
A.~Di Mauro\Irefn{org35}\And
P.~Di Nezza\Irefn{org73}\And
B.~Di Ruzza\Irefn{org108}\And
M.A.~Diaz Corchero\Irefn{org10}\And
T.~Dietel\Irefn{org91}\And
P.~Dillenseger\Irefn{org54}\And
R.~Divi\`{a}\Irefn{org35}\And
{\O}.~Djuvsland\Irefn{org18}\And
A.~Dobrin\Irefn{org83}\textsuperscript{,}\Irefn{org63}\And
D.~Domenicis Gimenez\Irefn{org121}\And
B.~D\"{o}nigus\Irefn{org54}\And
O.~Dordic\Irefn{org22}\And
T.~Drozhzhova\Irefn{org54}\And
A.K.~Dubey\Irefn{org134}\And
A.~Dubla\Irefn{org58}\And
L.~Ducroux\Irefn{org131}\And
P.~Dupieux\Irefn{org71}\And
R.J.~Ehlers\Irefn{org138}\And
D.~Elia\Irefn{org104}\And
E.~Endress\Irefn{org103}\And
H.~Engel\Irefn{org53}\And
E.~Epple\Irefn{org138}\And
B.~Erazmus\Irefn{org114}\And
I.~Erdemir\Irefn{org54}\And
F.~Erhardt\Irefn{org130}\And
B.~Espagnon\Irefn{org52}\And
M.~Estienne\Irefn{org114}\And
S.~Esumi\Irefn{org129}\And
J.~Eum\Irefn{org97}\And
D.~Evans\Irefn{org102}\And
S.~Evdokimov\Irefn{org112}\And
G.~Eyyubova\Irefn{org39}\And
L.~Fabbietti\Irefn{org94}\textsuperscript{,}\Irefn{org36}\And
D.~Fabris\Irefn{org108}\And
J.~Faivre\Irefn{org72}\And
A.~Fantoni\Irefn{org73}\And
M.~Fasel\Irefn{org75}\And
L.~Feldkamp\Irefn{org55}\And
A.~Feliciello\Irefn{org111}\And
G.~Feofilov\Irefn{org133}\And
J.~Ferencei\Irefn{org85}\And
A.~Fern\'{a}ndez T\'{e}llez\Irefn{org2}\And
E.G.~Ferreiro\Irefn{org17}\And
A.~Ferretti\Irefn{org26}\And
A.~Festanti\Irefn{org29}\And
V.J.G.~Feuillard\Irefn{org15}\textsuperscript{,}\Irefn{org71}\And
J.~Figiel\Irefn{org118}\And
M.A.S.~Figueredo\Irefn{org125}\textsuperscript{,}\Irefn{org121}\And
S.~Filchagin\Irefn{org100}\And
D.~Finogeev\Irefn{org57}\And
F.M.~Fionda\Irefn{org24}\And
E.M.~Fiore\Irefn{org32}\And
M.G.~Fleck\Irefn{org95}\And
M.~Floris\Irefn{org35}\And
S.~Foertsch\Irefn{org66}\And
P.~Foka\Irefn{org98}\And
S.~Fokin\Irefn{org81}\And
E.~Fragiacomo\Irefn{org110}\And
A.~Francescon\Irefn{org35}\And
A.~Francisco\Irefn{org114}\And
U.~Frankenfeld\Irefn{org98}\And
G.G.~Fronze\Irefn{org26}\And
U.~Fuchs\Irefn{org35}\And
C.~Furget\Irefn{org72}\And
A.~Furs\Irefn{org57}\And
M.~Fusco Girard\Irefn{org30}\And
J.J.~Gaardh{\o}je\Irefn{org82}\And
M.~Gagliardi\Irefn{org26}\And
A.M.~Gago\Irefn{org103}\And
K.~Gajdosova\Irefn{org82}\And
M.~Gallio\Irefn{org26}\And
C.D.~Galvan\Irefn{org120}\And
D.R.~Gangadharan\Irefn{org75}\And
P.~Ganoti\Irefn{org90}\And
C.~Gao\Irefn{org7}\And
C.~Garabatos\Irefn{org98}\And
E.~Garcia-Solis\Irefn{org13}\And
C.~Gargiulo\Irefn{org35}\And
P.~Gasik\Irefn{org94}\textsuperscript{,}\Irefn{org36}\And
E.F.~Gauger\Irefn{org119}\And
M.~Germain\Irefn{org114}\And
M.~Gheata\Irefn{org35}\textsuperscript{,}\Irefn{org63}\And
P.~Ghosh\Irefn{org134}\And
S.K.~Ghosh\Irefn{org4}\And
P.~Gianotti\Irefn{org73}\And
P.~Giubellino\Irefn{org111}\textsuperscript{,}\Irefn{org35}\And
P.~Giubilato\Irefn{org29}\And
E.~Gladysz-Dziadus\Irefn{org118}\And
P.~Gl\"{a}ssel\Irefn{org95}\And
D.M.~Gom\'{e}z Coral\Irefn{org65}\And
A.~Gomez Ramirez\Irefn{org53}\And
A.S.~Gonzalez\Irefn{org35}\And
V.~Gonzalez\Irefn{org10}\And
P.~Gonz\'{a}lez-Zamora\Irefn{org10}\And
S.~Gorbunov\Irefn{org42}\And
L.~G\"{o}rlich\Irefn{org118}\And
S.~Gotovac\Irefn{org117}\And
V.~Grabski\Irefn{org65}\And
O.A.~Grachov\Irefn{org138}\And
L.K.~Graczykowski\Irefn{org135}\And
K.L.~Graham\Irefn{org102}\And
A.~Grelli\Irefn{org58}\And
A.~Grigoras\Irefn{org35}\And
C.~Grigoras\Irefn{org35}\And
V.~Grigoriev\Irefn{org76}\And
A.~Grigoryan\Irefn{org1}\And
S.~Grigoryan\Irefn{org67}\And
B.~Grinyov\Irefn{org3}\And
N.~Grion\Irefn{org110}\And
J.M.~Gronefeld\Irefn{org98}\And
J.F.~Grosse-Oetringhaus\Irefn{org35}\And
R.~Grosso\Irefn{org98}\And
L.~Gruber\Irefn{org113}\And
F.~Guber\Irefn{org57}\And
R.~Guernane\Irefn{org72}\And
B.~Guerzoni\Irefn{org27}\And
K.~Gulbrandsen\Irefn{org82}\And
T.~Gunji\Irefn{org128}\And
A.~Gupta\Irefn{org92}\And
R.~Gupta\Irefn{org92}\And
R.~Haake\Irefn{org35}\And
{\O}.~Haaland\Irefn{org18}\And
C.~Hadjidakis\Irefn{org52}\And
M.~Haiduc\Irefn{org63}\And
H.~Hamagaki\Irefn{org128}\And
G.~Hamar\Irefn{org137}\And
J.C.~Hamon\Irefn{org56}\And
A.~Hansen\Irefn{org82}\And
J.W.~Harris\Irefn{org138}\And
A.~Harton\Irefn{org13}\And
D.~Hatzifotiadou\Irefn{org105}\And
S.~Hayashi\Irefn{org128}\And
S.T.~Heckel\Irefn{org54}\And
E.~Hellb\"{a}r\Irefn{org54}\And
H.~Helstrup\Irefn{org37}\And
A.~Herghelegiu\Irefn{org79}\And
G.~Herrera Corral\Irefn{org11}\And
B.A.~Hess\Irefn{org34}\And
K.F.~Hetland\Irefn{org37}\And
H.~Hillemanns\Irefn{org35}\And
B.~Hippolyte\Irefn{org56}\And
D.~Horak\Irefn{org39}\And
R.~Hosokawa\Irefn{org129}\And
P.~Hristov\Irefn{org35}\And
C.~Hughes\Irefn{org126}\And
T.J.~Humanic\Irefn{org20}\And
N.~Hussain\Irefn{org44}\And
T.~Hussain\Irefn{org19}\And
D.~Hutter\Irefn{org42}\And
D.S.~Hwang\Irefn{org21}\And
R.~Ilkaev\Irefn{org100}\And
M.~Inaba\Irefn{org129}\And
E.~Incani\Irefn{org24}\And
M.~Ippolitov\Irefn{org76}\textsuperscript{,}\Irefn{org81}\And
M.~Irfan\Irefn{org19}\And
M.~Ivanov\Irefn{org98}\And
V.~Ivanov\Irefn{org87}\And
V.~Izucheev\Irefn{org112}\And
B.~Jacak\Irefn{org75}\And
N.~Jacazio\Irefn{org27}\And
P.M.~Jacobs\Irefn{org75}\And
M.B.~Jadhav\Irefn{org48}\And
S.~Jadlovska\Irefn{org116}\And
J.~Jadlovsky\Irefn{org116}\textsuperscript{,}\Irefn{org60}\And
C.~Jahnke\Irefn{org121}\And
M.J.~Jakubowska\Irefn{org135}\And
H.J.~Jang\Irefn{org69}\And
M.A.~Janik\Irefn{org135}\And
P.H.S.Y.~Jayarathna\Irefn{org123}\And
C.~Jena\Irefn{org29}\And
S.~Jena\Irefn{org123}\And
R.T.~Jimenez Bustamante\Irefn{org98}\And
P.G.~Jones\Irefn{org102}\And
A.~Jusko\Irefn{org102}\And
P.~Kalinak\Irefn{org60}\And
A.~Kalweit\Irefn{org35}\And
J.H.~Kang\Irefn{org139}\And
V.~Kaplin\Irefn{org76}\And
S.~Kar\Irefn{org134}\And
A.~Karasu Uysal\Irefn{org70}\And
O.~Karavichev\Irefn{org57}\And
T.~Karavicheva\Irefn{org57}\And
L.~Karayan\Irefn{org98}\textsuperscript{,}\Irefn{org95}\And
E.~Karpechev\Irefn{org57}\And
U.~Kebschull\Irefn{org53}\And
R.~Keidel\Irefn{org140}\And
D.L.D.~Keijdener\Irefn{org58}\And
M.~Keil\Irefn{org35}\And
M. Mohisin~Khan\Aref{idp3203424}\textsuperscript{,}\Irefn{org19}\And
P.~Khan\Irefn{org101}\And
S.A.~Khan\Irefn{org134}\And
A.~Khanzadeev\Irefn{org87}\And
Y.~Kharlov\Irefn{org112}\And
B.~Kileng\Irefn{org37}\And
D.W.~Kim\Irefn{org43}\And
D.J.~Kim\Irefn{org124}\And
D.~Kim\Irefn{org139}\And
H.~Kim\Irefn{org139}\And
J.S.~Kim\Irefn{org43}\And
J.~Kim\Irefn{org95}\And
M.~Kim\Irefn{org139}\And
S.~Kim\Irefn{org21}\And
T.~Kim\Irefn{org139}\And
S.~Kirsch\Irefn{org42}\And
I.~Kisel\Irefn{org42}\And
S.~Kiselev\Irefn{org59}\And
A.~Kisiel\Irefn{org135}\And
G.~Kiss\Irefn{org137}\And
J.L.~Klay\Irefn{org6}\And
C.~Klein\Irefn{org54}\And
J.~Klein\Irefn{org35}\And
C.~Klein-B\"{o}sing\Irefn{org55}\And
S.~Klewin\Irefn{org95}\And
A.~Kluge\Irefn{org35}\And
M.L.~Knichel\Irefn{org95}\And
A.G.~Knospe\Irefn{org119}\textsuperscript{,}\Irefn{org123}\And
C.~Kobdaj\Irefn{org115}\And
M.~Kofarago\Irefn{org35}\And
T.~Kollegger\Irefn{org98}\And
A.~Kolojvari\Irefn{org133}\And
V.~Kondratiev\Irefn{org133}\And
N.~Kondratyeva\Irefn{org76}\And
E.~Kondratyuk\Irefn{org112}\And
A.~Konevskikh\Irefn{org57}\And
M.~Kopcik\Irefn{org116}\And
M.~Kour\Irefn{org92}\And
C.~Kouzinopoulos\Irefn{org35}\And
O.~Kovalenko\Irefn{org78}\And
V.~Kovalenko\Irefn{org133}\And
M.~Kowalski\Irefn{org118}\And
G.~Koyithatta Meethaleveedu\Irefn{org48}\And
I.~Kr\'{a}lik\Irefn{org60}\And
A.~Krav\v{c}\'{a}kov\'{a}\Irefn{org40}\And
M.~Krivda\Irefn{org60}\textsuperscript{,}\Irefn{org102}\And
F.~Krizek\Irefn{org85}\And
E.~Kryshen\Irefn{org87}\textsuperscript{,}\Irefn{org35}\And
M.~Krzewicki\Irefn{org42}\And
A.M.~Kubera\Irefn{org20}\And
V.~Ku\v{c}era\Irefn{org85}\And
C.~Kuhn\Irefn{org56}\And
P.G.~Kuijer\Irefn{org83}\And
A.~Kumar\Irefn{org92}\And
J.~Kumar\Irefn{org48}\And
L.~Kumar\Irefn{org89}\And
S.~Kumar\Irefn{org48}\And
P.~Kurashvili\Irefn{org78}\And
A.~Kurepin\Irefn{org57}\And
A.B.~Kurepin\Irefn{org57}\And
A.~Kuryakin\Irefn{org100}\And
M.J.~Kweon\Irefn{org51}\And
Y.~Kwon\Irefn{org139}\And
S.L.~La Pointe\Irefn{org111}\And
P.~La Rocca\Irefn{org28}\And
P.~Ladron de Guevara\Irefn{org11}\And
C.~Lagana Fernandes\Irefn{org121}\And
I.~Lakomov\Irefn{org35}\And
R.~Langoy\Irefn{org41}\And
K.~Lapidus\Irefn{org138}\textsuperscript{,}\Irefn{org36}\And
C.~Lara\Irefn{org53}\And
A.~Lardeux\Irefn{org15}\And
A.~Lattuca\Irefn{org26}\And
E.~Laudi\Irefn{org35}\And
R.~Lea\Irefn{org25}\And
L.~Leardini\Irefn{org95}\And
S.~Lee\Irefn{org139}\And
F.~Lehas\Irefn{org83}\And
S.~Lehner\Irefn{org113}\And
R.C.~Lemmon\Irefn{org84}\And
V.~Lenti\Irefn{org104}\And
E.~Leogrande\Irefn{org58}\And
I.~Le\'{o}n Monz\'{o}n\Irefn{org120}\And
H.~Le\'{o}n Vargas\Irefn{org65}\And
M.~Leoncino\Irefn{org26}\And
P.~L\'{e}vai\Irefn{org137}\And
S.~Li\Irefn{org71}\textsuperscript{,}\Irefn{org7}\And
X.~Li\Irefn{org14}\And
J.~Lien\Irefn{org41}\And
R.~Lietava\Irefn{org102}\And
S.~Lindal\Irefn{org22}\And
V.~Lindenstruth\Irefn{org42}\And
C.~Lippmann\Irefn{org98}\And
M.A.~Lisa\Irefn{org20}\And
H.M.~Ljunggren\Irefn{org33}\And
D.F.~Lodato\Irefn{org58}\And
P.I.~Loenne\Irefn{org18}\And
V.~Loginov\Irefn{org76}\And
C.~Loizides\Irefn{org75}\And
X.~Lopez\Irefn{org71}\And
E.~L\'{o}pez Torres\Irefn{org9}\And
A.~Lowe\Irefn{org137}\And
P.~Luettig\Irefn{org54}\And
M.~Lunardon\Irefn{org29}\And
G.~Luparello\Irefn{org25}\And
M.~Lupi\Irefn{org35}\And
T.H.~Lutz\Irefn{org138}\And
A.~Maevskaya\Irefn{org57}\And
M.~Mager\Irefn{org35}\And
S.~Mahajan\Irefn{org92}\And
S.M.~Mahmood\Irefn{org22}\And
A.~Maire\Irefn{org56}\And
R.D.~Majka\Irefn{org138}\And
M.~Malaev\Irefn{org87}\And
I.~Maldonado Cervantes\Irefn{org64}\And
L.~Malinina\Aref{idp3917824}\textsuperscript{,}\Irefn{org67}\And
D.~Mal'Kevich\Irefn{org59}\And
P.~Malzacher\Irefn{org98}\And
A.~Mamonov\Irefn{org100}\And
V.~Manko\Irefn{org81}\And
F.~Manso\Irefn{org71}\And
V.~Manzari\Irefn{org35}\textsuperscript{,}\Irefn{org104}\And
Y.~Mao\Irefn{org7}\And
M.~Marchisone\Irefn{org127}\textsuperscript{,}\Irefn{org66}\textsuperscript{,}\Irefn{org26}\And
J.~Mare\v{s}\Irefn{org61}\And
G.V.~Margagliotti\Irefn{org25}\And
A.~Margotti\Irefn{org105}\And
J.~Margutti\Irefn{org58}\And
A.~Mar\'{\i}n\Irefn{org98}\And
C.~Markert\Irefn{org119}\And
M.~Marquard\Irefn{org54}\And
N.A.~Martin\Irefn{org98}\And
J.~Martin Blanco\Irefn{org114}\And
P.~Martinengo\Irefn{org35}\And
M.I.~Mart\'{\i}nez\Irefn{org2}\And
G.~Mart\'{\i}nez Garc\'{\i}a\Irefn{org114}\And
M.~Martinez Pedreira\Irefn{org35}\And
A.~Mas\Irefn{org121}\And
S.~Masciocchi\Irefn{org98}\And
M.~Masera\Irefn{org26}\And
A.~Masoni\Irefn{org106}\And
A.~Mastroserio\Irefn{org32}\And
A.~Matyja\Irefn{org118}\And
C.~Mayer\Irefn{org118}\And
J.~Mazer\Irefn{org126}\And
M.A.~Mazzoni\Irefn{org109}\And
D.~Mcdonald\Irefn{org123}\And
F.~Meddi\Irefn{org23}\And
Y.~Melikyan\Irefn{org76}\And
A.~Menchaca-Rocha\Irefn{org65}\And
E.~Meninno\Irefn{org30}\And
J.~Mercado P\'erez\Irefn{org95}\And
M.~Meres\Irefn{org38}\And
S.~Mhlanga\Irefn{org91}\And
Y.~Miake\Irefn{org129}\And
M.M.~Mieskolainen\Irefn{org46}\And
K.~Mikhaylov\Irefn{org67}\textsuperscript{,}\Irefn{org59}\And
L.~Milano\Irefn{org75}\textsuperscript{,}\Irefn{org35}\And
J.~Milosevic\Irefn{org22}\And
A.~Mischke\Irefn{org58}\And
A.N.~Mishra\Irefn{org49}\And
D.~Mi\'{s}kowiec\Irefn{org98}\And
J.~Mitra\Irefn{org134}\And
C.M.~Mitu\Irefn{org63}\And
N.~Mohammadi\Irefn{org58}\And
B.~Mohanty\Irefn{org80}\And
L.~Molnar\Irefn{org56}\And
L.~Monta\~{n}o Zetina\Irefn{org11}\And
E.~Montes\Irefn{org10}\And
D.A.~Moreira De Godoy\Irefn{org55}\And
L.A.P.~Moreno\Irefn{org2}\And
S.~Moretto\Irefn{org29}\And
A.~Morreale\Irefn{org114}\And
A.~Morsch\Irefn{org35}\And
V.~Muccifora\Irefn{org73}\And
E.~Mudnic\Irefn{org117}\And
D.~M{\"u}hlheim\Irefn{org55}\And
S.~Muhuri\Irefn{org134}\And
M.~Mukherjee\Irefn{org134}\And
J.D.~Mulligan\Irefn{org138}\And
M.G.~Munhoz\Irefn{org121}\And
K.~M\"{u}nning\Irefn{org45}\And
R.H.~Munzer\Irefn{org36}\textsuperscript{,}\Irefn{org54}\textsuperscript{,}\Irefn{org94}\And
H.~Murakami\Irefn{org128}\And
S.~Murray\Irefn{org66}\And
L.~Musa\Irefn{org35}\And
J.~Musinsky\Irefn{org60}\And
B.~Naik\Irefn{org48}\And
R.~Nair\Irefn{org78}\And
B.K.~Nandi\Irefn{org48}\And
R.~Nania\Irefn{org105}\And
E.~Nappi\Irefn{org104}\And
M.U.~Naru\Irefn{org16}\And
H.~Natal da Luz\Irefn{org121}\And
C.~Nattrass\Irefn{org126}\And
S.R.~Navarro\Irefn{org2}\And
K.~Nayak\Irefn{org80}\And
R.~Nayak\Irefn{org48}\And
T.K.~Nayak\Irefn{org134}\And
S.~Nazarenko\Irefn{org100}\And
A.~Nedosekin\Irefn{org59}\And
R.A.~Negrao De Oliveira\Irefn{org35}\And
L.~Nellen\Irefn{org64}\And
F.~Ng\Irefn{org123}\And
M.~Nicassio\Irefn{org98}\And
M.~Niculescu\Irefn{org63}\And
J.~Niedziela\Irefn{org35}\And
B.S.~Nielsen\Irefn{org82}\And
S.~Nikolaev\Irefn{org81}\And
S.~Nikulin\Irefn{org81}\And
V.~Nikulin\Irefn{org87}\And
F.~Noferini\Irefn{org105}\textsuperscript{,}\Irefn{org12}\And
P.~Nomokonov\Irefn{org67}\And
G.~Nooren\Irefn{org58}\And
J.C.C.~Noris\Irefn{org2}\And
J.~Norman\Irefn{org125}\And
A.~Nyanin\Irefn{org81}\And
J.~Nystrand\Irefn{org18}\And
H.~Oeschler\Irefn{org95}\And
S.~Oh\Irefn{org138}\And
S.K.~Oh\Irefn{org68}\And
A.~Ohlson\Irefn{org35}\And
A.~Okatan\Irefn{org70}\And
T.~Okubo\Irefn{org47}\And
J.~Oleniacz\Irefn{org135}\And
A.C.~Oliveira Da Silva\Irefn{org121}\And
M.H.~Oliver\Irefn{org138}\And
J.~Onderwaater\Irefn{org98}\And
C.~Oppedisano\Irefn{org111}\And
R.~Orava\Irefn{org46}\And
M.~Oravec\Irefn{org116}\And
A.~Ortiz Velasquez\Irefn{org64}\And
A.~Oskarsson\Irefn{org33}\And
J.~Otwinowski\Irefn{org118}\And
K.~Oyama\Irefn{org95}\textsuperscript{,}\Irefn{org77}\And
M.~Ozdemir\Irefn{org54}\And
Y.~Pachmayer\Irefn{org95}\And
D.~Pagano\Irefn{org132}\And
P.~Pagano\Irefn{org30}\And
G.~Pai\'{c}\Irefn{org64}\And
S.K.~Pal\Irefn{org134}\And
J.~Pan\Irefn{org136}\And
A.K.~Pandey\Irefn{org48}\And
V.~Papikyan\Irefn{org1}\And
G.S.~Pappalardo\Irefn{org107}\And
P.~Pareek\Irefn{org49}\And
W.J.~Park\Irefn{org98}\And
S.~Parmar\Irefn{org89}\And
A.~Passfeld\Irefn{org55}\And
V.~Paticchio\Irefn{org104}\And
R.N.~Patra\Irefn{org134}\And
B.~Paul\Irefn{org111}\textsuperscript{,}\Irefn{org101}\And
H.~Pei\Irefn{org7}\And
T.~Peitzmann\Irefn{org58}\And
H.~Pereira Da Costa\Irefn{org15}\And
D.~Peresunko\Irefn{org81}\textsuperscript{,}\Irefn{org76}\And
E.~Perez Lezama\Irefn{org54}\And
V.~Peskov\Irefn{org54}\And
Y.~Pestov\Irefn{org5}\And
V.~Petr\'{a}\v{c}ek\Irefn{org39}\And
V.~Petrov\Irefn{org112}\And
M.~Petrovici\Irefn{org79}\And
C.~Petta\Irefn{org28}\And
S.~Piano\Irefn{org110}\And
M.~Pikna\Irefn{org38}\And
P.~Pillot\Irefn{org114}\And
L.O.D.L.~Pimentel\Irefn{org82}\And
O.~Pinazza\Irefn{org105}\textsuperscript{,}\Irefn{org35}\And
L.~Pinsky\Irefn{org123}\And
D.B.~Piyarathna\Irefn{org123}\And
M.~P\l osko\'{n}\Irefn{org75}\And
M.~Planinic\Irefn{org130}\And
J.~Pluta\Irefn{org135}\And
S.~Pochybova\Irefn{org137}\And
P.L.M.~Podesta-Lerma\Irefn{org120}\And
M.G.~Poghosyan\Irefn{org86}\textsuperscript{,}\Irefn{org88}\And
B.~Polichtchouk\Irefn{org112}\And
N.~Poljak\Irefn{org130}\And
W.~Poonsawat\Irefn{org115}\And
A.~Pop\Irefn{org79}\And
H.~Poppenborg\Irefn{org55}\And
S.~Porteboeuf-Houssais\Irefn{org71}\And
J.~Porter\Irefn{org75}\And
J.~Pospisil\Irefn{org85}\And
S.K.~Prasad\Irefn{org4}\And
R.~Preghenella\Irefn{org105}\textsuperscript{,}\Irefn{org35}\And
F.~Prino\Irefn{org111}\And
C.A.~Pruneau\Irefn{org136}\And
I.~Pshenichnov\Irefn{org57}\And
M.~Puccio\Irefn{org26}\And
G.~Puddu\Irefn{org24}\And
P.~Pujahari\Irefn{org136}\And
V.~Punin\Irefn{org100}\And
J.~Putschke\Irefn{org136}\And
H.~Qvigstad\Irefn{org22}\And
A.~Rachevski\Irefn{org110}\And
S.~Raha\Irefn{org4}\And
S.~Rajput\Irefn{org92}\And
J.~Rak\Irefn{org124}\And
A.~Rakotozafindrabe\Irefn{org15}\And
L.~Ramello\Irefn{org31}\And
F.~Rami\Irefn{org56}\And
R.~Raniwala\Irefn{org93}\And
S.~Raniwala\Irefn{org93}\And
S.S.~R\"{a}s\"{a}nen\Irefn{org46}\And
B.T.~Rascanu\Irefn{org54}\And
D.~Rathee\Irefn{org89}\And
K.F.~Read\Irefn{org126}\textsuperscript{,}\Irefn{org86}\And
K.~Redlich\Irefn{org78}\And
R.J.~Reed\Irefn{org136}\And
A.~Rehman\Irefn{org18}\And
P.~Reichelt\Irefn{org54}\And
F.~Reidt\Irefn{org35}\textsuperscript{,}\Irefn{org95}\And
X.~Ren\Irefn{org7}\And
R.~Renfordt\Irefn{org54}\And
A.R.~Reolon\Irefn{org73}\And
A.~Reshetin\Irefn{org57}\And
K.~Reygers\Irefn{org95}\And
V.~Riabov\Irefn{org87}\And
R.A.~Ricci\Irefn{org74}\And
T.~Richert\Irefn{org33}\And
M.~Richter\Irefn{org22}\And
P.~Riedler\Irefn{org35}\And
W.~Riegler\Irefn{org35}\And
F.~Riggi\Irefn{org28}\And
C.~Ristea\Irefn{org63}\And
E.~Rocco\Irefn{org58}\And
M.~Rodr\'{i}guez Cahuantzi\Irefn{org2}\And
A.~Rodriguez Manso\Irefn{org83}\And
K.~R{\o}ed\Irefn{org22}\And
E.~Rogochaya\Irefn{org67}\And
D.~Rohr\Irefn{org42}\And
D.~R\"ohrich\Irefn{org18}\And
F.~Ronchetti\Irefn{org73}\textsuperscript{,}\Irefn{org35}\And
L.~Ronflette\Irefn{org114}\And
P.~Rosnet\Irefn{org71}\And
A.~Rossi\Irefn{org29}\And
F.~Roukoutakis\Irefn{org90}\And
A.~Roy\Irefn{org49}\And
C.~Roy\Irefn{org56}\And
P.~Roy\Irefn{org101}\And
A.J.~Rubio Montero\Irefn{org10}\And
R.~Rui\Irefn{org25}\And
R.~Russo\Irefn{org26}\And
E.~Ryabinkin\Irefn{org81}\And
Y.~Ryabov\Irefn{org87}\And
A.~Rybicki\Irefn{org118}\And
S.~Saarinen\Irefn{org46}\And
S.~Sadhu\Irefn{org134}\And
S.~Sadovsky\Irefn{org112}\And
K.~\v{S}afa\v{r}\'{\i}k\Irefn{org35}\And
B.~Sahlmuller\Irefn{org54}\And
P.~Sahoo\Irefn{org49}\And
R.~Sahoo\Irefn{org49}\And
S.~Sahoo\Irefn{org62}\And
P.K.~Sahu\Irefn{org62}\And
J.~Saini\Irefn{org134}\And
S.~Sakai\Irefn{org73}\And
M.A.~Saleh\Irefn{org136}\And
J.~Salzwedel\Irefn{org20}\And
S.~Sambyal\Irefn{org92}\And
V.~Samsonov\Irefn{org76}\textsuperscript{,}\Irefn{org87}\And
L.~\v{S}\'{a}ndor\Irefn{org60}\And
A.~Sandoval\Irefn{org65}\And
M.~Sano\Irefn{org129}\And
D.~Sarkar\Irefn{org134}\And
N.~Sarkar\Irefn{org134}\And
P.~Sarma\Irefn{org44}\And
E.~Scapparone\Irefn{org105}\And
F.~Scarlassara\Irefn{org29}\And
C.~Schiaua\Irefn{org79}\And
R.~Schicker\Irefn{org95}\And
C.~Schmidt\Irefn{org98}\And
H.R.~Schmidt\Irefn{org34}\And
M.~Schmidt\Irefn{org34}\And
S.~Schuchmann\Irefn{org54}\textsuperscript{,}\Irefn{org95}\And
J.~Schukraft\Irefn{org35}\And
Y.~Schutz\Irefn{org35}\textsuperscript{,}\Irefn{org114}\And
K.~Schwarz\Irefn{org98}\And
K.~Schweda\Irefn{org98}\And
G.~Scioli\Irefn{org27}\And
E.~Scomparin\Irefn{org111}\And
R.~Scott\Irefn{org126}\And
M.~\v{S}ef\v{c}\'ik\Irefn{org40}\And
J.E.~Seger\Irefn{org88}\And
Y.~Sekiguchi\Irefn{org128}\And
D.~Sekihata\Irefn{org47}\And
I.~Selyuzhenkov\Irefn{org98}\And
K.~Senosi\Irefn{org66}\And
S.~Senyukov\Irefn{org3}\textsuperscript{,}\Irefn{org35}\And
E.~Serradilla\Irefn{org10}\textsuperscript{,}\Irefn{org65}\And
A.~Sevcenco\Irefn{org63}\And
A.~Shabanov\Irefn{org57}\And
A.~Shabetai\Irefn{org114}\And
O.~Shadura\Irefn{org3}\And
R.~Shahoyan\Irefn{org35}\And
M.I.~Shahzad\Irefn{org16}\And
A.~Shangaraev\Irefn{org112}\And
A.~Sharma\Irefn{org92}\And
M.~Sharma\Irefn{org92}\And
M.~Sharma\Irefn{org92}\And
N.~Sharma\Irefn{org126}\And
A.I.~Sheikh\Irefn{org134}\And
K.~Shigaki\Irefn{org47}\And
Q.~Shou\Irefn{org7}\And
K.~Shtejer\Irefn{org9}\textsuperscript{,}\Irefn{org26}\And
Y.~Sibiriak\Irefn{org81}\And
S.~Siddhanta\Irefn{org106}\And
K.M.~Sielewicz\Irefn{org35}\And
T.~Siemiarczuk\Irefn{org78}\And
D.~Silvermyr\Irefn{org33}\And
C.~Silvestre\Irefn{org72}\And
G.~Simatovic\Irefn{org130}\And
G.~Simonetti\Irefn{org35}\And
R.~Singaraju\Irefn{org134}\And
R.~Singh\Irefn{org80}\And
V.~Singhal\Irefn{org134}\And
T.~Sinha\Irefn{org101}\And
B.~Sitar\Irefn{org38}\And
M.~Sitta\Irefn{org31}\And
T.B.~Skaali\Irefn{org22}\And
M.~Slupecki\Irefn{org124}\And
N.~Smirnov\Irefn{org138}\And
R.J.M.~Snellings\Irefn{org58}\And
T.W.~Snellman\Irefn{org124}\And
J.~Song\Irefn{org97}\And
M.~Song\Irefn{org139}\And
Z.~Song\Irefn{org7}\And
F.~Soramel\Irefn{org29}\And
S.~Sorensen\Irefn{org126}\And
F.~Sozzi\Irefn{org98}\And
E.~Spiriti\Irefn{org73}\And
I.~Sputowska\Irefn{org118}\And
M.~Spyropoulou-Stassinaki\Irefn{org90}\And
J.~Stachel\Irefn{org95}\And
I.~Stan\Irefn{org63}\And
P.~Stankus\Irefn{org86}\And
E.~Stenlund\Irefn{org33}\And
G.~Steyn\Irefn{org66}\And
J.H.~Stiller\Irefn{org95}\And
D.~Stocco\Irefn{org114}\And
P.~Strmen\Irefn{org38}\And
A.A.P.~Suaide\Irefn{org121}\And
T.~Sugitate\Irefn{org47}\And
C.~Suire\Irefn{org52}\And
M.~Suleymanov\Irefn{org16}\And
M.~Suljic\Irefn{org25}\Aref{0}\And
R.~Sultanov\Irefn{org59}\And
M.~\v{S}umbera\Irefn{org85}\And
S.~Sumowidagdo\Irefn{org50}\And
A.~Szabo\Irefn{org38}\And
I.~Szarka\Irefn{org38}\And
A.~Szczepankiewicz\Irefn{org135}\And
M.~Szymanski\Irefn{org135}\And
U.~Tabassam\Irefn{org16}\And
J.~Takahashi\Irefn{org122}\And
G.J.~Tambave\Irefn{org18}\And
N.~Tanaka\Irefn{org129}\And
M.~Tarhini\Irefn{org52}\And
M.~Tariq\Irefn{org19}\And
M.G.~Tarzila\Irefn{org79}\And
A.~Tauro\Irefn{org35}\And
G.~Tejeda Mu\~{n}oz\Irefn{org2}\And
A.~Telesca\Irefn{org35}\And
K.~Terasaki\Irefn{org128}\And
C.~Terrevoli\Irefn{org29}\And
B.~Teyssier\Irefn{org131}\And
J.~Th\"{a}der\Irefn{org75}\And
D.~Thakur\Irefn{org49}\And
D.~Thomas\Irefn{org119}\And
R.~Tieulent\Irefn{org131}\And
A.~Tikhonov\Irefn{org57}\And
A.R.~Timmins\Irefn{org123}\And
A.~Toia\Irefn{org54}\And
S.~Trogolo\Irefn{org26}\And
G.~Trombetta\Irefn{org32}\And
V.~Trubnikov\Irefn{org3}\And
W.H.~Trzaska\Irefn{org124}\And
T.~Tsuji\Irefn{org128}\And
A.~Tumkin\Irefn{org100}\And
R.~Turrisi\Irefn{org108}\And
T.S.~Tveter\Irefn{org22}\And
K.~Ullaland\Irefn{org18}\And
A.~Uras\Irefn{org131}\And
G.L.~Usai\Irefn{org24}\And
A.~Utrobicic\Irefn{org130}\And
M.~Vala\Irefn{org60}\And
L.~Valencia Palomo\Irefn{org71}\And
S.~Vallero\Irefn{org26}\And
J.~Van Der Maarel\Irefn{org58}\And
J.W.~Van Hoorne\Irefn{org35}\textsuperscript{,}\Irefn{org113}\And
M.~van Leeuwen\Irefn{org58}\And
T.~Vanat\Irefn{org85}\And
P.~Vande Vyvre\Irefn{org35}\And
D.~Varga\Irefn{org137}\And
A.~Vargas\Irefn{org2}\And
M.~Vargyas\Irefn{org124}\And
R.~Varma\Irefn{org48}\And
M.~Vasileiou\Irefn{org90}\And
A.~Vasiliev\Irefn{org81}\And
A.~Vauthier\Irefn{org72}\And
O.~V\'azquez Doce\Irefn{org94}\textsuperscript{,}\Irefn{org36}\And
V.~Vechernin\Irefn{org133}\And
A.M.~Veen\Irefn{org58}\And
M.~Veldhoen\Irefn{org58}\And
A.~Velure\Irefn{org18}\And
E.~Vercellin\Irefn{org26}\And
S.~Vergara Lim\'on\Irefn{org2}\And
R.~Vernet\Irefn{org8}\And
M.~Verweij\Irefn{org136}\And
L.~Vickovic\Irefn{org117}\And
J.~Viinikainen\Irefn{org124}\And
Z.~Vilakazi\Irefn{org127}\And
O.~Villalobos Baillie\Irefn{org102}\And
A.~Villatoro Tello\Irefn{org2}\And
A.~Vinogradov\Irefn{org81}\And
L.~Vinogradov\Irefn{org133}\And
T.~Virgili\Irefn{org30}\And
V.~Vislavicius\Irefn{org33}\And
Y.P.~Viyogi\Irefn{org134}\And
A.~Vodopyanov\Irefn{org67}\And
M.A.~V\"{o}lkl\Irefn{org95}\And
K.~Voloshin\Irefn{org59}\And
S.A.~Voloshin\Irefn{org136}\And
G.~Volpe\Irefn{org32}\textsuperscript{,}\Irefn{org137}\And
B.~von Haller\Irefn{org35}\And
I.~Vorobyev\Irefn{org94}\textsuperscript{,}\Irefn{org36}\And
D.~Vranic\Irefn{org98}\textsuperscript{,}\Irefn{org35}\And
J.~Vrl\'{a}kov\'{a}\Irefn{org40}\And
B.~Vulpescu\Irefn{org71}\And
B.~Wagner\Irefn{org18}\And
J.~Wagner\Irefn{org98}\And
H.~Wang\Irefn{org58}\And
M.~Wang\Irefn{org7}\And
D.~Watanabe\Irefn{org129}\And
Y.~Watanabe\Irefn{org128}\And
M.~Weber\Irefn{org35}\textsuperscript{,}\Irefn{org113}\And
S.G.~Weber\Irefn{org98}\And
D.F.~Weiser\Irefn{org95}\And
J.P.~Wessels\Irefn{org55}\And
U.~Westerhoff\Irefn{org55}\And
A.M.~Whitehead\Irefn{org91}\And
J.~Wiechula\Irefn{org34}\And
J.~Wikne\Irefn{org22}\And
G.~Wilk\Irefn{org78}\And
J.~Wilkinson\Irefn{org95}\And
G.A.~Willems\Irefn{org55}\And
M.C.S.~Williams\Irefn{org105}\And
B.~Windelband\Irefn{org95}\And
M.~Winn\Irefn{org95}\And
P.~Yang\Irefn{org7}\And
S.~Yano\Irefn{org47}\And
Z.~Yasin\Irefn{org16}\And
Z.~Yin\Irefn{org7}\And
H.~Yokoyama\Irefn{org129}\And
I.-K.~Yoo\Irefn{org97}\And
J.H.~Yoon\Irefn{org51}\And
V.~Yurchenko\Irefn{org3}\And
A.~Zaborowska\Irefn{org135}\And
V.~Zaccolo\Irefn{org82}\And
A.~Zaman\Irefn{org16}\And
C.~Zampolli\Irefn{org105}\textsuperscript{,}\Irefn{org35}\And
H.J.C.~Zanoli\Irefn{org121}\And
S.~Zaporozhets\Irefn{org67}\And
N.~Zardoshti\Irefn{org102}\And
A.~Zarochentsev\Irefn{org133}\And
P.~Z\'{a}vada\Irefn{org61}\And
N.~Zaviyalov\Irefn{org100}\And
H.~Zbroszczyk\Irefn{org135}\And
I.S.~Zgura\Irefn{org63}\And
M.~Zhalov\Irefn{org87}\And
H.~Zhang\Irefn{org18}\textsuperscript{,}\Irefn{org7}\And
X.~Zhang\Irefn{org75}\textsuperscript{,}\Irefn{org7}\And
Y.~Zhang\Irefn{org7}\And
C.~Zhang\Irefn{org58}\And
Z.~Zhang\Irefn{org7}\And
C.~Zhao\Irefn{org22}\And
N.~Zhigareva\Irefn{org59}\And
D.~Zhou\Irefn{org7}\And
Y.~Zhou\Irefn{org82}\And
Z.~Zhou\Irefn{org18}\And
H.~Zhu\Irefn{org7}\textsuperscript{,}\Irefn{org18}\And
J.~Zhu\Irefn{org7}\textsuperscript{,}\Irefn{org114}\And
A.~Zichichi\Irefn{org27}\textsuperscript{,}\Irefn{org12}\And
A.~Zimmermann\Irefn{org95}\And
M.B.~Zimmermann\Irefn{org55}\textsuperscript{,}\Irefn{org35}\And
G.~Zinovjev\Irefn{org3}\And
M.~Zyzak\Irefn{org42}
\renewcommand\labelenumi{\textsuperscript{\theenumi}~}
\section*{Affiliation notes}
\renewcommand\theenumi{\roman{enumi}}
\begin{Authlist}
\item \Adef{0}Deceased
\item \Adef{idp1793936}{Also at: Georgia State University, Atlanta, Georgia, United States}
\item \Adef{idp3203424}{Also at: Also at Department of Applied Physics, Aligarh Muslim University, Aligarh, India}
\item \Adef{idp3917824}{Also at: M.V. Lomonosov Moscow State University, D.V. Skobeltsyn Institute of Nuclear, Physics, Moscow, Russia}
\end{Authlist}
\section*{Collaboration Institutes}
\renewcommand\theenumi{\arabic{enumi}~}
\begin{Authlist}
\item \Idef{org1}A.I. Alikhanyan National Science Laboratory (Yerevan Physics Institute) Foundation, Yerevan, Armenia
\item \Idef{org2}Benem\'{e}rita Universidad Aut\'{o}noma de Puebla, Puebla, Mexico
\item \Idef{org3}Bogolyubov Institute for Theoretical Physics, Kiev, Ukraine
\item \Idef{org4}Bose Institute, Department of Physics and Centre for Astroparticle Physics and Space Science (CAPSS), Kolkata, India
\item \Idef{org5}Budker Institute for Nuclear Physics, Novosibirsk, Russia
\item \Idef{org6}California Polytechnic State University, San Luis Obispo, California, United States
\item \Idef{org7}Central China Normal University, Wuhan, China
\item \Idef{org8}Centre de Calcul de l'IN2P3, Villeurbanne, France
\item \Idef{org9}Centro de Aplicaciones Tecnol\'{o}gicas y Desarrollo Nuclear (CEADEN), Havana, Cuba
\item \Idef{org10}Centro de Investigaciones Energ\'{e}ticas Medioambientales y Tecnol\'{o}gicas (CIEMAT), Madrid, Spain
\item \Idef{org11}Centro de Investigaci\'{o}n y de Estudios Avanzados (CINVESTAV), Mexico City and M\'{e}rida, Mexico
\item \Idef{org12}Centro Fermi - Museo Storico della Fisica e Centro Studi e Ricerche ``Enrico Fermi'', Rome, Italy
\item \Idef{org13}Chicago State University, Chicago, Illinois, USA
\item \Idef{org14}China Institute of Atomic Energy, Beijing, China
\item \Idef{org15}Commissariat \`{a} l'Energie Atomique, IRFU, Saclay, France
\item \Idef{org16}COMSATS Institute of Information Technology (CIIT), Islamabad, Pakistan
\item \Idef{org17}Departamento de F\'{\i}sica de Part\'{\i}culas and IGFAE, Universidad de Santiago de Compostela, Santiago de Compostela, Spain
\item \Idef{org18}Department of Physics and Technology, University of Bergen, Bergen, Norway
\item \Idef{org19}Department of Physics, Aligarh Muslim University, Aligarh, India
\item \Idef{org20}Department of Physics, Ohio State University, Columbus, Ohio, United States
\item \Idef{org21}Department of Physics, Sejong University, Seoul, South Korea
\item \Idef{org22}Department of Physics, University of Oslo, Oslo, Norway
\item \Idef{org23}Dipartimento di Fisica dell'Universit\`{a} 'La Sapienza' and Sezione INFN Rome, Italy
\item \Idef{org24}Dipartimento di Fisica dell'Universit\`{a} and Sezione INFN, Cagliari, Italy
\item \Idef{org25}Dipartimento di Fisica dell'Universit\`{a} and Sezione INFN, Trieste, Italy
\item \Idef{org26}Dipartimento di Fisica dell'Universit\`{a} and Sezione INFN, Turin, Italy
\item \Idef{org27}Dipartimento di Fisica e Astronomia dell'Universit\`{a} and Sezione INFN, Bologna, Italy
\item \Idef{org28}Dipartimento di Fisica e Astronomia dell'Universit\`{a} and Sezione INFN, Catania, Italy
\item \Idef{org29}Dipartimento di Fisica e Astronomia dell'Universit\`{a} and Sezione INFN, Padova, Italy
\item \Idef{org30}Dipartimento di Fisica `E.R.~Caianiello' dell'Universit\`{a} and Gruppo Collegato INFN, Salerno, Italy
\item \Idef{org31}Dipartimento di Scienze e Innovazione Tecnologica dell'Universit\`{a} del Piemonte Orientale and Gruppo Collegato INFN, Alessandria, Italy
\item \Idef{org32}Dipartimento Interateneo di Fisica `M.~Merlin' and Sezione INFN, Bari, Italy
\item \Idef{org33}Division of Experimental High Energy Physics, University of Lund, Lund, Sweden
\item \Idef{org34}Eberhard Karls Universit\"{a}t T\"{u}bingen, T\"{u}bingen, Germany
\item \Idef{org35}European Organization for Nuclear Research (CERN), Geneva, Switzerland
\item \Idef{org36}Excellence Cluster Universe, Technische Universit\"{a}t M\"{u}nchen, Munich, Germany
\item \Idef{org37}Faculty of Engineering, Bergen University College, Bergen, Norway
\item \Idef{org38}Faculty of Mathematics, Physics and Informatics, Comenius University, Bratislava, Slovakia
\item \Idef{org39}Faculty of Nuclear Sciences and Physical Engineering, Czech Technical University in Prague, Prague, Czech Republic
\item \Idef{org40}Faculty of Science, P.J.~\v{S}af\'{a}rik University, Ko\v{s}ice, Slovakia
\item \Idef{org41}Faculty of Technology, Buskerud and Vestfold University College, Vestfold, Norway
\item \Idef{org42}Frankfurt Institute for Advanced Studies, Johann Wolfgang Goethe-Universit\"{a}t Frankfurt, Frankfurt, Germany
\item \Idef{org43}Gangneung-Wonju National University, Gangneung, South Korea
\item \Idef{org44}Gauhati University, Department of Physics, Guwahati, India
\item \Idef{org45}Helmholtz-Institut f\"{u}r Strahlen- und Kernphysik, Rheinische Friedrich-Wilhelms-Universit\"{a}t Bonn, Bonn, Germany
\item \Idef{org46}Helsinki Institute of Physics (HIP), Helsinki, Finland
\item \Idef{org47}Hiroshima University, Hiroshima, Japan
\item \Idef{org48}Indian Institute of Technology Bombay (IIT), Mumbai, India
\item \Idef{org49}Indian Institute of Technology Indore, Indore (IITI), India
\item \Idef{org50}Indonesian Institute of Sciences, Jakarta, Indonesia
\item \Idef{org51}Inha University, Incheon, South Korea
\item \Idef{org52}Institut de Physique Nucl\'eaire d'Orsay (IPNO), Universit\'e Paris-Sud, CNRS-IN2P3, Orsay, France
\item \Idef{org53}Institut f\"{u}r Informatik, Johann Wolfgang Goethe-Universit\"{a}t Frankfurt, Frankfurt, Germany
\item \Idef{org54}Institut f\"{u}r Kernphysik, Johann Wolfgang Goethe-Universit\"{a}t Frankfurt, Frankfurt, Germany
\item \Idef{org55}Institut f\"{u}r Kernphysik, Westf\"{a}lische Wilhelms-Universit\"{a}t M\"{u}nster, M\"{u}nster, Germany
\item \Idef{org56}Institut Pluridisciplinaire Hubert Curien (IPHC), Universit\'{e} de Strasbourg, CNRS-IN2P3, Strasbourg, France
\item \Idef{org57}Institute for Nuclear Research, Academy of Sciences, Moscow, Russia
\item \Idef{org58}Institute for Subatomic Physics of Utrecht University, Utrecht, Netherlands
\item \Idef{org59}Institute for Theoretical and Experimental Physics, Moscow, Russia
\item \Idef{org60}Institute of Experimental Physics, Slovak Academy of Sciences, Ko\v{s}ice, Slovakia
\item \Idef{org61}Institute of Physics, Academy of Sciences of the Czech Republic, Prague, Czech Republic
\item \Idef{org62}Institute of Physics, Bhubaneswar, India
\item \Idef{org63}Institute of Space Science (ISS), Bucharest, Romania
\item \Idef{org64}Instituto de Ciencias Nucleares, Universidad Nacional Aut\'{o}noma de M\'{e}xico, Mexico City, Mexico
\item \Idef{org65}Instituto de F\'{\i}sica, Universidad Nacional Aut\'{o}noma de M\'{e}xico, Mexico City, Mexico
\item \Idef{org66}iThemba LABS, National Research Foundation, Somerset West, South Africa
\item \Idef{org67}Joint Institute for Nuclear Research (JINR), Dubna, Russia
\item \Idef{org68}Konkuk University, Seoul, South Korea
\item \Idef{org69}Korea Institute of Science and Technology Information, Daejeon, South Korea
\item \Idef{org70}KTO Karatay University, Konya, Turkey
\item \Idef{org71}Laboratoire de Physique Corpusculaire (LPC), Clermont Universit\'{e}, Universit\'{e} Blaise Pascal, CNRS--IN2P3, Clermont-Ferrand, France
\item \Idef{org72}Laboratoire de Physique Subatomique et de Cosmologie, Universit\'{e} Grenoble-Alpes, CNRS-IN2P3, Grenoble, France
\item \Idef{org73}Laboratori Nazionali di Frascati, INFN, Frascati, Italy
\item \Idef{org74}Laboratori Nazionali di Legnaro, INFN, Legnaro, Italy
\item \Idef{org75}Lawrence Berkeley National Laboratory, Berkeley, California, United States
\item \Idef{org76}Moscow Engineering Physics Institute, Moscow, Russia
\item \Idef{org77}Nagasaki Institute of Applied Science, Nagasaki, Japan
\item \Idef{org78}National Centre for Nuclear Studies, Warsaw, Poland
\item \Idef{org79}National Institute for Physics and Nuclear Engineering, Bucharest, Romania
\item \Idef{org80}National Institute of Science Education and Research, Bhubaneswar, India
\item \Idef{org81}National Research Centre Kurchatov Institute, Moscow, Russia
\item \Idef{org82}Niels Bohr Institute, University of Copenhagen, Copenhagen, Denmark
\item \Idef{org83}Nikhef, Nationaal instituut voor subatomaire fysica, Amsterdam, Netherlands
\item \Idef{org84}Nuclear Physics Group, STFC Daresbury Laboratory, Daresbury, United Kingdom
\item \Idef{org85}Nuclear Physics Institute, Academy of Sciences of the Czech Republic, \v{R}e\v{z} u Prahy, Czech Republic
\item \Idef{org86}Oak Ridge National Laboratory, Oak Ridge, Tennessee, United States
\item \Idef{org87}Petersburg Nuclear Physics Institute, Gatchina, Russia
\item \Idef{org88}Physics Department, Creighton University, Omaha, Nebraska, United States
\item \Idef{org89}Physics Department, Panjab University, Chandigarh, India
\item \Idef{org90}Physics Department, University of Athens, Athens, Greece
\item \Idef{org91}Physics Department, University of Cape Town, Cape Town, South Africa
\item \Idef{org92}Physics Department, University of Jammu, Jammu, India
\item \Idef{org93}Physics Department, University of Rajasthan, Jaipur, India
\item \Idef{org94}Physik Department, Technische Universit\"{a}t M\"{u}nchen, Munich, Germany
\item \Idef{org95}Physikalisches Institut, Ruprecht-Karls-Universit\"{a}t Heidelberg, Heidelberg, Germany
\item \Idef{org96}Purdue University, West Lafayette, Indiana, United States
\item \Idef{org97}Pusan National University, Pusan, South Korea
\item \Idef{org98}Research Division and ExtreMe Matter Institute EMMI, GSI Helmholtzzentrum f\"ur Schwerionenforschung, Darmstadt, Germany
\item \Idef{org99}Rudjer Bo\v{s}kovi\'{c} Institute, Zagreb, Croatia
\item \Idef{org100}Russian Federal Nuclear Center (VNIIEF), Sarov, Russia
\item \Idef{org101}Saha Institute of Nuclear Physics, Kolkata, India
\item \Idef{org102}School of Physics and Astronomy, University of Birmingham, Birmingham, United Kingdom
\item \Idef{org103}Secci\'{o}n F\'{\i}sica, Departamento de Ciencias, Pontificia Universidad Cat\'{o}lica del Per\'{u}, Lima, Peru
\item \Idef{org104}Sezione INFN, Bari, Italy
\item \Idef{org105}Sezione INFN, Bologna, Italy
\item \Idef{org106}Sezione INFN, Cagliari, Italy
\item \Idef{org107}Sezione INFN, Catania, Italy
\item \Idef{org108}Sezione INFN, Padova, Italy
\item \Idef{org109}Sezione INFN, Rome, Italy
\item \Idef{org110}Sezione INFN, Trieste, Italy
\item \Idef{org111}Sezione INFN, Turin, Italy
\item \Idef{org112}SSC IHEP of NRC Kurchatov institute, Protvino, Russia
\item \Idef{org113}Stefan Meyer Institut f\"{u}r Subatomare Physik (SMI), Vienna, Austria
\item \Idef{org114}SUBATECH, Ecole des Mines de Nantes, Universit\'{e} de Nantes, CNRS-IN2P3, Nantes, France
\item \Idef{org115}Suranaree University of Technology, Nakhon Ratchasima, Thailand
\item \Idef{org116}Technical University of Ko\v{s}ice, Ko\v{s}ice, Slovakia
\item \Idef{org117}Technical University of Split FESB, Split, Croatia
\item \Idef{org118}The Henryk Niewodniczanski Institute of Nuclear Physics, Polish Academy of Sciences, Cracow, Poland
\item \Idef{org119}The University of Texas at Austin, Physics Department, Austin, Texas, USA
\item \Idef{org120}Universidad Aut\'{o}noma de Sinaloa, Culiac\'{a}n, Mexico
\item \Idef{org121}Universidade de S\~{a}o Paulo (USP), S\~{a}o Paulo, Brazil
\item \Idef{org122}Universidade Estadual de Campinas (UNICAMP), Campinas, Brazil
\item \Idef{org123}University of Houston, Houston, Texas, United States
\item \Idef{org124}University of Jyv\"{a}skyl\"{a}, Jyv\"{a}skyl\"{a}, Finland
\item \Idef{org125}University of Liverpool, Liverpool, United Kingdom
\item \Idef{org126}University of Tennessee, Knoxville, Tennessee, United States
\item \Idef{org127}University of the Witwatersrand, Johannesburg, South Africa
\item \Idef{org128}University of Tokyo, Tokyo, Japan
\item \Idef{org129}University of Tsukuba, Tsukuba, Japan
\item \Idef{org130}University of Zagreb, Zagreb, Croatia
\item \Idef{org131}Universit\'{e} de Lyon, Universit\'{e} Lyon 1, CNRS/IN2P3, IPN-Lyon, Villeurbanne, France
\item \Idef{org132}Universit\`{a} di Brescia
\item \Idef{org133}V.~Fock Institute for Physics, St. Petersburg State University, St. Petersburg, Russia
\item \Idef{org134}Variable Energy Cyclotron Centre, Kolkata, India
\item \Idef{org135}Warsaw University of Technology, Warsaw, Poland
\item \Idef{org136}Wayne State University, Detroit, Michigan, United States
\item \Idef{org137}Wigner Research Centre for Physics, Hungarian Academy of Sciences, Budapest, Hungary
\item \Idef{org138}Yale University, New Haven, Connecticut, United States
\item \Idef{org139}Yonsei University, Seoul, South Korea
\item \Idef{org140}Zentrum f\"{u}r Technologietransfer und Telekommunikation (ZTT), Fachhochschule Worms, Worms, Germany
\end{Authlist}
\endgroup
\section{Introduction}
The main goal of the heavy-ion physics program at the Large Hadron Collider (LHC) is to study the quark-gluon plasma (QGP), a deconfined state of matter existing at extreme temperatures and energy-densities. Experimental results from RHIC were the first to suggest that the QGP behaves as a nearly perfect fluid \cite{Arsene:2004fa,Back:2004je,Adams:2005dq,Adcox:2004mh}. A particularly important observable when characterizing the QGP is anisotropic azimuthal flow. The anisotropic flow develops from pressure gradients originating from the initial spatial geometry of a collision and is observed as a momentum anisotropy in the final-state particles. It is usually described by flow harmonics, which are defined as the Fourier coefficients:
\begin{equation}
\mathrm{v}_n = \left\langle \cos \left[ n(\varphi-\Psi_n) \right] \right\rangle,
\end{equation}
where $n$ is the order of the flow harmonic, $\varphi$ is the azimuthal angle and $\Psi_n$ is the symmetry plane angle of harmonic $n$. The first three Fourier coefficients, $\mathrm{v}_1$, $\mathrm{v}_2$, and $\mathrm{v}_3$ are known as directed, elliptic and triangular flow, respectively. The flow harmonics $\mathrm{v}_1$ to $\mathrm{v}_6$ have been studied extensively at RHIC \cite{Arsene:2004fa,Back:2004je,Adams:2005dq,Adcox:2004mh,Adams:2003zg, Adare:2011tg, Adamczyk:2013waa} and the LHC \cite{Aamodt:2010pa,ALICE:2011ab,ATLAS:2011ah,Chatrchyan:2012wg,ATLAS:2012at,Chatrchyan:2012ta,Aad:2013xma,Abelev:2013cva,Chatrchyan:2013kba,Abelev:2014pua}. The observed anisotropic flow is considered to be a strong indication of collectivity \cite{Ollitrault:1992bk} and is described well by relativistic hydrodynamics \cite{Luzum:2008cw}.
Anisotropic flow studies at RHIC played a major role in establishing that the produced system is a strongly interacting quark-gluon plasma (sQGP) \cite{Arsene:2004fa,Back:2004je,Adams:2005dq,Adcox:2004mh} with a shear viscosity to entropy density ratio ($\eta/s$) close to the conjectured lower limit of $1/(4\pi)$ predicted by the AdS/CFT correspondence \cite{Kovtun:2004de}. The fact that higher order harmonics are increasingly suppressed by viscosity \cite{Alver:2010dn} makes it possible to use anisotropic flow measurements to estimate the $\eta/s$ of the produced system \cite{Luzum:2012wu,Gardim:2014tya}.
The pseudorapidity ($\eta$) dependence of the flow harmonics can play a key role in understanding the temperature dependence of $\eta/s$, something that can be determined using Quantum Chromodynamics (QCD) \cite{ Prakash:1993bt, Arnold:2003zc, Denicol:2015nhu}. At forward rapidities, the average temperature drops which implies $\eta/s$ will also change. In addition, the lower temperatures at forward rapidities mean the system will spend less time in the QGP phase leading to the hadronic viscosity playing a greater role in affecting the flow harmonics \cite{Molnar:2014zha, Denicol:2015nhu}. Recently, it has been suggested that the symmetry plane angles may depend on $\eta$ \cite{Gardim:2012im,Jia:2014vja,Khachatryan:2015oea}. While this effect is not directly studied in this Letter, considering that the reference particles are taken from mid-rapidity, the measured values of anisotropy coefficients at forward rapidity will be suppressed if the symmetry-plane angles fluctuate with $\eta$.
At RHIC, the PHOBOS experiment reported the pseudorapidity dependence of elliptic flow over a wide range ($-5.0 < \eta < 5.3$) and variety of collision energies \cite{Back:2002gz,Back:2004mh,Back:2004zg}, and system sizes \cite{Alver:2006wh}. It was found that in the rest frame of one of the colliding nuclei ($\eta-y_{\text{beam}}$), $\mathrm{v}_2$ is energy independent. This feature was also observed in multiplicity density distributions \cite{Bearden:2001qq,Alver:2010ck} and for $\mathrm{v}_1$ \cite{Back:2005pc}. This suggests that at forward rapidity, in the fragmentation region, particle production is independent of the collision energy, an effect known as extended longitudinal scaling.
In this Letter, we present measurements of $\mathrm{v}_2$, $\mathrm{v}_3$, and $\mathrm{v}_4$ over a wide pseudorapidity range ($-3.5 < \eta < 5.0$) in Pb--Pb collisions at $\sqrt{s_{\text{NN}}} = 2.76$ TeV using the ALICE detector. At the LHC, the pseudorapidity dependence of the flow harmonics has already been reported by ATLAS \cite{ATLAS:2012at,Aad:2014eoa} and CMS \cite{Chatrchyan:2012ta,Chatrchyan:2013kba} in a limited $\eta$-range ($|\eta| < 2.5$ and $|\eta| < 2.4$, respectively). The extended longitudinal scaling has been shown to hold for multiplicity densities \cite{Abbas:2013bpa} and directed flow \cite{Abelev:2013cva}, and appears to occur for elliptic flow \cite{Chatrchyan:2012ta,Aad:2014eoa}. Here, the $\eta$-range is extended considerably compared to the former results and we will investigate whether the extended longitudinal scaling of elliptic flow continues to hold. We will compare our data to hydrodynamical and transport models, and investigate the decrease of $\mathrm{v}_{n}$ in the forward regions relative to $\mathrm{d}N_{\mathrm{ch}}/\mathrm{d}\eta$.
\section{Experimental setup}
A detailed description of the ALICE detector is available elsewhere \cite{Aamodt:2008zz}. In this section, the sub-detectors used in this analysis are described: the V0 detector, the Time Projection Chamber (TPC), the Inner Tracking System (ITS) and the Forward Multiplicity Detector (FMD). The V0 detector consists of 2 arrays of scintillators located on opposite sides of the interaction point (IP) along the beam line. The detector has full azimuthal coverage in the ranges of $2.8 < \eta < 5.1$ (V0-A) and $-3.7 < \eta < -1.7$ (V0-C) \cite{Abbas:2013taa}. The detector acts as an online trigger and, with its large coverage, as a centrality estimator.
Charged particle tracks are reconstructed using the TPC, a large Time Projection Chamber \cite{Alme:2010ke}. The detector can provide position and momentum information. Particles that traverse the TPC volume leave ionization trails that drift towards the endcaps, where they are detected. Full length tracks can be reconstructed in the range $|\eta| < 0.8$. For this analysis, a transverse momentum range of $0.2 < p_{\text{T}} < 5.0$ GeV$/c$ was used. To ensure good track quality, the tracks are required to have at least 70 reconstructed TPC space points (cluster) out of 159 possible and an average $\chi^2$ per TPC cluster $\leq 4$. In addition, to reduce contamination from secondary particles (weak decays or interactions with material), a cut on the distance of closest approach (DCA) between the track and the primary vertex is applied both in the transverse plane (DCA$_{xy} < 2.4$ cm) and on the $z$-coordinate (DCA$_z < 3.2$ cm).
The ITS is made up of six cylindrical concentric silicon layers divided into three sub-systems, the Silicon Pixel Detector (SPD), the Silicon Drift Detector (SDD) and the Silicon Strip Detector (SSD), each consisting of two layers \cite{Aamodt:2008zz}. ITS clusters can be combined with the TPC information to improve track resolution. The SPD has additional applications \cite{Aamodt:2008zz}. Firstly, it is used to estimate the primary vertex as it is located close to the beam pipe. Secondly, clusters from the SPD inner layer, which consists of $3.3 \times 10^6$ pixels of size $50 \times 425$ $\mu$m$^2$, are used to estimate the number of charged particles in the range $|\eta| < 2.0$.
The FMD consists of five silicon rings, providing a pseudorapidity coverage in the ranges $-3.5 < \eta < -1.7$ and $1.7 < \eta < 5.0$ \cite{Christensen:2007yc}. The rings are single-layer detectors and only charged particle hits, not tracks, are measured. This means that primary and secondary particles cannot be distinguished. There are two types of FMD rings: inner ring and outer rings. Inner rings have 512 radial strips each covering $18^{\circ}$ in azimuth and outer rings have 256 radial strips each covering $9^{\circ}$ in azimuth. The charged particle estimation in the FMD is described in more detail elsewhere \cite{Abbas:2013bpa}. The inner layer of the SPD and the five FMD rings allow one to measure charged particle hits in the range $-3.5 < \eta < 5.0$.
\section{Data sample and analysis details}
We analysed $10$ million minimum bias Pb--Pb collisions at $\sqrt{s_{\text{NN}}}=2.76$ TeV. The sample was recorded during the first LHC heavy-ion data-taking period in 2010. A minimum-bias trigger requiring a coincidence between the signals from V0-A and V0-C was used. In addition, it is required that the primary vertex, determined by the SPD, be within $|\mathrm{v}_z| < 10.0$~cm, where $\mathrm{v}_{z} = 0$~cm is the location of the nominal interaction position. The measurements are grouped according to fractions of the inelastic cross section, and cover the $80\%$ most central collisions. The V0 detector is used for the centrality estimate which is described in more detail elsewhere \cite{Abelev:2013qoq}. For the most central to the most peripheral events, the V0 has a centrality resolution of $0.5\%$ to $2\%$, respectively.
The flow harmonics are estimated using the Q-cumulants method \cite{Bilandzic:2010jr} for two- and four-particle correlations, denoted as $\mathrm{v}_n\{2\}$ and $\mathrm{v}_n\{4\}$ respectively. The two- and four-particle cumulants respond differently to flow fluctuations. The two-particle cumulants are enhanced, while four-particle cumulants are suppressed. At forward rapidities, the pseudorapidity density is relatively low. This means that it is not always possible to get statistically significant results using only particles from a small region in $\eta$. To circumvent this using the Q-cumulants method, the reference flow measurement is performed using the charged particle tracks from the TPC, where the correlations at mid-rapidity are measured. As a systematic check, the charged particle tracks using a combination of the TPC and ITS are also used. Then, for the $v_n (\eta)$ analysis, the correlations between charged particle hits (from the SPD or FMD) and the tracks are measured in $\eta$-bins $0.5$ units of pseudorapidity wide. To avoid autocorrelations between the SPD clusters and tracks, the tracks for the reference particles are located in a different $\eta$-region than the SPD hits. Effectively, for SPD hits with $\eta < 0$, tracks are required to have $\eta > 0$ and vice versa. The same considerations apply for FMD hits. Due to the use of particle hits, only the $p_{\text{T}}$-integrated flow is measured. The $\phi$ distribution for the SPD or FMD clusters is not uniform, therefore a non-uniform acceptance correction is applied based on relations derived elsewhere \cite{Hansen:2014phd}.
As the inner rings of the FMD have only 20 azimuthal segments, the flow harmonics are slightly suppressed. The effect of this was recently calculated \cite{Bilandzic:2013kga} and found to be $1.6\%$, $3.7\%$ and $6.5\%$ for $\mathrm{v}_2$, $\mathrm{v}_3$ and $\mathrm{v}_4$ respectively. This suppression is taken into account in the final results. When using charged particle hits it is not possible to distinguish secondary particles (from material interactions and decays) from primary particles. For the regions covered by the SPD, the contamination from secondary particles is small ($< 10\%$), as the inner layer of the SPD is very close to the beam pipe. Away from mid-rapidity, in the FMD, dense material such as cooling tubes and read-out cables cause a very large production of secondary particles - up to twice the number of primary particles according to Monte Carlo (MC) studies. These secondary particles are deflected in $\varphi$ with respect to the mother particle, which causes a reduction in the observed flow. The reduction of flow caused by the secondary particles is estimated using an event generator containing particle yields, ratios, momentum spectra and flow coefficients, which are then subject to a full detector simulation using GEANT3 \cite{Brun:1994aa}. To make sure that the correction is not model dependent, the AMPT MC event generator \cite{Lin:2004en,Xu:2011fi} is used as an independent input, with GEANT3 again used to model the detector response. Using these simulations, the reduction is found to be larger for higher harmonics, up to $41\%$ for $\mathrm{v}_4$. Finally, the correction also accounts for missing very low $p_\text{T}$ particles, which increase the observed $v_{n}$ as these particles have a very small $v_n$. However, as the correction is always less than 1, the dominant effect comes from the secondary particles, which reduce $v_n$.
Few-particle correlations, not originating from the initial geometry termed non-flow (decays, jets, etc.), enhance the two-particle cumulant measurements. The non-flow contribution to the four-particle cumulant is found to be negligible \cite{Bilandzic:2010jr,Abelev:2014mda}, however, it is necessary to apply a correction to the two-particle cumulant. In the FMD and SPD, there is also a non-flow contribution from secondary particles, as they are sometimes produced in pairs. For the differential flow measurement, there is a rapidity-gap between the charged particle hits and the charged particle tracks. For the SPD, it is between 0 and 2 units in pseudorapidity, while for the FMD it is between 0.9 and 4.2 units in pseudorapidity. The large rapidity gap suppresses the non-flow contribution at forward rapidity. However, at mid-rapidities, this contribution is non-negligible and needs appropriate corrections. For the reference flow measurement there is no rapidity gap, and non-flow removal is important. For this analysis, the non-flow contributions are estimated using the HIJING event generator \cite{Wang:1991hta} and GEANT3 for the detector simulation. The non-flow contribution is estimated and subtracted separately for the reference and differential flow, before the correction for the deflection of secondary particles is applied and the $\mathrm{v}_n$ estimates are derived.
\section{Systematic uncertainties}
Numerous sources of systematic uncertainty were investigated, including effects due to detector cuts, choice of reference particles and uncertainties related to the secondary particle correction. Four major contributors to the systematic uncertainty were identified: the choice of reference tracks, the model dependence of the secondary particle correction, the description of the detector used for the simulations, and finally the non-flow correction. As the non-flow contribution to the four-particle cumulant is negligible, only the first three systematic uncertainties are considered for $\mathrm{v}_2\{4\}$. The systematic uncertainties assigned to each of the sources are shown in Table~\ref{tab:sys}, and are described in more detail below.
\begin{table}[th!]
\centering
\begin{tabular}{l c c c c c}
\hline
\hline
Source & $\mathrm{v}_2\{2\}$ & $\mathrm{v}_3\{2\}$ & $\mathrm{v}_4\{2\}$ & $\mathrm{v}_2\{4\}$ \\
\hline
Reference particle tracks & $2$-$4\%$ & $2$-$4\%$ & $2$-$6\%$ & $2$-$4\%$ \\
Model dependence & $5\%$ & $5\%$ & $7\%$ & $5\%$ \\
Material budget & $3$-$4\%$ & $3$-$4\%$ & $3$-$4\%$ & $3$-$4\%$ \\
Non-flow correction & $2$-$10\%$ & $2$-$10\%$ & $2$-$10\%$ & - \\
\hline
Total & $6$-$12\%$ & $6$-$13\%$ & $6$-$14\%$ & $6$-$8\%$\\
\hline
\hline
\end{tabular}
\caption{List of the systematic uncertainties for each observable.}
\label{tab:sys}
\end{table}
\begin{figure}[t!]
\centering
\includegraphics[width=\textwidth]{vn_panel}
\caption{Measurements of the pseudorapidity dependence of $\mathrm{v}_2$, $\mathrm{v}_3$ and $\mathrm{v}_4$ in each centrality bin. The vertical lines represent the statistical uncertainties and the boxes represent the systematic uncertainties. The statistical uncertainties are usually smaller than the marker size.}
\label{fig:panel}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=0.7\textwidth]{v2_rhic}
\caption{Elliptic flow for the $25$--$50\%$ centrality range. Boxes represent systematic uncertainties and errors bars represent statistical uncertainties. The results for $\mathrm{v}_2\{2\}$ from this analysis are compared to measurements using the event plane method from CMS \cite{Chatrchyan:2012ta} and ATLAS \cite{Aad:2014eoa} at the same energy and lower energy results from PHOBOS \cite{Back:2004mh}. For the comparable LHC energy, the $p_\text{T}$ range for ALICE is $p_\text{T} > 0 $ GeV/c, for CMS is $0.3 < p_\text{T} < 3 $ GeV/c, and for ATLAS is $p_\text{T} > 0.07 $ GeV/c.}
\label{fig:rhic}
\end{figure}
The dependence of the differential flow on the reference tracks was tested by using tracks with combined information from the TPC and ITS, rather than tracks with only TPC information. The systematic uncertainty from the choice of reference tracks was found to vary slightly with centrality, with the most central events having the largest uncertainty. To test the model dependence of secondary particle production, the correction from the toy-model described above is compared to the one derived from AMPT tuned to LHC data. Both the secondary particle correction and the non-flow correction derived from HIJING are sensitive to inaccuracies in the description of the detector used for the simulation. To test this sensitivity, the output of two HIJING simulations with a flow afterburner, one with $+7\%$ material density and one with $-7\%$ material density, are compared to the output from having normal material density. In this case the systematic uncertainty has a small $\eta$-dependence, as there are significantly fewer secondary particles at mid-rapidity. The 3\% uncertainty is applicable to the SPD, while the 4\% uncertainty is applicable to the FMD.
We assessed the systematic uncertainty associated with the non-flow correction in two ways. Firstly, following another method proposed to subtract non-flow \cite{Voloshin:2008dg}, the two-particle cumulants were obtained from minimum bias pp collisions, where it is assumed that there is negligible anisotropic flow. The pp reference and differential cumulants are then rescaled according to their multiplicity, $M$, using the ratio $M^{\text{pp}}/M^{\text{cent}}$, then subtracted from the corresponding A-A cumulants. Any differences found between this method and the default HIJING method are treated as systematic uncertainties. Secondly, by using only charged particle hits from the SPD and FMD, it is possible to construct a two-particle cumulant with a large rapidity-gap, $\mathrm{v}_n\{2,|\Delta\eta|>2.0\}$, which largely removes all non-flow contributions. Unfortunately, this observable is statistically stable only for $\mathrm{v}_2$ and $\mathrm{v}_3$, so it is used as a further cross check. In Table~\ref{tab:sys}, the 2\% uncertainties correspond to mid-central collisions where the ratio of flow to non-flow is largest, while the 10\% uncertainties correspond to very central and very peripheral collisions where the ratio of flow to non-flow is smallest. Finally, we used the AMPT model \cite{Lin:2004en,Xu:2011fi} to investigate if there are differences between $\mathrm{v}_{n}(\eta)$ and $\mathrm{v}_{n}(y)$, as $\eta$ is supposed to approximate $y$. We found there are 15\% differences in the flow coefficients at mid-rapidity, which reduced to 0\% for $\eta > 2$. We did not assign any systematic uncertainties due to these differences, as we are explicitly reporting measurements as a function of $\eta$ (as in the case of $\mathrm{d}N_{\mathrm{ch}}/\mathrm{d}\eta$ measurements).
The systematic uncertainty assigned to the non-flow correction is the largest contributor to the total systematic uncertainty, except for $\mathrm{v}_2\{4\}$ due to the four-particle cumulant's insensitivity to non-flow. The total systematic uncertainties are slightly dependent on centrality and pseudorapidity.
\begin{figure}[t!]
\centering
\includegraphics[width=0.7\textwidth]{v2_longscale}
\caption{The elliptic flow as observed in the rest frame of one of the projectiles by using the variable $|\eta|-y_{beam}$ ($y_{beam}=7.99$) for the event averaged $0$--$40\%$ centrality range. The results from $\mathrm{v}_2\{2\}$ from this analysis are compared to lower energy results from PHOBOS \cite{Back:2004zg}. The vertical lines represent the statistical uncertainties and the boxes represent the systematic uncertainties. For the PHOBOS results only statistical errors are shown.}
\label{fig:scaling}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[width=0.7\textwidth]{vn_cpratio}
\caption{Ratio of $\mathrm{v}_n\{2\}$ between central ($0$--$5\%$) and peripheral ($50$--$60\%$) events for $\mathrm{v}_2$, $\mathrm{v}_3$ and $\mathrm{v}_4$. The vertical lines represent the statistical uncertainties and the boxes represent the systematic uncertainties. The $\mathrm{v}_2$ results are multiplied by 3 to fit on the same scale as $\mathrm{v}_3$ and $\mathrm{v}_4$.}
\label{fig:cp}
\end{figure}
\begin{figure}[th!]
\centering
\includegraphics[width=0.7\textwidth]{vn_ratio}
\caption{Ratios between different harmonics for the $30$--$40\%$ centrality range. The vertical lines represent the statistical uncertainties and the boxes represent the common systematic uncertainties. In the bottom panel the ratios are rescaled to $1$ at mid-rapidity and the common systematic uncertainties are shown as the thick bars on the left.}
\label{fig:ratios}
\end{figure}
\section{Results}
An overview of the four observables in each centrality class is shown in Fig.~\ref{fig:panel}. Due to the changing overlap geometry, a strong centrality dependence of the elliptic flow is observed over the entire pseudorapidity range. The weaker centrality dependence of the higher order coefficients $\mathrm{v}_3$ and $\mathrm{v}_4$ is an indication that initial-state fluctuations play a prominent role, as the centrality dependence of the corresponding eccentricities are more modest relative to n=2 \cite{Alver:2010dn}. The different behaviour of $\mathrm{v}_2\{2\}$ and $\mathrm{v}_2\{4\}$ caused by flow fluctuations is also clearly seen. For the most peripheral events, there are not enough particles to get statistically stable results for $\mathrm{v}_2\{4\}$ and similarly for $\mathrm{v}_4\{2\}$ due to the relatively small quadrangular flow.
The $p_{\text{T}}$-integrated elliptic flow was also measured by CMS \cite{Chatrchyan:2012ta} and ATLAS \cite{Aad:2014eoa} in Pb--Pb collisions at $\sqrt{s_{\text{NN}}}=2.76$ TeV and by PHOBOS in Au--Au collisions at $\sqrt{s_{\text{NN}}}=200$ GeV \cite{Back:2004mh}. A comparison between those results and this analysis is shown for the $25$--$50\%$ centrality class in Fig.~\ref{fig:rhic}. In the common region of pseudorapidity acceptance, the results of present analysis are consistent with the results obtained by CMS and ATLAS experiments within the systematic uncertainties. The present analyses extends the measurements to a wider range of pseudorapidity. The values of $\mathrm{v}_{2}$ at all pseudorapidities measured at LHC energies are larger than the corresponding values at RHIC, as reported by PHOBOS. This increase in elliptic flow coincides with a larger $p_{\text{T}}$ at the LHC energy \cite{Aamodt:2010pa}.
The extended longitudinal scaling observed by PHOBOS in Au--Au collisions with center-of-mass energies from $19.6$ to $200$ GeV \cite{Back:2004zg} is found to hold up to the LHC energy (shown in Fig.~\ref{fig:scaling}). This is consistent with what was found by CMS \cite{Chatrchyan:2012ta} and ATLAS \cite{Aad:2014eoa}. Here it is shown as an event average for the $0$-$40\%$ most central events. The event average means that the analysis was performed in smaller centrality bins using multiplicity weights, and was then averaged over the centrality bins using the number of events as a weight \cite{Bilandzic:2010jr}. To examine boost invariance, it would be preferable to use rapidity ($y$) instead of pseudorapidity, unfortunately that is not possible using the FMD as the momentum cannot be measured.
PHOBOS found the shape of $\mathrm{v}_2(\eta)$ to be largely independent of centrality, with only the overall level changing between central and peripheral events \cite{Back:2004mh}. The ratios of central to peripheral events for $\mathrm{v}_2$, $\mathrm{v}_3$ and $\mathrm{v}_4$ using the two-particle cumulant are shown in Fig.~\ref{fig:cp}. Here it is observed that none of the harmonics show a clear centrality dependence in the shape of $\mathrm{v}_n(\eta)$ within uncertainties (albeit hints of such a dependence are present in the $\mathrm{v}_2$ ratio), consistent with the results from PHOBOS at lower energy.
\begin{figure}[t]
\centering
\includegraphics[width=1\textwidth]{vnOverdNdEta}
\caption{Ratios between $\mathrm{v}_{n}$ coefficients and $\mathrm{d}N_{\mathrm{ch}}/\mathrm{d}\eta$ values for different centralities. Measurements of $\mathrm{d}N_{\mathrm{ch}}/\mathrm{d}\eta$ are taken from a previous ALICE publication \cite{Abbas:2013bpa}. Only systematic uncertainties are shown, as the statistical uncertainties are smaller than the symbols.}
\label{fig:ratiosdndeta}
\end{figure}
It is known that the suppression from viscous effects to the flow harmonics increases with $n$ \cite{Alver:2010dn}. The hadronic phase is speculated to be more dominant at forward rapidity \cite{Molnar:2014zha, Denicol:2015nhu}. Therefore, the relative decrease of the flow harmonics may help to disentangle the viscous effects from the hadronic phase with those from the QGP phase. When the ratio $\mathrm{v}_{m}/\mathrm{v}_{n}$ $(n\neq m)$ is formed most of the common systematic uncertainties cancel, leaving the contribution from the non-flow correction. The ratios of $\mathrm{v}_3/\mathrm{v}_2$ and $\mathrm{v}_4/\mathrm{v}_3$ are shown for the $30$-$40\%$ most central events in Fig.~\ref{fig:ratios}. A small decrease with $\eta$ is observed for $\mathrm{v}_3/\mathrm{v}_2$, qualitatively consistent with the expectation from viscous effects suppressing higher harmonics. The $\mathrm{v}_4/\mathrm{v}_3$ ratio remains constant with $|\eta|$ within the uncertainties. The figure also shows $\mathrm{v}_4/\mathrm{v}_2^2$, which is commonly used to estimate the non-linear contribution to $\mathrm{v}_4$ from the elliptic anisotropy \cite{Adams:2003zg}. Given the uncertainties, it is difficult to conclude whether $\mathrm{v}_4/\mathrm{v}_2^2$ changes with respect to $\eta$.
As mentioned previously, at forward rapidities the steepness of $\mathrm{v}_{n}(\eta)$ has been linked to the hadronic contribution to the viscosity to entropy ratio \cite{Molnar:2014zha, Denicol:2015nhu}. The larger the hadronic $\eta / s$, the steeper the fall off. We also note that the pseudorapidity densities of charged particles decrease in this region. In order to investigate the correspondence of the latter, in Fig.~\ref{fig:ratiosdndeta} we show the ratio of various $\mathrm{v}_n$ coefficients to previous ALICE measurements of $\mathrm{d}N_{\mathrm{ch}}/\mathrm{d}\eta$ \cite{Abbas:2013bpa}. In order to avoid any influence of the Jacobian translation from $y$ to $\eta$, only the range $\eta > 2$ is shown. We find that this ratio is generally flat, with the exception of $\mathrm{v}_{2}$ at the larger values of $\eta$. This indicates that within a fixed centrality interval, $\mathrm{v}_{3}$ and $\mathrm{v}_{4}$ are largely driven by the local particle density. Indeed, when comparing p--Pb and Pb--Pb collisions at LHC energies, it was found that values of $\mathrm{v}_3\{2\}$ were similar for similar values of $\mathrm{d}N_{\mathrm{ch}}/\mathrm{d}\eta$ \cite{Abelev:2014mda}. The correlation found between both quantities may be simply attributed to the fact that both particle production and the development of anisotropic flow are driven by the number of interactions in the system.
In Fig.~\ref{fig:hydro}, we compare our data to hydrodynamic calculations tuned to RHIC data \cite{Denicol:2015nhu}. The tuning involves finding a parameterization of the temperature dependence of $\eta/s$, so that the hydrodynamical calculations describe PHOBOS measurements of $v_{2}(\eta)$ \cite{Back:2004mh,Back:2004zg}. It is clear that the same parameterization does not describe the LHC data as well. For both centralities, the elliptic flow coefficient $\mathrm{v}_2$ is generally underestimated, while the higher order coefficients $\mathrm{v}_3$ and $\mathrm{v}_4$ are generally overestimated. This points to the need for an either an alternative parameterization of $\eta/s$ that describes both the RHIC and LHC data simultaneously, or further investigations into whether the initial state model used is applicable for the LHC energies.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{vnHydro}
\caption{Comparisons to hydrodynamics predictions \cite{Denicol:2015nhu}, where input parameters (temperature dependence of $\eta/s$) have been tuned to RHIC data for the Pb-Pb 20-30\% (top) and 40-50\% (bottom) centralities. The predictions are for Pb-Pb $\sqrt{s_{\text{NN}}} = 2.76$ TeV collisions. }
\label{fig:hydro}
\end{figure}
In contrast to hydrodynamical models, AMPT is a non-equilibrium model that attempts to simulate parton production after the initial collision, and collective behaviour arises from parton and hadronic rescatterings. It has previously been tuned to agree with ALICE measurements of $\mathrm{v}_2$ vs. $p_{\text{T}}$ and multiplicity for the $40$-$50\%$ most central events. It was found to reproduce $\mathrm{v}_3(p_{\text{T}})$ well using the same parameters. In Fig.~\ref{fig:ampt} the results of this analysis are compared to the output of the AMPT model for two different centralities. For the centrality range of $40$--$50\%$, which AMPT is tuned to match, there is good agreement at mid-rapidity for all observables modulo $\mathrm{v}_2\{4\}$ at larger $|\eta|$, where AMPT underestimates the data. The underestimation at forward rapidity is found to be independent of the choice of reference particles, suggesting that it is unrelated to symmetry plane angle fluctuations with $\eta$. For more central events AMPT tends to overestimate flow at forward rapidities, except for $\mathrm{v}_4$ which it describes quite well over the entire range. At mid-rapidity AMPT agrees with the observed values of $\mathrm{v}_2$, $\mathrm{v}_3$ and $\mathrm{v}_4$ within the systematic uncertainties. Further tuning may lead to an improvement at forward rapidities, and should be investigated in future studies.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{vn_ampt}
\caption{Comparison to AMPT \cite{Lin:2004en,Xu:2011fi} for the centrality ranges $5$--$10\%$ and (top) and $40$--$50\%$ (bottom). The AMPT predictions are for Pb-Pb $\sqrt{s_{\text{NN}}} = 2.76$ TeV collisions.}
\label{fig:ampt}
\end{figure}
\section{Conclusions}
The pseudorapidity dependence of the anisotropic flow harmonics $\mathrm{v}_2$, $\mathrm{v}_3$ and $\mathrm{v}_4$ have been measured in Pb--Pb collisions at $\sqrt{s_{\text{NN}}}=2.76$ TeV using the ALICE detector. The measurement is performed over the widest $\eta$-range at the LHC, $-3.5 < \eta < 5.0$, in nine centrality bins covering $0$ to $80\%$ of the total inelastic cross section. It was found that the shape of $\mathrm{v}_n(\eta)$ does not depend obviously on centrality. Comparing to lower energy measurements at RHIC, elliptic flow is larger at the LHC over the entire pseudorapidity range and extended longitudinal scaling of $\mathrm{v}_2$ observed at lower collision energies is still valid up to the LHC energy. In the range $|\eta| < 2.5$ the results were found to be consistent with previous LHC measurements. At forward rapidities, the higher harmonic flow coefficients are proportional to the charged particle densities for a given centrality, while the ratio of $\mathrm{v}_2$ to $\mathrm{d}N_{\mathrm{ch}}/\mathrm{d}\eta$ rises with increasing $\eta$. A comparison to hydrodynamic calculations tuned to RHIC data has difficulties in describing our data in some $\eta$ regions, and this suggests that the LHC data play a key role in constraining either the temperature dependence of $\eta/s$ or the initial state. Finally, comparing our data to AMPT, the model describes the flow well at mid-rapidity, but fails for $\mathrm{v}_2$ at forward rapidities.
\newenvironment{acknowledgement}{\relax}{\relax}
\begin{acknowledgement}
\section*{Acknowledgements}
\input{acknowledgements.tex}
\end{acknowledgement}
\bibliographystyle{utphys}
|
\section{\textbf{Introduction and Main Results}}
It is assumed that the reader is familiar with the standard
notions used in the Nevanlinna value distribution theory such as
$T(r,f),m(r,f),N(r,f),S(r,f)$ etc., one may refer
to \cite{hayman-1}.
\medskip
A family $\mathcal F$ of meromorphic functions defined on a domain $D \subseteq \overline {\mC}$ is said to be normal in $D$ if every sequence of elements of $\mathcal F$ contains a subsequence which converges locally uniformly in $D$ with respect to the spherical metric, to a meromorphic function or $\infty$ (see \cite{schiff-1}).
\medskip
Two nonconstant meromorphic functions $f$ and $g$ defined on $D$ are said to share a meromorphic functions $\psi$ in $D$ if $\overline{E}_f(\psi)=\overline{E}_g(\psi)$, where
$$\overline{E}_f(\psi)=\{z\in D: f(z)=\psi(z)\}.$$
\medskip
The following Picard type theorem is one of the main result from Hayman's seminal paper \cite{hayman-2}:
\begin{theorem}(Hayman's alternative) Let $f$ be a nonconstant meromorphic function in $\mC$, $k$ a natural number and $c$ a nonzero complex number. Then $f$ or $f^{(k)}-c$ has a zero in $\mC$. If $f$ is transcendental, $f$ and $f^{(k)}-c$ has infinitely many zeros in $\mC$.
\end{theorem}
In 1979, Y.X. Gu \cite{gu-1} proved the following normality criterion corresponding to Hayman's alternative:
\begin{theorem}\label{theorem1}(Gu's normality criterion) Let $\mathcal{F}$ be a family of meromorphic functions defined in a domain $D$, and let $k$ be a positive integer. If, for every function $f\in\mathcal{F}$, $f\neq 0$ and $f^{(k)}\neq 1$ in $D$, then $F$ is normal in $D$.
\end{theorem}
Since then many variations of Theorem \ref{theorem1} have been obtained, for instance one can see \cite{fang-2, liu-1, nevo-1, schwick-2, yang-1,yang-2}.
In fact Schwick \cite{schwick-1} proved a more general version of Gu's result:
\begin{theorem}\label{theorem2} Let $\psi\not\equiv 0$ be a meromorphic function in a domain $D$ and $k \in \mN $. Let $\mathcal{F}$ be a family of meromorphic functions in $D$, such that $f\neq 0$ and $f^{(k)}\neq \psi $, and $f$ and $\psi$ have no common poles for each $f \in \mathcal{F}$. Then $\mathcal{F}$ is normal in $D$.
\end{theorem}
In 2004, Fang and Zalcman \cite{fang-1} proved the following generalization of Theorem \ref{theorem1} by considering the sharing of values:
\begin{theorem} \label{theorem3}
Let $k$ be a positive integer and $b$ be a nonzero complex constant. Let $\mathcal{F}$ be a family of meromorphic functions on $D$, all of whose zeros have multiplicity at least $k+2$, such that for each pair of functions $f$ and $g$ in $\mathcal{F}$, $f$ and $g$ share the value $0$, and $f^{(k)}$ and $g^{(k)}$ share the value $b$ in $D$, then $\mathcal{F}$ is normal in $D$.
\end{theorem}
Recently, J. Chang \cite{chang-1} proved the following result by replacing the constant $b$ by a holomorphic function:
\begin{theorem} \label{theorem4} Let $k \in \mN$ and $h(\not\equiv 0)$ be a function holomorphic on $D$. Let $\mathcal{F}$ be a family of meromorphic functions in $D$, all of whose zeros have multiplicity at least $k+2$, such that for each pair of functions $f$ and $g$ in $\mathcal{F}$, $f$ and $g$ share the value $0$, and $f^{(k)}$ and $g^{(k)}$ share the function $h$. Suppose additionally that at each common zero of $f$ and $h$ for every $f \in \mathcal{F}$, the multiplicities $m_f$ for $f$ and $m_h$ for $h$ satisfy $m_f\geq m_h +k+1$ for $k>1$ and $m_f \geq 2m_h +3$ for $k=1$. Then, $F$ is normal in $D$.
\end{theorem}
Examples are also given in \cite{chang-1} for the sharpness of conditions in Theorem \ref{theorem4}. Working in this direction, we prove the following generalization of Theorem \ref {theorem4} :
\begin{theorem}\label{theorem5} Let $\mathcal{F}$ be a family of meromorphic functions in a domain $D$, and let $k$ be a positive integer. Suppose that $\phi$ is a holomorphic function on $D$ and $\psi$ is a meromorphic function on $D$ such that $\phi^{(k)} (z)\not\equiv \psi (z)$. Suppose that for each pair of functions $f$ and $g$ in $\mathcal{F}$, $f$ and $g$ share $\phi$, and $f^{(k)}$ and $g^{(k)}$ share the function $\psi$.
Suppose further that
\begin{enumerate}
\item every $f \in\mathcal{F}$, $f-\phi$ has zeros of multiplicity at least $k+2$,
\item for every common zero of $f-\phi$ and $\psi -\phi^{(k)}$, the multiplicities $m_{f-\phi}$ for $f-\phi$ and $m_{\psi-\phi^{(k)}}$ for $\psi -\phi^{(k)}$ satisfy $$m_{f-\phi} \geq m_{\psi-\phi^{(k)}}+k+1 \text{~for~} k>1,\text{~and}$$ $$m_{f-\phi} \geq 2 m_{\psi-\phi^{(k)}}+3 \text{~for~} k=1, $$
\item for every $f \in \mathcal{F}$, $f$ and $\psi$ have no common poles in $D$.
\end{enumerate}
Then $\mathcal{F}$ is normal on $D$.
\end{theorem}
\begin{remark} If $\phi\equiv 0 $ and $\psi$ is a holomorphic function, then Theorem \ref{theorem5} reduces to Theorem \ref{theorem4}. Thus, the conditions (1) and (2) in Theorem \ref{theorem5} can easily be seen to be essential.
\end{remark}
\begin{example}Consider the family $$\mathcal F =\left\{\frac{1}{2mz}:m\in \mN\right\}$$ on the open unit disk $\mD$, and let $\phi(z)=1/z$ and $\psi (z)\equiv 0$. Then clearly, for every $f,g \in \mathcal F$, $f$ and $g$ share $\phi(z)$, and $f^{(k)}$ and $g^{(k)}$ share $\psi(z)$ in $\mD$. However, the family $F$ is not normal in $\mD$. This shows that $\phi$ cannot be taken meromorphic in Theorem \ref{theorem5}.
\medskip
Further, for the same family $\mathcal F$, if we take $\phi(z)\equiv 0$ and $\psi(z)=1/z^{k+1}$, then for every $f,g\in \mathcal F$, $f$ and $g$ share $\phi(z)$, and $f^{(k)}$ and $g^{(k)}$ share $\psi(z)$ in $\mD$. But, the family $\mathcal F$ is not normal in $\mD$. This shows that the condition (3) in Theorem \ref{theorem5} is essential.
\end{example}
\begin{example}Consider the family $$\mathcal F =\left\{z^{k+1}+\frac{1}{mz}:m\in \mN\right\}$$ on the open unit disk $\mD$, and let $\phi(z)=z^{k+1}$ and $\psi (z)=(k+1)!z$. Then clearly, for every $f,g \in \mathcal F$, $f$ and $g$ share $\phi(z)$, and $f^{(k)}$ and $g^{(k)}$ share $\psi(z)$ in $\mD$. However, the family $F$ is not normal in $\mD$. This shows that the condition $\phi^{(k)}(z)\not\equiv \psi(z)$ in Theorem \ref{theorem5} cannot be dropped.
\end{example}
\begin{corollary}\label{corollary1} Let $n \geq 2$ be a positive integer and $\psi(\not\equiv 0)$ be a function meromorphic on $D$. Let $\mathcal{F}$ be a family of meromorphic functions in $D$ such that for each pair of functions $f$ and $g$ in $\mathcal{F}$, $f$ and $g$ share the value $0$, and $f^n f'$ and $g^n g'$ share the function $\psi$. Suppose further
\begin{enumerate}
\item for every common zero of $f$ and $\psi$, the multiplicities $m_f$ for $f$ and $m_{\psi}$ for $\psi$ satisfy $$m_{f} \geq m_{\psi}+k+1 \text{~for~} k>1,\text{~and}$$ $$m_{f} \geq 2 m_{\psi}+3 \text{~for~} k=1, $$
\item for every $f \in \mathcal{F}$, $f$ and $\psi$ have no common poles in $D$.
\end{enumerate}
Then $\mathcal{F}$ is normal on $D$.
\end{corollary}
\medskip
Corollary \ref{corollary1} follows by setting $\mathcal {G}=\left\{f^{n+1}/(n+1): f\in \mathcal {F}\right\}$ and applying Theorem \ref{theorem5} to this family with $\phi(z)\equiv 0$ and $k=1$.
\section{Proof of the Main result}
For $z_0\in\mathbb{C}$ and $r>0$, we denote by $D_r(z_0)$ the open unit disk with centre $z_0$ and radius $r$, and $D' _r(z_0)$ the corresponding punctured disk. To prove our main result-Theorem \ref{theorem5}, we require the following lemma:
\begin{lemma}\label{lemma1}\cite{chen-1} Let $\mathcal{F}$ be a family of meromorphic functions in a domain $D$, all of whose zeros have multiplicity at least $k$. Then, if $\mathcal{F}$ is not normal at $z_0$, then for each $\beta:-1 < \beta <k$, there exist points $z_n \in D$ with $z_n\rightarrow z_0$, functions $f_n \in \mathcal{F}$ and positive numbers $\rho_n \rightarrow 0$ such that $$g_n(\zeta):=\rho^{-\beta}_n f_n (z_n+ \rho_n \zeta)$$ converges locally uniformly with respect to the spherical metric in $\mC$ to a nonconstant meromorphic function $g$ of finite order, all of whose zeros have multiplicity at least $k$.
\end{lemma}
Further, recall that the sets $\left\{E_{\lambda}\right\}_{\lambda \in \Lambda}$ are said to be {\it locally uniformly discrete} in $D$, if for each point $z_0 \in D$, there exists $\delta>0$ such that either $E_{\lambda}\cap D_\delta(z_0)$ is an empty set or a singleton, one may refer to \cite{chang-1}.
\begin{proof} [\textbf{Proof of Theorem \ref{theorem5}.}] Since normality is a local property, it is enough to show that $\mathcal{F}$ is normal at each $z_0\in D$. We distinguish the following cases.\\\\
\textbf{Case I.} Suppose that there exist $f\in \mathcal{F}$ such that $f(z_0)\neq \phi(z_0)$ and $f^{(k)}(z_0)\neq \psi(z_0)$. Then we can find $r>0$ such that $D_r(z_0)\subset D$, and $f(z)\neq \phi(z) $ and $f^{(k)}(z)\neq \psi(z)$ in $D_r(z_0)$ and so by the given sharing condition, $f\neq \phi$ and $f^{(k)} \neq \psi$ for every $f\in \mathcal F$ in $D_r(z_0)$. Now set $\alpha:=\psi - \phi^{(k)}$ and consider the family $$\mathcal G=\left\{g=f-\phi:f\in \mathcal{F}\right\}.$$ Then clearly $\alpha(z)\not\equiv 0$ and for every $g\in \mathcal G$, $g(z)\neq 0$ and $g^{(k)}(z)\neq \alpha (z)$. Thus by Theorem \ref{theorem2} $\mathcal G$ is normal in $D_r(z_0)$. Since $\mathcal G$ is normal if and only if $\mathcal F$ is normal, $\mathcal F$ is normal at $z_0$.\\\\
\textbf{Case II.} Suppose that there exist $f\in \mathcal F$ such that $f(z_0)=\phi(z_0)$ or $f^{(k)}(z_0)=\psi(z_0)$. Then we can find $r>0$ such that $D_r(z_0)\subset D$, and $f(z)\neq \phi(z)$ and $f^{(k)}(z)\neq \psi(z)$ in $D'_r(z_0)$ and so by the given sharing condition, $f(z)\neq \phi(z)$ and $f^{(k)}(z)\neq \psi(z)$ for every $f\in \mathcal F$ in $D'_r(z_0)$. Thus for any $z_1 \in D_r(z_0)$, there exists $\delta>0$ such that every $E_f$ has at most one point lying in $D_\delta (z_1)$, where $$E_f:=\left\{z\in D_r(z_0):f(z)=\phi(z)\right\}\cup \left\{z\in D_r(z_0):f^{(k)}(z)=\psi(z)\right\}.$$ Therefore, the sets $\left\{E_f\right\}_{f\in \mathcal F}$ are locally uniformly discrete in $D_r(z_0)$.
\medskip
As in Case I, consider the family $\mathcal G=\left\{g=f-\phi:f\in \mathcal{F}\right\}$ and set $\alpha:=\psi - \phi^{(k)}$. Then clearly the sets $\left\{E_g\right\}_{g\in \mathcal G}$ are locally uniformly discrete in $D_r(z_0)$, where $$E_g=\left\{z\in D_r(z_0):g(z)=0\right\}\cup \left\{z\in D_r(z_0):g^{(k)}(z)=\alpha(z)\right\}.$$ If $\alpha(z)$ is holomorphic in $D_r(z_0)$, then by \cite[Theorem 4, p.49]{chang-1}, $\mathcal G$ is normal in $D_r(z_0)$ and hence $\mathcal F$ is normal at $z_0$. Suppose that $\alpha(z)$ is not holomorphic in $D_r(z_0)$. Assume that $z_0$ is a pole of $\alpha(z)$. Then we can find $\delta>0$ such that $D_\delta (z_0) \subset D_r(z_0)$ and $\alpha(z)$ is holomorphic in $D'_\delta(z_0)$, and thus $\mathcal G$ is normal in $D'_\delta(z_0)$. Next, consider the family $$\mathcal H:=\left\{h(z)=\frac{g(z)}{\alpha(z)}:g\in \mathcal G\right\}.$$ Noting that $z_0$ is a pole of $\alpha(z)$ and for every $f\in \mathcal F$, $f$ and $\psi$ have no common poles implies that\\
$(a)$ for every $g \in \mathcal G$, $g$ and $\alpha$ have no common poles and hence for every $h\in \mathcal H$, $z_0$ is a zero of $h$ of multiplicity at least $k+3$,\\
$(b)$ there exists $\eta >0$ such that $D_{\eta}(z_0)\subset D_\delta(z_0)$ and for every $h \in \mathcal H$, $h\neq 1,\infty$ in $D_{\eta}(z_0)$.
\medskip
We first prove that $\mathcal H$ is normal at $z_0$. Suppose on contrary that $\mathcal H$ is not normal at $z_0$. Then by Lemma \ref{lemma1}, we can find a sequence $\left\{ h_j \right\}$ in $\mathcal{H}$, a sequence $\left\{ z_j\right\}$ of complex numbers with $z_j\rightarrow 0$ and a sequence $\left\{\rho_j\right\}$ of positive real numbers with $\rho_j \rightarrow 0 $ such that $$H_j (\zeta) = h_j (z_j +\rho_j \zeta)$$ converges locally uniformly with respect to the spherical metric to a nonconstant meromorphic function $H(\zeta)$ on $\mC$, all of whose zeros have multiplicity at least $k+3$. Also by Hurwitz theorem, we have $H\neq 1,\infty$ on $\mC$. Thus by second fundamental theorem of Nevanlinna, we have
\begin{align*} T(r,H)&\leq \overline N(r,H) + \overline N\left(r,\frac{1}{H}\right)+ \overline{N}\left(r,\frac{1}{H-1}\right)+S(r,H)\\
&\leq \frac{1}{k+3}N\left(r,\frac{1}{H}\right) + S(r,H)\\
& \leq \frac{1}{k+3}T(r,H)+ S(r,H),
\end{align*}
which is a contradiction. Therefore $\mathcal H$ is normal at $z_0$. Now we turn to prove the normality of $\mathcal G$ at $z_0$.
\medskip
Suppose that $\mathcal{G}$ is not normal at $z_0$. Since $\mathcal H$ is normal at $z_0$, it is equicontinuous at $z_0$ with respect to the spherical metric. Also $h(z_0)=0$ for every $h\in \mathcal H$. Thus there exists $\delta_1>0$ such that $D_{\delta_1}(z_0)\subset D_{\delta}(z_0)$ and $|h(z)|\leq 1$ for every $h\in \mathcal H$ in $D_{\delta_1}(z_0)$. It follows that $\mathcal G$ is a family of holomorphic functions in $D_{\delta_1}(z_0)$.
\medskip
Let $\left\{g_n\right\}$ be a sequence in $\mathcal G$. Since $\mathcal{G}$ is normal in $D'_{\delta_1} (z_0)$ but not at $z_0$, there exists a subsequence of $\left\{g_n\right\}$, which we may take as $\left\{g_n\right\}$ itself, which converges locally uniformly on $D'_{\delta_1} (z_0)$ but not on $D_{\delta_1} (z_0)$. By the maximum modulus principle, we have $\left\{g_n\right\}$ converges locally uniformly to $\infty$ in $D'_{\delta_1}(z_0)$ and hence $\left\{h_n\right\}$ converges locally uniformly to $\infty$ in $D'_{\delta_1}(z_0)$, which is a contradiction to the fact that $|h(z)|\leq 1$ for every $h\in \mathcal H$ in $D_{\delta_1}(z_0)$.
Thus $\mathcal G$ is normal at $z_0$ and hence $\mathcal F$ is normal at $z_0$.
\end{proof}
\bibliographystyle{amsplain}
|
\section{Introduction}
When materials are stretched in a particular direction, they typically contract in the directions orthogonal to the applied loading. The Poisson's ratio ($\nu$) is the material property that characterizes this behavior, and it is typically positive, in the range of $0.2< \nu < 0.5$ for most engineering materials. A negative Poisson's ratio (NPR), in which the material expands in the directions orthogonal to the applied loading, is allowed by classical elasticity theory, which sets a range of $-1< \nu < 0.5$ for the Poisson's ratio in an isotropic three-dimensional material.\cite{LandauLD}
Because most materials have a positive Poisson's ratio, significant efforts have been made to discover NPR materials starting with the seminal work of Lakes in 1987.\cite{LakesR1987sci} In this experiment, the NPR was induced by the permanent compression of a conventional low-density open cell polymer foam, which was explained by the re-entrant configuration of the cell. Since then, many works have achieved the NPR through structural engineering or via composite structures.\cite{RothenburgL1991nat,LakesR1993adm,BaughmanRH1993nat,EvansKE2000adm,YangW2004jmsci,RaviralaN2007jms,LethbridgeZAD2010am,BertoldiK2010am,GreavesGN2011nm,AldersonK2012pssb,clausenAM2015} Materials with NPR have become known as auxetic, as coined by Evans.\cite{EvansKE1991Endeavour}
Besides the NPR phenomena found in specifically engineered structures, the NPR has also been found to be intrinsic for some materials. For example, some cubic elemental (both FCC and BCC) metals have intrinsic NPR along a non-axial, i.e. $\langle110\rangle$ direction.\cite{MilsteinF1979prb,BaughmanRH1998nat} The Poisson's ratio for FCC metals can be negative along some principal directions by controlling the transverse loading.\cite{hoPSSB2016b} NPR was found to be intrinsic to single-layer black phosphorus due to its puckered configuration, which leads to NPR in the out-of-plane direction.\cite{JiangJW2014bpnpr} NPR was also predicted to be intrinsic for few-layer orthorhombic arsenic using first-principles calculations.\cite{HanJ2015ape}
While most existing works are on bulk auxetic structures, some theoretical works have recently emerged predicting NPR in nanomaterials, through a range of different mechanisms. For example, Yao et al. investigated the possibility of inducing NPR in carbon nanotubes, though significant, and likely non-physical changes to either the structural parameters or bonding strengths were found to be necessary for the NPR to appear.\cite{YaoYT2008pssb} The NPR for metal nanoplates was found due to a surface-induced phase transformation.\cite{HoDT2014nc} Furthermore, the NPR for graphene has been found through various means, including due to the rippling curvature induced by entropic effects at very high (1700 K) temperatures\cite{ZakharchenkoKV}, by creating periodic porous graphene structures\cite{hoPSSB2016a}, and introducing many vacancy defects\cite{GrimaJN2015adm}. More recently, two of the current authors found that the compressive edge stress-induced warping of the free edges can cause NPR in graphene nanoribbons with widths smaller than 10~nm.\cite{JiangJW2016npr_fbc}
The salient point in the above literature review is that it is important to uncover intrinsic NPR in specific materials, including graphene, the most widely studied material since its discovery in 2004. However, the literature to-date has shown that the NPR phenomenon can only be observed in graphene after specific engineering of its structure, such as thermally-induced ripples, vacancy defects, or free edges. More specifically, it is still unclear whether the NPR is an intrinsic property for graphene.
In this letter, we reveal that NPR is intrinsic to single-layer graphene, and is independent of its size and temperature. More specifically, the Poisson's ratio evolves from positive to negative when the applied tensile strain exceeds about 6\%. We find that this NPR is due to the interplay between two fundamental deformation pathways, which we term PW-I and PW-II, and which correspond to two characteristic interactions in graphene, i.e., the bond stretching and angle bending. The PW-I deformation mode yields a positive Poisson's ratio, while the PW-II deformation mode results in a NPR. We therefore propose a pathway-based energy criterion, which predicts that the PW-II mode becomes more important than the PW-I mode and dominates the deformation mechanism of graphene for strains larger than 6\%. Consequently, the Poisson's ratio becomes negative when the applied tensile strain is larger than 6\%.
\textbf{Results.} The crystal structure for single-layer graphene is shown in the inset of Figure~\ref{fig_poisson}(b). Periodic boundary conditions are applied in both x and y-directions such that our studies, and the properties we report, represent those of bulk graphene without edge effects. The carbon-carbon interactions are described by the Brenner potential,\cite{brennerJPCM2002} which has been widely used to study the mechanical response of graphene.\cite{MoY2009nat} The structure is stretched in the x (armchair)-direction while graphene is allowed to be fully relaxed in the y (zigzag)-direction, using both molecular dynamics (MD) or molecular statics (MS) simulations. For the MD simulations, the standard Newton equations of motion are integrated in time using the velocity Verlet algorithm with a time step of 1~{fs}, which is small enough to maintain energy conservation during the MD simulations. This time step is also small enough to accurately discretize the atomic trajectory corresponding to the highest-frequency vibration modes in graphene (with frequency around\cite{JiangJW2008prb} $4.8\times 10^{13}$~Hz). For the MS simulations, the conjugate gradient algorithm is used for energy minimization. Simulations are performed using the publicly available simulation code LAMMPS~\cite{PlimptonSJ}, while the OVITO package is used for visualization~\cite{ovito}.
Figure~\ref{fig_poisson}~(a) shows the resultant strain $\epsilon_y$ in the y-direction in graphene that is stretched by $\epsilon_x$ in the x-direction. The x-axis is along the horizontal direction, while the y-axis is in the vertical direction as shown in the inset of Figure~\ref{fig_poisson}~(b). The resultant strain in the y-direction is computed by $\epsilon_{y} = \frac{L_{y}-L_{y0}}{L_{y0}}$ with $L_{y0}$ and $L_y$ as the initial and deformed lengths in the y-direction. We simulate the tensile deformation of graphene using both MD and MS simulations. MD simulations are carried out for square graphene of size $L=200$~{\AA} and 300~{\AA} at 4.2~K and 300~K. MS simulations are performed for graphene of dimension $L=20$~{\AA} and 50~{\AA}.
As shown in Figure~\ref{fig_poisson}~(a), there is a robust valley point around $\epsilon_x=0.06$ (6\%) in all of these curves. This valley point indicates that the Poisson's ratio, calculated by\cite{HoDT2014nc} $\nu=-\partial \epsilon_y/\partial \epsilon_x$, is positive for $\epsilon_x<0.06$ but becomes negative for $\epsilon_x>0.06$. These results demonstrate that the NPR is robust, as it is observed for both low and room temperature conditions, as well as for all structural sizes we have considered. We note that graphene is highly stretchable and has been stretched in a wide strain range experimentally. A strain up to 0.15 has been applied on graphene to measure the nonlinear stiffness\cite{LeeC2008sci} or manipulate its rippled structure.\cite{BaoW2009nn} Furthermore, recent experiments reported a uniaxial strain up to 0.1, which can be controlled in a reversible and nondestructive manner in graphene.\cite{GarzaHHP2014nl} Thus, the critical strain of $\epsilon_x=0.06$ we find has been achieved in contemporary experiments on graphene, so that theoretical results in the present work are experimentally verifiable. The Poisson's ratio is about 0.3 at $\epsilon_x\approx 0$, which agrees with a recent numerical result with the realistic interatomic potential LCBOPII.\cite{LosJH2006prl}
Figure~\ref{fig_poisson}~(a) shows that results from MD simulations at 4.2~K coincide with the results from MS simulations, so we will concentrate on the MS simulation results for the rest of this paper. Figure~\ref{fig_poisson}~(b) shows the strain dependence for the Poisson's ratio extracted from these two curves from MS simulations in panel (a). It explicitly shows that the Poisson's ratio is negative for $\epsilon_x>0.06$.
We note that there have been previous reports of NPR in graphene. Specifically, large numbers of vacancy defects\cite{GrimaJN2015adm} or patterning periodic porous structures\cite{hoPSSB2016a} for bulk graphene or compressive edge stress-induced warping in graphene ribbons are three different mechanisms to achieve the NPR in graphene.\cite{JiangJW2016npr_fbc} The Poisson's ratio for graphene can also be driven into the negative regime by thermally induced ripples at high temperatures above 1700~K.\cite{ZakharchenkoKV} In contrast, the NPR revealed in the present work represents an intrinsic material property for single-layer graphene.
\textbf{Discussion.} To explore the underlying mechanism for the intrinsic NPR, we first illustrate two major deformation modes for the tensile deformation of graphene in Figure~\ref{fig_pathway}. These two deformation modes are fundamental deformation modes corresponding to the bond stretching and angle bending interactions;\cite{ChangT2003jmps} i.e., $V_{b} = \frac{K_{b}}{2}\left(b-b_{0}\right)^{2}$ and $V_{\theta} = \frac{K_{\theta}}{2}\left(\theta-\theta_{0}\right)^{2}$, where $b$ is the bond length and $\theta$ is the angle in the deformed graphene, and $\theta_0=120^{\circ}$ and $b_0=1.42$~{\AA} are material constants related to undeformed graphene. $K_{b}$ is the force constant that characterizes the resistance to bond stretching, and so a larger value of $K_{b}$ indicates a stiffer bond. $K_{\theta}$ characterizes the resistance to angle bending, and so a larger value of $K_{\theta}$ means a larger resistance to angle bending deformations. The values of these force constants can be obtained by using the value of the Young's modulus and the Poisson's ratio of graphene.\cite{ChangT2003jmps} The bond stretching and angle bending are two major interaction terms in graphene, especially in graphene without out-of-plane deformation. We note that the bond stretching and angle bending interactions were used to derived analytic expressions for the Poisson's ratio in carbon nanotubes.\cite{ChangT2003jmps,ShenL2004prb,ChangTC2005apl,YaoYT2008pssb} The analytic expressions illustrate the dependence of the Poisson's ratio on geometrical parameters and force constants. For example, Yao et al. performed a speculative investigation on the possibility of NPR for carbon nanotubes by varying one parameter (or ratio of parameters) while holding all other parameters unchanged.\cite{YaoYT2008pssb} However, it was determined that significant, and likely non-physical changes to the geometry or material constants would be necessary for the NPR to appear in CNTs.
The overall deformation process for graphene depends on the competition between the bond stretching and angle bending interactions. For $V_b\gg V_\theta$, the bonds are too stiff to be stretched ($\Delta b\approx 0$), so only the bond angles will change during the tension of graphene. This type of deformation will be referred to as the PW-I deformation mode, as shown in the top (blue online) of Figure~\ref{fig_pathway}. For $V_b\ll V_\theta$, the bond angles cannot be changed ($\Delta \theta\approx 0$), so the bond lengths will be stretched to accommodate the applied tension. This type of deformation will be referred to as the PW-II deformation mode, as shown in the bottom (red online) of Figure~\ref{fig_pathway}.
The Poisson's ratio corresponding to the PW-I and PW-II deformation modes can be derived by simple algebra. The unit cell is displayed by the parallelogram gray area in the left configuration in Figure~\ref{fig_pathway}. The size of the cell in Figure~\ref{fig_pathway} in the x and y-directions are
\begin{eqnarray}
L_{x} & = & 2\left(b_{1}+b_{2}\cos\frac{\theta_1}{2}\right)\\
L_{y} & = & 2b_{2}\sin\frac{\theta_{1}}{2},
\end{eqnarray}
which yields
\begin{eqnarray}
dL_{x} & = & 2\left(db_{1}+db_{2}\cos\frac{\theta_1}{2}-\frac{b_{2}}{2}\sin\frac{\theta_{1}}{2}d\theta_{1}\right)\\
dL_{y} & = & 2\left(db_{2}\sin\frac{\theta_{1}}{2}+\frac{b_{2}}{2}\cos\frac{\theta_{1}}{2}d\theta_{1}\right).
\end{eqnarray}
As a result, the Poisson's ratio is
\begin{eqnarray}
\nu & = & -\frac{\epsilon_{y}}{\epsilon_{x}}=-\frac{dL_{y}/L_{y}}{dL_{x}/L_{x}}\nonumber\\
& = & -\frac{b_{1}+b_{2}\cos\frac{\theta_{1}}{2}}{b_{2}\sin\frac{\theta_{1}}{2}}\times\frac{db_{2}\sin\frac{\theta_{1}}{2}+\frac{b_{2}}{2}\cos\frac{\theta_{1}}{2}d\theta_{1}}{db_{1}+db_{2}\cos\frac{\theta_{1}}{2}-\frac{b_{2}}{2}\sin\frac{\theta_{1}}{2}d\theta_{1}},
\label{eq_nu}
\end{eqnarray}
where $b_{1}=b_{2}=b_{0}$ and $\theta_{1}=\theta_{2}=\theta_{0}$ for undeformed graphene. We note that, for small strains, the definition of the Poisson's ratio in equation~(\ref{eq_nu}) is consistent with the numerical formula used in the above to extract the Poisson's ratio in Figure~\ref{fig_poisson}~(b), because $\epsilon_y$ and $\epsilon_x$ have a linear relationship for small strains.
For the PW-I mode, we have $\Delta b_{1}\approx0$, $\Delta b_{2}\approx0$, and $\Delta\theta_{1}\not=0$. As a result, we obtain the Poisson's ratio for PW-I mode as $\nu=1$ from equation~(\ref{eq_nu}). For the PW-II mode, we have $\Delta\theta_{1}\approx0$, and the force equilibrium condition leads to\cite{ChangT2003jmps} $db_{2}=\frac{1}{2}db_{1}\cos\frac{\theta_{1}}{2}$. Hence, the Poisson's ratio for the PW-II mode is $\nu=-1/3$ from equation~(\ref{eq_nu}). It is interesting to note that the Poisson's ratio ($\nu=-1/3$) for the PW-II mode coincides with the expectations of the self consistent screening approximation.\cite{LosJH2006prl}
Figure~\ref{fig_bondangle_energy}~(a) illustrates the interplay between the PW-I and PW-II deformation modes during the tensile deformation of graphene. Specifically, it shows the absolute value of the relative variation of the angles $\theta_{1}$ and $\theta_{2}$, and bond lengths $b_{1}$ and $b_{2}$; we note that the change in angle $\theta_{1}$ is negative in stretched graphene. For $\epsilon_x<0.035$, the variations of angles $\theta_{1}$ and $\theta_{2}$ are larger than the variations of bonds $b_{1}$ and $b_{2}$, respectively, which indicates PW-I to be the dominant deformation mode for graphene subject to small uniaxial tensile strains. For $0.035<\epsilon_x<0.085$, the variation of angle $\theta_{1}$ becomes less than the variation of bond $b_{1}$, while the variation of angle $\theta_{2}$ is still larger than the variation of bond $b_{2}$, which implies a competition between the PW-I and PW-II deformation modes for moderate strains. For $\epsilon_x>0.085$, variations for both bonds are larger than the variations of angles, so PW-II overcomes PW-I to be the dominant deformation mode for large tensile strains. Hence, the value of the Poisson's ratio will decrease with increasing strain, and will become negative at some critical strain between $[0.035, 0.085]$ when PW-II dominates the deformation process. The critical strain of 0.06 for the NPR in Figure~\ref{fig_poisson} falls in this strain range.
We perform an energy-based analysis, shown in Figure~\ref{fig_bondangle_energy}~(b), of the PW-I and PW-II deformation modes to gain further insight into the interplay between these two deformation modes governing the transition from positive to negative Poisson's ratio at $\epsilon_{x}=0.06$. The energy curve is computed as follows. For PW-I, the structure is manually deformed corresponding to the PW-I mode shown in the top of Figure~\ref{fig_pathway}. We then calculate the potential energy of this deformed structure, which is higher than the potential energy of undeformed graphene. The energy curve shown in Figure~\ref{fig_bondangle_energy}~(b) is the difference between the potential energy per atom of the deformed and undeformed graphene structures. The energy curve for PW-II is computed similarly, where angular distortions are allowed while the bond lengths are kept constant. Figure~\ref{fig_bondangle_energy}~(b) clearly shows a crossover around $\epsilon_x=0.06$ between the energy curves for PW-I and PW-II modes.
\textit{We thus propose a pathway energy based criteria: the tensile deformation process for graphene is governed by the deformation mode with lower pathway energy.} According to this criteria, PW-I mode will be the major deformation mode for $\epsilon_x<0.06$, in which the pathway energy for PW-I mode is lower than the pathway energy for PW-II mode. Similarly, the pathway energy criteria predicts the PW-II to be the major deformation mode for $\epsilon_x>0.06$, in which PW-II has lower pathway energy. We showed in Figure~\ref{fig_pathway} that the PW-I mode has a positive Poisson's ratio, while the PW-II mode has a NPR. As a consequence, the Poisson's ratio is positive for $\epsilon_x<0.06$, and will turn negative for $\epsilon_x>0.06$. This prediction is in excellent agreement with the numerical results in Figure~\ref{fig_poisson}, where the Poisson's ratio changes from positive to negative at $\epsilon_x=0.06$. We note that, to our knowledge, it is the first time the pathway energy criteria is proposed, which is based on the energetic competition for the two major in-plane deformation pathways. This criteria may be useful for future investigations into the mechanical properties of nano-materials similar as graphene.
\textbf{Conclusion.} In conclusion, we performed both MD and MS simulations to investigate the Poisson's ratio of graphene in the strain range 0 $<$ $\epsilon$ $<$ 0.15. We observed an intrinsic NPR for tensile strains exceeding $\epsilon_{x}=0.06$, which is independent of graphene's size and temperature. The appearance of the NPR is a direct result of the interplay between the PW-I (with positive Poisson's ratio) and PW-II (with NPR) modes during the tensile deformation of graphene. These two deformation modes correspond to the fundamental bond stretching and angle bending interactions, which are the two major in-plane interaction components in graphene. These results were further validated through a pathway energy criteria to predict positive or negative Poisson's ratio in graphene. Using this simple model, we found that the pathway energy for the PW-II deformation mode becomes lower than the PW-I mode for graphene tensile by strain above 0.06, leading to NPR above this strain range.
\textbf{Acknowledgements} The work is supported by the Recruitment Program of Global Youth Experts of China, the National Natural Science Foundation of China (NSFC) under Grant Nos. 11504225, 11472163, 11425209, and the start-up funding from Shanghai University. HSP acknowledges the support of the Mechanical Engineering department at Boston University.
\textbf{Author contributions} J.W.J performed the calculations and discussed the results with T.C, X.G, and H.S.P. J.W.J and H.S.P co-wrote the paper. All authors comment on the paper.
\textbf{Competing financial interests} The authors declare no competing financial interests.
\providecommand{\latin}[1]{#1}
\providecommand*\mcitethebibliography{\thebibliography}
\csname @ifundefined\endcsname{endmcitethebibliography}
{\let\endmcitethebibliography\endthebibliography}{}
\begin{mcitethebibliography}{38}
\providecommand*\natexlab[1]{#1}
\providecommand*\mciteSetBstSublistMode[1]{}
\providecommand*\mciteSetBstMaxWidthForm[2]{}
\providecommand*\mciteBstWouldAddEndPuncttrue
{\def\unskip.}{\unskip.}}
\providecommand*\mciteBstWouldAddEndPunctfalse
{\let\unskip.}\relax}
\providecommand*\mciteSetBstMidEndSepPunct[3]{}
\providecommand*\mciteSetBstSublistLabelBeginEnd[3]{}
\providecommand*\unskip.}{}
\mciteSetBstSublistMode{f}
\mciteSetBstMaxWidthForm{subitem}{(\alph{mcitesubitemcount})}
\mciteSetBstSublistLabelBeginEnd
{\mcitemaxwidthsubitemform\space}
{\relax}
{\relax}
\bibitem[Landau and Lifshitz(1995)Landau, and Lifshitz]{LandauLD}
Landau,~L.~D.; Lifshitz,~E.~M. \emph{Theory of Elasticity}; Pergamon,Oxford,
1995\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Lakes(1987)]{LakesR1987sci}
Lakes,~R. \emph{Science} \textbf{1987}, \emph{235}, 1038\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Rothenburg \latin{et~al.}(1991)Rothenburg, Berlint, and
Bathurst]{RothenburgL1991nat}
Rothenburg,~L.; Berlint,~A.~A.; Bathurst,~R.~J. \emph{Nature} \textbf{1991},
\emph{354}, 470\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Lakes(1993)]{LakesR1993adm}
Lakes,~R. \emph{Advanced Materials} \textbf{1993}, \emph{5}, 293--296\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Baughman and Galvao(1993)Baughman, and Galvao]{BaughmanRH1993nat}
Baughman,~R.~H.; Galvao,~D.~S. \emph{Nature} \textbf{1993}, \emph{365},
735\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Evans and Alderson(2000)Evans, and Alderson]{EvansKE2000adm}
Evans,~K.~E.; Alderson,~A. \emph{Advanced Materials} \textbf{2000}, \emph{12},
617\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Yang \latin{et~al.}(2004)Yang, Li, Shi, Xie, and Yang]{YangW2004jmsci}
Yang,~W.; Li,~Z.-M.; Shi,~W.; Xie,~B.-H.; Yang,~M.-B. \emph{Journal of
Materials Science} \textbf{2004}, \emph{39}, 3269--3279\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Ravirala \latin{et~al.}(2007)Ravirala, Alderson, and
Alderson]{RaviralaN2007jms}
Ravirala,~N.; Alderson,~A.; Alderson,~K.~L. \emph{Journal of Materials Science}
\textbf{2007}, \emph{42}, 7433--7445\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Lethbridge \latin{et~al.}(2010)Lethbridge, Walton, Marmier, Smith, and
Evans]{LethbridgeZAD2010am}
Lethbridge,~Z.~A.; Walton,~R.~I.; Marmier,~A.~S.; Smith,~C.~W.; Evans,~K.~E.
\emph{Acta Materialia} \textbf{2010}, \emph{58}, 6444--6451\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Bertoldi \latin{et~al.}(2010)Bertoldi, Reis, Willshaw, and
Mullin]{BertoldiK2010am}
Bertoldi,~K.; Reis,~P.~M.; Willshaw,~S.; Mullin,~T. \emph{Advanced Materials}
\textbf{2010}, \emph{22}, 361--366\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Greaves \latin{et~al.}(2011)Greaves, Greer, Lakes, and
Rouxel]{GreavesGN2011nm}
Greaves,~G.~N.; Greer,~A.~L.; Lakes,~R.~S.; Rouxel,~T. \emph{Nature Materials}
\textbf{2011}, \emph{10}, 823\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Alderson \latin{et~al.}(2012)Alderson, Alderson, Anand, Simkins,
Nazare, and Ravirala]{AldersonK2012pssb}
Alderson,~K.; Alderson,~A.; Anand,~S.; Simkins,~V.; Nazare,~S.; Ravirala,~N.
\emph{Phys. Status Solidi B} \textbf{2012}, \emph{249}, 1322--1329\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Clausen \latin{et~al.}(2015)Clausen, Wang, Jensen, Sigmund, and
Lewis]{clausenAM2015}
Clausen,~A.; Wang,~F.; Jensen,~J.~S.; Sigmund,~O.; Lewis,~J.~A. \emph{Advanced
Materials} \textbf{2015}, \emph{27}, 5523--5527\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Evans(1991)]{EvansKE1991Endeavour}
Evans,~K.~E. \emph{Endeavour} \textbf{1991}, \emph{15}, 170--174\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Milstein and Huang(1979)Milstein, and Huang]{MilsteinF1979prb}
Milstein,~F.; Huang,~K. \emph{Physical Review B} \textbf{1979}, \emph{19},
2030\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Baughman \latin{et~al.}(1998)Baughman, Shacklette, Zakhidov, and
Stafstrom]{BaughmanRH1998nat}
Baughman,~R.~H.; Shacklette,~J.~M.; Zakhidov,~A.~A.; Stafstrom,~S.
\emph{Nature} \textbf{1998}, \emph{392}, 362--365\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Ho \latin{et~al.}(2016)Ho, Ho, Kwon, and Kim]{hoPSSB2016b}
Ho,~V.~H.; Ho,~D.~T.; Kwon,~S.-Y.; Kim,~S.~Y. \emph{Phys. Status Solidi B}
\textbf{2016}, DOI:10.1002/pssb.201600017\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Jiang and Park(2014)Jiang, and Park]{JiangJW2014bpnpr}
Jiang,~J.-W.; Park,~H.~S. \emph{Nature Communications} \textbf{2014}, \emph{5},
4727\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Han \latin{et~al.}(2015)Han, Xie, Zhang, Yang, Si, and
Xue]{HanJ2015ape}
Han,~J.; Xie,~J.; Zhang,~Z.; Yang,~D.; Si,~M.; Xue,~D. \emph{Applied Physics
Express} \textbf{2015}, \emph{8}, 041801\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Yao \latin{et~al.}(2008)Yao, Alderson, and Alderson]{YaoYT2008pssb}
Yao,~Y.~T.; Alderson,~A.; Alderson,~K.~L. \emph{Phys. Status Solidi B}
\textbf{2008}, \emph{245}, 2373--2382\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Ho \latin{et~al.}(2014)Ho, Park, Kwon, Park, and Kim]{HoDT2014nc}
Ho,~D.~T.; Park,~S.-D.; Kwon,~S.-Y.; Park,~K.; Kim,~S.~Y. \emph{Nature
Communications} \textbf{2014}, \emph{5}, 3255\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Zakharchenko \latin{et~al.}(2009)Zakharchenko, Katsnelson, and
Fasolino]{ZakharchenkoKV}
Zakharchenko,~K.~V.; Katsnelson,~M.~I.; Fasolino,~A. \emph{Physical Review
Letters} \textbf{2009}, \emph{102}, 046808\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Ho \latin{et~al.}(2016)Ho, Ho, Kwon, and Kim]{hoPSSB2016a}
Ho,~V.~H.; Ho,~D.~T.; Kwon,~S.-Y.; Kim,~S.~Y. \emph{Phys. Status Solidi B}
\textbf{2016}, DOI:10.1002/pssb.201600061\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Grima \latin{et~al.}(2015)Grima, Winczewski, Mizzi, Grech, Cauchi,
Gatt, Attard, Wojciechowski, and Rybicki]{GrimaJN2015adm}
Grima,~J.~N.; Winczewski,~S.; Mizzi,~L.; Grech,~M.~C.; Cauchi,~R.; Gatt,~R.;
Attard,~D.; Wojciechowski,~K.~W.; Rybicki,~J. \emph{Advanced Materials}
\textbf{2015}, \emph{27}, 1455--1459\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Jiang and Park(2016)Jiang, and Park]{JiangJW2016npr_fbc}
Jiang,~J.-W.; Park,~H.~S. \emph{Nano Letters} \textbf{2016}, \emph{16},
2657--2662\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Brenner \latin{et~al.}(2002)Brenner, Shenderova, Harrison, Stuart, Ni,
and Sinnott]{brennerJPCM2002}
Brenner,~D.~W.; Shenderova,~O.~A.; Harrison,~J.~A.; Stuart,~S.~J.; Ni,~B.;
Sinnott,~S.~B. \emph{Journal of Physics: Condensed Matter} \textbf{2002},
\emph{14}, 783--802\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Mo \latin{et~al.}(2009)Mo, Turner, and Szlufarska]{MoY2009nat}
Mo,~Y.; Turner,~K.~T.; Szlufarska,~I. \emph{Nature} \textbf{2009}, \emph{457},
1116\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Jiang \latin{et~al.}(2008)Jiang, Tang, Wang, and Su]{JiangJW2008prb}
Jiang,~J.-W.; Tang,~H.; Wang,~B.-S.; Su,~Z.-B. \emph{Physical Review B}
\textbf{2008}, \emph{77}, 235421\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Plimpton(1995)]{PlimptonSJ}
Plimpton,~S.~J. \emph{Journal of Computational Physics} \textbf{1995},
\emph{117}, 1--19\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Stukowski(2010)]{ovito}
Stukowski,~A. \emph{Modelling and Simulation in Materials Science and
Engineering} \textbf{2010}, \emph{18}, 015012\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Lee \latin{et~al.}(2008)Lee, Wei, Kysar, and Hone]{LeeC2008sci}
Lee,~C.; Wei,~X.; Kysar,~J.~W.; Hone,~J. \emph{Science} \textbf{2008},
\emph{321}, 385\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Bao \latin{et~al.}(2009)Bao, Miao, Chen, Zhang, Jang, Dames, and
Lau]{BaoW2009nn}
Bao,~W.; Miao,~F.; Chen,~Z.; Zhang,~H.; Jang,~W.; Dames,~C.; Lau,~C.~N.
\emph{Nature Nanotechnology} \textbf{2009}, \emph{4}, 562--566\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Garza \latin{et~al.}(2014)Garza, Kievit, Schneider, and
Staufer]{GarzaHHP2014nl}
Garza,~H. H.~P.; Kievit,~E.~W.; Schneider,~G.~F.; Staufer,~U. \emph{Nano
Letters} \textbf{2014}, \emph{14}, 4107--4113\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Los \latin{et~al.}(2006)Los, Fasolino, and Katsnelson]{LosJH2006prl}
Los,~J.~H.; Fasolino,~A.; Katsnelson,~M.~I. \emph{Physical Review Letters}
\textbf{2006}, \emph{116}, 015901\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Chang and Gao(2003)Chang, and Gao]{ChangT2003jmps}
Chang,~T.; Gao,~H. \emph{Journal of the Mechanics and Physics of Solids}
\textbf{2003}, \emph{51}, 1059--1074\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Shen and Li(2004)Shen, and Li]{ShenL2004prb}
Shen,~L.; Li,~J. \emph{Physical Review B} \textbf{2004}, \emph{69},
045414\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\bibitem[Chang \latin{et~al.}(2005)Chang, Geng, and Guo]{ChangTC2005apl}
Chang,~T.; Geng,~J.; Guo,~X. \emph{Applied Physics Letters} \textbf{2005},
\emph{87}, 251929\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\unskip.}
\end{mcitethebibliography}
\begin{figure}[htpb]
\begin{center}
\scalebox{1.4}[1.4]{\includegraphics[width=8cm]{poisson.eps}}
\end{center}
\caption{(Color online) Negative Poisson's ratio in graphene. (a) The resultant strain $\epsilon_y$ versus the applied strain $\epsilon_x$. A robust valley point exists at $\epsilon_x=0.06$ in all curves for varying simulation parameters using both MD or MS approaches. (b) The Poisson's ratio extracted from the MS results in (a) through $\nu=-\partial \epsilon_y/\partial \epsilon_x$, which is negative for $\epsilon_x>0.06$.}
\label{fig_poisson}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\scalebox{1.8}[1.8]{\includegraphics[width=8cm]{pathway.eps}}
\end{center}
\caption{(Color online) Two typical ideal deformation pathways during the tensile deformation of graphene. The left atom cluster (black online) is stretched along the horizontal x-direction. The parallelogram gray area indicates the unit cell. PW-I (blue online): carbon-carbon bond lengths remain constant ($\Delta b=0$), while angles are altered to accommodate the external strain, which results in a Poisson's ratio of $\nu=1$. PW-II (red online): angles are unchanged and bond lengths are elongated to accommodate the external tension, resulting in a NPR of $\nu=-1/3$. The lighter shades show the undeformed structure.}
\label{fig_pathway}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\scalebox{1.5}[1.5]{\includegraphics[width=8cm]{bondangle_energy.eps}}
\end{center}
\caption{(Color online) Pathway energy criteria for PW-I and PW-II deformation modes. (a) The variation of key geometrical parameters (angles $\theta$ and bond lengths $b$) in graphene. The y-axis shows the relative variation, i.e., $\Delta b/b_0$ or $\Delta \theta/\theta_0$. (b) Pathway energy curve for PW-I and PW-II deformation modes. The curves show a crossover at $\epsilon_x=0.06$, which predicts a transition from PW-I mode (positive Poisson's ratio) to PW-II mode (negative Poisson's ratio) during the tensile deformation of graphene. Left bottom inset (black online) shows the undeformed structure. Top inset (blue online) displays the PW-I deformed structure. Right inset (red online) is the PW-II deformed structure.}
\label{fig_bondangle_energy}
\end{figure}
\end{document}
|
\section{Introduction}
Observations by the {\it Fermi Gamma-ray Space Telescope}, with its unprecedented energy coverage, have led to an exciting and productive period for gamma-ray burst (GRB) astronomy. The {\it Fermi} Gamma-ray Burst Monitor (GBM) has detected over 1500 GRBs in 5 yr of operations, with over 90 of these bursts detected by the {\em Fermi}~Large Area Telescope (LAT) above 40 MeV\footnote{\url{http://fermi.gsfc.nasa.gov/ssc/observations/types/grbs/lat_grbs/}}. The high-energy emission observed by the LAT is typically longer lasting and delayed in onset in comparison to emission at keV energies detected by the GBM.
The origin of the high-energy emission observed by the LAT has been much debated within the GRB community. Emission detected contemporaneous by the GBM and LAT during the prompt phase of several bursts (e.g. GRB~090217, GRB~090323, and GRB~130427A) suggests that some of the emission at MeV energies observed by the LAT may originate from a simple extension of the prompt GRB spectra into the LAT energy range \citep{Ackermann2013}. At the same time, the delayed onset and long-lived nature of the LAT emission has led to speculation \citep{Kumar2009, Ghisellini2010, DePasquale2010, Razzaque2010} that the afterglow components commonly observed at optical and X-ray wavelengths may also produce a significant amount of gamma-ray emission from the high-energy extension of the synchrotron spectrum of the external forward shock. Broadband fits to the simultaneous multiwavelength observations of GRB~110731A \citep{GRB110731A} and GRB~130427A \citep{GRB130427A_LAT}, which were detected by both {\em Swift}~and {\em Fermi}, show remarkably similar late-time spectral and temporal behavior, lending support to an external shock origin of their late-time GeV emission. In the case of GRB 130427A, the GeV emission persisted for almost a day after the burst and matched the temporal decay slope observed in the X-ray and optical wavelengths. Moreover, spectral and temporal {\it NuSTAR} observations taken 1.5 and 5 days after the event onset, combined with {\it Swift, Fermi}, and ground-based optical data, unambiguously establish a single afterglow spectral component from optical to multi-GeV energies, most certainly due to synchrotron radiation \citep{GRB130427A_NuSTAR}. Therefore, there is now a growing body of evidence suggesting that the high-energy emission observed by the LAT is likely due to a combination of both prompt and afterglow contributions in the LAT energy window.
Despite this converging picture as to the origin of the LAT-detected emission, there still remains a large number of bright bursts for which no high-energy emission was observed by the LAT. Of the over 1500 bursts detected by the GBM in the 8 keV to 40 MeV energy range in the first 5 yr of operation, only 8$\%$ of the bursts that have occurred within the LAT field of view (FOV) were detected above 40 MeV \citep{Ackermann2013}. An examination by \citet{UpperLimitsCatalog} of the prompt emission from bright GBM-detected bursts that fell within the LAT FOV, but were not detected by the instrument, showed that many of these bursts either require spectral breaks or have intrinsically steeper prompt spectra than inferred from fitting the GBM data alone in order to explain their nondetections by the LAT. Likewise, \citet{GRB130427A_NuSTAR} showed that multiwavelength data collected for GRB~130427A were best fit by a smoothly broken power law 1.5 days (and possibly up to 5 days) after the burst. These observations indicate that spectral breaks may persist between the {\it Swift's} X-ray Telescope (XRT) and LAT windows for thousands of seconds after the prompt emission.
In this paper, we present a stacking analysis of LAT data of {\it Swift}-detected GRBs that fell within the LAT FOV at the time of trigger, in order to search for subthreshold emission at MeV and GeV energies. By using well-localized {\em Swift}~bursts, we can eliminate any ambiguity that arises from the positional uncertainty of the burst in the LAT FOV. We examine a total of 79 GRBs by comparing the observed counts over a range of time intervals to that expected from designated background orbits, as well as by using a joint likelihood method to model the expected distribution of counts, in order to search for subthreshold emission. We find a significant excess above the combined background during time intervals including and exceeding the durations characterizing the prompt emission observed at keV energies using both methods. This analysis follows a similar study by \citet{Lange13}, who performed a counting analysis of GBM-detected bursts with no detectable emission in the LAT above 100 MeV. In their study, the authors found similar evidence for extended subthreshold emission. Here we use the joint likelihood analysis to obtain a robust estimate of the flux and spectral properties of this emission.
The paper is structured as follows: In \S\ref{sec:InstrumentOverview}, we review the characteristics of the {\em Fermi}~GBM, {\em Fermi}~LAT, {\em Swift}~BAT, and {\em Swift}~XRT instruments. In \S\ref{sec:SampleDefinition}, we define the GRB samples considered in this work, and we outline the analysis performed in \S\ref{sec:Analysis}. We present the results in \S\ref{sec:Results} and discuss the implications of our results in \S\ref{sec:Discussion}.
\section{Instrument Overview } \label{sec:InstrumentOverview}
\subsection{{\it Fermi~LAT and Fermi~GBM} } \label{sec:FermiInstrumentOverview}
The {\it Fermi Gamma-ray Space Telescope} consists of two primary instruments, the GBM and the LAT. The GBM has 14 scintillation detectors that together view the entire unocculted sky. Triggering and localization are performed using 12 sodium iodide (NaI) detectors with different orientations placed around the spacecraft in four clusters of three. Two bismuth germanate (BGO) scintillators are placed on opposite sides of the spacecraft so that at least one detector can see any triggered event. GBM spectroscopy uses both the NaI and BGO detectors, sensitive between 8 keV and 1 MeV and between 150 keV and 40 MeV, respectively. Together they provide nearly 4 decades of energy for unprecedented sensitive spectroscopic studies of GRBs \citep{Meegan:09}.
The LAT is a pair conversion telescope comprising a $4\times4$ array of silicon strip trackers and cesium iodide (CsI) calorimeters covered by a segmented anti-coincidence detector to reject charged-particle background events. The LAT covers the energy range from 20\,MeV to more than 300\,GeV with an FOV of $\sim 2.4$ sr. The deadtime of the LAT is nominally 26\,$\mu$s, which is crucial for observations of high-intensity transient events such as GRBs. The LAT triggers on many more background events than celestial gamma-rays. Onboard background rejection is supplemented on the ground using event class selections that accommodate the broad range of sources of interest \citep{Atwood:09}.
\subsection{{\it Swift~BAT and Swift~XRT}} \label{sec:SwiftInstrumentOverview}
The {\em Swift}~spacecraft consists of the Burst Alert Telescope (BAT; \citealt{Barthelmy05}), (XRT) \citep{Roming05}, and Optical and Ultraviolet Telescope (UVOT; \citealt{Roming05}). The BAT is a wide-field, solid-state gamma-ray detector, covering an FOV of 1.4 sr and an imaging energy range of 15--150 keV. The instrument's coded mask allows for positional accuracy of 1$'$ to 4$'$ within seconds of the burst trigger. The XRT is a focusing X-ray telescope covering an energy range from 0.2--10 keV and providing a typical localization accuracy of $\sim3"$. The UVOT is a clear-aperture Ritchey-Chretien telescope that provides optical and ultra-violet photometry and subarcsecond positional accuracy of the long-lived afterglow counterparts to the prompt emission from GRBs.
\newcommand{\mbox{$T_0$}}{\mbox{$T_0$}}
\begin{figure}
\includegraphics[width=1\columnwidth]{SwiftBAT_T90.pdf}
\caption{The $T_{90}$ duration distribution of the subsample of bursts for which prompt Swift BAT observations were available. The median duration is 48.9 s, with most bursts having a duration below 100 s}
\label{Fig:T90Distribution}
\end{figure}
\section{Sample Definition} \label{sec:SampleDefinition}
We compiled a sample of all GRBs observed by {\em Swift}~between the beginning of {\em Fermi}~science operations on 2008 August 4 and 2012 February 1, reported by \citet{Donato2012}, yielding a total of 369 GRBs. Of these, 121 bursts fell within 65$^\circ$ of the LAT z-axis (or boresight), which we define as the LAT FOV, at the time of trigger. The sensitivity of the LAT falls as a function of off-axis angles; therefore, bursts detected at angles greater than $65^\circ$ were not considered for this analysis. Of these 121 bursts, we excluded 16 GRBs that were detected by the LAT above 40 MeV and an additional 25 GRBs that occurred during spacecraft passages through the South Atlantic Anomaly (SAA) or at angles with respect to the Earth's zenith that were $\ga 105^\circ$, placing the burst near the Earth's horizon. Observations of such high zenith angles result in emission at the burst location that is dominated by gamma-rays from the Earth's limb produced by interactions of cosmic rays with the Earth's atmosphere. The remaining sample includes 79 GRBs.
Of the 79 GRBs in this sample, 68 were originally detected by BAT. The remaining 11 bursts were detected by other spacecraft and triggered subsequent follow-up observations by XRT. Of the total sample, 13 GRBs obtained their best localization through BAT detections, 28 through XRT detections, 18 through UVOT detections, 16 through ground-based followup observations, and 4 from other missions, yielding a median positional uncertainty of 120 arcsec, 1.6$''$, 0.$''$6, and 0.$''$5 respectively \citep{Donato2012}. All of these positional uncertainties are far smaller than the LAT point spread function (PSF), which has a $68\%$ containment (i.e., the radius of the circle containing 95$\%$ of the PSF) that varies from $5^\circ$ at 100 MeV to 0.1$^{\circ}$ at 100 GeV with the `P7REP\_SOURCE\_V15' instrument response functions. Finally, eight of these GRBs were detected by GBM and triggered an autonomous repoint request (ARR) of the {\em Fermi}~spacecraft and another six bursts occurred while the spacecraft was performing a target of opportunity (ToO) observation. Of the bursts with prompt BAT observations, the median $T_{90}$ duration is 48.9 s. The $T_{90}$ distribution of this subsample of bursts is shown in Figure \ref{Fig:T90Distribution} for reference.
The complete list of the bursts in our final sample, their positions, and positional uncertainty is given in Table 1.
\section{Analysis} \label{sec:Analysis}
\subsection{Counting Analysis Overview} \label{sec:CountingAnalysis}
The analysis presented here focuses on two distinct techniques with which to search for a signal in stacked data. The first method takes the sum of the observed counts collected over a specific duration, energy range, and region of interest (ROI) on the sky centered on each burst's best known position and compares it to the counts collected over an interval deemed to adequately represent the expected background at the time of trigger. The significance of the stacked counts compared to the stacked background is then estimated through Gaussian statistics.
The size of the ROI is typically chosen to reflect the 95\% containment radius of the LAT energy-dependent PSF at a particular energy. For this analysis, we examine both a fixed 10$^\circ$ and 12$^\circ$ energy-independent ROI for each burst, as well as an energy-dependent ROI with a size ranging from 12$^\circ$ at 100 MeV to 2.2$^\circ$ at 10 GeV. The size of the energy-dependent ROI reflects the 95$\%$ containment of the LAT PSF when considering the P7REP\_SOURCE\_V15 instrument response functions. From these ROIs, we select `Source' class events that occurred within the first 2700 s after each GRB's trigger time ($T_{0}$) with an energy between 75 MeV and 30 GeV. The ``Source" data class was specifically optimized for the study of point-like sources, with stricter cuts against non-photon-background contamination in comparison to the ``Transient" data class that is typically used to study GRBs on very short timescales \citep{LATPerformancePaper}. We also excluded events with an estimated angle from the Earth's zenith greater than 105$^\circ$ to guard against possible contamination from photons that originate from the Earth's limb. The selection of 75 MeV as the minimum energy is motivated by the possibility of detecting emission from bursts with significant attenuation emission below 100 MeV, (e.g.,~\citealt{Ackermann2013}). The choice of 2700 s reflects the longest amount of time that most sources on the sky can continuously stay in the LAT FOV before being occulted by the Earth (the exception being those located near the north orbital pole).
Each of our ROI selections has its own set of advantages and disadvantages. The energy-dependent ROI has the advantage of collecting photons within a radius that is commensurate with the LAT 95\% containment radius at a given energy, potentially reducing background contamination at high energies, where the LAT PSF is narrowest. The 12$^\circ$ energy-independent ROI, on the other hand, makes no assumptions regarding the exact shape of the LAT PSF, at the cost of collecting more background at higher energies. The 10$^\circ$ energy-independent ROI is an intermediate solution, encompassing less of the PSF tail at low energies, but still collecting higher levels of background at high energies compared to the energy-dependent ROI. The use of a 10$^\circ$ ROI also reflects the better PSF of the \emph{Source} class selection used in this analysis, compared to the \emph{Transient} class selection, for which a 12$^\circ$ ROI is typically used.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=5.35in]{OrbitalComparison_GRB110107A.eps}
\end{center}
\caption{A comparison of the orbits from which the signal and background counts are extracted. The blue diamonds correspond to the location of the spacecraft at the time of the burst trigger and its location exactly 30 sidereal orbits prior. The color of the orbits represents the angle from the burst position on the sky to the instrument boresight.}
\label{Fig:OrbitalComparison}
\end{figure*}
\subsection{Joint Likelihood Analysis Overview} \label{sec:CompositeLikelihoodAnalysis}
The second technique we employ to search for a subthreshold signal in the stacked LAT data consists of a joint likelihood analysis using the analysis tools developed by the LAT team (ScienceTools version v9r30p1).\footnote{{\tt http://fermi.gsfc.nasa.gov/ssc/}} In standard unbinned likelihood fitting of individual sources, the expected distribution of counts for each burst is modeled as a point source using an energy-dependent LAT PSF and a power-law source spectrum with a normalization and photon index that are left to vary as free parameters. Galactic and isotropic background components are also included in this model. The Galactic component, \emph{gll\_iem\_v05}, is a spatial and spectral template that accounts for interstellar diffuse gamma-ray emission from the Milky Way. The normalization of the Galactic component is kept fixed during the fit. The isotropic component, \emph{iso\_source\_v05}, provides a spectral template to account for all remaining isotropic emission that is not represented in the Galactic diffuse component and therefore accounts for contributions from both residual charged-particle backgrounds and the isotropic time-averaged celestial gamma-ray emission. The normalization of the isotropic component is allowed to vary during the fit. Both the Galactic and isotropic templates are publicly available\footnote{http://fermi.gsfc.nasa.gov/ssc/data/access/lat/BackgroundModels.html}.
We can then derive the probability density of the observed data given this model and create a ``likelihood" function by treating this probability density as a function of the model parameters. The likelihood function is essentially the probability of observing the data given the chosen model for a range of model parameters. By maximizing the likelihood function with respect to our parameters of interest, we can then estimate the model parameters that make the observed data the most probable.
The maximum likelihood technique can be expanded to apply to multiple sources to constrain or estimate a set of parameters thought to be common to those sources. In this case, the backgrounds and properties of each source can be modeled individually in source-specific likelihoods, as described above. The source-specific likelihoods can then be multiplied to yield a joint likelihood function that is used for inference on the common parameter of interest. In this way, the joint likelihood becomes a source stacking technique, which has been previously applied to LAT data in the search for dark matter (Ackermann et al. 2011) and to study the extragalactic background light (Ackermann et al. 2012b). This method is implemented in the {\em Fermi}~Science Tools as the \emph{Composite2} routine.
To quantify the significance of a potential excess above background, we employ the common likelihood-ratio test \citep{Neyman1928}. For this test, we form a test statistic (TS) that is the ratio of the likelihood evaluated at the best-fit parameters under a background-only, null hypothesis, i.e. a model that does not include a point-source component, to the likelihood evaluated at the best-fit model parameters when including a candidate point source at the center of the ROI. According to Wilks's theorem, this ratio is distributed approximately as $\chi^{2}$, so we choose to reject the null hypothesis when the TS is greater than TS $= 25$, roughly equivalent to a $5 \sigma$ rejection criterion.
The data selection for this method is identical to that performed for the counting analysis in the case of the 10$^\circ$ and 12$^\circ$ ROIs described in \S~\ref{sec:CountingAnalysis}. We note that the {\em Fermi}~Science Tools do not currently include a means of performing a likelihood analysis on data selected using an energy-dependent ROI, as is possible with the counting analysis.
\subsection{Background Selection} \label{sec:BackgroundSelection}
The proper selection of background regions with which to compare the stacked signal is a crucial component to the counting analysis and provides a control sample for the joint likelihood analysis. Since the spacecraft's geomagnetic coordinates and its celestial pointing can vary significantly over the 2700 s duration under consideration, the use of an off-and-on source method of background determination may not always adequately represent the true background. In addition, the use of intervals immediately prior to the burst trigger or after the end of the prompt emission, as observed by GBM or BAT, to estimate a background interval could serve to bias our investigation, as LAT-detected GRBs have exhibited both delayed and extended high-energy emission on timescales that exceed the durations traditionally defined by observations in the keV$-$MeV energy range \citep{Ackermann2013} and subthreshold emission prior to the burst trigger cannot be ruled out. Therefore, we avoid using intervals immediately preceding or succeeding the burst activity for background estimation.
Instead, we attempt to locate an interval at least one orbit prior to the GRB trigger that best matches the spacecraft's observing conditions at the time of the trigger. This includes matching the same off-axis angle between the GRB sky coordinates and the LAT boresight, the spacecraft's geomagnetic coordinates in orbit, and the angle of the GRB location to the Earth's zenith at the time of trigger. These three criteria serve to match the charged-particle background, the Galactic and isotropic backgrounds, and possible contamination from Earth limb photons, respectively, to that observed by the LAT at the time of the GRB trigger. A period of 30 sidereal orbits (171,915 s) prior to trigger adequately matches these geomagnetic and pointing criteria and provides an interval with which to estimate the background during the GRB. An example of the spacecraft's orbit and orientation with respect to the position of a GRB in our sample and its associated background orbit can be seen in Figure \ref{Fig:OrbitalComparison}.
We note that this selection does not adequately represent the expected background for the 15 GRBs that either triggered ARRs of the spacecraft or occurred while the spacecraft was performing a ToO, since maneuvers take the spacecraft out of survey mode and initiate a custom pointing that keeps the GRB in the LAT FOV. As a result, there is no previous interval that matches the spacecraft's orbit and orientation during the repoint, and as such, these bursts are excluded from the counting analysis. Since the background for each individual GRB is being modeled in the likelihood analysis, this method can adequately take into account the spacecraft motion for these bursts. This leaves a total of 64 bursts for which we can apply the counting analysis. Therefore, the application of the likelihood technique to the bursts for which we have the most comprehensive observations is a significant advantage of this method over the counting analysis.
In order to validate the effectiveness of this background selection for bursts that did not trigger an ARR or occurred during a ToO, we extract the observed photons over a 2700 s interval from 1000 random locations on the sky covering an energy range from 75 MeV to 300 GeV, using a 10$^\circ$, 12$^\circ$, and energy-dependent ROI. We then extract the observed photons at the same location, but 30 sidereal orbits prior to the original selection interval. We can then test whether the observed photons during the signal and background intervals are consistent with being drawn from the same distribution by examining the resulting significance distribution. We define the counting significance as $(S-B)/\sqrt{S+B}$, where $S$ and $B$ are the counts in the signal and background intervals, respectively \citep{LiMa83}.
The resulting distribution for the 12$^\circ$ ROI has a mean value of $\mu_{12} = 0.09$ and variance of $\sigma_{12} = 1.01$. Among our 1000 trials, we obtain three false positives above $3\sigma$, consistent with expectations, as we expect 99.7$\%$ of values drawn from our significance distribution to lie within $3\sigma$ of the mean. The results for the 10$^\circ$ and energy-dependent ROIs were consistent with these values. Therefore, our background selection is found to be robust for any single source and background interval selection.
\section{Results}
\label{sec:Results}
\subsection{Counting Analysis} \label{sec:ResultsCountingAnalysis}
\begin{figure}[t]
\includegraphics[width=1\columnwidth]{StackedLightCurve.pdf}
\caption{Top panel: a stacked light curve of signal (black solid line) and background (red dotted line) counts for a 10$^\circ$ ROI selection, binned to 100 s intervals. Middle panel: The significance of the signal compared to the background. Bottom panel: comparison of the total exposure during the signal and background orbits.}
\label{Fig:StackedLightCurve-ARRs}
\end{figure}
Using the data selection criteria described in Section~\ref{sec:CountingAnalysis}, we can directly compare the extracted signal to the counts accumulated using the same selection criteria, but collected during the background orbits described in Section~\ref{sec:BackgroundSelection}. When considering the entire interval from $T_{0}$ to 2700 s and an energy range from 75 MeV to 30 GeV, we obtain a total accumulated signal and background of $S_{10} = 1711$ and $B_{10} = 1476$, $S_{12} = 2365$ and $B_{12} = 2161$, and $S_{\rm EROI} = 750$ and $B_{\rm EROI} = 626$ counts for a significance of $4.16 \sigma$, $3.03 \sigma$, and $3.34 \sigma$ for the 10$^\circ$, 12$^\circ$, and energy-dependent ROIs, respectively.
Introducing cuts on the data, as we have here with the extraction radius, in order to maximize the observed significance introduces a bias in the analysis due to the so-called look-elsewhere effect. The chance that the observed significance could have arisen at random owing to the size of the parameter space that was searched can be accounted for by applying trial factor corrections to the final significance. Here we searched the data using three different extraction radii, and therefore the final significance needs to be attenuated as $1-(1-\alpha$)$^N$, where $\alpha$ is the probability of observing such a value by chance\footnote{Recall that for a normal distribution, a $3\sigma$ detection has a probability of chance occurrence of $\alpha = 1-99.7\% = 0.003$} of the detection and $N$ represents the number of trials. For the analysis presented above, our $4.16 \sigma$ detection therefore becomes an erf($\frac{4.16}{\sqrt{2}})^{3}$ $\sim 3.9 \sigma$ detection. We note that this correction factor assumes that each data set is statistically independent, which is overly conservative in our case, since the three different ROI selections result in data sets that are subsets of each other. Therefore, we regard the $3.9 \sigma$ detection significance as a conservative lower limit to the true significance of the signal over the background.
Focusing on the 10$^\circ$ ROI analysis, we can create a stacked light curve that is co-aligned to the trigger time of each GRB. This light curve of stacked signal and background counts, binned to 100 s intervals, is shown in the top panel of Figure \ref{Fig:StackedLightCurve-ARRs}, with the middle panel showing the Gaussian significance of their difference and the bottom panel showing the summed LAT effective area during the observations. The overall shape of the light curve reflects the evolution in the total effective area. The observed count rate rises as additional fields containing GRBs enter the LAT FOV and falls as they exit or as the spacecraft enters the (SAA), whereby data taking is disabled. The total effective area, and hence the stacked counts light curve, peaks near the co-aligned trigger, as all fields are predicated to be in the LAT FOV at this time owing to our sample selection criteria. The stacked light curve shows a $\sim3\sigma$ excess roughly 30 s after the co-aligned trigger, followed by additional periods of excess signal over background, although a consideration of trial factors lessens the significance of any one peak in such a time-resolved analysis.
\begin{figure}[t]
\includegraphics[width=1\columnwidth]{CumulativeLightCurve-ARRs_LiMa1983.pdf}
\caption{Top panel: cumulative signal (black solid line) and background (red dotted line) counts as a function of time. Bottom panel: significance of the cumulative signal compared to the cumulative background for a 10$^\circ$ ROI selection.}
\label{Fig:CumulativeLightCurve-ARRs}
\end{figure}
The median ratio of the exposure during the signal and background orbit is found to be $\mu = 1.05$, with a variance of $\sigma = 0.072$, highlighting that the exposures of the signal and background orbits are well matched, but not exactly equal. Small mismatches in exposure such as these may be a fundamental limitation of the stacked counting analysis performed here. Such differences are small for the source/background comparison for any one burst, but when adding 79 such comparisons, these small mismatches contribute to a non-negligible difference in the total summed exposure, which can be seen in the bottom panel of Figure \ref{Fig:StackedLightCurve-ARRs}. Ultimately, normalizing the observed counts by the estimated exposure for each burst and performing a summed rate comparison can account for these differences.
Next, we examine the difference between the cumulative signal and background over much wider time intervals. In the top panel of Figure \ref{Fig:CumulativeLightCurve-ARRs}, we show the cumulative signal and background counts as a function of time since $T_{0}$ - 1000 s for a 10$^\circ$ ROI selection, with the bottom panel again showing the significance of their difference. The effective area of the stacked observations drops to zero above $T_{0}$ + 2700 s as all fields exit the LAT FOV, and as such, the cumulative light curve levels off accordingly. The excesses seen in Figure \ref{Fig:StackedLightCurve-ARRs} can be clearly traced as local maxima in Figure \ref{Fig:CumulativeLightCurve-ARRs}. Although the difference between the signal and background varies above and below zero significance prior to the co-aligned trigger, the signal clearly begins to diverge from the background at $T_{0}$, above which the significance climbs to approximately $\sim3 \sigma$ at $T_{0}$ + 2700 s. We note that this cumulative significance differs from that quoted above because the integration period here begins at $T_{0}$ - 1000 s.
We also examine the dependence of the signal significance on the minimum energy $E_{\rm min}$ used in our selection criteria. In Figure~\ref{Fig:SignificancevsEnergy}, we plot the resulting signal significance as a function of $E_{\rm min}$ when considering the interval from $T_{0}$ to 2700 s for a 10$^\circ$ ROI selection, showing that the detection significance does not improve when only considering higher-energy photons. The dashed line, representing the number of bursts with photons contributing to the modified selection criteria, falls steadily with increasing $E_{\rm min}$.
Finally, an identical analysis of the preburst observations, covering an interval from $T_{0}$ - 2700 s to the co-aligned trigger, reveals no excess emission above the background. For this control sample, using a 10$^\circ$ ROI selection, we obtain a total accumulated signal and background of $S = 1704$ and $B = 1681$ counts, respectively, for a significance of $0.40 \sigma$.
\begin{figure}[t]
\includegraphics[width=1\columnwidth]{SignificancevsEnergy_10ROI_LiMa1983.pdf}
\caption{Counting analysis: the dependence of the signal significance (solid line) and the sample size (dashed line) on the minimum selection energy $E_{\rm min}$ for a 10$^\circ$ ROI.}
\label{Fig:SignificancevsEnergy}
\end{figure}
\subsection{Joint Likelihood Analysis} \label{sec:JointLikelihoodAnalysis}
Using the data selection criteria described in Section~\ref{sec:CountingAnalysis} and the analysis method outlined in Section~\ref{sec:CompositeLikelihoodAnalysis}, we first performed a joint likelihood analysis on the ensemble of GRB locations, including those that triggered ARRs of the spacecraft or occurred during ToOs, for an integration period covering the entire 2700 s post trigger and an energy range from 75 MeV to 30 GeV, using both a 10$^\circ$ and 12$^\circ$ ROI. The resulting TS of the joint likelihood fit is 59 and 58.6, respectively, roughly representing a $\sim7 \sigma$ detection significance of a point source being present in excess to the expected background using either ROI size. The TS of an identical analysis performed on the background orbits defined in Section~\ref{sec:BackgroundSelection} is consistent with zero.
Focusing on the 10$^\circ$ ROI analysis, we examine the dependence of the TS value on the minimum energy $E_{\rm min}$ used in our selection criteria. In Figure~\ref{Fig:TSvsEnergy}, we plot the resulting TS for a range of $E_{\rm min}$ values, showing that the detection significance rises as low-energy photons are excluded from the fit, peaking at $E_{\rm min} = 300$ MeV, before eventually leveling out above 400 MeV. The dashed line represents the number of analysis intervals for which we observed photons matching the selection criteria. As $E_{\rm min}$ rises, the number of intervals with photons contributing to the joint likelihood analysis falls, with nearly one-third of the burst positions being removed from the sample when $E_{\rm min} = 1000$ MeV. Because the detection significance peaks at $E_{\rm min} = 300$ MeV, we will focus the remainder of our joint likelihood analysis on an energy selection criterion covering 300 MeV to 30 GeV.
In order to investigate the time dependence of the excess signal, we also performed a joint likelihood analysis over a range of integration times before and after the trigger time, going from from $T_{0}$ - 1000 s to $T_{0}$ and from $T_{0}$ to $T_{0}$ + 2700 s, in 50 s intervals.. The TS as a function of integration time is shown in Figure~\ref{Fig:MaxTSVsTime}. The red dotted line represents the TS of an identical analysis performed on each burst's associated background orbits. The TS of the joint likelihood fit to the data collected over the 1000 s prior to the co-aligned trigger is consistent with zero. The TS rises slightly for shorter integration duration approaching $T_{0}$, but is below TS = 9, roughly equivalent to $3\sigma$, for all integration durations prior to $T_{0}$, except for the 50 second interval covering $T_{0}$ - 50 to $T_{0}$, at which point it approaches TS $\sim10$. At all subsequent integration times after $T_{0}$, the significance of the signal excess above background is greater than $3\sigma$, reaching $5\sigma$ within 200 s after $T_{0}$. The TS values exhibit local maxima as a function of time, reflecting the arrival of photons in excess to our background model, before leveling off at TS $= 59$ as the accumulated effective area drops to zero. The data collected from the same fields during the background intervals show no such excess when fit to our background models, being consistent with TS $\sim$ 0 for roughly all integration periods under consideration.
\begin{figure}[t]
\includegraphics[width=1\columnwidth]{TSvsEnergy.pdf}
\caption{Likelihood analysis: the dependence of the likelihood TS (solid line) and the sample size (dashed line) on the minimum selection energy $E_{\rm min}$ for a 10$^\circ$ ROI..}
\label{Fig:TSvsEnergy}
\end{figure}
Before we can assess the final significance of our likelihood analysis, we again need to take trial factors into account. Since we considered two different ROI sizes and 10 different minimum energy boundaries, this introduces $N = 12$ trials. Therefore, a TS $= 59$, or $\sqrt{59} \sim 7.68\sigma$, detection becomes an erf($\frac{7.68}{\sqrt{2}})^{12}$ $\sim 7.46 \sigma$ detection. Again, since the ROI and minimum boundary selections did not result in statistically independent data sets, we regard this attenuated significance as a conservative lower limit to the true significance of the detection.
By scanning the joint likelihood function over a range of photon indices and flux normalizations for the point source in our model, we can obtain a joint likelihood profile that is a function of these two common parameters of interest. Finding the maximum of this profile allows us to estimate the stacked source flux and characteristic photon index of the sample. A contour plot of the resulting joint likelihood profile for the analysis covering the entire 2700 s interval post trigger and an energy range from 300 MeV to 30 GeV is shown in Figure \ref{Fig:LikelihoodContourPlot}. The best-fit photon flux and photon index of the combined data is $F_{\rm ph} = 6.4\times10^{-8}$ photons cm$^{-2}$ s$^{-1}$ and $\Gamma = -1.92$, respectively. The 68$\%$, 95$\%$, and 99$\%$ confidence levels (C.L.) for the photon flux estimate are shown as the solid, dashed, and dash-dotted lines, respectively.
For comparison, a typical photon flux upper limit for an individual well-observed GRB that remained in the LAT FOV for an entire 2700 s duration is on the order of $F_{\rm ph,UL} \sim 10^{-6}$ \citep{Ackermann2012}, making it clear why none of the bursts in our sample were individually detected. The best-fit photon index is representative of indices measured at late times in previously detected LAT bursts (e.g. GRB~110731A) and consistent with photon indices measured by XRT of GRB afterglows at late times.
\begin{figure}[t]
\includegraphics[width=1\columnwidth]{MaxTSVsTime_GRBsNew.pdf}
\caption{Likelihood TS (black solid line) as a function of integration time from trigger, binned to 100 s intervals, for the entire GRB sample. The red dotted line represents the same analysis applied to designated background intervals exactly 30 sidereal orbits prior to trigger. The red dashed line represents the minimum TS that we assign to constitute a detection.}
\label{Fig:MaxTSVsTime}
\end{figure}
\begin{figure}[t]
\includegraphics[width=1\columnwidth]{LikelihoodContourPlot.pdf}
\caption{Contours of the joint likelihood profile as a function of photon flux and photon index, generated from data covering 2700 s post trigger and a 300 MeV to 30 GeV energy range.}
\label{Fig:LikelihoodContourPlot}
\end{figure}
In order to ensure that one or more GRBs are not dominating the observed excess seen in Figure~\ref{Fig:MaxTSVsTime}, we examine the TS distribution for the 2700 s second integration period succeeding $T_{0}$ in Figure~\ref{Fig:TSDistribution_GRBs_300Mevto30GeV_-10sto2700s}. The distribution peaks at TS $= 0$ for 41 GRBs, nearly half the sample under consideration. The remaining bursts are evenly distributed between $1 \lesssim$ TS $\lesssim 10$, with one burst (GRB~110903A) at TS $= 23$. This burst is just under the fiducial $5\sigma$ threshold for the a burst to be considered detected by the LAT. Removing this burst from the sample, we find that the TS of the joint likelihood analysis for the 2700 s second integration interval drops from TS = 55 to TS = 41, still yielding a strong detection of the remaining bursts. Moreover, the best-fit flux and index remain relatively unchanged, at $F_{\rm ph} = 6.0\times10^{-8}$ photons cm$^{-2}$ s$^{-1}$ and $\Gamma = -2.09$.
\subsection{Comparison of the Two Methods} \label{sec:MethodComparison}
\begin{figure}
\includegraphics[width=1\columnwidth]{TSDistribution_GRBs_300Mevto30GeV_-10sto2700s}
\caption{Distribution of individual TS values for a likelihood analysis covering the 2700 s after co-aligned trigger. The red dashed line represents the minimum TS that we assign to constitute a detection. }
\label{Fig:TSDistribution_GRBs_300Mevto30GeV_-10sto2700s}
\end{figure}
The counting and joint likelihood analyses employed here have both revealed the existence of a signal in excess of the expected background for the GRB locations in our sample. Here we compare the two methods to determine whether there are any inherent benefits to using one method over the other. The significance of the detected signal as a function of integration times, covering an energy range from 300 MeV to 30 GeV using a 10$^\circ$ ROI, for each method is displayed in Figure \ref{Fig:LikelihoodVsCounting_v2}. In order to ensure a proper comparison, we have removed the bursts that triggered ARRs of the spacecraft or occurred during ToOs from the likelihood analysis.
With the ARR and ToO bursts removed, the joint likelihood method provides stronger evidence for a signal excess when considering the entire 2700 s integration period, yielding a total signal significance of $\sim5 \sigma$, whereas the counting technique provides only a marginal detection at $\sim3.4 \sigma$. On smaller integration time, the two methods are roughly consistent, before diverging for longer integration intervals. The significance derived from the joint likelihood analysis rises gradually as more data are accumulated before leveling out to roughly $5 \sigma$, whereas the significance derived from the counting analysis remains largely unchanged through the inclusion of these data. Note that the final signal significance quoted here for the counting analysis differs from that reported in \S\ref{sec:ResultsCountingAnalysis} because of the different energy ranges under consideration. Here we are examining photons over a 300 MeV to 30 GeV energy range, in order to match the analysis performed using the joint likelihood technique.
\begin{figure}
\includegraphics[width=1\columnwidth]{LikelihoodVsCounting_v2_LiMa1983.pdf}
\caption{Comparison of the significance of the detected signal as a function of integration time, covering an energy range from 300 MeV to 30 GeV for a 10$^\circ$ ROI selection, for both the counting and likelihood techniques.}
\label{Fig:LikelihoodVsCounting_v2}
\end{figure}
\subsection{Population Demographics} \label{sec:PopulationDemographics}
In order to examine which bursts in our sample contribute most significantly to the observed signal, we calculate the cumulative TS by rerunning the likelihood analysis for each additional burst, sorted as a function of a selected burst property. An example of this method can be seen in Figure \ref{Fig:TSVsExposure}, where the cumulative TS is plotted as a function of burst exposure. Here exposure is defined as an integral of the total response over the entire ROI\footnote{http://fermi.gsfc.nasa.gov/ssc/data/analysis/documentation/Cicerone}. The TS is first calculated for the burst with the lowest exposure, covering an energy range from 300 MeV to 30 GeV and a period of 2700 s post trigger, then recalculated by including the data from the burst with the next highest exposure, and so on. The result is an increasing cumulative TS that eventually peaks at the value reported in Section~\ref{sec:JointLikelihoodAnalysis}. The dashed line in Figure \ref{Fig:TSVsExposure} represents the cumulative sample size as a function of exposure, and the dotted line demarcates the 50th percentile of the sample. If all the bursts contributed equally to the final TS value, independent of their exposure, then one would expect the cumulative TS to rise at the same rate as the cumulative sample size. This is not the case in Figure \ref{Fig:TSVsExposure}, where the bursts with exposures in the first 50th percentile of the sample contribute very little to the final signal significance. This would be expected, since longer observations of bursts closer to the instrument boresight would be more sensitive at detecting extended subthreshold emission. Still, Figure \ref{Fig:TSVsExposure} shows that exposure alone is not entirely indicative of whether a burst will contribute to the final TS value. Several bursts with the highest exposure are seen to contribute very little to the overall signal, whereas several of the bursts with the lowest exposure do contribute to the significance of the final signal. This is also reflected by the fact that the non-ARR sample of bursts discussed in Section~\S\ref{sec:MethodComparison} is still detected at roughly $\sim5 \sigma$ (TS$\sim26$), compared to $\sim7.7\sigma$ (TS$\sim59$) for the entire sample. Therefore, the ARR and ToO bursts contribute most, but not all, of the observed signal.
We can test whether the observed low-energy gamma-ray flux of a source plays an additional role in its detectability and hence contribution to the final TS value. To examine this, we calculated the cumulative TS as a function of the burst's peak photon flux in the 15-150 keV energy range as measured by BAT \citep{Donato2012}, and we show the results in Figure \ref{Fig:TSVsPeakFlux}. The bursts with the highest observed peak flux contribute strongly to the cumulative TS, although there remain bursts with flux values above 1.0 photon cm$^{-2}$ s$^{-1}$ that do not appear to contribute significantly to the final TS value. Note that the sample size presented in Figure \ref{Fig:TSVsExposure} differs from that in Figure \ref{Fig:TSVsPeakFlux} because not every burst in the likelihood sample was initially detected by BAT, and therefore a photon flux estimate of their prompt emission is not available.
Given the extended nature of the signal inferred in Figure \ref{Fig:LikelihoodVsCounting_v2}, we also calculate the cumulative TS as a function of the burst's X-ray brightness at 11 hrs, in the 0.3-10 keV energy range, as measured by XRT. The X-ray flux at 11 hr has become a standard measure of afterglow brightness, as it samples the afterglow light curve at a period where the steep decay and plateau phases have typically ceased \citep{Nousek06}. To calculate these values, we downloaded the XRT data for each from the {\em Swift}~XRT light-curve repository \citep{Evans09} and fit the flux light curves with a multi-segmented afterglow model defined by \citet{Zhang06}. The fitting procedure employed is outlined in \citet{Racusin09}. The result of the X-ray brightness versus cumulative TS analysis is presented in Figure \ref{Fig:TSVsXRTFlux_11hrs}. There exists an even stronger trend of bursts with bright X-ray emission 11 hrs post trigger contributing significantly to the final TS value, compared to the prompt gamma-ray flux measured by the BAT. In fact, fewer than half of the brightest bursts in the sample contribute most of the signal, with the addition of the remaining bursts actually decreasing the final signal significance.
\begin{figure}
\includegraphics[width=1\columnwidth]{TSVsExposure.pdf}
\caption{Cumulative likelihood TS as a function of burst exposure.}
\label{Fig:TSVsExposure}
\end{figure}
\begin{figure}
\includegraphics[width=1\columnwidth]{TSVsPeakFlux.pdf}
\caption{Cumulative likelihood TS as a function of peak flux as measured by BAT over a 15-150 keV energy range.}
\label{Fig:TSVsPeakFlux}
\end{figure}
\section{Discussion} \label{sec:Discussion}
The stacking analysis presented above shows significant evidence for subthreshold emission among LAT non-detected GRBs. These results are consistent with the conclusions drawn from the first {\em Fermi}~LAT GRB catalog \citep{Ackermann2013}, in which the authors find a significance correlation between the burst fluence as measured by GBM in the 10 keV$--$1 MeV energy range, the LAT boresight angle at the time of trigger, and burst detectability (see their Figure 31). The authors find that LAT-detected GRBs are among the most fluent bursts observed by the GBM, but that this fluence threshold for a LAT detection falls with decreasing boresight angle, tracing the instrument sensitivity as a function of off-axis angle.
The counting and likelihood analysis techniques employed above show evidence for subthreshold emission on both prompt and extended timescales, mirroring the range of emission timescales observed in the LAT-detected population. The observed signal detected by both techniques remains significant over the entire 2700 s period under consideration and suggests that the extended emission observed in some LAT-detected GRBs may be common among the population. The photon index measured through the likelihood analysis is consistent with the average value measured for the LAT-detected population \citep{Ackermann2013}. This value is also consistent with the photon index expected from the high-energy extension of the synchrotron spectrum due to the external forward shock in the standard afterglow theory, supporting an external shock origin of the extended emission.
Our detection of ubiquitous long-lived emission, albeit below the LAT detection threshold, is consistent with similar work performed by \citet{Lange13}. In their work, the authors examined a sample of 99 GBM localized GRBs that were not detected by LAT and found evidence for long-lived subthreshold emission that lasted for as much as 10 times the bursts' $T_{\rm 90}$ duration in the keV-MeV energy range. Unlike the conclusions drawn by \citet{Lange13}, though, we find that the photon index of the subthreshold population derived from our joint likelihood analysis is largely consistent with that observed in the LAT-detected bursts. The different results may be due to the different techniques employed, with the hardest bursts in our sample contributing most significantly to the joint likelihood results, and hence possibly providing a harder average photon index than that measured by \citet{Lange13}.
\begin{figure}
\includegraphics[width=1\columnwidth]{TSVsXRTFlux_11hrs.pdf}
\caption{Cumulative likelihood TS as a function of X-ray flux as measured at 11 hrs post trigger by XRT over a 0.3-10 keV energy range.}
\label{Fig:TSVsXRTFlux_11hrs}
\end{figure}
The comparison of the counting and likelihood techniques presented in Figure \ref{Fig:LikelihoodVsCounting_v2} reveals that the two methods produce consistent results when applied over short time intervals, but diverge when considering longer integration periods. This is likely due to the increased sensitivity of the likelihood method, which takes into consideration the energy of the arriving photons, their position within the ROI, and the instrument PSF as function of energy. High-energy photons that are consistent with the position of the candidate point source may only constitute a small percentage of the observed source flux over the expected background, but their arrival may be highly improbable (and hence significant) given the spectral shape of the assumed background model. The likelihood analysis inherently accounts for this when calculating the significance of the source detection and may explain the diverging results returned by the two methods for long integration intervals. As a result, we conclude that the joint likelihood method is more sensitive for source detection on intermediate and long timescales, but emphasize that the model assumptions and computational overhead inherent to the likelihood method still leave room for a counting analysis when searching for uncharacterized emission over a wide parameter space.
The cumulative likelihood analysis presented in Figures \ref{Fig:TSVsPeakFlux} and \ref{Fig:TSVsXRTFlux_11hrs} reveals that both a burst's prompt gamma-ray flux and afterglow X-ray flux strongly correlate with the strength of the subthreshold emission. These results are consistent with the interpretation that an extension of both the prompt and afterglow spectra contribute to the emission observed in the LAT energy range. An inspection of Figure \ref{Fig:TSVsPeakFlux} reveals, though, a number of bright BAT-detected GRBs that nonetheless do not contribute significantly to the stacked signal significance. These bursts are consistent with the population of bright GBM-detected bursts first reported in \citet{Ackermann2012} that are in the LAT FOV at the time of trigger but which produce no significant emission at MeV and GeV energies. A detailed broadband spectral analysis of these LAT dark bursts by \citet{Ackermann2012} attributes their non-detection to spectral curvature of their high-energy spectrum, possibly due to pair attenuation.
In contrast, the X-ray afterglow flux measured by XRT at 11 hrs correlates particularly strongly with the signal significance, with less than half the sample resulting in almost all of the observed excess. This suggests that the prompt and afterglow phases may not both contribute to the LAT-detected emission for individual bursts, with the prompt phase contribution being suppressed in some cases. The differences between Figures \ref{Fig:TSVsPeakFlux} and \ref{Fig:TSVsXRTFlux_11hrs} could imply that the LAT detection of the afterglow contribution could be due to simple threshold effects, whereas the detection of the prompt phase may be due to both threshold effects and an intrinsic suppression of the high-energy emission in some cases. Detailed broadband spectral fits of XRT derived spectra and LAT upper limits will be required to decipher whether the LAT nondetections of the high-energy component of bright X-ray afterglows are simply due to instrumental sensitivity, or whether a break between the XRT and LAT energy ranges is required to explain the LAT nondetections. Nonetheless, Figure \ref{Fig:TSVsXRTFlux_11hrs} shows that the X-ray afterglow flux is a strong predictor of the strength of the subthreshold LAT emission.
Finally, we note that the TS value for the integration period covering the 50 s prior to the co-aligned trigger presented in Figure \ref{Fig:MaxTSVsTime}, while statistically not significant, remains elevated in comparison to the composite likelihood analysis performed over the same interval during the background orbits. This may just reflect the varying trigger times with respect to the true start of the prompt emission, or it may suggest the presence of high-energy precursor emission prior to the prompt emission at keV energies. Such activity has not yet been observed at MeV or GeV energies in any of the LAT-detected GRBs \citep{Ackermann2013}. Ultimately, an analysis of a larger sample of bursts and/or the use of the upcoming Pass 8 event reconstruction \citep{Atwood2013}, which significantly improves the LAT sensitivity at low energies, will be needed to investigate whether such emission exists and how it compares temporally to activity at keV energies.
\section{Conclusions} \label{sec:Conclusions}
We perform a comprehensive stacking analysis of LAT data of {\em Swift}~localized GRBs that were not detected by the LAT, but which fell within the instrument's FOV at the time of trigger. We examine a total of 79 GRBs by comparing the observed counts over a range of time intervals to that expected from designated background orbits, as well as by using a joint likelihood technique to model the expected distribution of stacked counts, and find strong evidence for subthreshold emission at MeV to GeV energies using both techniques. This observed excess is detected during intervals that include and exceed the durations typically characterizing the prompt emission observed at keV energies and lasts at least 2700 s after the co-aligned burst trigger.
By utilizing a novel cumulative likelihood analysis, we are also able to identify which bursts contribute most significantly to the stacked signal. We find that although a burst's prompt gamma-ray flux and afterglow X-ray flux both correlate with the strength of the subthreshold emission, the X-ray afterglow flux measured by XRT at 11 hrs post trigger correlates far more significantly. This suggests that although the prompt and afterglow phases may both contribute to the LAT-detected emission for individual bursts, the prompt phase contribution may be suppressed in some cases. This is consistent with the population of bright GBM-detected bursts that are in the LAT FOV at the time of trigger, but which produce no significant emission at MeV energies and above.
Overall, the extended nature of the subthreshold emission and its connection to the burst's afterglow brightness lend further support to the external forward shock origin of the late-time emission detected by the LAT. These results suggest that the extended high-energy emission observed by the LAT may be a relatively common feature but remains undetected in a majority of bursts owing to instrumental threshold effects.
\acknowledgements
The {\em Fermi}\ LAT Collaboration acknowledges generous ongoing support
from a number of agencies and institutes that have supported both the
development and the operation of the LAT as well as scientific data analysis.
These include the National Aeronautics and Space Administration and the
Department of Energy in the United States, the Commissariat \`a l'Energie Atomique
and the Centre National de la Recherche Scientifique / Institut National de Physique
Nucl\'eaire et de Physique des Particules in France, the Agenzia Spaziale Italiana
and the Istituto Nazionale di Fisica Nucleare in Italy, the Ministry of Education,
Culture, Sports, Science and Technology (MEXT), high-energy Accelerator Research
Organization (KEK) and Japan Aerospace Exploration Agency (JAXA) in Japan, and
the K.~A.~Wallenberg Foundation, the Swedish Research Council and the
Swedish National Space Board in Sweden.
Additional support for science analysis during the operations phase is gratefully
acknowledged from the Istituto Nazionale di Astrofisica in Italy and the Centre National d'\'Etudes Spatiales in France.
\include{Table}
|
\section{Introduction}
In the last years, the interest on two-dimensional conformal field theories (CFT) with $W_n$ symmetry has been renewed, mainly because of two reasons: On the one hand, $W_n$ algebra has been identified as the charge algebra associated to the asymptotic symmetries of spin-$n$ theories in three-dimensional Anti-de Sitter spacetime \cite{Campoleoni, Henneaux}, generalizing the spin-$2$ result of Brown and Henneaux \cite{Brown-Henneaux}. On the other hand, it has been observed in \cite{AGTW} that conformal blocks of ${sl}(n)$ Toda conformal field theory, the archetypical example of a theory with $W_n$ symmetry, codify the information of the partition function of $\mathcal{N}=2$ four-dimensional superconformal quiver theories for the gauge group $SU(n)$, generalizing the Alday-Gaiotto-Tachikawa (AGT) correspondence \cite{AGT} to the case $n\geq 2$. Here, we will study the $sl(n)$ Toda field theory (TFT) from a different perspective: We will consider TFT as a generating functional of $N$-point correlation functions of a plethora of new non-rational conformal field theories that are defined ipso facto. This is in the line of the so-called $H^+_3$ Wess-Zumino-Witten -- Liouville correspondence \cite{Ribault-Teschner} or, more precisely, of the extension of it proposed in Ref. \cite{Ribault-family}, where correlation functions of a new family of non-rational CFTs were constructed in terms of Liouville field theory (namely, in terms of the $sl(2)$ Toda field theory). Here, we provide a generalization of these results to the case $sl(n)$.
The $H^+_3$ Wess-Zumino-Witten -- Liouville correspondence has been proposed in Ref. \cite{Ribault-Teschner}. It is based on a remarkable relation existing between the $sl(2)$ Knizhnik-Zamolodchikov (KZ) equation and the Belavin-Polyakov-Zamolodchikov (BPZ) equation \cite{Stoyanovsky}. The latter is the differential equation obeyed by Liouville correlation functions that involve degenerate fields, usually denoted by $V_{-1/(2b)}$. Using that solutions of the BPZ equation can be related to solutions of the KZ equation for $sl(2)$, it was shown in \cite{Ribault-Teschner} that $N$-point correlators in $H_3^+= SL(2,\mathbb{C})/SU(2)$ Wess-Zumino-Witten (WZW) on the sphere can be written in terms of $(2N-2)$-point correlators in Liouville field theory formulated also on the sphere, where the latter correlators include $N-2$ degenerate fields $V_{-1/(2b)}$. This correspondence has been later generalized to genus-$g$ \cite{Schomerus}, to surfaces with boundaries \cite{Ribault-boundary, Ribault2, Ribault3, Ribault-Hosomichi}, to spectrally flowed representations of $\hat{sl}(2,\mathbb{R})$ \cite{Ribault-winding}, to the irregular vertex representations of the Virasoro algebra \cite{Gaiotto}, and to supersymmetric theories \cite{Schomerus2}. In Ref. \cite{Ribault-family}, a generalization of the $H^+_3$ WZW--Liouville correspondence was proposed for the case in which the Liouville correlators include $N-2$ degenerate fields of higher level, $V_{-m/(2b)}$ with $m\in \mathbb{Z}_{>0}$. In the case $m=1$, these correlators yield correlators of $H^+_3$ WZW \cite{Ribault-Teschner, Stoyanovsky}; however, in the case $m>1$ the $(2N-2)$-point Liouville correlators including $N-2$ fields $V_{-m/(2b)}$ were argued to generate the $n$-point correlation functions of a new family of non-rational conformal field theories \cite{Ribault-family}. This family of CFTs is parameterized by $m$, which enters in the central charge as $c=3+6(b+(1-m)/b)^2$, with the Liouville central charge being $c_L=1+6(b+1/b)^2$. The consistency of these theories was later studied in Refs. \cite{Giribet-Babaro1, Giribet-Babaro2, Nicolas}, where correlation functions on different surfaces were computed. It is possible to speculate that such CFTs actually exist even for other real non-integer values $m$, where the Liouville correlators do not necessarily involve degenerate representations.
Another natural question that has been addressed in the literature is whether an extension of the $H^+_3$ WZW--Liouville correspondence is possible for higher rank algebras. That is, whether a relation of this sort also exists between, say, the $SL(n,\mathbb{R})$ WZW theory and the $sl(n)$ Toda theory, which would extend the results of \cite{Ribault-Teschner, Ribault-winding} to the case $n\geq 2$. This problem has been explored in Ref. \cite{Ribault-sl3} for the case $n=3$, where such a correspondence has been observed to emerge in the highly quantum limit $k\to 3$, where the WZW level equals the Coxeter number of the $sl(3)$ algebra. In the recent paper \cite{Hikida-sl3}, a more general correspondence for cases $n\geq 3$ was achieved by considering particular $sl(2)$ embeddings in (super)algebras.
Here, aimed at investigating whether a generalization of the construction of \cite{Ribault-family} to higher rank is possible, we will propose a larger family of non-rational CFTs whose $N$-point correlation functions are given by a subset of $(N+(N-2)r)$-point correlation functions of $sl(n)$ Toda field theories, with $r=n-1$ being the rank of $sl(n)$. These new CFTs are parameterized by a collection of $n-1$ real numbers $\{ m_1,m_2,...\ m_{n-1}\} $. They correspond to deformations of Toda field theories that, while preserve confomal invariance, seem to break the $W_n$ symmetry. We will see, however, that the CFTs constructed through this procedure may also preserve a larger symmetry, generated by an affine extension of a proper subalgebra of the Borel subalgebra of the affine Kac-Moody $\hat{sl}(n)_k$. The case $n=3$ is analyzed explicitly to illustrate such remnant infinite-dimensional symmetry in relation to the symmetries of the $sl(3)$ WZW model. We will present a Lagrangian representation of these new non-rational CFTs, which is the natural generalization of the Lagrangian proposed in \cite{Ribault-family} for $n=2$, here involving interaction operators between the $n-1$ scalar fields associated to the simple roots of $sl(n)$ and $n-1$ copies of the $\beta $-$\gamma $ ghost system. In the case $n=2$ the theories coincide with those defined in Ref. \cite{Ribault-family}; that is, for $n=2$ and $m_1=1$ it corresponds to the $H_3^+$ WZW model\footnote{One may wonder whether the non-rational CFTs defined in this way are unitary. Being $c>1$ theories defined as deformation of Toda field theories, it is probable that, at least in some particular cases, the theory can be render unitary provided one considers some restriction of the Hilbert space. This is the case, for instance, of the $SL(2,\mathbb{R})$ WZW model, which represents a unitary theory when one restricts the isospin $j$ of the discrete representations.}, while for $n=2$ and $m_1=0$ it reduces to Liouville theory. In general, the {\it undeformed} case $m_i=0$ with $i=1, 2, ...\ n-1$ for arbitrary $n$ it corresponds to $sl(n)$ TFT.
The paper is organized as follows: In section 2, we briefly review the theory of Toda for the Lie algebra $sl(n)$. In section 3, we propose the Lagrangian representation of a family of CFTs which consist of deformations of TFT Lagrangian. This is a generalization of the theories proposed in \cite{Ribault-family} to the case $n>2$. We compute the correlation functions for these theories in the path integral approach, following the techniques developed in \cite{Schomerus} adapted to this case. We show that these correlation functions are determined in terms of those of TFT. We also derive an integral representation for the $N$-point functions in the Coulomb gas formalism. In section 4, we study the remnant affine symmetry that the deformed TFTs exhibit. We consider the case of $sl(3)$ to illustrate the details of such symmetry. We discuss the relation between the deformation of the $sl(3)$ TFT and the $sl(3)$ WZW model, which leads to suggest other deformations of the former. Section 5 contains our conclusions.
\section{Toda conformal field theory}
The $sl(n)$ TFT is a conformal field theory whose degrees of freedom are represented by $n-1$ bosons living in the $(n-1)$-dimensional root space of $sl(n)$ Lie algebra; see for instance Refs. \cite{Fateev-Litvinov, Fateev-Litvinov2, Miura, Fateev-Litvinov3, Fateev-Ribault}. There are $n-1$ simple roots $e_1$, $e_2$, ..., $e_{n-1}$ in this Lie algebra and its Cartan matrix is given by
\begin{equation}
K_{ij}\ =\ \left(
\begin{array}{cccccc}
2 & -1 & 0 & \cdots & \cdots & 0 \\
-1 & 2 & -1 & \cdots & \cdots & \cdots \\
0 & -1 & 2 & \cdots & \cdots & \cdots \\
\cdots & \cdots & \cdots & \cdots & \cdots & \cdots \\
\cdots & \cdots & \cdots & \cdots & 2 & -1 \\
0 & \cdots & \cdots & \cdots & -1 & 2 \\
\end{array}
\right) .
\end{equation}
This defines the inner product $(. , . )$. That is to say, this defines the bilinear form $K_{i\ j}=(e_i,e_j)$ with $K_{i\ i+1}=K_{i+1\ i}=-1$, $K_{i\ i}=2$, and $0$ otherwise in the basis above. The $n-1$ fundamental weights $\omega _i$ are defined in such a way they span dual roots, i.e. $(\omega_i,e_j)=\delta_{ij}$.
The Lagrangian of the $sl(n)$ Toda CFT is
\begin{equation}
S_T\ =\ \frac{1}{2\pi}\int d^2z\bigg((\partial\varphi,\bar{\partial}\varphi)+\frac{(Q_T,\varphi)}{4}\mathcal{R}+
\sum_{k=1}^{n-1}e^{\sqrt{2}b(e_k,\varphi)}\bigg)\label{toda}
\end{equation}
where $b$ is an arbitrary (real) parameter, and the background charge is
\begin{equation}
Q_T\ =\ \sqrt{2}\bigg( b+\frac{1}{b}\bigg)\rho \ , \label{telamorfaste}
\end{equation}
where $\rho$ is the Weyl vector; that is, $\rho $ is the half-sum of all positive roots. Normalization of the action is such that the free field propagator of the fields are
\begin{equation}
\langle \varphi_{k }(z)\varphi_{\ell }(w)\rangle \ =\ -\frac{1}{2}\delta_{k\ell }\log (z-w) \ ,
\end{equation}
with $k, \ell = 1, 2, ...\ n-1$.
For every integer value $n$, action (\ref{toda}) defines a different theory. The case $n=2$ corresponds to Liouville field theory; in such case there is just one simple root $e_1=\sqrt{2}$ and just one highest weight $\omega_1=1/\sqrt{2}$. Liouville field theory is a conformal field theory with central charge $c=1+6Q_L^2$ and background charge $Q_L=(b+b^{-1})$. For $n>2$ the theory contains $n-1$ fields of the Liouville type interacting each other by specific couplings that are given by the $sl(n)$ structure. The resulting theory, the $sl(n)$ TFT, is also conformal invariant and its central charge is given by
\begin{equation}
c_T\ =\ n-1+6Q_T^2\ =\ (n-1)\left(1+n(n+1)\left(b+{b}^{-1}\right)^2\right),
\end{equation}
where $Q_T^2= (Q_T,Q_T)$.
These theories present spin-$n$ symmetry, represented by $n-1$ holomorphic and $n-1$ antiholomorphic conserved currents. These generate the ${ W_n}\oplus { \bar{W}_n}$ algebra, which contains Virasoro as a subalgebra (see (\ref{TToda})-(\ref{Virato}) below). Let us denote such currents by ${W}^s(z)$ and ${\bar{W}^s (\bar{z})}$ with ${s}=1,2,3,...\ n$. The index ${s}$ labels the spin of the current. These admit a representation in terms of Miura transformation \cite{Miura}
\begin{equation}
\prod_{r=0}^{n-1}\left((q+q^{-1})\partial+\sqrt{2}(\varrho_{n-r},\partial\varphi)\right)\ =\ \sum_{r=0}^n{W^{n-r}}(z)((q+q^{-1})\partial)^r
\end{equation}
where the vectors $\varrho_k$ are the weights of the first fundamental representation of the Lie algebra
$sl(n)$ with the highest weight $\omega_1$, namely $\varrho_k\ =\ \omega_1 - e_1-e_2 -...\ e_{k-1}.$
The first and second currents are taken to be ${ W^0}=1$ and ${ W^1}=0$; the third one is the first nontrivial one and corresponds to the energy-momentum tensor
\begin{equation}
{W^2}(z) =\ -\left(\partial\varphi,\partial\varphi\right)+\left(Q_T,\partial^2\varphi\right), \label{TToda}
\end{equation}
while $W^3(z)$ is the generator of the Zamolodchikov algebra \cite{Zamolodchikov}. It is customary to write these algebras (and their $n>3$ generalization) in terms of the modes of the currents, which are defined by the expansion
\begin{equation}
{W^s}(z) = \sum_{r\in \mathbb{Z}} {W^s}_r \ z^{-r-s} \ .
\end{equation}
In terms of the modes, and in the case $n=3$, in addition to the Virasoro algebra
\begin{eqnarray}
[W_n^2 , W_m^2 ] = (n-m) W_{n+m}^2 + \frac{c}{12} n(n^2-1)\delta_{n+m , 0} \label{Virato}
\end{eqnarray}
the algebra includes the Lie products
\begin{eqnarray}
[W_n^2 , W_m^3 ] = (2n-m) W_{n+m}^3
\end{eqnarray}
and
\begin{eqnarray}
[W_n^3 , W_m^3 ] &=& \frac{1}{15} (n-m) \left( n^2+ m^2 -\frac{1}{2} nm -4\right) W_{n+m}^2 + \nonumber \\
&&\frac{16}{22+5c}(n-m)\Lambda_{n+m}+
\frac{c}{360} n(n^2-1)(n^2-4) \delta_{n+m,0} \ , \label{coso}
\end{eqnarray}
where the quadratic piece in (\ref{coso}) is given by the normal ordered product
\begin{equation}
\Lambda_{n} = \sum_{m\in \mathbb{Z}} :W^2_{m}W^2_{n-m}: - \frac{1}{4} t_{n}W^2_n \ ,
\end{equation}
with $t_n = n^2 -4$ for $n$ even, and $t_n = n^2+2n-15$ for $n$ odd. As usual, the normal order is defined as $:W^2_nW^2_m:\ \equiv \ W^2_nW^2_m \theta (n-m)+W^2_mW^2_n \theta (m-n)$, with $\theta (x)=\int_{\mathbb{R}_{<x}} dy\delta (y)$ being the Heaviside step function with $\theta(0)=1/2$. The OPE that realizes the $W_3$ algebra (\ref{Virato})-(\ref{coso}) takes the form
\begin{eqnarray}
W^2(z)W^2(w) &=& \frac{c/2}{(z-w)^4} + \frac{2W^2(w)}{(z-w)^2} + \frac{\partial W^2(w)}{(z-w)} + ... \nonumber \\
W^2(z)W^3(w) &=& \frac{3W^3(w)}{(z-w)^2} + \frac{\partial W^3(w)}{(z-w)} + ... \nonumber \\
W^3(z)W^3(w) &=& \frac{c/3}{(z-w)^6} + \frac{2W^2(w)}{(z-w)^4} + \frac{\partial W^2(w)}{(z-w)^3} + \frac{1}{(z-w)^2}\bigg( \frac{32}{22+5c} \Lambda (w) + \frac{3}{10} \partial^2W^2(w) \bigg) + \nonumber \\
&+&
\frac{1}{(z-w)}\bigg( \frac{16}{22+5c} \partial \Lambda (w) + \frac{1}{15} \partial^3W^2(w) \bigg)+ ... \nonumber \\
\end{eqnarray}
with $\Lambda(z) = :(W^2(z))^2:-(3/10)\partial^2W^2(z)$.
The basic objects in the construction of conformal TFT are exponential fields parameterized by an $(n-1)$-component vector parameter $\alpha = (\alpha_1 , \alpha_2, ...\ \alpha_{n-1})$; namely
\begin{equation}
V_{\alpha}^T(z)\ =\ e^{2(\alpha,\varphi (z))} \ ,
\end{equation}
which are in correspondence with primary states of the theory $|\alpha \rangle = \lim_{z\to 0 } V^T_{\alpha }(z)| 0 \rangle $. The operator product expansion (OPE) between the currents and these primary fields read
\begin{equation}
{W^k}(z)V_{\alpha}^T(w)\ =\ \frac{h_{\alpha}^{k}V_{\alpha}^T(w)}{(z-w)^k}\, \left( 1+ \ldots \right)
\end{equation}
where the ellipses stand for terms that vanish when $w\to z$. In particular, it yields
\begin{equation}
{W^2}(z)V_{\alpha}^T(w)\ =\ \frac{h_{\alpha}^{2}V_{\alpha}^T(w)}{(z-w)^2}+ \frac{W^{2}_{-1} V_{\alpha}^T(w)}{(z-w)}+ ...
\end{equation}
with $W^{2}_{-1}V_{\alpha}^T(w) = \partial V_{\alpha}^T(w)$. From this, we obtain the conformal dimension
\begin{equation}
h_{\alpha}^{2} =\ \left(\alpha,Q_T-\alpha\right).
\end{equation}
The OPE with the $W^3(z)$ current reads
\begin{equation}
{W^3}(z)V_{\alpha}^T(w)\ =\ \frac{h_{\alpha}^{3}V_{\alpha}^T(w)}{(z-w)^3}+
\frac{W^{3}_{-1} V_{\alpha}^T(w)}{(z-w)^2}+ \frac{W^{3}_{-2} V_{\alpha}^T(w)}{(z-w)}+ ...
\end{equation}
where $h_{\alpha }^3$ is an expression cubic in $\alpha $ and quadratic $b$, and symmetric under $b\leftrightarrow b^{-1}$.
The $N$-point correlation functions in TFT on the Riemann sphere are defined by
\begin{equation}
\left\langle \prod_{i=1}^N V^T_{\alpha_i}(z_i) \right\rangle_{\text{TFT}} \ \equiv \ \int \prod_{a=1}^{n-1} {\mathcal D}\varphi _a \ e^{-S_T} \ \prod_{i=1}^{N} e^{2(\alpha_i , \varphi (z_i))} \label{BBBB}
\end{equation}
where the fields in the functional integral are defined on $\mathbb{CP}^1 \setminus \{z_1, z_2, ... \ z_N \}$. In the next section, we will propose a definition of a set of two-dimensional CFTs whose correlation functions can be expressed in terms of TFT correlation functions (\ref{BBBB}). Proving such correspondence between observables in the case of $sl(2)$ can be assisted by the modular differential equations that correlators that involve degenerate representations obey. In the case of TFT for $sl(n>2)$, the structure of degenerate and semi-degenerate is less restrictive \cite{Ribault-sl3, Fateev-Litvinov, Fateev-Litvinov2, Miura, Fateev-Litvinov3}. Because of that, and because we are interested in defining the theories for generic values of the parameters $m_i$, in this paper we will resort to the path integral approach, following the techniques of reference \cite{Schomerus}. We will also derive an integral representation for the $N$-point functions resorting to the free field approach.
\section{Deformations of Toda field theory}
\subsection{Lagrangian representation}
Generalizing $sl(n)$ conformal Toda theories, and generalizing at the same time the family of theories presented in \cite{Ribault-family}, here we will consider the theories defined by the following actions
\begin{equation}\begin{split}
S_{\{ m_k \}}\ =\ \frac{1}{2\pi }\int d^{2}z
\bigg( & (\partial \phi, \bar{\partial}\phi) \,+\,\sum_{k=1}^{n-1}\left(\beta_k\bar{\partial}\gamma_k\,+\
\bar{\beta}_k\partial \bar{\gamma}_k\right)\,+\,\frac{(Q_{\{m_k\}},\phi)}{4}\mathcal{R}\ +
\\& \,\sum_{k=1}^{n-1}(-\beta_k\bar{\beta}_k)^{m_k}\,e^{\sqrt{2}b(e_k,\phi)}\bigg)\label{nuevaaccion}
\end{split}\end{equation}
which, apart from the $(n-1)$-component boson field $\phi =(\phi_1, \phi_2, ...\ \phi_{n-1})$ multiplied with the $sl(n)$ based inner product, includes $n-1$ copies of the commutative $\beta $-$\gamma $ ghost system of conformal weight $(1,0)$. The definition of the theory also requires the specification of $n-1$ (real) parameters $m_k$, $k=1, 2, ...\ n-1$, and the parameter $b$. For each collection $\{ m_k\} =\{ m_1, m_2, ...\ m_{n-1}\} $ one obtains a different\footnote{It is, however, possible that a given CFT is represented by more than one set $\{ m_1, m_2, ...\ m_{n-1}\}$; see \cite{Nicolas}.} CFT, which is a deformation of TFT. The background charge depends on these parameters $m_k$; it is given by
\begin{equation}
Q_{\{m_k\}}=Q_T\,+\delta Q \ , \ \ \ \ \text{with}\ \ \ \delta Q =- \sum_{k=1}^{n-1}\frac{\sqrt{2}m_k}{b}\omega_k \ , \label{deformaste}
\end{equation}
where $Q_T$ is the Toda background charge (\ref{telamorfaste}). The specific value (\ref{deformaste}) is such that the interaction terms in (\ref{nuevaaccion}) are marginal. This value for the background charge yields the central charge $c_{\{m_k\}} = 3(n-1)+6Q_{\{m_k\}}^2$. This can be obtained from the OPE
\begin{equation}
W^2(z)W^2(w) = \frac{c_{\{m_k\}}/2}{(z-w)^4} + \frac{2W^2(w)}{(z-w)^2} + \frac{\partial W^2(w)}{(z-w)} + ...
\end{equation}
with the stress tensor being
\begin{equation}
W^2(z) = \sum_{\ell = 1}^{n-1} \beta_{\ell }\partial \gamma_{\ell} - (\partial \phi , \partial \phi) + ( Q_{\{m_k\}}, \partial^2 \phi) \ .
\end{equation}
From this one can also verify that the operators $\beta_{\ell}^{m_{\ell}}\bar{\beta}_{\ell}^{m_{\ell}}e^{\sqrt{2}b(e_{\ell},\phi)}$ in (\ref{nuevaaccion}) have actually conformal weight 1. This follows from $(e_k ,\delta Q)=-\sqrt{2}m_k /b$.
We will refer to the CFTs defined by action (\ref{nuevaaccion}) as $m$-deformed CFTs. As said in the introduction, these theories are the natural generalization of the Lagrangian proposed by Ribault in Ref. \cite{Ribault-family}. In fact, in the case $n=2$ with $m_1=1$ the theory corresponds to the level $k=b^{-2}+2$ $SL(2,\mathbb{R})$ WZW model, while for $n=2$ and $m_1=0$ it reduces to Liouville theory. In general, the {\it undeformed} case $m_k=0$ for all $k=1, 2, ...\ n-1$ for arbitrary $n$ corresponds to $sl(n)$ TFT.
We consider primary states defined by vertex operators that are exponential functions of the fields; namely
\begin{equation}
\Phi_{\alpha }(\{p_k\}|z)\ =\ {\mathcal N} \,
e^{[p,\gamma (z)]}\,e^{{2}(\alpha,\phi (z))} \label{Gmoco}
\end{equation}
where we define $[ p,\gamma ] \equiv \sum_{k=1}^{n-1} (p_k\gamma_k -\bar{p}_k \bar{\gamma }_k)$ with $\{ p_1, \ p_2, \ ... \ p_{n-1} \} $ being complex variables, and where ${\mathcal N}$ is a normalization factor, that, in principle, can depend on the momenta $\{ p_1, \ p_2, \ ... \ p_{n-1} \} $ and $\alpha $, and on the parameters of the theory $\{ m_1, \ m_2 , \ ... \ m_{n-1} \} $ and $b$.
\subsection{Correlation functions: Path integral}
The quantities we are interested to compute are the $N$-point correlation functions
\begin{equation}
\Omega\left(\{p_k^{\nu}\}, \{\alpha^{\nu}\} |\{ z_{\nu}\} \right) =
\left\langle \prod_{\nu =1}^{N}\Phi_{\alpha_{\nu}}(\{p_k\}^{\nu} |z_{\nu
})\right\rangle_{\{ m_k \} } \equiv \ \int \prod_{k=1}^{n-1} \mathcal{D}\phi_k \mathcal{D
^{2}\beta_k \mathcal{D}^{2}\gamma_k\,e^{-S_{\{m_k\}}}\prod_{\nu
=1}^{N}\Phi_{\alpha_{\nu }}(\{p_k\}^{\nu }|z_{\nu }), \label{Cosillas}
\end{equation}
where our notation is such that the symbol $\{ p_k^{\nu}\}$ in the argument of a function represents the collection of $N(n-1)$ elements $\{p_1^1, p_2^1, ... \ p_{n-1}^1 ; p_1^2, p_2^2, ... \ p_{n-1}^2 ; ...\ p_1^N, p_2^N, ... \ p_{n-1}^N \}$ on which the function depends, while the symbol $\{ p_k\}^{\nu}$ represents the collection of $(n-1)$ variables $\{p_1^{\nu }, p_2^{\nu }, ... \ p_{n-1}^{\nu } \}$ with $\nu $ fixed. Indices $\mu , \nu$ will be used to run over $\{ 1,2, ...\ N \}$, while indices $k,\ell $ run over $\{ 1,2, ...\ n-1 \}$, and indices $a,b$ run over $\{ 1,2, ...\ N-2 \}$.
The first step in computing (\ref{Cosillas}) is dealing with the ghost system: The integration over the fields $\gamma_k $ and $\bar{\gamma}_k$ produces a product of $n-1$ double $\delta$-functions
\begin{equation}
\delta^2\left(\bar{\partial}\beta_k(w)\,-\,2\pi\sum_{\nu=1}^Np_k^{\nu}\delta^2(w-z_{\nu})\right), \label{Gdelta}
\end{equation}
which set the conditions
\begin{eqnarray}
\bar{\partial}\beta_k(z) -2\pi \sum_{\nu =1}^{N}p_k^{\nu }\,\delta
^{2}(z-z_{\nu }) =0, \ \ \ \
\partial \bar{\beta}_k(\bar{z})+2\pi \sum_{\nu =1}^{N}\bar{p}_k^{\nu
}\,\delta ^{2}({z}-{z}_{\nu }) =0 ,
\end{eqnarray}
for each $k$.
These $2(n-1)$ equations have solution only if the sum of the momenta $\{ p_k\}^{\nu}$ vanishes. More precisely, being meromorphic on the Riemann sphere, the sum of the residues of $\beta_k(z)$ vanishes, and therefore
\begin{equation}
\sum_{\nu=1}^{N}p_k^{\nu}=0. \label{U37}
\end{equation}
In order to write the $\beta_k(z)$ in terms of its residues $\{ p_k^1$, $p_k^2$, ... $p_k^N \}$, one may resort to the representation $\bar{\partial}({1}/{z}) =\partial({1}/{\bar{z}})= 2\pi\delta^2(z)$ and integrate. Since $\beta_k(z)$ are 1-differentials, the general solution can be written as a rational function; namely
\begin{equation}\label{U37b}
\beta_k(z)\ =\ \sum_{\nu=1}^N\frac{p_k^{\nu}}{z-z_{\nu}}\ \equiv\ \frac{{\mathcal P}_k(z,\{ y_a\}^{k} )}{{\mathcal Q}(z,\{ z_{\nu } \} )},
\end{equation}
with ${\mathcal P}_k(z,\{ y_a\}^{k} )$ being $n-1$ polynomials of degree $N-2$ in $z$, and ${\mathcal Q}(z,\{ z\}_{\mu } )$ being a polynomial of degree $N$ in $z$. In fact, for a meromorphic 1-differential $\beta_k(z)$ on the Riemann sphere, the difference between the amount of its poles $\{y_a \}^k$ ($a=1, 2, ... \ $) and the amount of its zeroes $\{z_{\mu} \}$ ($\mu=1, 2, ... \ $) is $2$. Therefore, these polynomials have the form
\begin{eqnarray}
{\mathcal P}_k(z,\{ y_a\}^{k} ) = \kappa_k \prod_{a=1}^{N-2}(z-y^k_a) \ , \ \ \
{\mathcal Q}(z,\{ z_{\nu } \} ) = \prod_{\nu =1}^N (z-z_{\nu}).
\end{eqnarray}
From this, using (\ref{U37}) it is easy to show that $\kappa_k = \sum_{\nu = 1}^{N}p_{\nu }^{k}z_{\nu}$. This follows from multiplying (\ref{U37}) by ${\mathcal Q}(z,\{ z_{\nu } \} ) $ to obtain an identity between polynomials and then matching the coefficients of powers of $z$. More precisely, one finds
\begin{eqnarray}
{\mathcal P}_k(z,\{ y_a\}^{k} ) = \sum_{\nu =1}^{N} \prod_{\mu \neq \nu }^{N} p_{k}^{\nu } (z-z_{\mu }) \equiv \sum_{n =1}^{N-1} c_{n}\ z^{N-n} \ ,
\end{eqnarray}
which yields
\begin{eqnarray}
0=c_{1} = \sum_{\mu=1}^{N} p_{k}^{\mu } \ ,\ \ \ \ \ \ \ \kappa_k = c_{2} = -\sum_{\mu \neq \nu }^{N} p_{k}^{\nu }z_{\mu } = \sum_{\mu=1}^{N} z_{\mu } p_{k}^{\mu } .
\end{eqnarray}
It is easy to keep track of the dependence on $\kappa_k$. This is gathered by an overall factor $\prod_{k=1}^{n-1}|\kappa_{k} |^{2\lambda_k}$ in the final result. This can be seen by shifting the fields as $\phi \to \phi -\sum_{k=1}^{n-1}\sqrt{2}b^{-1}m_{k} \omega_{k} \log \kappa_{k}$ to absorb other dependence of $\kappa_k$. Then, we can omit the explicit dependence on $\kappa_k$ and restore it in the final result (\ref{relacion}).
By evaluating the residues, one finds
\begin{equation}
p_k^{\nu}\ =\ \kappa_{k} \frac{\prod_{a=1}^{N-2}(z_{\nu}-y^k_a)}{\prod_{\mu\neq\nu}^N
(z_{\nu }-z_{\mu })}. \label{Gusala}
\end{equation}
Taking all this into account, the $\delta $-function (\ref{Gdelta}) can be replaced by
\begin{equation}
\delta^2\left(\sum_{\nu=1}^Np_k^{\nu}\right)\,\prod_{k=1}^{n-1}\delta^2\left(\beta_k -
\frac{{\mathcal P}_k(z,\{ y_a\}^{k} )}{{\mathcal Q}(z,\{ z_{\mu } \} )}
\right).
\end{equation}
provided the determinant factor ${\det ^{1-n} \square }$ is trivial on the sphere (with $\square = \partial\bar{\partial}$). Then, one can integrate over $\beta_k $ and $\bar{\beta}_k$ to obtain
\begin{equation}\begin{split}
\Omega\left(\{p_k^{\nu}\}, \{\alpha_{\nu }\}|\{ z_{\nu}\} \right)\ =\ \int\prod_{k=1}^{n-1}\mathcal{D}\phi_k \,e^{-S_{\text{eff}}} \ \prod_{\nu=1}^N{\mathcal N}_{\nu }\ e^{{2}(\alpha^{\nu},\phi (z_{\nu }))}. \label{Gfinal}
\end{split}\end{equation
with the effective action
\begin{equation}
S_{\text{eff}} = -\frac{1}{2\pi}\int d^2z \bigg((\partial\phi,
\bar{\partial}\phi)+\frac{(Q_{\{m_k\}},\phi)}{4}\mathcal{R}\,+\,b^2\sum_{k=1}^{n-1} { | {\mathcal P}_k(z,\{ y_a\}^{k} ) / {\mathcal Q}(z,\{ z\}_{\mu } )} |^{2m_k}
e^{\sqrt{2}b(e_k,\phi)}\bigg) , \label{Gfinal2}
\end{equation}
which depends on the $N$ inserting points $\{ z_{\mu } \}$ and the $(n-1)(N-2)$ variables $\{ y_{a}^{k}\}$, although the latter are determined by the $(n-1)N$ algebraic equations (\ref{Gusala}) with the $n-1$ constraints (\ref{U37}).
For further purpose, in order to achieve a regularization of divergence coming from coincident points in the path integral, it is convenient to write the action in the conformal frame $ds^2\ =\ e^{2\sigma }dz\,d\bar{z}$
which yield $({1}/{4})\sqrt{g}\mathcal{R} = -2 \square \sigma $. This will enable us to regularize the propagator $\langle \phi (z)\phi(w) \rangle $ in the limit $z\to w$.
As previously said, the normalization of the vertices ${\mathcal N}$ in (\ref{Gmoco}) can in principle depend on the momenta $\{ p_1, \ p_2, \ ... \ p_{n-1} \} $ and $\alpha $. For instance, the natural generalization of the normalization considered in \cite{Ribault-family} to the $sl(n\geq 2)$ case would be
\begin{equation}
{\mathcal N_{\nu}} = \prod_{k=1}^{n-1}|p^{\nu }_k|^{\frac{2m_k}{b}(\alpha^{\nu},\omega_k)}, \label{Cucote}
\end{equation}
which, after $\delta$-function evaluation and considering (\ref{Gusala}), would produce in (\ref{Gfinal}) a factor
\begin{equation}
\prod_{\nu =1}^{N}\prod_{k=1}^{n-1}\bigg| \frac{{\mathcal P}_k(z_{\nu },\{ y_a\}^{k} )}{{\mathcal Q}(z_{\nu } ,\{ z_{\mu } \} )}\bigg|^{\frac{2m_k}{b}(\alpha_{\nu} ,\omega_k )}\
\end{equation}
accompanying the $N$ vertices. On the other hand, a convenient normalization also seems to be demanding ${\mathcal N}_{\nu }$ not to depend on the momenta at all. In order to consider a more general case that, in particular, includes (\ref{Cucote}), we might consider ${\mathcal N_{\nu}} = \prod_{k=1}^{n-1}|p^{\nu }_k|^{\frac{2m_k}{b}(\alpha^{\nu},\omega_k)(1-t)}$ with $0\leq t \leq 1$. This would affect the final result only in the exponent of the factor that we call $\Theta $ in (\ref{Tuto}) below.
In order to absorb the factors $| {\mathcal P}_k(z,\{ y_a\}^{k} ) / {\mathcal Q}(z,\{ z\}_{\mu } ) |^{2m_k}$ in the terms in the effective action (\ref{Gfinal2}), it is convenient to shift the scalar field $\phi $ -- which actually is a vector in the space generated by the simple roots $\{ e_{\ell } \}$ ($\ell =1, 2, ... n-1$) -- by defining the new field $\varphi $
\begin{equation}
\varphi (z)\ \equiv \phi (z) \,+\,\sum_{k=1}^{n-1}\omega_k\,\frac{\sqrt{2}m_k}{b}\left(
\log \vert { {\mathcal P}_k(z,\{ y_a\}^{k} ) } /{ {\mathcal Q}(z,\{ z_{\nu }\} )}\vert -\,\sigma \right). \label{Gshifting}
\end{equation}
That is
\begin{equation}
\phi (z)\ = \varphi (z)\,-\,\sum_{k=1}^{n-1}\omega_k\,\frac{\sqrt{2}m_k}{b}\left(\sum_{a=1}^{N-2}\log|z-y^k_a|^2\,-\,
\sum_{\nu=1}^N\log|z-z_{\nu}|^2\,-\,\sigma \right). \label{definito}
\end{equation}
Indeed, this produces the rescaling in the exponential potential terms, yielding
\begin{equation}
e^{\sqrt{2}b(e_k,\phi(z))}\ = \ e^{\sqrt{2}b(e_k,\varphi(z))}
\bigg| \frac{ {\mathcal Q}(z,\{ z_{\nu } \} ) }{ {\mathcal P}_k(z,\{ y_a\}^{k} ) }\bigg|^{2m_k} e^{2m_k \sigma } \ ,
\end{equation}
which cancels the factors $| {\mathcal P}_k(z,\{ y_a\}^{k} ) / {\mathcal Q}(z,\{ z\}_{\mu } ) |^{2m_k}$. It also produces the rescaling
\begin{equation}
{\mathcal N_{\nu}}\ e^{\sqrt{2}(\alpha_{\nu},\phi(z_{\nu}))}\ = \ e^{\sqrt{2}(\alpha_{\nu},\varphi(z_{\nu}))}
\prod_{k=1}
^{n-1}
\bigg| \frac{ {\mathcal Q}(z,\{ z_{\mu }\neq z_{\nu } \} ) }{ {\mathcal P}_k(z,\{ y_a\}^{k} ) }\bigg|^{\frac{2tm_k}{b} (\alpha_{\nu} , \omega_k )},\label{moCo}
\end{equation}
where the divergences coming from coincident points is regularized by defining the finite part of $\underset{z_{\mu}\rightarrow z_{\nu}}{\lim}\log|z_{\mu }-z_{\nu}|^2 $ absorbing the divergence in the exponent of the conformal factor $ -2\sigma $, \cite{Polyakovo}. Let us consider the normalization (\ref{Cucote}), corresponding to $t=0$.
Definition (\ref{definito}) also produces a shifting in the kinetic term; namely
\begin{equation}
(\phi,\square \phi)\, = \,(\varphi,\square \varphi)\,-\sum_{k=1}^{n-1}\frac{m_k}{\sqrt{2}b}(\omega_k,a_k) +
\sum_{j=1}^{n-1}\sum_{k=1}^{n-1} \frac{m_j m_k}{2b^2} (\omega _j , b_{jk})
\end{equation}
with
\begin{eqnarray}
a_k(z, \{z_{\mu}\}, \{y_{a}\}^{k}) &=&
\varphi (z)\ \left(2\pi\sum_{a=1}^{N-2} \delta^2(z-y_a^k)-2\pi \sum_{\nu=1}^N\delta^2(z-z_{\nu})- 2\square \sigma \right) + \nonumber \\
&&
\ \partial\bar{\partial}\varphi (z) \ \left(\sum_{a=1}^{N-2}\log|z-y_a^k|^2-\sum_{\nu=1}^N
\log|z-z_{\nu}|^2-2\sigma \right) \label{GasA}
\end{eqnarray}
and
\begin{eqnarray}
b_{jk}(z, \{z_{\mu}\}, \{y_{a}\}^{k}, \{y_{a}\}^{j}) &=& 2\pi\omega_k \ \left(\sum_{a=1}^{N-2}\log|z-y_a^k|^2-\sum_{\mu=1}^N\log|z-z_{\mu}|^2-2\sigma \right)\,\times \nonumber \\
&&\times \left(\sum_{b=1}^{N-2}\delta^2(z-y_b^j)-\sum_{\nu=1}^N\delta^2(z-z_{\nu})\right) - \label{GasB} \\
&& 2\omega_k \left(\sum_{a=1}^{N-2}\log|z-y_a^k|^2-
\sum_{\nu=1}^N\log|z-z_{\nu}|^2-2\sigma \right) \square \sigma . \nonumber
\end{eqnarray}
Considering again the regularization of coincident points, and replacing the last equation into the effective action (\ref{Gfinal2}), one obtains the following contributions in the correlation function: From the first two terms (\ref{GasA}), one obtains
\begin{equation}
\prod_{k=1}^{n-1}\prod_{i=1}^{N-2}e^{-\frac{\sqrt{2}m_k}{b}(\omega_k,\varphi(y_i^k))} \times \prod_{\nu=1}^Ne^{\frac{\sqrt{2}}{b}\sum_{k=1}^{n-1}m_k(\omega_k,\varphi(z_{\nu}))}.
\end{equation}
From the third term, and taking into account $({1}/{4})\sqrt{g}\mathcal{R} = -2 \square \sigma$, one observes that the background charge term in the action gets shifted as $Q_{\{m_k\}}\to Q_{\{m_k\}}-\delta Q = Q_T$. The terms (\ref{GasB}), after a proper regularization of coincident insertions, yields the factor
\begin{eqnarray}
\vert \Theta (\{ m_k \} | \{ z_{\mu } \} , \{ y_a^k \} ) \vert &=& \prod_{a=1}^{N-2} \prod_{b=1}^{N-1} \prod_{k=1}^{n-1} \prod_{l=1}^{n-1} |y_{a}^{k}-y_{b}^{l}|^{\frac{2m_k m_l}{b^2}(\omega_k , \omega_l )}
\prod_{\mu < \nu }^{N}
|z_{\mu }-z_{\nu }|^{\frac{2}{b^2}\sum_{k=1}^{n-1}\sum_{l=1}^{n-1}m_k m_l(\omega_k , \omega_l )} \times \nonumber \\
&& \times \prod_{a=1}^{N-2} \prod_{k=1}^{n-1} \prod_{\mu =1}^{N}
|y_{a}^{k}-z_{\mu }|^{-\frac{2m_k}{b^2}\sum_{l=1}^{n-1}m_l(\omega_k , \omega_l )} \ . \label{Tuto}
\end{eqnarray}
Notice that this contribution corresponds exactly to the correlator of exponential vertices of a free boson $\chi $ with a background charge \cite{Giribeto}; namely
\begin{equation}
\Theta (\{ m_k \} | \{ z_{\mu } \} , \{ y_a^k \} ) = \left\langle \prod_{\mu =1}^{N} V_{ -\sum_{k=1}^{n-1} q_k}(z_{\mu }) \prod_{a=1}^{N-2}\prod_{l=1}^{n-1} V_{q_l}(y_{a }^l) \right\rangle_{\chi }
\end{equation}
where
\begin{equation}
V_{P }(z) \ = \ e^{2iP \chi (z)} ,
\end{equation}
and where $q_{\ell }=m_{\ell }\omega_{\ell }/(\sqrt{2}b)$, Wick contracted the free field propagator $\langle \chi (z) \chi (w) \rangle \ = \ -(1/2) \log (z-w)$. The value of the background charge $Q_{\chi }$ associated to the field $\chi (z)$ can be read from the conservation law that would follow from the integration over the zero-mode $\langle \chi \rangle $. This yields the result $Q_{\chi }=\delta Q$.
Therefore, we finally arrive to the conclusion that the $N$-point correlation functions (\ref{Cosillas}) take the form
\begin{equation}\begin{split}
\left\langle \prod_{\nu =1}^{N}\Phi_{\alpha_{\nu}}(\mu _{\nu }|z_{\nu})\right\rangle _{\{m_k \} } =
\ \prod_{k=1}^{n-1}|\kappa_{k} |^{2 \lambda_k} \prod_{k=1}^{n-1} & \delta\left(\sum_{\nu=1}^Np_k^{\nu}\right)\,
\left\langle \prod_{\mu =1}^{N} V_{-\sum_{k=1}^{n-1} q_k}(z_{\mu }) \prod_{a=1}^{N-2}\prod_{l=1}^{n-1} V_{q_l}(y_{a }^l) \right\rangle_{\chi } \,\times\\
&\times\, \left\langle\prod_{\nu=1}^NV_{\alpha_{\nu}+\sum_{k=1}
^{n-1}q_k}^T(z_{\nu})\prod_{l=1}^{n-1}\prod_{a=1}^{N-2}V_{-q_l}^T (y^l_a)\right
\rangle_{\text{TFT}}\label{relacion}
\end{split}\end{equation}
where $\lambda_k = m_k (1+b^{-2}-\sum_{j=1}^{n-1}m_jb^{-2}(\omega_k , \omega_j ))$; notice that here we have reintroduced the dependence on $\kappa_k$. In other words, $N$-point correlation functions $\Omega (\{p^{\mu}_k\},\{\alpha_{\mu}\}|\{z_{\mu}\})$ of the theories defined by the Lagrangian representation (\ref{nuevaaccion}) turn out to be given by $((N-2)n+2)$-point correlation functions of the CFT given by the product of $sl(n)$ TFT times a free $U(1)$ boson. Having the expression (\ref{relacion}) for the correlation functions $\Omega (\{p_k^{\nu}\}, \{\alpha ^{\nu}\} |\{z_{\nu }\})$, the $m$-deformed CFTs (\ref{nuevaaccion}) are defined {ipso facto}. In the next section, we will analyze the symmetries of the new theories.
\subsection{Correlation functions: Coulomb gas}
Lagrangian representation (\ref{nuevaaccion}) enables us to compute correlation functions in the so-called Coulomb gas approach and thus provide an integral representation of $N$-point correlation functions. This amounts to consider the expectation values (\ref{Cosillas}) and first integrate over the zero-modes of the fields $\phi_k$. This yields
\begin{equation}
\Omega\left(\{p_k^{\nu}\}, \{\alpha^{\nu}\} |\{ z_{\nu}\} \right) = \prod_{l =1}^{n-1} (-1)^{m_l s_l}\Gamma(-s_l) \int \prod_{k=1}^{n-1} \prod_{r =1}^{s_k} d^2w_{r}^{k}
\ \left\langle \prod_{\nu =1}^{N}\Phi_{\alpha_{\nu}}(\{p_k\}^{\nu} |z_{\nu
}) \prod^{n-1}_{k=1} \prod_{r = 1}^{s_k} S^{k}(w^k_{r})
\right\rangle_{\text{free} } \nonumber
\end{equation}
where $\{w_{r}^{k}\} = \{ w_{1}^{1}, w_{2}^{1}, ...\ w_{s_1}^{1};w_{1}^{2}, w_{2}^{2},...\ w_{s_2}^{2}; ... \ ;w_{1}^{{n-1}}, w_{2}^{{n-1}}, ... w_{s_{n-1}}^{{n-1}} \}$ are complex variables, where the screening operators $S^{k}(w)$ are given by
\begin{equation}
S^{k}(w)\equiv \beta^{m_{k}}{(w)}\bar{\beta}^{m_{k}}{(\bar{w})} e^{\sqrt{2}b(e_k , \phi(w))} \ ,
\end{equation}
and where the expectation value $\langle ... \rangle _{\text{free}}$ is defined in terms of the free action
\begin{equation}\begin{split}
S_{\text{free}}\ =\ \frac{1}{2\pi }\int d^{2}z
\bigg( (\partial \phi, \bar{\partial}\phi) \,+\,\sum_{k=1}^{n-1}\left(\beta_k\bar{\partial}\gamma_k\,+\
\bar{\beta}_k\partial \bar{\gamma}_k\right)\,+\,\frac{(Q_{\{m_k\}},\phi)}{4}\mathcal{R} \bigg) \ . \label{Fret}
\end{split}\end{equation}
The dependence on $\{m_k\}$ is through the number of integrals to be performed, and it is given by
\begin{equation}
s_k = {\sqrt{2}}{b}^{-1}( \Delta , \omega_k) \ , \ \ \ \ \Delta = \sqrt{2} (b+b^{-1})\rho -\sqrt{2}b^{-1}\sum_{l=1}^{n-1} m_l\omega_l -\sum_{i=1}^{N} \alpha_i . \label{conference52}
\end{equation}
This equation, as well as the $\Gamma (-s_l)$ factors above, comes from the $\delta$-functions arising in the integration over the zero-modes $\langle \phi_k \rangle$, what selects in the path integral the terms satisfying the vector equation
\begin{equation}
\sqrt{2}\sum_{i=1}^{N} \alpha_i + \sum_{k=1}^{n-1} (s_k e_k b + {2}b^{-1}m_k \omega _k) -{2} (b+b^{-1})\rho = 0 .
\end{equation}
The integrals are over the complex hyperplane $\mathbb{C}^{\sum_{l=1}^{n-1}s_l}$, where $w_{r}^{k}$ represent $s_1 + s_2 + ...\ s_{n-1}$ complex integration variables, with $k=1$,$2$,$...$ $n-1$ and $r=1$,$2$,$...$ $s_k$.
The fact that the expectation value is now defined in terms of action (\ref{Fret}) enables one to use the free field propagators. In particular, one has the Coulomb propagator $\langle \phi_{\ell} (z_i)\phi_k (z_j)\rangle = -(\delta_{\ell , k}/2)\log (z_i -z_i)$. For the set of correlators whose vertices obey the kinematic condition $p_k^{\mu}=\bar{p}^{\mu}_k=0$ for all $k$ such that $m_k\neq 0$ and for all $\mu =1, 2, ... \ N$, one eventually finds
\begin{eqnarray}
\Omega\left(\{p_k^{\nu}\}, \{\alpha^{\nu}\} |\{ z_{\nu}\} \right) = \prod_{\mu =1}^N{\mathcal N}_{\mu } \ \prod_{l=1}^{n-1} (-1)^{s_lm_l}\Gamma(-s_l) \prod_{\mu <\nu }^{N} |z_{\mu } -z_{\nu }|^{-4(\alpha _{\mu } , \alpha_{\nu } )}\ I_N(\{\alpha_{\mu}\},\{m_k\}|\{z_{\mu}\}) \nonumber
\end{eqnarray}
with
\begin{eqnarray}
I_N(\{\alpha_{\mu}\},\{m_k\}|\{z_{\mu}\})&=&\int \prod_{k=1}^{n-1}\prod_{r =1}^{s_k} d^2w_{r}^{k} \prod_{k=1}^{n-1} \prod_{t=1}^{s_k} \prod_{l=1}^{s_{k+1}} | w^k_{t} - w^{k+1}_{l} |^{2b^2} \prod_{k=1}^{n-1} \prod_{r < t}^{s_k} |w^k_{r} -w^k_{t} |^{-4b^2} \times \nonumber \\
&&
\times \prod_{\mu =1}^{N} \prod_{k=1}^{n-1} \prod_{r= 1}^{s_k} |w^k_{r}-z_{\mu }|^{-2\sqrt{2}b(e_{k},\alpha_{\mu })}
\label{integrator}
\end{eqnarray}
where $K_{ij}=(e_i , e_j)$ has been used. In the cases in which for some value of $k$ it happens that $p_k\neq 0\neq m_k$, then a similar representation exists, although the combinatorics when performing the Wick contraction is more involved because of the $\beta $-$\gamma$ systems, which in particular yield $\beta_i (w)e^{[p,\gamma (z)]}\simeq p_i e^{[p,\gamma (w)]}/(w-z)+...$. This is similar to the contraction between string theory tachyon and graviton vertices. The Wick contraction for these operators in the case $n=2$ is discussed in \cite{Schomerus}. The same can be applied here. For instance, in the case of the 2-point function the Wick contraction of the holomorphic piece of the $\beta$-$\gamma$ system yields a factor
\begin{equation}
\prod_{k=1}^{n-1} \bigg( \prod_{i=1}^{2}\prod_{r=1}^{s_{k}} (z_i - w_r^k)^{-1} \bigg( \prod_{l=1}^{s_k} (-w_l^k) (p_1+p_2)^{s_k} + ... \ \bigg) \bigg)
\end{equation}
where the ellipses represent contributions with less $w$'s factors. Using (\ref{U37}), one observes that, eventually, only a contribution proportional to
\begin{equation}
\propto \prod_{k=1}^{n-1}\prod_{i=1}^{2}\prod_{r=1}^{s_{k}} |z_i - w_r^k|^{-2} \label{tormentita}
\end{equation}
survives. This produces an additional shift in the exponent in the last factor of (\ref{integrator}).
Integral representation (\ref{integrator}), which is similar to the one that originally appears in the context of Minimal Models \cite{DF}, can be solved in some very special cases using the techniques of Refs. \cite{Fateev-Litvinov, Fateev-Litvinov2}. This yields closed expressions for reflection coefficients, structure constants, and spherical partition function in several CFTs, including non-rational ones \cite{GN3}. Specially in the latter case, an analytic continuation is required as the expression (\ref{integrator}) only makes sense for $s_k\in \mathbb{Z}_{\geq 0}$. Such extension for values $s_k \in \mathbb{C}$ has been successfully carried out in diverse examples, including non-compact timelike CFTs \cite{Gasto}.
Notice that, even when the integral representation (\ref{integrator}) looks very similar to the one that would appear in the analogous computation for TFT, the integrals appearing in both cases are not exactly the same, one of the differences being the amount of integrations to be performed: While for the $m$-deformed theory one finds (\ref{conference52}), the analogous quantity in an $M$-point TFT correlation function requires not $s_k$ but
\begin{equation}
s_k^T = s_k+ 2b^{-2} \sum_{l=1}^{n-1}m_l (\omega_l , \omega_k) = -\sqrt{2}\sum_{i=1}^{M}(\alpha^T_i ,\omega_k ) + 2(b+b^{-1})(\rho ,\omega _k) \label{tormentita2}
\end{equation}
integrals. This is consistent with the shifting in the momenta $\{ \alpha _i \}$ and the presence of additional $M-N=(N-2)(n-1)$ vertices in the TFT correlator on the right hand side of equation (\ref{relacion}), as well as with the shifting $\delta Q$ in the background charge. To see this in a concise example, let us compute the 2-point functions in the theory (\ref{nuevaaccion}) using the integral representation above. That is, let us compute
\begin{equation}
\langle \Phi_{\alpha } (\{p_k\}^1|z_1) \Phi_{\alpha } (\{p_k\}^2|z_2) \rangle _{\{m_k\}} = \frac{R(\alpha )}{|z_1-z_2|^{4\tilde{h}^2_{\alpha}}},
\end{equation}
where $\tilde{h}^2_{\alpha}= (\alpha , Q+\delta Q - \alpha )$ is the conformal dimension. The relevant information here is given by the reflection coefficient $R(\alpha)$. Conformal invariance permits to fix three vertex insertions on the Riemann sphere; as usual, let us fix the inserting points of the vertex operators $\Phi_{\alpha } (\{p_k\}^i|z_i)$ at $z_1=0$ and $z_2=1$, together with one screening operator at $\infty $. Taking into account (\ref{tormentita}), one observes that (\ref{tormentita2}) maps into (\ref{conference52}) if one identifies the TFT momenta $\alpha_i^T$ as follows
\begin{equation}
\alpha_i^T = \alpha_i + \sum_{l=1}^{n-1} q_l , \label{shiftedmomentum}
\end{equation}
which turns out to be in perfect agreement with (\ref{relacion}) for $N=2$. On the one hand, this shows that the Coulomb gas realization above is consistent with the relation between correlators given in (\ref{relacion}). On the other hand, this gives the expression for the reflection coefficients of the CFTs defined by (\ref{nuevaaccion}), which turns out to be given by the TFT analogous quantity \cite{FateevR} evaluated in the shifted momentum (\ref{shiftedmomentum}). The explicit form of TFT reflection coefficient is such that making the replacing (\ref{shiftedmomentum}) results in the shifting $Q_T\to Q$, as expected. This is because the TFT reflection coefficient depends on the momentum\footnote{See Eqs. (1.14)-(1.17) of Ref. \cite{FateevR}. Our convention relates to the one there by performing the changes $\phi \to \phi / \sqrt{2} $, $\alpha^T \to \alpha / \sqrt{2} $, and $Q_T \to \sqrt{2} Q $.} through the combination $2\alpha^T - Q_T$.
\section{Remnant affine symmetry}
\subsection{Remnant symmetry}
Let us consider the case $n=3$ with deformation parameters $m_1$ and $m_2$. It turns out that, remarkably, in that case the theory (\ref{nuevaaccion}) results to be invariant under the symmetry generated by the currents
\begin{eqnarray}
J_1^+ (z) &=& {m_1}^{-1} \beta_1 \ , \ \ \ \ \ J_1^0 (z) = {\sqrt{2}m_1}{b}^{-1}(e_1,\partial\phi)-{2\gamma_1\beta_1}+{m_1}{m_2}^{-1}\gamma_2\beta_2 \nonumber \\
J_2^+ (z) &=& {m_2}^{-1} \beta_2 \ , \ \ \ \ \ J_2^0 (z) = {\sqrt{2}m_2}{b}^{-1}(e_2,\partial\phi)+{m_2}{m_1}^{-1}\gamma_1\beta_1-{2}\gamma_2\beta_2
\label{nuevascorrientes}
\end{eqnarray}
and their anti-holomorphic counterparts $\bar{J}^A_i$ ($A=0,\pm $, $i=1,2$), with the free field correlators
\begin{equation}
\langle \phi _i (z) \phi_{j} (w)\rangle = -\frac{1}{2}\delta_{ij } \log (z-w) \ , \ \ \ \ \langle \beta_i (z) \gamma_j (w) \rangle = \frac{\delta_{ij }}{(z-w)}
\end{equation}
with $i,j\ = 1,2 $. It is possible to verify that the OPE between the interaction term of the action (\ref{nuevaaccion}) and the currents (\ref{nuevascorrientes}) has no singular term up to a total derivatives.
The symmetry algebra is encoded in the singular terms of the OPE between the currents (\ref{nuevascorrientes}). The non-regular OPEs read
\begin{equation}
J_i^+(z)J_j^0(w)\ \sim\ \frac{(3\delta_{ij}-1)}{(z-w)}\ J_i^+(w) \,+\ ... ,\ \ \
J_i^0(z)J_j^0(w)\ \sim\ -\frac{c_{ij}}{(z-w)^2}\,+\ ... ,\nonumber
\end{equation}
and regular otherwise. The coefficients of the central terms of the algebra are given by
\begin{equation}
c_{ij}= m_i m_jb^{-2} (e_i , e_j ) + 4\delta_{ij} -2(m_i^2 + m_j^2){m^{-1}_i m^{-1}_j} |\varepsilon_{ij}| + m_i m_j \sum_{k} m_k^{-2} |\varepsilon_{ik} \varepsilon_{jk} |
\end{equation}
where $\varepsilon_{12}=-\varepsilon_{21}=1$, $\varepsilon_{11} = \varepsilon_{22} = 0$, and $i,j,k=1,2$. Some of these coefficients, however, can be changed by changing the normalization of the currents. In terms of the modes $J_{i,n}^a$, which are defined by $J_{i}^a(z) = \sum_{n\in \mathbb{Z}} J_{i,n}^a\ z^{-n-1}$ (with $i=1,2$ and $a=0,+$), the symmetry algebra reads $[J_{i,n}^{0},J_{j,m}^{0}]=(n/2)c_{ij} \delta_{n+m,0}$, $[J_{i,n}^{+},J_{i,m}^{0}]= 2 J_{i, n+m}^{+}$, $[J_{i,n}^{+},J_{i\neq j,m}^{0}]= - J_{i, n+m}^{+}$, $[J_{i,n}^{+},J_{j,m}^{+}]= 0$.
This can be extended to the $sl(n)$ case, which one can actually verify to be symmetric under the $2n-2$ currents
\begin{eqnarray}
J_k^+ (z)= {m_k}^{-1} \beta_k \ , \ \ \ \ \ J_k^0 (z)= {\sqrt{2}m_k}{b}^{-1}(e_k,\partial\phi)-{3\gamma_k\beta_k}+{m_k}\sum_{l=k-1}^{k+1}{m_l}^{-1}\gamma_l\beta_l
\label{nuevascoientes}
\end{eqnarray}
with $k,l=1,2, ... n-1$ and where $m^{-1}_{0}=m^{-1}_{n}=0$.
Let us denote by $\hat{\mathcal A}_n \oplus \hat{\mathcal A}_n$ the algebra generated by (\ref{nuevascoientes}) and by its anti-holomorphic counterparts. Some properties of this algebra are the following: Algebra $\hat{\mathcal{A}}_n$ is the affine extension of the semi-direct sum of two Abelian Lie algebras ${\mathcal{A}}_n^{+}$ and ${\mathcal{A}}_n^{0}$; that is, $\hat{\mathcal{A}}_n = \hat{\mathcal{A}}_n^+ \oplus _s \hat{\mathcal{A}}_n^0$ with $J^+_k (z)$ and $J^0_k (z)$ generating each of the two pieces respectively. Algebras $\mathcal{A}_n^+$ and $\mathcal{A}_n^0$ are Abelian and of dimension $n-1$. $\mathcal{A}_n^0$ is the Cartan subalgebra of $sl(n)$. While $\hat{\mathcal{A}}_n^+$ is the loop algebra associated to $\mathcal{A}_n^+$, algebra $\hat{\mathcal{A}}_n^0$ is the affine Kac-Moody extension of $\mathcal{A}_n^0$ with non-vanishing central extensions. Algebra $\hat{\mathcal{A}}_n^{+}$ is an ideal of the semi-direct sum $\hat{\mathcal{A}}_n^{+} \oplus _s \hat{\mathcal{A}}_n^{0}$; the latter is not semi-simple. The semi-direct sum $\mathcal{A}^{+}_n \oplus _s \mathcal{A}^{0}_n$ is included in the Borel subalgebra of $sl(n)$, coinciding with the latter in the case $n=2$. This means that the $m$-deformed CFTs defined by (\ref{nuevaaccion}) may exhibit an infinite-dimensional symmetry apart from local conformal invariance.
One could still raise the question as to whether the full $W_n$ symmetry is actually broken. That is, as it happens with conformal symmetry, which is preserved after the introduction of the $\beta $-$\gamma $ system and the shifting $\delta Q$ in the background charge, one may wonder whether the $W^{n>2}(z)$ currents do not suffer from similar modifications and still represent a symmetry of the theory. A naive attempt to construct such modified currents would be shifting the background charge contribution in the $W^3(z)$ current of the $sl(3)$ TFT and adding to it a piece
\begin{equation}
W^3_{\beta \gamma}(w) = \frac{1}{\sqrt{6}} (\partial \beta_1 \partial \gamma_1 - \beta_1 \partial^2\gamma_1 + \partial \beta_2 \partial \gamma_2 - \beta_2 \partial^2\gamma_2),
\end{equation}
which indeed yields
\begin{equation}
W^3_{\beta \gamma}(z)W^3_{\beta \gamma}(w) =\frac{4/3}{(z-w)^6} + \frac{2W_{\beta \gamma}^2(w)}{(z-w)^4} + \frac{\partial W^2_{\beta \gamma}(w)}{(z-w)^3} + ...
\end{equation}
where the ellipses stand for quadratic and simple poles, and where
\begin{equation}
W_{\beta \gamma}^2(z)= \beta_1\partial\gamma_1 + \beta_2\partial\gamma_2
\end{equation}
is the correct contribution of the ghost system to the stress-tensor. However, this direct sum proposal can be seen not to work, the reason being the non-linear nature of the $W_3$ algebra. To the best of our knowledge, there is no evident systematic manner to deform the $W^{n>2}$ currents and find $W$-symmetry in the $m$-deformed CFT. At least in the case of $W_3$ the question about whether such enhanced $W$-symmetry exists is motivated by the fact that the theory seems to have {\it too many} fields for a current algebra such as $\hat{\mathcal A}_3 \oplus \hat{\mathcal A}_3$. One would expect the CFT to be well defined -at least in the sense of the conformal bootstrap- if it has enough symmetry generators\footnote{G.G. thanks Sylvain Ribault for conversations about this point.}, and therefore it is certainly an interesting question whether the model (\ref{nuevaaccion}) exhibits larger symmetry.
A related question is the following: Since the $m$-deformed CFT seems to break the original $W_n$ symmetry to Virasoro symmetry, it would be fully defined only after a complete list of Virasoro primaries together with an algorithm to compute their 3-point point functions are provided. The question arises as to whether the Virasoro primaries considered here form a complete basis or, at least, a sector closed. While equation (\ref{Gmoco}) provides a collection of such primaries whose correlation functions are defined in terms of the TFT observables (\ref{relacion}), the spectrum of Virasoro primary operators could be a priori larger, since in the undeformed theory the $W_n$ module can be decomposed in multiple $W_2$ modules. The situation would be somehow more problematic if the theory happens to exhibit full $W$-symmetry, as in that case it is not sufficient to consider only $W_n>2$ primaries to fully solve the CFT \cite{SexPistols}. This is precisely why providing techniques alternative to the bootstrap, such as the path integral techniques of \cite{Schomerus} is important, specially in the case of non-rational CFTs.
\subsection{Hidden Kac-Moody symmetry}
As suggested in \cite{Babaro}, the existence of this hidden symmetry generated by $\hat{{\mathcal A}}_n\oplus \hat{{\mathcal A}}_n$, which is a subalgebra of $\hat{sl}(3)\oplus \hat{sl}(3)$, invites to look for a generalization of the deformation (\ref{nuevaaccion}) that, if supplemented with the additional fields in order to realize the additional generators, happens to exhibit full $\hat{sl}(3) \oplus \hat{sl}(3)$ affine Kac-Moody symmetry. In order to look for such a theory, let us consider the action
\begin{equation}\begin{split}
S_{\{ m_k; {\delta}\}}\ =\ &\frac{1}{2\pi }\int d^{2}w
\bigg( (\partial \phi, \bar{\partial}\phi) \,+\,\sum_{k=1}^{2}\left(\beta_k\bar{\partial}\gamma_k\,+\
\bar{\beta}_k\partial \bar{\gamma}_k\right)\,+\,\frac{(Q_{T}+\delta Q,\phi)}{4}\mathcal{R}\,+\\&
(-1)^{m_1}\left(\beta_1- \delta \right)^{m_1}\left(\bar{\beta}_1- \bar{\delta } \right)^{m_1}\,e^{\sqrt{2}b(e_1,
\phi)}\,+ (-\beta_2\bar{\beta}_2)^{m_2}\,e^{\sqrt{2}b(e_2,\phi)}\bigg) , \label{HHH}
\end{split}\end{equation}
which is a deformation similar to the one considered before for the case $n=3$ that, apart from the kind of deformation of the type (\ref{nuevaaccion}), also includes a shifting in the ghost field $\beta_1$.
The theory defined by action (\ref{HHH}) may represent a conformal field theory exhibiting a larger algebra that the one generated by the current (\ref{nuevascoientes}) above, provided an adequate relation between $\delta $ and the fields of the theory is prescribed. For instance, if one introduces a third copy of the $\beta $-$\gamma $ system, by adding to (\ref{HHH}) a piece
\begin{equation}
S_{\beta_3 \gamma_3 }=\frac{1}{2\pi }\int d^2z \ \left( \beta_3\bar{\partial }\gamma_3 + \bar{\beta}_3{\partial }\bar{\gamma}_3 \right) , \label{HJK}
\end{equation}
and considers the deformation $\delta = -\gamma_2\beta_3 $, making the fields (\ref{HJK}) to interact, and chooses $m_1=m_2=1$, for which $Q_{T}+\delta Q=b\rho$, then one finds that the action (\ref{HHH}) exhibits a hidden full $\hat{sl}(3)_k\oplus \hat{sl}(3)_k$ affine symmetry with Kac-Moody level $k=b^{-2}+3$ \cite{BershadskyOoguri, Bershadsky}. To see this explicitly, one writes down the $\hat{sl}(3)$ currents
\begin{eqnarray}
J_1^+ (z)&=& \beta_1 \ , \ \ \ \ J_1^0 (z)= {\sqrt{2}}{b}^{-1}(e_1,\partial\phi)-2\gamma_1\beta_1+\gamma_2\beta_2-\gamma_3\beta_3 \nonumber \\
J_2^+ (z)&=& \beta_2+\gamma_1\beta_3 \ , \ \ J_2^0 (z)= {\sqrt{2}}{b}^{-1}(e_2,\partial\phi)+\gamma_1\beta_1-2\gamma_2\beta_2-\gamma_3\beta_3\nonumber
\end{eqnarray}
together with $J_3^+ (z)= \beta_3$ and
\begin{eqnarray}
J_1^- (z)&=& {\sqrt{2}}{b}^{-1}(e_1,\partial\phi)\gamma_1-k\partial\gamma_1-\gamma_3\beta_2-\gamma_1\gamma_1\beta_1
+\gamma_1\gamma_2\beta_2-\gamma_1\gamma_3\beta_3\nonumber \\
J_2^- (z)&=& {\sqrt{2}}{b}^{-1}(e_2,\partial\phi)\gamma_2-(k-1)\partial\gamma_2+\gamma_3\beta_1-\gamma_2
\gamma_2\beta_2\nonumber \\
J_3^- (z)&=& {\sqrt{2}}{b}^{-1}(e_1,\partial\phi)\gamma_3
+{\sqrt{2}}{b}^{-1}(e_3,\partial\phi)\gamma_3
-{\sqrt{2}}{b}^{-1}(e_2,\partial\phi)\gamma_1\gamma_2-k\partial
\gamma_3+\nonumber \\
&&(k-1)\gamma_1\partial\gamma_2-\gamma_1\gamma_3\beta_1\,-\,\gamma_2\gamma_3\beta_2
-\gamma_3\gamma_3\beta_3-\gamma_1\gamma_2\gamma_2\beta_2 ,
\end{eqnarray}
with $b^{-2}=k-3$, and with the free field correlators $\langle \phi_{k} (z)\phi_{\ell } (w)\rangle \sim -({1}/{2})\delta_{k,\ell }\log (z-w)$ and $\langle \beta_i(z)\gamma_j(w)\rangle \sim {\delta_{i,j}}/{(z-w)}$, now with $k,\ell =1,2$ and $i,j = 1,2,3$. It is possible to verify that the OPE between these eight currents and the interaction operators
\begin{equation}
\tilde{S}^1 = (\beta_1 +\gamma_2 \beta _3) (\bar{\beta}_1 +\bar{\gamma}_2 \bar{\beta}_3) e^{\sqrt{2}b(e_1,\phi)} \ , \ \ \ \ \
{S}^2 = \beta _2 \bar{\beta}_2 e^{\sqrt{2}b(e_2,\phi)}
\end{equation}
is regular, up to total derivatives. Notice that, excluding the contribution of the third ghost system (\ref{HJK}), the currents $J_1^0$, $J_2^0$, $J_1^+$, and $J_2^+$ above coincide with the currents (\ref{nuevascorrientes}) in the case $m_1=m_2=1$.
Other deformations of this type, such as $m_1=m_2=b^2$, also enjoy full $\hat{sl}(3)_{\hat{k}}\oplus \hat{sl}(3)_{\hat{k}}$ symmetry, in such case with $\hat{k}=(3k-8)/(k-3)$. However, the case $m_1=m_2=1$ is special: in this case, action (\ref{HHH}) augmented with the system (\ref{HJK}) actually corresponds to the $SL(3,\mathbb{R})$ WZW model at level $k=b^{-2}+3$ written in Wakimoto variables. In other words, $S_{\text{ WZW}} = S_{\{ m_{1,2}=1;\delta =-{\gamma_2 \beta_3 } \}} + S_{\beta_3 \gamma_3 }$. The central charge in this case is given by $c_{\text{WZW}}= 8+24b^{2}=8k/(k-3)$. The relation between the level $k$ and the parameter $b$ of the undeformed TFT is the same as in the Drinfeld-Sokolov Hamiltonian reduction \cite{Drinfeld-Sokolov, BershadskyOoguri, Bershadsky}. To see this explicitly, consider the WZW action
\be
S_{\text{WZW}}[g]=\f{k}{2\pi}\int_{S} d^2 z \, \text{Tr} \big{(}g^{-1}\partial g\, g^{-1}\bar{\partial} g \big{)}+
\f{k}{12\pi}\int_{B} d^3{x} \, \epsilon^{\nu\sigma\kappa}\, \text{Tr}\big{(} \hat{g}^{-1}\partial_\nu \hat{g}\, \hat{g}^{-1}\partial_\sigma \hat{g}\, \hat{g}^{-1}\partial_\kappa \hat{g} \big{)} \label{WZW56}
\ee
where $k$ is the WZW level, $g(z)$ is a group valued field on $S$, $g\in SL(3,\mathbb{R})$; $S$ is a 2-dimensional surface that coincides with the boundary of $B$, i.e. $S=\partial B$. $\hat{g}(x)$ is the extension of $g(z)$ in the 3-dimensional ambient $B$. To parameterize the group element, consider the Jordan-Chevalley decomposition
\begin{equation}
g=e^{-\gamma_1 T_1^- -\gamma_2 T_2^- -(\gamma_3-\frac{1}{2}\gamma_1 \gamma_2) T^-_3} e^{\phi_1 \tilde{T}^0_1+\phi_2 \tilde{T}^0_2} e^{-\bar{\gamma}_1 T^+_1-\bar{\gamma}_2 T^+_2-(\bar{\gamma}_3-\frac{1}{2}\bar{\gamma}_1\bar{\gamma}_2) T^+_3},
\end{equation}
with $T_i^A$ being the generators of $sl(3)$ given by the upper triangular matrices
\bea
T^+_1=\begin{pmatrix} 0 & 1 & 0 \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{pmatrix},\qquad
T^+_2&=&\begin{pmatrix} 0 & 0 & 0 \\ 0 & 0 & 1 \\ 0 & 0 & 0 \end{pmatrix},\qquad
T^+_3=\begin{pmatrix} 0 & 0 & 1 \\ 0 & 0 & 0 \\ 0 & 0 & 0 \end{pmatrix},\nonumber
\eea
together with the lower triangular
\bea
T^-_1=\begin{pmatrix} 0 & 0 & 0 \\ 1 & 0 & 0 \\ 0 & 0 & 0 \end{pmatrix},\qquad
T^-_2&=&\begin{pmatrix} 0 & 0 & 0 \\ 0 & 0 & 0 \\ 0 & 1 & 0 \end{pmatrix},\qquad
T^-_3=\begin{pmatrix} 0 & 0 & 0 \\ 0 & 0 & 0 \\ 1 & 0 & 0 \end{pmatrix},\nonumber
\eea
and the two Cartan elements
\bea
T^0_1&=&\begin{pmatrix} 1 & 0 & 0 \\ 0 & -1 & 0 \\ 0 & 0 & 0 \end{pmatrix},\qquad
T^0_2=\begin{pmatrix} 0 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & -1 \end{pmatrix};\nonumber
\eea
it is convenient to define the basis $\tilde{T}^0_1=({T^0_1+T^0_2})/{\sqrt{2}}$, $\tilde{T}^0_2=({T^0_1-T^0_2})/{\sqrt{6}}$. The holomorphic conserved currents that generate the $\hat{sl}(3)$ affine algebra are
\begin{equation}
J_{i}^A(z)=\text{Tr}(J(z) T^A_i ) \ , \ \ \ \ \ \ \ J(z) = k\partial g g^{-1} \label{currento}
\end{equation}
with $i=1,2,3$ for $A=\pm $ and $i=1,2$ for $A=0$. The anti-holomorphic currents can be written in a similar manner, with $\bar{J}(z) = -kg^{-1}\bar{\partial }g $. The simple roots of $sl(3)$ are
\bea
{e}_1=\frac{1}{\sqrt{2}}\left(1,\sqrt{3}\right),\qquad
{e}_2=\frac{1}{\sqrt{2}}\left(1,-\sqrt{3}\right),
\eea
with fundamental weights
\be
{\omega}_1=\frac{1}{\sqrt{2}}\left(1,\frac{1}{\sqrt{3}}\right),\qquad
{\omega}_2=\frac{1}{\sqrt{2}}\left(1,-\frac{1}{\sqrt{3}}\right) ,
\ee
from which one easily verifies $(\omega_i ,e_j)=\delta_{ij}$ and $(e_1,e_1)=(e_2,e_2)=2$, $(e_1,e_2)=(e_2,e_1)=-1$; the Weyl vector reads $\rho = \omega_1+\omega_2=(\sqrt{2},0)$. Plugging this representation in (\ref{currento}), defining the fields $\beta_i$
\bea
\beta_1&=&-k e^{\sqrt{2}b(e_1,\phi)}\partial \bar{\gamma}_1 -\gamma_2 \beta_3,\nonumber\\
\beta_2&=&-k e^{\sqrt{2}b(e_2,\phi)}\partial \bar{\gamma}_2,\nonumber\\
\beta_3&=&-k e^{\sqrt{2}b(\rho,\phi)} \left(\partial\bar{\gamma}_3-\bar{\gamma}_2 \partial \bar{\gamma}_1\right),
\eea
(similarly for $\bar{\beta }_i$), and taking into account quantum corrections that amount to renormalize $\phi \to \sqrt{2} b \phi $, one eventually verifies that the $SL(3,\mathbb{R})$ WZW action (\ref{WZW56}) takes the form\footnote{Where we also shifted the zero-modes of the fields $\phi_1$ and $\phi_2$ in order to absorb an overall factor $b^{-2}$ in the interaction terms.}
\bea
S_{\text{WZW}}&=&\f{1}{2\pi}\int d^2 z \Bigg( (\partial \phi, \bar{\partial} \phi)+\sum_{i=1}^{3}(\beta_i \bar{\partial} \gamma_i+\bar{\beta}_i \partial \bar{\gamma}_i)+\f{b(\rho ,\phi )}{4} {\mathcal R} - \nonumber\\
&& (\beta_1+\beta_3\gamma_2)(\bar{\beta}_1+\bar{\beta}_3\bar{\gamma}_2) e^{\sqrt{2}b(e_1,\phi)}-\beta_2 \bar{\beta}_2 e^{\sqrt{2}b(e_2,\phi)}+ \text{c.t.} \Bigg). \label{cosito}
\eea
where $\text{c.t.}$ stands for a contact term; more precisely, for a term
\begin{equation}
\text{c.t.} = -\frac{1}{2\pi}\int d^2z \beta_3 \bar{\beta}_3 e^{\sqrt{2}b(\rho,\phi)} . \label{contacto}
\end{equation}
In Ref. \cite{Hikida-sl3}, a different parameterization of the $SL(3,\mathbb{R})$ elements is considered. Such parameterization leads to a more symmetric form of the action (\ref{cosito}), introducing a shift $\delta $ also in the last term what makes the two screening operators to look similar. The parameterization we considered here has some advantage for the Coulomb gas computation. The contact term in \cite{Hikida-sl3} takes, however, exactly the same form as the one here, namely (\ref{contacto}). The same interpretation for such term as the one given in \cite{Hikida-sl3} holds here. Then, we observe that, up contact terms, action (\ref{cosito}) actually agrees with $S_{\{ m_{1,2}=1;\delta =-{\gamma_2 \beta_3 } \}} + S_{\beta_3 \gamma_3 }$.
\section{Discussion}
In this paper, we have constructed an infinite-dimensional family of two-dimensional conformal field theories that admit Lagrangian representation. These theories consist of particular deformations of $sl(n)$ Toda field theories. Such deformations preserve conformal invariance and deform the full $W_n$ symmetry. As a recognition for having being indulgent with the conformal symmetry, we have been left with a remnant infinite-dimensional symmetry $\hat{\mathcal A}_n\oplus \hat{\mathcal A}_n$ which, in the particular case $n=3$, can be enhanced to full $\hat{sl}(3) \oplus \hat{sl}(3)$ affine symmetry if other deformation operators are allowed. We have explored here the simplest cases of deformation of TFT, which basically consist in the most direct extension of the results of \cite{Ribault-family} to the $sl(n)$ case with $n\geq 2$. It would be interesting to explore other types of deformations and their possible physical applications. Some open questions regarding this are the following: First, whether a systematic way of deforming $sl(n)$ TFT is possible such that one obtains a full $\hat{sl}(n)\oplus \hat{sl}(n)$ affine symmetry for $n>3$. This would extend the WZW-Liouville correspondence to higher rank and for more general $sl(2)$ embeddings. Secondly, it would be interesting to have a full understanding of the relation between the Hamiltonian reduction at quantum level and the correspondence between correlators of theories with $W_n$ symmetry and of theories with $\hat{sl}(n)$ affine symmetry. Interesting results in this direction have been obtained recently in \cite{Hikida-sl3}. A third question that remains open is the aforementioned problem of proving whether or not the deformed theory exhibits hidden $W_n$ symmetry. It appears to us that there is no obvious, systematic way of showing this. This could be seen as an obstruction, since the theory seems not to have a symmetry algebra as large as needed to be solved by bootstrap methods. This is precisely why alternative techniques such as the path integral approach of \cite{Schomerus} are important, in particular when dealing with non-rational CFTs.
Lastly, it would be interesting to see whether there exists a concrete application of the deformed TFT to study gauge theories via AGT and its generalization. It turns out that the $m$-deformed theories do offer a potential application within this context, which is the description of defects in the ${\mathcal N}=2$ superconformal $SU(n)$ quiver theories: According to the Wyllard's generalization of AGT conjecture, the Nekrasov partition function of such $SU(n)$ theories is in correspondence with $sl(n)$ TFT correlation functions. In the case $n=2$, it is known how to generalize the correspondence in order to describe not only the partition function but also expectation values of a whole set of surface and loop operators in the gauge theory side \cite{gordos1, gordos2}. Such observables are also given by Liouville correlation functions, but including degenerate fields, namely fields that contain null Virasoro descendants. The vacua of surface operators in the gauge theory are labeled by integer numbers that are in correspondence with the level of the null vectors in the 2-dimensional CFT. Non-fundamental surface operators of this type are believed to exist in generic $SU(n)$ ${\mathcal N}=2$ gauge theories too, and the expectation value of such operators would also admit a 2-dimensional CFT description in terms of TFT observables. TFT contains degenerate and semi-degenerate representations, and the possibility of the correlators of the theory (\ref{nuevaaccion}) for the appropriate values of $\{ m_1 , m_2 , ... ,m_{n-1} \}$ describing expectation values of non-fundamental operators in $SU(n)$ gauge theories in certainly interesting. This idea has been discussed in \cite{Giribet-Babaro2} for the case $n=2$, where it was argued that defects in the ${\mathcal N}=2^*$ $SU(2)$ gauge theory could be associated to a theory with affine symmetry. The affine CFT description of surface operators in ${\mathcal N}=2$ theories was suggested in \cite{AT}. Another possibly related result is that of reference \cite{ultimatum}, where, based on previous results \cite{ultimatum2} for the TFT correlation functions, it was argued how the inclusion of a semi-degenerate primary operator in the TFT 3-point function corresponds in the gauge theory side to a particular Higgsing of the non-Lagrangian theory on $S^4$. It would be interesting to make these ideas precise and study the potential applications of these deformed theories within the context of the 2D/4D correspondence.
\[
\]
The authors thank Simone Giacomelli and Sylvain Ribault for interesting remarks. This work has been partially funded by FNRS-Belgium (convention FRFC PDR T.1025.14 and convention IISN 4.4503.15), by CONICET of Argentina, by the Communaut\'{e} Fran\c{c}aise de Belgique through the ARC program and by a donation from the Solvay family. The Centro de Estudios Cient\'{\i}ficos (CECs) is funded by the Chilean Government through the Centers of Excellence Base Financing Program of CONICYT-Chile.
|
\section{Introduction}
Let $S$ be a surface of genus $g$ with $p$ marked points and let $\QD(S)$
be the space of quadratic differentials on $S$. Each element $q \in
\QD(S)$ naturally endows $S$ with a locally Euclidean metric with
isolated conical singularities and linear holonomy restricted to
$\set{\pm \id}$. We also refer to elements of $\QD(S)$ as
half-translation surfaces. There is a natural action of $\ensuremath{\mathrm{SL}(2,\R)}\xspace$ on $\QD(S)$
preserving the signature of the singularities. A half-translation surface
$q$ is called a \emph{Veech surface} if its group of (derivatives of)
affine self-diffeomorphisms is a lattice in $\ensuremath{\mathrm{SL}(2,\R)}\xspace$. Veech surfaces possess
remarkable dynamical properties akin to flat tori, and arise naturally in
the contexts of rational billiards and Teichm\"uller curves in moduli
space \cite{Vee89}.
In this paper, we study $\QD(S)$ from the point of view of simple closed
curves on $S$. On a half-translation surface $q$, a simple closed curve
$\alpha$ either has a unique geodesic representative or there is a
maximal flat cylinder on $q$ foliated by closed geodesics in the homotopy
class of $\alpha$. In the former case, the geodesic representative
$\alpha^q$ of $\alpha$ is a concatenation of saddle connections. We say
that $\alpha$ is a \emph{crooked curve} on $q$ if $\alpha^q$ has at least
two saddle connections whose associated holonomy vectors are not
parallel. We define the $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--\emph{infimal length} of $\alpha$ on $q$
to be
\[
\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q) = \inf_{q' \in \, \ensuremath{\mathrm{SL}(2,\R)}\xspace \, \cdot \, q} l_{\alpha}(q'),
\]
where $l_\alpha(q)$ denotes the geodesic length of $\alpha$ on $q$. A
curve $\alpha$ is crooked on $q$ if and only if $\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q)$ is
positive (see Proposition \ref{polygon}). Our first main result is a
characterization of Veech surfaces in terms of their $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--infimal
length spectra.
\begin{theorem}\label{nscc}
Let $q$ be a half-translation surface. Then $q$ is a Veech surface if
and only if it has no short crooked curves: there is an $\epsilon >0$
such that $\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q) \geq \epsilon$ for every crooked curve
$\alpha$.
\end{theorem}
This result is reminiscent of the no-small-(virtual)-triangles theorem
due to Smillie and Weiss \cite{SW10}. They characterize Veech surfaces as
the half-translation surfaces which possess a positive lower bound on the
areas of Euclidean triangles on $q$ with edges formed by saddle
connections. One advantage of working with simple closed curves is that
they are topological objects; they do not depend on the half-translation
structure, and therefore we can study them at once over the entire
quadratic differential space. In contrast, a saddle connection on $q$
persists only in an open subset of the relevant stratum
of $\QD(S)$.
One of the main tools we use in this paper is the \emph{auxiliary
polygon} $P_\alpha(q)$ associated to a simple closed curve $\alpha$ on
$q$, as introduced by the second author and Webb in \cite{TW15}. The area
of $P_\alpha(q)$ gives an estimate for $\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q)^2$ up to bounded
multiplicative error (see Proposition \ref{polygon}\ref{tw3}). It follows
that Theorem \ref{nscc} is equivalent to the statement that $q$ is a
Veech surface if and only if the \emph{polygonal area spectrum}
\[
\polyspec(q) = \set{\area(P_\alpha(q)) \st \alpha \text{ is a simple
closed curve}}
\]
has a gap above zero. In fact, our arguments will prove a slightly
stronger statement: either there is a gap (exactly when $q$ is a Veech
surface), or this spectrum is dense in a neighborhood $[0,a)$ for some $a
> 0$. This statement also holds for the \ensuremath{\mathrm{SL}(2,\R)}\xspace--infimal length spectrum.
The forward implication of Theorem \ref{nscc} will follow from a
relatively straightforward application of the no-small-virtual-triangles
theorem.
The reverse implication of Theorem \ref{nscc} will make use of the
auxiliary polygon mentioned above, as well as two other ingredients: the
orbit closure theorem of Eskin, Mirzakhani, and Mohammadi \cite{EMM15}
and a rigidity statement for $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--orbits of quadratic differentials due
to Duchin, Leininger, and Rafi \cite{DLR10}. We remark that the orbit
closure theorem is used only to deduce local path connectedness of
$\ensuremath{\mathrm{SL}(2,\R)}\xspace$--orbit closures in strata of half-translation surfaces.
To prove the reverse implication we use an elementary continuity
argument; see Proposition \ref{notveech-dense}. An essential step is
establishing that the auxiliary polygon $P_\alpha(q)$ is continuous in
$q$, with respect to the Hausdorff topology. This is achieved in several
steps. By continuity of the intersection pairing between measured
foliations on $S$, we obtain a continuous function {$\QD^1(S) \to
\ensuremath{\mathscr{C}}\xspace(\R \mathrm{P}^1, \ensuremath{\mathbb{R}}\xspace)$ of the form $q \mapsto i\big(\nu_q^{\frac{\pi}{2}
+\theta}, \alpha\big)$ for any fixed curve $\alpha$ on $S$. The value
$i\big(\nu_q^{\frac{\pi}{2} +\theta}, \alpha\big)$ coincides with the
\emph{width} of the auxiliary polygon in direction $\theta$, and so the
\emph{width function} $w_{P_\alpha(q)}\in \ensuremath{\mathscr{C}}\xspace(\R \mathrm{P}^1,\ensuremath{\mathbb{R}}\xspace)$ of the polygon
is continuous in $q$. Finally, standard results from convex geometry on
centrally symmetric sets yield continuity of $P_\alpha(q)$ itself.
As a further application of continuity of the auxiliary polygon, we
show that $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--infimal length is continuous in $q$; see Proposition
\ref{infimalcontinuity}.
Next consider the \emph{virtual triangle area spectrum}, defined as
follows:
\[
\VT(q) = \set{ \, \abs{u \wedge v} \st u, v \in \hol(q)}
\]
where $\hol(q)$ is the set of holonomy vectors of saddle
connections in $q$. The no-small-virtual-triangles theorem of
\cite{SW10} states that $\VT(q)$ has a gap above zero if and only if
$q$ is a Veech surface. In our second theorem we show further that
$\VT(q)$ resembles the polygonal area spectrum as discussed above:
\begin{theorem} \label{nsvt}
Let $q$ be a half-translation surface. Then $\VT(q)$ either has a
gap above zero (exactly when $q$ is a Veech surface) or is
dense in a neighborhood $[0,a)$ for some $a>0$.
\end{theorem}
In fact, our argument provides a new proof of the ``gap
implies Veech'' direction of the no-small-virtual-triangles
theorem. Moreover, the virtual triangles yielding the
dense subset of $[0,a)$ in the non-Veech case are \emph{based}
virtual triangles; see Section \ref{sec:triangles} and Proposition
\ref{densetriangles}.
The non-Veech case of Theorem \ref{nsvt} is proved very similarly to
Theorem \ref{nscc}.
Finally we have some additional results on polygonal area and
$\ensuremath{\mathrm{SL}(2,\R)}\xspace$--infimal length when $q$
is a Veech surface. The first of these will be derived from the
analogous result for $\VT(q)$ in \cite{SW10}.
\begin{theorem} \label{discreteness}
If $q$ is a Veech surface then $\polyspec(q)$ is a discrete subset
of\/ $\ensuremath{\mathbb{R}}\xspace$.
\end{theorem}
Next, for $a > 0$ define the sets
\begin{eqnarray*}
\PA(a) &=& \set{q \in \QD^1(S) \st \area(\ensuremath{\mathrm{P}}\xspace_\alpha(q)) \geq a \text{
for every crooked curve } \alpha \text{ on } q}, \\
\IL(a) &=& \set{q \in \QD^1(S) \st \ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q) \geq a \text{
for every crooked curve } \alpha \text{ on } q}.
\end{eqnarray*}
We say that two half-translation surfaces are \emph{affinely
equivalent} if they are related by the actions of the mapping class
group and $\ensuremath{\mathrm{SL}(2,\R)}\xspace$.
\begin{theorem} \label{finiteness}
For any $a > 0$, the sets $\PA(a)$ and $\IL(a)$ both contain only finitely
many affine equivalence classes of half-translation surfaces.
\end{theorem}
This result is an application of Theorem 2.2 of \cite{EMM15}, which
classifies the closed $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--invariant subsets of strata of
half-translation surfaces: namely, any such subset is a finite union of
orbit closures.
\subsection*{Acknowledgements} The authors thank Alex Wright for many
helpful conversations regarding $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--orbit closures. We also thank
Jenya Sapir for encouraging us to prove continuity of $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--infimal
length, and the referee for providing helpful feedback. Forester was
partially supported by NSF award DMS-1105765,
and Tao by NSF award DMS-1311834.
\section{Background}
\subsection{Quadratic differentials and half-translation surfaces}
We begin by recalling relevant background regarding quadratic
differentials and half-translation surfaces. For further details, consult
\cite{Str84}.
Let $S$ be a surface of genus $g$ and let $\varrho \subset S$ be $p$
marked points with $3g-3+p \ge 1$. A \emph{half-translation
structure} on $S$ consists of a finite set $\varsigma$ of singular
points on $S$ (possibly including marked points),
together with an atlas of charts to $\ensuremath{\mathbb{C}}\xspace$ defined away from $\varsigma$,
where the transition maps are of the form $z \mapsto \pm z + c$ for some
$c \in \ensuremath{\mathbb{C}}\xspace$. By pulling back the standard Euclidean metric on $\ensuremath{\mathbb{C}}\xspace$, one
obtains a locally Euclidean metric on $S - \varsigma$. We require that
the metric completion of this metric yields $S$ with a singular Euclidean
structure, where every singularity has a cone angle $n \pi$ for some $n
\ge 1$, and those that are not marked points must have cone angle at
least $3\pi$. The atlas determines a preferred vertical direction on $S$,
and we consider this to be part of the data in the half-translation
structure.
One can construct a half-translation surface by taking a finite
collection of disjoint Euclidean polygons in $\ensuremath{\mathbb{C}}\xspace$, with pairs of edges
identified by gluing maps of the form $z \mapsto \pm z + c$ \cite{FM14}.
By a \emph{quadratic differential} $q$ on $S$, we mean a complex
structure on $S$ equipped with an integrable holomorphic quadratic
differential on $S - \varrho$ with finitely many zeros. This means that
$q$ can extended to a meromorphic quadratic differential on $S$ with at
worst simple poles at the marked points. The number of zeros minus the
number of poles of $q$, counted with multiplicity, is $4g-4$. The union
of the set of zeros and poles will constitute the singularities of $q$.
There is a \emph{natural} holomorphic coordinate system $z$ on $S$ such
that, in a neighborhood of a point away from a singularity, $q$ is given
by $q = dz^2$. In a neighborhood of a zero of order $k \ge 1$, $q =
z^kdz^2$, and in a neighborhood of a pole, $q = \frac{1}{z}dz^2$. A zero
of order $k$ has cone angle $(k+2)\pi$ and a pole has cone angle $\pi$.
Thus $q$ induces a half-translation structure on $S$, where the singular
Euclidean metric is given locally by $|dz|^2$. We shall use $q$ to
denote a quadratic differential on $S$, as well as $S$ equipped with the
corresponding half-translation structure. The assumptions on $q$ ensure
that the area of the half-translation surface, equal to $\int_S
\abs{q}dz^2$, is finite.
Let $\QD(S)$ be the space of quadratic differentials on $S$. This is a
complex manifold of dimension $6g-6+2p$, and can be identified with the
cotangent bundle to Teichm\"uller space $\T$ via the natural
projection $\QD(S) \rightarrow \T$ obtained by taking the underlying
complex structure of each half-translation surface. The space
$\QD^1(S)$ of \emph{unit area} quadratic differentials on $S$ can be
identified with the unit cotangent bundle to $\T$.
There is a natural $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--action on $\QD(S)$ defined by post-composing
the coordinate charts to $\ensuremath{\mathbb{C}}\xspace$ by an $\ensuremath{\mathbb{R}}\xspace$--linear transformation. One can
view this action by applying an element ${A \in \ensuremath{\mathrm{SL}(2,\R)}\xspace}$ to a defining set
of polygons for a half-translation surface $q$ to obtain a new
half-translation structure $A \cdot q$, noting that parallel edges of the
same length remain so under $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--deformations. Note that
$\ensuremath{\mathrm{SL}(2,\R)}\xspace$--deformations preserve area, and so $\ensuremath{\mathrm{SL}(2,\R)}\xspace$ also acts naturally on
$\QD^1(S)$.
We will write $e^{i\theta}\cdot q$ as shorthand for
$\begin{pmatrix}\cos\theta & -\sin\theta \\ \sin\theta & \cos\theta
\end{pmatrix}\cdot q$; or in natural coordinates at a regular point,
$e^{i\theta}\cdot q =
e^{i\theta} dz^2$.
\subsection{Geodesic representatives and measured foliations}
Let $\scc$ be the set of (free homotopy classes of) essential simple
closed curves on $S$. For any $\alpha \in \scc$, either $\alpha$ has a
unique geodesic representative on $q$, or there is a unique maximal flat
cylinder on $q$ foliated by the closed geodesics in the homotopy class of
$\alpha$. In the former case, the geodesic representative of $\alpha$ is
a concatenation of \emph{saddle connections} -- embedded geodesic arcs or
loops with endpoints at singularities with no singularities on their
interior. The angle between consecutive saddle connections is always
at least $\pi$ on both sides. We shall use $\alpha^q$ to denote any
geodesic representative of $\alpha$ on $q$.
If $\alpha^q$ is a core curve of a flat cylinder, then we call $\alpha$ a
\emph{cylinder curve} on $q$. Let $\ensuremath{\mathsf{cyl}}\xspace(q)$ denote the set of cylinder
curves on $q$, and $\ensuremath{\widehat{\mathsf{cyl}}}\xspace(q)$ the set of curves whose geodesic
representatives have constant direction on $q$. Any curve in $\scc -
\ensuremath{\widehat{\mathsf{cyl}}}\xspace(q)$ is called a \emph{crooked curve} on $q$.
Let $|dz|^2 = dx^2+dy^2$ be the singular Euclidean metric associated to
$q$. We can consider several notions of length of a curve $\alpha$ on
$q$. The Euclidean length $l_\alpha(q)$ of $\alpha^q$ is given by
integrating $\alpha^q$ with respect to $|dz|$ (in local coordinates).
Integrating $\alpha^q$ with respect to $\abs{dx}$ and $\abs{dy}$ (in
local coordinates) gives the \emph{horizontal} and \emph{vertical}
lengths $l_\alpha^H(q)$ and $l_\alpha^V(q)$ respectively. Finally, define
the $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--\emph{infimal length} of $\alpha$ with respect to $q$ to be
\[
\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q) = \inf_{q' \in \, \ensuremath{\mathrm{SL}(2,\R)}\xspace \, \cdot \, q} l_\alpha(q').
\]
This should be viewed as a measure of length of $\alpha$ with respect to
the $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--orbit of $q$, rather than with respect to $q$ itself.
Using the natural coordinate of $q$, one can pull back the foliation of
$\ensuremath{\mathbb{C}}\xspace$ by lines in the direction $\theta \in \R \mathrm{P}^1$ to obtain a measured
foliation $\nu_q^\theta$ on $q$, where the transverse measure is the
Euclidean distance between leaves. In particular, the horizontal and
vertical directions respectively give rise to the \emph{horizontal} and
\emph{vertical} foliations $\nu_q^H = \nu_q^0$ and $\nu_q^V =
\nu_q^{\pi/2}$. The map $\QD(S) \times \R \mathrm{P}^1 \rightarrow \MF(S)$ defined by
${(q, \theta) \mapsto \nu_q^\theta}$ is continuous, where $\MF(S)$ is the
space of measured foliations on $S$ \cite{HM79}. Let $\MF(q) = \set{t
\cdot \nu_q^\theta \st \theta \in \R \mathrm{P}^1, t \in \ensuremath{\mathbb{R}}\xspace_+}$. Let $\PMF(S)$ and
$\PMF(q)$ be the projectivizations of $\MF(S)$ and $\MF(q)$,
respectively. Note that these sets are invariant under
$\ensuremath{\mathrm{SL}(2,\R)}\xspace$--deformations.
The \emph{geometric intersection number} $i \from \scc \times \scc \to
\ensuremath{\mathbb{R}}\xspace$ extends continuously to ${\MF(S) \times \MF(S)}$. For any curve
$\alpha$, we have $i(\nu_q^H, \alpha) = l_\alpha^V(q)$ and $i(\nu_q^V,
\alpha) = l_\alpha^H(q)$. See \cite{FLP79} for additional background on
measured foliations.
\subsection{Auxiliary polygons}
We now recall the construction from \cite{TW15} of the \emph{auxiliary
polygon} $P_\alpha(q)$ associated to a curve $\alpha$ and a quadratic
differential $q \in \QD(S)$.
To each saddle connection $e$ on $q$, we assign a \emph{holonomy vector}
${v}_e \in \ensuremath{\mathbb{R}}\xspace^2$ which is parallel to it and is of the same length. This
is well defined up to scaling by $\pm 1$. For consistency, we require
that the direction $\theta(v_e)$ (that is, the oriented angle from the
positive $x$--axis to $v_e$) lies in the interval $[0, \pi)$. Consider a
geodesic representative $\alpha^q$ on $q$. If $\alpha$ is a cylinder
curve, we may choose $\alpha^q$ to be a boundary component of the maximal
flat cylinder with core curve $\alpha$. Let $m_e$ be the number of times
$\alpha^q$ runs over the saddle connection $e$, in either direction.
Define
\[
P_\alpha(q) = \set{\sum_e t_e m_e {v}_e \st -\frac{1}{2} \leq
t_e \leq \frac{1}{2}} \subset \ensuremath{\mathbb{R}}\xspace^2,
\]
where the sum is taken over all saddle connections used by
$\alpha^q$.
The set $P_\alpha(q)$ is a convex Euclidean polygon,
unless $\alpha^q$ has constant direction, in which case $P_\alpha(q)$
degenerates to a line segment parallel to $\alpha^q$ of length
$l_\alpha(q)$. In particular, in the case of a cylinder curve, the
definition of $P_\alpha (q)$ does not depend on the choice of
boundary component. Moreover, this construction commutes with
$\ensuremath{\mathrm{SL}(2,\R)}\xspace$--deformations: for all $A \in \ensuremath{\mathrm{SL}(2,\R)}\xspace$ we have $P_\alpha(A \cdot q) = A
\cdot P_\alpha(q)$.
\begin{proposition}[\cite{TW15}]\label{polygon}
Given a curve $\alpha \in \scc$ and $q \in \QD^1(S)$, the auxiliary
polygon $P_\alpha(q)$ defined above satisfies the following:
\begin{enumerate}[label=\textup{(\roman*)}]
\item The perimeter of $P_\alpha(q)$ is $2
l_\alpha(q)$,\label{tw1}
\item $\height(P_\alpha(q)) = l_\alpha^V(q)$ and
$\width(P_\alpha(q))
= l_\alpha^H(q)$, \label{tw2}
\item $\pi \area(P_\alpha(q)) \leq \ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q)^2 \leq
8\area(P_\alpha(q))$, \label{tw3}
\item $\area(P_\alpha(q)) = 0$ if and only if $\alpha \in
\ensuremath{\widehat{\mathsf{cyl}}}\xspace(q)$. \label{tw4} \qed
\end{enumerate}
\end{proposition}
Here, the \emph{perimeter} of a polygon $P$ is the length of its
boundary $\partial P$. In the situation where $P$ degenerates to a line
segment, we view $\partial P$ as a closed path traversing the line
segment once in each direction.
The next property of $P_\alpha(q)$ that we need is really
a statement about centrally symmetric convex polygons in
general. Let $v_1, \dotsc, v_k$ be non-zero vectors in $\ensuremath{\mathbb{R}}\xspace^2$ whose
directions $\theta(v_i)$ satisfy
\[
0 \leq \theta(v_1) \leq \dotsb \leq \theta(v_k) < \pi.
\]
Let $P_k$ be the convex set $\set{ \sum_{i=1}^k t_i v_i \st - 1/2
\leq t_i \leq 1/2}$.
\begin{lemma}\label{tiling}
The points $p_k = \sum_{i=1}^k v_i/2$ and $-p_k$ lie on the boundary
$\partial P_k$. Furthermore,
\begin{enumerate}[label=\textup{(\roman*)}]
\item \label{t1} the boundary arc $\partial^+P_k$, traveling
counter-clockwise from $-p_k$ to $p_k$, is the path obtained by
concatenating the vectors $v_1, \dotsc, v_k$, in order. Similarly,
the boundary arc $\partial^- P_k$ from $p_k$ to $-p_k$ is the
concatenation of $-v_1, \dotsc, -v_k$.
\item \label{t2} $P_k$ admits a tiling by the (possibly degenerate)
parallelograms $v_i \times v_j$, with one copy of $v_i
\times v_j$ for each unordered pair $i \not= j$.
\end{enumerate}
\end{lemma}
\begin{proof}
We proceed by induction on $k$, the case $k=1$ being trivial (with
$P_1$ a line segment).
Recall that the \emph{Minkowski sum} of sets
$A, B \subset \ensuremath{\mathbb{R}}\xspace^2$ is the set $A + B = \set{a + b \st a \in A, b \in
B}$. Observe that $P_k$ is the Minkowski sum of
$P_{k-1}$ with the line segment $S_k$ with endpoints $\pm
v_k/2$. The description of $P_{k-1}$ given by the
induction hypothesis allows us to see this sum clearly. The boundary
of $P_k$ decomposes into four parts: the path $\partial^+ P_{k-1} +
\set{ -v_k/2}$, the path $\partial^- P_{k-1} + \set{ v_k / 2}$,
and two shifted copies of $S_k$; see Figure \ref{polyfig}. It is
the assumption that $\theta(v_k)$ lies between $\theta(v_i)$ and
$\pi$ for all $i$ that ensures that $\partial P_k$ is as
described.
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}
\path (0,0) coordinate (origin);
\path (7,0) coordinate (rightorigin);
\path (1.4,0) coordinate (V1);
\path (.9,.3) coordinate (V2);
\path (.6,.5) coordinate (V3);
\path (.8, 1.3) coordinate (V4);
\path (0, 1.1) coordinate (V5);
\path (-.6, .6) coordinate (V6);
\path (-1.4,0) coordinate (V1-);
\path (-.9,-.3) coordinate (V2-);
\path (-.6,-.5) coordinate (V3-);
\path (-.8, -1.3) coordinate (V4-);
\path (0, -1.1) coordinate (V5-);
\path (.6, -.6) coordinate (V6-);
\path (.6, -.2) coordinate (W-);
\path (-.6, .2) coordinate (W+);
\path ($2*(W-)$) coordinate (V7-);
\path ($2*(W+)$) coordinate (V7+);
\path (origin) ++($.5*(V1-)$) ++($.5*(V2-)$) ++($.5*(V3-)$)
++($.5*(V4-)$) ++($.5*(V5-)$) ++($.5*(V6-)$) coordinate (P6-);
\path ($-1*(P6-)$) coordinate (P6+);
\draw[thick] (P6-) -- ++(W-) -- ++(V1) --
++(V2) -- ++(V3) -- ++(V4) -- ++(V5) -- ++(V6) -- ++(W+) -- ++(W+) --
++(V1-) -- ++(V2-) -- ++(V3-) -- ++(V4-) -- ++(V5-) -- ++(V6-) --
++(W-) -- cycle;
\begin{scope}
\clip (P6-) -- ++(W-) -- ++(V1) --
++(V2) -- ++(V3) -- ++(V4) -- ++(V5) -- ++(V6) -- ++(W+) -- ++(W+) --
++(V1-) -- ++(V2-) -- ++(V3-) -- ++(V4-) -- ++(V5-) -- ++(V6-) --
++(W-) -- cycle;
\foreach \x in {-20,...,40}
{\draw[blue!40]
(-3,.1*\x) -- (3,.1*\x-2);
}
\end{scope}
\filldraw[fill=white, draw=black] (P6-) -- ++(V1) --
++(V2) -- ++(V3) -- ++(V4) -- ++(V5) -- ++(V6) --
++(V1-) -- ++(V2-) -- ++(V3-) -- ++(V4-) -- ++(V5-) -- ++(V6-) --
cycle;
\small
\draw (origin) ++(W-) node[anchor=north] {$S_k$};
\scriptsize
\draw[very thick,blue] (origin) ++(W-) -- ++(V7+);
\filldraw (origin) circle (.4mm) node[anchor=south west] {$0$};
\draw[->,>=angle 90,black] (P6-) -- ++(V1);
\draw (P6-) -- ++(V1) +($-.5*(V1)$) node[anchor=south] {$v_1$};
\draw[->,>=angle 90,black] (P6-) ++(V1) -- ++(V2);
\draw (P6-) ++(V1) ++(V2) +($-.5*(V2)$) +(-0.6,-.15) node[anchor=south] {$v_2$};
\draw[->,>=angle 90,black] (P6-) ++(V1) ++(V2) -- ++(V3);
\draw[->,>=angle 90,black] (P6-) ++(V1) ++(V2) ++(V3) -- ++(V4);
\draw[->,>=angle 90,black] (P6-) ++(V1) ++(V2) ++(V3) ++(V4) --
++(V5);
\draw[->,>=angle 90,black] (P6-) ++(V1) ++(V2) ++(V3) ++(V4) ++(V5) -- ++(V6);
\draw[->,>=angle 90,black] (P6+) -- ++(V1-);
\draw (P6+) -- ++(V1-) +($-.5*(V1-)$) node[anchor=north] {$-v_1$};
\draw[->,>=angle 90,black] (P6+) ++(V1-) -- ++(V2-);
\draw (P6+) ++(V1-) ++(V2-) +($-.5*(V2-)$) +(0.6,.15)
node[anchor=north] {$-v_2$};
\draw[->,>=angle 90,black] (P6+) ++(V1-) ++(V2-) -- ++(V3-);
\draw[->,>=angle 90,black] (P6+) ++(V1-) ++(V2-) ++(V3-) -- ++(V4-);
\draw[->,>=angle 90,black] (P6+) ++(V1-) ++(V2-) ++(V3-) ++(V4-) --
++(V5-);
\draw[->,>=angle 90,black] (P6+) ++(V1-) ++(V2-) ++(V3-) ++(V4-)
++(V5-) -- ++(V6-);
\draw (P6+) +(0.05,-.3) node[anchor=north] {$v_{k-1}$};
\filldraw (origin) circle (.4mm) node[anchor=south west] {$0$};
\draw[very thick,blue] (P6-) ++(W-) -- ++(V7+) (P6+) ++(W-) -- ++(V7+);
\filldraw (P6+) circle (.4mm) node[anchor=south west] {$p_{k-1}$};
\filldraw (P6+) +(W+) circle (.4mm) node[anchor=south west] {$p_k$};
\filldraw (P6-) circle (.4mm) node[anchor=north east] {$-p_{k-1}$};
\filldraw (P6-) +(W-) circle (.4mm) node[anchor=north east] {$-p_k$};
\draw (rightorigin) ++(P6-) -- ++(W-) -- ++(V1) --
++(V2) -- ++(V3) -- ++(V4) -- ++(V5) -- ++(V6) -- ++(W+) -- ++(W+) --
++(V1-) -- ++(V2-) -- ++(V3-) -- ++(V4-) -- ++(V5-) -- ++(V6-) --
++(W-) -- cycle;
\filldraw[fill=white, draw=black] (rightorigin) ++(P6-) ++(W+) -- ++(V1) --
++(V2) -- ++(V3) -- ++(V4) -- ++(V5) -- ++(V6) --
++(V1-) -- ++(V2-) -- ++(V3-) -- ++(V4-) -- ++(V5-) -- ++(V6-) --
cycle;
\draw[black!35] (rightorigin) ++(P6-) ++(W+) -- +(V7-) ++(V1) -- +(V7-)
++(V2) -- +(V7-) ++(V3) -- +(V7-) ++(V4) -- +(V7-) ++(V5) -- +(V7-) ;
\draw[very thick,blue] (rightorigin) ++(P6-) ++(W-) -- ++(V7+)
(rightorigin) ++(P6+) ++(W-) -- ++(V7+);
\small
\draw (origin) ++(1.5,.3) node {$P_{k-1}$};
\draw (P6-) ++(W-) ++(V1) ++(V2) ++(V3) node[anchor=north west] {$P_k$};
\draw[black!35] (rightorigin) ++(P6-) ++(W+) ++(V6) -- ++(V1) --
++(V2) -- ++(V3) -- ++(V4) -- ++(V5);
\draw[black!35] (rightorigin) ++(P6-) ++(W+) ++(V6) ++(V1) -- +(V6-)
++(V2) -- +(V6-) ++(V3) -- +(V6-) ++(V4) -- +(V6-) ;
\draw[black!35] (rightorigin) ++(P6-) ++(W+) ++(V6) ++(V5) -- ++(V1) --
++(V2) -- ++(V3) -- ++(V4) ;
\draw[black!35] (rightorigin) ++(P6-) ++(W+) ++(V6) ++(V5) ++(V1) -- +(V5-)
++(V2) -- +(V5-) ++(V3) -- +(V5-) ;
\draw[black!35] (rightorigin) ++(P6-) ++(W+) ++(V6) ++(V5) ++(V4) --
++(V1) -- ++(V2) -- ++(V3) ;
\draw[black!35] (rightorigin) ++(P6-) ++(W+) ++(V6) ++(V5) ++(V4)
++(V1) -- +(V4-) ++(V2) -- +(V4-) ;
\draw[black!35] (rightorigin) ++(P6-) ++(W+) ++(V6) ++(V5) ++(V4)
++(V3) -- ++(V1) -- ++(V2) ;
\draw[black!35] (rightorigin) ++(P6-) ++(W+) ++(V6) ++(V5) ++(V4)
++(V3) ++(V1) -- +(V3-) ;
\end{tikzpicture}
\end{center}
\caption{The polygon $P_k$ expressed as $P_{k-1} + S_k$,
and tiled by parallelograms $v_i \times v_j$.}
\label{polyfig}
\end{figure}
The point $p_k$ is the initial point of the path $\partial^- P_{k-1}
+ \set{ v_k / 2}$, which is on the boundary. Similarly, $-p_k$ is the
initial point of $\partial^+ P_{k-1} + \set{ -v_k/2}$, and conclusion
\ref{t1} is now clear.
Next divide $P_k$ into two regions, one bounded by $\partial
P_{k-1} + \set{v_k/2}$, and the other bounded by the paths
$\partial^+ P_{k-1} + \set{v_k/2}$, $\partial^+ P_{k-1} +
\set{-v_k/2}$, and the two copies of $S_k$. The first region is
isometric to $P_{k-1}$ and can be tiled by the parallelograms $v_i
\times v_j$, $i < j < k$, by the induction hypothesis. The second
region is the sum $\partial^+ P_{k-1} + S_k$, and is tiled by the
parallelograms $v_i \times v_k$, for all $i < k$, since
$\partial^+ P_{k-1}$ is a concatenation of the vectors $v_i$, $i <
k$. \qedhere
\end{proof}
Since $P_\alpha(q)$ is $P_k$ for the vectors $\set{m_e v_e \st e
\text{ appears in }\alpha}$, ordered appropriately, we immediately
deduce the following from Lemma \ref{tiling}\ref{t2}.
\begin{lemma}\label{area-sum}
For any $\alpha \in \scc$ and $q \in \QD^1(S)$, we have
\[
\area(P_\alpha(q)) = \sum_{e, e'} m_em_{e'} \abs{v_e \wedge v_{e'}},
\]
where the sum is taken over all unordered pairs of distinct saddle
connections $e, e'$ appearing in $\alpha^q$. \qed
\end{lemma}
\subsection{Rigidity}
A key tool that we use is the following result of Duchin, Leininger, and
Rafi. Our formulation of the statement is slightly different from theirs,
but their proof still works without modification.
\begin{proposition}[\cite{DLR10}, Lemma 22]\label{rigidity}
Let $q, q' \in \QD^1(S)$ be half-translation surfaces. If\/
$\ensuremath{\widehat{\mathsf{cyl}}}\xspace(q) \subset \ensuremath{\widehat{\mathsf{cyl}}}\xspace(q')$ then $\ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q = \ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot
q'$. \qed
\end{proposition}
Therefore, if $\ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q \not= \ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q'$ then both
$\ensuremath{\widehat{\mathsf{cyl}}}\xspace(q) - \ensuremath{\widehat{\mathsf{cyl}}}\xspace(q')$ and $\ensuremath{\widehat{\mathsf{cyl}}}\xspace(q') - \ensuremath{\widehat{\mathsf{cyl}}}\xspace(q)$ are non-empty.
\begin{remark}
One consequence of Proposition \ref{rigidity} is that when $3g-3+p >
1$, every $q \in \QD^1(S)$ has a crooked curve. This is true
because $\QD^1(S)$ always has real dimension greater than $3$, and
hence cannot be a single $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--orbit.
\end{remark}
\begin{remark}
Theorem 1 of \cite{DLR10} states that the marked length spectrum of
simple closed curves determines the half-translation surface $q$.
Associated to $q$ is the \emph{marked polygonal area spectrum}, which
is the $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--invariant function $\scc \to \ensuremath{\mathbb{R}}\xspace$ given by $\alpha \mapsto
\area(P_\alpha(q))$. We observe that, by Proposition
\ref{polygon}\ref{tw4} and Proposition \ref{rigidity}, the marked
polygonal area spectrum of $q$ determines its $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--orbit.
\end{remark}
\subsection{$\ensuremath{\mathrm{SL}(2,\R)}\xspace$--orbit closures}
The space $\QD^1(S)$ of unit-area half translation structures on $S$ is
naturally partitioned into strata $\Qk$, where $\kappa$ is a partition of
$4g - 4$ specifying the orders of the singularities. The mapping class
group $\mcg(S)$ acts on $\QD^1(S)$ and each stratum $\Qk$ by change of
marking. The $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--action on $\QD^1(S)$ preserves each stratum, and also
descends naturally to the moduli space of half-translation surfaces
$\QMod(S) = \QD^1(S) / \mcg(S)$, as well as each unmarked stratum $\QkMod
= \Qk / \mcg(S)$. Let $\pi \from \QD^1(S) \to \QMod(S)$ be the natural
projection, which is an orbifold covering map. Given $q \in \Qk$, let
$\ensuremath{\mathrm{M}}\xspace_q$ be the closure of $\pi(\ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q)$ in $\QkMod$. We call
$\ensuremath{\mathrm{M}}\xspace_q$ the \emph{orbit closure} associated to $q$.
The structure of $\ensuremath{\mathrm{M}}\xspace_q$ has been elucidated in the work of Eskin,
Mirzakhani, and Mohammadi, as follows:
\begin{theorem}[\cite{EMM15}, Theorem 2.1]\label{EMM}
For any $q \in \Qk$, the orbit closure $\ensuremath{\mathrm{M}}\xspace_q$ is an affine
invariant submanifold of\/ $\QkMod$. \qed
\end{theorem}
The statement given in \cite{EMM15} actually refers to abelian
differentials rather than quadratic differentials. However, the result
applies equally well to the setting of quadratic differentials, by
considering an appropriate two-fold branched covering of the surface.
The precise definition of ``affine invariant submanifold'' is rather
involved and we shall not repeat it here in its entirety. It
suffices to note that it includes the following:
\begin{itemize}
\item $\ensuremath{\mathrm{M}}\xspace_q$ is $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--invariant,
\item $\ensuremath{\mathrm{M}}\xspace_q$ is the image of a properly immersed orbifold
$f \from N \to \QkMod$.
\end{itemize}
Since $\pi \from \Qk \to \QkMod$ is an orbifold covering, the
preimage $\pi^{-1} (\ensuremath{\mathrm{M}}\xspace_q)$ is also the image of a properly
immersed orbifold (in fact, manifold) of the same dimension as
$N$.
The main conclusion we need to draw from Theorem \ref{EMM} is that
$\pi^{-1}(\ensuremath{\mathrm{M}}\xspace_q)$ is locally path connected. Let $\ensuremath{\widetilde{\mathrm{M}}}\xspace_q$ be the
connected component of $\pi^{-1}(\ensuremath{\mathrm{M}}\xspace_q)$ containing $q$. Then local
path connectedness of $\pi^{-1}(\ensuremath{\mathrm{M}}\xspace_q)$ implies that $\ensuremath{\widetilde{\mathrm{M}}}\xspace_q$ is
open and closed in $\pi^{-1}(\ensuremath{\mathrm{M}}\xspace_q)$.
Now let $\Gamma_q \leq \mcg(S)$ be the (setwise) stabilizer of
$\ensuremath{\widetilde{\mathrm{M}}}\xspace_q$, and define
\[
\ensuremath{\mathcal{O}}\xspace_q = \Gamma_q \cdot (\ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q).
\]
\begin{lemma}\label{EMMlemma}
$\ensuremath{\widetilde{\mathrm{M}}}\xspace_q$ is the closure of\/ $\ensuremath{\mathcal{O}}\xspace_q$ in $\Qk$.
\end{lemma}
\begin{proof}
Certainly, $\ensuremath{\widetilde{\mathrm{M}}}\xspace_q$ contains $\ensuremath{\mathcal{O}}\xspace_q$ and is closed in $\Qk$, so
it contains the closure of $\ensuremath{\mathcal{O}}\xspace_q$. Also, $\pi^{-1}(\ensuremath{\mathrm{M}}\xspace_q)$ is
the closure of $\mcg(S) \cdot (\ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q)$. Note that if $g \in
\mcg(S) - \Gamma_q$ then $g \cdot (\ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q)$ is contained in
the component $g \ensuremath{\widetilde{\mathrm{M}}}\xspace_q$ of $\pi^{-1}(\ensuremath{\mathrm{M}}\xspace_q)$, which is
disjoint from $\ensuremath{\widetilde{\mathrm{M}}}\xspace_q$. Now if $q' \in \ensuremath{\widetilde{\mathrm{M}}}\xspace_q$ is a limit of a
sequence of points $q_i \in g_i \cdot (\ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q)$, the open set
$\ensuremath{\widetilde{\mathrm{M}}}\xspace_q$ must contain almost all $q_i$, and therefore $g_i \in
\Gamma_q$ for almost all $i$, and $q'$ is in the closure of
$\ensuremath{\mathcal{O}}\xspace_q$. \qedhere
\end{proof}
We will also make use of the following finiteness result.
\begin{theorem}[\cite{EMM15}, Theorem 2.2]\label{EMMfinite}
Any closed $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--invariant subset of\/ $\Qk$ is a finite union
of\/ $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--orbit closures. \qed
\end{theorem}
\subsection{Veech surfaces}
Recall that $q\in \QD(S)$ is a \emph{Veech surface} if its group of
affine automorphisms is a lattice in $\ensuremath{\mathrm{SL}(2,\R)}\xspace$. We now state several
characterizations of Veech surfaces due to Smillie and Weiss
\cite{SW10}, which builds on work of Vorobets \cite{Vor96}. By a
\emph{triangle} on $q$, we mean a Euclidean triangle on $q$ with
isometrically embedded interior, whose sides are saddle connections
on $q$. Let $\hol(q)$ be the set of holonomy vectors arising from
saddle connections on $q$.
\begin{theorem}[\cite{SW10}]\label{nst}
For any $q \in \Qk$, the following are equivalent.
\begin{enumerate}[label=\textup{(\roman*)}]
\item $q$ is a Veech surface, \label{sw1}
\item $q$ has no small triangles: there is a lower bound $\epsilon >
0$ on the areas of all triangles on $q$,
\item $q$ has no small virtual triangles: there exists $\epsilon > 0$
such that $\abs{{u} \wedge {v}} > \epsilon$ for all pairs
of non-parallel holonomy vectors $ {u}, {v} \in
\hol(q)$, \label{sw3}
\item the virtual triangle area spectrum $\VT(q) = \set{\abs{{u}
\wedge {v}} \st {u}, {v} \in
\hol(q)}$ is discrete, \label{sw3a}
\item $\pi(\ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q)$ is closed in $\QkMod$. \label{sw4}
\qed
\end{enumerate}
\end{theorem}
Note that condition \ref{sw4} is the same as saying that
$\ensuremath{\mathrm{M}}\xspace_q = \pi(\ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q)$. Applying Lemma \ref{EMMlemma}, we deduce
the following.
\begin{corollary}\label{veechcor}
A half-translation surface $q$ is a Veech surface if and only
if\/ $\ensuremath{\widetilde{\mathrm{M}}}\xspace_q = \ensuremath{\mathcal{O}}\xspace_q$. \qed
\end{corollary}
\section{Continuity results}
\label{sec:continuity}
\subsection{Continuity of the auxiliary polygon}
Here we show that the polygon $P_\alpha(q)$ is continuous in $q$,
with respect to the Hausdorff topology in the plane. First we recall
some basic notions from convex geometry. See, for instance, Sections
1.7 and 1.8 of \cite{Sch14}.
For any non-empty compact convex subset $K \subset \ensuremath{\mathbb{R}}\xspace^2$, the
\emph{support function} $h_K \from S^1 \to \ensuremath{\mathbb{R}}\xspace$ is defined by
\[
h_K(u) = \sup \set{\langle x, u\rangle \st x \in K}.
\]
Here $\langle \, \param \, , \param \, \rangle$ is the usual
inner product on $\ensuremath{\mathbb{R}}\xspace^2$. The \emph{width function} $w_K \from S^1
\to \ensuremath{\mathbb{R}}\xspace$ is defined by
\[
w_K(u) = h_K(u) + h_K(-u).
\]
Note that $w_K$ is even, and descends to a function on $\R \mathrm{P}^1$ which we
also denote by $w_K$. Now let $d_H$ denote Hausdorff distance. We have
the following standard fact:
\begin{lemma}[\cite{Sch14},Lemma~1.8.14]\label{hausdorff}
Suppose $K$ and $L$ are non-empty compact convex subsets of\/
$\ensuremath{\mathbb{R}}\xspace^2$. Then
\begin{equation}
d_H(K,L) = \sup_{u \in S^1} \abs{h_K(u) - h_L(u)}. \tag*{\qed}
\end{equation}
\end{lemma}
Next consider \emph{centrally symmetric} convex sets: these are
convex sets $K$ such that $K = -K$. Note that the auxiliary polygons
$P_\alpha(q)$ are both convex and centrally symmetric. If $K$ is
centrally symmetric then $h_K(u) = h_K(-u)$ for all $u$, and
therefore
\begin{equation}\label{heightwidth}
w_K = 2h_K.
\end{equation}
Now let $\norm{f}$ denote the sup norm for functions $f \from \R \mathrm{P}^1
\to \ensuremath{\mathbb{R}}\xspace$, and let $\ensuremath{\mathscr{C}}\xspace(\R \mathrm{P}^1,\ensuremath{\mathbb{R}}\xspace)$ be the space of continuous
functions, with the sup metric. Let $\ensuremath{\mathscr{K}_0}\xspace$ be the space of
non-empty centrally symmetric compact convex sets in $\ensuremath{\mathbb{R}}\xspace^2$, with
the Hausdorff metric. The next lemma follows directly from Lemma
\ref{hausdorff} and equation \eqref{heightwidth}.
\begin{lemma}\label{hausdorffwidth}
If $K, L \in \ensuremath{\mathscr{K}_0}\xspace$ then
\begin{equation}
2d_H(K,L) = \sup_{\theta \in \R \mathrm{P}^1} \abs{w_K(\theta) - w_L(\theta)} =
\norm{w_K - w_L}. \tag*{\qed}
\end{equation}
\end{lemma}
\begin{corollary}\label{embedding}
The map $W \from \ensuremath{\mathscr{K}_0}\xspace \to \ensuremath{\mathscr{C}}\xspace(\R \mathrm{P}^1,\ensuremath{\mathbb{R}}\xspace)$ given by $K \mapsto
\frac{1}{2}w_K$ is an isometric embedding. \qed
\end{corollary}
Let us now return our attention to the auxiliary polygons.
\begin{theorem}\label{polygoncontinuity}
The map $\QD^1(S) \times \scc \to \ensuremath{\mathscr{K}_0}\xspace$ defined by $(q,
\alpha) \mapsto P_\alpha(q)$ is continuous in the first factor.
\end{theorem}
\begin{proof}
Let $\alpha \in \scc$ be fixed. Applying Proposition \ref{polygon},
we see that
\[
w_{P_\alpha(q)}(\theta) = \width(e^{-i\theta}\cdot P_\alpha(q)) =
l_\alpha^H(e^{-i\theta}\cdot q) = i\big(\nu_q^{\frac{\pi}{2}+\theta},
\alpha\big)
\]
for all $q \in \QD^1(S)$ and $\theta \in \R \mathrm{P}^1$. The map $\QD^1(S) \times
\R \mathrm{P}^1 \to \MF(S)$ given by $(q, \theta) \mapsto
\nu_q^{\frac{\pi}{2}+\theta}$ is continuous. By continuity of
intersection number on $\MF(S) \times \MF(S)$ and compactness of
$\R \mathrm{P}^1$, the map $q \mapsto w_{P_\alpha(q)}$ defines a continuous
function from $\QD^1(S)$ to $\ensuremath{\mathscr{C}}\xspace(\R \mathrm{P}^1,\ensuremath{\mathbb{R}}\xspace)$. Moreover, its image is
contained in $W(\ensuremath{\mathscr{K}_0}\xspace)$, and composing this map with
$\frac{1}{2}W^{-1}$ yields the function $q \mapsto
P_\alpha(q)$. This map is continuous by Corollary
\ref{embedding}. \qedhere
\end{proof}
Finally, applying continuity of $\area \from \ensuremath{\mathscr{K}_0}\xspace \to \ensuremath{\mathbb{R}}\xspace$
\cite[Theorem~1.8.20]{Sch14} yields the desired result.
\begin{corollary}\label{areacontinuity}
The function $\area \from \QD^1(S) \times \scc \rightarrow \ensuremath{\mathbb{R}}\xspace_{\geq 0}$
defined by $\area(q, \alpha) = \area(P_\alpha(q))$ is continuous and
$\ensuremath{\mathrm{SL}(2,\R)}\xspace$--invariant in the first factor. \qed
\end{corollary}
\subsection{Continuity of \ensuremath{\mathrm{SL}(2,\R)}\xspace--infimal length}
In this section, we apply continuity of the auxiliary polygon to deduce
continuity of\/ $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--infimal length.
Given $K \in \ensuremath{\mathscr{K}_0}\xspace$, let $r^-(K) = \inf_{u \in S^1} h_K(u)$ and
$r^+(K) = \sup_{u \in S^1} h_K(u)$. One can show that these two numbers
coincide with the minimum and maximum distances to the origin of
points on $\partial K$. Define the \emph{eccentricity} of $K$
to be $\ecc(K) = \frac{r^+(K)}{r^-(K)} \geq 1$. Note that $r^-(K) =
0$ if and only if $K$ is degenerate.
Recall that any matrix $A \in \ensuremath{\mathrm{SL}(2,\R)}\xspace$ has a \emph{singular value
decomposition}
\[
A \ = \ e^{i\theta_1} \begin{pmatrix}\lambda & 0 \\
0 & \lambda^{-1} \end{pmatrix} e^{i\theta_2},
\]
for some \emph{stretch factor} $\lambda \geq 1$ and $\theta_1,
\theta_2 \in \ensuremath{\mathbb{R}}\xspace$. Moreover, $\lambda = \ensuremath{\lambda}\xspace(A)$ is unique, and
$\ensuremath{\lambda}\xspace(AB) \leq \ensuremath{\lambda}\xspace(A)\ensuremath{\lambda}\xspace(B)$ and $\ensuremath{\lambda}\xspace(A^{-1}) = \ensuremath{\lambda}\xspace(A)$ for all $A,B \in
\ensuremath{\mathrm{SL}(2,\R)}\xspace$. Also note that
\begin{equation}\label{radiuslambda}
\ensuremath{\lambda}\xspace(A)r^-(K) \leq r^+(A \cdot K) \leq \ensuremath{\lambda}\xspace(A)r^+(K) \quad
\textrm{and} \quad
\frac{r^-(K)}{\ensuremath{\lambda}\xspace(A)} \leq r^-(A \cdot K) \leq \frac{r^+(K)}{\ensuremath{\lambda}\xspace(A)}
\end{equation}
for all $K \in \ensuremath{\mathscr{K}_0}\xspace$.
Since $A$ acts as a $\ensuremath{\lambda}\xspace(A)$--Lipschitz map from $\ensuremath{\mathbb{R}}\xspace^2$ to itself, we deduce
for all $K,L \in \ensuremath{\mathscr{K}_0}\xspace$ that
\begin{equation}\label{hdstretch}
d_H(A \cdot K, A \cdot L) \ \leq \ \ensuremath{\lambda}\xspace(A)~ d_H(K,L).
\end{equation}
\begin{lemma}\label{ecchaus}
Fix $r_0 > 0$, and suppose $K,L \in \ensuremath{\mathscr{K}_0}\xspace$ satisfy $r^-(K),
r^-(L) > r_0$. Then
\[
\left(1 + \frac{d_H(K,L)}{r_0}\right)^{-2} < \
\frac{\ecc(K)}{\ecc(L)} \ < \ \left(1 +
\frac{d_H(K,L)}{r_0}\right)^2.
\]
\end{lemma}
\begin{proof}
Let $D = d_H(K,L)$. Applying Lemma \ref{hausdorff} gives
\[
\frac{\ecc(K)}{\ecc(L)} \ = \
\frac{r^+(K)}{r^+(L)}\frac{r^-(L)}{r^-(K)} \ \leq \
\left(\frac{r^+(L) + D}{r^+(L)}\right)\left(\frac{r^-(K) +
D}{r^-(K)}\right) \ < \ \left(1 + \frac{D}{r_0}\right)^2.
\]
The other bound can be deduced similarly. \qedhere
\end{proof}
Define the \emph{eccentricity} of a curve $\alpha \in \scc$ on $q \in
\QD^1(S)$ to be $\ecc_\alpha(q) = \ecc(P_\alpha(q))$. Note that
$\ecc_\alpha(q) = \infty$ if and only if $\alpha \in \ensuremath{\widehat{\mathsf{cyl}}}\xspace(q)$. By
Theorem \ref{polygoncontinuity}, $\ecc_\alpha$ is continuous on the set
of quadratic differentials on which $\alpha$ is crooked.
\begin{lemma}\label{lambdagrowth}
Let $\alpha$ be a crooked curve on $q \in \QD^1(S)$, and suppose the
\ensuremath{\mathrm{SL}(2,\R)}\xspace--infimal length of $\alpha$ is attained at $m \in \ensuremath{\mathrm{SL}(2,\R)}\xspace\cdot q$.
There is a constant $c \geq 1$ (independent of\/ $S$, $q$, and
$\alpha$) such that
\[
\frac{\ensuremath{\lambda}\xspace(A)^2}{c} \ \leq \ \ecc_\alpha(A \cdot m) \ \leq \
c\ensuremath{\lambda}\xspace(A)^2
\]
for all $A \in \ensuremath{\mathrm{SL}(2,\R)}\xspace$. In particular, $\ecc_\alpha(m) \leq c$.
\end{lemma}
\begin{proof}
By Lemma 6.4 and Corollary 7.2 of \cite{TW15}, there exists $A_0 \in
\ensuremath{\mathrm{SL}(2,\R)}\xspace$ such that $\ecc_\alpha(A_0 \cdot m) \leq 2$ and $\ensuremath{\lambda}\xspace(A_0) < c'$
for a constant $c'$ independent of $S$, $q$, and
$\alpha$. Taking $c \geq 2(c')^2$, the result
follows by applying the inequalities in
(\ref{radiuslambda}), with $K = P_\alpha (A_0 \cdot m)$. \qedhere
\end{proof}
\begin{lemma}\label{fellow}
Let $\alpha$ be a crooked curve on $q \in \QD^1(S)$. Then for all
$\epsilon > 0$ and $\lambda_0 \geq 1$, there exists an open
neighborhood $U \subseteq \QD^1(S)$ of $q$ such that for all $q'
\in U$ and $A \in \ensuremath{\mathrm{SL}(2,\R)}\xspace$ satisfying $\ensuremath{\lambda}\xspace(A) \leq \lambda_0$, we have:
\begin{enumerate}[label=\textup{(\roman*)}]
\item $(1+\epsilon)^{-2} < \frac{\ecc_\alpha(A \cdot
q')}{\ecc_\alpha(A \cdot q)} < (1 + \epsilon)^2$, \label{f2}
\item $|l_\alpha(A \cdot q) - l_\alpha(A \cdot q')| < \pi
\epsilon$. \label{f3}
\end{enumerate}
\end{lemma}
\begin{proof}
Set $r_0 = \frac{r^-(P_\alpha(q))}{2\lambda_0}$, and choose $\delta <
\min \set{\frac{r^-(P_\alpha(q))}{2},
\frac{r_0\epsilon}{\lambda_0}, \frac{\epsilon}{\lambda_0}}$.
By Theorem \ref{polygoncontinuity}, there is an open neighborhood $U$
of $q$ with $d_H(P_\alpha(q), P_\alpha(q')) < \delta$ for all $q' \in
U$. Choose any $q' \in U$ and $A \in \ensuremath{\mathrm{SL}(2,\R)}\xspace$ satisfying $\ensuremath{\lambda}\xspace(A) \le
\lambda_0$. Then
\[
d_H(P_\alpha(A \cdot q), P_\alpha(A \cdot q')) < \min
\set{ r_0 \epsilon, \epsilon}
\]
by \eqref{hdstretch}. Now by \eqref{radiuslambda}, we deduce
\[
r^-(P_\alpha(A \cdot q')) \ \geq \
\frac{r^-(P_\alpha(q'))}{\ensuremath{\lambda}\xspace(A)} \
> \ \frac{r^-(P_\alpha(q))-\delta}{\lambda_0} \ > \
\frac{r^-(P_\alpha(q))}{2\lambda_0} \ = \ r_0.
\]
Applying Lemma \ref{ecchaus} yields the first claim. For the second
claim, we use the well-known fact that the perimeter of any convex
region $K \subset \ensuremath{\mathbb{R}}\xspace^2$ is $\int_0^\pi w_K(u) du$, where $w_K(u)$ is
the width of $K$ at $u \in S^1$. Set $w_q(u)$ to be the width of
$P_\alpha(q)$ at $u$. Then using Proposition
\ref{polygon}\ref{tw1} and Lemma \ref{hausdorffwidth} we have
\begin{align*}
|l_\alpha(A\cdot q) - l_\alpha(A \cdot q')| \
& = \ \frac{1}{2} \left| \int_0^\pi w_{A\cdot q}(u) - w_{A \cdot
q'}(u) du \right| \\
& \le \ \frac{1}{2} \int_0^\pi \left| w_{A\cdot q}(u) - w_{A \cdot
q'}(u) \right| du \\
& \le \ \pi d_H(P_\alpha(A\cdot q), P_\alpha(A \cdot q')) \ < \
\pi \epsilon. \qedhere
\end{align*}
\end{proof}
\begin{proposition}\label{infimalcontinuity}
For any curve $\alpha \in \scc$, the function $\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha \from
\QD^1(S) \rightarrow \ensuremath{\mathbb{R}}\xspace_{\geq 0}$ is continuous and $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--invariant.
\end{proposition}
\begin{proof}
Fix $\alpha \in \scc$ and $q \in \QD^1(S)$. We now prove that
$\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha$ is continuous at $q$. In the case where $\alpha \in
\ensuremath{\widehat{\mathsf{cyl}}}\xspace(q)$, we have $\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q) = 0 = \area(q,\alpha)$. Continuity
of $\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha$ at $q$ then follows from Proposition
\ref{polygon}\ref{tw3} and Corollary \ref{areacontinuity}.
Assume $\alpha$ is crooked on $q$. Given any $\epsilon \in (0,1)$,
we shall show that $|\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q) - \ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q')| < \pi
\epsilon$ for all $q'$ in a sufficiently small neighborhood of
$q$. Suppose $\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q)$ is attained at $A_1 \cdot q$, for
some $A_1 \in \ensuremath{\mathrm{SL}(2,\R)}\xspace$. Choose $\lambda_0 \geq 2c\ensuremath{\lambda}\xspace(A_1)$,
where $c$ is the constant from Lemma \ref{lambdagrowth}. Let $U$ be an
open neighborhood of $q$ which satisfies the conclusion of Lemma
\ref{fellow}. For $q' \in U$, let $A_2 \in \ensuremath{\mathrm{SL}(2,\R)}\xspace$ be such that
$\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q') = l_\alpha(A_2 \cdot q')$. By Lemmas
\ref{lambdagrowth} and \ref{fellow}\ref{f2}, we have
\[
\ecc_\alpha(A_1 \cdot q') \ < \ (1+\epsilon)^2 \ecc_\alpha(A_1
\cdot q) \ < \ (1+\epsilon)^2 c \ < \ 4c.
\]
Applying Lemma \ref{lambdagrowth} with $m = A_2 \cdot q'$, we
obtain $\lambda(A_1 A_2^{-1})^2 \leq c \cdot\ecc_\alpha (A_1 \cdot
q')$, and therefore $\ensuremath{\lambda}\xspace(A_1A_2^{-1}) \leq 2c$. Thus,
\[
\ensuremath{\lambda}\xspace(A_2) \ \leq \ \ensuremath{\lambda}\xspace(A_2A_1^{-1})\ensuremath{\lambda}\xspace(A_1) \ = \
\ensuremath{\lambda}\xspace(A_1A_2^{-1})\ensuremath{\lambda}\xspace(A_1) \ \leq \ 2c\ensuremath{\lambda}\xspace(A_1) \ \leq \
\lambda_0.
\]
Finally, applying Lemma \ref{fellow}\ref{f3} yields
\begin{multline*}
\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q) - \pi \epsilon \ \leq \ l_\alpha(A_2 \cdot q) - \pi
\epsilon \ < \ l_\alpha(A_2 \cdot q') \ = \ \ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q') \\
\leq \ l_\alpha(A_1 \cdot q') \ < \ l_\alpha(A_1 \cdot q) + \pi
\epsilon \ = \ \ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q) + \pi \epsilon
\end{multline*}
as required. \qedhere
\end{proof}
\section{The polygonal area spectrum}
We are now ready to prove Theorem \ref{nscc}. The first step is
Theorem \ref{discreteness}, which says that $\polyspec(q)$ is
discrete if $q$ is a Veech surface.
\proof[Proof of Theorem \ref{discreteness}]
By Theorem \ref{nst}\ref{sw3a}, the virtual triangle area spectrum
$\VT(q)$ is discrete. For each simple closed curve $\alpha$,
$\area(P_{\alpha}(q))$ is a positive integer combination of numbers
from the set $\VT(q)$, by Lemma \ref{area-sum}. The result follows.
\qedhere
\endproof
Using Proposition \ref{polygon}\ref{tw3}, we obtain:
\begin{corollary}
If $q \in \QD^1(S)$ is a Veech surface then
\begin{equation}\inf\set{\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q) \st \alpha \text{ is
crooked on } q} > 0.\tag*{\qed}\end{equation}
\end{corollary}
For the converse, it is worth remarking that the existence of short
crooked curves is not an immediately obvious consequence of having small
virtual triangles. From a given collection of saddle connections on
$q$ for which $\abs{{u} \wedge {v}}$ can be taken to be arbitrarily
small, it appears difficult to construct a sequence of saddle
connections to satisfy the following:
\begin{itemize}
\item no two saddle connections intersect (in their interiors),
\item consecutive saddle connections meet with an angle of at least
$\pi$ on both sides,
\item their concatenation is homotopic to an essential simple
closed curve.
\end{itemize}
(Small triangles on $q$ are not particularly useful since their sides
must meet at an angle of less than $\pi$.) In our proof below, the
auxiliary polygon plays a key role in bypassing this difficulty.
For $q \in \Qk$, recall that $\ensuremath{\mathcal{O}}\xspace_q = \Gamma_q \cdot (\ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q)$
is a dense subset of $\ensuremath{\widetilde{\mathrm{M}}}\xspace_q$ in $\Qk$, where $\Gamma_q \leq \mcg(S)$
is the stabilizer of $\ensuremath{\widetilde{\mathrm{M}}}\xspace_q$.
\begin{proposition}\label{notveech-dense}
Suppose $q\in \QD^1(S)$ is not a Veech surface. Then there is a
number $a > 0$ such that the polygonal area spectrum $\polyspec(q)$
contains a dense subset of\/ $[0,a]$.
\end{proposition}
\proof
Applying Corollary \ref{veechcor}, we have $\ensuremath{\widetilde{\mathrm{M}}}\xspace_q \neq \ensuremath{\mathcal{O}}\xspace_q$
and so we may choose $q' \in \ensuremath{\widetilde{\mathrm{M}}}\xspace_q - \ensuremath{\mathcal{O}}\xspace_q$. Then $\ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q \neq
\ensuremath{\mathrm{SL}(2,\R)}\xspace \cdot q'$, and so by Proposition \ref{rigidity}, there exists
a curve $\alpha \in \ensuremath{\widehat{\mathsf{cyl}}}\xspace(q) - \ensuremath{\widehat{\mathsf{cyl}}}\xspace(q')$. By Proposition
\ref{polygon}\ref{tw4}, we have $\area(q, \alpha) = 0$ and
$\area(q', \alpha) = a > 0$. Since $\ensuremath{\widetilde{\mathrm{M}}}\xspace_q$ is connected and
$\area(\, \param \, , \alpha)$ is continuous, by Corollary
\ref{areacontinuity}, we deduce that $[0,a] \subseteq
\area(\ensuremath{\widetilde{\mathrm{M}}}\xspace_q, \alpha)$. It follows that $\area(\ensuremath{\mathcal{O}}\xspace_q, \alpha)$,
and hence $\area(\ensuremath{\mathcal{O}}\xspace_q, \scc)$, contains a dense subset of
$[0,a]$. Finally, the polygonal area spectrum $\polyspec(q) =
\area(q, \scc)$ is invariant under $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--deformations and changes
of markings, and therefore $\area(q, \scc) = \area(\ensuremath{\mathcal{O}}\xspace_q,
\scc)$. \qedhere
\endproof
Note that this proof uses only the fact that $\area(\, \param \, , \alpha)$
is a continuous \ensuremath{\mathrm{SL}(2,\R)}\xspace--invariant function to $\ensuremath{\mathbb{R}}\xspace_{\geq 0}$ which takes
the value 0 precisely when $\alpha \in \ensuremath{\widehat{\mathsf{cyl}}}\xspace(q)$. Since the same properties
hold for $\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha$, we may argue as above to deduce:
\begin{proposition}
If $q \in \QD^1(S)$ is not a Veech surface then there exists
$a > 0$ such that the \ensuremath{\mathrm{SL}(2,\R)}\xspace--infimal length spectrum
$\set{\ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q) \st \alpha \in \scc}$
contains a dense subset of \/ $[0,a]$. \qed
\end{proposition}
We conclude this section with a proof of Theorem \ref{finiteness},
which is restated below. For $a > 0$, recall that
\begin{eqnarray*}
\PA(a) &=& \set{q \in \QD^1(S) \st \area(\ensuremath{\mathrm{P}}\xspace_\alpha(q)) \geq a \text{
for every crooked curve } \alpha \text{ on } q}, \\
\IL(a) &=& \set{q \in \QD^1(S) \st \ensuremath{l^{\,\mathrm{SL}}\xspace}_\alpha(q) \geq a \text{
for every crooked curve } \alpha \text{ on } q}.
\end{eqnarray*}
\begin{theorem}
For any $a>0$, the sets $\PA(a)$ and $\IL(a)$ both contain only finitely
many affine equivalence classes of half-translation surfaces.
\end{theorem}
\proof
Applying Corollary \ref{areacontinuity}, we deduce that $\PA(a)$ is a
closed and $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--invariant subset of $\QD^1(S)$. Moreover, $\PA(a)$
is invariant under the action of $\mcg(S)$, and so $\PA(a)$ descends
to a closed $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--invariant subset $C$ in $\QMod(S)$. It follows that
$C \cap \Qk$ is closed and $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--invariant in each stratum $\Qk$
under the subspace topology. By Theorem \ref{EMMfinite}, $C \cap \Qk$
is a finite union of $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--orbit closures. Since elements of $\PA(a)$
are necessarily Veech surfaces by Theorem \ref{nscc}, $C \cap \Qk$
must be a finite union of $\ensuremath{\mathrm{SL}(2,\R)}\xspace$--orbits. Finally, there are only finitely
many strata for a given genus, and so the desired result follows.
The proof for $\IL(a)$ proceeds identically using Proposition
\ref{infimalcontinuity} in place of Corollary
\ref{areacontinuity}.\qedhere
\endproof
\section{The virtual triangle area spectrum}
\label{sec:triangles}
We are almost ready to prove Theorem \ref{nsvt}. In the introduction we
defined the virtual triangle area spectrum $\VT(q)$. Let $\VT_0(q)
\subset \VT(q)$ be the subset consisting of the numbers $\abs{u \wedge
v}$ such that $u$ and $v$ are the holonomy vectors of a pair of saddle
connections with a common endpoint (that is, a \emph{based} virtual
triangle). Note that saddle connections forming a based virtual triangle
need not form a triangle, since they may have angle $\pi$ or more on both
sides. We also define $\VT_0(X) = \bigcup_{q \in X}\VT_0(q)$ for any set
$X\subset \QD^1(S)$.
We know from the implication $\ref{sw1} \Rightarrow \ref{sw3}$ of Theorem
\ref{nst} that if $q$ is a Veech surface then $\VT(q)$ has a gap above
zero. The remainder of Theorem \ref{nsvt} follows from the next
proposition:
\begin{proposition}\label{densetriangles}
Suppose $q \in \QD^1(S)$ is not a Veech surface. Then there is a
number $a > 0$ such that $\VT_0(q)$ contains a dense subset of\/
$[0,a]$.
\end{proposition}
\begin{proof}
As in the proof of Proposition \ref{notveech-dense}, there exist a
half-translation surface $q' \in \ensuremath{\widetilde{\mathrm{M}}}\xspace_q - \ensuremath{\mathcal{O}}\xspace_q$ and a curve
$\alpha \in \ensuremath{\widehat{\mathsf{cyl}}}\xspace(q) - \ensuremath{\widehat{\mathsf{cyl}}}\xspace(q')$, and we know that $\area(q,
\alpha) = 0$ and $\area(q',\alpha) > 0$. Let $q_t$ be a path in
$\ensuremath{\widetilde{\mathrm{M}}}\xspace_q$ from $q_0 = q$ to $q_1 = q'$. Consider the set
\[
\set{t \in [0,1] \st \area(q_t, \alpha) > 0},
\]
which is an open neighborhood of $1$ in $[0,1]$ that does not
contain $0$. It has a connected component $(t_0, 1]$. Replacing
the path $q_t$ by its restriction to $[t_0, 1]$ and
reparametrizing over $[0,1]$, we have $\area(q_0, \alpha) = 0$ and
$\area(q_t, \alpha) > 0$ for all $t \in (0,1]$.
Now consider the geodesic representative $\alpha^{q_0}$ and
express it as a concatenation of saddle connections $e_1 \dotsm
e_k$. Let $\alpha_i$ be the topological arc represented by
$e_i$; it is an isotopy class rel endpoints, where the interior of
the arc is required to avoid the singularities. Consider the set
\[
\set{t \in [0,1] \st \text{each } \alpha_i \text{ is represented
by a saddle connection in } q_t}.
\]
This set is an open neighborhood of $0$ in $[0,1]$. Again,
replacing the path $q_t$ by its restriction to an interval $[0,
\epsilon]$ and reparametrizing over $[0,1]$, we may assume that
the arcs $\alpha_1, \dotsc, \alpha_k$ are represented by saddle
connections for all $t \in [0,1]$. Let $e_i(t)$ denote the saddle
connection in $q_t$ representing $\alpha_i$. Let $v_i(t)$ be the
holonomy vector of $e_i(t)$.
Define functions $\phi_i(t) = \abs{v_i(t) \wedge v_{i+1}(t)}$ for
each $i$ (with indices taken mod $k$).
These are continuous because holonomy vectors vary continuously
where defined. Since all $e_i$ are parallel on $q_0$,
we have $\phi_i(0) = 0$ for all $i$.
We claim that $\phi_i(t_1) > 0$ for some $t_1>0$ and some $i$.
If not, then for all $t \in [0,1]$ the saddle
connections $e_i(t)$ are all parallel. The angles
between consecutive saddle connections (on either side) must
remain constant, since they are constrained to lie in the discrete
set $\pi\ensuremath{\mathbb{Z}}\xspace$, and therefore the concatenation $e_1(t) \dotsm
e_k(t)$ remains a geodesic representative for $\alpha$ on $q_t$.
But this contradicts the fact that $\area(q_t, \alpha)
> 0$ for all $t \in (0,1]$.
Restricting $q_t$ to $[0,t_1]$ and reparametrizing over $[0,1]$ one
last time, we have a path $q_t$ in $\ensuremath{\widetilde{\mathrm{M}}}\xspace_q$ and a pair of saddle
connections $e_i(t)$, $e_{i+1}(t)$ which persist on $q_t$ throughout
the path, such that $\phi_i(0) = \abs{v_i(0) \wedge v_{i+1}(0)} = 0$
and $\phi_i(1) = \abs{v_i(1) \wedge v_{i+1}(1)} = a > 0$. The function
$\phi_i$ is defined and continuous on the open set $U \subset \ensuremath{\widetilde{\mathrm{M}}}\xspace_q$
where $e_i$ and $e_{i+1}$ persist. This set contains the path $q_t$ and
hence $\phi_i(U)$ contains $[0,a]$. Since $U \cap \ensuremath{\mathcal{O}}\xspace_q$ is dense in
$U$, it follows that $\phi_i(U \cap \ensuremath{\mathcal{O}}\xspace_q)$, and hence
$\VT_0(\ensuremath{\mathcal{O}}\xspace_q)$, contains a dense subset of $[0,a]$. Finally, the based
virtual triangle area spectrum is invariant under change of marking and
$\ensuremath{\mathrm{SL}(2,\R)}\xspace$, and so $\VT_0(q) = \VT_0(\ensuremath{\mathcal{O}}\xspace_q)$. \qedhere
\end{proof}
\small
\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}
\label{sec:introduction}
Spin-orbit-coupled Bose-Einstein condensates (BECs) are characterized by a rich variety of quantum phases,
which have already been the subject of theoretical and experimental investigations (see the recent
reviews~\cite{Dalibard2011review,Galitski2013review,Zhou2013review,Goldman2014review,Zhai2015review,
Li2015review,Zhang2016review} and references therein). In particular, in the case of a BEC of pseudospin $1/2$
with equal Rashba~\cite{Bychkov1984} and Dresselhaus~\cite{Dresselhaus1955} spin-orbit couplings, by tuning
the value of the Raman coupling between the two pseudospin states one can explore different phase transitions.
For relatively large values of the Raman coupling, of the order of the recoil energy, these systems exhibit a
second-order transition between a plane-wave and a single-minimum phase~\cite{Lin2011,Li2012PRL}.
The former is characterized by the macroscopic occupation of a single-particle state with finite momentum and
magnetic polarization, while in the latter the atoms populate an unpolarized state with vanishing momentum.
The above transition is associated with a divergent behavior of the magnetic susceptibility and with a large increase
of the effective mass. At a dynamic level it is characterized by the softening of the sound velocity~\cite{Martone2012,
Zheng2013,Ji2015} and of the frequency of the collective oscillations in the presence of harmonic trapping~\cite{Zhang2012,
Li2012EPL}. When one decreases the value of the Raman coupling the plane-wave phase eventually disappears in favor
of the so-called striped phase~\cite{Wang2010,Wu2011,Ho2011,Sinha2011,Li2012PRL,Ozawa2012,Li2013,Zezyulin2013,Lan2014,
Han2015,Natu2015,Sun2015,Liao2015,Sun2016,Yu2016,Martone2016}. In this configuration one has the appearance of periodic
modulations in the density profile, whose contrast depends on the value of the Raman coupling. In the absence of an
effective magnetic field the magnetic polarization vanishes in the striped phase. The phase transition between the
plane-wave and the striped phases has a first-order nature. It is characterized by the occurrence, on the side of the
plane-wave phase, of a roton minimum in the excitation spectrum, whose energy becomes smaller and smaller as one approaches
the transition~\cite{Martone2012,Zheng2013,Khamehchi2014,Ji2015}. The presence of the striped phase is one of the most
interesting features exhibited by spin-orbit-coupled BECs, due to its direct link to the long sought phenomenon of
supersolidity~\cite{Boninsegni2012}, which takes place when two continuous symmetries (gauge and translational invariance)
are spontaneously broken. So far the striped phase has not been identified directly in experiments because of the very small
contrast of the density fringes in the available experimental conditions~\cite{Martone2014}.
The occurrence of the roton excitation in the plane-wave phase and the emergence of the striped phase are deeply related
physical phenomena. This link becomes more evident by looking at the effects of a one-dimensional static periodic potential
applied to the BEC in the plane-wave phase. The linear response of the gas to a perturbation with wave vector close to the
roton momentum is greatly enhanced when approaching the transition to the striped phase~\cite{Martone2012}. Beyond the linear
regime, it has been shown that the application of the external potential can induce a magnetic phase transition to a fully
unpolarized configuration~\cite{Chen2016}. In this paper we investigate in detail the connection between the properties of
the response of the BEC at the linear and the nonlinear level. This results in a strong dependence of the behavior of the
system on the parameters of the external periodic potential. For example, when its wave vector is close to the roton minimum,
even a tiny static field is capable of inducing the transition to the unpolarized phase, with the appearance of highly contrasted
density modulations and strong magnetic fluctuations. For wave vectors different from the roton momentum, the magnetic phase
transition can take place at larger intensities of the optical lattice.
The paper is structured as follows. In Sec.~\ref{sec:soc_becs_unif} we review some relevant properties of the quantum phases
of a spin-orbit-coupled BEC in uniform matter and we point out the peculiar behavior of the static density response in the
plane-wave phase. Section~\ref{sec:soc_becs_ol} deals with the ground state of the system in the presence of a static optical
lattice; we mainly focus on the occurrence of a magnetic phase transition, which takes place at a critical lattice intensity whose
value depends in a nontrivial way on the Raman coupling and the lattice wave vector. The role of magnetic fluctuations and the effects
of an external trapping potential are also investigated. We summarize in Sec.~\ref{sec:conclusion}. Finally, in the Appendix we provide
a brief description of the features of the band structure of an ideal Bose gas with spin-orbit coupling.
\section{Spin-orbit-coupled Bose-Einstein condensates in uniform matter}
\label{sec:soc_becs_unif}
\subsection{Single-particle Hamiltonian}
\label{subsec:sp_ham}
The single-particle Hamiltonian first realized in the experiment of Ref.~\cite{Lin2011} reads (we set $\hbar = 1$)
\begin{equation}
h_0 = \frac{1}{2m} \left[ (p_x - k_0 \sigma_z)^2 + p_\perp^2 \right]
+ \frac{\Omega}{2}\sigma_x + \frac{\delta}{2}\sigma_z \, .
\label{eq:soc_ham}
\end{equation}
It acts on two-component spinors describing bosons with pseudospin up ($\uparrow$) and down ($\downarrow$);
for simplicity, we will refer to these degrees of freedom just as the two spin states of the system.
Equation~\eqref{eq:soc_ham} accounts for the presence of two counterpropagating and linearly polarized Raman lasers,
which provide transitions between different hyperfine levels of the atoms, and of a bias magnetic field.
The Raman coupling strength is quantified by $\Omega$, while $k_0$ is the momentum transfer due to the lasers.
The latter also fixes the value of the Raman recoil energy $E_r \equiv k_0^2 / 2m$. The linear Zeeman term $\delta$
represents an effective magnetic field, given by the sum of the Raman detuning and of the physical external magnetic
field (see, for example, Ref.~\cite{Martone2012}). Finally, $\sigma_i$ with $i=x,y,z$ denotes the $i$th Pauli matrix.
For vanishing $\delta$, when the Raman coupling $\Omega$ is smaller than $4E_r$, the lower branch of the single-particle
dispersion exhibits two degenerate minima at $p_x = \pm k_1^0$, with
\begin{equation}
k_1^0 \equiv k_0 \sqrt{1 - \left(\frac{\Omega}{4 E_r}\right)^2} \, ,
\label{eq:k10}
\end{equation}
while for $\Omega \geq 4 E_r$ it has one single minimum at $p_x = 0$.
\subsection{Many-body ground state}
\label{subsec:mb_ground_state}
We now discuss the effects of the two-body interactions. In the Gross-Pitaevskii mean-field approach, the energy
of an interacting system of $N$ spin-orbit-coupled bosons enclosed in a volume $V$ is given by
\begin{equation}
\begin{split}
E\left[\Psi\right] = \int_V \mathrm{d}\vec{r} \big\{ & \Psi^\dagger(\vec{r}) h_0 \Psi(\vec{r}) \\
&{} + g_1 n^2(\vec{r}) + g_2 s_z^2(\vec{r}) + g_3 n(\vec{r}) s_z(\vec{r}) \big\} \, ,
\end{split}
\label{eq:soc_energy}
\end{equation}
where $\Psi$ denotes the two-component condensate wave function, $n(\vec{r}) = \Psi^\dagger(\vec{r}) \Psi(\vec{r})$
is the total density obeying the normalization condition $\int_V \mathrm{d}\vec{r} \, n(\vec{r}) = N$, and $s_z(\vec{r}) =
\Psi^\dagger(\vec{r}) \sigma_z \Psi(\vec{r})$ is the longitudinal spin density. The coupling constants in
Eq.~\eqref{eq:soc_energy} correspond to the combinations $g_1 \equiv (g_{\uparrow\uparrow} + g_{\downarrow\downarrow}
+ 2 g_{\uparrow\downarrow})/8$, $g_2 \equiv (g_{\uparrow\uparrow} + g_{\downarrow\downarrow} - 2 g_{\uparrow\downarrow})/8$,
and $g_3 \equiv (g_{\uparrow\uparrow} - g_{\downarrow\downarrow})/4$ of the intraspecies and interspecies interaction strengths
$g_{\alpha\beta}$ ($\alpha,\beta = \uparrow,\downarrow$), which are related to the corresponding $s$-wave scattering lengths
$a_{\alpha\beta}$ via $g_{\alpha\beta} = 4\pi a_{\alpha\beta}/m$. In the following, unless otherwise specified, we will assume
equal intraspecies interactions $g_{\uparrow \uparrow} = g_{\downarrow \downarrow}$, yielding $g_3 = 0$. We will also take
$\delta = 0$, although the effects of a nonvanishing effective magnetic field will also be discussed qualitatively.
In the presence of antiferromagnetic spin-dependent interactions ($g_2 > 0$), when the Raman coupling $\Omega$ is small,
the ground state of the many-body system corresponds to the so-called striped phase~\cite{Wang2010,Wu2011,Ho2011,Sinha2011,
Li2012PRL,Ozawa2012,Li2013,Zezyulin2013,Lan2014,Han2015,Natu2015,Sun2015,Liao2015,Sun2016,Yu2016,Martone2016}. In this
configuration the density profile of the gas exhibits periodic modulations in the form of stripes, which appear as a consequence
of the spontaneous breaking of translational invariance. Another relevant feature of the striped phase is the absence of magnetic
polarization along $z$, that is, $\langle \sigma_z \rangle \equiv \int_V \mathrm{d}\vec{r} \, s_z(\vec{r}) = 0$.
By increasing $\Omega$ the system enters the polarized plane-wave phase. In this phase the condensate order parameter is given
by the plane-wave function $\exp(\mathrm{i} k_1 x)$ [or $\exp(- \mathrm{i} k_1 x)$] times a real spinor. The components of the latter have
relative weights fixed by the value of the Raman coupling. The plane-wave phase is characterized by a uniform density and by a
finite value $\langle \sigma_z \rangle = N k_1/k_0$ (or $\langle \sigma_z \rangle = - N k_1/k_0$) of the longitudinal magnetic
polarization~\cite{Ho2011,Li2012PRL}. The momentum $k_1$ in the previous formulas does not actually coincide with its
single-particle value $k_1^0$ given by Eq.~\eqref{eq:k10} because of the spin-dependent interactions proportional to $g_2$;
one finds~\cite{Li2012PRL}
\begin{equation}
k_1 = k_0 \sqrt{1 - \left[\frac{\Omega}{4(E_r - \bar{n}g_2)}\right]^2} \, ,
\label{eq:k1}
\end{equation}
with $\bar{n} \equiv N/V$ the average density of the gas. To simplify the discussion, in this work we will not account for the small
difference between $k_1$ and $k_1^0$, which is negligible in the conditions of current experiments. The twofold degeneracy of the
plane-wave phase highlighted above stems from a spontaneous breaking mechanism of two $\mathbb{Z}_2$ symmetries of the energy
functional~\eqref{eq:soc_energy}, similarly to what happens in usual ferromagnetic configurations. More specifically, if both $\delta$
and $g_3$ are vanishing, the energy~\eqref{eq:soc_energy} is invariant under the separate action of $\sigma_x \mathcal{P}$
and $\sigma_z \mathcal{T}$, with $\mathcal{P}$ and $\mathcal{T}$ the parity and time-reversal operator, respectively. The plane-wave
phase breaks both the above symmetries, while being invariant under their product
$(\sigma_x \mathcal{P})(\sigma_z \mathcal{T})$.\footnote{Notice that the energy functional~\eqref{eq:soc_energy} keeps its invariance
under $(\sigma_x \mathcal{P}) (\sigma_z \mathcal{T})$ even if $\delta$ and $g_3$ are nonvanishing.}
The transition between the striped and the plane-wave phases occurs at a critical Raman coupling $\Omega^0_\mathrm{tr}$ whose value is given,
in the low-density limit, by the density-independent expression~\cite{Ho2011,Li2012PRL}
\begin{equation}
\Omega^0_\mathrm{tr} = 4E_r \sqrt{\frac{2 g_2}{g_1+2 g_2}} \, .
\label{eq:Omega_tr_0}
\end{equation}
This transition is of first order and has an important magnetic character~\cite{Martone2012,Li2012EPL} that has been pointed
out experimentally in~\cite{Lin2011,Ji2014}. At even larger values of the Raman coupling $\Omega$ the BEC undergoes a
second-order transition to the single-minimum phase, in which the condensation momentum and the magnetic polarization are
both vanishing.\footnote{This description of the phase diagram of a spin-orbit-coupled BEC with $g_2 > 0$ holds for densities
$\bar{n}$ smaller than the critical value $\bar{n}_\mathrm{cr} = E_r g_1/[2 g_2 (g_1 + g_2)]$. For $\bar{n} > \bar{n}_\mathrm{cr}$
the plane-wave phase disappears and one has a direct first-order transition between the striped and the single-minimum phases.}
We finally mention that, if the spin-dependent interactions have a ferromagnetic nature ($g_2 < 0$), only the plane-wave
and the single-minimum phases appear in the phase diagram of the BEC, the striped phase being always energetically unfavored.
\subsection{Rotonic excitations and compressibility}
\label{subsec:roton_dens_resp}
A peculiar property of the plane-wave phase is the existence of a low-energy roton minimum in the excitation spectrum, occurring
at a momentum close to $2k_1$~\cite{Martone2012,Zheng2013,Khamehchi2014,Ji2015}. The energy of the roton becomes smaller and smaller
as $\Omega$ approaches (from above) the critical value $\Omega^0_\mathrm{tr}$, providing the onset of the transition to the striped
phase. The roton minimum is responsible for a peculiar behavior of the static density response, i.e., the compressibility $\chi_\rho(q)$.
According to the linear response theory, the compressibility can be calculated by adding, to the single-particle
Hamiltonian~\eqref{eq:soc_ham}, a small static periodic perturbation of the form $V_\lambda = -\lambda \rho_q + \mathrm{H.c.}$,
where $\rho_q \equiv \int_V \mathrm{d} \vec{r} \, \mathrm{e}^{-\mathrm{i} q x} \hat{n}(\vec{r})$ and $\hat{n}(\vec{r})$ is the density operator. Then
$\chi_\rho(q) \equiv \lim_{\lambda \to 0} \langle \rho_q \rangle_\lambda/\lambda$, where $\langle \rho_q \rangle_\lambda$ is the expectation
value of $\rho_q$ on the perturbed ground state. One finds that the function $\chi_\rho(q)$ exhibits a significant enhancement
when $q$ is close to the roton minimum~\cite{Martone2012}. A similar effect is known to characterize the static response of superfluid
helium~\cite{Dalfovo1992}. The enhancement becomes particularly strong when one approaches the phase transition to the striped phase at $\Omega =
\Omega^0_\mathrm{tr}$. In Fig.~\ref{fig:chi_dens_x_q} we plot the compressibility calculated at $q = 2k_1^0$ as a function of $\Omega$,
in the presence (red solid line) and in the absence (red dotted line) of spin-orbit coupling. The values of the average density $\bar{n}$
and of the ratio $g_2/g_1 = 0.0012$ employed in the figure correspond to current experiments with $^{87}$Rb BECs (see also the discussion
in Sec.~\ref{subsec:harm_trap}). In such conditions, the low-density expression~\eqref{eq:Omega_tr_0} yields the value
$\Omega^0_\mathrm{tr}/E_r = 0.19$ for the critical Raman coupling, which is raised to $\Omega^0_\mathrm{tr}/E_r = 0.21$ after including
the small corrections due to the finite density of the system. One can notice that, for small $\Omega$, the static response of a
spin-orbit-coupled BEC at $q = 2k_1^0$ takes extremely high values as compared to those of a Bose gas without spin-orbit coupling.
In the latter case the value of $\chi_\rho(q=2k_1^0)$ for small $\Omega$ is well approximated by the asymptotic large-$q$ behavior
$\chi_\rho(q) \to 4m/q^2$. The strong enhancement of $\chi_\rho(q)$ does not take place if the momentum transfer $q$ is significantly
smaller or larger than the roton momentum; to show this, in Fig.~\ref{fig:chi_dens_x_q} we also report the values of the compressibility
in the presence of spin-orbit coupling at $q = 1.5k_1^0$ (green dash-dotted line) and $q = 3k_1^0$ (blue dashed line).
\begin{figure}
\includegraphics{figure1_chi_dens_x_q.eps}
\caption{(Color online) Static density response $\chi_\rho(q)$ of a spin-orbit-coupled BEC in the plane-wave phase as a function of
the Raman coupling $\Omega$, at three different values $q = 1.5k_1^0$ (green dash-dotted line), $q = 2k_1^0$ (red solid line),
and $q = 3k_1^0$ (blue dashed line) of the momentum transfer. The red dotted curve corresponds to the response at $q = 2k_1^0$
of a standard Bose gas without spin-orbit coupling. The vertical black dash-dotted line identifies the critical Raman coupling
$\Omega^0_\mathrm{tr}$ at which the transition to the striped phase takes place. The other parameters are $\bar{n} g_1 / E_r = 0.4$
and $g_2/g_1 = 0.0012$.}
\label{fig:chi_dens_x_q}
\end{figure}
The significant dependence of the static response on the momentum transfer is reflected in the peculiar behavior of the system
when entering the nonlinear regime. In particular, the large value of the compressibility close to the roton momentum suggests
that nonlinear effects will emerge soon in the response, even in the presence of a tiny periodic perturbation. They are expected
to give rise to highly contrasted density modulations and to important effects in the magnetic polarization of the gas, as we will
discuss in the next section.
\section{Magnetic phase transition in the presence of an optical lattice}
\label{sec:soc_becs_ol}
\subsection{Variational ansatz for the ground state}
\label{subsec:var_ansatz}
Let us now explore the behavior of a spin-orbit-coupled BEC when a one-dimensional sinusoidal periodic potential of the form
\begin{equation}
V(\vec{r}) = s E_\mathrm{latt} \sin^2 \frac{q x}{2}
\label{eq:periodic_pot}
\end{equation}
is added to the single-particle Hamiltonian~\eqref{eq:soc_ham}. In the previous expression, $q$ denotes the momentum transferred
by the two lasers, oriented along the $x$ axis, which generate the optical lattice. The dimensionless parameter $s$ fixes the
strength of the periodic potential in units of the lattice recoil energy $E_\mathrm{latt} \equiv (q/2)^2/2m$. We will focus on
the case of relatively weak lattice intensities, such that the Gross-Pitaevskii mean-field description is still applicable.
There have been several studies on spin-orbit-coupled BECs in shallow optical lattices~\cite{Larson2010,Sakaguchi2013,Zhang2013,
Kartashov2013,Cheng2014,Han2015,Salerno2015,Hamner2015,Li2015,Zhang2015,Poon2015,Chen2016,Hurst2016}. In particular,
the ground state in the presence of the periodic potential~\eqref{eq:periodic_pot} has been investigated in Ref.~\cite{Chen2016}.
In this work we focus on the properties of a magnetic phase transition induced by the periodic potential.
First, we recall that, if the atoms condense in a state with a well-defined value of the quasimomentum $k$, the condensate
wave function can be written in the usual Bloch form
\begin{equation}
\tilde{\Psi}_k\left(\vec{r}\right)
= \sqrt{\bar{n}} \, \mathrm{e}^{\mathrm{i} k x} \sum_{\bar{K}} C_{k+\bar{K}}
\begin{pmatrix}
\phantom{+}\cos\theta_{k+\bar{K}} \\
- \sin\theta_{k+\bar{K}}
\end{pmatrix}
\mathrm{e}^{\mathrm{i} \bar{K} x}.
\label{eq:bloch_waves}
\end{equation}
Here $\bar{K}$ are the reciprocal lattice vectors having values $\left\{\bar{m} q\right\}_{\bar{m}\in\mathbb{Z}}$,
while $\theta_{k+\bar{K}}$ and $C_{k+\bar{K}}$ are parameters characterizing the coefficients of the Bloch wave expansion.
The coefficients $C_{k+\bar{K}}$ satisfy the constraint $\sum_{\bar{K}} | C_{k+\bar{K}} |^2 = 1$, which ensures that
the order parameter~\eqref{eq:bloch_waves} is normalized to the total number of particles in the condensate, i.e.,
$\int_V \mathrm{d}\vec{r} \, \tilde{\Psi}_k^\dagger(\vec{r}) \tilde{\Psi}_k(\vec{r}) = N$. Henceforth we will take $k$ in the
first Brillouin zone.
A good starting point for the study of the ground state of the BEC is represented by the properties of the band structure
of the system in the noninteracting limit~\cite{Zhang2013,Chen2016}, which are summarized in the Appendix. In particular,
in the first Brillouin zone the lowest-lying energy band exhibits two degenerate minima, which occur at two opposite finite values
of the quasimomentum [see Figs.~\ref{fig:band_structure}(a) and \ref{fig:band_structure}(b)]. These considerations suggest
that one can write an ansatz for the ground-state wave function as a superposition of two Bloch waves of the form~\cite{Chen2016}
\begin{equation}
\Psi\left(\vec{r}\right) =
\tilde{C}_+ \tilde{\Psi}_{+k_s}\left(\vec{r}\right)
+ \tilde{C}_- \tilde{\Psi}_{-k_s}\left(\vec{r}\right) \, .
\label{eq:ansatz}
\end{equation}
Here $|\tilde{C}_+|^2 + |\tilde{C}_-|^2 = 1$ and the expressions of $\tilde{\Psi}_{\pm k_s}$ are given by
Eq.~\eqref{eq:bloch_waves} with $k = \pm k_s$, $k_s$ being the magnitude of the quasimomentum in the ground state.
For an ideal Bose gas, Eq.~\eqref{eq:ansatz} is able to reproduce the exact ground-state wave function of the
system; in particular, the quasimomenta $\pm k_s$ correspond to the minima of the lowest band mentioned above, while
the relative values of the coefficients $\tilde{C}_+$ and $\tilde{C}_-$ remain arbitrary.
The same ansatz~\eqref{eq:ansatz} is also well suited to study the ground state of the system in the presence of two-body
interactions~\cite{Chen2016}. Before proceeding, it is worth noticing that it allows one to recover all the results presented
in Sec.~\ref{sec:soc_becs_unif} for the quantum phases of the BEC in the absence of the optical lattice (i.e., for $s=0$).
In this limit the Bloch wave~\eqref{eq:bloch_waves} has $C_k = 1$ and $C_{k+\bar{K}} = 0$ for $\bar{K} \neq 0$, that is,
it reduces to a simple plane wave with momentum $k$. Consequently, the wave function~\eqref{eq:ansatz} becomes a
superposition of two counterpropagating plane waves with momenta $\pm k_s$, which coincides with the ansatz
employed in the variational analysis of Ref.~\cite{Li2012PRL}. The plane-wave phase is reproduced by setting
$k_s = k_1$, with $k_1$ given by Eq.~\eqref{eq:k1}, $\tilde{C}_+ = 1$, and $\tilde{C}_- = 0$ (or $\tilde{C}_+ = 0$
and $\tilde{C}_- = 1$ for the degenerate state with opposite momentum and magnetic polarization); if additionally $k_1$
vanishes, one gets instead the single-minimum phase. The striped phase corresponds to the choice $|\tilde{C}_+| = |\tilde{C}_-|
= 1/\sqrt{2}$.
When the optical lattice is turned on one expects that also the plane-wave components with $\bar{K} \neq 0$ in Eq.~\eqref{eq:bloch_waves}
are populated. In order to study the ground state of the system at finite $s$, we first insert the ansatz~\eqref{eq:ansatz} into the energy
functional~\eqref{eq:soc_energy} and we deduce an expression for the energy as a function of the variational parameters $k_s$,
$\theta_{\pm k_s+\bar{K}}$, $C_{\pm k_s+\bar{K}}$, and $\tilde{C}_\pm$. Then we minimize the energy at fixed values of $k_0$,
$\Omega$, $\bar{n}$, the ratio $g_2/g_1$ of the interaction strengths, and the lattice parameters $q$ and $s$. We have checked that the
results given by this variational procedure agree with those obtained by directly solving the Gross-Pitaevskii equation in a box configuration.
\subsection{Mixed regime}
\label{subsec:mixed_regime}
At sufficiently small values of the Raman coupling $\Omega$, if $g_2 > 0$, the ground-state wave function of the BEC contains both the Bloch wave
terms in the superposition~\eqref{eq:ansatz} with equal weights $|\tilde{C}_+| = |\tilde{C}_-| = 1/\sqrt{2}$, giving rise to the so-called mixed
phase. The properties of this phase have been studied in detail in Ref.~\cite{Chen2016}. Here we mention that it is characterized by vanishing magnetic
polarization and by modulations in the density profile at two different wavelengths: The first one, equal to $2\pi/q$, is fixed by the external lattice
potential, while the second one arises because of the spin-orbit coupling and is given by $\pi/k_s$. The presence of the latter entails a spontaneous
breaking of the discrete translational symmetry exhibited by the energy functional~\eqref{eq:soc_energy} after the addition of the lattice
potential~\eqref{eq:periodic_pot}. Depending on whether the two wavelengths are commensurate or not, the global oscillation of the density can be periodic
or nonperiodic. The ansatz~\eqref{eq:ansatz} actually provides only a first approximation for the wave function of the mixed phase, as it neglects
the higher-order Bloch wave terms caused by the nonlinear interactions in the energy functional~\eqref{eq:soc_energy}~\cite{Li2013,Chen2016}.
Notice that at $s = 0$ the mixed phase smoothly connects to the striped phase, where only modulations with period $\pi/k_s$ are present.
\subsection{Unmixed regime: Magnetic phase transition}
\label{subsec:unmixed_regime}
The situation becomes dramatically different if one increases $\Omega$. Above a critical value $\Omega_\mathrm{tr}$ of the Raman coupling, the mixed
phase becomes energetically unfavored and a first-order transition occurs to an unmixed regime. In the latter the procedure of energy minimization yields,
for the coefficients in the ansatz~\eqref{eq:ansatz}, the values $\tilde{C}_+ = 1$ and $\tilde{C}_- = 0$ or the values $\tilde{C}_+ = 0$ and $\tilde{C}_- = 1$,
the two choices corresponding to the same value of the energy~\cite{Chen2016}. Hence, in the unmixed regime the wave function of the BEC is given by
a single Bloch wave of the kind~\eqref{eq:bloch_waves}, whose quasimomentum can be assumed equal to either $+k_s$ or $-k_s$. The value of $\Omega_\mathrm{tr}$
separating the mixed and unmixed configurations depends on the lattice strength $s$ (see Fig.~\ref{fig:phase_diag_Omega_s}); for $s = 0$, it reduces
to the critical Raman coupling $\Omega^0_\mathrm{tr}$ at which the transition from the striped to the plane-wave phase occurs [see Eq.~\eqref{eq:Omega_tr_0}
and the related discussion]. In the rest of the present paper we will focus principally on the physics of the unmixed regime, which represents
the central part of our work.
The quantum phases appearing in the unmixed regime, which we discuss below, display modulations in the density profile with the same periodicity
$2\pi/q$ as the external lattice potential~\eqref{eq:periodic_pot}. For such periodic fringes one can define the contrast as
\begin{equation}
I \equiv \frac{n_\mathrm{max} - n_\mathrm{min}}{n_\mathrm{max} + n_\mathrm{min}} \, ,
\label{eq:contrast}
\end{equation}
where $n_\mathrm{max}$ and $n_\mathrm{min}$ are the maximum and the minimum value, respectively, taken by the density during its spatial oscillations.
For a fixed value of the Raman coupling $\Omega > \Omega_\mathrm{tr}$, the properties of the ground state are determined by the competition
between the density-density interaction term in the energy~\eqref{eq:soc_energy}, proportional to $g_1$, and the lattice potential with strength $s$.
For very small values of $s$, the interactions favor a configuration where the atoms dominantly occupy the $\bar{K} = 0$ state in the
superposition~\eqref{eq:bloch_waves}, the populations $|C_{\pm k_s+\bar{K}}|^2$ of the terms with $\bar{K} \neq 0$ being much smaller. In this phase,
which smoothly connects to the plane-wave phase at $s = 0$, the ground state is twofold degenerate. In particular, the magnetic polarization $\left\langle \sigma_z
\right\rangle$ is finite and takes opposite values in the two states with quasimomentum $+k_s$ and $-k_s$. As in the plane-wave phase at $s = 0$, this
degeneracy stems from the spontaneous breaking of the $\sigma_x \mathcal{P}$ and $\sigma_z \mathcal{T}$ symmetries discussed in Sec.~\ref{sec:soc_becs_unif},
which are not affected by the addition of the periodic potential~\eqref{eq:periodic_pot}. Notice that the quasimomentum $k_s$ approaches $k_1$
[see Eq.~\eqref{eq:k1}] in the limit of vanishing lattice strength, provided that $k_1$ belongs to the first Brillouin zone, i.e., $k_1 \leq q/2$; more in general,
if the value of $k_1$ falls in the $\ell$th Brillouin zone, then $k_s$ tends to $|k_1 - (\ell - 1) q|$ [see, for example, the results for $q = 1.5k_1^0$ in
Fig.~\ref{fig:ks_sz_contr_bulk}(a), for which $\ell = 2$]. We will refer to the state described above as the polarized Bloch-wave phase.
\begin{figure}
\includegraphics{figure2_ks_sz_contr_bulk.eps}
\caption{(a) Condensation quasimomentum, (b) magnetic polarization, and (c) contrast of the fringes as a function of the dimensionless lattice
strength $s$ for $\Omega/E_r = 1.0$. In each panel we show the results for three different values $q = 1.5k_1^0$ (green dash-dotted line),
$q = 2k_1^0$ (red solid line), and $q = 3k_1^0$ (blue dashed line) of the momentum transfer. The other parameters are $\bar{n} g_1 / E_r = 0.4$
and $g_2/g_1 = 0.0012$.}
\label{fig:ks_sz_contr_bulk}
\end{figure}
As one increases $s$, the states with $\bar{K} \neq 0$ in Eq.~\eqref{eq:bloch_waves} become more and more populated at the expense
of the $\bar{K} = 0$ state. At the same time, the quasimomentum $k_s$, where Bose-Einstein condensation takes place, moves in
the direction of the wave vector $q/2$, dictated by the optical potential~\cite{Zhang2013,Chen2016}.\footnote{The last statement is true
as long as the value of $q$ is not much smaller or much larger than $2k_1^0$. In the opposite case a different behavior can take place, i.e.,
the condensation quasimomentum $k_s$ moves toward the center of the Brillouin zone at $k=0$. This effect never occurs for the values
of $q$ considered in the present work, hence we will not discuss it further.} This results in the decrease of the magnitude of the magnetic
polarization $\langle\sigma_z\rangle$, as well as in the increase of the contrast of the fringes~\eqref{eq:contrast}. In Fig.~\ref{fig:ks_sz_contr_bulk}
we plot these quantities as a function of $s$, for a fixed value $\Omega/E_r = 1.0$ of the Raman coupling and for several choices of the lattice wave
vector. In particular, for $q = 2k_1^0$ (red solid line), i.e., for a momentum transfer very close to the roton wave vector, the decrease of the magnetic
polarization occurs very rapidly as one increases the optical lattice strength $s$. This behavior of the nonlinear response of the system to the external
lattice potential~\eqref{eq:periodic_pot} is deeply connected to the enhancement of the compressibility discussed in Sec.~\ref{subsec:roton_dens_resp}.
Notice also that the quasimomentum $k_s$ lies extremely close to the edge of the first Brillouin zone for any $s$ in the $q = 2k_1^0$ case. Instead, for
$q = 1.5k_1^0$ (green dash-dotted line) and $q = 3k_1^0$ (blue dashed line) the reduction of $|\langle\sigma_z\rangle|$ and the shift of $k_s$ towards
the value $q/2$ take place for larger lattice intensities.
At the critical lattice strength $s_\mathrm{cr}$, whose value depends on $\Omega$ (see Fig.~\ref{fig:phase_diag_Omega_s}),
the condensation quasimomentum $k_s$ coincides with the wave vector of the periodic potential $q/2$, i.e., with the edge of
the first Brillouin zone, as shown in Fig.~\ref{fig:ks_sz_contr_bulk}(a). Since the two Bloch states having quasimomentum $\pm q/2$
are physically identical, the corresponding magnetic polarization must vanish [see Fig.~\ref{fig:ks_sz_contr_bulk}(b)]. As a consequence,
the system undergoes a second-order transition to an unpolarized Bloch wave phase. In this latter phase the populations of the various
plane-wave states of Eq.~\eqref{eq:bloch_waves} are balanced such that $|C_{-q/2-\bar{K}}|^2 = |C_{q/2+\bar{K}}|^2$ and
$\theta_{-q/2-\bar{K}} = \pi/2 - \theta_{q/2+\bar{K}}$ for any $\bar{K}$; hence, the $\sigma_x \mathcal{P}$ and the
$\sigma_z \mathcal{T}$ symmetries are restored in the unpolarized state. Concerning the contrast of the density
modulations~\eqref{eq:contrast}, which we plot in Fig.~\ref{fig:ks_sz_contr_bulk}(c), it keeps growing with $s$, although more slowly
than in the polarized phase. Notice that in the unpolarized Bloch wave phase the contrast is much larger than the typical values exhibited
in the striped phase in the absence of the lattice~\cite{Martone2014}; this is true even in the case $q = 2k_1^0$, where the magnetic phase
transition, for small values of $\Omega$ [see Fig.~\ref{fig:phase_diag_Omega_s}(b)], takes place at extremely small values $s$ of the optical
lattice strength.
\subsection{Phase diagram}
\label{subsec:phase_diag}
In Fig.~\ref{fig:phase_diag_Omega_s} we show the phase diagram of the system in the $\Omega$-$s$ plane, for the same values of the average
density $\bar{n}$ and of the interaction parameters $g_1$ and $g_2$ as the previous figures. Each panel of Fig.~\ref{fig:phase_diag_Omega_s}
corresponds to a different value of $q$ [$q = 1.5k_1^0$ in Fig.~\ref{fig:phase_diag_Omega_s}(a), $q = 2k_1^0$
in Fig.~\ref{fig:phase_diag_Omega_s}(b), and $q = 3k_1^0$ in Fig.~\ref{fig:phase_diag_Omega_s}(c)]. In each diagram, the blue solid line
indicates the critical Raman coupling $\Omega_\mathrm{tr}$ at which, for a given value of $s$, the first-order transition from the mixed to
the polarized Bloch wave phase occurs. The red dashed lines show instead the behavior, as a function of $\Omega$, of the critical lattice
strength $s_\mathrm{cr}$ separating the polarized and unpolarized Bloch wave states. The predictions for $s_\mathrm{cr}$ obtained using the
ideal Bose gas model (see the Appendix) are also reported (red dotted lines).
\begin{figure}
\includegraphics{figure3_phase_diag_Omega_s.eps}
\caption{(Color online) Phase diagram as a function of the Raman coupling $\Omega$ and of the dimensionless lattice strength $s$
for (a) $q = 1.5k_1^0$, (b) $q = 2k_1^0$, and (c) $q = 3k_1^0$. The blue solid line represents the first-order transition from
the mixed to the polarized Bloch wave phase. The red dashed line corresponds to the second-order transition from the polarized
to the unpolarized Bloch wave phase. For the latter, the prediction of the ideal gas model is also given (red dashed line).
The other parameters are $\bar{n} g_1/ E_r = 0.4$ and $g_2/g_1 = 0.0012$.}
\label{fig:phase_diag_Omega_s}
\end{figure}
The mixed phase discussed in Sec.~\ref{subsec:mixed_regime} appears in a very narrow region close to the left edge of the diagrams of
Figs.~\ref{fig:phase_diag_Omega_s}(a) and \ref{fig:phase_diag_Omega_s}(c). The value of $\Omega_\mathrm{tr}$ is maximum at $s = 0$,
where it coincides with $\Omega^0_\mathrm{tr}$, and becomes smaller and smaller as the lattice intensity $s$ increases.
The rest of the diagram is occupied by the unmixed polarized and unpolarized configurations. One can notice that, if interactions are included,
the critical lattice strength $s_\mathrm{cr}$ shows a nonmonotonic behavior as a function of $\Omega$ in the $q = 1.5k_1^0$ and $q = 3k_1^0$
cases. The large values exhibited by $s_\mathrm{cr}$ at small $\Omega$ can be understood by recalling that, in this limit, the system is strongly
polarized in the absence of the lattice; hence, large lattice intensities are required to achieve the transition to the unpolarized phase.
In this regime of small Raman coupling the contribution of the interaction energy, being of the order of $\bar{n} g_1$ [see Eq.~\eqref{eq:soc_energy}],
is less important than that of the lattice potential~\eqref{eq:periodic_pot}, proportional to $s E_\mathrm{latt}$; consequently, the interactions
do not change qualitatively the behavior with respect to the prediction of the ideal Bose gas model (dotted lines). In the opposite limit of large
$\Omega$, the transition takes place at low values of $s$ in the ideal gas model. Since $k_1^0/k_0 \ll 1$, the energy scale $E_\mathrm{latt}$
associated with the lattice potential is very small and the interactions play an important role in determining the properties of the gas.
This leads to an increase of $s_\mathrm{cr}$ with respect to the values obtained within the ideal Bose gas model.
As $q \to 2 k_1^0$, from both above and below, the regions of the mixed and of the unpolarized Bloch wave phases get closer and closer and eventually
they touch, without the occurrence of the polarized Bloch wave phase in between. This process results in a different behavior of the phase diagram of
Fig.~\ref{fig:phase_diag_Omega_s}(b) for $q = 2k_1^0$ with respect to those of Figs.~\ref{fig:phase_diag_Omega_s}(a) and \ref{fig:phase_diag_Omega_s}(c).
In particular, the value of $s_\mathrm{cr}$ exhibits a monotonic increasing dependence on $\Omega$ for $\Omega > \Omega^0_\mathrm{tr}$. The difference
with respect to the $q = 1.5k_1^0$ and $q = 3k_1^0$ cases can be better understood by taking into account that the ideal gas model, for $q = 2k_1^0$,
gives rise to an unpolarized phase as soon as $s \neq 0$. The increase of $s_\mathrm{cr}$ with $\Omega$ is hence a pure effect of the two-body
interactions. In Fig.~\ref{fig:phase_diag_Omega_s}(b) we have not included the line separating the mixed and the unpolarized Bloch wave phases; its
calculation would require an accurate estimate of the energy difference between such phases, which, however, is extremely small if $q = 2k_1^0$ [this is
due to the closeness of the wave vectors $k_s$ and $q/2$ appearing in the ansatz~\eqref{eq:ansatz} for the wave function].
It is also interesting to understand how the phase diagram discussed above depends on the density and on the interaction strengths of the BEC.
For a fixed ratio $g_2/g_1$, a larger density $\bar{n}$ tends to enhance the effects of the interactions. In particular, they favor the polarized
Bloch wave phase with respect to the unpolarized one, thus increasing the critical lattice strength $s_\mathrm{cr}$ for any given $\Omega$.
The same also happens for the critical Raman coupling $\Omega_\mathrm{tr}$ at fixed $s$, which leads to the enlargement of the region of
the mixed phase, in agreement with Ref.~\cite{Chen2016}. Concerning the role of the coupling constants, an increase of the ratio $g_2/g_1$
for a given density favors the configurations with vanishing magnetic polarization, i.e., the mixed and the unpolarized Bloch wave phases. This
yields a smaller value of $s_\mathrm{cr}$ and a larger critical Raman coupling $\Omega_\mathrm{tr}$. Notice that the latter result agrees with
the prediction of Eq.~\eqref{eq:Omega_tr_0} for the $s = 0$ value of $\Omega_\mathrm{tr}$. We finally point that, unlike the mixed phase, the
unpolarized Bloch wave phase can also appear in the presence of ferromagnetic spin-dependent interactions, i.e., when $g_2 < 0$.
\subsection{Role of magnetic fluctuations}
\label{subsec:mag_fluct}
A further important signature of the magnetic phase transition occurring in our system is given by the behavior of the magnetic susceptibility
$\chi^{}_M$. This can be evaluated in a way similar to the one described in Sec.~\ref{subsec:roton_dens_resp} for the compressibility:
One adds a small perturbation $V_\lambda = -\lambda \sigma_z$ to the single-particle Hamiltonian~\eqref{eq:soc_ham} and calculates the
expectation value $\langle\sigma_z\rangle_\lambda$ of the longitudinal spin operator on the perturbed ground state. Then the magnetic
susceptibility is given by $\chi^{}_M \equiv \mathrm{d}\langle\sigma_z\rangle_\lambda/\mathrm{d}\lambda|_{\lambda=0}$. The results show that $\chi^{}_M$
exhibits a divergent behavior at the transition between the polarized and the unpolarized Bloch wave phases (see Fig.~\ref{fig:chi_mag_bulk}),
revealing the occurrence of strong magnetic fluctuations.
\begin{figure}
\includegraphics{figure4_chi_mag_bulk.eps}
\caption{Magnetic susceptibility $\chi^{}_M$ as a function of the dimensionless lattice strength $s$. The parameters are the same
as in Fig.~\ref{fig:ks_sz_contr_bulk}, i.e., $\Omega/E_r = 1.0$; $q = 1.5k_1^0$ (green dash-dotted line), $q = 2k_1^0$ (red solid line),
and $q = 3k_1^0$ (blue dashed line); $\bar{n} g_1/ E_r = 0.4$; and $g_2/g_1 = 0.0012$.}
\label{fig:chi_mag_bulk}
\end{figure}
Notice that, if one considers values of $q$ too close to the roton momentum, such as $q=2k_1^0$ (red solid line in Fig.~\ref{fig:chi_mag_bulk}),
the value of the magnetic susceptibility grows very rapidly as one ramps up the lattice intensity starting from $s = 0$. Furthermore, $\chi^{}_M$
remains large for a wide range of values of $s$ much bigger than $s_\mathrm{cr}$. As a consequence, the curve $\langle\sigma_z\rangle$ vs $s$ of
Fig.~\ref{fig:ks_sz_contr_bulk}(b) and the phase diagram of Fig.~\ref{fig:phase_diag_Omega_s}(b) are significantly affected by the presence of
a nonvanishing effective magnetic field $\delta$ or by an even tiny difference between the coupling parameters $g_{\uparrow\uparrow}$ and
$g_{\downarrow\downarrow}$, as in the case of $^{87}$Rb. Figure~\ref{fig:chi_mag_bulk} also shows that the increase of the magnetic susceptibility
is notably slower in the $q=1.5k_1^0$ (green dash-dotted line) and especially in the $q=3k_1^0$ (blue dashed line) cases. Hence, the phase diagram
is much more stable if one takes the momentum transfer $q$ significantly different from the roton wave vector. We finally mention that the enhancement
of $\chi^{}_M$ can be further reduced by working at larger values of the Raman coupling $\Omega$, as we do in the following section.
\subsection{Magnetic phase transition in harmonic traps}
\label{subsec:harm_trap}
The results discussed so far were based on a calculation in the thermodynamic limit. However, it is important to understand how they are modified
if one considers a finite system in the presence of a trapping potential of harmonic type. For this purpose we have solved numerically the
three-dimensional Gross-Pitaevskii equation for a gas of $N = 1.8 \times 10^5$ $^{87}$Rb atoms, in the presence of an elongated trap with frequencies
$(\omega_x,\omega_y,\omega_z) = 2\pi \times (50,50,140)$ Hz. The scattering lengths are $a_{\uparrow\uparrow} = 100.83 a_B$ and
$a_{\downarrow\downarrow} = a_{\uparrow\downarrow} = 100.37 a_B$, $a_B$ being the Bohr radius. The value of the momentum transfer due to the Raman
laser is set to $k_0 = 5.52$ $\mu$m$^{-1}$. All the above parameters correspond to the conditions of the experiment of Ref.~\cite{Lin2011}.
For this calculation we have taken the Raman coupling $\Omega/E_r = 3.0$ and the momentum transfer $q = 3k_1^0 = 1.98 k_0$. This choice corresponds
to working close to the minimum of the critical lattice intensity $s_\mathrm{cr}$ appearing in Fig.~\ref{fig:phase_diag_Omega_s}(c); furthermore,
following the discussion in Sec.~\ref{subsec:mag_fluct}, one can expect that magnetic fluctuations will be less important in this regime of the
parameters.
In Fig.~\ref{fig:ks_sz_contr_trap} we report our predictions (blue squares) for the quasimomentum [Fig.~\ref{fig:ks_sz_contr_trap}(a)], the magnetic
polarization [Fig.~\ref{fig:ks_sz_contr_trap}(b)], and the contrast of the fringes in the integrated density $n_1(x) = \int \mathrm{d} y \, \mathrm{d} z \, n(\vec{r})$
close to the center of the sample [Fig.~\ref{fig:ks_sz_contr_trap}(c)]. Notice that, since $g_3 > 0$ for the above values of the scattering lengths,
the quasimomentum and the magnetic polarization must actually be negative; in the figure we only plot their magnitudes. Our results clearly show that
the emergence of the magnetic phase transition can be detected also in trapped configurations. Furthermore, we have checked that, at variance with
the choice of small $\Omega$ and of $q$ too close to the roton momentum, the results of the simulations are not significantly perturbed by the inclusion
of a small but finite effective magnetic field $\delta$.
\begin{figure}
\includegraphics{figure5_ks_sz_contr_trap.eps}
\caption{(a) Condensation quasimomentum, (b) magnetic polarization, and (c) contrast of the fringes as a function of the dimensionless
lattice strength $s$ for $\Omega / E_r = 3.0$ and $q = 3k_1^0 = 1.98 k_0$. The dashed line corresponds to the case of an infinite system with
$\bar{n} g_1 / E_r = 0.4$, $g_2/g_1 = 0.0012$, and $g_3/g_2 = 2.0$. The squares are the results for a three-dimensional trapped gas in the
conditions described in the text, which yield a density at the center of the trap equal to the average density $\bar{n}$ of the infinite system.}
\label{fig:ks_sz_contr_trap}
\end{figure}
Also in Fig.~\ref{fig:ks_sz_contr_trap} we also report the results (blue dashed lines) for an infinite system whose average density $\bar{n}$
coincides with the density at the center of the trapped BEC described above. By comparing the two sets of data, one immediately sees that the
results for the two systems are slightly different from the quantitative point of view. Indeed, as a consequence of the spatially varying density,
the values of the magnetic polarization in the trapped gas are smaller than in the infinite system and they decrease more rapidly with increasing
$s$. However, the qualitative behavior is very similar to the one in the absence of the trap.
\section{Conclusion}
\label{sec:conclusion}
The static response to a weak density perturbation of a spin-orbit-coupled Bose-Einstein condensate
in the plane-wave phase exhibits an important dependence on the wave vector of the external potential.
If the latter is close to the momentum at which the roton minimum occurs, the response acquires a very
large value close to the transition to the striped phase. This phenomenon is related to the vanishing of
the energy of the roton minimum and reveals the presence of important highly nonlinear effects. By
studying the ground state of the system in the presence of a one-dimensional optical lattice, we have shown
that such effects are strongly connected with the occurrence of a second-order magnetic phase transition
to an unpolarized Bloch wave phase. In this state the contrast of the density modulations can achieve much
larger values than in the striped phase appearing at low Raman couplings in the absence of the lattice. The
critical lattice intensity needed to achieve the magnetic phase transition can assume very tiny values if
the lattice wave vector is close to the roton momentum; in the opposite case, the transition can take place
at larger lattice strengths. The phase transition is accompanied by a divergent behavior of the magnetic
susceptibility. The stability of the system against magnetic fluctuations can be enhanced by choosing the
momentum transfer due to the optical lattice far enough from the roton momentum and considering large values
of the Raman coupling. Our predictions for infinite systems have been confirmed also by Gross-Pitaevskii
calculations in a three-dimensional trap with realistic parameters, which opens the possibility of exploring
these nonlinear phenomena in current experiments.
\begin{acknowledgments}
Useful correspondence and discussions with Peter Engels, Amin Khamehchi, and Yun Li are acknowledged.
G.I.M. was partially supported by the PRIN Grant No.~2010LLKJBX ``Collective quantum phenomena:
From strongly correlated systems to quantum simulators'' and by INFN through the project ``QUANTUM.''
The research leading to these results received funding from the European Research Council under
European Community's Seventh Framework Programme (FR7/2007-2013 Grant Agreement No. 341197).
T.O., C.Q., and S.S. were supported by ERC through the QGBE grant, by the QUIC grant of the Horizon2020
FET program, and by Provincia Autonoma di Trento. T.O. was also supported by the EU-FET Proactive grant
AQuS, Project No. 640800.
\end{acknowledgments}
|
\section{Introduction}
For any metric theory of gravity what is required is construction of a divergence free second rank symmetric tensor from the metric and curvature alone. It can be achieved in two possible ways: (a) varying the action, which is an invariant quantity constructed from Riemann curvature and the metric, (b) by making trace of Bianchi derivative of a properly defined ``Riemann tensor'' to be vanishing. For Einstein gravity, variation of the Einstein-Hilbert action with respect to metric tensor or trace of Bianchi identity of Riemann tensor, both lead to Einstein's equations \cite{gravitation,Carroll:1997ar,Parattu:2015gga}. In order to obtain physically meaningful solutions with proper initial value problem ensuring unique evolution as well as absence of ghosts, it is imperative that equations of motion continue to remain second order as is the case for Einstein gravity. The key requirement is that it should remain so even when we go to higher orders in the Riemann tensor. This requirement uniquely singles out Lovelock
polynomial action that always yields second order gravitational field equations \cite{Lovelock:1971yv,Kastor:2008xb,Padmanabhan:2013xyr}. Lovelock action is essentially a sum of dimensionally continued Euler densities, being of order $m$ ($\geq 0$) in curvature, such that $m=1$ gives the Einstein-Hilbert Lagrangian, $m=2$ provides the Gauss-Bonnet Lagrangian and so on. Alternatively, it is possible to define $m$th order Lovelock ``Riemann'' tensor \cite{Dadhich:2008df,Kastor:2012se,Camanho:2015hea} and trace of its Bianchi derivative yields a divergence free second rank symmetric tensor, an analogue of Einstein tensor, as obtained from variation of $m$th order Lovelock action.
Lanczos-Lovelock gravity is important for various reasons. It is generally believed that Einstein-Hilbert action is an effective action, valid at small enough energy (or large length) scales. At high energy, near the Planck scale gravity cannot be described by Einstein-Hilbert action alone but should be supplemented by higher curvature terms\footnote{Note that, requiring the field equations to be of second order uniquely singles out Lanczos-Lovelock Lagrangian from all other possibilities.}. For example, supersymmetric string theory \emph{exactly} reproduces Gauss-Bonnet term as quadratic correction to Einstein-Hilbert action \cite{Boulware:1985wk}. Further thermodynamic interpretation of Einstein's equations generalize in a straightforward and natural \emph{but} non-trivial manner to Lanczos-Lovelock gravity. The same is true for other correspondences between thermodynamics and gravity in the context of general relativity\ as well \cite{Chakraborty:2015wma,Chakraborty:2014joa,Chakraborty:2014rga,Chakraborty:2015hna}. Finally pure Lovelock theories, i.e., a single term in the Lovelock polynomial,
exhibit very interesting features. These include --- (a) there is a close connection between pure Lovelock and dimensionally continued black holes \cite{Dadhich:2012ma,Garraffo:2008hu}, (b) gravity is kinematic in all critical $d=2m+1$ dimensions, i.e., vacuum is pure Lovelock flat \cite{Dadhich:2012cv}, (c) bound orbits exist for a given $m$ in all $2m+1<d<4m+1$ dimensions, in contrast, for Einstein gravity they do so only in four dimensions \cite{Dadhich:2013moa} and finally (d) equipartition of gravitational and non-gravitational energy defines location of black hole horizon \cite{Chakraborty:2015kva}.
There have been various other interesting results in the context of cosmology within the Lovelock gravity, in particular the Einstein-Gauss-Bonnet gravity. For example, the scheme of dynamical compactification in higher curvature gravity introduced in \cite{MuellerHoissen:1985mm} has been applied to Einstein-Gauss-Bonnet gravity in higher dimensions to obtain exact solutions in various cosmological models \cite{Pavluchenko:2009wn,Maeda:2014nua,Canfora:2013xsa,Alexeyev:2000eb,Elizalde:2006ub}. The dynamical compactification scheme has also been applied in numerical investigations of cosmological scenarios in
\cite{Canfora:2014iga,Canfora:2008iu}. In particular it has been observed in \cite{Canfora:2016umq} that one can have standard Friedmann dynamics even though the gravity theory is that of Einstein-Gauss-Bonnet. It has also been noticed that in higher curvature gravity in higher dimensions the approach to big bang singularity has some discerning features in comparison to four dimensional gravity \cite{Deruelle:1989he}. However all the above works had Einstein-Gauss-Bonnet gravity as the background stage, also the results derived depend crucially on the compactification scheme while here we are concerned with a single term in the Lovelock polynomial and without any compactification scheme.
To these interesting features alluded above we wish to add in this note yet another interesting property of pure Lovelock gravity. In Einstein gravity potential goes as $1/r^{d-3}$ that takes the familiar $1/r$ form only in four dimensions and none else. In contrast, for pure Lovelock it goes as $1/r^\alpha$ with $\alpha=(d-2m-1)/m$ \cite{Dadhich:2012ma}, and hence there exists dimension spectrum $d=3m+1$ for $1/r$ potential. It should be emphasized that we are interested in exact solutions of gravitational field equations, not just Newtonian limit and hence must be contrasted with earlier approaches in \cite{ArkaniHamed:1998rs,ArkaniHamed:1999hk}. It is also possible to arrive at a similar situation in the weak field if the higher dimensional spacetime is assumed to be fractal in nature. However as demonstrated in \cite{Rami_CTP}, in order to arrive at a $1/r$ gravitational potential it was necessary to work in five dimensions with some particular choices related to the fractal structure of
the spacetime (see also \cite{Tarasov_2015,Tarasov:2006bc,Tarasov:2014gua,Stillinger_1977}). However, the scenario we are interested in is completely different conceptually, as the spacetime manifold in consideration does not posses any fractal structure and the solution is derived from the fully non-linear theory in a series of spacetime dimensions. In our case the existence of $1/r$ potential is solely due to modified gravity theory rather than modified structure of spacetime manifold.
Therefore, the above feature can happen only in pure Lovelock gravity and hence this is also a nice distinguishing property for pure Lovelock theories. What it means is that, from a purely gravitational standpoint there is no way to fathom from solar system observations whether it is four dimensional Einstein or seven dimensional pure Gauss-Bonnet or in general $(3m+1)$-dimensional pure $m$th order Lovelock gravity. Further, it also turns out, that the same result applies to cosmological scenarios as well. In particular, in this work we have demonstrated that not only for dust but also for a linear equation of state parameter, which is being regularly considered in standard cosmology, the behaviour of the universe (as far as the gravitational sector is concerned) in pure Lovelock theories in $d=3m+1$ is indistinguishable from that of general relativity\ in four dimensions, except for propagation of gravitational waves which would have two polarizations only in four dimensions and none else.
Hence from purely gravitational standpoint, one cannot distinguish between any member of the $m$th order Lovelock in $d=3m+1$ dimensions. The discerning features do however come from two counts: (a) propagation of gravitational wave, and (b) introduction of matter fields. It turns out that gravitational wave can have two polarizations only in four dimensions. On the other hand, for existence of atoms, it is necessary that Maxwell field should admit bound orbits around a static charge which again can happen only in four dimensions. Thus these two features root us firmly to four dimensions which is in consonance with the standard paradigm of higher dimensional theories, where it is envisaged that all matter fields remain confined to the usual four dimension (termed as $3$-brane) while gravity is however free to propagate in higher dimensions. Therefore, at least, whether from a gravitational viewpoint we may as well be living in a higher dimensional spacetime rather than the usual four with gravity being
described by pure Lovelock instead of Einstein gravity, is the question we wish to pose and wonder about in this note.
\section{Lovelock Gravity}
In a $d$-dimensional spacetime, gravity in general, can be described by an action functional involving arbitrary scalar functions of metric and curvature, but not derivatives of curvature. In general, variation of this arbitrary Lagrangian would lead to an equation having fourth order derivatives of the metric. For it to be of second order, gravitational Lagrangian is constrained to be of the following Lovelock form,
\begin{equation}
L=\sum _{m}c_{m}L_{m}=\sum _{m}c_{m}\frac{1}{2^{m}}\delta ^{a_{1}b_{1}a_{2}b_{2}\ldots a_{m}b_{m}}_{c_{1}d_{1}c_{2}d_{2}\ldots c_{m}d_{m}}R^{c_{1}d_{1}}_{a_{1}b_{1}}R^{c_{2}d_{2}}_{a_{2}b_{2}}\ldots R^{c_{m}d_{m}}_{a_{m}b_{m}},
\end{equation}
where $\delta ^{pq\ldots}_{rs\ldots}$ stands for completely antisymmetric determinant tensor. The case $m=2$ is Gauss-Bonnet Lagrangian which is quadratic in curvature and reads as $L_{\rm GB}=(1/2)(R^{abcd}R_{abcd}-4R^{ab}R_{ab}+R^{2})$. Remarkably it also appears in the low energy effective theory of supersymmetric strings, as first pointed out in \cite{Boulware:1985wk}. Lovelock Lagrangian is a sum over $m$ where each term is a homogeneous polynomial in curvature and has a dimensionful coupling constant. Further, complete antisymmetry of the determinant tensor demands $d\geq 2m$, else it would vanish identically. Even for $d=2m$ Lanczos-Lovelock Lagrangian reduces to total derivative --- pure surface term. Lovelock Lagrangian $L_{m}$ is therefore non-trivial only in dimension $d\geq 2m+1$. Note that pure Lovelock gravity is kinematic in all critical odd $d=2m+1$ dimensions because $m$th order Riemann is entirely given in terms of corresponding Ricci, hence it has no non-trivial vacuum solution
\cite{Dadhich:2012ma,Camanho:2015hea}. Thus non-trivial vacuum solutions only exist in dimensions $d\geq 2m+2$. Finally, variation of the Lagrangian with respect to metric variation, for pure Lovelock theories lead to the following second order equation,
\begin{equation}\label{pure_Eq_01}
{}^{(m)}E^{a}_{b}\equiv -\frac{1}{2^{m+1}}\delta ^{ac_{1}d_{1}\ldots c_{m}d_{m}}_{ba_{1}b_{1}\ldots a_{m}b_{m}}R^{a_{1}b_{1}}_{c_{1}d_{1}}\ldots R^{a_{m}b_{m}}_{c_{m}d_{m}}=8\pi T^{a}_{b}.
\end{equation}
Since there appears no derivatives of curvature, hence the equation is as expected of second order. Though not directly visible, second derivative too appears linearly; i.e., it is quasilinear, thereby ensuring unique evolution.
\section{Black Holes}
Let us now consider pure Lovelock vacuum spacetime and write spherically symmetric metric in Schwarzschild gauge incorporating the null energy condition, as given by
\begin{equation}\label{pure_ansatz}
ds^{2}=-f(r)dt^{2}+\frac{dr^{2}}{f(r)}+r^{2}d\Omega_{d-2}^{2},
\end{equation}
where $f(r)=1+2\Phi(r)$ and $d\Omega _{d-2}^{2}$ is metric on $(d-2)$-sphere. Then the equation, ${}^{(m)}E^{t}_{t}=0$ could be trivially integrated as it takes the simple form $(r^{d-2m-1} \Phi^m)' =0$ where prime is derivative relative to $r$, and so we obtain
\begin{equation}
\Phi (r)=-\frac{M}{r^\alpha} , \, \, \, \alpha=(d-2m-1)/m,
\end{equation}
where $M$ is a constant of integration which identifies to $m$th root of the ADM mass \cite{Kastor:2008xb}. In general for static spherically symmetric metric in Schwarzschild gauge, the equation ${}^{(m)}E^{t}_{t}=0$ is the first integral of ${}^{(m)}E^{\theta}_{\theta}=0$, and so we have complete vacuum solution for pure Lovelock static black hole \cite{Dadhich:2012ma}.
Clearly there is no way to distinguish between four dimensional Schwarzschild black hole and $(3m+1)$ dimensional $m$th order pure Lovelock black hole because the gravitational potential goes as $1/r$ for all of them. It is worth emphasizing that it is not only the Newtonian limit that behaves as $1/r$, but the full non-linear Lovelock field equation that has the following exact solution,
\begin{equation}
ds^{2}_{d=3m+1}=-\left(1-\frac{2M}{r}\right)dt^{2}+\left(1-\frac{2M}{r}\right)^{-1}dr^{2}
+r^{2}d\Omega _{3m-1}^{2}
\end{equation}
such that on the four-dimensional hypersurface characterized by $\theta _{3}=\theta _{4}=\cdots =\theta _{3m-3}=\pi/2$, the spacetime metric is \emph{identical} to the Schwarzschild solution in four spacetime dimensions. Note that we have not employed any compactification scheme in order to arrive at this solution. Hence any observable tests in solar system physics cannot distinguish between $m$th order Lovelock gravity in $d=3m+1$ dimensions and Einstein gravity in four dimensions.
As a matter of fact dimensional spectrum is given by $d=(2+\alpha)m+1$ corresponding to a given value of $\alpha$. The choice $\alpha=2$ at the first level corresponds to five dimensional Einstein with dimensional spectrum $d=4m+1$ while $\alpha=1/2$ for six dimensional Gauss-Bonnet with dimensional spectrum $d=5m/2+1$. Also note that since horizon structure and potential are the same, all properties associated with horizon, e.g., Hawking radiation, should remain the same. Another test that can distinguish these two black holes may be \emph{black hole entropy}. But \emph{remarkably} in the case under consideration, entropy density of the black hole goes as product of $(m-1)$ curvature, each of which scales as $r^{-3(m-1)}$ for $d=3m+1$, while area element scales as $r^{3m-1}$ and hence entropy would always scale as $r_h^{2}$, where $r_{h}$ is horizon radius irrespective of Lovelock order $m$. For pure Lovelock static black holes, it has been shown that thermodynamical parameters, temperature and entropy bear
the same relation to horizon radius in all critical odd and even $d=2m+1, 2m+2$ dimensions \cite{Dadhich:2012eg}.
The next question arises, would black hole, like Schwarzschild in four dimension, be stable in higher dimensions? In stability analysis one decomposes metric perturbations into three types, tensor, vector and scalar perturbations \cite{Regge:1957td,Vishveshwara:1970cc,Ishibashi:2003ap}. Stability analysis of Lovelock gravity black holes under scalar, vector and tensor perturbations was carried out in \cite{Takahashi:2009xh}, which was further particularized to pure Lovelock black holes in \cite{Gannouji:2013eka}. Under metric perturbation, Riemann tensor and hence all geometric constructs will be perturbed. Using which one can obtain a master differential equation for perturbation, which is like Schr\"{o}dinger equation with a potential. Only when corresponding differential operator is a positive (or zero) self-adjoint operator, solutions will be perturbatively stable if and only if there are no negative eigenvalues. This in turn leads to the following condition
\begin{equation}
\frac{d-3m-1}{mr}\geq 0.
\end{equation}
Clearly it would be stable for $d\geq3m+1$ which means $\alpha =(d-2m-1)/m\geq1$. Also note that the same conclusion was drawn from a different perspective for stability of small black holes \cite{Camanho:2013kfa} in generic Lovelock theories. This is another interesting feature of our work. The above result explicitly shows that black hole for Einstein gravity in four or higher dimensions is always stable because $\alpha\geq1$, while it is unstable for pure Lovelock gravity in all the critical even $d=2m+2$ dimensions \footnote{Note that $d=3m+1$ is not the critical dimension for $m$-th order pure Lovelock.} for which $\alpha = 1/m <1$ \cite{Gannouji:2013eka}. Thus the determining factor is $\alpha\geq1$. Since it is stable under tensor perturbations and hence it would be so for vector perturbations as well.
The final question corresponds to the weak field limit of gravity, which is generally obtained by assuming the metric to be a perturbation over a flat background, i.e., $g_{ab}=\eta _{ab}+h_{ab}$ with $|h_{ab}|\ll 1$. In general Lovelock polynomial the Einstein-Hilbert term is always present, thus the weak field limit will contain linear order terms in the perturbation. However in the case of pure Lovelock theories only a single term in the Lovelock polynomial is present, there is no term linear in $h_{ab}$ because curvature tensor for the background geometry vanishes. For ``m''-th order pure Lovelock theories the weak field limit would correspond to a differential equation having m powers of the perturbation $h_{ab}$. For example, in the case of pure Gauss-Bonnet gravity the gravitational field equation in the weak field limit would involve terms like $(1/4)\square \bar{h}_{ab}\square \bar{h}^{ab}$, $\partial _{a}\partial _{b}\bar{h}_{cd}\partial ^{m}\partial ^{b}\bar{h}^{cd}$ and so on, where $\bar{h}_{ab}
=h_{ab}-(1/2)\eta _{ab}h$ \cite{rado}. Hence the Newtonian limit is not reproduced.
This suggests that for Lovelock gravity, perhaps one should consider perturbation around constant, not zero, curvature spacetime; i.e., dS/AdS which would yield non-trivial equation. For that purpose we write $g_{ab}=\bar{g}_{ab}+h_{ab}$, where $\bar{g}_{ab}$ is the background dS/AdS spacetime and $h_{ab}$ is the perturbation. Then the curvature tensor to leading order in the perturbation leads to,
\begin{equation}
R^{ab}_{cd}=M\delta ^{a}_{c}\delta ^{b}_{d}+\frac{1}{2}\nabla _{c}\left(-\nabla ^{a}h^{b}_{d}+\nabla _{d}h^{ab}+\nabla ^{b}h^{a}_{d}\right)-\left(c\leftrightarrow d\right)
\end{equation}
where $M$ is proportional to the Ricci scalar and hence to the cosmological constant. Use of this expansion leads to the following expression of the field equation expanded upto first order in the perturbation from \ref{pure_Eq_01} as,
\begin{eqnarray}
-\frac{M^{m-1}}{2}&\times&\delta ^{aa_{1}b_{1}\ldots a_{m}b_{m}}_{bc_{1}d_{1}\ldots c_{m}d_{m}}\delta ^{c_{2}}_{a_{2}}\delta ^{d_{2}}_{b_{2}}\ldots \delta ^{c_{m}}_{a_{m}}\delta ^{d_{m}}_{b_{m}}\Bigg[\delta ^{c_{1}}_{a_{1}}\delta ^{d_{1}}_{b_{1}}
\nonumber
\\
&+&\frac{1}{2}\nabla _{a_{1}}\left(-\nabla ^{c_{1}}h^{d_{1}}_{b_{1}}+\nabla _{b_{1}}h^{c_{1}d_{1}}+\nabla ^{d_{1}}h^{c_{1}}_{b_{1}}\right)\Bigg]=8\pi T^{a}_{b}
\end{eqnarray}
One can evaluate the left hand side explicitly. The first term will lead to the standard cosmological constant contribution, which as subtracted away will lead to the following evolution equation of the perturbation in vacuum spacetime
\begin{equation}
\square \bar{h}_{ab}=0
\end{equation}
where $\bar{h}_{ab}=h_{ab}-(1/2)\bar{g}_{ab}h$ and we have used the gauge condition $\nabla _{a}\bar{h}^{ab}=0$. The above expression is essentially the same as the one obtained from Einstein gravity in the weak field limit. Thus in presence of cosmological constant the weak field limit of Einstein gravity can be reproduced.
This clearly shows that pure Lovelock gravity picks out constant curvature spacetime as background rather than the flat Minkowski. In this context it is interesting to note that on inclusion of $\Lambda$ we shall have $f(r)=1-(\Lambda r^{2m} + M/r^{d-2m-1})^{1/m}$ which for large $r$ goes to $f(r)= 1 - \bar\Lambda r^2 - 2\bar{M}/r^{d-3}$ \cite{Dadhich:2012ma}. This is the $d$-dimensional Einstein-dS solution. It is remarkable that even though there was no Einstein term in the equation of motion yet the solution goes over to Einstein-dS asymptotically.
At this stage, let us summarize the key results that we have arrived at in the context of black holes in pure Lovelock theories. First of all, it is possible to have \emph{exact} Schwarzschild solution on a four dimensional hypersurface for $m$th order pure Lovelock gravity in $d=3m+1$ dimensions. Secondly, the semi-classical properties of the spacetime, e.g., Hawking radiation and black hole entropy for this $(3m+1)$ dimensional Lovelock black hole remain indistinguishable from the four dimensional Einstein gravity. Moreover it turns out that the black hole in $(3m+1)$ dimension is indeed stable alike the Schwarzschild solution in four dimensions. Finally correct weak field limit in these theories can also be obtained when expanded around de Sitter or anti-de Sitter solution.
Since the weak field limit of pure Lovelock gravity is identical to that of Einstein, it would be interesting to understand the propagation of gravitational waves in higher dimensional pure Lovelock background, in particular the structure of the corresponding quadrupole formula will be worth pursuing. Even without going into details, which we keep for the future, one can provide some general comments. Note that presence of cosmological constant is mandatory in pure Lovelock theories in order to arrive at the correct weak field limit, while the quadrupole moment formula is derived in general relativity\ without a cosmological constant. Recently, there have been renewed interest in the gravitational wave propagation in de Sitter spacetime, whose formulation differs quite significantly from the standard scenario and results in a modified quadrupole moment formula \cite{Ashtekar:2015lxa,Ashtekar:2015ooa,Ashtekar:2014zfa}. Thus even in the context of pure Lovelock theories the quadrupole formula will be different from the general relativity\
counterpart. However it would be interesting to see does it bear any resemblance with the quadrupole formula for de Sitter spacetime or not? In particular whether the dimensional spectrum also holds in this context?
\section{FLRW Cosmology}
From black hole we now come to cosmology and ask the question, would the standard FLRW cosmology remain the same for the dimensional spectrum $d=3m+1$? That means this feature is not only restricted to vacuum solutions but is true for cosmology as well. Thus the standard FLRW metric for homogeneous and isotropic universe reads,
\begin{equation}
ds^{2}=-dt^{2}+a^{2}(t)\left[dr^{2}+r^{2}d\Omega _{d-2}^{2}\right].
\end{equation}
From Eq. (2) we obtain for Hubble parameter, $H=\dot{a}/a$ as,
\begin{equation}\label{pure_cosmo_01}
H^{2m}=\frac{(d-2m-1)!}{(d-1)!}16\pi G\rho,
\end{equation}
and for isotropic pressure
\begin{eqnarray}\label{pure_cosmo_02}
\frac{1}{2}\frac{(d-2)!}{(d-2m-1)!}H^{2m-2}&\times&\left[2m\frac{\ddot{a}}{a}+(d-2m-1)H^{2}\right]=-8\pi Gp
\end{eqnarray}
Deriving the conservation equation from \ref{pure_cosmo_01} and \ref{pure_cosmo_02} is straightforward, one needs to take time derivative of \ref{pure_cosmo_01} and then use \ref{pure_cosmo_02} to eliminate $\dot{H}$ term, resulting in $\dot{\rho}+(d-1)(\rho +p)H=0$. Assuming an equation of state $p=\omega \rho$ with a constant parameter $\omega$ we obtain time evolution of scale factor as $a(t)\sim t^{2m/[(d-1)(1+w)]}$. It is immediate that for $d=3m+1$, one obtains $a(t)\sim t^{2/[3(1+w)]}$, identical to the standard general relativistic cosmology in these contexts.
Note that in the above cosmological scenario the equation of state parameter, $\omega$ can have any value including $1/3, 0, -1$ respectively for early time radiation, non-relativistic dust fluid and the cosmological constant -- the dark energy. Thus it fully accords to standard cosmological model which does however have open problems of dark matter and dark energy. For addressing these problems, one either seeks modification of gravity theory or some exotic matter field. Our aim was not to address these questions instead to point out that there can exist exactly the same cosmological scenario in higher dimensions in an appropriate gravity theory.
However all this would have been fine, had matter fields could have also existed in higher dimensions, which is however not the case. Hence even though at face value the cosmological solutions are identical to their general relativistic counterparts, inclusion of matter (in particular, radiation) acts as the discriminator. That is, our Universe may be four dimensional with Einstein gravity or seven dimensional with Gauss-Bonnet gravity, and there is no way to break this dilemma from a pure gravitational standpoint. But, introduction of matter clearly breaks it in favour of 4-dimensional Einstein gravity.
\section{Discussion}
There is a very strong and compelling observational evidence for $1/r$ potential in four dimensions in the framework of general relativity which is a metric theory of gravity. It is curious to ask, could gravitational potential have the same behaviour in any other dimension in some other metric theory? It is remarkable that the answer is yes, and it is uniquely the pure Lovelock theory that offers a spectrum of dimensions, $d=3m+1$ for which potential is indeed $1/r$. This means all gravitational observations, except the gravitational waves, involving stellar objects and black holes will not be able to distinguish between any two members of this spectrum. Not only that, cosmological dynamics is also driven by the same inverse square law and hence it is expected that should also remain the same. And that indeed is the case as we have just shown above. Below we point out the main outcomes of this work:
\begin{itemize}
\item We have explicitly demonstrated that the two most fundamental problems for a gravitational theory, namely compact objects and cosmology might have the same dynamics for this dimensional spectrum from a purely gravitational standpoint. This is a very important conclusion that simply follows from pure Lovelock gravity.
\item We have also demonstrated that all the properties associated with the black hole horizon, e.g., Hawking radiation, remains identical with the four-dimensional counterpart. Moreover, the horizon entropy would always scale as $r_h^{2}$, where $r_{h}$ is the horizon radius, irrespective of the Lovelock order $m$ (also see, \cite{Dadhich:2012eg}). This is another intriguing fact associated with the pure Lovelock theories.
\item We have also shown that, when perturbed with respect to constant curvature background, the weak field limit of Einstein gravity can be reproduced. It also follows that pure Lovelock static black hole solution with $\Lambda$, asymptotically goes over to $d$-dimensional Einstein-dS solution \cite{Dadhich:2012ma}. It is intriguing that even though there was no Einstein term present in the equation of motion, yet the solution goes over to Einstein-dS asymptotically.
\end{itemize}
In particular, purely based on gravitational observations it would be difficult to decipher whether it is the usual four dimensional spacetime with Einstein gravity or ten dimensional spacetime with pure third order Lovelock gravity.
Besides the question of existence of matter in higher dimensions, there is yet another gravitational feature that may also turn out to be a discriminator. It is the propagation of gravitational wave and its number of polarizations and degrees of freedom. It had been argued that number of degrees of freedom in a $d$ dimensional spacetime remain the same as given by $d(d-3)/2$ for Einstein as well as Lovelock theory \cite{Teitelboim:1987zz,Deser:2011zk}, and it has recently been further reinforced for pure Lovelock gravity \cite{Dadhich:2015ivt}. In that case two polarizations for gravitational wave could only occur in four dimension and none else. Let us see how count for degrees of freedom works in a $d$ dimensional spacetime by separating out the usual graviton degrees, gauge vector, and scalar degrees in the light of Kaluza-Klein decomposition. Let us envisage the metric in $d$ dimensions with parametrization: $g_{AB} =({g_{ab}, g_{ai}, g_{ij}})$ where $A,B=0,\ldots,(d-1); a,b=0,1,2,3; i,j=4,
5,\ldots,(d-1)$. Then there is the usual graviton $g_{ab}$ with two degrees, $(d-4)$ gauge vector fields with each two degrees and $(d-4)(d-3)/2$ scalars. One can easily check that they add up to give the correct number of degrees of freedom $d(d-3)/2$. Thus extra degrees of freedom would be $(d+1)(d-4)/2$, which becomes, $3(m-1)(3m+2)/2$ for $d=3m+1$.
For seven dimensional Gauss-Bonnet spacetime, extra degrees of freedom include, three gauge vector fields, each having two degrees and six scalars. So there would appear extra vector and scalar fields which could possibly distinguish between the usual four dimensional Einstein and seven dimensional Gauss-Bonnet spacetime. However for a static source in seven dimensions, the metric is orthogonal and thus eliminating all the gauge fields and three scalar fields, and the remaining three scalars will be washed out on dimensional reduction, $\theta_i = \pi/2$. Despite all this a comprehensive field theoretic analysis has to be undertaken on the lines of Kaluza-Klein theory \cite{Appelquist:1983vs} before anything definitive could be said. However, it seems that propagation of gravitational waves or inclusion of matter in the theory will vote for four-dimensional Schwarzschild spacetime alone, unless existence of matter in higher dimensions is assumed.
We thus conclude that even though it is difficult gravitationally to distinguish between any two members of the dimensional spectrum, except for the number of degrees of freedom, raising the question, ``Could we have as well lived in higher dimensions?", perhaps yes, if matter could have existed there. It possibly points to the fact that gravity is abundant in complexity and richness of structure that one is never disappointed as one probes deeper and wider. The present question offers an excellent example of this richness as it coaxes one to wonder, could the universe been higher dimensional or not? This is an important question that would have broad scientific as well as philosophical fallout.
\section*{Acknowledgements}
Research of S.C. is funded by the SERB-NPDF grant (PDF/2016/001589) from SERB, Government of India. He also thanks Kinjalk Lochan for fruitful discussions. N.D. thanks Xian Camanho for a very enlightening discussion on degrees of freedom and also acknowledges warm hospitality of Albert Einstein Institute where the manuscript was finalized.
|
\section{Introduction}
\label{sec:introduction}
Supersymmetry (SUSY) (for a review see e.g. \cite{Chung:2003fi})
remains an attractive candidate for new physics beyond the Standard Model (SM),
even if there is to date no direct evidence for it at colliders, most notably the Large Hadron Collider (LHC). However, there remain good motivations for considering SUSY, which are worth repeating, namely that it opens up the possibility for gauge coupling unification, provides a viable dark matter (DM) candidate such as the
R-parity stabilized lightest neutralino, and addresses the big hierarchy problem of the SM. Despite the lack of evidence for SUSY at the LHC, including the lack of non-standard flavour signals in LHCb detector, almost for two decades there remains one stubborn experimental inconsistency in the SM coming from the anomalous magnetic moment of the muon,
which is often overlooked or ignored for one or another reason.
It is well known that SUSY can account for this inconsistency, provided that there are light
sleptons and charginos, which by themselves are not inconsistent with LHC constraints on new coloured particles.
It remains an intriguing question, which we shall address in this paper,
whether this data can be accounted for by a well motivated unified SUSY model
consistent with other collider and non-collider constraints including DM.
The magnetic moment of the muon, as predicted by the Dirac equation, is related to the particle's spin $\bm{S}$ by
\begin{align}
\bm{M} = g_{\mu} \dfrac{e}{2 m_{\mu}} \bm{S}
\label{eq:Muon_magnetic}\, ,
\end{align}
where, at classical level, the gyromagnetic ratio is $g_{\mu} = 2$. Small deviations from this value are induced at the quantum level and can be parametrized by the so called anomalous magnetic moment of the muon
%
\begin{align}
a_{\mu} = \dfrac{g_{\mu} - 2}{2}
\label{eq:amu}\, .
\end{align}
$a_\mu$ is one of the most precisely measured quantities in modern particle physics. The E821 experiment at the Brookhaven National Laboratory has measured $a_\mu$ to $0.54\,$ppm~\cite{Bennett:2006fi,Agashe:2014kda}, resulting in
\begin{align}
a_\mu^{\rm exp} =
11 659 2091(63) \times 10^{-11}.
\end{align}
New experiments at Fermilab~\cite{Grange:2015fou} and J-PARC \cite{Saito:2012zz} promise to improve this accuracy by a factor of four. The SM theory prediction is of a comparable accuracy (for useful reviews, see \cite{Jegerlehner:2009ry,Blum:2013xva,Benayoun:2014tra,Knecht:2014sea}). This prediction includes QED corrections to five loops~\cite{Aoyama:2012wk} (see also \cite{Kataev:2012kn,Lee:2013sx,Kurz:2013exa,Kurz:2015bia}) as well as weak corrections to two loops~\cite{Czarnecki:2002nt,Gnendiger:2013pva} and hadronic corrections~\cite{Nyffeler:2009tw,Davier:2010nc,Hagiwara:2011af,Benayoun:2012wc,Colangelo:2014dfa,Kurz:2014wya,Colangelo:2014qya,Colangelo:2014pva,Pauk:2014rfa,Colangelo:2015ama,Benayoun:2015gxa} (see also \cite{Blum:2014oka,Jin:2015eua,Chakraborty:2015ugp,Aubin:2015rzx,Blum:2015you} for lattice QCD evaluations).
The uncertainties in the hadronic corrections, which rely on data for $e^+e^- \to $ hadrons, vary somewhat between authors. In all combinations, there remains a significant tension between experiment and theoretical prediction. This discrepancy ranges from
\begin{align}
\Delta a_\mu = a_\mu^{\rm exp} - a_\mu^{\rm SM} = 237(86) \times 10^{-11}
\label{eq:Damu_1}
\end{align}
to
\begin{align}
\Delta a_\mu = a_\mu^{\rm exp} - a_\mu^{\rm SM} = 278(80) \times 10^{-11},
\label{eq:Damu_2}
\end{align}
which are $2.8\,\sigma$ and $3.4\,\sigma$ tensions respectively~\cite{Knecht:2014sea}. In the interest of compatibility with other studies, here we will use the deviation of experiment from the SM prediction quoted in Ref.~\cite{Agashe:2014kda}, which is
\begin{align}
\Delta a_{\mu} = a_\mu^{\rm exp} - a_\mu^{\rm SM} = 288(80) \times 10^{-11}
\label{eq:Damu_3}\, .
\end{align}
If this discrepancy persists and reaches even higher significance
when confronted with new experiments and/or improvements to the SM hadronic contributions,
it may become a sign of new physics beyond the SM. In particular, within supersymmtric models, the deviation from the SM prediction may be totally or partially attributed to smuon-neutralino and sneutrino-chargino loops~\cite{Grifols:1982vx,Ellis:1982by,Chakrabortty:2013voa,Chakrabortty:2015ika,Barbieri:1982aj,Kosower:1983yw,Yuan:1984ww,Romao:1984pn,Lopez:1993vi,Moroi:1995yh,Martin:2000cr,Czarnecki:2001pv,Cho:2011rk,Endo:2011mc,Endo:2011xq,Endo:2011gy,Evans:2012hg,Endo:2013bba,Mohanty:2013soa,Ibe:2013oha,Akula:2013ioa,Okada:2013ija,Endo:2013lva,Bhattacharyya:2013xma,Gogoladze:2014cha,Kersten:2014xaa,Li:2014dna,Chiu:2014oma,Badziak:2014kea,Calibbi:2015kja,Kowalska:2015zja,Wang:2015rli}. Although $\Delta a_\mu$ may be accommodated in the Minimal Supersymmetric Standard Model (MSSM) (see e.g.\ \cite{Ibe:2013oha,Endo:2013bba}) with its large number of free parameters, finding a suitable value in more constrained supersymmetric models can be challenging. For example, in the well studied Constrained MSSM (CMSSM), in which the supersymmetric soft-breaking masses are given common values at some high energy scale, it is difficult to achieve the desired value of $a_\mu$~\cite{Bechtle:2012zk,Buchmueller:2012hv,Balazs:2013qva}. Of course, if one is willing to attribute only part of the discrepancy to supersymmetric effects, then simple models of Grand Unification that satisfy all constraints become viable (see e.g.\ \cite{Miller:2013jra,Miller:2014jza}) but are no more attractive for explaining $a_\mu$ than the SM.
Another possible class of models which could address $(g-2)_{\mu}$
are the SUSY GUT models with normal mass hierarchy with non-universal
scalar masses for the first two and the third generation of sfermions~\cite{Baer:2004xx}.
Also, the $(g-2)_{\mu}$ problem can be addressed in the essentialy non-universal
model such as the phenomenological MSSM (pMSSM) scenario
\cite{Djouadi:1998di} which is based on the following simplifying assumptions:
\begin{itemize}
\item First and second generation universality for low energy soft masses
$m_{Q_1},m_{U_1},m_{D_1},m_{L_1},m_{E_1}$ (equal to $m_{Q_2},m_{U_2},m_{D_2},m_{L_2},m_{E_2}$,
respectively)
\item Separate low energy soft masses for third generation scalar masses
$m_{Q_3},m_{U_3},m_{D_3},m_{L_3},m_{E_3}$
\item Separate low energy gaugino masses $M_1,M_2,M_3$
\item Separate trilinear parameters $A_t,A_b,A_{\tau}$
\end{itemize}
In this paper we will investigate contributions to $a_\mu$ that arise from a
conceptually different MSSM model
based on a high energy (GUT scale) Pati-Salam gauge group combined with an $A_4 \times Z_5$ family symmetry~\cite{King:2014iia}.
The point is that this model was initially motivated not by $(g-2)_\mu$
but by the fact that it provides an excellent description of quark and lepton masses, mixing and CP violation. The model predicts the following high energy (GUT scale) soft mass parameters:
\begin{itemize}
\item A universal high energy soft scalar mass for all left-handed squarks and sleptons of all three families, $m_0$
(i.e. $m_{Q_i}$ and $m_{L_i}$ are unified into $m_0$ at the GUT scale )
\item Three high energy soft mass parameters for the right-handed squarks and leptons, one for each
family $m_1$, $m_2$, $m_3$ (i.e. $m_{U_i}$, $m_{D_i}$ and $m_{E_i}$ are unified into $m_i$ at the GUT scale,
respectively for $i=1,2,3$ )
\item Separate high energy gaugino masses $M_1,M_2,M_3$
\item Separate trilinear parameters $A_t,A_b,A_{\tau}$
\end{itemize}
These soft mass boundary conditions are consistent with the (s)particle groupings
dictated by the model as shown in figure \ref{A2Z}.
We will show that this model has also a great potential
to predict $a_\mu$ that is in agreement with the experimental value, while simultaneously providing a viable Dark Matter candidate, maintaining vacuum stability and remaining consistent with all experimental constraints.
In section~\ref{sec:the-model} we will describe the model in some detail, and in section~\ref{sec:g-2} we clarify the leading contributions to $\Delta a_\mu$. We will discuss constraints from experiment, including collider constraints and those on the Dark Matter relic density, in section~\ref{sec:constraints}. We present our results, including some example scenarios, in section~\ref{sec:results}. Finally we investigate vacuum stability for these example scenarios in section~\ref{sec:vacuum-stability}, before concluding in section~\ref{sec:conclusion}.
\section{The Model}
\label{sec:the-model}
An ``A to Z of flavour with Pati-Salam'' based on the Pati-Salam gauge group has been proposed
\cite{King:2014iia} as sketched in figure \ref{A2Z}.
The Pati-Salam symmetry leads to $Y^u=Y^{\nu}$, where the columns of the Yukawa matrices
are determined by flavon alignments.
The first column is proportional to the alignment $(0,e,e)$,
the second column proportional to the orthogonal alignment
$(a,4a,2a)$, and the third column is proportional to the alignment $(0,0,c)$,
where $e\ll a\ll c$ gives the hierarchy $m_u\ll m_c\ll m_t$.
This structure predicts a Cabibbo angle
$\theta_C\approx 1/4$
in the diagonal $Y^d\sim Y^e$ basis enforced by the
first three alignments.
It also predicts a normal neutrino mass hierarchy with
$\theta_{13}\approx 9^{\circ}$, $\theta_{23}\approx 45^{\circ}$
and $\delta \approx 260^{\circ}$ \cite{King:2014iia}.
The model is based on the Pati-Salam (PS) gauge group,
with $A_4\times Z_5$ (A to Z) family symmetry,
\begin{equation}
SU(4)_{C} \times SU(2)_L \times SU(2)_R\times A_4 \times Z_5.
\label{422A4Z5}
\end{equation}
The quarks and leptons are unified in the PS representations as follows,
\begin{equation}
\begin{aligned}
F &= (4,2,1)_i = \left(\begin{array}{cccc}u&u&u&\nu\\
d&d&d&e\end{array}\right)_i \rightarrow (Q_i,L_i), \\
F^c_i &= (\bar{4},1,2)_i =
\left(\begin{array}{cccc}u^c&u^c&u^c&\nu^c\\
d^c&d^c&d^c&e^c\end{array}\right)_i \rightarrow (u^c_i,d^c_i,\nu^c_i,e^c_i ),
\label{ql}
\end{aligned}
\end{equation}
\noindent where the SM multiplets $Q_i,L_i,u^c_i,d^c_i,\nu^c_i,e^c_i$
resulting from PS breaking are also shown and the
subscript $i\ (=1,2,3)$ denotes the family index.
The left-handed quarks and leptons form an $A_4$ triplet $F$,
while the three (CP conjugated) right-handed fields $F^c_i$ are $A_4$ singlets, distinguished by $Z_5$ charges $\alpha, \alpha^3,1$, for $i=1,2,3$, respectively.
Clearly the Pati-Salam model cannot be embedded into an $SO(10)$ Grand Unified Theory (GUT)
since different components
of the 16-dimensional representation of $SO(10)$ would have to transform differently under $A_4\times Z_5$,
which is impossible, but the PS gauge group and $A_4$ could emerge directly from
string theory.
\begin{figure}[tbp]
\centering
\includegraphics[width=0.4\textwidth]{figures/cube2.pdf}
\includegraphics[width=0.4\textwidth]{figures/cube2_PS.pdf}
\vspace*{-4mm}
\caption{$A$ to $Z$ of flavour with Pati-Salam, where $A\equiv A_4$ and $Z\equiv Z_5$.
The left-handed families form a triplet of $A_4$ and are
doublets of $SU(2)_L$.
The right-handed families are distinguished by $Z_5$
and are doublets of $SU(2)_R$.
The $SU(4)_C
$ unifies the quarks and leptons with leptons as the fourth colour,
depicted here as white.
} \label{A2Z}
\vspace*{-2mm}
\end{figure}
In the SUSY theory at the GUT scale, from \eqref{ql} there are therefore four different matter
multiplets: $F, F^c_1,F^c_2,F^c_3$, corresponding to the left-handed block and the three
distinct
right-handed blocks
in figure \ref{A2Z} respectively. The GUT-scale scalar soft mass of $F$ will be called $m_0$, while the soft masses of
$F^c_1,F^c_2,F^c_3$ will be denoted $m_1,m_2,m_3$, respectively,
as discussed in the introduction. The model therefore provides novel
SUSY boundary conditions for soft masses at the GUT scale, more constrained than the general
MSSM, but less so than the CMSSM. As we shall see, this allows us to account for the
experimentally observed g-2 of the muon, and will lead to a distinctive and novel low energy
superpartner mass spectrum, with characteristic signatures at the LHC.
The Pati-Salam gauge group is broken at the GUT scale to the SM gauge group,
\begin{equation}
SU(4)_{C}\times SU(2)_L \times SU(2)_R\rightarrow SU(3)_C\times SU(2)_L \times U(1)_Y,
\end{equation}
by PS Higgs,
$H^c$ and $\overline{H^c}$,
\begin{equation}
\begin{aligned}
{H^c} & = & (u^c_H,d^c_H,\nu^c_H,e^c_H ) \in (\bar{4},1,2), \\
{\overline{H}^c} & = & (\bar{u}^c_H,\bar{d}^c_H,\bar{\nu}^c_H,\bar{e}^c_H ) \in (4,1,2).
\end{aligned}
\end{equation}
These acquire vacuum expectation values (VEVs) in the ``right-handed neutrino'' directions, with equal VEVs
close to the GUT scale $2\times 10^{16}$ GeV,
\begin{equation}
\langle {H^c}\rangle = \langle {\nu^c_H}\rangle
=\langle{\overline{H^c}}\rangle=\langle{\bar{\nu}^c_H}\rangle \sim 2\times 10^{16} \ {\rm GeV},
\label{PS}
\end{equation}
so as to maintain supersymmetric gauge coupling unification.
The model will involve Higgs bi-doublets of two kinds, $h_u$ which lead to up-type quark and neutrino Yukawa couplings and $h_d$ which lead to down-type quark and charged lepton Yukawa couplings.
In addition a Higgs bidoublet $h_3$, which is also an $A_4$ triplet, is used to give the third family
Yukawa couplings.
After the PS and $A_4$ breaking, most of these Higgs bi-doublets
will get high scale masses and will not appear in the low energy spectrum. In fact only two
light Higgs doublets will survive down to the TeV scale, namely $H_u$ and $H_d$.
The light Higgs doublet $H_u$ with hypercharge $Y=+1/2$, which couples to up-type quarks and neutrinos,
is a linear combination of components of the Higgs bi-doublets of the kind $h_u$ and $h_3$,
while the light Higgs doublet $H_d$ with hypercharge $Y=-1/2$,
which couples to down-type quarks and charged leptons,
is a linear combination of components of Higgs bi-doublets of the kind $h_d$ and $h_3$,
\begin{equation}
h_u,h_3 \rightarrow H_u, \ \ \ \
h_d,h_3 \rightarrow H_d.
\label{H}
\end{equation}
Therefore, below the GUT scale, the model reduces to the usual MSSM, but with GUT scale boundary conditions
for soft scalar masses as discussed above.
\section{Anomalous magnetic moment of the muon}
\section{\texorpdfstring{One-loop contributions to $\Delta a_{\mu}$}{One-loop contributions to Delta a\_mu}}
\label{sec:g-2}
The magnetic moment of a massive charged particle is a result of the interaction of its spin with the electromagnetic field. At zeroth order in perturbation theory, the gyromagnetic ratio is predicted to be 2 for every massive particle with semi-integer spin. Deviations from this classical value emerge at the loop-level, where besides SM corrections, new physics contributions may also be relevant. This is indeed the case for the anomalous magnetic moment of the muon, where one-loop supersymmetric contributions are represented in the Feynman diagrams of figure \ref{fig:1-loop-amu}.
\begin{figure}[t!]
\centering
\includegraphics[width=0.45\textwidth]{figures/Diagram5g-2.jpg}
\includegraphics[width=0.45\textwidth]{figures/Diagram4g-2.jpg}
\includegraphics[width=0.45\textwidth]{figures/Diagram3g-2.jpg}
\includegraphics[width=0.45\textwidth]{figures/Diagram2g-2.jpg}
\includegraphics[width=0.45\textwidth]{figures/Diagram1g-2.jpg}
\caption{One-loop contributions to the anomalous magnetic moment of the muon for supersymmetric models with low-scale MSSM.}
\label{fig:1-loop-amu}
\end{figure}
These diagrams were computed in \cite{Moroi:1995yh,Endo:2013bba} and give contributions
\begin{subequations}
\label{eq:loops}
\begin{align}
\Delta a_{\mu}^{(A)}&=\left(\dfrac{M_{1}\mu}{m_{\tilde{\mu}_{L}}^{2}m_{\tilde{\mu}_{R}}^{2}}\right)\dfrac{\alpha_{1}}{4\pi}m_{\mu}^{2}\tan\beta\cdot f^{(A)}_{{\rm N}}\left(\dfrac{m_{\tilde{\mu}_{L}}^{2}}{M_{1}^{2}},\dfrac{m_{\tilde{\mu}_{R}}^{2}}{M_{1}^{2}}\right)\, , \label{eq:loops:A} \\
\Delta a_{\mu}^{(B)}&=-\left(\dfrac{1}{M_{1}\mu}\right)\dfrac{\alpha_{1}}{4\pi}m_{\mu}^{2}\tan\beta\cdot f^{(B)}_{{\rm N}}\left(\dfrac{M_{1}^{2}}{m_{\tilde{\mu}_{R}}^{2}},\dfrac{\mu^{2}}{m_{\tilde{\mu}_{R}}^{2}}\right)\, , \label{eq:loops:B} \\
\Delta a_{\mu}^{(C)}&=\left(\dfrac{1}{M_{1}\mu}\right)\dfrac{\alpha_{1}}{8\pi}m_{\mu}^{2}\tan\beta\cdot f^{(C)}_{{\rm N}}\left(\dfrac{M_{1}^{2}}{m_{\tilde{\mu}_{L}}^{2}},\dfrac{\mu^{2}}{m_{\tilde{\mu}_{L}}^{2}}\right)\, , \label{eq:loops:C} \\
\Delta a_{\mu}^{(D)}&=-\left(\dfrac{1}{M_{2}\mu}\right)\dfrac{\alpha_{2}}{8\pi}m_{\mu}^{2}\tan\beta\cdot f^{(D)}_{{\rm N}}\left(\dfrac{M_{2}^{2}}{m_{\tilde{\mu}_{L}}^{2}},\dfrac{\mu^{2}}{m_{\tilde{\mu}_{L}}^{2}}\right)\, , \label{eq:loops:D} \\
\Delta a_{\mu}^{(E)}&=\left(\dfrac{1}{M_{2}\mu}\right)\dfrac{\alpha_{2}}{4\pi}m_{\mu}^{2}\tan\beta\cdot f^{(E)}_{{\rm C}}\left(\dfrac{M_{2}^{2}}{m_{\tilde{\nu}_{\mu}}^{2}},\dfrac{\mu^{2}}{m_{\tilde{\nu}_{\mu}}^{2}}\right)\, , \label{eq:loops:E}
\end{align}
\end{subequations}
with $\alpha_1$ and $\alpha_2$ the $U(1)_{Y}$ and $SU(2)_L$ fine structure constants respectively. The functions $f^{(A,B,C,D)}_{{\rm N}}\left(x,y\right)$ and $f^{(E)}_{{\rm C}}\left(x,y\right)$ are given by
\begin{subequations}
\label{eq:f}
\begin{align}
f^{(A,B,C,D)}_{{\rm N}}\left(x,y\right)&=xy\left[\dfrac{-3+x+y+xy}{\left(x-1\right)^{2}\left(y-1\right)^{2}}+\dfrac{2x\log x}{\left(x-y\right)\left(x-1\right)^{3}}-\dfrac{2y\log y}{\left(x-y\right)\left(y-1\right)^{3}}\right]\, , \label{eq:f:1} \\
f^{(E)}_{{\rm C}}\left(x,y\right)&=xy\left[\dfrac{5-3\left(x+y\right)+xy}{\left(x-1\right)^{2}\left(y-1\right)^{2}}-\dfrac{2\log x}{\left(x-y\right)\left(x-1\right)^{3}}+\dfrac{2\log y}{\left(x-y\right)\left(y-1\right)^{3}}\right]\, , \label{eq:f:2}
\end{align}
\end{subequations}
where we use the superscripts $(A,B,C,D)$ and $(E)$ as a short notation to allow omission of the mass ratio arguments.
As described in \cite{Endo:2013bba}, the loop-functions $f^{(A,B,C,D)}_{{\rm N}}$ and $f^{(E)}_{{\rm C}}$ are monotonically increasing for both $x$ and $y$ and are defined in $0 \leq f_{{\rm N, C}} \leq 1 $. From \eqref{eq:loops}, we see that the size of each $\Delta a_{\mu}^{(i)}$ contribution is largely governed by the pre-factor between brackets on the RHS. Therefore, a large $\mu$ combined with light smuons enhances $\Delta a_{\mu}$ via diagram $(A)$ in figure \ref{fig:1-loop-amu}, while keeping the remaining contributions suppressed. However this solution is not unique and in the limit of small $\mu$ the size of the functions $f^{(A,B,C,D)}_{\rm N}$ and $f^{(E)}_{\rm C}$ themselves may distinguish the dominant contributions among diagrams $(B)$ to $(E)$. In particular, we see from the contour plots of figure \ref{fig:contour}, that for a fixed $(x,y)$, say $x \sim y \sim 0.2$, $f^{(E)}_{\rm C} \sim 0.2$ is approximately one order of magnitude larger than $f^{(A,B,C,D)}_{\rm N} \sim 0.02$. We will see in section~\ref{sec:results} the importance of these functions for the explanation of $\Delta a_\mu$.
\begin{figure}[htb!]
\centering
\includegraphics[width=0.495\textwidth]{figures/fn.jpg}
\includegraphics[width=0.495\textwidth]{figures/fc.jpg}
\caption{Contour plots for $f_{\rm N}\left(x,y \right)$ (left) and $f_{\rm C}\left(x,y \right)$ (right).}
\label{fig:contour}
\end{figure}
\section{Experimental Constraints}
\label{sec:constraints}
Any successful high-energy completion of the SM should satisfy all known low-energy experimental constraints. In
particular, we require our scenarios to conform to measurements of the Dark Matter (DM) relic density and obey
constraints from the direct detection of DM. The current combined best fit of the DM relic density to data from
\textsc{Planck} and \textsc{Wmap} is $\Omega h^2 = 0.1198 \pm 0.0026$ \cite{Ade:2013zuv}. We will also consider smaller
values of $\Omega h^2$, allowing the possibility that our model does not account for DM in its entirety, which opens up
the bound to $\Omega h^2 \in [0.06,0.1224]$. For DM direct detection constraints, we apply the current 90\% upper
confidence level cross-sections for spin-independent models with a WIMP mass of $33\,\textrm{GeV}$, which are given by
$\sigma^{\text{DD-SI}} \leq 7.6 \times 10^{-46} \ \text{cm}^{-2} = 7.6 \times 10^{-10} \ \text{pb}$
\cite{Akerib:2013tjd}. For WIMP masses less or greater than $33\,\textrm{GeV}$ the direct detection bound is weaker, so this
choice is conservative.
Furthermore we require agreement with the recently measured Higgs mass, the correct branching ratios for the decays $b
\to s \gamma$ and $B_s \to \mu^+\mu^-$, and agreement with the $\rho$-parameter. The current combined ATLAS and CMS
measurement of the Higgs boson mass is $m_H = (125.09 \pm 0.21\, (\text{stat.}) \pm 0.11\, (\text{sys.})) \ \text{GeV}$
\cite{Aad:2015zhl}. However, these experimental uncertainties are dominated by our much larger theoretical uncertainty,
and consequently we relax our constraint to scenarios with $m_H = (125.09 \pm 1.5)\ \text{GeV}$. We directly apply
limits on the branching ratios $\text{Br}(b \to s \gamma) = (3.29 \pm 0.19 \pm 0.48) \times 10^{-4}$
\cite{Lees:2012wg} and $\text{Br}(B_s \to \mu^+ \mu^-)$$ = 3.0^{+1.0}_{-0.9} \times 10^{-9}$ \cite{Chatrchyan:2013bka}.
Apart from the fixed experimental constraints, we are free to further modify the parameter space in order to include
some useful features. For example, having light sleptons, especially smuons, is one of these features. This is reasoned
by the fact that light smuons heavily increase the $\Delta a_\mu$ contribution from diagram $(A)$ (see eq.
\ref{eq:loops:A}). Also, having light sleptons grants a suitably higher possibility to explore them during current or
upcoming experimental studies, e.g. at the LHC, due to the comparably clean muonic signals. The corresponding
parameters for the smuons are $m_0$ and $m_2$, which need to be light in order to get light smuons. For actual
parameter choices, see tables \ref{table:Range1}, \ref{table:RangeSmallMu} and \ref{tab:input-parameters}.
Two other useful features are a bino-like LSP (denoted by $\tilde{\chi}$) and a large mass gap between the LSP and the
smuon masses. These characteristics are helpful to provide the correct dark matter relic density while preventing
leptons arising from $\tilde{\mu}^\pm \to \tilde{\chi}\, \mu^\pm$ decays to be soft, which would render them nearly
undetectable at any collider. None of the parameters of our model is directly responsible for these features, so
analysing different scans with different parameter choices is necessary.
As a last point, we have verified that benchmarks we consider below do not violate any of the 8 TeV ATLAS and CMS analyses. This is necessary, since one of the scenarios we have found -- the small $\mu$ scenario -- could give rise to light $\tilde{\chi}^0_1$, $\tilde{\chi}^0_2$ and $\tilde{\chi}^\pm_1$ with comparatively low (few dozen GeV) mass splittings. This region of the parameter space provides distinctive di-lepton or tri-lepton signatures at the LHC which are not observed and which therefore rule out the respective parameter space. To do this verification we have used the chain consisting of \texttt{MadGraph 5.2.2.3}~\cite{Alwall:2014hca} to generate all relevant combinations for chargino-neutralino pair production, \texttt{PYTHIA 6.4}~\cite{Sjostrand:2006za} linked to \texttt{MadGraph} to simulate the parton showering and hadronisation and \texttt{CheckMATE 1.2.1}~\cite{Drees:2013wra} to perform fast detector simulations with \texttt{DELPHES 3.0}~\cite{deFavereau:2013fsa} and event analysis. Using the same set of cuts as the experimental analyses (either CMS or ATLAS), \texttt{CheckMATE} allowed us to establish whether a given point from the parameter space is ruled out or not making use of the data given by the collaborations in their published analyses which are validated in \texttt{CheckMATE}. In particular, we found that tri-lepton signatures explored in Refs. \cite{Aad:2014nua,ATLAS:2013rla} are the most constraining ones for the small $\mu$ region. On the other hand, di-lepton signatures are also worth mentioning, albeit turning out to be less constraining for the parameter space under study.
\section{Results}
\label{sec:results}
After selecting a certain point in parameter space by choosing all relevant model parameters (cf. section \ref{sec:introduction} and \ref{sec:the-model}), we use \texttt{SoftSUSY 3.5.2} \cite{Allanach:2001kg} to generate the mass spectrum of that point and exclude any point with a Higgs mass out of the bounds chosen in section \ref{sec:constraints}. In case the Higgs mass is in bounds, we use \texttt{micrOMEGAs\_3.6.9.2} \cite{Belanger:2013oya} to compute the relic density as well as the remaining constraints described in section \ref{sec:constraints}.
\subsection{An inclusive scan}
\label{subsec:inclusive}
The lack of evidence for strongly interacting superpartners at the LHC puts low scale supersymmetry under pressure. However, while gluinos and squarks of the first two generations need to be heavier than $\sim1.5~\rm{TeV}$, electroweak sector searches are still rather weak. As light supersymmetric particles could be the source for a sizable $\Delta a_{\mu}$ deviation, we investigate scenarios with light smuons and light selectrons in our low scale spectrum that avoid conflict with current experimental exclusion limits. To do this we first preform an inclusive scan on the parameter space varying the GUT scale parameters as shown in table \ref{table:Range1}.
\begin{table}[h]
\centering
\begin{tabular}{|c|rcl|}
\hline
Parameter & \multicolumn{3}{c|}{range} \\
\hline
$\left|A_{\text{tri}}\right|$ & $1$ &--& $3000$ \\
$m_0$, $m_1$, $m_2$ & $1$ &--& $500$ \\
$m_3$ & $1$ &--& $3000$ \\
$m_{H_1}$, $m_{H_2}$& $1$ &--& $3000$ \\
\hline
\end{tabular}
\begin{tabular}{|c|rcl|}
\hline
Parameter & \multicolumn{3}{c|}{range} \\\hline
$\left|M_1\right|$, $\left|M_2\right|$ & $1$ &--& $600$\\
$\left|M_3\right|$ & $1$ &--& $6000$\\
$\tan \beta$ & $5$ &--& $50$ \\
$\operatorname{sgn}\(\mu\)$ & & $\pm 1$ & \\
\hline
\end{tabular}
\caption{Model parameters at the GUT scale. Dimensionful parameters are in GeV.}
\label{table:Range1}
\end{table}
We allow the $SU(3)_C$ gaugino mass, $M_3$, and the third generation right-handed scalar mass, $m_3$, to acquire large values so the stops may provide a significant contribution to the Higgs mass via loops. In figure \ref{fig:gmuon_vs_mu_M1_M2}, we show viable scenarios in the $\Delta a_{\mu} \-- \mu$ (top), $\Delta a_{\mu} \-- M_1$ (bottom-left) and $\Delta a_{\mu} \-- M_2$ (bottom-right) planes, where the light green and orange triangles have too low relic density, the turquoise and salmon circles have only the relic density in bounds and the dark blue and red diamonds have $\Delta a_\mu$ as well as the relic density in bounds.
\begin{figure}[h]
\centering
\includegraphics[width=0.49\textwidth]{figures/gmuon_vs_mu.pdf}\\
\includegraphics[width=0.49\textwidth]{figures/gmuon_vs_M1.pdf}
\includegraphics[width=0.49\textwidth]{figures/gmuon_vs_M2.pdf}
\caption{Viable scenarios in the $\Delta a_{\mu} \-- \mu$ (top) and $\Delta a_{\mu} \-- M_1$ (bottom-left), $\Delta a_{\mu} \-- M_2$ (bottom-right) planes. Dark blue and red diamonds are scenarios with bino-like DM, whereas the light green and orange triangles and turquoise and salmon circles are scenarios with mainly wino and partially higgsino-like DM. The reddish points correspond to a separate scan around the isolated dark blue point in the top plot at small $\mu$. The input parameters are shown in table \ref{table:RangeSmallMu}.}
\label{fig:gmuon_vs_mu_M1_M2}
\end{figure}
It turns out there are two classes of solutions for the correct values of $\Delta a_\mu$, which can be distinguished as a large $\mu$ (the v-shaped bands at $|\mu| \gtrsim 2$ TeV) and a small $\mu$ (the single blue diamond at $\mu \approx 0$ and the red band around it) region. As can be seen, the first class of solutions requires not only a rather large SUSY-preserving mass parameter $\lvert \mu \rvert \gtrsim 2~\rm{TeV}$, but also a soft breaking gaugino mass $\lvert M_1 \rvert \gtrsim 100~\rm{GeV}$. However, if we relax the relic density requirement, we find solutions from the latter class with small $\mu$ and satisfactory values of $\Delta a_{\mu}$. In particular, the isolated dark blue point in the top of figure \ref{fig:gmuon_vs_mu_M1_M2} at small $\mu$ has $\Delta a_{\mu}=25.96 \times 10^{-10},~\mu = 262.5~\rm{GeV},~M_1 = -475.8~\rm{GeV}~{\rm{and}}~M_2 = 588.9~\rm{GeV}$, and predicts a LSP (bino) with mass $m_{B^0} = 200.1~\rm{GeV}$.
In figure \ref{fig:mu_vs_M1}, we display the correlation between $\mu$, $M_1$ and $m_{\tilde{\chi}^0_1}$, where we have selected only those points where the lightest neutralino wave function is dominated by the bino component. In this figure, we show that for the rare points with light $\mu$, the smallness of the $U(1)_Y$ gaugino mass at the GUT scale ensures that the LSP is predominantly bino via RGE running to the electroweak scale.
\begin{figure}[H]
\centering
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=.95\textwidth]{figures/mu_vs_M1.pdf}
\end{minipage}%
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=1.225\textwidth]{figures/mu_vs_M1-bino.pdf}
\end{minipage}
\caption{Correlation between $\mu(Q)$ and $M_1$. The left plot shows points satisfying the given constraints, where the dark blue and red diamonds show points fulfilling both the relic density and $\Delta a_\mu$ constraints, turquoise and salmon circles only have the relic density in bounds, but not $\Delta a_\mu$, and light green and orange triangles refer to points with neither the relic density nor $\Delta a_\mu$ in bounds. The plot on the right shows the same data, but the colour gradients now correspond to the LSP mass $m_{\tilde{\chi}^0_1}$ for the inclusive scan (blue points, see table \ref{table:Range1}) and the small $\mu$ scan (red points, see table \ref{table:RangeSmallMu}), as indicated by the colour bars on the right.}
\label{fig:mu_vs_M1}
\end{figure}
\subsection{\texorpdfstring{Small $\mu$}{Small mu}}
\label{subsec:small_mu}
As we verified in section \ref{subsec:inclusive}, there are two preferred regions compatible with the correct value for the anomalous magnetic moment of the muon. We first investigate the small $\mu$ region corresponding to solutions in the vicinity of the isolated band on the top panel of figure \ref{fig:gmuon_vs_mu_M1_M2}. We perform a dedicated scan to generate small $\mu$ and the ranges used for the input parameters at the GUT scale can be found in table~\ref{table:RangeSmallMu}.
\begin{table}[h]
\centering
\begin{tabular}{|c|rcl|}
\hline
Parameter & \multicolumn{3}{c|}{Range} \\
\hline
$A_{\rm{tri}}$ & $-4000$ &--& $-2300$ \\
$m_0$ & $400$ &--& $700$ \\
$m_1$ & $300$ &--& $500$ \\
$m_2$ & $200$ &--& $400$ \\
$m_3$ & $200$ &--& $2000$ \\
$m_{H_1}$, $m_{H_2}$& $1500$ &--& $2500$ \\
\hline
\end{tabular}
\begin{tabular}{|c|rcl|}
\hline
Parameter & \multicolumn{3}{c|}{Range} \\\hline
$M_1$ & $-500$ &--& $-100$\\
$M_2$ & $100$ &--& $600$\\
$M_3$ & $750$ &--& $1200$\\
$\tan \beta$ & $15$ &--& $35$ \\
$\operatorname{sgn}\( \mu \)$& & $+ 1$ & \\
&&&\\
\hline
\end{tabular}
\caption{Theory parameters at the GUT scale. The soft-SUSY breaking parameters are given in GeV.}
\label{table:RangeSmallMu}
\end{table}
We show in figure \ref{fig:mneut_vs_smuon_small_mu} the results obtained for this scan in the $m_{\tilde{\chi}^0_1}$ vs $m_{\tilde{\mu}_R}$ plane, where only points with positive contribution to the anomalous magnetic moment of the muon are displayed. We observe two clear bands and a bulk region corresponding to distinct regions, where dark matter efficiently annihilates due to different physics processes. In particular, the vertical band with $m_{\tilde{\chi}^0_1}\lesssim 50~{\rm GeV}$ corresponds to LSP annihilation via $Z$ boson resonant decay, whereas the band with $m_{\tilde{\chi}^0_1}\gtrsim 60~{\rm GeV}$ the annihilation into visible SM particles is possible due to Higgs boson exchange.
\begin{figure}[h]
\centering
\includegraphics[scale=.45]{figures/mneut_vs_smuon_small_mu-L.pdf}
\includegraphics[scale=.45]{figures/mneut_vs_smuon_small_mu-R.pdf}
\caption{Lightest neutralino mass vs. smuon masses. All dark blue diamonds are bino-like, whereas the light green triangles and turquoise circles are wino-like. The orange pentagons represent the benchmark points defined in table \ref{tab:benchmark_tab_small}.}
\label{fig:mneut_vs_smuon_small_mu}
\end{figure}
The lower diagonal band with $m_{\tilde{\chi}^0_1} \sim m_{\tilde{\mu}_R}$ corresponds to the neutralino-smuon co-annihilation region whereas the bulk region on top of this band shows scenarios where dark matter co-annihilates with non-smuon NLSP.
\begin{figure}[h]
\centering
\includegraphics[scale=.45]{figures/mass_gaps_small_mu-L.pdf}
\includegraphics[scale=.45]{figures/mass_gaps_small_mu-R.pdf}
\caption{Mass gaps between the smuon and lightest neutralino masses $m_{\tilde{\mu}_{L/R}}$ and $m_{\tilde{\chi}^0_1}$. All dark blue diamonds are bino-like, whereas the light green triangles and turquoise circles are wino-like. The orange pentagons represent the benchmark points defined in table \ref{tab:benchmark_tab_small}.}
\label{fig:mass_gaps_small_mu}
\end{figure}
In figure \ref{fig:mass_gaps_small_mu}, we show the mass differences for $m_{\tilde{\mu}_{L/R}} - m_{\tilde{\chi}^0_1}$ versus the lightest neutralino mass. While the mass gap for the left handed smuon never deceeds 200 GeV, mass gaps for the right handed smuon can be as small as 1 GeV, thus rendering any muons emerging from smuon decays nearly undetectable. However, the orange benchmark points have both smuon masses $\gtrsim$ 100 GeV, which prevents the muons from smuon decays to be soft.
For such small values of $\mu$ it may seem that the leading contributions to $\Delta a_\mu$ arise from diagrams $(B),~(C),~(D)$ and $(E)$ in figure \ref{fig:1-loop-amu} as the factor of $\tfrac{1}{\mu}$ in equations \eqref{eq:loops:B} to \eqref{eq:loops:E} becomes large for small $\mu$. However, as we discussed in section \ref{sec:g-2}, the functions $f^{(A,B,C,D)}_{\rm N} \left( x,y \right)$ and $f^{(E)}_{\rm C} \left( x,y \right)$ may also play an important role and should not be disregarded in this analysis. In order to understand which diagrams are indeed relevant for enhancing $\Delta a_{\mu}$ we show in figure \ref{fig:f_vs_gmuon} each individual contribution $\Delta a^{(X)}_{\mu}$ against the corresponding $f_{N,C}\left( x, y \right)$ function and the total $\Delta a_{\mu}$ in the color scale.
\begin{figure}[H]
\centering
\includegraphics[scale=0.43]{figures/f_vs_gmuon-a.pdf}
\includegraphics[scale=0.43]{figures/f_vs_gmuon-b.pdf}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.43]{figures/f_vs_gmuon-c.pdf}
\includegraphics[scale=0.43]{figures/f_vs_gmuon-d.pdf}
\includegraphics[scale=0.43]{figures/f_vs_gmuon-e.pdf}
\caption{Individual contributions for the $\Delta a^{(i)}_{\mu}$ terms, with $i=\left\{A,B,C,D,E\right\}$ in equations \eqref{eq:loops}, vs. the $f_{N,C}\left( x, y \right)$ functions. The color scale indicates the total value value of $\Delta a_{\mu}$. The orange pentagons represent the benchmark points defined in table \ref{tab:benchmark_tab_small}.}
\label{fig:f_vs_gmuon}
\end{figure}
The only relevant positive contributions are coming from diagrams $(B)$ and $(E)$. This agrees with equations \eqref{eq:loops:B} and \eqref{eq:loops:E} as in our scan $M_1$ is negative and both $\mu$ and $M_2$ are positive. Furthermore, the leading contributions to $\Delta a_\mu$ are also coming from these two diagrams and the reason for such an enhancement is the dependency on the $f_{\rm N,C} \left( x,y \right)$ functions. In particular, as the right-handed smuon is always lighter than its left-handed counterpart, we have that $f^{(B)}_{{\rm N}}\left(\tfrac{M_{1}^{2}}{m_{\tilde{\mu}_{R}}^{2}},\tfrac{\mu^{2}}{m_{\tilde{\mu}_{R}}^{2}}\right) \gg f^{(C,D)}_{{\rm N}}\left(\tfrac{M_{1,2}^{2}}{m_{\tilde{\mu}_{L}}^{2}},\tfrac{\mu^{2}}{m_{\tilde{\mu}_{L}}^{2}}\right)$, which explains the enhancement of digram $(B)$ and the suppression of the absolute value of $\Delta a^{(X)}_{\mu}$ for diagrams $(C)$ and $(D)$. For the particular case of diagram $(E)$, one could also expect a strong suppression as the muon sneutrino and the left-handed smuon are very close in mass and the $f_{N,C}\left( x, y \right)$ functions share the same asymptotic limits. However this is not what we observe, and if we refer back to the contour plots of figure \ref{fig:contour} and the discussion carried out in section \ref{sec:g-2}, we realise that for the same values of $(x,y)$ we have in general that $f^{(E)}_{C}\left( x, y \right) \gg f^{(A,B,C,D)}_{N}\left( x, y \right)$. Therefore, in diagram $(E)$, it is the function $f^{(E)}_{{\rm C}}\left(\tfrac{M_{2}^{2}}{m_{\tilde{\nu}_{\mu}}^{2}},\tfrac{\mu^{2}}{m_{\tilde{\nu}_{\mu}}^{2}}\right)$ that is responsible for the enhancement of $\Delta a_\mu$, explaining our results. Benchmark points for the small $\mu$ scenario can be found in table \ref{tab:benchmark_tab_small}.
\begin{table}[H]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{cccccccccc}
\toprule
& Benchmark: & BP1 & BP2 & BP3 & BP4 & BP5 & \\
\midrule
\multirow{12}{*}{\rotatebox{90}{\textsc{Input at GUT scale}}} & $\tan \beta$ & 26.48 & 21.20 & 22.89 & 29.52 & 25.88 \\
& sgn$(\mu)$ & + & + & + & + & + & \\\cline{2-8}
& $m_0$ & 681.1 & 490.4 & 689.0 & 691.4 & 688.4 & \multirow{10}{*}{\rotatebox{-90}{[GeV]}} \\
& $m_1$ & 402.0 & 327.5 & 447.0 & 364.4 & 417.9 & \\
& $m_2$ & 397.4 & 273.0 & 394.2 & 342.2 & 390.7 & \\
& $m_3$ & 1204.7 & 871.8 & 1085.4 & 987.4 & 1192.3 & \\
& $M_1$ & -100.1 & -124.1 & -123.8 & -224.9 & -255.1 & \\
& $M_2$ & 294.9 & 367.5 & 449.9 & 168.6 & 177.9 & \\
& $M_3$ & 1004.6 & 1085.7 & 1109.8 & 1066.5 & 947.6 & \\
& $M_{h_1}$ &2204.8 & 2108.4 & 2246.6 & 2127.3 & 2007.2 & \\
& $M_{h_2}$ & 2385.7 & 2350.9 & 2455.7 & 2330.2 & 2344.7 & \\
& $A_{\text{tri}}$ & -2839.1 & -2762.5 & -2838.5 & -2764.0 & -3090.0 & \\
\hline
\multirow{24}{*}{\rotatebox{90}{\textsc{Masses}}} & $m_{h_0}$ & 125.2 & 125.2 & 125.2 & 125.1 & 125.1 & \multirow{26}{*}{\rotatebox{-90}{[GeV]}} \\
& $m_{\tilde{g}}$ & 2220.9 & 2373.5 & 2427.3 & 2349.4 & 2108.5 & \\
& $m_{\tilde{q}^{1,2}_L}$ & 2040.6 & 2122.7 & 2220.1 & 2149.1 & 1949.0 & \\
& $m_{\tilde{b}_1}$ & 1424.1 & 1537.5 & 1592.3 & 1506.8 & 1234.0 & \\
& $m_{\tilde{t}_1}$ & 1120.3 & 1117.4 & 1207.9 & 1184.6 & 962.3 & \\
& $m_{\tilde{q}^1_R}$ & 1963.9 & 2086.2 & 2149.9 & 2070.3 & 1872.7 & \\
& $m_{\tilde{q}^2_R}$ &1962.9 & 2078.1 & 2136.3 & 2066.3 & 1866.5 & \\
& $m_{\tilde{b}_2}$ & 2164.4 & 2108.7 & 2209.8 & 2026.6 & 1984.0 & \\
& $m_{\tilde{t}_2}$ & 1488.6 & 1584.3 & 1641.0 & 1561.4 & 1323.4 & \\
& $m_{\tilde{e}_L}$ & 710.5 & 555.8 & 752.4 & 705.3 & 715.6 & \\
& $m_{\tilde{e}_R}$ & 352.7 & 244.2 & 396.3 & 313.5 & 335.2 & \\
& $m_{\tilde{\mu}_L}$ & 710.1 & 555.2 & 751.8 & 704.5 & 714.9 & \\
& $m_{\tilde{\mu}_R}$ & 346.1 & 160.7 & 333.5 & 283.9 & 297.6 & \\
& $m_{\tilde{\tau}_1}$ & 594.8 & 375.0 & 589.5 & 424.9 & 483.8 & \\
& $m_{\tilde{\tau}_2}$ & 1054.1 & 612.5 & 834.6 & 560.1 & 894.9 & \\
& $m_{\tilde{\chi}^0_1}$ & -48.58 & -59.58 & -60.00 & -101.0 & -113.2 & \\
& $m_{\tilde{\chi}^0_2}$ & 169.5 & 215.5 & 243.3 & 115.9 & 127.9 & \\
& $m_{\tilde{\chi}^0_3}$ & -228.2 & -265.1 & -277.4 & -350.7 & -411.9 & \\
& $m_{\tilde{\chi}^0_4}$ & 287.7 & 337.3 & 391.5 & 357.2 & 416.9 & \\
& $m_{\tilde{\chi}^\pm_1}$ & 171.3 & 217.3 & 245.0 & 116.3 & 128.2 & \\
& $m_{\tilde{\chi}^\pm_2}$ & 287.4 & 336.9 & 390.8 & 360.4 & 419.9 & \\
& $m_{\tilde{\nu}^e_L}$ & 705.8 & 549.9 & 747.9 & 700.5 & 711.0 & \\
& $m_{\tilde{\nu}^\mu_L}$ & 705.5 & 549.4 & 747.5 & 704.5 & 710.4 & \\
& $m_{\tilde{\nu}^\tau_L}$ & 589.5 & 367.5 & 584.5 & 421.6 & 478.1 & \\\cline{1-7}
& $Q$ & 1293.4 & 1337.0 & 1409.0 & 1360.4 & 1143.6 & \\
& $\mu(Q)$ & 212.3 & 250.5 & 263.2 & 335.2 & 397.9 & \\
\hline
\multirow{5}{*}{\rotatebox{90}{\textsc{Constraints}}} & Br$(b \to s \gamma)$ & $2.89 \times 10^{-4}$ & $2.91 \times 10^{-4}$ & $2.91 \times 10^{-4}$ & $3.25 \times 10^{-4}$ & $ 3.25 \times 10^{-4}$ & \\
& Br$(B_s \to \mu^+ \mu^-)$ & $2.69 \times 10^{-9}$ & $2.97 \times 10^{-9}$ & $2.97 \times 10^{-9}$ & $ 3.06 \times 10^{-9}$ & $3.11\times 10^{-9}$ & \\
& $\sigma^{\text{DD SI}}$ & $1.31 \times 10^{-11}$ & $1.28 \times 10^{-11}$ & $1.18 \times 10^{-11}$ & $ 2.42 \times 10^{-11}$ & $1.06 \times 10^{-11}$ & [pb] \\
& $\Omega h^2$ & $1.05 \times 10^{-1}$ & $1.25 \times 10^{-1}$ & $1.23 \times 10^{-1}$ & $8.32 \times 10^{-2}$ & $ 8.47 \times 10^{-2}$ & \\
& $\Delta a_\mu$ & $1.37 \times 10^{-9}$ & $2.28 \times 10^{-9}$ & $1.30 \times 10^{-9}$ & $1.99 \times 10^{-9}$ & $1.52 \times 10^{-9}$ & \\
\bottomrule
\end{tabular}}
\caption{Input and Output parameters for the benchmark points with the most accurate $\Delta a_\mu$ and $\Omega h^2$ in the case of small $\mu(Q)$ and all other constraints being fulfilled. $\tilde{q}^{i}$ labels the $i$-th generation of squarks.}
\label{tab:benchmark_tab_small}
\end{table}
\subsection{\texorpdfstring{Large $\mu$}{Large mu}}
\label{subsec:large_mu}
The other class of solutions that provides the full contribution to the anomalous magnetic moment of the muon requires $\lvert \mu \rvert \gtrsim 2~{\rm TeV}$. In order to study this region in detail we perform an enhanced scan on the parameter space around the points in figure \ref{fig:gmuon_vs_mu_M1_M2} that better approach the value of $\Delta a_{\mu}$ as given in \eqref{eq:Damu_3}. The new scenarios were generated with the GUT scale parameters as in table \ref{tab:input-parameters}.
\begin{table}[h]
\centering
\begin{tabular}{|c|rcl|}
\hline
Parameter & \multicolumn{3}{c|}{Range} \\
\hline
$A_{\rm{tri}}$ & $-3000$ &--& $0$ \\
$m_0$ & $100$ &--& $300$ \\
$m_1$ & $500$ &--& $1500$ \\
$m_2$ & $100$ &--& $400$ \\
$m_3$ & $1000$ &--& $2000$ \\
$m_{H_1}$, $m_{H_2}$& $100$ &--& $3000$ \\
\hline
\end{tabular}
\begin{tabular}{|c|rcl|}
\hline
Parameter & \multicolumn{3}{c|}{Range} \\\hline
$M_1$ & $-1000$ &--& $1000$\\
$M_2$ & $-2000$ &--& $2000$\\
$M_3$ & $2000$ &--& $3000$\\
$\tan \beta$ & $5$ &--& $50$ \\
$\operatorname{sgn}\( \mu \)$& & $\pm 1$ & \\
&&&\\
\hline
\end{tabular}
\caption{Theory parameters at the GUT scale. The soft-SUSY breaking parameters are given in GeV.}
\label{tab:input-parameters}
\end{table}
Analogue to figure \ref{fig:f_vs_gmuon}, we first investigate which loop diagrams from equation \ref{eq:loops} contribute most to $\Delta a_\mu$. This is shown in figure \ref{fig:diag_vs_mu}.
\begin{figure}[H]
\centering
\includegraphics[scale=0.43]{figures/diag_vs_mu-a.pdf}
\includegraphics[scale=0.43]{figures/diag_vs_mu-b.pdf}\\
\includegraphics[scale=0.43]{figures/diag_vs_mu-c.pdf}
\includegraphics[scale=0.43]{figures/diag_vs_mu-d.pdf}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.43]{figures/diag_vs_mu-e.pdf}
\caption{Individual contributions for the $\Delta a^{(i)}_{\mu}$ terms, with $i=\left\{A,B,C,D,E\right\}$ in equations \eqref{eq:loops}, vs. $\mu(Q)$. The color scale indicates the total value value of $\Delta a_{\mu}$, while the black bars in the top left panel show the 1$\sigma$ bound of $\Delta a_\mu$. The orange pentagons represent the benchmark points defined in table \ref{tab:benchmark_tab1}.}
\label{fig:diag_vs_mu}
\end{figure}
It is clear that, in this case, diagram $(A)$ yields the main contribution to $\Delta a_\mu$. This is mainly due to the prefactor $\(\frac{M_1 \mu}{m^2_{\tilde{\mu}_L} m^2_{\tilde{\mu}_R}}\)$ from equation \ref{eq:loops:A} being large for large $\mu$ and small smuon masses. Additionally, eqauations \ref{eq:loops:B} \-- \ref{eq:loops:E} all feature $\mu$ in the denominator, thus leading to highly suppressed contributions from these diagrams.
In this scenario, dark matter is entirely bino-dominated for points with $\Delta a_\mu$ in the 1$\sigma$ bound (dark blue diamonds), leading to a viable relic density. This is visualised in figure \ref{fig:M1_vs_M2}.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{figures/M1_vs_M2-LR.pdf}
\caption{$M_1(Q)$ vs. $M_2(Q)$ (left) and a smaller excerpt of it (right). All dark blue diamonds are bino-like, whereas the light green triangles and turquoise circles are wino-like. The orange pentagons represent the benchmark points defined in table \ref{tab:benchmark_tab1}.}
\label{fig:M1_vs_M2}
\end{figure}
In figure \ref{fig:mass_gaps}, we show the mass gaps between smuons and the LSP, which always is the lightest neutralino in this scenario. In case of left handed smuons, the mass gap for points featuring good $\Delta a_\mu$ and relic density (dark blue diamonds) is in a range of roughly 50 - 700 GeV, which is important for any collider phenomenology (cf. section \ref{sec:constraints}). As an example, muons emitted in the decay $\tilde{\mu}_L \to \tilde{\chi}^0_1 \, \mu_L$ would be very energetic, but most likely soft in the case of right handed smuons, as there are plenty of points with a mass gap below 50 GeV.
\begin{figure}[h]
\centering
\includegraphics[scale=.45]{figures/mass_gaps-L.pdf}
\includegraphics[scale=.45]{figures/mass_gaps-R.pdf}
\caption{Mass gaps between the smuon and lightest neutralino masses $m_{\tilde{\mu}_{L/R}}$ and $m_{\tilde{\chi}^0_1}$. All dark blue diamonds are bino-like, whereas the light green triangles and turquoise circles are wino-like. The orange pentagons represent the benchmark points defined in table \ref{tab:benchmark_tab1}.}
\label{fig:mass_gaps}
\end{figure}
Figure \ref{fig:all_NLSP} shows the mass gaps between the LSP and NLSP vs. the LSP mass for too low relic density (top left plot), relic density in bounds (top right) and relic density as well as $\Delta a_\mu$ in bounds (bottom). In case of too low relic density, the first chargino is the NLSP for the majority of points and is degenerated in mass with the LSP. If the relic density increases, there are almost no chargino-NLSP's left and the NLSP changes to the right-handed smuon, but the first stauon and the $\tau$-sneutrino also yield significant amounts of NLSP's for this scenario. Also, all three of them are mass degenerated up to roughly 10 GeV with the LSP. In case of both relic density and $\Delta a_\mu$ being in the 1$\sigma$ bound, this picture does not change, but the favoured LSP mass is narrowed down from 100 \-- 400 GeV to 200 \-- 300 GeV for right-handed smuons. In case of $\tilde{\tau}_1$ or $\tilde{\nu}_\tau$, the LSP mass range is only slightly reduced.
\begin{figure}[H]
\centering
\includegraphics[scale=.45]{figures/mLSP_vs_mNLSP-RD-NOTOK.pdf}
\includegraphics[scale=.45]{figures/mLSP_vs_mNLSP-RD-OK.pdf}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=.45]{figures/mLSP_vs_mNLSP-g-2-OK.pdf}
\caption{Mass differences between the LSP and NLSP compared to the LSP mass. The top left plot has too small relic density, whereas the top right plot has the relic density in bounds and the lower central plot additionally has $\Delta a_\mu$ in bounds. For this plot, the LSP always is the lightest neutralino $\tilde{\chi}^0_1$.}
\label{fig:all_NLSP}
\end{figure}
In figure \ref{fig:gmuon_vs_RD}, we show the $\Delta a_\mu \-- \Omega h^2$ plane and the respective 1$\sigma$ bounds as a grey shaded area. There are plenty of points lying close to the 1$\sigma$ bound w.r.t. $\Omega h^2$ and still many points in both 1$\sigma$ bounds. Based on the best points in the 1$\sigma$ bound (lower plot), we set up benchmark points (shown as orange pentagons in figures \ref{fig:diag_vs_mu} \-- \ref{fig:gmuon_vs_RD}) for the upcoming analysis for vacuum stability.
All benchmark points and their respective input parameters as well as a selection of the output parameters are shown below in table \ref{tab:benchmark_tab1}.
\begin{figure}[H]
\centering
\includegraphics[scale=.45]{figures/gmuon_vs_RD_full.pdf}
\includegraphics[scale=.45]{figures/gmuon_vs_RD_zoom.pdf}
\caption{$\Delta a_\mu$ vs. $\Omega h^2$. The top left plot shows the full parameter spectrum, the top right plot a smaller excerpt of it with the grey shaded area being the 1$\sigma$ bound of $\Delta a_\mu$ and $\Omega h^2$. All dark blue diamonds are bino-like, whereas the light green triangles and turquoise circles are wino-like. The orange pentagons represent the benchmark points defined in table \ref{tab:benchmark_tab1}.}
\label{fig:gmuon_vs_RD}
\end{figure}
\begin{table}[H]
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{cccccccccc}
\toprule
& Benchmark: & BP6 & BP7 & BP8 & BP9 & BP10 & \\
\midrule
\multirow{12}{*}{\rotatebox{90}{\textsc{Input at GUT scale}}} & $\tan \beta$ & 16.96 & 26.88 & 32.15 & 22.21 & 40.22 \\
& sgn$(\mu)$ & + & + & + & + & + & \\\cline{2-8}
& $m_0$ & 238.8 & 149.6 & 106.5 & 271.5 & 137.5 & \multirow{10}{*}{\rotatebox{-90}{[GeV]}} \\
& $m_1$ & 1426.7 & 1131.1 & 626.5 & 508.9 & 1470.7 & \\
& $m_2$ & 239.2 & 302.7 & 125.3 & 193.5 & 178.4 & \\
& $m_3$ & 1458.7 & 1631.9 & 1076.3 & 1434.2 & 1847.8 & \\
& $M_1$ & 577.9 & 292.3 & 711.6 & 579.8 & 760.7 & \\
& $M_2$ & 412.8 & 612.4 & 948.8 & -436.4 & 982.8 & \\
& $M_3$ & 2195.7 & 2055.2 & 2680.5 & 2456.0 & 2524.6 & \\
& $M_{h_1}$ & 670.6 & 2924.4 & 577.0 & 1512.8 & 1577.3 & \\
& $M_{h_2}$ & 814.9 & 925.9 & 918.8 & 1306.2 & 1362.7 & \\
& $A_{\text{tri}}$ & -2244.8 & -2776.6 & -1113.2 & -2896.2 & -2370.1 & \\
\hline
\multirow{24}{*}{\rotatebox{90}{\textsc{Masses}}} & $m_{h_0}$ & 124.1 & 124.1 & 123.5 & 124.5 & 123.6 & \multirow{26}{*}{\rotatebox{-90}{[GeV]}} \\
& $m_{\tilde{g}}$ & 4595.1 & 4308.9 & 5497.4 & 5089.9 & 5201.7 & \\
& $m_{\tilde{q}^{1,2}_L}$ & 3931.0 & 3697.9 & 4709.6 & 4356.0 & 4468.8 & \\
& $m_{\tilde{b}_1}$ & 3527.8 & 3216.6 & 4257.5 & 3893.5 & 3878.2 & \\
& $m_{\tilde{t}_1}$ & 3412.9 & 3154.1 & 4068.0 & 3743.2 & 3842.3 & \\
& $m_{\tilde{q}^1_R}$ & 4183.9 & 3859.4 & 4731.4 & 4378.0 & 4683.5 & \\
& $m_{\tilde{q}^2_R}$ & 3936.0 & 3699.4 & 4690.4 & 4352.2 & 4445.5 & \\
& $m_{\tilde{b}_2}$ & 4137.3 & 3891.1 & 4637.2 & 4478.3 & 4510.4 & \\
& $m_{\tilde{t}_2}$ & 3586.5 & 3334.7 & 4286.0 & 3936.8 & 4038.8 & \\
& $m_{\tilde{e}_L}$ & 328.2 & 393.0 & 588.3 & 375.9 & 627.6 & \\
& $m_{\tilde{e}_R}$ & 1442.2 & 1136.2 & 684.3 & 552.4 & 1497.9 & \\
& $m_{\tilde{\mu}_L}$ & 328.2 & 393.0 & 588.1 & 375.9 & 627.7 & \\
& $m_{\tilde{\mu}_R}$ & 315.0 & 318.7 & 298.4 & 289.1 & 328.5 & \\
& $m_{\tilde{\tau}_1}$ & 248.1 & 120.0 & 485.0 & 244.9 & 328.2 & \\
& $m_{\tilde{\tau}_2}$ & 1445.0 & 1553.8 & 1052.4 & 1399.6 & 1720.5 & \\
& $m_{\tilde{\chi}^0_1}$ & 235.5 & 113.0 & 294.8 & 237.6 & 319.7 & \\
& $m_{\tilde{\chi}^0_2}$ & 310.4 & 483.2 & 758.4 & -426.1 & 792.1 & \\
& $m_{\tilde{\chi}^0_3}$ & -2942.2 & -2921.4 & -3116.3 & 3226.7 & -3273.1 & \\
& $m_{\tilde{\chi}^0_4}$ & 2942.6 & 2921.6 & 3116.9 & -3226.9 & 3273.5 & \\
& $m_{\tilde{\chi}^\pm_1}$ & 310.6 & 483.4 & 758.5 & 426.3 & 792.2 & \\
& $m_{\tilde{\chi}^\pm_2}$ & 2943.5 & 2922.6 & 3117.6 & 3227.8 & 3274.3 & \\
& $m_{\tilde{\nu}^e_L}$ & 318.5 & 384.8 & 582.7 & 367.4 & 622.4 & \\
& $m_{\tilde{\nu}^\mu_L}$ & 318.5 & 384.8 & 582.7 & 367.4 & 622.5 & \\
& $m_{\tilde{\nu}^\tau_L}$ & 243.3 & 129.8 & 517.2 & 247.0 & 350.5 & \\\cline{1-7}
& $Q$ & 3409.7 & 3163.2 & 4072.1 & 3742.4 & 3845.1 & \\
& $\mu(Q)$ & 2932.7 & 2917.6 & 3105.9 & 3217.7 & 3271.1 & \\
\hline
\multirow{5}{*}{\rotatebox{90}{\textsc{Constraints}}} & Br$(b \to s \gamma)$ & $3.32 \times 10^{-4}$ & $3.29 \times 10^{-4}$ & $3.30 \times 10^{-4}$ & $3.32 \times 10^{-4}$ & $3.28 \times 10^{-4}$ & \\
& Br$(B_s \to \mu^+ \mu^-)$ & $3.07 \times 10^{-9}$ & $3.13 \times 10^{-9}$ & $3.14 \times 10^{-9}$ & $3.08 \times 10^{-9}$ & $3.32 \times 10^{-9}$ & \\
& $\sigma^{\text{DD SI}}$ & $9.69 \times 10^{-13}$ & $4.44 \times 10^{-13}$ & $6.65 \times 10^{-13}$ & $5.50 \times 10^{-13}$ & $6.31 \times 10^{-13}$ & [pb] \\
& $\Omega h^2$ & $1.20 \times 10^{-1}$ & $1.22 \times 10^{-1}$ & $1.20 \times 10^{-1}$ & $1.20 \times 10^{-1}$ & $1.19 \times 10^{-1}$ & \\
& $\Delta a_\mu$ & $2.71 \times 10^{-9}$ & $3.06 \times 10^{-9}$ & $2.23 \times 10^{-9}$ & $2.98 \times 10^{-9}$ & $2.36 \times 10^{-9}$ & \\
\bottomrule
\end{tabular}}
\caption{Input and Output parameters for the benchmark points with the most accurate $\Delta a_\mu$ and $\Omega h^2$ in the case of large $\mu(Q)$ and all other constraints being fulfilled. $\tilde{q}^{i}$ labels the $i$-th generation of squarks.}
\label{tab:benchmark_tab1}
\end{table}
\section{Vacuum Stability}
\label{sec:vacuum-stability}
\texttt{SoftSUSY} implements two-loop tadpole contributions to the minimization conditions to ensure the breaking of electroweak symmetry by Higgs VEVs. As with other spectrum generators, the minimization conditions are used to fix parameters of the theory in such a way that the desired vacuum is a minimum of the scalar potential. One downside of this procedure is that other solutions to the minimization conditions might exist and lie lower in the scalar potential of the theory. At the same time, color- and charge- breaking (CCB) VEVs are usually ignored and such minima might also exist and lie lower than the desired vacuum.
It is then interesting to understand if the points in our scans suffer from CCB minima, whether they are lower than the desired vacuum and in that case if the desired vacuum is sufficiently long-lived (meta-stable). Although approximate analytical conditions for the avoidance of CCB minima exist for the MSSM, a full numerical study of the one-loop effective potential is often needed as the conditions are neither sufficient nor necessary to ensure the absence of such minima \cite{Camargo-Molina:2013sta}. In addition, such analytical rules are based on a tree-level analysis and are thus irrelevant for points where the symmetry breaking occurs only at one-loop.
Using \texttt{Vevacious} \cite{Camargo-Molina:2013qva} we performed a numerical analysis of the tree and one-loop effective potential for a set of benchmark parameter points allowing for stop and stau VEVs. Due to the fact that the desired vacuum comes as a solution of two-loop minimization conditions we found that quite often the EWSB minimum only appears after two-loop contributions to the effective potential are considered. For such parameter points an analysis with \texttt{Vevacious}, which uses the one-loop effective potential, was not possible and thus the vacuum stability analysis was inconclusive. However, it was still possible to find parameter points where the EWSB minimum (the desired vacuum) develops at tree-level or one-loop. In the case of minima appearing only at one-loop, a careful numerical minimization of the one-loop effective potential was required, as \texttt{Vevacious} uses the tree-level minima as starting points for numerical minimization therefore missing such cases out of the box. It was possible however to study the vacuum stability in a point by point basis by starting the numerical minimization around the field values for the EWSB minimum that develops once two-loop contributions are considered.
For the points considered in section \ref{subsec:small_mu} and shown in table \ref{tab:benchmark_tab_small}, the desired vacuum was the global minimum of the one-loop effective potential. For the points considered in section \ref{subsec:large_mu}, we started with a set of benchmark points satisfying all the constraints considered in the previous sections, and after performing the vacuum stability analysis we selected those where the desired vacuum was either the global minimum (and thus stable) or long-lived after considering tunneling to deeper CCB minima at zero and non-zero temperature. The result of the analysis is shown in figure \ref{fig:VS_mu_vs_tanb}. In this figure we can see that points with larger $\mu$ roughly correspond to those for which the desired vacuum develops once two-loop corrections are considered, as could be naively expected. In addition the stable and long-lived points tend to have larger $|A_t|$ and $A_0$ together with lower $\tan \beta$ values. This comes from the fact that the larger $\tan \beta$ is the smaller $m_{\tilde{\tau}_R}^2 $, increasing the chance for $\tilde{\tau}$ VEVs. Conversely, lower $\tan \beta$ thus allows for higher values of $A_0$ and $|A_t|$ a combination that allows the points to fulfill all other constrains together with vacuum stability. The points for which the desired vacuum was either global or long-lived minimum (red and orange in fig. \ref{fig:VS_mu_vs_tanb}) correspond to the benchmark points shown in light orange in figures \ref{fig:M1_vs_M2}, \ref{fig:mass_gaps} and \ref{fig:gmuon_vs_RD}.
\begin{figure}[H]
\centering
\includegraphics[width=0.49\linewidth]{figures/VS_mu_vs_At-crop.pdf}
\includegraphics[width=0.463\linewidth]{figures/VS_A0_vs_tanbeta-crop.pdf}
\caption{Vacuum stability analysis for set of points fulfilling all other constraints in the large-$\mu$ region: Orange and red points correspond to the final benchmark points for which the desired vacua are stable or long-lived respectively. For light blue points (labeled ``Minimum at two-loop'' ) the desired vacuum appears only when two-loop corrections are included and thus could not be studied with current tools. Black points showed deeper CCB minima with $< 1 \%$ survival probability of the desired vacuum. Gray points showed CCB minima at 1-loop but the desired vacua appear after two-loop corrections are included. }
\label{fig:VS_mu_vs_tanb}
\end{figure}
\section{Conclusions} \label{sec:conclusion}
The anomalous magnetic moment of the muon continues to show a disagreement with the SM which suggests new physics at a relatively low mass scale. The leading candidate for such new physics is the MSSM with light sleptons and light charginos and neutralinos, which can contribute substantially to $\Delta a_\mu$ at one-loop and explain the experimental $\Delta a_\mu$ measurements. Such a SUSY spectrum as low as a few hundreds GeV requires to explain $\Delta a_\mu$ contrasts with the failure of the LHC to discover coloured superpartners such as squarks and gluinos, leading to stringent bounds on such sparticles, requiring their masses to typically lie above the TeV scale. The Higgs boson mass also requires at least some stop masses above the TeV scale in the MSSM.
From the experimental side, these constraints are not inconsistent with having light sleptons and gauginos down to about 100 GeV, since the LHC sensitivity to colour singlets is significantly lower than to coloured particles. At the same time, from the theory side it is very hard to accommodate light sleptons and heavy squarks for all generations at the weak scale in the cMSSM or mSUGRA model with universal sfermion masses at the GUT scale. This is especially difficult if one takes into account combined collider and non-collider constraints including those from the dark matter relic density.
Such a tension strongly favours MSSM models with non-universal sfermion masses at the GUT scale like the pMSSM which relax the constraints of the CMSSM without introducing excessive flavour changing neutral currents and without unleashing all the 100 or so parameters of the MSSM. However, the pMSSM still contains 19 SUSY parameters and is not particularly well theoretically motivated. In this paper, we have considered a theoretically very well motivated scenario, which involves just four soft scalar masses at the GUT scale, namely $m_0$ (a universal left-handed scalar mass) and $m_1$, $m_2$, $m_3$ (three universal right-handed scalar masses, one for each family), together with non-universal gaugino and trilinear soft masses. In this model, the first and second family sleptons can be light to explain $\Delta a_\mu$ while simultaneously, $m_3$ can be large enough to provide enough mass for the Higgs boson and the agreement with other observables such as $\text{Br}(b \to s \gamma)$ and $\text{Br}(B_s \to \mu^+ \mu^-)$ stays valid.
The comprehensive scan over the soft parameter space of the model, exploiting the relatively small number of soft input masses (as compared for example to the pMSSM), has confirmed the existence of viable points which satisfy both $\Delta a_\mu$ and dark matter constaints neatly dividing into two sets: small $\mu$ and large $\mu$, which we subsequently investigated in detail separately. For these two parameter regions, we were able to understand the dominant effects leading to successful $\Delta a_\mu$ as well as the characteristics of the dark matter candidate, while satisfying all other experimental constraints. For example we investigated the NLSP to understand which SUSY particle is responsible for the effective co-annihilation as well as the LSP-NLSP mass splitting, which is very important experimentally. We also proposed sets of benchmark points for each scenario and checked the vacuum stability for all benchmark points, especially for the large $\mu$ case where vacuum stability is an issue.
The small $\mu \lesssim 400$ GeV region involves a bino-like neutralino LSP which annihilates in the early Universe either resonantly, if its
mass is around half the mass of the $Z$ or Higgs boson, or via co-annihilation with the higgsino states if the $\mu$ parameter is about 15 GeV higher than the LSP mass. The benchmarks are chosen such that there is a large mass gap of around 100 GeV between the LSP and the smuon mass, so that the smuon decay will involve a hard muon, providing a clear signal at the LHC. For all these small $\mu$ cases, $\Delta a_\mu$ is dominated by diagrams (B) and (E) of figure~\ref{fig:1-loop-amu}. The large $\mu \sim 3$ TeV region also involves a bino-like neutralino LSP which co-annihilates in the early Universe with an NLSP which may be $\tilde{\tau}_1$, $\tilde{\nu}_{\tau}$, $\tilde{\chi}_2$ or $\tilde{\mu}$, depending on the precise parameters. In all these cases, the dominant contribution to $\Delta a_\mu$ comes from diagram (A) of figure~\ref{fig:1-loop-amu}. In both scenarios, heavy gluinos (above 2 TeV) help to split the squark and slepton masses of the first two generations, yielding heavy squark masses satisfying the LHC bounds on the first and second family squarks, while allowing light sleptons. These scenarios both predict light smuons (100-300 GeV), which can be probed via leptonic signatures and even potentially explain di-lepton excesses reported by the ATLAS and CMS collaborations. In addition, the small $\mu$ scenario also predicts quite light charginos and second neutralinos exhibiting
di-lepton or tri-lepton signatures which can be tested in the near future and/or explain the di-lepton excesses mentioned above.
In conclusion, the MSSM with a Pati-Salam gauge group broken at the GUT scale and flavour symmetries $A_4$ and $Z_5$, which unify the soft masses of the left-handed (but not right-handed) squarks and sleptons, provides a well motivated framework with a relatively low number of input soft masses, which is capable of accounting for $\Delta a_\mu$ as well as providing good dark matter candidates, consistently with all other experimental and theoretical constraints. We emphasise that (unlike some other models) the A to Z Pati-Salam model initially was not designed to explain $\Delta a_\mu$, since its primary motivation was to explain the flavour mass and mixing of quarks and leptons, in particular neutrinos. Nevertheless, we have seen that the model is well suited to account for $\Delta a_\mu$, while simultaneously providing a good dark matter candidate, namely the lightest neutralino, which is consistently bino-like in nature. The characteristic SUSY spectra presented here should enable this model to be distinguished from other less well motivated models such as the pMSSM.
\section{Some title}
\acknowledgments
The authors acknowledge the use of the IRIDIS High Performance Computing Facility, and
associated support services at the University of Southampton, in the completion of this work.
ASB, SFK and PBS acknowledge partial support from the InvisiblesPlus RISE from the European
Union Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant
agreement No 690575. SFK acknowledges partial support from the Elusives ITN from the European
Union Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant
agreement No 674896.
AB and SFK acknowledges partial support from the STFC grant ST/L000296/1.
DJM acknowledges patial support from the STFC grant ST/L000446/1.
AB also thanks the NExT Institute, Royal Society Leverhulme Trust Senior Research Fellowship LT140094 and
Soton-FAPESP grant for patial support.
AM is supported by the FCT grant SFRH/BPD/97126/2013 and partially by the H2020-MSCA-RISE-2015 Grant agreement No StronGrHEP-690904, and by the CIDMA project UID/MAT/04106/2013. AM also acknowledges the THEP group at Lund University for all hospitality and support provided for the development of this work. JECM wants to thank Ben O'Leary for discussions in the early stages of this work.
\newpage
\bibliographystyle{JHEP}
|
\section{Implementation of the FONLL scheme with
$\overline{\mbox{MS}}$ masses}\label{app:implementation}
In this appendix we will describe in some detail the implementation of
the FONLL scheme in terms of the $\overline{\mbox{MS}}$ heavy-quark
running masses in {\tt APFEL}. Starting from the more used definition
of structure functions and in terms of pole masses, our goal is to
consistently replace them with the $\overline{\mbox{MS}}$ mass
definition.
\subsection{$\overline{\mbox{MS}}$ mass vs. pole mass}
The (scale independent) pole mass $M$ and the (scale dependent)
$\overline{\mbox{MS}}$ mass $m(\mu)$ arise from two different
renormalization procedures and in perturbation theory they can be
expressed one in terms of the other. The relation connecting pole and
$\overline{\mbox{MS}}$ mass definitions has been computed in
Ref.~\cite{Chetyrkin:1999qi} up to four loops. However, in the
following we will only need to go up to one loop and thus we report
here the corresponding relation:
\begin{equation}\label{eq:PoleToMSbar}
\frac{M}{m(\mu)} = 1 + h^{(1)}a_s+\mathcal{O}(a_s^2)\,,
\end{equation}
with:
\begin{equation}\label{eq:CoeffPoleToMSbar}
\begin{array}{l}
\displaystyle h^{(1)}(\mu,m(\mu)) = C_F\left(4 + 3L_{\mu m}\right)\,
\end{array}
\end{equation}
where $C_F=4/3$ is one of the usual QCD color factors. Moreover, we
have defined:
\begin{equation}
a_s\equiv a_s(\mu) = \frac{\alpha_s(\mu)}{4\pi}\,.
\end{equation}
and:
\begin{equation}
L_{\mu m} = \ln\frac{\mu^2}{m^2(\mu)}\,.
\end{equation}
In the following we will use eq. (\ref{eq:PoleToMSbar}) to replace the
pole mass $M$ with the $\overline{\mbox{MS}}$ mass $m(\mu)$.
\subsection{RGE solution for the $\overline{\mbox{MS}}$ running mass}
In order to evaluate the running of $m(\mu)$ with the renormalization
scale $\mu$ we have to solve the corresponding renormalization-group
equation (RGE):
\begin{equation}\label{eq:massRGE}
\mu^2\frac{dm}{d\mu^2} = m(\mu)\gamma_m(a_s) = -m(\mu)\sum_{n=0}^{\infty}\gamma_m^{(n)}a_s^{n+1}\,,
\end{equation}
whose first three coefficients can be taken from
Ref.~\cite{Chetyrkin:1999pq}\footnote{The following expressions have
been adjusted taking into account our definition of $a_s$ which
differs by a factor of 4 with respect to that of Ref.~\cite{Chetyrkin:1999pq}.}:
\begin{subequations}
\begin{equation}
\gamma_m^{(0)} = 4 \,,
\end{equation}
\begin{equation}
\gamma_m^{(1)} = \frac{202}3 - \frac{20}{9}N_f\,,
\end{equation}
\begin{equation}
\gamma_m^{(2)} = 1249 - \left(\frac{2216}{27}+\frac{160}{3}\zeta_3\right)N_f-\frac{140}{81}N_f^2\,,
\end{equation}
\end{subequations}
where $N_f$ is the number of active flavours. In addition the RGE for
the running of $\alpha_s$ reads:
\begin{equation}\label{eq:alphasRGE}
\mu^2\frac{da_s}{d\mu^2} = \beta(a_s) = -\sum_{n=0}^{\infty}\beta_n a_s^{n+2}\,,
\end{equation}
with:
\begin{subequations}
\begin{equation}
\beta_0 = 11-\frac23 N_f \,,
\end{equation}
\begin{equation}
\beta_1 = 102 - \frac{38}3 N_f\,.
\end{equation}
\begin{equation}
\beta_2 = \frac{2857}{2} - \frac{5033}{18}N_f + \frac{325}{54}N_f^2\,.
\end{equation}
\end{subequations}
Combining eqs.~(\ref{eq:massRGE}) and~(\ref{eq:massRGE}) we obtain the
following differential equation:
\begin{equation}\label{runmass}
\frac{dm}{da_s} = \frac{\gamma_m(a_s)}{\beta(a_s)}m(a_s)\,,
\end{equation}
whose solution is:
\begin{equation}\label{eq:massRGEsolution}
m(\mu) = m(\mu_0)\exp\left[\int_{a_s(\mu_0)}^{a_s(\mu)}\frac{\gamma_m(a_s)}{\beta(a_s)}da_s\right]\,.
\end{equation}
In order to get an analytical expression out of
eq.~(\ref{eq:massRGEsolution}), we expand the integrand in the
l.h.s. using the perturbative expansions of $\gamma_m(a_s)$ and
$\beta(a_s)$ given in eqs.~(\ref{eq:massRGE})
and~(\ref{eq:alphasRGE}). This allows us to solve the integral
analytically, obtaining:
\begin{equation}\label{eq:massRGEsolutionAnal}
\begin{array}{rcl}
m(\mu)&=&\displaystyle m(\mu_0)\left(\frac{a}{a_0}\right)^{c_0}\\
\\
&\times&\displaystyle
\frac{1+(c_1-b_1c_0)a+\frac12[c_2-c_1b_1-b_2c_0+b_1^2c_0+(c_1-b_1c_0)^2]a^2}{1+(c_1-b_1c_0)a_0+\frac12[c_2-c_1b_1-b_2c_0+b_1^2c_0+(c_1-b_1c_0)^2]a_0^2}\,,
\end{array}
\end{equation}
where we have defined:
\begin{equation}\label{eq:SolCoefs}
b_i = \frac{\beta_i}{\beta_0}\quad\mbox{and}\quad c_i = \frac{\gamma_m^{(i)}}{\beta_0}\,,
\end{equation}
and $a\equiv a_s(\mu)$ and $a_0\equiv a_s(\mu_0)$.
Eq.~(\ref{eq:massRGEsolutionAnal}) represents the
next-to-next-lo-leading order (NNLO) solution of the RGE for the
$\overline{\mbox{MS}}$ mass $m(\mu)$.
Of course, the NLO and the LO solutions can be easily extracted from
eq.~(\ref{eq:massRGEsolutionAnal}) just by disregarding the terms
proportional to $a^2$ and $a_0^2$ for the NLO solution and also the
terms proportional to $a$ and $a_0$ for the LO solution\footnote{In
order to be consistent, the evaluation of $a$ and $a_0$
eq.~(\ref{eq:massRGEsolutionAnal}) must be performed at the same
perturbative order of $m(\mu)$. So, for instance, if one wants to
evaluate the NNLO running of $m(\mu)$ also the value of $a$ and
$a_0$ must be computed using the NNLO running.}.
\subsection{Matching conditions}\label{sec:MatchingConditions}
When working in the context of a variable-flavour-number scheme
(VFNS), all running quantities are often required to cross the
heavy-quark thresholds to evolve from one scale to another. Such a
transition in turn requires to match in correspondence of a given
heavy-quark threshold different factorization schemes whose content of
light quarks differs by one unity. In other words, if the perturbative
evolution leads from an energy region where (by definition) there are
$N_f-1$ light flavours to another region where there are $N_f$ light
flavours, the two regions must be consistently connected and such a
connection can be perturbatively evaluated and goes under the name of
\textit{matching conditions}.
In general, the matching conditions give rise to discontinuities of
the running quantities in correspondence of the matching scale and in
the following we will report the matching conditions up to NNLO in
terms of the $\overline{\mbox{MS}}$ heavy-quark thresholds for:
$\alpha_s(\mu)$, $m(\mu)$ and PDFs.
\subsubsection*{Matching of $\alpha_s(\mu)$}
The matching condition for $\alpha_s$ were evaluated in
Ref.~\cite{Chetyrkin:1997sg} to three loops. We report here the
relation up to two loops (again taking into account the factor 4
coming from the different definitions of $a$):
\begin{equation}\label{eq:alphaspole}
\frac{a^{(N_f-1)}(\mu)}{a^{(N_f)}(\mu)}=1-\frac23 L_{\mu M}a^{(N_f)}(\mu)+\left(\frac49L_{\mu M}^2-\frac{38}3L_{\mu M}-\frac{14}3\right)[a^{(N_f)}(\mu)]^2\,.
\end{equation}
$M$ being the pole mass of the $n$-th flavour. From
eq.~(\ref{eq:PoleToMSbar}) we can easily infer that:
\begin{equation}
\ln M^2 = \ln m^2(\mu) + 2\ln[1+h^{(1)}(\mu)a^{(N_f)}(\mu)]= \ln m^2(\mu) + 2h^{(1)}(\mu)a^{(N_f)}(\mu)+\mathcal{O}([a^{(N_f)}]^2)\,.
\end{equation}
Therefore it is straightforward to see that:
\begin{equation}\label{eq:conversionLog}
L_{\mu M} = L_{\mu m} - 2h^{(1)}a^{(N_f)}=L_{\mu m}-\left(\frac{32}3+8L_{\mu m}\right)a^{(N_f)}\,,
\end{equation}
so that:
\begin{equation}\label{eq:alphasmsbar}
\frac{a^{(N_f-1)}(\mu)}{a^{(N_f)}(\mu)}=1-\frac23 L_{\mu m}a^{(N_f)}(\mu)+\left(\frac49L_{\mu m}^2-\frac{22}3L_{\mu m}+\frac{22}9\right)[a^{(N_f)}(\mu)]^2\,.
\end{equation}
In order to simplify the expression, it is a common procedure to
perform the matching in the point where the logarithms vanish. In this
particular case, choosing $\mu=m(\mu)=m(m)$, we get:
\begin{equation}
a^{(N_f-1)}(m)=a^{(N_f)}(m)\left(1+\frac{22}9[a^{(N_f)}(m)]^2\right)\,,
\end{equation}
which can be easily inverted obtaining:
\begin{equation}\label{eq:alphasatthrs}
a^{(N_f)}(m)=a^{(N_f-1)}(m)\left(1-\frac{22}9[a^{(N_f-1)}(m)]^2\right)\,.
\end{equation}
It is interesting to observe that, in order to perform the matching as
described above, one just needs to know the value of $m(m)$. This is
the so-called RG-invariant $\overline{\mbox{MS}}$ mass. This finally
allows us to evaluate the values of $a$ and $a_0$ that enter in
eq. (\ref{eq:massRGEsolutionAnal}) needed to compute the running of
$m(\mu)$.
\subsubsection*{Matching of $m(\mu)$}
Also the running of $\overline{\mbox{MS}}$ needs to be matched at the
heavy-quark thresholds. In particular, one needs to match the
$(N_f-1)$ with $(N_f)$ scheme for the mass $m_q(\mu)$, with $q=c,b,t$,
at the threshold $m_h(\mu)$, with again $h=c,b,t$ the $n$-th
threshold. From Ref.~\cite{Chetyrkin:2000yt} we read:
\begin{equation}\label{eq:mqmc}
\frac{m_q^{(N_f-1)}(\mu)}{m_q^{(N_f)}(\mu)}=1+\left(\frac43L_{\mu m}^{(h)2}-\frac{20}9L_{\mu m}^{(h)}+\frac{89}{27}\right)[a^{(N_f)}(\mu)]^2\,,
\end{equation}
where:
\begin{equation}
L_{\mu m}^{(h)} =\ln\frac{\mu^2}{m_h^2(\mu)}\,.
\end{equation}
Exactly as before, if we choose to match the two schemes at the scale $\mu=m_h(\mu)=m_h(m_h)$, the logarithmic terms vanish and we are left with:
\begin{equation}
m_q^{(N_f-1)}(m_h)=\left(1+\frac{89}{27}[a^{(N_f)}(m_h)]^2\right)m_q^{(N_f)}(m_h)\,,
\end{equation}
whose inverse is:
\begin{equation}
m_q^{(N_f)}(m_h)=\left(1-\frac{89}{27}[a^{(N_f-1)}(m_h)]^2\right)m_q^{(N_f-1)}(m_h)\,.
\end{equation}
\subsubsection*{Matching of PDFs}
To conclude the section on matching conditions, we finally consider
PDFs. One can write the singlet and the gluon in the $N_f$ scheme in
terms of singlet and gluon in the $N_f-1$ scheme at any scale $\mu$ in
the pole-mass scheme as follows:
\begin{equation}\label{eq:MatchPDFsPole}
\begin{array}{c}
\displaystyle {\Sigma^{(N_f)} \choose g^{(N_f)}}=\begin{pmatrix}1+a_s^2[A_{qq,h}^{N\!S,(2)}+\tilde{A}^{S,(2)}_{hq}] & a_s\tilde{A}^{S,(1)}_{hg}+a_s^2\tilde{A}^{S,(2)}_{hg}\\
a_s^2A^{S,(2)}_{gq,h} & 1+a_sA_{gg,h}^{S,(1)}+a_s^2A_{gg,h}^{S,(2)}\end{pmatrix}{\Sigma^{(N_f-1)} \choose g^{(N_f-1)}}\,,
\end{array}
\end{equation}
where the form of the functions entering the transformation matrix
above are given in Appendix B of Ref.~\cite{Buza:1996wv}. We omit the
matching conditions for the non-singlet PDF combinations because there
is no $\mathcal{O}(a_s)$ correction and the first correction appears
at $\mathcal{O}(a_s^2)$. This makes the conversion from the pole to
the $\overline{\mbox{MS}}$ mass scheme ineffective up to NNLO
In order to replace the pole mass $M$ with the $\overline{\mbox{MS}}$
mass $m(\mu)$, we just have to plug eq.~(\ref{eq:conversionLog}) into
eq.~(\ref{eq:MatchPDFsPole}). In doing so, only the $\mathcal{O}(a_s)$
terms proportional to $\ln(\mu^2/M^2)$ play a role in the conversion
up to NNLO. Since the functions $\tilde{A}^{S,(1)}_{hg}$ and
$A_{gg,h}^{S,(1)}$ can be written as:
\begin{equation}\label{eq:OasMatchPDFs}
\begin{array}{l}
\displaystyle \tilde{A}^{S,(1)}_{hg}\left(x,\frac{\mu^2}{M^2}\right) = f_1(x)\ln\frac{\mu^2}{M^2}\,,\\
\\
\displaystyle A^{S,(1)}_{gg,h}\left(x,\frac{\mu^2}{M^2}\right) = f_2(x)\ln\frac{\mu^2}{M^2}
\end{array}\,,
\end{equation}
where:
\begin{equation}\label{eq:OasPDFsMatchCoeff}
\begin{array}{l}
\displaystyle f_1(x)= 4 T_R[x^2+(1-x)^2]\,,\\
\\
\displaystyle f_2(x)= -\frac43 T_R \delta(1-x)\,.
\end{array}
\end{equation}
Replacing $M$ with $m$ in eq.~(\ref{eq:OasMatchPDFs}) using
eq. (\ref{eq:conversionLog}), we get:
\begin{equation}
\begin{array}{l}
\displaystyle \tilde{A}^{S,(1)}_{hg}\left(x,\frac{\mu^2}{m^2}\right) = f_1(x)\ln\frac{\mu^2}{m^2}-2h^{(1)}(\mu)f_1(x)a_s(\mu)\\
\\
\displaystyle A^{S,(1)}_{gg,h}\left(x,\frac{\mu^2}{m^2}\right) = f_2(x)\ln\frac{\mu^2}{m^2}-2h^{(1)}(\mu)f_2(x)a_s(\mu)
\end{array}\,.
\end{equation}
Therefore eq. (\ref{eq:MatchPDFsPole}) in terms of $m$ becomes:
\begin{equation}
\begin{array}{c}
\displaystyle {\Sigma^{(N_f)}\choose g^{(N_f)}}=\begin{pmatrix}1+a_s^2[A_{qq,h}^{N\!S,(2)}+\tilde{A}^{S,(2)}_{hq}] & a_s\tilde{A}^{S,(1)}_{hg}+a_s^2[\tilde{A}^{S,(2)}_{hg}-2h^{(1)}f_1]\,,\\
a_s^2A^{S,(2)}_{gq,h} & 1+a_sA_{gg,h}^{S,(1)}+a_s^2[A_{gg,h}^{S,(2)}-2h^{(1)}f_2]\end{pmatrix}{\Sigma^{(N_f-1)} \choose g^{(N_f-1)}}\,.
\end{array}
\end{equation}
As usual, we choose to match the $N_f$ scheme to the $N_f-1$ scheme at
$\mu = m(\mu) = m(m)$ so that all the logarithmic terms vanish,
obtaining:
\begin{equation}
\begin{array}{c}
\displaystyle {\Sigma^{(N_f)}\choose g^{(N_f)}}=\begin{pmatrix}1+a_s^2[A_{qq,h}^{N\!S,(2)}+\tilde{A}^{S,(2)}_{hq}] & a_s^2[\tilde{A}^{S,(2)}_{hg}-2h^{(1)}f_1]\\
a_s^2A^{S,(2)}_{gq,h} & 1+a_s^2[A_{gg,h}^{S,(2)}-2h^{(1)}f_2]\end{pmatrix}{\Sigma^{(N_f-1)} \choose g^{(N_f-1)}}\,.
\end{array}
\end{equation}
\subsubsection*{Renormalization scale variation}
The scale $\mu$ that appears in $a_s$ and $m_q$ is the
\textit{renormalization} scale, which in this section we will indicate
with $\mu_R$. On the other hand, the scale that explicitly appears in
the PDFs is the \textit{factorization} scale, which we will now denote
with $\mu_F$. In principle, renormalization and factorization scales
are different but in general one can write $\mu_R = \kappa \mu_F$,
where $\kappa$ can be any real number\footnote{It should be noticed
that in the case $\kappa\neq1$ PDFs aquire an implicit dependence on
$\mu_R$ that comes from a redefinition of the splitting functions
that in turn derives from the expansion of $\alpha_s(\mu_R)$ around
$\mu_R=\mu_F$ that appears in the DGLAP equation.}.
The most common choice when matching the $N_f-1$ scheme with the $N_f$
scheme is to set $\mu_F$ equal to heavy quark thresholds ($M_c$, $M_b$
and $M_t$ in the pole-mass scheme and $m_c(m_c)$, $m_b(m_b)$ and
$m_t(m_t)$ in the $\overline{\mbox{MS}}$ scheme). In doing this the
logarithmic terms in the PDF matching conditions are assured to
vanish. On the other hand, if $\kappa$ is different from one, the
logarithmic terms in the matching conditions for $a_s(\mu_R)$ and
$m_q(\mu_R)$ do not vanish anymore. This is exactly the case when
renormalization scale variations are considered and in the
following we will show how matching condition
for $a_s$ and $m_q$ change for $\kappa\neq1$.
Let us start with $\alpha_s$. Inverting eq.~(\ref{eq:alphasmsbar}) we
obtain:
\begin{equation}
\frac{a^{(N_f)}(\mu_R)}{a^{(N_f-1)}(\mu_R)} = 1 + c_1a^{(N_f-1)}(\mu_R) + c_2 [a^{(N_f-1)}(\mu_R)]^2\,,
\end{equation}
where:
\begin{equation}
c_1 = \frac23 L_{\mu m}
\quad\mbox{and}\quad
c_2 = \frac49L_{\mu m}^2+\frac{22}3L_{\mu m}-\frac{22}9\,.
\end{equation}
Setting $\mu_F=\kappa \mu_F$ we have that:
\begin{equation}
L_{\mu m} = \ln\frac{\mu_R}{m(\mu_R)}=\ln\frac{\kappa\mu_F}{m(\kappa \mu_F)}\,.
\end{equation}
As usual, the matching scale is chosen to be $\mu_F = m(m)$, so that:
\begin{equation}
L_{\mu m} \rightarrow \ln\kappa + \ln\frac{m(m)}{m(\kappa m)}\,.
\end{equation}
But using eq.~(\ref{eq:massRGEsolution}), it easy to see that:
\begin{equation}
\ln\frac{m(m)}{m(\kappa m)}=a_s(\kappa m)\gamma_m^{(0)}\ln\kappa+\mathcal{O}[a_s^2(\kappa m)]\,,
\end{equation}
so that:
\begin{equation}\label{eq:LmumExp}
L_{\mu m} \rightarrow [1+\gamma_m^{(0)}a_s(\kappa m)]\ln\kappa\,.
\end{equation}
It should be noticed that in the eq.~(\ref{eq:LmumExp}), since
$a_s^{(N_f-1)}=a_s^{(N_f)}+\mathcal{O}([a_s^{(N_f)}]^2)$, it does not
matter whether one puts $a_s^{(N_f)}(\kappa m)$ or
$a_s^{(N_f-1)}(\kappa m)$ because the difference would be subleading.
Therefore, setting $\mu=\mu_R=\kappa m(m) = \kappa m$ in
eq.~(\ref{eq:alphasmsbar}) and using eq. (\ref{eq:LmumExp}), one gets:
\begin{equation}
\begin{array}{rcl}
a^{(N_f-1)}(\kappa m)&=&\displaystyle a^{(N_f)}(\kappa m)\bigg\{1-\frac23
\ln\kappa\,a^{(N_f)}(\kappa
m)\\
\\
&+&\displaystyle \left[\frac49\ln^2\kappa-\frac{2}3\left(\gamma_m^{(0)}+11\right)\ln\kappa+\frac{22}9\right][a^{(N_f)}(\kappa
m)]^2\bigg\}\,,
\end{array}
\end{equation}
whose inverse is:
\begin{equation}
\begin{array}{rcl}
a^{(N_f)}(\kappa m)&=&\displaystyle a^{(N_f-1)}(\kappa m)\bigg\{1+\frac23
\ln\kappa\,a^{(N_f-1)}(\kappa
m)\\
\\
&+&\displaystyle \left[\frac49\ln^2\kappa+\frac{2}3\left(\gamma_m^{(0)}+11\right)\ln\kappa-\frac{22}9\right][a^{(N_f-1)}(\kappa
m)]^2\bigg\}\,.
\end{array}
\end{equation}
Now let us turn to $m_q$. In this case there is not much to do. In
fact, for an arbitary matching point the matching condition of the
$\overline{\mbox{MS}}$ mass starts at $\mathcal{O}(\alpha_s^2)$
(cfr. eq. (\ref{eq:mqmc})), threfore writing $L_{\mu m}$ in terms of
$\ln\kappa$ would give rise to subleading terms (see
eq. (\ref{eq:LmumExp})). As a consequence, up to NNLO, we have that:
\begin{equation}
m_q^{(N_f-1)}(\kappa m_h)=\left[1+\left(\frac43\ln^2\kappa-\frac{20}9\ln\kappa+\frac{89}{27}\right)[a^{(N_f)}(\kappa m_h)]^2\right]m_q^{(N_f)}(\kappa m_h)\,,
\end{equation}
and its inverse is:
\begin{equation}
m_q^{(N_f)}(\kappa
m_h)=\left[1-\left(\frac43\ln^2\kappa-\frac{20}9\ln\kappa+\frac{89}{27}\right)[a^{(N_f-1)}(\kappa
m_h)]^2\right]m_q^{(N_f-1)}(\kappa m_h)\,.
\end{equation}
\subsection{Structure functions}
We finally turn to discuss how the $\overline{\mbox{MS}}$ masses
change the DIS massive structure functions. We will first consider the
neutral-current (NC) massive structure functions up to
$\mathcal{O}(\alpha_s^2)$, which is the highest perturbative order at
which corrections are known exactly, and then we will consider the
charged-current (CC) massive structure functions again up to the
highest perturbative order exactly known, that is
$\mathcal{O}(\alpha_s)$. In order to shorten the notation, in the
following we will adopt the following defintions:
\begin{equation}
M =\;\mbox{pole mass},\quad m\equiv m(\mu) =\;\overline{\mbox{MS}}\mbox{ mass},\quad a_s\equiv a_s(\mu),\quad h^{(l)}\equiv h^{(l)}(\mu,m(\mu))\,.
\end{equation}
\subsubsection{Neutral current}
Dropping all the unnecessary dependences, the NC massive structure
functions up to $\mathcal{O}(a_s^2)$ have the form:
\begin{equation}
F(M) = a_sF^{(0)}(M) + a_s^2F^{(1)}(M) + \mathcal{O}(a_s^3)\,.
\end{equation}
The goal is to replace explicitly the pole mass $M$ with the
$\overline{\mbox{MS}}$ mass $m$ using eq.~(\ref{eq:PoleToMSbar}). To
this end, we expand $F^{(0)}(M)$ and $F^{(1)}(M)$ around $M=m$:
\begin{equation}
F^{(l)}(M) = \sum_{n=0}^{\infty}\frac1{n!}\frac{d^n F^{(l)}}{dM^n}\bigg|_{M=m}(M-m)^n\,,
\end{equation}
so that, up to $\mathcal{O}(a_s^2)$, what we need is:
\begin{equation}
\begin{array}{l}
\displaystyle F^{(0)}(m) = F^{(0)}(m) + a_smh^{(1)}\frac{dF^{(0)}}{dM}\bigg|_{M=m}\,,\\
\\
\displaystyle F^{(1)}(M) = F^{(1)}(m)\,.
\end{array}
\end{equation}
Finally, we have that:
\begin{equation}\label{eq:changescheme}
F(m) = a_sF^{(0)}(m) + a_s^2\left[F^{(1)}(m)+mh^{(1)}\frac{dF^{(0)}}{dM}\bigg|_{M=m}\right]\,.
\end{equation}
We now need to evaluate explicitly the derivative in
eq. (\ref{eq:changescheme}). First of all we observe that:
\begin{equation}\label{eq:convolution0}
F^{(0)}(M) = x\int_x^{x_{\mbox{\tiny max}}(M)} \frac{dz}{z}g\left(\frac{x}{z}\right)C_g^{(0)}(\eta(z,M),\xi(M),\chi(M))\,,
\end{equation}
being $g$ the gluon distribution function and where we have used the
following definitions:
\begin{equation}
x_{\mbox{\tiny max}}(M)=\frac1{1+\frac{4M^2}{Q^2}},\quad\eta(z,M) = \frac{Q^2}{4M^2}\left(\frac1z-1\right)-1,\quad \xi(M) =\frac{Q^2}{M^2},\quad \chi(M) =\frac{\mu^2}{M^2}\,.
\end{equation}
Defining:
\begin{equation}
G(z,M)=\frac{x}{z}g\left(\frac{x}{z}\right)C_g^{(0)}(\eta(z,M),\xi(M),\chi(M))\,,
\end{equation}
the derivative of eq.~(\ref{eq:convolution0}) can be written as:
\begin{equation}
\frac{dF^{(0)}}{dM} = \frac{d}{dM}\int_x^{x_{\mbox{\tiny max}}(M)} dzG(z,M) = \frac{d\widetilde{G}(x_{\mbox{\tiny max}}(M),M)}{dM}-\frac{d\widetilde{G}(x,M)}{dM}\,,
\end{equation}
where $\widetilde{G}(z,M)$ is the primitive of $G(z,M)$ with respect of $z$ (i.e. $\partial\widetilde{G}/\partial z = G$). But:
\begin{equation}
\frac{d\widetilde{G}(x_{\mbox{\tiny max}}(M),M)}{dM} = \frac{d \widetilde{G}(x_{\mbox{\tiny max}},M)}{d M}+\frac{dx_{\mbox{\tiny max}}}{dM}G(x_{\mbox{\tiny max}},M)\,,
\end{equation}
thus:
\begin{equation}\label{eq:withbound}
\begin{array}{c}
\displaystyle \frac{dF^{(0)}}{dM} = \frac{\partial \widetilde{G}(x_{\mbox{\tiny max}},M)}{\partial M}-\frac{\partial\widetilde{G}(x,M)}{\partial M}+\frac{dx_{\mbox{\tiny max}}}{dM}G(x_{\mbox{\tiny max}},M) =\\
\\
\displaystyle \int_x^{x_{\mbox{\tiny max}}(M)} dz\frac{\partial G(z,M)}{\partial M}+\frac{dx_{\mbox{\tiny max}}}{dM}G(x_{\mbox{\tiny max}},M)\,.
\end{array}
\end{equation}
It can be shown that bounduary term in eq.~(\ref{eq:withbound})
vanishes ($e.g.$ see Ref.~\cite{Alekhin:2010sv}), thus it can be
omitted.
Gathering all pieces, since:
\begin{equation}
\frac{\partial G(z,M)}{\partial M} = \frac{x}{z}g\left(\frac{x}{z}\right)\frac{\partial C_g^{(0)}}{\partial M}\,,
\end{equation}
we have that:
\begin{equation}\label{eq:withoutbound}
\begin{array}{rcl}
\displaystyle \frac{dF^{(0)}}{dM}\bigg|_{M=m}&=&\displaystyle
\left[x\int_x^{x_{\mbox{\tiny
max}}(M)}\frac{dz}{z}g\left(\frac{x}{z}\right)\frac{\partial
C_g^{(0)}}{\partial
M}\right]\Bigg|_{M=m}\\
\\
&=&\displaystyle x\int_x^{x_{\mbox{\tiny
max}}(m)}\frac{dz}{z}g\left(\frac{x}{z}\right)\left[\frac{\partial
C_g^{(0)}}{\partial M}\right]\Bigg|_{M=m}\,.
\end{array}
\end{equation}
Finally, taking into account that:
\begin{equation}
F^{(1)}(M) = \sum_{i=q,\overline{q},g}x\int_x^{x_{\mbox{\tiny max}}(M)} \frac{dz}{z}q_i\left(\frac{x}{z}\right)C_i^{(1)}(z,M)
\end{equation}
and using eqs.~(\ref{eq:changescheme}) and~(\ref{eq:withoutbound}),
one can explicitly write down the full structure of the NNLO massive
structure functions ($F_2$ and $F_L$) in terms of $\overline{\mbox{MS}}$ masses as
follows:
\begin{equation}\label{eq:masterNC}
\begin{array}{c}
\displaystyle F(m) = x\int_x^{x_{\mbox{\tiny max}}(m)} \frac{dz}{z}g\left(\frac{x}{z}\right)\left[a_sC_g^{(0)}(z,m)+a_s^2\left(C_g^{(1)}(z,m)+mh^{(1)}\left[\frac{\partial C_g^{(0)}}{\partial M}\right]\Bigg|_{M=m}\right)\right]+\\
\\
\displaystyle \sum_{i=q,\overline{q}}x\int_x^{x_{\mbox{\tiny max}}(M)} \frac{dz}{z}q_i\left(\frac{x}{z}\right)a_s^2C_i^{(1)}(z,M)\,.
\end{array}
\end{equation}
In order to carry out the implementation, we need to evaluate
explicitly the derivative of $C_g^{(0)}$ in eq. (\ref{eq:masterNC})
and this must be done separately for $F_2$ and $F_L$.
We consider first $F_2$. The general structure of $C_{2,g}^{(0)}$ is
the following:
\begin{equation}
\begin{array}{rl}
\displaystyle C_{2,g}^{(0)}(z,Q^2,M^2)=&\displaystyle T_R\Big\{2(1-6\epsilon-4\epsilon^2)I_2(\epsilon,z)-2(1-2\epsilon)I_1(\epsilon,z)+I_0(\epsilon,z)+\\
\\
& \displaystyle -4(2-\epsilon)J_2(\epsilon,z)+4(2-\epsilon)J_1(\epsilon,z)-J_0(\epsilon,z)\Big\}\,,
\end{array}
\end{equation}
where:
\begin{equation}\label{eq:iq}
I_q(\epsilon,z) = z^q\ln\left(\frac{1+v}{1-v}\right)\,.
\end{equation}
\begin{equation}\label{eq:jq}
J_q(\epsilon,x) = z^q v\,,
\end{equation}
with:
\begin{equation}\label{eq:definitions1}
\epsilon = \frac{M^2}{Q^2}\,,\quad
a=\frac1{1+4\epsilon}\quad\mbox{and}\quad v=\sqrt{1-4\epsilon\frac{z}{1-z}}\,.
\end{equation}
From the definitions in eq.~(\ref{eq:definitions1}), we obtain:
\begin{equation}\label{eq:derivatives}
\begin{array}{rl}
\displaystyle \frac{\partial}{\partial M} &\displaystyle = \frac{\partial \epsilon}{\partial M} \frac{\partial}{\partial \epsilon} = \frac{2\epsilon}{M}\frac{\partial}{\partial \epsilon}\,,\\
\\
\displaystyle \frac{\partial}{\partial M} &\displaystyle = \frac{\partial \epsilon}{\partial M} \frac{\partial v}{\partial \epsilon} \frac{\partial}{\partial v} = -\frac{1-v^2}{Mv}\frac{\partial}{\partial v}\,.
\end{array}
\end{equation}
Therefore:
\begin{equation}
\begin{array}{rl}
\displaystyle \frac{\partial C_{2,g}^{(0)}}{\partial M}=&\displaystyle \frac{1}{M}T_R\Bigg\{2\epsilon\Big[2(-6-8\epsilon)I_2+4I_1+4J_2-4J_1\Big]\\
\\
&\displaystyle -\frac{1-v^2}{v}\Bigg[2(1-6\epsilon-4\epsilon^2)\frac{\partial I_2}{\partial v}-2(1-2\epsilon)\frac{\partial I_1}{\partial v}+\frac{\partial I_0}{\partial v}\\
\\
& \displaystyle -4(2-\epsilon)\frac{\partial J_2}{\partial v}+4(2-\epsilon)\frac{\partial J_1}{\partial v}-\frac{\partial J_0}{\partial v}\Bigg]\Bigg\}\,.
\end{array}
\end{equation}
To find the explicit expression, we just need to evaluate the
derivative of $I_q$ and $J_q$ starting from eqs. (\ref{eq:iq}) and
(\ref{eq:jq}) which is easily done:
\begin{equation}
\begin{array}{rcl}
\displaystyle \frac{\partial I_q}{\partial v} &=& \displaystyle\frac{2 z^q}{1-v^2}\,,\\
\\
\displaystyle \frac{\partial J_q}{\partial v} &=& \displaystyle z^q\,.
\end{array}
\end{equation}
In the end we get:
\begin{equation}\label{eq:dC2g0}
\begin{array}{rl}
\displaystyle \frac{\partial C_{2,g}^{(0)}}{\partial M}=&\displaystyle
\frac{1}{M}T_R\Bigg\{4\epsilon\left[(-6-8\epsilon)z^2+2z\right]\ln\left(\frac{1+v}{1-v}\right)+8\epsilon
z(z-1)v\\
\\
&\displaystyle -\frac{2}{v}\left[2(1-6\epsilon-4\epsilon^2)z^2-2(1-2\epsilon)z+1\right]\\
\\
& \displaystyle -\frac{1-v^2}{v}\left[-4(2-\epsilon)z^2+4(2-\epsilon)z-1\right]\Bigg\}\,.
\end{array}
\end{equation}
As is well known, the implementation of the FONLL scheme requires the
massless limit of the massive structure functions. In practice this
means that we need to compute the limit $M\rightarrow 0$ of the
massive coefficient functions retaining the logarithmic enhanced terms
and droping the power suppressed ones. To apply this recipe to
eq.~(\ref{eq:dC2g0}), we need to keep in mind that:
\begin{equation}
\epsilon \mathop{\longrightarrow}_{M\rightarrow 0} 0 \,,\quad v \mathop{\longrightarrow}_{M\rightarrow 0} 1\,,
\end{equation}
and that:
\begin{equation}
\ln\left(\frac{1+v}{1-v}\right) \mathop{\longrightarrow}_{M\rightarrow
0} \ln\frac{Q^2(1-z)}{M^2z}\,,
\end{equation}
so that:
\begin{equation}\label{eq:dC2g00}
\displaystyle \frac{\partial C_{2,g}^{(0)}}{\partial M}\mathop{\longrightarrow}_{M\rightarrow 0}\displaystyle \frac{\partial C_{2,g}^{0,(0)}}{\partial M}=
-\frac{2}{M}T_R\left(2z^2-2z+1\right)\,.
\end{equation}
We now turn to consider $F_L$. In this case the the gluon coefficient
function takes the simpler form:
\begin{equation}
C_{L,g}^{(0)}\left(z,Q^2,M^2\right)= T_R\left[-8\epsilon I_2(\epsilon,z)-4J_2(\epsilon,z)+4J_1(\epsilon,z)\right]\,.
\end{equation}
Therefore, using eq. (\ref{eq:derivatives}), we immediately get:
\begin{equation}
\frac{\partial C_{L,g}^{(0)}}{\partial M} = \frac1{M}T_R\left[-16\epsilon
z^2\ln\left(\frac{1+v}{1-v}\right) +\frac{8\epsilon z^2}{v}-\frac{1-v^2}{v}\left(-4z^2+4z\right)\right]\,.
\end{equation}
It is finally easy to realize that:
\begin{equation}
\frac{\partial C_{L,g}^{(0)}}{\partial
M}\mathop{\longrightarrow}_{M\rightarrow 0} \frac{\partial C_{L,g}^{0,(0)}}{\partial M}=0\,.
\end{equation}
This concludes the treatment of the NC massive structure functions.
\subsubsection{Charged current}
In this section we consider the generic CC massive structure
function. The treatment is exactly the same of the NC structure
functions, with the only difference that in the CC case they start at
$\mathcal{O}(1)$ and they are presently known up to
$\mathcal{O}(a_s)$. This means that their perturbative expansion in
terms of the pole mass $M$ looks like this:
\begin{equation}
F_k(M) = F_k^{(0)}(M) + a_sF_k^{(1)}(M) + \mathcal{O}(a_s^2)\,,
\end{equation}
with $k=2,3,L$. Therefore, expanding $F^{(0)}$ and $F^{(1)}$ around $M=m$ and keeping only the terms up to $\mathcal{O}(a_s)$, one obtains:
\begin{equation}\label{eq:expCC}
F_k(m) = F_k^{(0)}(m) + a_s\left[F_k^{(1)}(m)+mh^{(1)}\frac{dF_k^{(0)}}{dM}\bigg|_{M=m}\right]\,.
\end{equation}
The leading-order contribution to the coefficient function can be
written as follows:
\begin{equation}
F^{(0)}_k(M) = b_k(M)s'(\xi(M))\,,
\end{equation}
where:
\begin{equation}\label{eq:definitions}
\xi = x\underbrace{\left(1+\frac{M^2}{Q^2}\right)}_{\frac1\lambda}=\frac{x}\lambda\quad\mbox{and}\quad
\left\{\begin{array}{l}
b_2 = \xi\\
b_3 = 1\\
b_L = (1-\lambda)\xi
\end{array}
\right.
\end{equation}
and where we have also defined:
\begin{equation}
s'\equiv 2|V_{cs}|^2s+2|V_{cd}|^2d\,.
\end{equation}
Therefore:
\begin{equation}\label{eq:derivCC}
\begin{array}{rcl}
\displaystyle mh^{(1)}\frac{dF^{(0)}_k}{dM}\bigg|_{M=m} &=&\displaystyle
mh^{(1)}\frac{d\xi}{dM}\frac{dF^{(0)}_k}{d\xi}\bigg|_{M=m} \\
\\
&=&\displaystyle
2h^{(1)}(1-\lambda)\xi\left[\frac{db_k}{d\xi}s'(\xi)+b_k(\xi)\frac{ds'}{d\xi}\right]\bigg|_{M=m}\,,
\end{array}
\end{equation}
that can be conveniently rewritten as:
\begin{equation}
mh^{(1)}\frac{dF^{(0)}_k}{dM}\bigg|_{M=m} = 2h^{(1)}(1-\lambda)\left[\left(\frac{db_k}{d\xi}-\frac{b_k}{\xi}\right)+b_k(\xi)\frac{d}{d\xi}\right]\xi s'(\xi)\bigg|_{M=m}\,,
\end{equation}
so that, using eq.~(\ref{eq:definitions}), we have that:
\begin{equation}\label{eq:CCcorrections}
\begin{array}{rcl}
\displaystyle mh^{(1)}\frac{dF^{(0)}_2}{dM}\bigg|_{M=m} &=& \displaystyle
2h^{(1)}(1-\lambda)\xi\frac{d}{d\xi} \xi s'(\xi)\bigg|_{M=m}\,,\\
\\
\displaystyle mh^{(1)}\frac{dF^{(0)}_3}{dM}\bigg|_{M=m} &=& \displaystyle
2h^{(1)}(1-\lambda)\frac{1}{\xi}\left[ \xi\frac{d}{d\xi}-1\right]\xi
s'(\xi)\bigg|_{M=m}\,,\\
\\
\displaystyle mh^{(1)}\frac{dF^{(0)}_L}{dM}\bigg|_{M=m} &=& \displaystyle
2h^{(1)}(1-\lambda)^2\xi \frac{d}{d\xi}\xi s'(\xi)\bigg|_{M=m}\,.
\end{array}
\end{equation}
Finally, we notice that in the massless limit, where
$\lambda\rightarrow 1$, all expressions in
eq.~(\ref{eq:CCcorrections}) vanish, with the consequence that the CC
massive structure functions up to $\mathcal{O}(a_s)$ in terms of the
pole mass $M$ or the $\overline{\mbox{MS}}$ mass $m$ look exactly the
same.
As a last comment, the derivative of the PDF $s'$ that appears in
eq.~(\ref{eq:CCcorrections}) is implemented in {\tt APFEL} by means
the derivative of the Lagrange polynomials that are used to
interpolate PDFs.
\section{QCD fit settings} \label{sec:fitsettings}
The QCD fits were performed to the combined H1 and ZEUS charm
production cross-section measurements~\cite{Abramowicz:1900rp}
together with the combined HERA1+2 H1 and ZEUS inclusive DIS
cross-section data~\cite{Abramowicz:2015mha}, accounting
for all given sources of systematic uncertainties.
The kinematic region covered by HERA is constrained by the invariant
mass of the hadronic system of $W > 15$ GeV and the Bjorken scaling
variable of $x < 0.65$, therefore target mass corrections are expected
to have negligible effects and are not discussed in this paper. The
settings of the QCD fits in {\tt xFitter} closely follow those used
for the HERAPDF2.0 PDF extraction~\cite{Abramowicz:2015mha}, with a
few differences related to the specifics of the current analysis which
are motivated in the following.
The nominal result is extracted using the FONLL-C variant of the FONLL
scheme discussed in Sect.~\ref{sec:fonll}. It should be pointed out
that, while being accurate at NNLO for the inclusive DIS cross
sections, the sensitivity to mass corrections of the FONLL-C scheme is
actually NLO. The reason is that at $\mathcal{O}(\alpha_s^0)$ the
FONLL scheme reduces to the parton model which is insensitive to
heavy-quark mass effects. Therefore, the first mass-sensitive term is
$\mathcal{O}(\alpha_s)$ which is the accuracy of the FONLL-A scheme
which would thus provide a LO determination of the charm mass. Both
the FONLL-B and the FONLL-C schemes, instead, include the
$\mathcal{O}(\alpha_s^2)$ massive corrections and thus would both
produce determinations of the mass of the charm accurate at NLO. The
advantage of FONLL-C with respect to FONLL-B is that it is accurate at
$\mathcal{O}(\alpha_s^2)$ also in the massless sector and thus it is
supposed to provide a better description of the data. In other words,
FONLL-C is the most accurate variant of the FONLL scheme presently
available and as such it will be employed for our determination of
$m_c(m_c)$.
The result obtained in the FONLL scheme is accompanied by an analogous
determination of $m_c(m_c)$ obtained using the FFN scheme with
$\overline{\rm MS}$ masses~\cite{Alekhin:2013nda} at NLO. Access to
the structure functions calculated with the FFN scheme is possible via
the {\tt xFitter} interface to the {\tt OPENQCDRAD}
program~\cite{openqcdrad} using the {\tt QCDNUM} program for the PDF
evolution~\cite{qcdnum}.
The procedure to determine the $\overline{\rm MS}$ charm mass follows
closely the methodology described in Ref.~\cite{Abramowicz:1900rp}. It
involves a series of fits in each of which a set of PDFs is determined
corresponding to numerical values of charm mass ranging between
$m_c(m_c)=1.15$ GeV and $m_c(m_c)=1.60$ GeV with steps of $0.05$ GeV.
For each value of $m_c(m_c)$ a value of global $\chi^2$ is obtained.
The best fit value of $m_c(m_c)$ is determined from the minimum of the
parabolic fit to the resulting $\chi^2$ distribution
and the associated 1-$\sigma$ uncertainty, which reflects the
sensitivity of the data set to the charm mass, is determined as the
$\Delta\chi^2=1$ variation around the minimum.
We now discuss the settings of the nominal fits and the variations
that we performed to assess the different sources of uncertainty
deriving from: the PDF parametrization, the model parameters, and the
theoretical assumptions.
The assumption that heavy-quark PDFs are dynamically generated via
gluon splitting at the respective thresholds requires that the
starting scale $Q_0$ at which PDFs are parametrized is below the charm
threshold, which in turn is identified with $m_c(m_c)$. Given the
range in which the scan of $m_c(m_c)$ is done (from 1.1 to 1.6 GeV),
we have chosen to set $Q_0 = 1$ GeV to allow all fits to be
parametrized at the same starting scale. The combinations and the
relative functional forms of the initial scale PDFs have been chosen
following the parametrization scan procedure as performed for the
HERAPDF2.0 determination~\cite{Abramowicz:2015mha}, and the optimal
configuration has been found to be:
\begin{equation}\label{eq:param}
\begin{array}{rclcl}
xg(x) & & &=& A_gx^{B_g}(1-x)^{C_g}-A'_gx^{B'_g}(1-x)^{25},\\
xu_v(x) &=& xu(x)-x\overline{u}(x) &=& A_{u_v}x^{B_{u_v}}(1-x)^{C_{u_v}}(1+E_{u_v}x^2),\\
xd_v(x) &=& xd(x)-x\overline{d}(x) &=& A_{d_v}x^{B_{d_v}}(1-x)^{C_{d_v}},\\
x\bar{U}(x) &=& x\overline{u}(x) &=& A_{\bar{U}}x^{B_{\bar{U}}}(1-x)^{C_{\bar{U}}}(1+D_{\bar{U}}x),\\
x\bar{D}(x) &=& x\overline{d}(x)+x\overline{s}(x) &=& A_{\bar{D}}x^{B_{\bar{D}}}(1-x)^{C_{\bar{D}}}.
\end{array}
\end{equation}
There are $14$ free parameters, since additional constraints were
applied as follows. The QCD sum rules are imposed at the starting
scale and constrain the normalisation parameters $A_g$, $A_{u_v}$,
$A_{d_v}$. The light-sea quark parameters that affect the low-$x$
kinematic region $B_{\bar{U}}$ and $B_{\bar{D}}$, as well as the
normalisation parameters $A_{\bar{U}}$ and $A_{\bar{D}}$, are
constrained by the requirement that $\bar{u}\rightarrow \bar{d}$ as $x
\to 0$, leading to the following constraints:
\begin{eqnarray}
B_{\bar{U}} &=& B_{\bar{D}} ,\\
A_{\bar{U}} &=& A_{\bar{D}}(1-f_s),
\label{eq:constr}
\end{eqnarray}
with $f_s$ being the strangeness fraction of $\bar{D}$ assumed at the
starting scale, $i.e.$ $f_s=\bar{s} / \bar{D}$, because HERA data alone
are not able to provide a precise light-sea flavor separation. The
strangeness fraction for the nominal fits is set to $f_s = 0.4$, as in
the HERAPDF2.0 analysis~\cite{Abramowicz:2015mha}.
In order to estimate the uncertainty associated to the PDF
parametrization, we have considered the following variations with
respect to the nominal configuration:
\begin{itemize}
\item we have moved up the initial scale $Q_0$ from 1 to $\sqrt{1.5}$
GeV. In the FONLL scheme, this restricted the $m_c(m_c)$ range in
which we did the scan because we could not use values of the charm
mass such that $m_c(m_c) < \sqrt{1.5}$ GeV. We were however able to
perform the parabolic fit in order to find the best fit value of
$m_c(m_c)$. This complication does not arise in the FFN scheme in
which there is no threshold crossing.
\item In the $xu_v$ distribution we have included an additional linear
term so that the last factor in second line of eq.~(\ref{eq:param})
reads $(1+D_{u_v} x + E_{u_v} x^2)$. After trying different
variations of the parametrization, we found that this particular
choice leads to the largest differences.
\end{itemize}
Moving to the model parameters, the values of the bottom and top quark
masses for the nominal fits are chosen to be equal to the PDG values,
defined in the $\overline{\rm MS}$ scheme, $i.e.$ $m_b(m_b)=4.18$ GeV
and $m_t(m_t)=160$ GeV~\cite{Agashe:2014kda}. The value of the strong
coupling is set to $\alpha_s(M_Z) = 0.118$. It should be pointed out
that this value of $\alpha_s$ assumes 5 active flavors. For the FFN
scheme fits, though, one needs to use the value of $\alpha_s$ with 3
active flavors. In order to find this value one has to evolve
$\alpha_s(M_Z)$ down to below $m_c(m_c)$ in the VFN scheme and evolve
back to $M_Z$ with 3 active flavors. We have computed the value of
$\alpha_s$ with 3 active flavors for each of the values of $m_c(m_c)$
considered.
The uncertainty associated to model parameters will be estimated by
considering the following variations:
\begin{itemize}
\item the bottom mass has been moved up and down by 0.25 GeV, $i.e.$
$m_b(m_b) = 3.93$ GeV and $m_b(m_b) = 4.43$ GeV. The magnitude of
the variation is actually much larger than the present uncertainty
on the bottom mass and thus our choice is meant to provide a
conservative estimate of the associated uncertainty.
\item The variation of the strong coupling follows the recent PDF4LHC
prescription~\cite{Butterworth:2015oua}. In particular, we have
considered the conservative variation up and down by 0.0015 with
respect to the nominal value, $i.e.$ $\alpha_s(M_Z) = 0.1165$ and
$\alpha_s(M_Z) = 0.1195$.
\item Finally, we considered the value of the strangeness fraction
introduced in eq.~(\ref{eq:constr}) as being a model parameter and
we have thus varied it up and down by 0.1 around the nominal value
considering $f_s=0.3$ and $f_s=0.5$.
\end{itemize}
We finally turn to the theory assumptions and their variations. These
mostly concern unknown higher-order corrections and the most common
way to estimate them is by varying the renormalization and the
factorization scales $\mu_R$ and $\mu_F$. As nominal scales in our
analysis we have chosen $\mu_R^2 =\mu_F^2 = Q^2$ for both the
FONLL\footnote{A scale choice involving the heavy-quark mass would
lead to technical complications with the FONLL matching as
implemented in {\tt APFEL}. However, we have checked that the more
commonly used scales $\mu_R^2 =\mu_F^2 = Q^2+4m_c(m_c)^2$ produce a
very marginal difference in the determination of $m_c(m_c)$ in the
FFN scheme.} and the FFN scheme analyses. Another possible source of
theoretical uncertainty in the FONLL scheme is the presence of the
damping factor discussed in Sect.~\ref{sec:fonll} which is meant to
suppress unwanted subleading terms and whose explicit form in the
nominal fits is given in eq.~(\ref{eq:dampingfactor}).
The theoretical uncertainty associated to the missing higher-order
corrections has been estimated as follows:
\begin{itemize}
\item the factorization and renormalization scales were varied by a
factor 2 up and down with respect to the nominal values, that is
choosing $\mu_R^2 =\mu_F^2 = Q^2 / 2$ and $\mu_R^2 =\mu_F^2 = 2Q^2$.
Such variations have been applied only to the heavy-quark components
of the structure functions, while the light part has been left
unchanged. The reason for this is that, in order to estimate the
theoretical uncertainty associated to the determination of
$m_c(m_c)$, we want to perform scale variations only in the part of
the calculation sensitive to this parameter, which is clearly the
charm structure function (for consistency, the same variation was
applied also to the bottom structure functions).
\item As already mentioned, the FONLL damping factor represents a
further source of uncertainty. It has the role of suppressing
unwanted subleading terms but the particular way in which this
suppression is implemented is somewhat arbitrary. To assess the
impact of our particular choice on the determination of $m_c(m_c)$,
we have changed the suppression power around the nominal one,
considering the following functional form:
\begin{equation}
D_p(Q,m_c)=\theta(Q^2-m_c^2)\left(1-\frac{m_c^2}{Q^2}\right)^p\,,
\end{equation}
with $p=1,4$.
\end{itemize}
In addition, to assure the applicability of perturbative QCD and to
keep higher-twist corrections under control, a cut on $Q^2$ is imposed
on the fitted data. Our nominal cut is $Q^2 > Q_{\rm min}^2=3.5$
GeV$^2$. The choice of the value of $Q_{\rm min}^2$ requires some
care; an extensive discussion on the impact of varying it on the
determination of $m_c(m_c)$ is given in
Sect.~\ref{sec:Q2minDependence}.
To conclude this section, we observe that the self-consistency of the
input data set and the good control of the systematic uncertainties
enable the determination of the experimental uncertainties in the PDF
fits using the tolerance criterion of $\Delta \chi^2=1$.
\section{FONLL with $\overline{\rm MS}$ heavy-quark masses}\label{sec:fonll}
In this section we discuss how the FONLL general-mass
variable-flavor-number scheme for DIS structure functions can be
expressed in terms of $\overline{\rm MS}$ heavy-quark masses. We also
describe the subsequent implementation in the public code {\tt APFEL},
and present a number of benchmark comparisons with other public codes.
In general, higher-order calculations are affected by ambiguities in
the prediction for the physical quantities due to the choice of the
subtraction scheme used to remove divergences. In fact, different
prescriptions imply different numerical values of the parameters of
the underlying theory.
As far as the mass parameters are concerned, the pole mass definition
is usually more common in the calculation of massive higher-order QCD
corrections to heavy-quark production processes. The main reason for
this is that the pole mass is, by its own definition, more closely
connected to what is measured in the experiments. On the other hand,
it is well known that observables expressed in terms of the pole mass
present a slow perturbative convergence. This is caused by the fact
that the pole mass definition suffers from non-perturbative effects
which result in an intrinsic uncertainty of order
$\Lambda_{\rm QCD}$~\cite{Chetyrkin:1999qi}. The $\overline{\rm MS}$
scheme, which stands for modified minimal subtraction scheme, is
instead free of such ambiguities and as a matter of fact massive
computations expressed in terms of heavy-quark masses normalized in
this scheme present a better perturbative
convergence~\cite{Alekhin:2010sv}. As a consequence, the results
obtained in the $\overline{\rm MS}$ scheme are more appropriate to
achieve a reliable determination of the numerical value of the charm
mass.
The FONLL scheme, as any other GM-VFN scheme, aims at improving the
accuracy of fixed-order calculations at high scales by matching them
to resummed computations. In DIS this results in the combination of
massive (fixed-order) calculations, that are more reliable at scales
closer to the heavy-quark masses, with resummed calculations that are
instead more accurate at scales much larger than the heavy-quark
masses. However, in the original derivation, the massive component of
the FONLL scheme was expressed in terms of the pole
masses~\cite{Forte:2010ta}.
It is then one of the goals of this paper to provide a full
formulation of the FONLL scheme applied to DIS structure functions in
terms of $\overline{\rm MS}$ masses. A detailed discussion on such a
formulation is given below in Sect.~\ref{sec:implementation}. Here, we
limit ourselves to describing the main steps needed.
The generic form of the DIS structure functions in the FONLL approach
applied to charm production is:
\begin{equation}\label{eq:FONLLdef}
\begin{array}{rcl}
F(x,Q,m_c) &=& F^{(3)}(x,Q,m_c) + F^{(d)}(x,Q,m_c)\\
\\
F^{(d)}(x,Q,m_c) &=& F^{(4)}(x,Q) - F^{(3,0)}(x,Q,m_c)\,,
\end{array}
\end{equation}
where $x$, $Q$, and $m_c$ are the Bjorken variable, the virtuality of
the photon, and the mass of the charm quark, respectively. In
eq.~(\ref{eq:FONLLdef}) the three-flavor structure function $F^{(3)}$
is evaluated retaining the full charm-mass dependence and with no
charm in the initial state. The four-flavor structure function
$F^{(4)}$ is instead computed by setting $m_c$ to zero and allowing
for charm in the initial state, and its associated PDF reabsorbs the
mass (collinear) divergences which are in turn resummed by means of
the DGLAP evolution. Finally, $F^{(3,0)}$ represents the massless
limit of $F^{(3)}$ where all the massive power corrections are set to
zero and only the logarithmically enhanced terms are retained. This
last term is meant to subtract the double counting terms resulting
from the sum of $F^{(3)}$ and $F^{(4)}$. In fact, the role of
$F^{(3,0)}$ is twofold: for $Q\gg m_c$, by definition $F^{(3)}$ and
$F^{(3,0)}$ tend to the same value so that the FONLL structure
function reduces to $F^{(4)}$. By contrast, in the region where
$Q \simeq m_c$ it can be shown that $F^{(d)}$ becomes subleading in
$\alpha_s$ reducing the FONLL structure function to $F^{(3)}$ up to
terms beyond the nominal perturbative accuracy.
It should be noticed that, even though $F^{(d)}$ in
eq.~(\ref{eq:FONLLdef}) becomes subleading in the low-energy region,
it might become numerically relevant and it is advisable to suppress
it. To this end, the term $F^{(d)}$ in eq.~(\ref{eq:FONLLdef}) is
usually replaced by:
\begin{equation}
F^{(d')}(x,Q,m_c) = D(Q,m_c)F^{(d)}(x,Q,m_c)\,,
\end{equation}
where the function $D(Q,m_c)$ is usually referred to as the
\textit{damping factor} and has the explicit form:
\begin{equation}\label{eq:dampingfactor}
D(Q,m_c)=\theta(Q^2-m_c^2)\left(1-\frac{m_c^2}{Q^2}\right)^2\,.
\end{equation}
The role of the damping factor is clearly that of setting $F^{(d')}$
to zero for $Q<m_c$, suppressing it for $Q\gtrsim m_c$, and reducing
it to $F^{(d)}$ for $Q\gg m_c$. It should be pointed out that the
particular functional form of the damping factor given in
eq.~(\ref{eq:dampingfactor}) is somewhat arbitrary. In fact, any
function $D$ such that $F^{(d')}$ and $F^{(d)}$ only differ by
power-suppressed terms, namely:
\begin{equation}
D(Q,m_c) = 1 + \mathcal{O}\left(\frac{m_c^2}{Q_2}\right)\,,
\end{equation}
is a formally suitable choice. In the results section we will also
consider the effect of varying the functional form of the damping
factor in order to estimate the associated theoretical uncertainty on
$m_c(m_c)$.
Given the possible different perturbative structure of the elements
that compose the FONLL structure function in eq.~(\ref{eq:FONLLdef}),
two possibilities for the definition of the perturbative ordering are
possible: the \textit{relative} and the \textit{absolute}
definitions. In the relative definition $F^{(4)}$ and $F^{(3)}$ are
combined using the same relative perturbative accuracy, that is LO
with LO, NLO with NLO, and so on. The absolute definition, instead, is
such that LO refers to $\mathcal{O}(\alpha_s^0)$ (parton model), NLO
to $\mathcal{O}(\alpha_s)$, and so forth. This issue is relevant in
the neutral-current case where the lowest non-vanishing order is
$\mathcal{O}(\alpha_s^0)$ for $F^{(4)}$ and $\mathcal{O}(\alpha_s)$
for $F^{(3)}$\footnote{This is strictly true only if the heavy-quark
PDFs are dynamically generated via gluon splitting. In fact, the
presence of an intrinsic heavy-quark component would introduce a
$\mathcal{O}(\alpha_s^0)$ contribution also in $F^{(3)}$ leading to
a ``realignment'' of the perturbative structure between $F^{(4)}$
and $F^{(3)}$ (see Refs.~\cite{Ball:2015tna,Ball:2015dpa}).} such
that the relative and absolute orderings lead to different
prescriptions.
Beyond LO, there are currently three possible variants of the FONLL
scheme, all of them implemented in {\tt APFEL}:
\begin{itemize}
\item the FONLL-A variant adopts the absolute ordering at
$\mathcal{O}(\alpha_s)$ and thus only terms up to this accuracy are
included. This variant is formally NLO and thus also PDFs should be
evolved using the same accuracy in the DGLAP evolution.
\item The FONLL-B variant is instead computed using the relative
ordering at NLO. Therefore, $F^{(4)}$ is computed at
$\mathcal{O}(\alpha_s)$ and combined with $F^{(3)}$ at
$\mathcal{O}(\alpha_s^2)$. $F^{(3,0)}$ is instead computed dropping
the non-logarithmic $\mathcal{O}(\alpha_s^2)$ term to match the
accuracy of $F^{(4)}$ in the low-energy region. PDFs are again
evolved at NLO.
\item Finally, the FONLL-C scheme adopts the absolute ordering at
$\mathcal{O}(\alpha_s^2)$. This is formally a NNLO scheme thus PDFs
should be evolved using the same accuracy.
\end{itemize}
Presently, no other variant beyond FONLL-C can be pursued because the
$\mathcal{O}(\alpha_s^3)$ massive coefficient functions are not known
yet. Approximate NNLO corrections valid near the partonic threshold,
in the high-energy (small-$x$) limit, and at high scales $Q^2 \gg m^2$
have been derived in Ref.~\cite{Kawamura:2012cr} and they are
currently employed by the ABM group to determine NNLO
PDFs~\cite{Alekhin:2013nda}.
As clear from the description above, the computations for the
three-flavor structure functions $F^{(3)}$ and $F^{(3,0)}$ depend
explicitly on the charm mass, while $F^{(4)}$ does not. In addition,
as already mentioned, the expressions needed to compute $F^{(3)}$ and
$F^{(3,0)}$ are usually given in terms of the pole mass. As a
consequence, one of the steps required to achieve a full formulation
of the FONLL structure functions in terms of $\overline{\rm MS}$
masses is the adaptation of the heavy-flavor contributions to the
structure functions. A thorough explanation of the procedure adopted
to perform such transformation can be found in
Ref.~\cite{Alekhin:2010sv} for both neutral- and charged-current
structure functions. In Sect.~\ref{sec:implementation} we re-derive
the main formulae and report the full expressions for the relevant
coefficient functions. It should be pointed out that the derivation
presented in Ref.~\cite{Alekhin:2010sv} is performed assuming
$\mu_R=m_c(m_c)$, $\mu_R$ being the renormalisation scale, and the
renormalisation scale dependence of $\alpha_s$ is restored only at the
end using the expansion of the solution of the relative RG
equation. Such a procedure implies that the heavy-quark mass is not
subject to the relative RG equation: in other words, the mass running
is not expressed explicitly. The reason is that in the running of the
heavy-quark mass in $\overline{\rm MS}$ one can resum logarithms of
$\mu_R/m_c(m_c)$ and this is not required in a fixed-order
calculation. On the contrary, when dealing with a GM-VFN scheme like
FONLL, such a resummation is an important ingredient and thus should
be consistently incorporated into the derivation. For this reason, the
transition from pole to $\overline{\rm MS}$ masses of the massive
structure functions presented in Sect.~\ref{sec:implementation} is
done at the generic renormalisation scale $\mu_R$ and the connection
between $m_c(m_c)$ and $m_c(\mu_R)$ is established solving the
appropriate RG equation.
A further complication that arises in FONLL as a VFN scheme is the
fact that the involved running quantities, that is PDFs, $\alpha_s$
and the mass itself, have to be properly matched when crossing a
heavy-quark threshold in their evolution. The matching conditions for
PDFs and $\alpha_s$ are presently known up to
$\mathcal{O}(\alpha_s^2)$~\cite{Larin:1994va} and
$\mathcal{O}(\alpha_s^3)$~\cite{Chetyrkin:1997sg}, respectively, but
those for PDFs are given in terms of the pole mass. In the next
section we will show how to express them in terms of the
$\overline{\rm MS}$ mass up to the relevant accuracy. As far as the
matching of the mass is concerned, the expressions for the matching
conditions are given in Ref.~\cite{Chetyrkin:2000yt} up to
$\mathcal{O}(\alpha_s^3)$ also in terms of $\overline{\rm MS}$ mass.
\subsection{Implementation}\label{sec:implementation}
In this section we will describe in some detail the implementation of
the FONLL scheme in terms of the $\overline{\mbox{MS}}$ heavy-quark
masses in {\tt APFEL}. Starting from the more usual definition of
structure functions in terms of pole masses, our goal is to
consistently replace them with the $\overline{\mbox{MS}}$ mass
definition.
\subsubsection{$\overline{\mbox{MS}}$ mass vs. pole mass}
The (scale independent) pole mass $M$ and the (scale dependent)
$\overline{\mbox{MS}}$ mass $m(\mu)$ arise from two different
renormalization procedures and, as already mentioned, in perturbation
theory they can be expressed one in terms of the other. The relation
connecting pole and $\overline{\mbox{MS}}$ mass definitions has been
computed in Ref.~\cite{Chetyrkin:1999qi} up to four loops. However, in
the following we will only need to go up to one loop and thus we
report here the corresponding relation:
\begin{equation}\label{eq:PoleToMSbar}
\frac{M}{m(\mu)} = 1 + h^{(1)}a_s+\mathcal{O}(a_s^2)\,,
\end{equation}
with:
\begin{equation}\label{eq:CoeffPoleToMSbar}
\begin{array}{l}
\displaystyle h^{(1)}(\mu,m(\mu)) = C_F\left(4 + 3L_{\mu m}\right)\,
\end{array}
\end{equation}
where $C_F=4/3$ is one of the usual QCD color factors. Moreover, we
have defined:
\begin{equation}
a_s\equiv a_s(\mu) = \frac{\alpha_s(\mu)}{4\pi}\,,
\end{equation}
and:
\begin{equation}
L_{\mu m} = \ln\frac{\mu^2}{m^2(\mu)}\,.
\end{equation}
In the following we will use eq. (\ref{eq:PoleToMSbar}) to replace the
pole mass $M$ with the $\overline{\mbox{MS}}$ mass $m(\mu)$.
\subsubsection{Solution of the RGE for the running of the $\overline{\mbox{MS}}$ mass}
In order to evaluate the running of $m(\mu)$ with the renormalization
scale $\mu$ we have to solve the corresponding renormalization-group
equation (RGE):
\begin{equation}\label{eq:massRGE}
\mu^2\frac{dm}{d\mu^2} = m(\mu)\gamma_m(a_s) = -m(\mu)\sum_{n=0}^{\infty}\gamma_m^{(n)}a_s^{n+1}\,,
\end{equation}
whose first three coefficients can be taken from
Ref.~\cite{Chetyrkin:1999pq}\footnote{The following expressions have
been adjusted taking into account our definition of $a_s$ which
differs by a factor of 4 with respect to that of Ref.~\cite{Chetyrkin:1999pq}.}:
\begin{subequations}
\begin{equation}
\gamma_m^{(0)} = 4 \,,
\end{equation}
\begin{equation}
\gamma_m^{(1)} = \frac{202}3 - \frac{20}{9}N_f\,,
\end{equation}
\begin{equation}
\gamma_m^{(2)} = 1249 - \left(\frac{2216}{27}+\frac{160}{3}\zeta_3\right)N_f-\frac{140}{81}N_f^2\,,
\end{equation}
\end{subequations}
where $N_f$ is the number of active flavors. In addition, the RGE for
the running of $\alpha_s$ reads:
\begin{equation}\label{eq:alphasRGE}
\mu^2\frac{da_s}{d\mu^2} = \beta(a_s) = -\sum_{n=0}^{\infty}\beta_n a_s^{n+2}\,,
\end{equation}
with:
\begin{subequations}
\begin{equation}
\beta_0 = 11-\frac23 N_f \,,
\end{equation}
\begin{equation}
\beta_1 = 102 - \frac{38}3 N_f\,.
\end{equation}
\begin{equation}
\beta_2 = \frac{2857}{2} - \frac{5033}{18}N_f + \frac{325}{54}N_f^2\,.
\end{equation}
\end{subequations}
Combining eqs.~(\ref{eq:massRGE}) and~(\ref{eq:alphasRGE}) we obtain the
following differential equation:
\begin{equation}\label{runmass}
\frac{dm}{da_s} = \frac{\gamma_m(a_s)}{\beta(a_s)}m(a_s)\,,
\end{equation}
whose solution is:
\begin{equation}\label{eq:massRGEsolution}
m(\mu) = m(\mu_0)\exp\left[\int_{a_s(\mu_0)}^{a_s(\mu)}\frac{\gamma_m(a_s)}{\beta(a_s)}da_s\right]\,.
\end{equation}
In order to get an analytical expression out of
eq.~(\ref{eq:massRGEsolution}), one can expand the integrand in the
r.h.s. using the perturbative expansions of $\gamma_m(a_s)$ and
$\beta(a_s)$ given in eqs.~(\ref{eq:massRGE})
and~(\ref{eq:alphasRGE}). This allows us to solve the integral
analytically, obtaining:
\begin{equation}\label{eq:massRGEsolutionAnal}
\begin{array}{rcl}
m(\mu)&=&\displaystyle m(\mu_0)\left(\frac{a}{a_0}\right)^{c_0}\\
\\
&\times&\displaystyle
\frac{1+(c_1-b_1c_0)a+\frac12[c_2-c_1b_1-b_2c_0+b_1^2c_0+(c_1-b_1c_0)^2]a^2}{1+(c_1-b_1c_0)a_0+\frac12[c_2-c_1b_1-b_2c_0+b_1^2c_0+(c_1-b_1c_0)^2]a_0^2}\,,
\end{array}
\end{equation}
where we have defined:
\begin{equation}\label{eq:SolCoefs}
b_i = \frac{\beta_i}{\beta_0}\quad\mbox{and}\quad c_i = \frac{\gamma_m^{(i)}}{\beta_0}\,,
\end{equation}
and $a\equiv a_s(\mu)$ and $a_0\equiv a_s(\mu_0)$.
Eq.~(\ref{eq:massRGEsolutionAnal}) represents the NNLO solution of the
RGE for the $\overline{\mbox{MS}}$ mass $m(\mu)$.
Of course, the NLO and the LO solutions can be easily extracted from
eq.~(\ref{eq:massRGEsolutionAnal}) just by disregarding the terms
proportional to $a^2$ and $a_0^2$ for the NLO solution and also the
terms proportional to $a$ and $a_0$ for the LO solution\footnote{In
order to be consistent, the evaluation of $a$ and $a_0$
eq.~(\ref{eq:massRGEsolutionAnal}) must be performed at the same
perturbative order of $m(\mu)$. So, for instance, if one wants to
evaluate the NNLO running of $m(\mu)$ also the value of $a$ and
$a_0$ must be computed using the NNLO running.}.
\subsubsection{Matching conditions}\label{sec:MatchingConditions}
When working in the context of a VFN scheme, all running quantities
are often required to cross heavy-quark thresholds when evolving from
one scale to another. Such a transition in turn requires the matching
different factorization schemes whose content of active flavors
differs by one unit. In other words, if the perturbative evolution
leads from an energy region where (by definition) there are $N_f-1$
active flavors to another region where there are $N_f$ active flavors,
the two regions must be consistently connected and such a connection
can be evaluated perturbatively. This goes under the name of
\textit{matching conditions}.
In general, matching conditions give rise to discontinuities of the
running quantities at the matching scales and in the following we will
report the matching conditions up to NNLO in terms of the
$\overline{\mbox{MS}}$ heavy-quark thresholds for: $\alpha_s(\mu)$,
$m(\mu)$ and PDFs.
\subsubsection*{Matching of $\alpha_s(\mu)$}
The matching conditions for $\alpha_s$ were evaluated in
Ref.~\cite{Chetyrkin:1997sg} to three loops. We report here the
relation up to two loops (again taking into account the factor 4
coming from the different definitions of $a$):
\begin{equation}\label{eq:alphaspole}
\frac{a^{(N_f-1)}(\mu)}{a^{(N_f)}(\mu)}=1-\frac23 L_{\mu M}a^{(N_f)}(\mu)+\left(\frac49L_{\mu M}^2-\frac{38}3L_{\mu M}-\frac{14}3\right)[a^{(N_f)}(\mu)]^2\,.
\end{equation}
$M$ being the pole mass of the $n$-th flavor. From
eq.~(\ref{eq:PoleToMSbar}) we can easily infer that:
\begin{equation}
\ln M^2 = \ln m^2(\mu) + 2\ln[1+h^{(1)}(\mu)a^{(N_f)}(\mu)]= \ln m^2(\mu) + 2h^{(1)}(\mu)a^{(N_f)}(\mu)+\mathcal{O}([a^{(N_f)}]^2)\,.
\end{equation}
Therefore, it is straightforward to see that:
\begin{equation}\label{eq:conversionLog}
L_{\mu M} = L_{\mu m} - 2h^{(1)}a^{(N_f)}=L_{\mu m}-\left(\frac{32}3+8L_{\mu m}\right)a^{(N_f)}\,,
\end{equation}
so that:
\begin{equation}\label{eq:alphasmsbar}
\frac{a^{(N_f-1)}(\mu)}{a^{(N_f)}(\mu)}=1-\frac23 L_{\mu m}a^{(N_f)}(\mu)+\left(\frac49L_{\mu m}^2-\frac{22}3L_{\mu m}+\frac{22}9\right)[a^{(N_f)}(\mu)]^2\,,
\end{equation}
consistently with eq.~(20) of Ref.~\cite{Chetyrkin:2000yt}.
In order to simplify this expression, it is a common procedure to
perform the matching at the point where the logarithms vanish. In this
particular case, choosing $\mu=m(\mu)=m(m)$, we get:
\begin{equation}
a^{(N_f-1)}(m)=a^{(N_f)}(m)\left(1+\frac{22}9[a^{(N_f)}(m)]^2\right)\,,
\end{equation}
which can be easily inverted obtaining:
\begin{equation}\label{eq:alphasatthrs}
a^{(N_f)}(m)=a^{(N_f-1)}(m)\left(1-\frac{22}9[a^{(N_f-1)}(m)]^2\right)\,.
\end{equation}
It is interesting to observe that, in order to perform the matching as
described above, one just needs to know the value of $m(m)$. This is
the so-called RG-invariant $\overline{\mbox{MS}}$ mass.
\subsubsection*{Matching of $m(\mu)$}
The running of the $\overline{\mbox{MS}}$ masses also needs to be
matched at the heavy-quark thresholds. In particular, one needs to
match the $(N_f-1)$- with $(N_f)$-scheme for the mass $m_q(\mu)$, with
$q=c,b,t$, at the threshold $m_h(\mu)$, where $h=c,b,t$. From
Ref.~\cite{Chetyrkin:2000yt} we read:
\begin{equation}\label{eq:mqmc}
\frac{m_q^{(N_f-1)}(\mu)}{m_q^{(N_f)}(\mu)}=1+\left(\frac43L_{\mu m}^{(h)2}-\frac{20}9L_{\mu m}^{(h)}+\frac{89}{27}\right)[a^{(N_f)}(\mu)]^2\,,
\end{equation}
where:
\begin{equation}
L_{\mu m}^{(h)} =\ln\frac{\mu^2}{m_h^2(\mu)}\,.
\end{equation}
Exactly as before, if we choose to match the two schemes at the scale
$\mu=m_h(\mu)=m_h(m_h)$, the logarithmic terms vanish and we are left
with:
\begin{equation}\label{eq:MarchMhUp}
m_q^{(N_f-1)}(m_h)=\left(1+\frac{89}{27}[a^{(N_f)}(m_h)]^2\right)m_q^{(N_f)}(m_h)\,,
\end{equation}
whose inverse is:
\begin{equation}\label{eq:MarchMhDown}
m_q^{(N_f)}(m_h)=\left(1-\frac{89}{27}[a^{(N_f-1)}(m_h)]^2\right)m_q^{(N_f-1)}(m_h)\,.
\end{equation}
\subsubsection*{Matching of PDFs}
To conclude the section on the matching conditions, we finally
consider PDFs. One can write the singlet and the gluon in the
$(N_f)$-scheme in terms of singlet and gluon in the $(N_f-1)$-scheme
at any scale $\mu$ as follows:
\begin{equation}\label{eq:MatchPDFsPole}
\begin{array}{c}
\displaystyle {\Sigma^{(N_f)} \choose g^{(N_f)}}=\begin{pmatrix}1+a_s^2[A_{qq,h}^{N\!S,(2)}+\tilde{A}^{S,(2)}_{hq}] & a_s\tilde{A}^{S,(1)}_{hg}+a_s^2\tilde{A}^{S,(2)}_{hg}\\
a_s^2A^{S,(2)}_{gq,h} & 1+a_sA_{gg,h}^{S,(1)}+a_s^2A_{gg,h}^{S,(2)}\end{pmatrix}{\Sigma^{(N_f-1)} \choose g^{(N_f-1)}}\,,
\end{array}
\end{equation}
where the form of the functions entering the transformation matrix
above are given in Appendix B of Ref.~\cite{Buza:1996wv} in terms of
the pole mass. We omit the matching conditions for the non-singlet PDF
combinations because they have no $\mathcal{O}(a_s)$ correction and
the first correction appears at $\mathcal{O}(a_s^2)$. This leaves the
conversion from the pole to the $\overline{\mbox{MS}}$ mass scheme
unaffected up to NNLO.
In order to replace the pole mass $M$ with the $\overline{\mbox{MS}}$
mass $m(\mu)$, we just have to plug eq.~(\ref{eq:conversionLog}) into
eq.~(\ref{eq:MatchPDFsPole}). In doing so, only the $\mathcal{O}(a_s)$
terms proportional to $\ln(\mu^2/M^2)$ play a role in the conversion
up to NNLO. Since the functions $\tilde{A}^{S,(1)}_{hg}$ and
$A_{gg,h}^{S,(1)}$ can be written as:
\begin{equation}\label{eq:OasMatchPDFs}
\begin{array}{l}
\displaystyle \tilde{A}^{S,(1)}_{hg}\left(x,\frac{\mu^2}{M^2}\right) = f_1(x)\ln\frac{\mu^2}{M^2}\,,\\
\\
\displaystyle A^{S,(1)}_{gg,h}\left(x,\frac{\mu^2}{M^2}\right) = f_2(x)\ln\frac{\mu^2}{M^2}
\end{array}\,,
\end{equation}
where:
\begin{equation}\label{eq:OasPDFsMatchCoeff}
\begin{array}{l}
\displaystyle f_1(x)= 4 T_R[x^2+(1-x)^2]\,,\\
\\
\displaystyle f_2(x)= -\frac43 T_R \delta(1-x)\,,
\end{array}
\end{equation}
replacing $M$ with $m$ in eq.~(\ref{eq:OasMatchPDFs}) using
eq. (\ref{eq:conversionLog}) leads to:
\begin{equation}
\begin{array}{l}
\displaystyle \tilde{A}^{S,(1)}_{hg}\left(x,\frac{\mu^2}{m^2}\right) = f_1(x)\ln\frac{\mu^2}{m^2}-2h^{(1)}(\mu)f_1(x)a_s(\mu)\,,\\
\\
\displaystyle A^{S,(1)}_{gg,h}\left(x,\frac{\mu^2}{m^2}\right) = f_2(x)\ln\frac{\mu^2}{m^2}-2h^{(1)}(\mu)f_2(x)a_s(\mu)\,.
\end{array}
\end{equation}
Therefore eq. (\ref{eq:MatchPDFsPole}) in terms of $m$ becomes:
\begin{equation}
\begin{array}{c}
\displaystyle {\Sigma^{(N_f)}\choose g^{(N_f)}}=\begin{pmatrix}1+a_s^2[A_{qq,h}^{N\!S,(2)}+\tilde{A}^{S,(2)}_{hq}] & a_s\tilde{A}^{S,(1)}_{hg}+a_s^2[\tilde{A}^{S,(2)}_{hg}-2h^{(1)}f_1]\,,\\
a_s^2A^{S,(2)}_{gq,h} & 1+a_sA_{gg,h}^{S,(1)}+a_s^2[A_{gg,h}^{S,(2)}-2h^{(1)}f_2]\end{pmatrix}{\Sigma^{(N_f-1)} \choose g^{(N_f-1)}}\,.
\end{array}
\end{equation}
As usual, we choose to match the $(N_f)$-scheme to the
$(N_f-1)$-scheme at $\mu = m(\mu) = m(m)$ so that all the logarithmic
terms vanish, obtaining:
\begin{equation}
\begin{array}{c}
\displaystyle {\Sigma^{(N_f)}\choose g^{(N_f)}}=\begin{pmatrix}1+a_s^2[A_{qq,h}^{N\!S,(2)}+\tilde{A}^{S,(2)}_{hq}] & a_s^2[\tilde{A}^{S,(2)}_{hg}-2h^{(1)}f_1]\\
a_s^2A^{S,(2)}_{gq,h} & 1+a_s^2[A_{gg,h}^{S,(2)}-2h^{(1)}f_2]\end{pmatrix}{\Sigma^{(N_f-1)} \choose g^{(N_f-1)}}\,.
\end{array}
\end{equation}
\subsubsection*{Renormalization scale variation}
The scale $\mu$ that appears in $a_s$ and $m_q$ is the
\textit{renormalization} scale, which we will now denote as $\mu_R$.
The scale that explicitly appears in the PDFs is instead the
\textit{factorization} scale, which we will now denote with
$\mu_F$. In principle, renormalization and factorization scales are
different but one usually takes them to be proportional to each other,
as $\mu_R = \kappa \mu_F$, where $\kappa$ can be any real
number\footnote{It should be noticed that in the case $\kappa\neq1$
PDFs acquire an implicit dependence on $\mu_R$ that comes from a
redefinition of the splitting functions that in turn derives from
the expansion of $\alpha_s(\mu_R)$ around $\mu_R=\mu_F$.}.
The most common choice when matching the $(N_f-1)$-scheme to the
$(N_f)$-scheme is to set $\mu_F$ equal to heavy-quark thresholds
($M_c$, $M_b$ and $M_t$ in the pole-mass scheme and $m_c(m_c)$,
$m_b(m_b)$ and $m_t(m_t)$ in the $\overline{\mbox{MS}}$ scheme). In
doing so, the logarithmic terms in the PDF matching conditions are
assured to vanish. However, if $\kappa$ is different from one, the
logarithmic terms in the matching conditions for $a_s(\mu_R)$ and
$m_q(\mu_R)$ do not vanish anymore. In the following we will show how
the matching conditions for $a_s$ and $m_q$ change for $\kappa\neq1$.
Let us start with $\alpha_s$. Inverting eq.~(\ref{eq:alphasmsbar}) we
obtain:
\begin{equation}
\frac{a^{(N_f)}(\mu_R)}{a^{(N_f-1)}(\mu_R)} = 1 + c_1a^{(N_f-1)}(\mu_R) + c_2 [a^{(N_f-1)}(\mu_R)]^2\,,
\end{equation}
where:
\begin{equation}
c_1 = \frac23 L_{\mu m}
\quad\mbox{and}\quad
c_2 = \frac49L_{\mu m}^2+\frac{22}3L_{\mu m}-\frac{22}9\,.
\end{equation}
Setting $\mu_F=\kappa \mu_F$, we have that:
\begin{equation}
L_{\mu m} = \ln\frac{\mu_R}{m(\mu_R)}=\ln\frac{\kappa\mu_F}{m(\kappa \mu_F)}\,.
\end{equation}
As usual, the matching scale is chosen to be $\mu_F = m(m)$, so that:
\begin{equation}
L_{\mu m} \rightarrow \ln\kappa + \ln\frac{m(m)}{m(\kappa m)}\,.
\end{equation}
But using eq.~(\ref{eq:massRGEsolution}), it is easy to see that:
\begin{equation}
\ln\frac{m(m)}{m(\kappa m)}=a_s(\kappa m)\gamma_m^{(0)}\ln\kappa+\mathcal{O}[a_s^2(\kappa m)]\,,
\end{equation}
so that:
\begin{equation}\label{eq:LmumExp}
L_{\mu m} \rightarrow [1+\gamma_m^{(0)}a_s(\kappa m)]\ln\kappa\,.
\end{equation}
It should be noticed that in the eq.~(\ref{eq:LmumExp}), since
$a_s^{(N_f-1)}=a_s^{(N_f)}+\mathcal{O}([a_s^{(N_f)}]^2)$, it does not
matter whether one uses $a_s^{(N_f)}(\kappa m)$ or
$a_s^{(N_f-1)}(\kappa m)$ because the difference would be subleading
up to NNLO.
Therefore, setting $\mu=\mu_R=\kappa m(m) = \kappa m$ in
eq.~(\ref{eq:alphasmsbar}) and using eq. (\ref{eq:LmumExp}), one gets:
\begin{equation}
\begin{array}{rcl}
a^{(N_f-1)}(\kappa m)&=&\displaystyle a^{(N_f)}(\kappa m)\bigg\{1-\frac23
\ln\kappa\,a^{(N_f)}(\kappa
m)\\
\\
&+&\displaystyle \left[\frac49\ln^2\kappa-\frac{2}3\left(\gamma_m^{(0)}+11\right)\ln\kappa+\frac{22}9\right][a^{(N_f)}(\kappa
m)]^2\bigg\}\,,
\end{array}
\end{equation}
whose inverse is:
\begin{equation}
\begin{array}{rcl}
a^{(N_f)}(\kappa m)&=&\displaystyle a^{(N_f-1)}(\kappa m)\bigg\{1+\frac23
\ln\kappa\,a^{(N_f-1)}(\kappa
m)\\
\\
&+&\displaystyle \left[\frac49\ln^2\kappa+\frac{2}3\left(\gamma_m^{(0)}+11\right)\ln\kappa-\frac{22}9\right][a^{(N_f-1)}(\kappa
m)]^2\bigg\}\,.
\end{array}
\end{equation}
Now let us turn to $m_q$. In this case there is not much to do. In
fact, for an arbitrary matching point the matching condition of the
$\overline{\mbox{MS}}$ mass starts at $\mathcal{O}(\alpha_s^2)$
(cfr. eq. (\ref{eq:mqmc})), therefore writing $L_{\mu m}$ in terms of
$\ln\kappa$ would give rise to subleading terms up to NNLO (see
eq. (\ref{eq:LmumExp})). As a consequence, we have that:
\begin{equation}
m_q^{(N_f-1)}(\kappa m_h)=\left[1+\left(\frac43\ln^2\kappa-\frac{20}9\ln\kappa+\frac{89}{27}\right)[a^{(N_f)}(\kappa m_h)]^2\right]m_q^{(N_f)}(\kappa m_h)\,,
\end{equation}
whose inverse is:
\begin{equation}
m_q^{(N_f)}(\kappa
m_h)=\left[1-\left(\frac43\ln^2\kappa-\frac{20}9\ln\kappa+\frac{89}{27}\right)[a^{(N_f-1)}(\kappa
m_h)]^2\right]m_q^{(N_f-1)}(\kappa m_h)\,.
\end{equation}
\subsubsection{Structure functions}
We finally turn to discuss how the DIS massive structure functions
change when expressing them in terms of the $\overline{\mbox{MS}}$
masses. We will first consider the neutral-current (NC) massive
structure functions up to $\mathcal{O}(\alpha_s^2)$, which is the
highest perturbative order at which corrections are known exactly, and
then we will consider the charged-current (CC) massive structure
functions again up to the highest perturbative order exactly
known\footnote{In a recent publication~\cite{Berger:2016inr} the
$\mathcal{O}(\alpha_s^2)$ corrections (NNLO) to charm production in
CC DIS were presented. However, no analytical expression was
provided.}, that is $\mathcal{O}(\alpha_s)$. In order to shorten the
notation, we will adopt the following definitions:
$$
M =\;\mbox{pole mass},\quad m\equiv m(\mu) =\;\overline{\mbox{MS}}\mbox{ mass},\quad a_s\equiv a_s(\mu),\quad h^{(l)}\equiv h^{(l)}(\mu,m(\mu))\,.
$$
\subsubsection*{Neutral current}
Dropping all the unnecessary dependences, the NC massive structure
functions up to $\mathcal{O}(a_s^2)$ have the form:
\begin{equation}
F = a_sF^{(0)}(M) + a_s^2F^{(1)}(M) + \mathcal{O}(a_s^3)\,.
\end{equation}
The goal is to replace explicitly the pole mass $M$ with the
$\overline{\mbox{MS}}$ mass $m$ using eq.~(\ref{eq:PoleToMSbar}). To
this end, following the procedure adopted in
Ref.~\cite{Alekhin:2010sv}, we expand $F^{(0)}(M)$ and $F^{(1)}(M)$
around $M=m$:
\begin{equation}
F^{(l)}(M) = \sum_{n=0}^{\infty}\frac1{n!}\frac{d^n F^{(l)}}{dM^n}\bigg|_{M=m}(M-m)^n\,,
\end{equation}
so that, up to $\mathcal{O}(a_s^2)$, what we need is:
\begin{equation}
\begin{array}{l}
\displaystyle F^{(0)}(M) = F^{(0)}(m) + a_smh^{(1)}\frac{dF^{(0)}}{dM}\bigg|_{M=m}\,,\\
\\
\displaystyle F^{(1)}(M) = F^{(1)}(m)\,.
\end{array}
\end{equation}
Finally, we have that:
\begin{equation}\label{eq:changescheme}
F = a_sF^{(0)}(m) + a_s^2\left[F^{(1)}(m)+mh^{(1)}\frac{dF^{(0)}}{dM}\bigg|_{M=m}\right]\,.
\end{equation}
We now need to evaluate explicitly the derivative in
eq. (\ref{eq:changescheme}). First of all we observe that:
\begin{equation}\label{eq:convolution0}
F^{(0)}(M) = x\int_x^{x_{\mbox{\tiny max}}(M)} \frac{dz}{z}g\left(\frac{x}{z}\right)C_g^{(0)}(\eta(z,M),\xi(M),\chi(M))\,,
\end{equation}
where $g$ is the gluon distribution and we have used the following
definitions:
\begin{equation}
x_{\mbox{\tiny max}}(M)=\frac1{1+\frac{4M^2}{Q^2}},\quad\eta(z,M) = \frac{Q^2}{4M^2}\left(\frac1z-1\right)-1,\quad \xi(M) =\frac{Q^2}{M^2},\quad \chi(M) =\frac{\mu^2}{M^2}\,.
\end{equation}
Defining:
\begin{equation}
G(z,M)=\frac{x}{z}g\left(\frac{x}{z}\right)C_g^{(0)}(\eta(z,M),\xi(M),\chi(M))\,,
\end{equation}
the derivative of eq.~(\ref{eq:convolution0}) can be written as:
\begin{equation}
\frac{dF^{(0)}}{dM} = \frac{d}{dM}\int_x^{x_{\mbox{\tiny max}}(M)} dzG(z,M) = \frac{d\widetilde{G}(x_{\mbox{\tiny max}}(M),M)}{dM}-\frac{d\widetilde{G}(x,M)}{dM}\,,
\end{equation}
where $\widetilde{G}(z,M)$ is the primitive of $G(z,M)$ with respect to $z$ (i.e. $\partial\widetilde{G}/\partial z = G$). But:
\begin{equation}
\frac{d\widetilde{G}(x_{\mbox{\tiny max}}(M),M)}{dM} = \frac{d \widetilde{G}(x_{\mbox{\tiny max}},M)}{d M}+\frac{dx_{\mbox{\tiny max}}}{dM}G(x_{\mbox{\tiny max}},M)\,,
\end{equation}
thus:
\begin{equation}\label{eq:withbound}
\begin{array}{c}
\displaystyle \frac{dF^{(0)}}{dM} = \frac{\partial \widetilde{G}(x_{\mbox{\tiny max}},M)}{\partial M}-\frac{\partial\widetilde{G}(x,M)}{\partial M}+\frac{dx_{\mbox{\tiny max}}}{dM}G(x_{\mbox{\tiny max}},M) =\\
\\
\displaystyle \int_x^{x_{\mbox{\tiny max}}(M)} dz\frac{\partial G(z,M)}{\partial M}+\frac{dx_{\mbox{\tiny max}}}{dM}G(x_{\mbox{\tiny max}},M)\,.
\end{array}
\end{equation}
It can be shown that the boundary term in eq.~(\ref{eq:withbound})
vanishes (see Ref.~\cite{Alekhin:2010sv}), thus it can be omitted.
Gathering all pieces and taking into account that:
\begin{equation}
\frac{\partial G(z,M)}{\partial M} = \frac{x}{z}g\left(\frac{x}{z}\right)\frac{\partial C_g^{(0)}}{\partial M}\,,
\end{equation}
we have that:
\begin{equation}\label{eq:withoutbound}
\begin{array}{rcl}
\displaystyle \frac{dF^{(0)}}{dM}\bigg|_{M=m}&=&\displaystyle
\left[x\int_x^{x_{\mbox{\tiny
max}}(M)}\frac{dz}{z}g\left(\frac{x}{z}\right)\frac{\partial
C_g^{(0)}}{\partial
M}\right]\Bigg|_{M=m}\\
\\
&=&\displaystyle x\int_x^{x_{\mbox{\tiny
max}}(m)}\frac{dz}{z}g\left(\frac{x}{z}\right)\left[\frac{\partial
C_g^{(0)}}{\partial M}\right]\Bigg|_{M=m}\,.
\end{array}
\end{equation}
Finally, considering that:
\begin{equation}
F^{(1)}(M) = \sum_{i=q,\overline{q},g}x\int_x^{x_{\mbox{\tiny max}}(M)} \frac{dz}{z}q_i\left(\frac{x}{z}\right)C_i^{(1)}(z,M)
\end{equation}
and using eqs.~(\ref{eq:changescheme}) and~(\ref{eq:withoutbound}),
one can explicitly write down the full structure of the massive
structure functions ($F_2$ and $F_L$) in terms of
$\overline{\mbox{MS}}$ masses up to $\mathcal{O}(\alpha_s^2)$ as
follows:
\begin{equation}\label{eq:masterNC}
\begin{array}{c}
\displaystyle F = x\int_x^{x_{\mbox{\tiny max}}(m)} \frac{dz}{z}g\left(\frac{x}{z}\right)\left[a_sC_g^{(0)}(z,m)+a_s^2\left(C_g^{(1)}(z,m)+mh^{(1)}\left[\frac{\partial C_g^{(0)}}{\partial M}\right]\Bigg|_{M=m}\right)\right]+\\
\\
\displaystyle \sum_{i=q,\overline{q}}x\int_x^{x_{\mbox{\tiny max}}(M)} \frac{dz}{z}q_i\left(\frac{x}{z}\right)a_s^2C_i^{(1)}(z,M)\,.
\end{array}
\end{equation}
In order to carry out the implementation, we need to evaluate
explicitly the derivative of $C_g^{(0)}$ in eq. (\ref{eq:masterNC})
and this must be done separately for $F_2$ and $F_L$.
We consider $F_2$ first. The explicit expression of $C_{2,g}^{(0)}$ is
the following:
\begin{equation}
\begin{array}{rl}
\displaystyle C_{2,g}^{(0)}(z,Q^2,M^2)=&\displaystyle T_R\Big\{2(1-6\epsilon-4\epsilon^2)I_2(\epsilon,z)-2(1-2\epsilon)I_1(\epsilon,z)+I_0(\epsilon,z)+\\
\\
& \displaystyle -4(2-\epsilon)J_2(\epsilon,z)+4(2-\epsilon)J_1(\epsilon,z)-J_0(\epsilon,z)\Big\}\,,
\end{array}
\end{equation}
where:
\begin{equation}\label{eq:iq}
I_q(\epsilon,z) = z^q\ln\left(\frac{1+v}{1-v}\right)\,.
\end{equation}
\begin{equation}\label{eq:jq}
J_q(\epsilon,x) = z^q v\,,
\end{equation}
with:
\begin{equation}\label{eq:definitions1}
\epsilon = \frac{M^2}{Q^2}\,,\quad
a=\frac1{1+4\epsilon}\quad\mbox{and}\quad v=\sqrt{1-4\epsilon\frac{z}{1-z}}\,.
\end{equation}
From the definitions in eq.~(\ref{eq:definitions1}), we obtain:
\begin{equation}\label{eq:derivatives}
\begin{array}{rl}
\displaystyle \frac{\partial}{\partial M} &\displaystyle = \frac{\partial \epsilon}{\partial M} \frac{\partial}{\partial \epsilon} = \frac{2\epsilon}{M}\frac{\partial}{\partial \epsilon}\,,\\
\\
\displaystyle \frac{\partial}{\partial M} &\displaystyle = \frac{\partial \epsilon}{\partial M} \frac{\partial v}{\partial \epsilon} \frac{\partial}{\partial v} = -\frac{1-v^2}{Mv}\frac{\partial}{\partial v}\,.
\end{array}
\end{equation}
Therefore:
\begin{equation}
\begin{array}{rl}
\displaystyle \frac{\partial C_{2,g}^{(0)}}{\partial M}=&\displaystyle \frac{1}{M}T_R\Bigg\{2\epsilon\Big[2(-6-8\epsilon)I_2+4I_1+4J_2-4J_1\Big]\\
\\
&\displaystyle -\frac{1-v^2}{v}\Bigg[2(1-6\epsilon-4\epsilon^2)\frac{\partial I_2}{\partial v}-2(1-2\epsilon)\frac{\partial I_1}{\partial v}+\frac{\partial I_0}{\partial v}\\
\\
& \displaystyle -4(2-\epsilon)\frac{\partial J_2}{\partial v}+4(2-\epsilon)\frac{\partial J_1}{\partial v}-\frac{\partial J_0}{\partial v}\Bigg]\Bigg\}\,.
\end{array}
\end{equation}
To find the explicit expression, we just need to evaluate the
derivative of $I_q$ and $J_q$ starting from eqs. (\ref{eq:iq}) and
(\ref{eq:jq}) which is easily done:
\begin{equation}
\begin{array}{rcl}
\displaystyle \frac{\partial I_q}{\partial v} &=& \displaystyle\frac{2 z^q}{1-v^2}\,,\\
\\
\displaystyle \frac{\partial J_q}{\partial v} &=& \displaystyle z^q\,.
\end{array}
\end{equation}
In the end we get:
\begin{equation}\label{eq:dC2g0}
\begin{array}{rl}
\displaystyle \frac{\partial C_{2,g}^{(0)}}{\partial M}=&\displaystyle
\frac{1}{M}T_R\Bigg\{4\epsilon\left[(-6-8\epsilon)z^2+2z\right]\ln\left(\frac{1+v}{1-v}\right)+8\epsilon
z(z-1)v\\
\\
&\displaystyle -\frac{2}{v}\left[2(1-6\epsilon-4\epsilon^2)z^2-2(1-2\epsilon)z+1\right]\\
\\
& \displaystyle -\frac{1-v^2}{v}\left[-4(2-\epsilon)z^2+4(2-\epsilon)z-1\right]\Bigg\}\,.
\end{array}
\end{equation}
The implementation of the FONLL scheme given in
eq.~(\ref{eq:FONLLdef}) requires the massless limit of the massive
structure functions. In practice this means that one needs to compute
the limit $M\rightarrow 0$ of the massive coefficient functions
retaining the logarithmic enhanced terms. In order to apply this
recipe to eq.~(\ref{eq:dC2g0}), we observe that:
\begin{equation}
\epsilon \mathop{\longrightarrow}_{M\rightarrow 0} 0 \,,\quad v \mathop{\longrightarrow}_{M\rightarrow 0} 1\,,
\end{equation}
and that:
\begin{equation}
\ln\left(\frac{1+v}{1-v}\right) \mathop{\longrightarrow}_{M\rightarrow
0} \ln\frac{Q^2(1-z)}{M^2z}\,,
\end{equation}
so that:
\begin{equation}\label{eq:dC2g00}
\displaystyle \frac{\partial C_{2,g}^{(0)}}{\partial M}\mathop{\longrightarrow}_{M\rightarrow 0}\displaystyle \frac{\partial C_{2,g}^{0,(0)}}{\partial M}=
-\frac{2}{M}T_R\left(2z^2-2z+1\right)\,.
\end{equation}
We now turn to consider $F_L$. In this case the the gluon coefficient
function takes the simpler form:
\begin{equation}
C_{L,g}^{(0)}\left(z,Q^2,M^2\right)= T_R\left[-8\epsilon I_2(\epsilon,z)-4J_2(\epsilon,z)+4J_1(\epsilon,z)\right]\,.
\end{equation}
Therefore, using eq. (\ref{eq:derivatives}), we immediately get:
\begin{equation}
\frac{\partial C_{L,g}^{(0)}}{\partial M} = \frac1{M}T_R\left[-16\epsilon
z^2\ln\left(\frac{1+v}{1-v}\right) +\frac{8\epsilon z^2}{v}-\frac{1-v^2}{v}\left(-4z^2+4z\right)\right]\,.
\end{equation}
It is finally easy to realize that:
\begin{equation}
\frac{\partial C_{L,g}^{(0)}}{\partial
M}\mathop{\longrightarrow}_{M\rightarrow 0} \frac{\partial C_{L,g}^{0,(0)}}{\partial M}=0\,.
\end{equation}
\subsubsection*{Charged current}
In this section we consider the CC massive structure functions. The
treatment follows the exact same steps as the NC structure functions,
with the only difference being that in the CC case the first
non-vanishing term is $\mathcal{O}(a_s^0)$. This means that,
truncating the perturbative expansion at $\mathcal{O}(a_s)$, we have:
\begin{equation}
F_k = F_k^{(0)}(M) + a_sF_k^{(1)}(M) + \mathcal{O}(a_s^2)\,,
\end{equation}
with $k=2,3,L$. Therefore, expanding $F^{(0)}$ and $F^{(1)}$ around $M=m$ and keeping only the terms up to $\mathcal{O}(a_s)$, one obtains:
\begin{equation}\label{eq:expCC}
F_k = F_k^{(0)}(m) + a_s\left[F_k^{(1)}(m)+mh^{(1)}\frac{dF_k^{(0)}}{dM}\bigg|_{M=m}\right]\,.
\end{equation}
The leading-order contribution can be written as follows:
\begin{equation}
F^{(0)}_k(M) = b_k(M)s'(\xi(M))\,,
\end{equation}
where:
\begin{equation}\label{eq:definitions}
\xi = x\underbrace{\left(1+\frac{M^2}{Q^2}\right)}_{\frac1\lambda}=\frac{x}\lambda\quad\mbox{and}\quad
\left\{\begin{array}{l}
b_2 = \xi\\
b_3 = 1\\
b_L = (1-\lambda)\xi
\end{array}
\right.\,,
\end{equation}
where we have also defined:
\begin{equation}
s' = 2|V_{cs}|^2s+2|V_{cd}|^2d\,.
\end{equation}
Therefore:
\begin{equation}\label{eq:derivCC}
\begin{array}{rcl}
\displaystyle mh^{(1)}\frac{dF^{(0)}_k}{dM}\bigg|_{M=m} &=&\displaystyle
mh^{(1)}\frac{d\xi}{dM}\frac{dF^{(0)}_k}{d\xi}\bigg|_{M=m} \\
\\
&=&\displaystyle
2h^{(1)}(1-\lambda)\xi\left[\frac{db_k}{d\xi}s'(\xi)+b_k(\xi)\frac{ds'}{d\xi}\right]\bigg|_{M=m}\,,
\end{array}
\end{equation}
that can be conveniently rewritten as:
\begin{equation}
mh^{(1)}\frac{dF^{(0)}_k}{dM}\bigg|_{M=m} = 2h^{(1)}(1-\lambda)\left[\left(\frac{db_k}{d\xi}-\frac{b_k}{\xi}\right)+b_k(\xi)\frac{d}{d\xi}\right]\xi s'(\xi)\bigg|_{M=m}\,,
\end{equation}
so that, using eq.~(\ref{eq:definitions}), we have that:
\begin{equation}\label{eq:CCcorrections}
\begin{array}{rcl}
\displaystyle mh^{(1)}\frac{dF^{(0)}_2}{dM}\bigg|_{M=m} &=& \displaystyle
2h^{(1)}(1-\lambda)\xi\frac{d}{d\xi} \xi s'(\xi)\bigg|_{M=m}\,,\\
\\
\displaystyle mh^{(1)}\frac{dF^{(0)}_3}{dM}\bigg|_{M=m} &=& \displaystyle
2h^{(1)}(1-\lambda)\frac{1}{\xi}\left[ \xi\frac{d}{d\xi}-1\right]\xi
s'(\xi)\bigg|_{M=m}\,,\\
\\
\displaystyle mh^{(1)}\frac{dF^{(0)}_L}{dM}\bigg|_{M=m} &=& \displaystyle
2h^{(1)}(1-\lambda)^2\xi \frac{d}{d\xi}\xi s'(\xi)\bigg|_{M=m}\,.
\end{array}
\end{equation}
Finally, we notice that in the massless limit, where
$\lambda\rightarrow 1$, all expressions in
eq.~(\ref{eq:CCcorrections}) vanish, with the consequence that the CC
massive structure functions up to $\mathcal{O}(a_s)$ in terms of the
pole mass $M$ or the $\overline{\mbox{MS}}$ mass $m$ are exactly the
same.
\subsection{Benchmark}
In order to validate the implementation in {\tt APFEL}, we have
benchmarked it against public codes. To the best of our knowledge,
there exist no public codes able to compute structure functions in the
FONLL scheme with $\overline{\rm MS}$ masses. For this reason the best
we could do is to benchmark the various ingredients separately.
As a first step, we present the benchmark of the running of PDFs,
$\alpha_s$ and $m_c$\footnote{The running of $m_b$ and $m_t$ has also
been checked finding the same lavel of accuracy found for $m_c$.} in
the VFN scheme with $\overline{\rm MS}$ heavy-quark thresholds. The
difference with respect to the more common pole-mass formulation
arises from the fact that the matching of the evolutions at the
heavy-quark thresholds needs to be adapted to take into account the
different scheme used to renormalize the masses. The full set of such
matching conditions for PDFs, $\alpha_s$ and $m_c$ has been collected
in Sect.~\ref{sec:implementation}.
We start with the DGLAP PDF evolution in the VFN scheme with
$\overline{\rm MS}$ heavy-quark thresholds. A careful benchmark was
already presented in the original {\tt APFEL} publication. In
particular, the {\tt APFEL} evolution has been checked against the
{\tt HOPPET} code~\cite{Salam:2008qg} v1.1.5, finding a very good
agreement at the $\mathcal{O}\left(10^{-4}\right)$ level or
better. Since then, {\tt APFEL} has undergone several changes and
improvements and thus we repeated the benchmark using the same
settings and finding the same level of agreement with {\tt HOPPET}, as
shown in Fig.~\ref{fig:APFELvsHOPPET} for a representative set of
combinations of PDFs\footnote{We observe that, thanks to a better
interpolation strategy, the predictions at the transition regions
between internal $x$-space subgrids is now smoother.}.
\begin{figure}
\begin{center}
\includegraphics[width=0.65\textwidth]{APFEL_vs_HOPPET_MSbar.pdf}
\end{center}
\caption{\label{fig:APFELvsHOPPET} Comparison between {\tt APFEL}
v2.7.0 and {\tt HOPPET} v1.1.5 for the VFNS DGLAP evolution at
NNLO with $\overline{\rm MS}$ heavy-quark thresholds. The
evolution settings, $i.e.$ initial scale PDFs, reference value of
$\alpha_s$, and heavy-quark thresholds, are the same as used in
the Les Houches PDF evolution benchmark~\cite{Dittmar:2005ed}. The
upper inset shows the gluon PDF $xg$, the valence up and down PDFs
$xu_v \equiv xu - x\overline{u}$ and
$xd_v \equiv xd - x\overline{d}$, respectively, and the total
strangeness $xs^+ \equiv xs + x\overline{s}$ at $\mu_F = 100$ GeV
as functions of the Bjorken variable $x$ as returned by {\tt
APFEL}. In the lower inset the ratio to {\tt HOPPET} is
displayed showing a relative difference of $10^{-4}$ or better all
over the considered range.}
\end{figure}
Although the benchmark of the DGLAP evolution already provides an
indirect check of the evolution of $\alpha_s$, we have also performed
a direct check of the VFNS evolution with $\overline{\rm MS}$
heavy-quark thresholds of $\alpha_s$ along with the evolution of the
$\overline{\rm MS}$ charm mass. To this end, we have used the {\tt
CRunDec} code~\cite{Schmidt:2012az}, which is the {\tt C++} version
of the {\tt Mathematica} package {\tt
RunDec}~\cite{Chetyrkin:2000yt}. In Fig.~\ref{fig:APFELvsRunDec} we
show the comparison between {\tt APFEL} and {\tt CRunDec} for the
three-loop evolution (NNLO) of the strong coupling $\alpha_s$ (left
plot) and the charm mass $m_c$ (right plot). As is clear from the
lower insets, the agreement between the two codes is excellent. Also
the one- and two-loop evolutions have been checked finding the same
level of agreement.
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{APFEL_vs_RunDec_alphas.pdf}
\includegraphics[width=0.49\textwidth]{APFEL_vs_RunDec_mc.pdf}
\end{center}
\caption{\label{fig:APFELvsRunDec} Comparison between {\tt APFEL}
v2.7.0 and {\tt CRunDec} v1.1 for the VFNS RG three-loop evolution
with $\overline{\rm MS}$ heavy-quark thresholds of the strong
coupling $\alpha_s$ (left plots) and the $\overline{\rm MS}$ charm
mass $m_c$ (right plot). The evolution settings are:
$\alpha_s^{(n_f=3)}(\sqrt{2}\mbox{ GeV}) = 0.35$,
$m_c^{(n_f=4)}(m_c) = \sqrt{2}\mbox{ GeV}$, and
$m_b^{(n_f=5)}(m_b) = 4.5\mbox{ GeV}$. The upper insets show the
strong coupling $\alpha_s$ (left) and the charm mass $m_c$ (right)
as functions of the renormalization scale $\mu_R$ as returned by
{\tt APFEL}. In the lower insets the ratios to {\tt CRunDec} are
displayed showing a relative difference well below $10^{-6}$ over
the complete range considered.}
\end{figure}
Finally, we benchmarked the implementation of massive DIS structure
functions ($i.e.$ $F^{(3)}$ in eq.~\ref{eq:FONLLdef}) with
$\overline{\rm MS}$ masses against the public code {\tt OPENQCDRAD}
v1.6~{\cite{openqcdrad}}. {\tt OPENQCDRAD} implements DIS structure
functions in terms of the $\overline{\rm MS}$ heavy-quark masses
following the formalism discussed in
Ref.~\cite{Alekhin:2010sv}. However, as already mentioned above, such
a procedure does not directly correspond to what is needed for the
implementation of the FONLL scheme. In order to make the comparison
with {\tt OPENQCDRAD} possible, we have implemented in {\tt APFEL} a
variant of the FONLL scheme with $\overline{\rm MS}$ masses where, as
done in {\tt OPENQCDRAD} , the RG running of the heavy-quark masses is
expanded and truncated to the appropriate order. In
Fig.~\ref{fig:APFELvsOpenQCDrad} we show the comparison between {\tt
APFEL} and {\tt OPENQCDRAD} for the exclusive charm neutral-current
structure functions $F_2^c$ (left plot) and $F_L^c$ (right plot) at
$\mathcal{O}(\alpha_s^2)$ for three different values of $Q^2$ and over
a wide range of $x$. As is clear from the lower ratio plots, the
agreement is typically at the per-mil level except in the very
large-$x$ region where, due to the smallness of the predictions, the
relative difference tends to increase but maintains a good level of
absolute accuracy.
\begin{figure}
\begin{center}
\includegraphics[width=0.49\textwidth]{APFEL_vs_openQCDrad_MSbar_F2c.pdf}
\includegraphics[width=0.49\textwidth]{APFEL_vs_openQCDrad_MSbar_FLc.pdf}
\end{center}
\caption{\label{fig:APFELvsOpenQCDrad} Comparison between {\tt
APFEL} v2.7.0 and {\tt OPENQCDRAD} v1.6 for the neutral-currents
massive charm structure functions with $\overline{\rm MS}$
heavy-quark masses at $\mathcal{O}(\alpha_s^2)$. As an input PDF
set we have used {\tt MSTW2008nlo68cl\_nf3}~\cite{Martin:2010db}
from which also the numerical values of $\alpha_s$ and $m_c$ are
taken. The upper insets show $F_2^c$ (left) and $F_L^c$ (right) as
functions of $x$ for $Q^2 = 10, 100, 1000$ GeV$^2$ as returned by
{\tt APFEL}. In the lower insets the ratios to {\tt OPENQCDRAD}
are displayed showing a relative difference at the per-mil level
except in the very large-$x$ region where, due to the smallness of
the predictions, the relative differences tend to increase but
maintain a good level of absolute accuracy.}
\end{figure}
To conclude this section, we observe that, referring to
eq.~(\ref{eq:FONLLdef}), the introduction of the $\overline{\rm MS}$
masses does not affect the four-flavor structure function
$F^{(4)}$. The structure function $F^{(3,0)}$ is instead affected by
the transition from pole to $\overline{\rm MS}$ masses. Since we are
not aware of any public code that computes such structure functions, a
direct bechmark has not been possible. However, as a sanity check we
have checked that $F^{(3,0)}$ and $F^{(3)}$ for large values of $Q^2$
tend to the same value, as the definition of $F^{(3,0)}$ requires.
\section{Introduction}\label{sec:intro}
The masses of the heavy quarks, charm, bottom and top, are fundamental
parameters of the Standard Model~\cite{Moch:2014tta}.
A precise determination of their values is of utmost importance; as an
example, the fate of the electroweak vacuum depends crucially on the
exact value of $m_t$~\cite{Degrassi:2012ry}.
In the case of the charm quark, since its mass $m_c$ is larger than
the scale $\Lambda_{\rm QCD}$ of Quantum Chromodynamics (QCD), its
value is a direct input of many perturbative calculations involving
charm quarks in the initial and/or in the final state.
Differences in the value of the charm quark mass and in the
treatment of its effects in deep-inelastic-scattering structure functions can
lead to differences in modern analyses of parton
distribution functions
(PDFs)~\cite{Ball:2014uwa,Harland-Lang:2014zoa,Dulat:2015mca,Alekhin:2013nda,Abramowicz:2015mha},
with implications for precision phenomenology at the Large
Hadron Collider (LHC).
As a consequence, a high-precision determination of the charm quark
mass is of interest both in principle, as a fundamental test of the
Standard Model and a measurement of one of its fundamental parameters,
and in practice, as input for LHC calculations.
The current global-average value of the charm mass in the
$\overline{\rm MS}$ renormalization scheme is
$m_c(\mu_R = m_c)=1.275 \pm 0.025$ GeV~\cite{Agashe:2014kda}, where
the result is dominated by high-precision data from charm production
in $e^+e^-$ collisions.
It is therefore interesting to provide alternative determinations of
the charm mass from other processes, both to test the robustness of
the global average and to attempt to further reduce the present
uncertainty.
A process directly sensitive to the charm mass is open-charm
production in lepton-proton deep-inelastic scattering (DIS).
This process has been measured with high accuracy at the HERA collider and the
results of different measurements implying various charm-tagging
techniques are combined~\cite{Abramowicz:1900rp}. The charm
contribution to the inclusive structure functions can be determined
through the measurement of the charm-pair production cross section.
In addition, the final combination of inclusive measurements from Runs
I and II at HERA has been recently presented
in~\cite{Abramowicz:2015mha}.
DIS structure functions can be described using a variety of
theoretical schemes, including the fixed-flavor number (FFN) scheme,
where charm mass effects are included to a fixed perturbative order,
the zero-mass variable-flavor number (ZM-VFN) scheme that neglects
power-suppressed terms in the charm mass but resums to all orders
large collinear logarithms, and the so-called matched general-mass
variable-flavor-number (GM-VFN) schemes, which interpolate smoothly
between the two regimes. A recent discussion and summary of the
application of these schemes to heavy-flavor data at HERA can be
found $e.g.$ in \cite{Behnke:2015qja}.
Examples of matched general-mass schemes in electro-, photo- and hadroproduction
include FONLL~\cite{Forte:2010ta,Ball:2013gsa,Cacciari:1998it},
TR~\cite{thornehq,Thorne:,Thorne:2014toa}, ACOT~\cite{Guzzi:2011ew},
and a scheme generically referred to as
GMVFNS~\cite{Kramer:2003jw,Kniehl:2009mh,Kneesch:2007ey,Kniehl:2004fy,Kniehl:2012ti}.
In this work we will mostly concentrate on the FONLL scheme and on
its implications for the determination of the charm mass. For
the sake of comparison with previous studies~\cite{Alekhin:2012un,Abramowicz:1900rp,Alekhin:2012vu,Alekhin:2013qqa}, a determination
of the charm mass in the FFN scheme at NLO is also performed.
The original formulation of the FONLL general-mass scheme for DIS
structure functions was derived in the pole (on-shell)
heavy quark scheme~\cite{Forte:2010ta}.
In Ref.~\cite{Alekhin:2010sv} it was shown how DIS structure functions
in the FFN scheme can be modified to include $\overline{\rm MS}$
heavy-quark masses.
The same scheme conversion can be applied to any GM-VFN scheme,
and in this work we provide the relevant expressions for FONLL structure
functions with $\overline{\rm MS}$ running masses.
The main advantage of the use of $\overline{\rm MS}$ masses
is the possibility of direct
connection with the precise determinations from low-energy
experimental data~\cite{Agashe:2014kda}.
In this work we will use the {\tt xFitter} open-source
framework~\cite{Alekhin:2014irh} (previously known as {\tt HERAfitter})
to extract the
$\overline{\rm MS}$ charm mass from a PDF fit to the most up-to-date
inclusive and charm data from HERA.
Structure functions are computed using the FONLL scheme as implemented
in the {\tt APFEL}~\cite{Bertone:2013vaa} code.
Our results have been obtained employing the most accurate
perturbative calculations presently available and will include a
detailed characterization of the different sources of uncertainties on
$m_c(m_c)$ from data, theory and fitting methodology.
As we will show, the results are consistent with the global PDG average as
well as with previous determinations based on the
FFN~\cite{Alekhin:2012un,Abramowicz:1900rp,Alekhin:2012vu,Alekhin:2013qqa}
and in the S-ACOT~\cite{Gao:2013wwa} schemes.\footnote{See also~\cite{Harland-Lang:2015qea} for a recent determination of the pole charm mass from a global
PDF fit.}
The uncertainty in our results turns out to be competitive with that
of previous determinations based on DIS structure functions.
The outline of this paper is the following.
In Sect.~\ref{sec:fonll} we discuss how FONLL can be formulated in
terms of $\overline{\rm MS}$ masses and present a benchmark of its
implementation in {\tt APFEL}.
In Sect.~\ref{sec:fitsettings} we describe the settings of the PDF
fits and the treatment of the uncertainties.
Results for the determination of $m_c(m_c)$ are presented in
Sect.~\ref{sec:results}, where we also compare with previous
determinations.
We conclude and discuss possible next steps in
Sect.~\ref{sec:summary}.
\section{Results}\label{sec:results}
In this section we will present the result for our the determination
of the value $m_c(m_c)$ in the $\overline{\mbox{MS}}$ renormalization
scheme using the FONLL scheme with its associated set of
uncertainties.
The parabolic fit to the global $\chi^2$ as a function of $m_c(m_c)$
is shown in Fig.~\ref{fig:fonllC} and yields a best fit value and its
1-$\sigma$ experimental uncertainty equal to
$m_c(m_c)=1.335 \pm 0.043$ GeV. An estimate of the parametric, model,
and the theoretical uncertainties, performed following the procedure
described in Sect.~\ref{sec:fitsettings}, is summarised in the second
column of Tab.~\ref{tab:model} and leads to our final result:
\begin{equation}\label{eq:fonll}
m_c(m_c)=1.335 \pm 0.043\mbox{(exp)}^{+0.019}_{-0.000}\mbox{(param)}^{+0.011}_{-0.008}\mbox{(mod)}^{+0.033}_{-0.008}\mbox{(th)} \mbox{ GeV.}
\end{equation}
An illustration of the deviations, again determined through parabolic
fits, caused by the variations employed to determine the parametric,
model, and theoretical uncertainties is given in
Fig.~\ref{fig:variations}.
\begin{figure*}
\begin{center}
\includegraphics[angle=270,width=0.9\textwidth]{HERAFitterMcFit_central_FONLLC.pdf}
\end{center}
\caption{\label{fig:fonllC} Parabolic fit to the global $\chi^2$ as
a function of $m_c(m_c)$ in the FONLL-C scheme with nominal
settings.}
\end{figure*}
\begin{table*}
\centering
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}lcc@{}}
\hline
variation & FONLL-C & FFN \\
\hline
\hline
central & $1.335 \pm 0.043$ & $1.318\pm 0.054$\\
\hline
\hline
$Q_0^2=1.5$ & $1.354\;[+0.019]$ & $1.329\;[+0.011]$ \\
\hline
$D_{uv}$ non-zero & $1.340\;[+0.005]$ & $1.308\;[-0.010]$ \\
\hline
\hline
$f_s=0.3$ & $1.338\;[+0.003]$ & $1.320\;[+0.002]$ \\
$f_s=0.5$ & $1.332\;[-0.003]$ & $1.315\;[-0.003]$ \\
\hline
$m_b(m_b)=3.93$ GeV & $1.330\;[-0.005]$ & $1.312\;[-0.006]$ \\
$m_b(m_b)=4.43$ GeV & $1.343\;[+0.008]$ & $1.324\;[+0.006]$ \\
\hline
$\alpha_s(M_Z)=0.1165$ & $1.342\;[+0.007]$ & $1.332\;[+0.014]$ \\
$\alpha_s(M_Z)=0.1195$ & $1.329\;[-0.006]$ & $1.300\;[-0.018]$\\
\hline
\hline
$\mu_F^2=\mu_R^2=2\cdot Q^2$ & $1.347\;[+0.012]$ & $1.314\;[-0.004]$ \\
$\mu_F^2=\mu_R^2= Q^2/2$ & $1.361\;[+0.026]$ & $1.363\;[+0.045]$ \\
\hline
FONLL Damping power = 1 & $1.352\;[+0.017]$ & -- \\
FONLL Damping power = 4 & $1.327\;[-0.008]$ & -- \\
\hline
\end{tabular*}
\caption{\label{tab:model} List of the variations performed to
estimate the non-experimental uncertainties on $m_c(m_c)$ with
the respective results obtained in the FONLL-C scheme and in the FFN scheme at NLO.}
\end{table*}
\begin{figure*}
\begin{center}
\includegraphics[angle=270,width=0.45\textwidth]{HERAFitterMcFit_Q02_FONLLC.pdf}\includegraphics[angle=270,width=0.45\textwidth]{HERAFitterMcFit_Duv_FONLLC.pdf}\\
\includegraphics[angle=270,width=0.45\textwidth]{HERAFitterMcFit_fs_FONLLC.pdf}\includegraphics[angle=270,width=0.45\textwidth]{HERAFitterMcFit_mb_FONLLC.pdf}\\
\includegraphics[angle=270,width=0.45\textwidth]{HERAFitterMcFit_as1_FONLLC.pdf}\includegraphics[angle=270,width=0.45\textwidth]{HERAFitterMcFit_mu2_FONLLC.pdf}\\
\includegraphics[angle=270,width=0.45\textwidth]{HERAFitterMcFit_DampPow_FONLLC.pdf}
\end{center}
\caption{\label{fig:variations} Parabolic fits to the global
$\chi^2$'s as functions of $m_c(m_c)$ in the FONLL-C scheme for
all variations performed to estimate the non-experimental
uncertainties on $m_c(m_c)$.}
\end{figure*}
After we have determined the best fit value of the charm mass in
eq.~(\ref{eq:fonll}), we have used the central value to perform a
further fit in the FONLL-C scheme (nominal fit). In
Tab.~\ref{tab:chi2} we report the partial $\chi^2$'s over the number
of data points for each subset along with the total correlated
$\chi^2$, the logarithmic penalty, and the total $\chi^2$ per degree
of freedom.
\begin{table}
\begin{center}
\begin{tabular}{lp{2.57cm}}
\hline
Data Set & $\chi^2$ \\
\hline
Charm cross section H1-ZEUS combined & 44 / 47 \\
HERA1+2 CCep & 43 / 39 \\
HERA1+2 CCem & 55 / 42 \\
HERA1+2 NCem & 218 / 159 \\
HERA1+2 NCep 820 & 67 / 70 \\
HERA1+2 NCep 920 & 439 / 377 \\
HERA1+2 NCep 460 & 220 / 204 \\
HERA1+2 NCep 575 & 219 / 254 \\
Correlated $\chi^2$ & 104 \\
Log penalty $\chi^2$ & +12 \\
\hline
Total $\chi^2$ / d.o.f. & 1420 / 1178 \\
\hline
\end{tabular}
\end{center}
\caption{\label{tab:chi2} $\chi^2$'s resulting from the fit in the
FONLL-C scheme using the best fit value of the charm mass
$m_c(m_c) = 1.335$ GeV. The partial $\chi^2$'s per data point
along with the total correlated $\chi^2$, the logarithmic penalty,
and the total $\chi^2$ / d.o.f. are reported, as defined in Ref.~\cite{Aaron:2012qi}.}
\end{table}
As an illustration, the singlet and the gluon PDFs extracted from the
nominal fits are compared with other GM-VFNS PDF sets:
CT14\cite{Dulat:2015mca}, HERAPDF2.0\cite{Abramowicz:2015mha},
MMHT14\cite{Thorne:2015caa}, NNPDF3.0\cite{Ball:2014uwa}. They are
shown in Fig.~\ref{fig:pdfs} at the scale $Q^2=10$ GeV$^2$, where the
the experimental uncertainties from the nominal fits on PDFs are estimated using Monte Carlo
procedure with the root mean square estimated from 500 replica. An overall
good agreement is observed.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=0.48\textwidth]{q2_10_pdf_Sea.pdf}\includegraphics[angle=0,width=0.48\textwidth]{q2_10_pdf_g.pdf}
\end{center}
\caption{\label{fig:pdfs} Comparison at $Q^2 = 10$ GeV$^2$ of the
singlet (left plot) and gluon (right plot) distributions from the
nominal FONLL-C fit with other PDF sets determined using GM-VFN
schemes: HERAPDF2.0, CT14, MMHT14, NNPDF3.0.}
\end{figure*}
The FONLL determination of $m_c(m_c)$ presented above is supported by
an analogous determination in the FFN scheme at NLO. The corresponding
parabolic fit with the associated experimental uncertainty is shown in
Fig.~\ref{fig:ffABM}. Also in this case a full characterization of the
non-experimental uncertainty has beed achieved by carrying out the
same parametric, model, and theory variations (except for the
variation of the damping factor which is specific of the FONLL
scheme). The results of the variation in the FFN scheme are reported
in the third column of Tab.~\ref{tab:model}. The final result is:
\begin{equation}\label{eq:ffab}
m_c(m_c) = 1.318\pm 0.054\mbox{(exp)}^{+0.011}_{-0.010}\mbox{(param)}^{+0.015}_{-0.019}\mbox{(mod)}^{+0.045}_{-0.004}\mbox{(th)}\mbox{ GeV}\,,
\end{equation}
which is in agreement with the FONLL determination given in
eq.~(\ref{eq:fonll}).
\begin{figure*}
\begin{center}
\includegraphics[width=0.7\textwidth]{FF-ABM-central.pdf}
\end{center}
\caption{\label{fig:ffABM} Parabolic fit to the global $\chi^2$ as
a function of $m_c(m_c)$ in the FFN scheme at NLO with nominal
settings.}
\end{figure*}
It is interesting to notice that we observe a reduced scale dependence
in the FONLL scheme as compared to the FFN scheme. We ascribe this
effect to the fact that the leading contributions in the FONLL scheme
involve both gluon- and quark-initiated processes; typically the
contributions from gluon processes decrease with the scale, while the
contributions from quark processes tend to increase. Conversely, the
FFN scheme is mostly driven by gluon processes the contributions of
which (along with $\alpha_s$) tend to be monotonic in $\mu$ leading to
larger scale variations\footnote{We thank Fred Olness for this
interesting observation.}.
As discussed Sect.~\ref{sec:MatchingConditions}, the running of the
$\overline{\mbox{MS}}$ heavy-quark masses in the VFN scheme, exactly
like the running of $\alpha_s$ and PDFs, is not univocally defined at
the heavy-quark thresholds due to the presence of the so-called
matching conditions. In particular, when giving the value of the mass
at one of the heavy-quark thresholds, one should also specify whether
this corresponds to the value immediately below or above the threshold
itself. This is typically done by complementing the value with the
number of active flavors used in the computation. In fact, in general
$m_c^{(N_f=3)}(m_c) \neq m_c^{(N_f=4)}(m_c)$. On theoretical grounds,
this difference is relevant when comparing a determination obtained in
a VFN scheme like FONLL with a determination obtained in the ($N_f=3$)
FFN scheme: in the latter one automatically determines
$m_c^{(N_f=3)}(m_c)$, while in the former it is more natural to
extract $m_c^{(N_f=4)}(m_c)$. However, eqs.~(\ref{eq:MarchMhUp})
and~(\ref{eq:MarchMhDown}) tell us how the two values are connected up
to $\mathcal{O}(\alpha_s^2)$ and applying eq.~(\ref{eq:MarchMhUp}) to
the central value eq.~(\ref{eq:fonll}) one gets
$m_c^{(N_f=3)}(m_c) = 1.339$ GeV, that is a difference of 0.004 GeV as
compared to the nominal value which is well within the current
uncertainty on $m_c(m_c)$. We can then conclude that, even though
providing a value $m_c(m_c)$ is ambiguous if the number of active
flavors is not specified, the magnitude of the ambiguity is currently
not large enough to significantly affect the current determinations.
\subsection{Comparison to other results}\label{sec:comparison}
It is interesting to compare our results with the past determinations
of $\overline{\mbox{MS}}$ charm mass $m_c(m_c)$ using a similar
methodology (also see Ref.~\cite{Behnke:2015qja,Alekhin:2013qqa,Gao:2013wwa}
for previous comparisons).
The analysis of Ref.~\cite{Alekhin:2012vu} was performed in the ABM11
framework~\cite{Alekhin:2012ig} using the FFN scheme at NLO and at
approximate NNLO and based on world data for DIS from HERA, and
fixed-target DIS experiments and Tevatron Drell-Yan data. While the
analysis in Ref.~\cite{Alekhin:2012vu} was performed including the
same exclusive charm cross-section data used in this study, it did not
include the HERA1+2 combined inclusive cross-section data set which
was not available at the time, but used instead the HERA combined data
from run 1 only. An earlier analysis~\cite{Alekhin:2012un} used a
partial charm dataset only, with correspondingly larger uncertainties,
while a subsequent analysis~\cite{Alekhin:2013qqa} investigated the
correlation between the measurement of $m_c(m_c)$ and the strong
coupling constant.
The analysis of Ref.~\cite{Gao:2013wwa} is instead based on the
CT10NNLO global analysis, and uses the S-ACOT-$\chi$ GM-VFN scheme
discussed, $e.g.$, in Ref.~\cite{Guzzi:2011ew}. It is based on a
slightly wider data set as it includes LHC jet production data and
also a set of older $F_2^c$ measurements at HERA~\cite{Adloff:2001zj}
that are not included in the more recent combined charm data. The
authors of Ref.~\cite{Gao:2013wwa} provide a set of four
determinations deriving from different strategies to convert the
pole-mass definition into $\overline{\mbox{MS}}$. They also provide a
separate estimate of the uncertainty due to the
$\mathcal{O}(\alpha_s^3)$ corrections for one of the four strategies
essentially by varying the parameter that governs a generalized
version of the rescaling variable $\chi$.
Finally, a determination of the charm mass $m_c(m_c)$ was produced by
the H1 and ZEUS collaborations in the framework of the HERAPDF QCD
analysis in the same publication in which the charm cross-section
measurements employed in our study were
presented~\cite{Abramowicz:1900rp}. That determination also used only
the HERA combined inclusive data from run 1~\cite{Aaron:2009aa}.
In Tab.~\ref{tab:results} we report the numerical values for the
$m_c(m_c)$ determinations listed above along with our results
and the world average value~\cite{PhysRevD.86.010001}.
A short clarification about the nomenclature of the uncertainties
reported in Tab.~\ref{tab:results} is in order. In
Sect.~\ref{sec:fitsettings} we discussed extensively the meaning of
the uncertainties associated to our determinations. In doing so, we
tried to be consistent with the previous determinations, nevertheless
some differences remain. As far as the determination in
Ref.~\cite{Abramowicz:1900rp} is concerned, while their definition of
``(exp)'' and ``(param)'' essentially coincides with ours, their
``(model)'' uncertainty includes the variation of the cut in $Q^2$
(that we will discuss separately in Sect.~\ref{sec:Q2minDependence})
but does not include the $\alpha_s$ variation, which is instead quoted
separately. In addition, the authors do not quote any scale variation
uncertainty. The nomenclature of Ref.~\cite{Alekhin:2012vu} is also
different from ours. Apart from the common ``(exp)'' uncertainty, for
the NLO determination the authors only quote the ``(scale)''
uncertainty, which essentially coincides with our ``(th)'' (even
though the FONLL ``(th)'' uncertainty also accounts for the variation
of the damping factor), while for the approximate NNLO determination
they also quote a ``(th)'' uncertainty which, differently from our
nomenclature, accounts for the uncertainty on the approximated
expressions used at $\mathcal{O}(\alpha_s^3)$. Finally, the
determinations in Ref.~\cite{Gao:2013wwa} only quote the experimental
uncertainty (the asymmetric uncertainties are due to the use of a
generic second-degree polynomial to fit the $\chi^2$ profiles).
A graphical representation of Tab.~\ref{tab:results} is shown in
Fig.~\ref{fig:McComparison} where the inner error bars display the
experimental uncertainty while the outer error bars (when present) are
obtained as a sum in quadrature of all uncertainty sources. The blue
vertical band represents the world average and provides a reference
for all other determinations. It is clear that, while the spread of
the current determinations of $m_c(m_c)$ from DIS data covers a pretty
large range, they are generally in agreement with the world
average. As far as our determinations in particular are concerned, we
observe that, apart from being consistent with each other and with the
world average, they also present competitive uncertainties. This is
particularly relevant for the FONLL determination because this is the
first time that this scheme is employed for a direct determination of
the charm mass.
\begin{table*}
\centering
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}ll@{}}
\hline
scheme & $m_c(m_c)$ [GeV] \\
\hline
\hline
FONLL (this work) & $1.335\pm0.043\mbox{(exp)}^{+0.019}_{-0.000}\mbox{(param)}^{+0.011}_{-0.008}\mbox{(mod)}^{+0.033}_{-0.008}\mbox{(th)}$\\
FFN (this work) & $1.318\pm0.054\mbox{(exp)}^{+0.011}_{-0.010}\mbox{(param)}^{+0.015}_{-0.019}\mbox{(mod)}^{+0.045}_{-0.004}\mbox{(th)}$\\
\hline
FFN (HERA)~\cite{Abramowicz:1900rp} & $1.26 \pm 0.05\mbox{(exp)} \pm 0.03\mbox{(mod)} \pm 0.02\mbox{(param)} \pm 0.02(\alpha_s)$ \\
\hline
FFN (Alekhin \textit{et al.})~\cite{Alekhin:2012vu} & $1.24 \pm 0.03({\rm
exp})^{+0.03}_{-0.02}({\rm
scale})^{+0.00}_{-0.07}({\rm th})$
(approx. NNLO)\\
& $1.15 \pm 0.04({\rm exp})^{+0.04}_{-0.00}({\rm scale})$ (NLO)\\
\hline
S-ACOT-$\chi$ (CT10)~\cite{Gao:2013wwa} &
$1.12^{+0.05}_{-0.11}$
(strategy 1)\\
& $1.18^{+0.05}_{-0.11}$ (strategy 2)\\
& $1.19^{+0.06}_{-0.15}$ (strategy 3)\\
& $1.24^{+0.06}_{-0.15}$ (strategy 4)\\
\hline
\hline
World average~\cite{PhysRevD.86.010001} & $1.275 \pm 0.025$ \\
\hline
\end{tabular*}
\caption{\label{tab:results} List of the recent determinations of
$m_c(m_c)$ from fits to DIS data along with the
determinations extracted in this work. The PDG world average value
is also reported for reference.}
\end{table*}
\begin{figure*}
\begin{center}
\includegraphics[angle=180,width=0.5\textwidth]{mc_comparison.pdf}
\end{center}
\caption{\label{fig:McComparison} Graphical representation of the
determinations reported in Tab.~\ref{tab:results}. The inner error
bars display the experimental uncertainty while the outer error
bars (when present) are obtained as a sum in quadrature of all
uncertainty sources. The blue vertical band represents the world
average and provides a reference for all other determinations.}
\end{figure*}
Fig.~\ref{fig:McComparison} shows that our determinations tend to be
larger than the world average while most of the previous
determinations place themselves below it. Detailed investigations
show that the largest contribution to this difference arises from the
use of to the new combined HERA1+2 combined inclusive cross section
measurements that are employed for the first time to determine the
charm mass and that, as we will discuss in
Sect.~\ref{sec:Q2minDependence}, tend to prefer larger values of
$m_c(m_c)$.
\subsection{Cross-checks}\label{sec:crosschecks}
It is worth mentioning that we have also employed the variants A and B
of the FONLL scheme discussed in Sect.~\ref{sec:fonll} to determine
$m_c(m_c)$. While the FONLL-A scheme is accurate to LO in the massive
sector and thus does not produce a reliable determination of the charm
mass, the FONLL-B has the same formal accuracy in the massive sector
as FONLL-C and indeed it leads to a determination comparable to that
given in eq.~(\ref{eq:fonll}) both for the central value and the
uncertainties. It is interesting to notice that the FONLL-B scheme in
the low-energy region resembles very closely the FFN scheme at NLO.
In particular, both schemes are accurate to $\mathcal{O}(\alpha_s^2)$
in the massive sector and to $\mathcal{O}(\alpha_s)$ in the light
sector. As a matter of fact, we find that the experimental uncertainty
associated to the FONLL-B determination is very close to the FFN one
quoted in eq.~(\ref{eq:ffab}), which in turn is around 20\% larger
than that associated to the FONLL-C determination. This suggests that
the $\mathcal{O}(\alpha_s^2)$ corrections to the light sector that are
present in the FONLL-C scheme, which depend on the heavy-quark mass by
means of diagrams in which a gluon plits into a pair of heavy quarks,
provide a further constraint on $m_c(m_c)$.
Finally, we have also attempted a determination in the FFN scheme
using the approximate NNLO massive structure functions as implemented
in {\tt OPENQCDRAD}. However, we did not pursue a full
characterization of the uncertainties because we believe that this
determination, while giving a quantitative indication of the effect of
the NNLO corrections, cannot claim an NNLO accuracy and thus does not
add anything to our NLO determinations.
\subsection{Discussion on the $Q_{\rm min}^2$ dependence of the mass
determination}\label{sec:Q2minDependence}
Our determination of $m_c(m_c)$ given in eq.~(\ref{eq:fonll}) was
obtained cutting off all data with $Q^2 < Q_{\rm min}^2 = 3.5$
GeV$^2$. The necessity of such a cut stems from the fact that
low-energy data are hard to describe for two main reasons: the large
value of $\alpha_s$ with consequent large higher-order corrections,
and sizable higher-twist corrections. In addition, as pointed out in
Ref.~\cite{Caola:2009iy}, the low-$Q^2$ region (low-$x$, in fact)
might be affected by deviations from the fixed-order DGLAP evolution
whose description might require small-$x$ perturbative resummation.
The dependence on $Q_{\rm min}^2$ of fits to HERA data has already
been discussed in the context of the inclusive measurements only. In
this section, we will address this issue considering also the HERA
charm production data.
The particular value of $Q_{\rm min}^2$ used in our analysis (3.5
GeV$^2$) was determined by requiring a good fit quality but
maintaining a good sensitivity to $m_c(m_c)$. This is illustrated in
Fig.~\ref{fig:q2min} where the global $\chi^2$ per degree of freedom
is plotted as a function of $Q_{\rm min}^2$ in the left panel while
the best fit of $m_c(m_c)$ is plotted as a function of $Q_{\rm min}^2$
in the right panel. Looking at the left panel it is clear that, as
expected, the global $\chi^2$ improves as more and more low-energy
data are excluded from the fit. On the other hand, the right plot
shows that the experimental uncertainty associated to $m_c(m_c)$ gets
larger and larger as $Q_{\rm min}^2$ increases indicating that, again
as expected, the sensitivity to $m_c(m_c)$ deteriorates if low-energy
data are excluded. In the light of the plots in Fig.~\ref{fig:q2min},
we conclude that $Q_{\rm min}^2 = 3.5$ GeV$^2$ represents a good
compromise between a good description of the full data set and a good
sensitivity to $m_c(m_c)$.
\begin{figure*}
\begin{center}
\includegraphics[angle=270,width=0.49\textwidth]{q2cut_fonllC_chi2.pdf}
\includegraphics[angle=270,width=0.49\textwidth]{q2cut_fonllC_mc.pdf}
\end{center}
\caption{\label{fig:q2min} Left plot: dependence of the global
$\chi^2$ / d.o.f. as a function of $Q_{\rm min}^2$. Right plot:
dependence of the global best fit value of $m_c(m_c)$ with the
associated experimental uncertainty as a function of
$Q_{\rm min}^2$. Both plots have been obtained using the FONLL-C
scheme.}
\end{figure*}
In this context, it is interesting to look at the behaviour of the
partial $\chi^2$'s as a function of $Q_{\rm min}^2$ of the charm and
inclusive cross-section data separately to assess in a more specific
way which nominal value of $Q_{\rm min}^2$ is more convenient. Since
the meaning of ``degrees of freedom'' is unclear for a subset of the
full data set, in order to quantify the degree of improvement in the
partial $\chi^2$'s, we consider the following quantity:
\begin{equation}\label{eq:DeriveChi2}
\frac{\Delta\chi^2}{\Delta N_{\rm points}}(Q_{\rm min}^2) =
\frac{\chi^2(Q_{\rm min}^2) - \chi^2(Q_{\rm min}^2 = 2.5\mbox{ GeV}^2)}{N_{\rm points}(Q_{\rm min}^2) - N_{\rm points}(Q_{\rm min}^2 = 2.5\mbox{ GeV}^2)}\,,
\end{equation}
which provides an estimate of the improvement of the $\chi^2$ per data
point with respect to our lowest cut $Q_{\rm min}^2 = 2.5$ GeV$^2$. If
for a given value of $Q_{\rm min}^2$ this quantity is larger than one,
this means that that specific cut leads to an improvement of the
$\chi^2$ which is larger than the degrees of freedom subtracted by
excluding a given number of data points and thus the excluded data
points with respect of the reference cut (2.5 GeV$^2$) are poorly
described. On the contrary, if the quantity in
eq.~(\ref{eq:DeriveChi2}) is smaller than one, this means that the
excluded data points are better described than the fitted ones. In the
left panel of Fig.~\ref{fig:q2min_f2ch} we show the behaviour of the
contribution to the global $\Delta\chi^2/\Delta N_{\rm points}$
originating from the charm data points only. It is clear that any cut
between 3.5 and 5 GeV$^2$ improves drastically the partial $\chi^2$
while cuts above 5 GeV$^2$ either cause a much less significant
improvement or even lead to a deterioration. This provides a further
confirmation of the fact that our nominal cut (3.5 GeV$^2$) is a
sensible choice.
It is also interesting to look at the best fit values of $m_c(m_c)$
and the relative uncertainty preferred by a given subset as a function
of $Q_{\rm min}^2$ to quantify the sensitivity to $m_c(m_c)$ as more
and more data are excluded from the fit. This is plotted in the right
panel of Fig.~\ref{fig:q2min_f2ch} for the charm cross-section
data. It is clear that this particular subset of data tends to prefer
values of $m_c(m_c)$ around 1.23 GeV which is substantially lower than
the global value given in eq.~(\ref{eq:fonll}). The stability of the
central value of $m_c(m_c)$ for different values of $Q_{\rm min}^2$ is
remarkable and, as expected, the experimental uncertainty tends to
increase for larger value of $Q_{\rm min}^2$ indicating a loss of
sensitivity.
\begin{figure*}
\begin{center}
\includegraphics[angle=270,width=0.49\textwidth]{q2min_fonll_chi_fc.pdf}
\includegraphics[angle=270,width=0.49\textwidth]{q2min_fonll_mc_fc.pdf}
\end{center}
\caption{\label{fig:q2min_f2ch} Left plot: dependence of
$\Delta\chi^2/\Delta N_{\rm points}$ as a function of $Q_{\rm
min}^2$ for the charm cross-section subset. Right plot:
dependence of the best fit value of $m_c(m_c)$ with the associated
experimental uncertainty as a function of $Q_{\rm min}^2$ for the
charm cross-section subset in the combined fit. Both plots have
been obtained using the FONLL-C scheme.}
\end{figure*}
Finally, we have done the same exercise for the HERA1+2 inclusive
cross-section data and in Fig.~\ref{fig:q2min_incl} we present the
relative plots. In the left panel we observe that the $\chi^2$ of this
subset improves essentially monotonically as $Q_{\rm min}^2$ increases
while from the right panel it is clear that the preferred value of
$m_c(m_c)$ of the inclusive cross sections is substantially larger
than that preferred by the charm cross sections with, again,
uncertainties than become broader for larger values of
$Q_{\rm min}^2$. It is finally clear that our best value for
$m_c(m_c)$ quoted in eq.~(\ref{eq:fonll}) is a compromise between the
lower value preferred by the exclusive charm data and the larger value
preferred by the inclusive data.
\begin{figure*}
\begin{center}
\includegraphics[angle=270,width=0.49\textwidth]{q2min_fonll_chi_inc.pdf}
\includegraphics[angle=270,width=0.49\textwidth]{q2min_fonll_mc_inc.pdf}
\end{center}
\caption{\label{fig:q2min_incl} Same as Fig.~\ref{fig:q2min_f2ch}
for the inclusive cross-section subset in the combined fit.}
\end{figure*}
\subsection{Discussion on the sensitivity to $m_c(m_c)$ of the
inclusive data}\label{sec:IncDataSens}
It is clear from the right panels of Figs.~\ref{fig:q2min_f2ch}
and~\ref{fig:q2min_incl} that the exclusive charm and inclusive data
subsets prefer somewhat different values of $m_c(m_c)$. However, the
values shown in these figures are clearly correlated because they were
obtained in a simultaneous fit to all data. In order to investigate a
possible tension, we have performed a fit to the inclusive data only
using both the FONLL-C and FFN schemes. The $\chi^2$ profiles are
shown in Fig.~\ref{fig:chi2inclusive}. In contrast to
Figs.~\ref{fig:fonllC} and~\ref{fig:ffABM}, in both schemes the scan
in $m_c(m_c)$ of the fits to inclusive data only yielded a shallow
$\chi^2$ dependences with a minimum around $1.7$ GeV. This
demonstrates that the inclusive data alone cannot constrain $m_c(m_c)$
reasonably well, but also why this data exerts an upwards pull on the
$m_c(m_c)$ value in the combined fit. Furthermore, since
Figs.~\ref{fig:q2min}, \ref{fig:q2min_f2ch}, and~\ref{fig:q2min_incl}
in Sect.~\ref{sec:Q2minDependence} present an overall remarkable
stability of the central value of $m_c(m_c)$ for different values of
$Q^2_{\rm min}$, the observed feature cannot be attributed to the low
$Q^2$ part of the inclusive data.
\begin{figure*}
\begin{center}
\includegraphics[angle=270,width=0.7\textwidth]{chi2incl.pdf}
\end{center}
\caption{$\chi^2$ vs. $m_c(m_c)$ profile for the fits to the HERA1+2
inclusive cross sections only. The red circles indicate the
FONLL-C scheme while the blue squares the FFN scheme at
NLO.\label{fig:chi2inclusive}}
\end{figure*}
\section{Conclusions}\label{sec:summary}
In this work we have presented a new determination of the
$\overline{\rm MS}$ charm quark mass $m_c(m_c)$ obtained by fitting
HERA charm and inclusive DIS data. In particular, we included in our
fits the combined H1 and ZEUS charm production cross-section
measurements~\cite{Abramowicz:1900rp} and the final combination of
HERA1+2 H1 and ZEUS inclusive DIS cross-section
data~\cite{Abramowicz:2015mha}, the latter being used in this work for
the first time for the extraction of the charm mass.
Our determination is based on the FONLL general-mass
variable-flavor-number scheme, and has required the generalization of
the FONLL structure functions, originally constructed in the pole-mass
scheme, in terms of $\overline{\rm MS}$ heavy quark masses.
A detailed estimate of the various sources of uncertainty that affect
our determination of $m_c(m_c)$ has been performed. In particular, we
estimated the uncertainties due to the choice of the PDF
parametrization, the model parameters used as input for the
theoretical computations, and the missing higher-order corrections. We
found that those sources of uncertainty are smaller than the
experimental uncertainty, resulting in a competitive determination of
the charm mass.
We complemented the FONLL extraction of the charm mass with an
analogous determination based on the fixed-flavour number scheme at
next-to-leading order, finding a good agreement between the two. In
addition, we compared our results with previous determinations also
based on fits to DIS data and with the PDG world average finding again
a generally good agreement. We find that the values extracted in this
work, although compatible within uncertainties, tend to be slightly
higher than previous determinations from HERA data. This feature seems
to be associated to the final HERA1+2 combined inclusive dataset,
which tends to prefer larger values of $m_c(m_c)$ as compared to the
charm structure function data, and thus increases the best-fit value.
In the future, it would be interesting to repeat the FONLL
determination in the context of a global PDF analysis, since, in
addition to the inclusive and charm HERA data, other experiments are
expected to have some sensitivity to the value of the $\overline{\rm
MS}$ charm mass.
In addition, the use of a wider dataset might lead to a reduction of
the experimental uncertainties of the $m_c(m_c)$ determination.
Moreover, our analysis is based on the standard assumption that the
charm PDF is dynamically generated by collinear splitting from gluons
and light quarks.
In this respect, it would be useful to redo the determination of
$m_c(m_c)$ in the presence of a possible non-perturbative charm PDF,
for which the generalized FONLL structure functions accounting for a
fitted heavy quark PDF are available~\cite{Ball:2015tna}.
|
\section{Introduction}
Color images are frequently presented in grayscale.
This is common with digital ink displays and grayscale-only printing,
but synthesizing a consistent grayscale image is also important for color blindness, and for prosthetic vision devices such as retinal implants\cite{AytonPLOSONE14} and sensory substitution devices \cite{Stronks2015} which mostly convey brightness only.
One may also wish to present color information in grayscale as part of a virtual or augmented reality application.
Such conversion is sometimes considered in image processing tasks such as tone-mapping and compression.
For these reasons converting color images to grayscale images (C2G) has attracted a stream of research attention (\textit{e}.\textit{g}., \cite{Bala03,Gooch05,Yoo15}).
Despite its common usage, C2G is non-trivial.
Normal human vision perceives color with three peak spectral sensitivities \cite{Pokorny2004}.
Hence, perceivable color-space is homogeneous to $\mathbb{R}^3$, whereas gray-scale space is homogeneous to $\mathbb{R}$.
Standard displays present an image in a smaller discrete range though dynamic range and quantization restrictions.
Perception is restricted further by human observers' ability to perceive brightness and color difference \cite{Stevens57}.
However, as perception of the presented grayscale image is also restricted, it is inevitable that information loss will occur in C2G.
More importantly, grayscale space has strong perceptual meaning.
Simply quantizing the 3D signal and presenting it as a 1D signal often will not lead to a grayscale image that is perceptually consistent with the original color image .
An example is shown in \fref{Fig:Teaser}, Claude Monet's masterpiece "Impression Sunrise", which is included in {\^C}ad{\'\i}k~ dataset \cite{Cadik08} and has been actively used for validating C2G algorithms.
As can be seen, while the original color image gives the feeling of sunrise, that feeling can easily be lost in conversion to grayscale.
A direct grayscale conversion without considering color contrast can change the feeling of the image to fog or haze.
On the other hand, strongly enhancing contrast while allowing the brightness to change aggressively can change the feeling of the image to moonlight.
Further, some representations of color can lead to an impression of the scene where the appearance is not natural.
For example, consider the first row of \fref{Fig:VisualNeo}, where as a result of methods that emphasize contrast over perceptual consistency, orange juice may appear black.
\begin{figure}[tb]
\includegraphics[width=\linewidth]{Fig/Teaser.pdf}
\vspace{-20pt}
\caption{Converting Claude Monet's masterpiece "Impression Sunrise" to grayscale.} A conversion without considering the color contrast will change the feeling to haze or fog while a conversion without considering the brightness will change the feeling to moonlight.
\label{Fig:Teaser}
\vspace{-12pt}
\end{figure}
Another important consideration is that "contrast" is a relative concept that is highly reliant on context. A famous example is the shadow illusion by Edward Adelson \textit{et~al.}\cite{Adelson00}.
While earlier research in C2G focused more on contrast in color space, current research considers spatial context. However, perceiving spatial context is rather complex. Thus, a perceptional consistent contrast model is desired.
In this paper, we propose a C2G algorithm which aims to preserve perceptual consistency as much as possible.
To this end, we propose two perceptual consistency measurements based on contemporary research in vision science: brightness and multi-scale contrast measurement.
The brightness measurement is based on the idea that brightness in a grayscale image does affect the perception of the probability of color information.
The color contrast measurement is based on the finding that the contrast of a given pixel to its surroundings can be measured as a linear combination of color contrast in different scales.
Based on these measures, we propose graphical model based optimization framework to balance the brightness and contrast measurement. To solve the optimization, an $\ell_1$-norm based method is provided.
To validate our methods, we also propose a new dataset, called NeoColor, that improves on existing datasets.
The {\^C}ad{\'\i}k~ dataset \cite{Cadik08} has only 24 images. Further, the Color 250 dataset \cite{Lu14} is a subset of Berkeley segmentation dataset which was not designed to evaluate C2G algorithms. For NeoColor, we collect about 300 images from typical C2G scenarios including: natural scenes from commercial printing, books, magazines, masterpiece artworks, and computer designed graphics.
The contribution of our work is three-fold.
Firstly, we propose a rationale for perceptual consistency of C2G algorithms.
Based on perceptual consistency, we propose brightness and multi-scale color contrast measurements.
Secondly, we propose an efficient algorithm for C2G which best preserves the perceptual consistency measurement.
Thirdly, a new dataset which is specially designed for C2G evaluation is provided to address some shortcomings of the {\^C}ad{\'\i}k~ and Color250 datasets.
The rest of the paper is organized as follows: Section 2 summarizes related work on C2G. In Section 3, we discuss key aspects of human perception.
Based on this, in Section 4, the perceptual consistency C2G algorithm is described. Section 5 demonstrates the effectiveness of the proposed method using both automatic evaluation and a user study. Section 6 concludes the paper.
\section{Related Work}
Existing C2G research can be categorized as global and local methods. Generally, global methods seek a global mapping for each color value to an intensity value, aiming to maximize the overall contrast, whereas local methods enhance spatially local contrast.
\vspace{-12pt}
\paragraph{Global methods:}
Bala \textit{et~al.} \cite{Bala03} start with an image with a limited number of colors, the colors are sorted and brightness values are assigned accordingly.
Gooch \textit{et~al.} \cite{Gooch05} considered brightness variation according to hue, the orientation of the hue-ring was determined by maximizing the global contrast.
Kim \textit{et~al.} \cite{Kim09} extended Gooch \textit{et~al.}'s method: variation from the hue-ring is represented by a Fourier series instead of the original linear mapping.
Ancunti \textit{et~al.} \cite{Ancuti11} also extended the method of Gooch \textit{et~al.}, so that salient and non-salient areas have different hue-ring variation. A user study was also included to evaluate the effectiveness.
Rosche \textit{et~al.} \cite{Rasche05,Rasche05re} sought a continuous global mapping to assign similar colors to similar grayscale values, using pair-wise comparison between sampled colors.
Using the same framework, Grundland \textit{et~al.} \cite{Grundland07} improved the computational efficiency by using PCA to find the 1D component that maximizes contrast.
In more recent work, spatial context has been taken in consideration for global methods. Kuk \textit{et~al.} \cite{Kuk10} proposed method to find landmark colors though K-means. Adopting similar strategy for spatial context, Lu \textit{et~al.} \cite{Lu12,Lu14} proposed a bimodal contrast preserving energy function, along with local and non-local constraints. An evaluation metric and dataset were also introduced to automatically evaluate C2G algorithms.
Ji \textit{et~al.} \cite{Ji15} proposed a hybrid method where color mapping is performance by modulating the hue ring and edges are sharpened by bandpass filtering.
Yoo \textit{et~al.} \cite{Yoo15} considered video C2G instead of single image C2G. They clustered the image sequence in spatio-temporal space to a 1D manifold, and assigned brightness accordingly.
\vspace{-12pt}
\paragraph{Local methods:}
Bala \textit{et~al.} \cite{Bala04} proposed high pass filtering the color image to strengthen local contrast when converting to grayscale.
Subsequently, Smith \textit{et~al.} \cite{Smith08} presented a method to pre-assign color variation based on vision science studies, and applied the variation to enhance local contrast.
In recent work, Liu \textit{et~al.} \cite{Liu15} proposed an enhancement which preserves gradient, while attenuating the halo between inconsistent boundaries and
local method of Du \textit{et~al.}\cite{Du15} emphasised color contrast in salient areas.
Existing methods focus on best preserving contrast. However, inadequate focus has been given to ensuring that the image is consistent with human perception.
\section{Human Perception of Color and Measurement}
Modeling human perception of color in an algorithm is not trivial. It relies not only on physical modeling of object luminance and spectrum, but also on studies of human visual perception.
Before we introduce our algorithm, which aims to produce a grayscale image that is consistent with human perception of the original color image, in this section, we discuss the factors that affect human color perception.
We first introduce the basic concepts in color space and our notation.
In Sec. 3.1 we discuss the perception of color brightness with a probability model, and introduce an energy function to model brightness perception.
Then, in Sec. 3.2 we discuss perception of color contrast in complex spatial contexts and introduce an energy function to model complex color contrast perception.
\\
\\
\textbf{Background and notation}
\vspace{-12pt}
\paragraph{Color space:} Since Young-Helmholz theory in the mid-1800s, it has been generally agreed that human perception of color is within a three dimensional space \cite{Campbell68,Lotto02,Peli90,Haun13,Stevens57,Majumder07,Jaroensri15}.
In images, this can be represented in several common color spaces, \textit{e}.\textit{g}., RGB, HSL, YUV, CIEL*a*b*.
Because CIE-LAB preserves the perceived difference between colors, in this paper, we will use Euclidean distance in L*a*b* space as the contrast metric, which is common practice for recent C2G research \cite{Lu14,Du15,Ma15,Liu15}.
\Fref{Fig:ColorRing}.a is an illustration of the CIEL*a*b* color space: a spherical space where lightness $l$ varies from 0 for black to 100 for white. $a$ varies from 100 for red to -100 for green; and $b$ varies from 100 for yellow to -100 for blue.
\vspace{-6pt}
\paragraph{Notation:} Notation is listed in Table \ref{T:Denotation}, for the convenience of the reader. Specifically, we denote the color channels of a pixel as $l,a,b$ respectively and the targeted grayscale value as $g$. We use lower case for a single pixel, capitalized letter for image matrix and bold lower case for vectors.
\begin{table}[tb]
\caption{Notation}
\vspace{-6pt}
\includegraphics[width=\linewidth]{Fig/Denotation.pdf}
\label{T:Denotation}
\vspace{-12pt}
\end{table}
\subsection{Perception of Color Brightness and Perception Measurement}
\paragraph{Color as function of brightness:}
\begin{figure}[tb]
\includegraphics[width=\linewidth]{Fig/ColorRing.pdf}
\vspace{-12pt}
\caption{Color variance as function of brightness.} (a) The CIE L*a*b* color space. (b) A color slice in L*a*b* space with high, middle, low brightness. A high/low brightness color does not give same level of color contrast as mid-brightness color.
(c) Desaturated slice of (b), a high/low brighness indicates a small color variation, while middle brightness indicates a large range for color variation.
\label{Fig:ColorRing}
\vspace{-12pt}
\end{figure}
As can be inferred from L*a*b* color space, color components with very low or very high brightness do not provide the same level of color variance as those with mid brightness levels. We show a simple example in \fref{Fig:ColorRing}.
Unlike previous C2G, especially global methods, \cite{Lu14,Grundland07}, where brightness is considered to be independent of color, we consider brightness as an importance cue for indicating color.
As we presented in \fref{Fig:ColorRing}, although brightness does not allow us to infer the $a,b$ value of a color, it does allow us to infer the possibility of color. This is important for generating a grayscale image that preserves the perceivable feeling of color.
Here we discuss the possibility of infereing color $a, b$ as function of brightness $l$. Using the metric in CIEL*a*b* space, a color $(l, a, b)$ with brightness $l \in [0, 100]$ has a color variance, denoted as $v$, limited in the spherical color space:
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
v = \parallel (a(l), b(l))\parallel_2 \leq \sqrt{100 ^2 - 4 * (l - 50) ^ 2},
\label{Eq:BCM2}
\end{equation}
Assuming a uniform distribution of color in CIEL*a*b* space, the partial probability of $v$ by $l$ is:\footnote{We use proportion instead of equal mark and neglect a constant for normalization.}
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
\begin{split}
p(v = \tilde{v} | l) \propto \tilde{v}\sqrt{100 ^2 - 4 * (l - 50) ^ 2},
\\
0 \leq \tilde{v} \leq \sqrt{100 ^2 - 4 * (l - 50) ^ 2}.
\end{split}
\label{Eq:BCM2}
\end{equation}
and similarly:
\vspace{-12pt}
\begin{equation}
\vspace{-0pt}
\begin{split}
p(l = \tilde{l} | v) = \frac{2}{\sqrt{100^2 - v^2}},
\\
50 - \frac{\sqrt{100^2 - v^2}}{2} < \tilde{l} < + \frac{\sqrt{100^2 - v^2}}{2} .
\end{split}
\label{Eq:BCM3}
\end{equation}
As can be seen, a color with a large $v$ value is less likely to have a large variance in $l$, whereas a color with large $l$ is less likely to have a large variance in $v$.
\vspace{-0pt}
\paragraph{Brightness Perceptual Energy}
Considering \eref{Eq:BCM2} and \eref{Eq:BCM3}, we define the brightness perceptual energy for inconsistent brightness between the color and grayscale image as:
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
\begin{split}
E_B(g) &= w(l,a,b)\parallel l - g \parallel _ {\ell}
\\
& = \frac{1}{\sqrt{100^2 - a^2 - b^2} + \epsilon} \frac{1}{\sqrt{100^2 - (2l - 100)^2} + \epsilon}\parallel l - g \parallel _ {\ell},
\end{split}
\label{Eq:BCM1}
\end{equation}
where $\epsilon$ is a small constant regulator.
Generally, when the brightness in a color image $l$ is not mapped to same grayscale value $g$, error energy increases. The error energy is reweighted by $w$ which penalizes color with large variance $v$ (because high probablity color is not sensed); and color with high or low brightness $l$ (because low probablity of color is not sensed).
\vspace{-6pt}
\subsection{Perception of Multi-Scale Spatial Contrast and Measurement}
In the previous subsection, we assume that the two areas of color appear close together spatially, and color contrast is only evaluated between the two.
Much of the existing C2G research is based on such an assumption.
However, color contrast is not absolute, but depends on spatial context.
In this section, we start by considering pairwise color contrast and simple spatial context.
Specifically, we introduce the Campbell-Robson curve, \cite{Campbell68,Oliva06}.
Based on this, we introduce perception of complex spatial contrast, specifically considering recent research from vision science \cite{Peli90,Haun13}.
Finally, we study the perception complex spatial color contrast as it appears in many real images.
\begin{figure}[tb]
\vspace{0pt}
\includegraphics[width=\linewidth]{Fig/Campbell.pdf}
\caption{Campbell-Robson curve.} Left, the original Campbell-Robson curve, brightness contrast sensitivity is a function of spatial frequency. Middle, Campbell-Robson curve on single color. We assume color brightness contrast follows the same curve. Right, curve on hue ring, the color hue contrast follows the same curve.
\label{Fig:Campbell}
\vspace{-12pt}
\end{figure}
\vspace{-6pt}
\paragraph{Peception of simple spatial contrast:}
As illustrated in \fref{Fig:Campbell}.a, the Campbell-Robson curve demonstrates that the perception of contrast is a function of spatial frequency. Specifically, human perception of contrast is most sensitive at 1-2 cycles per degree (cpd). The sensitivity drops for spatial frequency above and below the peak.
The Campbell-Robson study is only of luminance contrast. No equivalent study is available for the perception of color contrast.
Although in practice, colour contrast is likely to exhibit spatial sensitivities that differ from luminance contrast, we take the simplest assumption that the perception of color contrast follows the same curve.
This is shown in \fref{Fig:Campbell}.b for the perceptional sensitivity of color-brightness contrast with respect to spatial frequency, and in \fref{Fig:Campbell}.c, for the perception of hue contrast.
Our subsequent algorithm could be adapted if curves like that Campbell-Robson became available for color contrasat.
Thus,
\noindent\textbf{Assumption 1} the perception of color contrast (Brightness, Hue, Saturation) with respect to spatial frequency follows the same Campbell-Robson curve.
\begin{figure}[tb]
\vspace{-0pt}
\includegraphics[width=\linewidth]{Fig/Complex.pdf}
\caption{The Haun-Peli curve for perceiving complex contrast \cite{Haun13}. Left, two gray scale images with difference combinations of intensity of spatial frequency component. Users are asked to choose the one with higher contrast. Right, statistics of importance of different frequency components for the overall perceived contrast.
}
\label{Fig:Haun}
\vspace{-6pt}
\end{figure}
\begin{figure}[tb]
\includegraphics[width=\linewidth]{Fig/Contrast.pdf}
\vspace{-12pt}
\caption{Complex color contrast. (a) Color image, (b) b channel of L*a*b* spece. (c - g) contrast in different scale, using DoG filter. (h) Combined complex contrast according to Haun-Peli curve.
}
\label{Fig:Contrast}
\vspace{-12pt}
\end{figure}
\vspace{-6pt}
\paragraph{Perception of complex spatial contrast:}
However, the spatial contrast of a natural image is often far more complex that the Campbell-Robson pattern.
According to Haun \textit{et~al.}\cite{Haun13}, human perception of complex contrast follows a linear of combination of different spatial frequencies.
In \fref{Fig:Haun}, we include a typical result of such a combination. Note that the coefficients for such a combination need not all be positive.
\vspace{-12pt}
\paragraph{Perception of complex spatial color contrast:}
Unfortunately, no systematic vision science model is available for complex color contrast. Therefore, here again, we make an assumption that the perception of complex brightness contrast can be extended to complex color contrast:
\noindent\textbf{Assumption 2} the perception of complex color contrast with respect to spatial frequency follows the same Haun-Peli curve.
\vspace{-12pt}
\paragraph{Complex Color Contrast Consistency Measurement:}
According to the above assumptions, we introduce the Brightness Contrast Measurement. For a single channel (grayscale) image the complex contrast is defined as:
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
c(G) = \sum\beta_i k_i * G,
\label{Eq:CCM1}
\end{equation}
where $k_i$ is the Difference of Gaussian (DoG) Kernel at scale $2^i$:
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
k_i(j, k) = \frac{1}{2\pi 2^{2i}} e^{-\frac{j^2 + k^2}{2 ^{2i}}} - \frac{1}{2\pi 2^{2(i+1)}} e^{-\frac{1}{2}\frac{k^2 + k^2}{2^{2(i+1)}}}.
\label{Eq:CCM2}
\end{equation}
$\beta_i$ are the coefficients derived from the Haun-Peli curve. For example, a screen with 72 dots per inch (dpi), and viewed from 60cm away.
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
\beta_1 = -4, \beta_2 = 1, \beta_3 = 4, \beta_4 = 4, \beta_5 = 1, \beta_6 = -2.
\label{Eq:CCM3}
\end{equation}
When varying the dpi or the viewing distance, the scale by pixels shifts proportionally. \Fref{Fig:Contrast} shows an example of the contrast at different scales and the combined complex contrast.
We define the complex contrast for the color channels similarly and denote them as $c(L), c(A), c(B)$. The contrast measure is defined as:
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
E_c(G) = \alpha_L\parallel c(L) - c(G) \parallel_{\ell}
+ \alpha_{AB}\parallel c(A) - c(G) \parallel_{\ell}
+ \alpha_{AB}\parallel c(B) - c(G) \parallel_{\ell},
\label{Eq:CCM4}
\end{equation}
where $\alpha_L, \alpha_{AB}$ balances the importance between different channels.
As can be seen, the contrast energy functions has three terms that represent the contrast lost for the $l,a,b$ channels respectively. Note that the contrast loss for the $a,b$ channels are directional, this is because we incorporate the understanding of human vision that the brightness of same color is not always perceived in the same way \cite{Corney09}.
\section{Methodology}
In this section, we introduce a graphical model based method which aims to find a gray-scale image which best fits the brightness perceptual consistency as well as the contrast consistency.
\vspace{-6pt}
\paragraph{Objective Function}
Combining \eref{Eq:BCM1} and \eref{Eq:CCM4}, the proposed objective function seeks the gray-scale image $G$ that optimizes the trade-off between brightness consistency and contrast:
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
E(G) = E_B(G) + E_C(G)
\label{Eq:OBJ1}
\end{equation}
Similarly, \eref{Eq:OBJ1} can be represented as linear system:
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
\begin{split}
E(\bm{g}) &= E_B(\bm{d}) + E_C(\bm{g})
\\
&= \parallel \Lambda(\bm{w})(\bm{l} - \bm{g}) \parallel_{\ell}
\\
&~~~~+ \alpha_L\parallel C\bm{l} - C\bm{g} \parallel_{\ell}
+ \alpha_{AB}\parallel C\bm{a} - C\bm{g} \parallel_{\ell}
+ \alpha_{AB}\parallel C\bm{b} - C\bm{g} \parallel_{\ell},
\end{split}
\label{Eq:OBJ3}
\end{equation}
where the bold letters are the vectorized representation of the image matrix. $C$ is the complex contrast operator in matrix form. $\Lambda$ converts the coefficients array $\bm{w}$ to a diagonal matrix. A typical choise of channel importance is $\alpha_L = 0.5, \alpha_{AB} = 1.5$. We will fix such setting for all the experiments.
\vspace{-6pt}
\paragraph{$\ell_2$-norm Solution:}
When $\ell = 2$, \eref{Eq:OBJ3} takes a quadratic form, which is minimized when:
\begin{equation}
\begin{split}
\frac{\partial E(\bm{g})}{ \partial \bm{g} } & = 0
\\
& = -\Lambda(\bm{w})(\bm{l} - \bm{g})
\\
& ~~~~ -\alpha_L C^\top (C \bm{l} - C\bm{g})
-\alpha_{AB} C^\top (C \bm{a} - C\bm{g})
-\alpha_{AB} C^\top (C \bm{b} - C\bm{g}),
\end{split}
\label{Eq:SOL1}
\end{equation}
This can be written as:
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
M \bm{g} = \bm{u},
\label{Eq:SOL2}
\end{equation}
where:
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
M = \Lambda(\bm{w}) + (\alpha_L + 2\alpha_{AB})C^\top C,
\label{Eq:SOL3}
\end{equation}
and
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
\bm{u} = \Lambda(\bm{w}) \bm{l} + \alpha_L C^\top C \bm{l} + \alpha_{AB} C^\top C \bm{a} + \alpha_{AB} C^\top C \bm{b}.
\label{Eq:SOL4}
\end{equation}
Notice that $M$ is invertable because:
1. $\Lambda(\bm{w})$ is a positive definite diagonal matrix.
2. $C^\top C$ is semi-positive definite.
3. $\alpha_L, \alpha_{AB}$ are positive.
1, 2, 3 together infer that $M$ is a positive-definite symmetric matrix and is thus invertable \cite{Lang02}. Therefore, the grayscale image is found by:
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
\bm{g} = M^{-1} \bm{u}.
\label{Eq:SOL5}
\end{equation}
\paragraph{$\ell_1$-norm Solution:}
It is well known that $\ell_2$ norm based objective functions tend to over-smooth boundaries, this problem is significant for color-to-gray conversion.
This is because a grayscale image is intended to preserve opposing terms: brightness and multi-channel color contrast.
Unfortunately, it is unlikely to find a solution to suit the contrast for all three channels simultaneously.
The $\ell_1$ solution allows us to neglect a relatively insignificant channel and strongly emphasize the contrast of a more significant channel.
In addition we take advantage of using $\ell_1$ where discontinuity at boundaries is implicitly handled in the graphical model.
To solve the $\ell_1$ objective function, we use the Iteratively Re-weighted Least Square method \cite{Candes08}.
For $i$-th iteration, we solve a reweighted least square problem of \eref{Eq:OBJ3}:
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
\begin{split}
\min E(\bm{g}_i) &= \min \Lambda(\bm{q}_i)[E_B(\bm{d}) + E_C(\bm{g}_i)],
\end{split}
\label{Eq:SOL6}
\end{equation}
where the weight is defined as:
\vspace{-12pt}
\begin{equation}
\vspace{-6pt}
\bm{q}_{i+1} = \frac{1}{|\bm{r}_i| + 1},
\label{Eq:SOL7}
\end{equation}
$w_0$ is set to 1 for all the pixels, and
$\bm{r}$ is the residual of each iteration:
\vspace{-6pt}
\begin{equation}
\vspace{-6pt}
\bm{r}_i = \Lambda(\bm{q}_i)[| \Lambda(\bm{w})(\bm{l} - \bm{g}_i) | + \alpha_L | C\bm{l} - C\bm{g}_i |
+ \alpha_{AB} | C\bm{a} - C\bm{g}_i | + \alpha_{AB} | C\bm{b} - C\bm{g}_i |].
\label{Eq:SOL8}
\end{equation}
\vspace{-6pt}
\paragraph{Speedup:}
Constructing the complex contrast matrix $C$ will yield a large dense matrix with $N^2$ elements where $N$ is the number of pixels, as a large scale DoG compares the difference over a large range. Fortunately, we find that the blurring matrix $C$ can be pre-computed using decomposition of the Gaussian kernel and the construction of an image pyramid. Details of the speedup can be found in the supplementary material.
\begin{figure}[tb]
\includegraphics[width=\linewidth]{Fig/VisualCadik.pdf}
\vspace{-24pt}
\caption{Visual results on {\^C}ad{\'\i}k~ dataset. (Optimized for A4 paper viewed from 40cm away).}
\label{Fig:VisualCadik}
\end{figure}
\begin{figure}[tb]
\vspace{-6pt}
\includegraphics[width=\linewidth]{Fig/Visual250.pdf}
\vspace{-24pt}
\caption{Visual results on Color250 dataset. (Optimized for A4 paper viewed from 40cm away).}
\label{Fig:Visual250}
\vspace{-12pt}
\end{figure}
\begin{figure}[htbp]
\includegraphics[width=\linewidth]{Fig/VisualNeo.pdf}
\caption{Visual results on NeoColor dataset. (Optimized for A4 paper viewed from 40cm away).}
\label{Fig:VisualNeo}
\end{figure}
\section{Experiments}
We validate the proposed method systematically.
We compare it with methods for which code are available, on existing datasets as well as the new proposed dataset.
We evaluate qualitative performance, as well as performance with respect to metrics.
Further, we include a user study to validate the method and the proposed metric.
\subsection{Datasets}
\vspace{-6pt}
{\em {\^C}ad{\'\i}k~}
\cite{Cadik08} is the earliest dataset for C2G evaluation, which contains 25 highly saturated images. The contents varies from geometric pattern to real scenes.
Because the {\^C}ad{\'\i}k~ dataset has only 25 images, as is common practice, we evaluate the performance on the {\em Color250} dataset\cite{Lu14}.
This data set is a subset of the 2001 Berkeley Segmentation Dataset \cite{Martin01}.
{\em CSDD}
\cite{Du15} contains 22 highly saturated and fully chromatic geometric patterns. Because the number of images is limited and no natural image are included, we do not use it in our evaluation. However, we include the results on the CSDD dataset in the supplementary material.
\vspace{-6pt}
\paragraph{NeoColor}
The Berkeley Segmentation Dataset was designed for evaluating segmentation performance rather than evaluating C2G algorithms.
Images in the dataset tend to have strong contrast on object boundaries, only contain a limited number of segments with different colors, and the contents are relatively simple.
On the other hand, C2G algorithms must be able to deal with more complex images that frequently arise.
To fully evaluate the performance of C2G algorithms, we have collected high quality digital images of natural scenes, advertisements, computer graphic designs as well as fine art.
We consider that this new dataset, NeoColor, extends upon the existing datasets by including images of greater color complexity.
A full collection of the dataset appears in the supplementary material.
Some examples are shown in \fref{Fig:VisualNeo}.
\vspace{-6pt}
\subsection{Evaluation metric}
We evaluate our methods using {\em EScore} proposed by Lu \textit{et~al.}
\cite{Lu14} which has a joint evaluation between color contrast preservation and color fidelity.
However, the metric emphasizes preservation of high contrast, and measurement of spatial contrast preservation is based on random sampling which does not fully reflect the concept on visual perception preservation.
The {\em QScore} Ma \textit{et~al.}
\cite{Ma15}, has better quantization for color contrast measurement, however, scale dependent contrast is not emphasized.
\vspace{-9pt}
\paragraph{Visual Perception Metric (PScore)}
Because both metrics focus more on contrast preservation rather than the overall perception of color, we provide further evaluation using the proposed perceptually consistent color energy function \eref{Eq:OBJ1}. For a fair comparison, we weight all scales of contrast preservation separately, we set $\beta_i = 1$ (\eref{Eq:CCM1}) for the evaluated scale and rest as 0.
\vspace{-6pt}
\subsection{Results}
\vspace{-6pt}
We compare with classic methods of Gooch \textit{et~al.} \cite{Gooch05} and Grundland \textit{et~al.} \cite{Grundland07} as well as the state-of-the-art methods of Lu \textit{et~al.} \cite{Lu14} and Du \textit{et~al.} \cite{Du15}. We use the default setting for all the datasets. For our method, scale is set to optimize contrast for A4 paper viewed from 40cm away, and all parameters are fixed for all the experiments.
Code is not available\footnote{Not available on webpage as well as through email contact} for the methods of Liu \textit{et~al.} \cite{Liu15} and Ji \textit{et~al.} \cite{Ji15} and thus they are not evaluated.
We report both visual and quantitative results.
\vspace{-6pt}
\paragraph{Visual results}
Visual results on the {\^C}ad{\'\i}k~ dataset, the Color250 dataset and NeoColor dataset are shown in \fref{Fig:VisualCadik}, \fref{Fig:Visual250} and \fref{Fig:VisualNeo} respectively.
A full collection of the results can be found in the supplementary material.
As can be seen, methods that do not consider perceptual consistency are likely to produce strong contrast but unnatural images.
For example in \fref{Fig:VisualNeo}, black orange juice, a rainbow with black and white stripes, and a black sun in a sunset image, all lose their natural impression.
Further, the impressionist painting "Lotus" which now has more emphasis on weeds has lost the original feeling of the image, and the intention of the artist.
On the contrary, the proposed methods allows a reasonable sense of color contrast as well as the original feeling of the images.
\begin{figure}[tb]
\includegraphics[width=\linewidth]{Fig/EScore.pdf}
\vspace{-24pt}
\caption{Quantitative evaluation using the EScore \cite{Lu14}.} The y-axis the Escore, where as the x-axis the contrast threshold level.
\label{Fig:EScore}
\vspace{-12pt}
\end{figure}
\begin{figure}[tb]
\includegraphics[width=\linewidth]{Fig/QScore.pdf}
\vspace{-24pt}
\caption{Quantitative evaluation using the QScore \cite{Ma15}. Higher score is better. Average score and standard variance is shown.}
\vspace{-0pt}
\label{Fig:QScore}
\end{figure}
\begin{figure}[tb]
\vspace{-6pt}
\includegraphics[width=\linewidth]{Fig/NScore.pdf}
\vspace{-24pt}
\caption{Quantitative evaluation using proposed PScore. The y-axis is the PScore. The x-axis represents the natural contrast preservation at scale $2 ^ {\frac{i + 2}{2}}$ pixels.}
\label{Fig:PScore}
\vspace{-12pt}
\end{figure}
\vspace{-6pt}
\paragraph{Quantitative results:}
Quantitative evaluation using the E-score is
shown in \fref{Fig:EScore}, a higher curve indicates better performance. As can be seen, the {\^C}ad{\'\i}k~ dataset is not large enough to produce a smooth curve. On the Color250 dataset, our method performs well on contrast in the range 0-20 (where total black to white was a scale of 100).
The Q-score proposed by Ma \textit{et~al.} \cite{Ma15} is shown in \fref{Fig:QScore}.
Whereas the evaluation for most C2G algorithms was reasonable, CIE Y which does not consider color contrast is evaluated as the best on all the datasets.
Results for consistency (P-Score) are shown in \fref{Fig:PScore}.
Again, a higher curve shows better performance.
The x-axis represents contrast preservation at scale $2 ^ {\frac{i + 2}{2}}$ pixels.
We plot the contrast preservation for different scales. As can be seen, the proposed method best preserves the color contrast at the scale which is most sensitive for perception. Whereas for the scales which are less contributive to overall contrast, the proposed method has lower contrast preservation in comparing with other methods.
The ordering is accordance with the feeling on visual results.
\vspace{-6pt}
\subsection{User Study}
Ten volunteers (aged 22-41) with normal or corrected-to-normal vision took part, with data collection in March 2016. Written consent was obtained from participants before they began the experiment. The research was approved by the {\em institution removed} Ethics Committee, and adhered to the tenets of the Declaration of Helsinki.
Considering the complexity of the task, four methods were compared: the proposed, CIE Y (as a widely used method), Lu \textit{et~al.} (as the state-of-the-art), Grundland \textit{et~al.} (as reported relatively worse from the quantitative metric).
Participants were masked to the algorithm.
The color image was presented in the middle of the screen, with a white background, and the four grayscale images presented at identical distances from the original image. Please refer to supplementary material for screen layout and experimental details.
We selected 50 hard images from Color250, and 200 images from NeoColor, excluding more difficult images. Then, 150 images were randomly selected from these subsets with a 30\% selection probability from Color250 and a 70\% probability from NeoColor.
Each user was shown these images in random order, and placement of the grayscale images was randomly assigned for each trial.
Participants were required to rank the grayscale images from best (1) to worst (4) according to the instruction: ``Which black-and-white image best represents the original color image". When answering, the participants were asked to consider: which grayscale image best retains the contrast between colors of the original color image; and, which grayscale image best retains the overall sense or feeling of the original color image.
The chance rate of ranking a image as "best" (1 and 2) representation of the color image was 50\%, and 75\% was the criterion set as a benchmark as a reliable grayscale representation of the color image (\textit{i}.\textit{e}., participants ranking a Color2Gray vision processing method ``1" or ``2" for 75\% of the trials).
Descriptive analyses were used to characterize the counts and percentages of responses. Comparisons between participants and color-to-gray methods for the average rankings of the preferred images were calculated using non-parametric statistics (e.g., Kruskal-Wallis H test). Windows SPSS v23 (IBM Corporation, Somers, NY) was used for all statistical analyses.
\vspace{-6pt}
\paragraph{Results:}
The data for ten participants was pooled for further analyses as there was no significant difference in overall preferences between participants ($p < 1.00$).
Overall preference is shown in \fref{Fig:UserStudy}.
Our proposed C2G method was ranked as a significantly ($p<0.0001$) better representation of the original color images compared to the Grundland and Lu methods. However, the CIE-Y approach was selected significantly ($p<0.0001$) more often compared to the proposed method.
Our proposed method was ranked as being a better (ranked 1 or 2) representation of the original color image for 64.5\% of responses, and CIE-Y achieved 75.0\% of responses ranked as 1 or 2.
The CIE-Y approach is similar to approaches that are ubiquitous in print media (\textit{e}.\textit{g}., newspapers, kindle).
Among the specialized C2G algorithms, our proposed method was highest ranked.
The Grundland and Lu methods were consistently ranked as the least preferred methods with only 12.8\% and 47.7\% of responses ranked 1 or 2.
\begin{figure}[tb]
\includegraphics[width=\linewidth]{Fig/UserStudy.pdf}
\vspace{-18pt}
\caption{User preference on C2G algorithms.} Left, statics of users' ranking by number of choices. Parenthesis list the percentage. Right, ranking frequency distribution of all the methods.
\vspace{-12pt}
\label{Fig:UserStudy}
\end{figure}
\section{Conclusion}
\vspace{-12pt}
In this paper, we proposed a method for converting color images to gray scale while preserving brightness consistency as much as possible.
The key ideas of brightness consistency is that the feeling and naturalness of the original color is preserved, and that color contrast is preserved.
These key ideas were used to derive quantitative metrics based on recent studies in vision science.
An $\ell_1$ optimization framework was proposed to find the grayscale image which optimised the proposed brightness consistency metric.
To evaluate the proposed methods, we used both existing datasets and a proposed new dataset.
We also validated both the algorithm and the metric with a user study.
For future work, we will exploit brightness consistency with more emphasis on visual augmentation.
\clearpage
\bibliographystyle{splncs}
|
\section{INTRODUCTION}
\label{sec:intro}
Hot exoplanets, i.e., exoplanets with orbital periods of approximately a few days, are unique
laboratories for the study of star-planet interactions (hereafter SPIs). The large amounts of
stellar UV and EUV radiation that these planets receive heats the planetary atmosphere and can
result in significant photoevaporative mass loss across a broad range of planetary masses
\citep[e.g.][]{vidal,murray09,desetangs,owen12,lopez13,ehren15}. If the mass loss is significant and
the planet overflows its Roche lobe, the outflowing planetary material can accrete onto the star
\citep{lai,matsakos,pillitteri15}. The short orbital distances of these planets can also result in
large tidal interactions, generating enhanced activity levels on the star that are directly related
to the orbital phase of the planet \citep{cuntz,shkolnik08,pillitteri15}. Direct magnetic
interactions via reconnection events or magnetic torques are also possible and can affect the
orbital and spin evolution of the planet and star \citep{lanza10,cohen10,strugarek}. Understanding
these interactions is important in order to move towards a full characterization of exoplanets and a
complete knowledge of their evolution under extreme conditions.
Two particularly interesting classes of SPIs are the direct interaction of escaping planetary
material (i.e., the planetary wind) and the stellar wind and the interaction of the planet's
magnetosphere with the stellar wind. Both scenarios are capable of producing bow shocks at the
interaction region which results in a significant density enhancement of the shocked material
\citep{vidotto10,bisikalo13,matsakos}. For parameters relevant to most hot planet systems, the
interaction region in both cases occurs ahead of the planet in its orbit. Observationally this can
produce a pre-transit signature (for transiting or very nearly transiting planets) if the material
ahead of the planet is sufficiently opaque to produce measurable absorption
\citep{vidotto10,vidotto11,llama11}. In the case of a bow shock mediated by the planet's
magnetosphere, it is possible to estimate the strength of the field given some assumptions about the
stellar wind \citep{vidotto10,llama13,cauley15}. While estimates of this sort are rough, this method
of measuring exoplanet magnetic fields may be worth pursuing considering the challenges of directly
detecting emission from electrons in the planetary magnetosphere and the current lack of such
confirmed detections \citep[e.g.,][]{murphy15,vidotto15}.
A third scenario capable of producing pre-transit absorption is the loss of planetary material that
then accretes onto the central star. Accretion streams or time-variable blobs are predicted from
some simulations of hot planet mass loss \citep{cohen11,matsakos} and have been suggested as the cause of
observed UV flares for HD 189733 b \citep{pillitteri15}. Accretion streams or blobs can be viewed in
transmission if the optical depth is sufficiently high in the transition of interest.
Besides material transiting ahead of the planet, changes in the stellar activity level can be also
produce observable changes in the pre-transit transmission spectrum: if spectra from a period of low
activity are compared to spectra from a period of higher activity, the relative difference can
produce features similar to what is seen in tranmission spectrum absorption lines. HD 189733 is an
active early K-star \citep{boisse09} and, especially for the Balmer line analysis presented here,
contributions from the stellar chromosphere may be non-negligible. These contributions need to be
taken into account when interpreting transmission spectra of lines produced in stellar active
regions.
Evidence for pre-transit material has now been observed in a handful of hot planet systems, namely
WASP-12 b \citep{fossati}, HD 189733 b \citep{benjaffel,bourrier13,cauley15}, and GJ 436 b
\citep{ehren15} \citep[see Section 1 of][for a brief overview]{cauley15}. Most recently,
\citet{ehren15} reported a large pre-transit absorption signature, as well as enhanced in-transit
absorption, in Lyman-$\alpha$ around the hot Neptune GJ 436 b. They model the early ingress as an
extended cloud of hydrogen which has escaped from the planet. The escaping material is subject
to radiation pressure from the star, which in the case of GJ 436 is too weak to overcome the
stellar gravity and instead acts as a radiative braking agent on the gas \citep{bourrier15a}.
We note that \citep{ehren15} do not include the interaction of the
planetary wind with the stellar wind. In \defcitealias{cauley15}{Paper I}\citet[][hereafter Paper
I]{cauley15} we describe a pre-transit absorption measurement in the Balmer lines for the hot
Jupiter HD 189733 b \citep{bouchy}. We showed that the absorption strength and time series evolution
of the absorption is consistent with the geometry of a thin bow shock at $\sim$13 $R_p$ ahead of the
planet. There are difficulties, however, in maintaining a population of hot neutral hydrogen at
large distances from the planet (Huang \& Christie, private communication). Furthermore, it is
unlikely that the densities produced by a compressed stellar wind at typical hot Jupiter orbital
distances are high enough in most atomic species to generate the necessary opacity, although
escaping planetary material trapped in the magnetosphere may be sufficient \citep{turner16}.
Regardless of their precise interpretation, it appears that pre-transit absorption signals should be
fairly common for hot planets. However, detections of these phenomena seem to be limited to specific
atomic transitions; they have not yet been detected using near-UV broadband methods
\citep[e.g.,][]{turner13,bento14,zellem15} nor, to the best of our knowledge, optical photometry.
High-resolution spectroscopic observations are expensive and are limited to the brightest transiting
targets. As a result, the detection and characterization of these signals, including their
variability, will require a significant observational investment.
In this paper we present followup observations to the 2013 transit of HD 189733 b presented in
\citetalias{cauley15}. Our observations of the 2013 transit missed a large portion of the
pre-transit phase due to an observing strategy that was not designed to monitor pre-transit signals.
The 2015 transit observations provide complete coverage of the $\sim$4 hours of visible pre-transit
phase. The instrumental setup and experimental design are almost identical to the 2013 transit,
facilitating a direct comparison between the two measurement sets. In \autoref{sec:observations} we
briefly describe the transit observations and data reduction. \autoref{sec:transspec} details the
construction of the Balmer line, \ion{Na}{1} D lines, and \ion{Mg}{1} 5184 \AA\ transmission spectra
and the line absorption time series for each line. We present an analysis of the \ion{Ca}{2} H and K
lines, which serve as a proxy of the stellar activity level, in \autoref{sec:calcium2}. The
residual core flux analysis in \autoref{sec:calcium2} differs from the Mt. Wilson $S_{HK}$ analysis
presented in \citetalias{cauley15} and provides a more precise estimate of the relative stellar
activity level. Contrary to the $S_{HK}$ index analysis presented in \citetalias{cauley15}, the core
flux analysis shows that there is some relationship between the stellar activity level and the
measured absorption. We present possible model scenarios to describe the new pre-transit absorption
in \autoref{sec:model}. We will present a detailed analysis of the in-transit absorption in a
forthcoming paper. In \autoref{sec:summary} we give a general summary of the results and comment on
future work that is required to clarify the nature of the pre-transit signature.
\section{Observations and data reduction}
\label{sec:observations}
A single transit epoch of HD 189733 b was obtained on the night of August, 4 2015 using HiRES on
Keck I \citep{vogt}. The instrument setup and experimental design were identical to the 2013 transit
observations described in \citetalias{cauley15}. We briefly review them here. The B2 decker was used
and the resolving power of the observations is $R$$\sim$68,000 at H$\alpha$. Exposure times were 5
minutes for all observations resulting in an average signal-to-noise of 450 for the extracted
spectra at H$\alpha$, 200 at H$\beta$, and 150 at H$\gamma$. The data were reduced using the HiRES
Redux package written by Jason X. Prochaska\footnote{http://www.ucolick.org/$\sim$xavier/HIRedux/}.
All standard reduction steps were taken and the barycentric velocity of the observatory is removed,
along with the radial velocity of the HD 189733 system which we take to be $-2.24$ km s$^{-1}$
\citep{digloria15}, leaving the spectra in the rest frame of the star. Wavelength solutions are
performed on Th-Ar exposures taken at the beginning and end of the night. Telluric absorption is
removed from the H$\alpha$ and \ion{Na}{1} spectra using the telluric fitting program Molecfit
\citep{kausch}. A telluric standard was observed at the beginning of the night. The telluric model
is first applied to the standard and then scaled and shifted to fit each individual observation
using a least-squares minimization routine.
\section{Transmission spectra}
\label{sec:transspec}
The transmission spectrum is defined here as
\begin{equation}\label{eq:strans}
S_T=\frac{F_{i}}{F_{out}}-1
\end{equation}
\noindent where $F_i$ is a single observation and $F_{out}$ is the master comparison spectrum. The
master comparison spectrum is a weighted average of a set of spectra that are, ideally, free of any
absorption contamination, i.e., they represent the pure stellar spectrum. In order for each $F_i$ to
be compared to $F_{out}$, the spectra need to be normalized and aligned in wavelength space. The
required constant wavelength shifts from one observation to the next, which are calculated by
cross-correlating a large number of metal lines in the order of interest, are of the order 0.01-0.02
\AA\ or 0.5-1.0 km s$^{-1}$. The corrections applied to the H$\alpha$ order, relative to the first
observed spectrum, are shown in \autoref{fig:halphavcorrs}. Higher order functions, e.g., linear or
spline fits, are not necessary to align the spectra. After the spectra are aligned and divided by
$F_{out}$ they are renormalized to remove any residual slope in the continuum.
\begin{figure}[htbp]
\centering
\includegraphics[scale=.5,clip,trim=50mm 25mm 50mm 40mm,angle=0]{halpha_linecorr_20150804.pdf}
\figcaption{Velocity corrections applied to the H$\alpha$ order, relative to the first spectrum (not shown) for each observation.
The comparison spectrum corrections are shown with red bow ties.
The abrupt changes at $-200$ and $-160$ minutes are due to small guiding errors that resulted in the telescope
being re-centered on the target.\label{fig:halphavcorrs}}
\end{figure}
We note that HiRES is a very stable spectrograph and does not
produce large instrumental effects on the observed spectrum. This is important for having confidence
in the transmission spectrum signal as being due to real changes in the star-planet system and not
from instrument and telescope systematics. Furthermore, any systematic shifting of the spectrum on
the detector or changes in resolution would uniformly affect all stellar lines in a single order. This is
never observed: all of the transmission line profiles reported here occur only in the cores of the stellar lines of
interest and not in the plethora of other stellar lines present in every order. Thus we are confident all of the
reported transmission signals are intrinsic to the star-planet system.
In order to create the transmission spectrum, the comparison spectra used to generate the master
comparison spectrum must be selected. Since all of the
absorption is measured relative to the master comparison spectrum, choosing different spectra
will increase or decrease the measured absorption values. Thus the true zero-point is difficult to
determine for the small window of time that a single night of observing provides. We have chosen to use
the eight spectra from $t-t_{mid}=-278$ minutes to $t-t_{mid}=-233$ as the comparison spectra for the analysis
presented below in the rest of \autoref{sec:transspec}. These spectra are chosen since they are the earliest spectra
obtained during the night and are farthest from the planet transit in time. If pre-transit absorption is due to
extended structures around the planet, the farther the observations occur from the planet transit the less likely
those observations are to contain a signal from optically thick circumplanetary material. Of course, this does not
guarantee that the spectra are uncontaminated by such material. Our choice of comparison spectra and the
consequences for interpreting the absorption measurements will be discussed in \autoref{sec:compspec}.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.68,clip,trim=5mm 25mm 5mm 30mm,angle=0]{preinpost_nomod_20150804.pdf}
\figcaption{The averaged transmission spectra for the pre-, in-, and post-transit spectra for H$\alpha$ (top row), H$\beta$
(middle row), and H$\gamma$ (bottom row). The cores of identified stellar lines other than the lines of interest are
masked in gray. These points are not included in the calculation of the absorbed flux. The number of spectra used
to create the average spectrum is listed in the first three panels of the top row. The fourth frame in each
row shows the in--transit empirical Monte Carlo (EMC) $W_\lambda$ distributions for each line. The master
absorption measurement is marked with a vertical dashed red line. The uncertainty in W$_\lambda$ derived from
the In-In EMC procedure is marked in black in the upper-left; the propagated flux uncertainty is labeled in orange.
The larger of the two is adopted for determining the detection significance. Each in-transit absorption measurement is
detected above the 3$\sigma$ level. The H$\alpha$ and H$\beta$ pre-transit measurements are detected at
$>3\sigma$.\label{fig:tspecs_all}}
\end{figure*}
\begin{figure*}[h]
\centering
\includegraphics[scale=.68,clip,trim=8mm 55mm 5mm 40mm,angle=0]{preinpost_z_20150804.pdf}
\figcaption{The averaged transmission spectra for the pre-, in-, and post-transit spectra for \ion{Na}{1} 5895.9 \AA\ (top row) and
\ion{Mg}{1} 5183.6 \AA\ (bottom row). All figure markings and colors
are the same as \autoref{fig:tspecs_all}. The \ion{Na}{1} in-transit absorption is measured at 4.4$\sigma$ while
the \ion{Mg}{1} absorption is marginal at 2.4$\sigma$. Neither line is detected in the pre-transit transmission spectra.\label{fig:tspecs_z}}
\end{figure*}
\subsection{Average transmission spectra}
\label{sec:ave_tspecs}
The average transmission spectra are shown in \autoref{fig:tspecs_all} for H$\alpha$ 6562.79 \AA\ (first row),
H$\beta$ 4861.35 \AA\ (second row), and H$\gamma$ 4340.47 \AA\ (third row)
for the pre-transit (first column), in-transit (second column), and out-of-transit comparison
observations (third column). The number of spectra used to compute the average spectrum is shown in the
bottom right corner of each panel in the first row. The fourth column shows the empirical Monte Carlo (EMC) distributions,
which highlight the influence of systematic effects on the absorption measurements, of the
absorption for each line. The fifth column shows the EMC for the pre-transit signal. \autoref{fig:tspecs_z} shows
the same thing for the \ion{Na}{1} 5895.92 \AA\ and \ion{Mg}{1} 5183.60 \AA\ lines. Due to its location at the very edge of the order, we do not
extract and analyze the \ion{Na}{1} 5889 \AA\ line.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.75,clip,trim=25mm 25mm 0mm 40mm,angle=0]{preinpost_ctrl_20150804.pdf}
\figcaption{The averaged transmission spectra for the pre-, in-, and post-transit spectra for the \ion{Ca}{1} control
lines in each of the Balmer line orders. None of the control lines are detected at a significant level.The scale
of the EMC panels is the same as those in \autoref{fig:tspecs_all} to facilitate
a direct comparison of the distributions. All of the colors and labels are the same as those in \autoref{fig:tspecs_all}.\label{fig:tspecs_ctrl}}
\end{figure*}
We perform two different EMC procedures: an In-In and In-Out procedure \citep[see][for similar applications]{redfield,jensen12,wyttenbach,cauley15}.
The Pre-Pre and Pre-Out procedures are identical but for the pre-transit spectra.
The In-Out EMC compares randomly selected subsets of the in-transit spectra to the master comparison
spectrum. The number of random spectra varies from $N$=5 to $N$=18 and the process is repeated
for $N$=5000 iterations. The In-In procedure compares a random subset of in-transit spectra with
another random subset of in-transit spectra. The random subsets are chosen without replacement
so that spectra are never compared to themselves and the number of random spectra varies from
$N$=2 to $N$=16. The In-In distributions should be centered at zero absorption and should be
broader than the In-Out distributions due to the effect of comparing two different random subsets
instead of one random subset with a fixed comparison spectrum. We choose the standard deviation of the In-Out distribution
as the 1$\sigma$ uncertainty associated with the EMC procedure since this is directly related to
the variation in the absorption measurement.
As expected, the In-In distributions (green) are centered near zero absorption.
The 1$\sigma$ values from the In-Out EMC procedure are shown in black. We also calculate the propagated
flux uncertainty associated with the master absorption measurement (vertical red dashed line).
The flux uncertainty is the standard deviation of two 50 km s$^{-1}$ wide portions of the transmission spectrum
from $-250$-$-200$ km s$^{-1}$ and $+200$-$+250$ km s$^{-1}$ weighted by the square root
of the normalized observed spectrum in order to account for the larger uncertainties in the line cores.
This value is shown in orange. The larger of the EMC and flux uncertainty measurements is used
for determining the significance of the absorption. The significance level $n\sigma$ is shown in red
next to the master absorption measurement.
The absorption is the negative of the equivalent width of the transmission spectrum integrated
from $-200$ km s$^{-1}$ to $+200$ km s$^{-1}$. More formally,
\begin{equation}\label{eq:wlambda}
W_\lambda = \sum\limits_{v=-200}^{+200} \left(1-\frac{F_v}{F_v^{out}} \right) \Delta\lambda_v
\end{equation}
\noindent where $F_v$ is the flux in the spectrum of interest at velocity $v$, $F_v^{out}$ is the flux in the
comparison spectrum at velocity $v$, and $\Delta\lambda_v$ is the wavelength difference at velocity
$v$. The gray regions in \autoref{fig:tspecs_all} indicate the cores of identified stellar absorption lines other than
the line of interest. These regions are ignored when calculating $W_\lambda$. The master absorption
detections are presented in \autoref{tab:tab1}.
The in-transit and pre-transit absorption for H$\alpha$ and H$\beta$ is detected at $>3\sigma$, confirming the
detections made in \citetalias{cauley15} and the H$\alpha$ detection from \citet{jensen12}. The level of in-transit
H$\alpha$ absorption is almost identical to that from \citet{jensen12} ($W_{H\alpha}$=13.7$\times$10$^{-3}$ \AA)
and 2.3$\times$ the value measured in \citetalias{cauley15} ($W_{H\alpha}$=5.9$\times$10$^{-3}$ \AA). The in-transit
H$\gamma$ absorption is detected at 4.0$\sigma$; the measurement is marginal (2.2$\sigma$) for the pre-transit
absorption.
The in-transit \ion{Na}{1} absorption is detected at 4.4$\sigma$ while the \ion{Mg}{1} absorption is marginal
at 2.4$\sigma$. There is little evidence for pre-transit absorption in either line. In order to compare
our \ion{Na}{1} 5896 \AA\ in-transit absorption value to previously reported high resolution \ion{Na}{1} absorption, we have
applied \autoref{eq:wlambda} to the \ion{Na}{1} 5896 \AA\ transmission spectra presented in \citet{redfield},
\citet{jensen11}, and \citet{wyttenbach}. In the case of \citet{wyttenbach}, we have used their reported Gaussian fit parameters
for the 5896 \AA\ line in order to approximate the absorption. For \citet{jensen11} we find $W_{NaI}$=2.85$\pm$0.85$\times10^{-3}$ \AA;
for \citet{redfield}, $W_{NaI}$=3.59$\pm$1.61$\times$10$^{-3}$ \AA; for \citet{wyttenbach}, $W_{NaI}$=2.22$\times$10$^{-3}$ \AA.
Our reported value of $W_{NaI}=$1.50$\pm$0.34$\times$10$^{-3}$ \AA\ is consistently lower than the previous measurements.
Although the \ion{Mg}{1} 5184 \AA\ in-transit absorption is marginal, the measurement is suggestive of a real signal.
The \ion{Mg}{1} UV transition at 2853 \AA\ has been detected previously for HD 209458 b by \citet{vidal13}, providing
evidence that \ion{Mg}{1} is present at detectable levels in hot Jupiter atmospheres. The
HD 209458 b signal is much stronger (7.5\%) than that reported here.
Although the \ion{Mg}{1} 5184 \AA\ line arises from an excited \ion{Mg}{1} state and thus the
line strength is not favorable compared to the UV resonance line, future investigations should consider examining the
\ion{Mg}{1} 5184 \AA\ line due to its ease of observability compared to the UV \ion{Mg}{1} lines \citep[see ][]{bourrier15}
and to confirm or reject the measurement presented here.
\begin{deluxetable}{lccccc}
\tablecaption{Master absorption line detections \label{tab:tab1}}
\tablehead{\colhead{}&\multicolumn{2}{c}{Pre-transit}&\colhead{}&\multicolumn{2}{c}{In-transit}\\
\cline{2-3}\cline{5-6}\\
\colhead{}&\colhead{W$_\lambda$}&\colhead{$\sigma$}&\colhead{}&\colhead{W$_\lambda$}&\colhead{$\sigma$}\\
\colhead{Spectral line}&\colhead{(10$^{-3}$ \AA)}&\colhead{(10$^{-3}$ \AA)}&\colhead{}&\colhead{(10$^{-3}$ \AA)}&\colhead{(10$^{-3}$ \AA)}\\
\colhead{(1)}&\colhead{(2)}&\colhead{(3)}&\colhead{}&\colhead{(4)}&\colhead{(5)}}
\tabletypesize{\scriptsize}
\startdata
H$\alpha$ & 5.25 & 0.75 & & 13.44 & 0.56 \\
H$\beta$ & 2.16 & 0.40 & & 4.17 & 0.49\\
H$\gamma$ & 0.97 & 0.44 & & 1.96 & 0.49 \\
\ion{Na}{1} 5896 & 0.34 & 0.31 & & 1.50 & 0.34 \\
\ion{Mg}{1} 5184 & 0.67 & 0.42 & & 1.15 & 0.48 \\
& & & & & \\
\ion{Ca}{1} 6439.1 & \nodata & \nodata & & 0.02 & 0.49 \\
\ion{Ca}{1} 4878.1 & \nodata & \nodata & & 0.40 & 0.33 \\
\ion{Ca}{1} 4355.1 & \nodata & \nodata & & $-$0.46 & 0.76 \\
\enddata
\end{deluxetable}
\subsection{\ion{Ca}{1} control lines}
\label{sec:cai_control}
\autoref{fig:tspecs_ctrl} shows a set of \ion{Ca}{1} control lines that are used to verify the transmission
spectrum analysis. Calcium is expected to condense out of hot Jupiter atmospheres and so should
not be present in any significant quantity \citep{lodders}. Each of the \ion{Ca}{1} lines is located in the
same order as the Balmer line closest to it in wavelength. \autoref{fig:tspecs_ctrl} is identical to
\autoref{fig:tspecs_all} except we do not perform a pre-transit EMC for the control lines. Using the
flux errors as the 1$\sigma$ uncertainties, none of the control lines are measured at a significant
level (see \autoref{tab:tab1}). The non-detections in the control lines provide reassurance that the
measurement algorithms are being properly applied and not resulting in spurious absorption
signatures.
\subsection{Absorption time series}
\label{sec:indtranspec}
Transmission spectra for individual exposures are shown in \autoref{sec:tspecs_app}.
The values of W$_\lambda$ for the individual spectra are shown as a time series in \autoref{fig:mabs_all}, \autoref{fig:mabs_nai},
and \autoref{fig:mabs_mgi}. Each of the values is calculated using \autoref{eq:wlambda}. For all measurements, we choose to use the more
conservative average flux uncertainties for the individual points rather than the standard deviation of the EMC
procedure. In \autoref{fig:mabs_all}, the mean uncertainty for the
individual measurements is shown with the solid colored bars in the upper left. Individual $1\sigma$ uncertainties
are shown for \ion{Na}{1} and \ion{Mg}{1} in \autoref{fig:mabs_nai} and \autoref{fig:mabs_mgi}.
The pre-transit Balmer line signature in \autoref{fig:mabs_all} begins abruptly at $-217$ minutes, which corresponds to a linear distance of
$\sim$17 R$_p$ to the stellar limb, and lasts for $\sim$70 minutes until it abruptly disappears. It then returns to a
similar level after $\sim$40 minutes and remains relatively constant until immediately before first contact. The in-transit absorption appears to begin
immediately before first contact and increases between first and second contact. The abrupt changes in the level
of in-transit absorption, e.g., at $-30$, 5, and 40 minutes, may be due to the planet transiting the inhomogeneous
stellar surface. \added{The first event begins at $\sim-$70 minutes, or $\sim$15 minutes before $t_I$. This suggests
that the atmosphere is occulting an extended prominence above the stellar disk. This event ends at $t_{II}$, immediately after
the planetary disk fully occults the star. The second event, from $\sim+$5 to $+40$ minutes, has a duration similar
to the first event. The duration of the events correspond to stellar features with a linear extent of $\sim$2 $R_p$. These
abrupt changes in the absorbed flux could also be due to changes in the stellar activity level rather than transits
of static features. While this cannot be definitively ruled out, we believe this is less likely due to the very similar duration
of both events.} The effect of active regions and changes in the stellar activity level on the line absorption
will be discussed more fully in \autoref{sec:activityimpact}.
An interesting feature of the 2015 transit is that the absorption appears to persist immediately post-transit for $\sim$30 minutes. The subsequent decrease
in absorption at $\sim$90 minutes, and appearance of emission in the line cores of H$\beta$ and H$\gamma$, may be due to a change
in the stellar activity level, e.g., a mild flaring event. After the sharp increase, the final two points in the time series
decrease from the maximum. The linear extent of the feature causing the post-transit absorption is $>$3 $R_p$
if we take the beginning of the small flaring event to be the end of the post-transit absorption. The depth of
the measured post-transit absorption indicates a disk coverage fraction larger than the nominal extended atmosphere
we use to model the in-transit absorption in \autoref{sec:model} since the opaque planet is no longer blocking the
stellar disk. The post-transit absorption may be the base of the extended evaporative flows previously detected
around HD 189733 b in neutral hydrogen \citep{desetangs,desetangs12,bourrier13}.
Only one individual $W_\lambda$ value at $t-t_{mid}$=$-48$ minutes, is detected at $>$3$\sigma$ for \ion{Na}{1}.
None of the individual absorption values for \ion{Mg}{1} is detected at $>2\sigma$. There is a notable decrease, however,
in the \ion{Na}{1} in-transit points compared to the pre- and post-transit values and the \ion{Mg}{1} values are
consistently below zero beginning at $\sim$$-255$ minutes. The structure of the in-transit \ion{Na}{1} absorption
shows a similar shape to the center-to-limb variations (CLVs) described by \citet{czesla15}. CLVs may be
important in determining the precise level of atmospheric \ion{Na}{1} absorption and should be taken
into account for a detailed analysis of the in-transit absorption. Such an analysis is beyond the scope of
this paper. The consistently high post-transit values of $W_{NaI}$ in \autoref{fig:mabs_nai} appear to
be due to the large core residuals and not to any real emission features.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.80,clip,trim=25mm 25mm 5mm 20mm,angle=0]{mabs_all_20150804.pdf}
\figcaption{Absorption time series calculated for each of the transmission spectra shown in \autoref{fig:tspecs_halpha},
\autoref{fig:tspecs_hbeta}, and \autoref{fig:tspecs_hgamma}. Values of W$_\lambda$ are calculated using
\autoref{eq:wlambda}. The average 1$\sigma$ uncertainties in W$_\lambda$ for each line
are marked with the solid bars in the upper-left. Optical transit contact points are marked with
vertical green dashed-dotted lines. The solid lines are model $W_\lambda$ values (see \autoref{sec:model}). Note
the two distinct pre-transit dips in W$_{H\alpha}$ between $-220$ minutes
and $-155$ minutes and then again between $-110$ and $-60$ minutes. The in-transit absorption strength is $\sim$2
times stronger than observed in the 2013 transit and there appears to be sustained absorption for $\sim$30 minutes
post-transit.\label{fig:mabs_all}}
\end{figure*}
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.80,clip,trim=25mm 25mm 5mm 20mm,angle=0]{mabs_all_nai1_20150804.pdf}
\figcaption{Absorption time series calculated for each of the \ion{Na}{1} transmission spectra shown in \autoref{fig:tspecs_nai}.\label{fig:mabs_nai}}
\end{figure*}
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.80,clip,trim=25mm 25mm 5mm 20mm,angle=0]{mabs_all_mgi_20150804.pdf}
\figcaption{Absorption time series calculated for each of the \ion{Mg}{1} transmission spectra shown in \autoref{fig:tspecs_mgi}.\label{fig:mabs_mgi}}
\end{figure*}
\subsection{H$\alpha$ line centroid velocities}
\label{sec:havels}
The signal-to-noise in the individual H$\alpha$ transmission spectra is high enough to consider the overall velocity
shift of the absorption. Measurements of the line velocity can be useful in understanding the mass motions of the
absorbing material. In particular, measurements of in-transit line velocities can provide information about
the planetary rotation and atmospheric dynamics \citep[e.g.,][]{showman02,menou10,rauscher14,snellen14,louden,brogi15}.
Here we will focus on the pre-transit line velocities and a detailed analysis of the in-transit velocities
will be presented in a future paper.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.80,clip,trim=30mm 25mm 5mm 20mm,angle=0]{halpha_linevels.pdf}
\figcaption{Line centroid velocities calculated for H$\alpha$ transmission spectra that display 2$\sigma$ absorption in \autoref{fig:tspecs_halpha}.
Values calculated for the 2013 data from \citetalias{cauley15} are also shown. The green vertical dashed lines show the transit contact points.
The full range of $t-t_{mid}$ is shown for easy comparison with \autoref{fig:mabs_all}. The pre-transit lines show mostly positive velocities,
indicating material moving towards the star. An analysis of the in-transit velocities will be presented in a future paper.\label{fig:halinevels}}
\end{figure*}
The velocity is measured for each H$\alpha$ transmission spectrum which has a W$_\lambda$ value significant at
the 2$\sigma$ level. We use a flux-weighted average from $-40$ to $+40$ km s$^{-1}$ to measure the velocity of the absorption:
\begin{equation}\label{eq:vhalpha}
v_{H\alpha} = \frac{\sum\limits_{v=-40}^{+40} v \left(1-f_{v}\right)^2}{\sum\limits_{v=-40}^{+40} \left(1-f_{v}\right)^2}
\end{equation}
\noindent where $(1-f_{v})^2$ is the square of the transmission spectrum flux at velocity $v$. This
ensures that deeper portions of the line are heavily weighted. We also add back in the velocity corrections from \autoref{fig:halphavcorrs}.
We note that these corrections are small ($\sim$-1.0 to 1.0 km s$^{-1}$) and do not significantly affect any of the conclusions
based on the line velocities discussed here. The uncertainty in each velocity is the standard deviation of the
distribution derived from calculating the line velocity for each of the $N=255$ combinations ($N=511$ for the 2013 data set) used
to produce the individual spectra in \autoref{fig:tspecs_halpha}. This is essentially the EMC process introduced in \autoref{sec:indtranspec}.
We add an extra 1.0 km s$^{-1}$ to each uncertainty to approximate the error in the original velocity correction.
The H$\alpha$ velocities for both the 2013 and 2015 data sets are shown in \autoref{fig:halinevels}. Most of the 2015 pre-transit
velocities are red-shifted, indicating material moving away from the observer. However, immediately before the transit the
velocities are grouped relatively close to zero with a sharp transition down from $\sim+$4 km s$^{-1}$ between $t-t_{mid}=-85$ and
$-79$ minutes. There is no obvious explanation for this abrupt change in the line velocity. Also note the consistent blue-shift
of the 2013 in-transit points compared to the symmetric in-transit velocities of the 2015 data. We will revisit the pre-transit
velocities during \autoref{sec:model} in the context of the clumpy accretion stream model.
\section{The residual \ion{Ca}{2} H core flux}
\label{sec:calcium2}
The \ion{Ca}{2} H and K lines are frequently used as stellar activity indicators \citep[e.g.,][]{duncan,wright04,isaacson,gomes14} and are standard
diagnostics of solar flare energetics and dynamics \citep[e.g.,][]{johnskrull97,garcia05,lalitha}. H$\alpha$, which is in emission
in stellar active regions, is also used as an activity measure \citep{meunier09,gomes14,kuridze} and correlates directly with \ion{Ca}{2} H and K
across solar cycles \citep{livingston07}. Much variation exists, however, in the H$\alpha$-\ion{Ca}{2} H and K correlation
for other stars \citep[e.g.,][]{cincunegui}. Surface H$\alpha$ activity can therefore mimic an absorption
signature: if the comparison spectra capture the star in a more active state compared to the spectra of interest (e.g., the
pre- or in-transit spectra), the filling in of the line cores in the active spectra will result in ``absorption'' in the transmission spectrum
of a non-active spectrum. Thus it is critical to understand the stellar activity level as a function of time in order
to attempt to separate any activity contribution to the absorption signature.
While measurements of \ion{Ca}{2} and H$\alpha$
core flux have been used to study the long term (days to years) influence of hot planets on their host stars
\citep[e.g.,][]{shkolnik05,shkolnik08,fares10,scandariato}, there has been little effort to characterize the very short term variations
(minutes to hours) in the line cores of stars other than the Sun. Ideally, short-cadence observations of the star would be carried out at
multiple phases of a single planetary orbit in order to establish a baseline of the short-term stellar activity level and
characterize typical changes in the line core flux. This baseline could be used to separate the purely stellar component
from any variations caused by circumstellar material during the phase of interest.
H$\alpha$ and \ion{Ca}{2} H and K emission is spatially coincident in active regions on the Sun
\citep[e.g., see][]{meunier09,kuridze}, although the two line fluxes do not necessarily correlate
across the entire solar disk. Filaments, which are cooler than plages and can absorb active region
H$\alpha$ emission, can affect the level of observed H$\alpha$ flux and weaken the correlation
between H$\alpha$ and \ion{Ca}{2} H and K \citep{meunier09}. Short cadence observations of flaring
regions on the Sun show a direct correlation between \ion{Ca}{2} and H$\alpha$ excess emission
equivalent widths \citep[e.g.,][]{johnskrull97,garcia05}. A similar correspondence is seen in dMe
flares. In the case of dMe objects, the peak \ion{Ca}{2} K flux occurs $\sim$30 minutes after the
peak in the Balmer line fluxes and variations of $\sim$5\% are common on 10-minute timescales
\citep{kowalski13}. Although HD 189733 is not as active as a typical dMe star, it is an active
K-dwarf and so it is likely that it would exhibit low-level chromospheric variability on
short timescales. To the best of our knowledge, no continuous short-cadence observations of HD
189733's H$\alpha$ and the \ion{Ca}{2} H and K lines, similar to the observations presented here and
in \citetalias{cauley15}, have been published. Thus it is difficult, outside of the transits
presented in this work, to quantify the stellar contribution to variations in the line fluxes on
short timescales.
\begin{figure}[htbp]
\centering
\includegraphics[scale=.65,clip,trim=65mm 35mm 5mm 70mm,angle=0]{caii_epoch_comp.pdf}
\figcaption{Comparison of the normalized \ion{Ca}{2} H line cores of the comparison spectrum for the 2013 transit (red)
and 2015 transit (blue). The 2015 line core has $\sim$12\% more flux than the 2013 line core, indicating that the
star is in a more active state during the 2015 transit.\label{fig:caii_comp}}
\end{figure}
In \citetalias{cauley15} we used the time series structure of the \ion{Ca}{2} Mt. Wilson S$_{HK}$
index to argue that the pre-transit signal was not caused by varying stellar activity and was
instead due to material occulting the star. This argument was based on the lack of correlation
between the S$_{HK}$ index and W$_{H\alpha}$. Furthermore, the mean S$_{HK}$ value of the comparison
spectra was almost identical to that of the pre-transit spectra, suggesting a similar activity level
pre- and post-transit. However, the S$_{HK}$ index compares the core flux with a wide continuum
window and, as a result, is subject to variations in the continuum flux. Small variations in the
continuum window flux, which are not necessarily related to the same physical processes causing flux
changes in the line cores, can thus mimic changes in the stellar activity level.
In order to mitigate any contribution to the activity index by the continuum windows, we have reanalyzed the
2013 \ion{Ca}{2} lines using the residual core flux method of \citet{shkolnik05}. We have also applied this
analysis to the 2015 \ion{Ca}{2} lines. To be clear, the S$_{HK}$ index is well suited for measuring average
activity levels across different epochs and, indeed, we present the average S$_{HK}$ values for each night
in \autoref{fig:caii_comp}. However, a more precise measurement of variations in the stellar activity level,
is better achieved by measuring only the core flux.
The residual profiles are constructed by normalizing the line core to 0.1 \AA\ wide spectral regions
centered at 3967.05 \AA\ and 3969.95 \AA\ and subtracting the mean profile of the same comparison
spectra used to generate the average Balmer line transmission spectra. Individual values of
W$_{CaH}$ are calculated by summing over the residual spectrum from 3967 \AA\ to 3970 \AA\ . The
mean \ion{Ca}{2} H profiles for the comparison spectra from each night are shown in
\autoref{fig:caii_comp}. The \ion{Ca}{2} H residual profiles are shown for the 2013 transit in
\autoref{fig:caiires_0704} and for the 2015 transit in \autoref{fig:caiires_0804_1} and
\autoref{fig:caiires_0804_2}. We exclude the \ion{Ca}{2} K profiles due to lower
signal-to-noise.\footnote{The \ion{Ca}{2} H and K lines are located near the edges of separate
orders in our spectrograph setup.} We note that the residual profiles are not identical to the
transmission spectra, since they are not normalized by the comparison spectrum, and so values of
W$_{H\alpha}$ and W$_{CaH}$ cannot be directly compared.
\autoref{fig:caiihatime_0704} and \autoref{fig:caiihatime_0804} show the W$_{CaH}$ timeseries for
the 2013 and 2015 transits, respectively. The values of W$_{H\alpha}$ from each transit are shown
for comparison. Representative 1$\sigma$ uncertainties are shown with the solid colored bars. The
residual core flux in \autoref{fig:caiihatime_0704} shows different structure compared to the
S$_{HK}$ index from \citetalias{cauley15} \citepalias[see Figure 4 from ][]{cauley15}. Notably, the
pre-transit W$_{CaH}$ values appear to trace the pre-transit W$_{H\alpha}$ values and the mean
pre-transit W$_{CaH}$ value is lower than the mean post-transit value, suggesting similar levels of
stellar activity both pre- and post-transit. This contradicts the behavior of the S$_{HK}$ index
from \citetalias{cauley15} which showed very similar pre- and post-transit values. As described
above, this must be due to small changes in the continuum window flux used in the index calculation
since the core flux remains the same in both measurements. The 2013 W$_{CaH}$\ values are plotted against
the W$_{H\alpha}$\ values in the top panel of \autoref{fig:caii_corr}. The Spearman's $\rho$ value, $\rho_S$,
and the corresponding $p$-value are given in the upper left of the panel. There is a moderate
correlation between W$_{CaH}$\ and W$_{H\alpha}$, although we note that this is mainly driven by the in-transit
values (green crosses) which have $\rho_S$=0.51 ($p$=0.016). The \ion{Ca}{2} H residual values for
the 2015 transit are shown in \autoref{fig:caiihatime_0804}. The 2015 W$_{CaH}$ and W$_{H\alpha}$ do
not show any obvious correspondence, although the in-transit W$_{CaH}$\ values are lower on average than
the pre-transit values. There is a weak but significant correlation present between the 2015 values
of W$_{CaH}$\ and W$_{H\alpha}$.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.80,clip,trim=25mm 25mm 5mm 20mm,angle=0]{caii_ha_timeseries_20130704.pdf}
\figcaption{\ion{Ca}{2} H residual core flux time series for the 2013 transit. The W$_{H\alpha}$ values are over plotted
for comparison. Typical 1$\sigma$ uncertainties are shown with solid colored lines in the upper left. The core flux
alone shows different behavior than the S$_{HK}$ index used in \citetalias{cauley15}.
Notably, the core flux appears to trace the H$\alpha$ absorption, suggesting that some of the signal in H$\alpha$
attributed to absorption may be caused by varying stellar activity. In addition, the residual core flux of the
comparison spectra (the nine post-transit spectra) is higher than the pre-transit spectra. However, the correlation is weak (see \autoref{fig:caii_corr}
and \autoref{fig:rhos_dist}) so it is difficult to quantify how much of the signal is purely due to stellar activity.\label{fig:caiihatime_0704}}
\end{figure*}
In order to test the significance of the correlations seen in both data sets, we have run a simple Monte Carlo
procedure that randomly draws values of W$_{CaH}$\ and W$_{H\alpha}$\ from a normal distribution defined by the 1$\sigma$
flux errors for each point. Representative values of the uncertainties are shown with the solid red lines in
\autoref{fig:caii_corr}. A simulation of 10,000 random draws is shown in \autoref{fig:rhos_dist} where the 2015
distribution of $\rho_S$ is shown in magenta and the 2013 distribution is shown in brown. The nominal $\rho_S$ values
calculated for the measurements in \autoref{fig:caiihatime_0704} and \autoref{fig:caiihatime_0804} are marked
with vertical dashed lines. It is clear that the nominal values lie on the high end of the distribution for both dates,
suggesting that the true correlation between W$_{CaH}$\ and W$_{H\alpha}$\ is weaker than the nominal values indicate.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.80,clip,trim=25mm 25mm 5mm 20mm,angle=0]{caii_ha_timeseries_20150804.pdf}
\figcaption{The same as \autoref{fig:caiihatime_0704} except for the 2015 transit. Note the larger plot range for
W$_\lambda$. The scatter in W$_{CaH}$ is larger than for the 2013 transit and there is little correlation between
W$_{CaH}$ and W$_{H\alpha}$ (see \autoref{fig:caii_corr}), although the mean in-transit value of W$_{CaH}$
is lower than the mean pre-transit value. Note the low values of W$_{CaH}$ for the first three comparison spectra
near $\sim-280$ minutes.\label{fig:caiihatime_0804}}
\end{figure*}
\begin{figure}[htbp]
\centering
\includegraphics[scale=.58,clip,trim=35mm 15mm 5mm 10mm,angle=0]{caii_ha_corrplots.pdf}
\figcaption{W$_\lambda$ for \ion{Ca}{2} H versus H$\alpha$ for the 2013 and 2015 transits. Note the larger plot range for W$_{CaH}$
in the bottom panel. There is a modest correlation
in the 2013 data but it is driven almost entirely by the in-transit points ($\rho_s$=0.52, $p$=0.016; green crosses). There is a weak but significant
correlation in the 2015 data. Typical 1$\sigma$ uncertainties in W$_\lambda$ are shown with the red solid lines.\label{fig:caii_corr}}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=.56,clip,trim=49mm 30mm 35mm 45mm,angle=0]{rhos_distributions.pdf}
\figcaption{Distributions of $\rho_s$ for the W$_{H\alpha}$ versus W$_{CaH}$ correlations shown in \autoref{fig:caii_corr}.
The distributions are generated by running 10,000 simulations and letting the values of W$_{CaH}$ and W$_{H\alpha}$
vary according to their 1$\sigma$ uncertainties, which are assumed to be normal errors. The measured values
of $\rho_S$ for both dates (vertical dashed lines) are much higher than the average values of the distributions,
suggesting that the true correlations are weaker than the measured correlation.\label{fig:rhos_dist}}
\end{figure}
While we have focused on stellar activity being the main driver of the \ion{Ca}{2} core flux, it is also possible
that the changes in the core flux are in part due to absorption in the pre-transit material. This would
weaken any interpretation of the $W_{CaII}$-$W_{H\alpha}$ correlation as testing the relationship
between stellar activity in the two lines. On the other hand, any absorption by pre-transit material
in one line but not the other has a similar effect. Thus interpretation of the \ion{Ca}{2} residual
spectrum will also benefit from the activity baseline observations suggested for H$\alpha$ earlier in
this section.
\subsection{The impact of stellar activity}
\label{sec:activityimpact}
The analysis presented in \autoref{sec:calcium2} suggests that the impact of stellar activity on the
Balmer line transmission signal is non-negligible and needs to be taken into account. The
correlation between W$_{CaH}$\ and W$_{H\alpha}$, although modest, weakens the interpretation of the pre-transit
signal presented in \citetalias{cauley15} as being caused by material compressed in a bow shock as
opposed to variations in the stellar activity level. Due to the level of scatter in the W$_{CaH}$\
values, and also the fact that \ion{Ca}{2} and H$\alpha$ do not track exactly the same physical
conditions or locations in the stellar chromosphere, we cannot empirically separate the contribution
of normal stellar activity from the observed W$_\lambda$ time series.
\begin{figure}[htbp]
\centering
\includegraphics[scale=.53,clip,trim=45mm 30mm 35mm 45mm,angle=0]{activity_tests.pdf}
\figcaption{W$_\lambda$ as a function of active region surface area fraction ($f_{act}$) for various normalized emission line strengths ($A_0$).
W$_\lambda$ is computed according to \autoref{eq:wlambda} for the transmission spectra defined in \autoref{eq:stactivity}.
The horizontal dashed brown line is the mean value of W$_{H\alpha}$ for the pre-transit absorption values in
\autoref{fig:mabs_all}; the dashed orange line is the mean value of the first three pre-transit points from
\citetalias{cauley15}. While the strength of the emission line is degenerate with f$_{act}$, it is clear that for
a range of reasonable parameters the measured change in the pre-transit values of W$_\lambda$ for both the 2013 and 2015 transits
can be accounted for by low levels of stellar activity.\label{fig:acttests}}
\end{figure}
In lieu of an empirical method for separating the stellar activity level from any true absorption, we can explore
the parameter space of active region surface coverage and emission line intensity required to reproduce
the observed changes in W$_\lambda$. We take an approach similar to that of \citet{berta} and examine
differences between weighted combinations of active region spectra, i.e., emission lines, and quiet star
spectra, in this case a flat continuum. The emission lines are Gaussians with FWHM $\sigma$ and height
$A_0$. We focus on H$\alpha$ and take $\sigma$=0.79 \AA, the mean FWHM of the pre-transit
absorption lines. The height $A_0$ is varied from 0.1 to 1.0. Low values of the emission line heights are
chosen to mimic the low levels of variation that we see here. Moderate solar and stellar flares can cause increases
in the emission line flux by factors of $\sim$2 relative to the local continuum \citep{johnskrull97,kowalski13} and
we do not see evidence (e.g., sharp increases in line strength followed by gradual decay) for large flaring
events in our spectra. In addition, large emission line strengths are not necessary to reproduce the observed
values of W$_{H\alpha}$, as shown below.
Active region surface area coverage of the observable hemisphere is varied
from 0.1\% up to 10\% \citep[e.g.,][]{andretta95,meunier09}. Spot coverage for HD 189733 has been
estimated at $\sim$1\% \citep{winn07}. For the Sun, \citet{shapiro15} showed that the disk spot coverage fraction ($A_S$),
during peaks in the solar activity cycle, is $\sim$0.2\% while the disk active region coverage fraction ($A_F$) is
$\sim$3\%. They also showed that the ratio of $A_S$/$A_F$ increases as the level of solar activity increases.
This suggests that for HD 189733 the ratio $A_S$/$A_F$ is greater than that for the sun, although the
precise relationship is uncertain for very active stars. For the Sun during an activity maximum, $A_S$/$A_F$=0.07.
Thus we can assume with some confidence that the maximum active region coverage fraction of
HD 189733 b is less than 15\%. Furthermore, \citet{shapiro15} showed that brightness changes
in more active stars are dominated by spots rather than faculae (i.e., bright active regions), although
this does not rule out active region variability as the cause of small changes in spectral line flux.
The transmission spectrum is calculated as
\begin{equation}\label{eq:stactivity}
S_T = \frac{S_* (1-f_{act}) + f_{act} S_{act}}{S_*}
\end{equation}
\noindent where $f_{act}$ is the fraction of the observable hemisphere covered by active regions with the
corresponding spectrum $S_{act}$. The quiet star spectrum is $S_*$, which can be considered to be the
comparison spectrum. W$_{H\alpha}$ is then calculated according to \autoref{eq:wlambda}.
\autoref{fig:acttests} shows the dependence of W$_{H\alpha}$ on the active region surface coverage and the
strength of the active region emission lines. The average observed value of W$_{H\alpha}$ for both dates is shown with a
horizontal dashed line. The strength of the active region emission line, $A_0$, is degenerate with $f_{act}$.
This results in many different combinations of $A_0$ and $f_{act}$ being able to reproduce the observed
changes in W$_{H\alpha}$. Regardless of the specific values of $A_0$ and $f_{act}$,
the measured levels of pre- and in-transit absorption in both the 2013 and 2015 data can be
approximated by relatively small changes in the stellar activity level between the comparison spectra and the spectrum
of interest. The smaller pre-transit values for the 2015 transit
can be reproduced by almost any combination of $f_{act}$ and $A_0$. The large values from the 2013 transit
can be reproduced by most of the parameter combinations but weak line strengths require large surface
coverage fractions. Although HD 189733 is an active star, it is unclear whether 8-10\% of the visible
hemisphere can be covered in bright active regions. High-cadence monitoring of HD 189733 while
HD 189733 b is not transiting will provide a better statistical understanding of activity variations
on $\sim$5 minutes timescales and help determine the true nature of the pre-transit signal.
For in-transit measurements, contrast between the comparison spectra, which are integrated across
the entire stellar disk, and the in-transit spectrum of interest, which is integrated across the
stellar disk not covered by the opaque planet, can result in line strength differences as the
spectrum is weighted towards active or spotted regions of the star \citep{berta}. The contrast
effect can be demonstrated by a planet with no atmosphere transiting a chord that is free of spots
or active-regions. In this case, the in-transit spectrum will be weighted towards the spotted or
active stellar surface by the ratio ($R_p$/$R_*$)$^2$. For the HD 189733 system,
($R_p$/$R_*$)$^2$=0.024. If the out-of-transit spectrum is composed of 1.00\% spotted/active region
spectrum, the weighting of the in-transit spectrum, ignoring limb darkening, will change to 1.02\%
spotted/active and 98.98\% quiescent. Thus any difference between the spotted/active region spectrum
and the quiescent stellar spectrum will affect the transmission spectrum at the level of 0.02\%.
This is much smaller than the observed in-transit line depth. We note that this effect changes
weakly with increased spot/active region coverage and is $<$1\% even for coverage fractions of 30\%.
In addition, the in-transit signal begins before the optical transit and continues after the opaque
planetary disk leaves the stellar disk. This is further evidence that the absorption is caused by
gas in the planet's atmosphere and not the contrast effect since the contrast effect requires some
of the stellar surface to be blocked by the planet.
\subsection{The choice of comparison spectra}
\label{sec:compspec}
The analysis in \autoref{sec:activityimpact} shows that differing levels of stellar activity at different times
during the period of observation can affect the relative level of measured absorption in the line core. This
is directly related to the choice of comparison spectra. If instead of choosing spectra numbers 2-9 for the 2015 transit we select the
observations between $t-t_{mid}=-230$ and $-160$ minutes as the comparison spectra, all of the measured
W$_\lambda$ (\autoref{fig:mabs_all}) values would shift up by the relative average difference between the points in those time ranges.
Thus instead of pre-transit \textit{absorption} we would observe an initial period of enhanced
stellar activity, followed by a relative period of quiescence, then another brief activity increase between
$t-t_{mid}=-150$ and $-116$ minutes, and finally relative quiescence immediately before the transit. If
instead we chose the post-transit spectra, as we chose to do in \citetalias{cauley15}, the in-transit
absorption would be very weak and all of the pre-transit W$_\lambda$ measurements would be positive,
suggesting sustained levels of higher stellar activity compared to in- and post-transit times. Thus clearly
the choice of comparison spectra can influence the interpretation of the W$_\lambda$ values.
Our main justification for choosing the earliest observations in the 2015 data as the comparison spectra
is that they are located furthest from the transit and thus have a lower likelihood of being contaminated
by high-density pre-transit material capable of producing an absorption signature. It would not be unreasonable,
however, to select a different set of pre-transit spectra. We believe the low post-transit W$_\lambda$ values, and
their significant variability, rule these spectra out as good out-of-transit baseline values. The up-and-down
nature of the 2015 pre-transit W$_\lambda$ values do not offer much evidence for or against absorbing
material versus changes in the stellar activity level as the cause of the signal. However, if the large pre-transit
variations of $\sim$0.005-0.007 \AA\ at $-230$, $-155$, and $-110$ minutes were caused by changes in the stellar activity level, this
would indicate such variations are common and we might expect to see similar changes on similar timescales
during and after the transit. This is not the case: most of the in-transit variations are of smaller magnitude
($\sim$0.003-0.005 \AA) and shorter timescales, which may be due to transits of active regions.
Furthermore, the duration of the dip in W$_{H\alpha}$ starting at $-220$ minutes is $\sim$80 minutes which
is very close to the transit duration of a ``narrow" feature. It would be
surprising if the star's activity level happened to abruptly change on similar timescales as the transit
duration.
Finally, although the individual $W_{H\beta}$ values are marginal, if the pre-transit
spectra between $-230$ and $-155$ minutes or between $-110$ and $-60$ minutes are chosen as
the comparison spectra, the in-transit H$\beta$ absorption becomes much weaker or disappears entirely.
Our results demonstrate the need for a longer baseline of short cadence observations in order to establish the
true out-of-transit stellar activity baseline.
\section{Modeling the pre-transit absorption}
\label{sec:model}
In \citetalias{cauley15} we modeled the pre-transit absorption as arising in a thin bow shock orbiting $\sim$13 $R_p$
ahead of the planet. While this simple geometric model was able to account for the absorption time series and
the strength of the measured absorption, the favored model parameters resulted in two relatively unlikely
conclusions: 1. A low stellar wind speed of $\sim$40 km s$^{-1}$ at the planet's orbital radius; 2. A very strong
equatorial surface planetary magnetic field strength of $\sim$28 G. The low stellar wind speed is at odds with
most MHD simulations of the slow solar and stellar winds \citep[e.g.][]{cohen07,llama13,johnstone15} which reach speeds of
$\sim$200 km s$^{-1}$ at $\sim$10 $R_*$. We note, however, that observational constraints on the stellar winds of low mass
stars besides the Sun are weak. The large magnetic field strength from our model, derived assuming pressure balance
between the incoming stellar wind and planetary magnetosphere, is $\sim$4 times larger\footnote{In \citetalias{cauley15} the
magnetic field strength was given as $\sim$2 times stronger than the \citet{reiners10} estimate. This was incorrect since the
\citet{reiners10} given value is 14 G at the pole and not the equator. For a dipolar field, the equatorial value is half of
the polar value.} than that estimated by \citet{reiners10} for HD 189733 b based on magnetic field strength scaling relations
for brown dwarfs and giant planets. While field strengths of this magnitude are estimated for more massive planets, it is difficult to produce
such a strong field for a Jupiter-mass object \citep{reiners09,reiners10}.
The absorption time series shown in \autoref{fig:mabs_all} does not match the specific prediction of the bow shock
model from \citetalias{cauley15}. Furthermore, we were unsuccessful at finding a suitable bow shock model
to describe the 2015 time series. As demonstrated in \autoref{sec:activityimpact}, the magnitude of the pre-transit
absorption can be influenced by low level variations in the stellar activity level. If the observed signal is assumed to
be solely associated with the planet, it is possible to explain the absorption as arising in transiting circumplanetary
material. This motivated us to search for a geometry that is capable of reproducing the absorption.
\subsection{Clumpy accretion}
\label{subsec:clump}
One such geometry is an accretion stream escaping from the L1 point as the planet overflows
its Roche lobe \citep{lai,li10}. The relatively constant pre-transit W$_\lambda$ absorption values, the very early appearance
of the absorption, and the duration of the pre-transit signals are suggestive of a clumpy accretion flow.
This type of accretion flow, i.e., non-uniform and time variable spiraling in from the planet's orbit,
is seen in specific 2D MHD simulations of \citep[][; see also \citet{bisikalo13} for simulations of Roche lobe overflow
where the accretion stream is halted by the stellar wind pressure]{matsakos}.
We have modeled the pre-transit absorption as being caused by two distinct clumps of material that are
spiraling in from the planet's orbital radius towards the central star. The choice of two clumps is
motivated by the two distinct pre-transit absorption signatures seen in the W$_{H\alpha}$ values in \autoref{fig:mabs_all}.
The spiral trajectory in this case is approximated as a straight line since the clumps are located relatively close to the planet.
In fact, due to the combination of the Coriolis force and the gravitational attraction of the
planet, the trajectory of the material near the L1 point can be approximated by a straight line
\citep{lubow75,lai}. Any deviations from a straight line trajectory are secondary effects on the
transit light curve. Each clump is released from rest from the L1 point at 4.25 $R_p$ at some
time $t_{pre}$. The clump is then accelerated along a straight line trajectory that forms an angle
$\theta_c$ with the planet's orbit. The angle $\theta_c$ is determined by the system mass ratio and
in this case is $\sim$60$^{\circ}$. The acceleration of the clump is $a_c$=0.0019 km
s$^{-2}$=2.3$\times$10$^{-8}$ $R_p$ s$^{-2}$, which is approximated from the
$q$=$M_p$/$M_*$=0.001 case ($q=$0.0013 for the HD 189733 system) presented in \citet{lai}. For simplicity, we
choose a cylindrical geometry for each clump defined by a radius $r_c$ and length $l_c$. The density
of each clump is uniform. The line profiles of the transiting clump are calculated identically to
the bow shock profiles in \citetalias{cauley15}. The optical depth at line center for each grid point is
\begin{equation}\label{eq:optdep}
\tau_0=\frac{\sqrt{\pi}e^2f_{lu}N_l\lambda_{lu}}{m_e c b}
\end{equation}
\noindent where $f_{lu}$ is the oscillator strength of the transition, $\lambda_{lu}$ is the central wavelength,
$m_e$ is the electron mass, $e$ is the electron charge, $c$ is the speed of light in vacuum, and $N_l$ is the column
density of the lower energy level of the transition. The line profile is a Doppler-broadened delta function, $\tau_v=\tau_0 e^{-(v/b)^2}$,
where $b=\sqrt{2}\sigma_v$ and $\sigma_v$ is the dispersion of a 1D Gaussian velocity distribution \citep{draine}.
The line broadening parameter for the clumps, used to approximate the profile shape according to
\autoref{eq:optdep}, is $b$=5.7 km s$^{-1}$.
\subsection{In-transit absorption}
\label{subsec:intran}
In order to maintain focus on the pre-transit signal, we will present a detailed discussion of the
in-transit data, including an analysis of velocities in the planetary atmosphere, in a subsequent
paper. A brief summary of the in-transit model, also shown over-plotted in \autoref{fig:mabs_all},
is given here. We model the 2015 in-transit absorption as arising in an extended uniform density
atmosphere. The choice of uniform density is motivated by the results of \citet{christie} who find
that the density of the dominant $n$=2 state (the 2s state) remains fairly constant over almost
three orders of magnitude in pressure (10$^{-6}$-10$^{-9}$ bar). This can be understood as the
result of outwardly increasing temperatures offsetting the decreasing neutral hydrogen abundance. We
assign a temperature to the base of the atmosphere and this sets the scale height, $H_{atm}$. The 3D
density grid is filled with material out to 10$H_{atm}$ and the column density is calculated for
each atmospheric point that occults the star. We find a good approximation of the average in-transit
Balmer line profile shape and absorption values with a density $\rho=4.0\times10^{-23}$ g cm$^{-3}$
and $T=6,000$ K, which translates to a scale height of $H_{exo}$=0.028 $R_p$.
\subsection{Model results}
\label{subsec:modresults}
When finding a model that roughly describes the pre-transit signal, we need to consider the strength
and shape of the $W_\lambda$ time series, the ratios of the Balmer line $W_\lambda$ values, and the
shape of the observed line profiles. While these constraints taken together significantly narrow
the available parameter space, degeneracies exist between the model parameters. For example,
increasing the length of the clump by 50\% while decreasing the density by a factor of $\sim$3
results in similar average $W_\lambda$ values. The increase in clump length, however, lengthens the
egress and ingress time of the pre-transit features, which appear to be very brief ($\sim$10
minutes). In addition, the decreased density produces a narrower line profile as opacity broadening
becomes less prominent. This results in less acceptable line profile matches between the data and
the model. The final clump model values in \autoref{tab:modpars} reflect our attempt to balance
these constraints.
\begin{deluxetable}{lccc}
\tablecaption{Model parameters \label{tab:modpars}}
\tablehead{\colhead{Parameter}&\colhead{Symbol}&\colhead{Value$^{a,b}$}&\colhead{Units}\\
\colhead{(1)}&\colhead{(2)}&\colhead{(3)}&\colhead{(4)}}
\tabletypesize{\scriptsize}
\tablewidth{2000pt}
\startdata
Stellar radius & $R_*$ & 0.756 & $R_\Sun$\\
Stellar rotational & $v$sin$i$ & 3.10 & km s$^{-1}$ \\
Impact parameters & $b$ & 0.680 & $R_*$ \\
Orbital period & $P_{orb}$ & 2.218573 & days \\
Orbital velocity & $v_{orb}$ & 151.96 & km s$^{-1}$ \\
Orbital radius & $a$ & 0.03099 & AU \\
Planetary radius & $R_p$ & 1.138 & $R_J$ \\
& & & \\
Time of clump initiation & $t_{pre}$ & 16.7,11.4 & hours before $t=0$ \\
Clump radius & $r_c$ & 0.5,0.5 & $R_p$ \\
Clump length & $l_c$ & 1.0,0.5 & $R_p$ \\
Clump density & $\rho_c$ & 1.5,4.0 & 10$^{-22}$ g cm$^{-3}$ \\
Line broadening & $b_c$ & 5.7,5.7 & km s$^{-1}$ \\
\enddata
\tablenotetext{a}{With the exception of $v$sin$i$, all stellar and planetary parameters taken from \citet{torres08}. The
$v$sin$i$ value is taken from \citet{collier10}.}
\tablenotetext{b}{Clump parameters listed for clump 1 and then clump 2.}
\end{deluxetable}
The model W$_\lambda$ values are shown as solid lines in \autoref{fig:mabs_all}. The final clump
parameters and the systems parameters used in the model are listed in \autoref{tab:modpars}. A view
of the clump transit is shown in \autoref{fig:tranview} for $t-t_{mid}=-158$ minutes\footnote{An animation
of the clump transit is available at pwcastro.com/research/.}. The clump
model is successful at describing both the timing of the pre-transit events and their absorption
levels. We note that if the clumps are not allowed to accelerate, the transit duration is too long
and the transit timing is significantly off. The densities of the clumps also roughly reproduce the
Balmer line absorption ratios and line shapes, although most of the individual H$\beta$ and
H$\gamma$ points are not measured at a significant level.
One piece of evidence that could support the accretion clump interpretation is the large $S_{HK}$
value measured for the 2015 transit. If the star is in an active state, the planet will be subjected
to higher levels of high energy stellar radiation, increasing the mass loss rate \citep[e.g.,
][]{murray09,trammell,owen,valsecchi15,owen16}. If enough material is evaporated from the atmosphere, the planet's
magnetosphere can limit the outflow rate if the magnetic pressure dominates the gas pressure,
creating a ``dead zone'' that extends many planetary radii \citep{trammell,trammell14,khodachenko}. Periodic
mass loss, i.e., the accretion clumps, is not implausible from a filled magnetosphere that is being
subjected to variable heating rates \citep[e.g., ][]{pillitteri15}. Furthermore, mass loss from HD 189733 b is observed to be
variable \citep{desetangs12}. Although we do not explore the in-transit signal in detail
here, the in-transit absorption appears to begin $\sim$15 minutes before $t_I$ and last at least 30
minutes after $t_{IV}$. The planet moves across the star at $\sim$0.11 $R_p$ minute$^{-1}$ which
means the feature causing the absorption immediately pre- and post-transit extends $\sim$1.5 $R_p$
ahead of the planet and $\sim$3.0 $R_p$ behind the planet. This extended cloud of material is
suggestive of a partially filled Roche lobe.
\begin{figure}[htbp]
\centering
\includegraphics[scale=.51,clip,trim=10mm 15mm 75mm 10mm,angle=0]{tranim_stream.pdf}
\figcaption{Views of the first transiting clump in the plane of the orbit (top panel) and perpendicular to the orbital plane (bottom
panel) for the snapshot at $t-t_{mid}=-158$ minutes. The densities are scaled to the planetary atmosphere. The clumps
represent stochastic mass loss events from the planet that are dense and extended enough to cause significant absorption
in the cores of the Balmer lines.\label{fig:tranview}}
\end{figure}
\subsection{Limitations to the clumpy accretion interpretation}
\label{subsec:modlims}
While the accretion clump model can reproduce the observed absorption levels, and in addition to the
parameter degeneracies discussed above, there are serious limitations to this interpretation. The
first is the lack of large red-shifted velocities in the pre-transit H$\alpha$ transmission spectra
(\autoref{fig:tspecs_halpha} and \autoref{fig:halinevels}). By the time the clumps transit the star,
they are moving with line-of-sight velocities of $\sim$100 km s$^{-1}$ and $\sim$70 km s$^{-1}$,
much larger than the 5-10 km s$^{-1}$ line centroids seen for the individual spectra. This is the
result of the large angle (60$^{\circ}$) that the clump trajectory makes with the planet's orbit.
This angle would need to be $\sim$5-10$^{\circ}$ in order to produce red-shifted velocities of 5-10
km s$^{-1}$. A possible resolution to this problem is if the trajectory of the accreting material is
altered by interactions with the stellar wind \citep[e.g., see ][]{matsakos}, although modeling
these interactions is beyond the scope of this paper. It is also not clear if simple hydrodynamic
Roche-lobe overflow, as we've assumed here, is directly applicable in the hot Jupiter regime,
especially when the confining effect of the planet's magnetosphere is taken into account
\citep{trammell,owen}. Thus a more realistic treatment of the mass loss could allow the trajectory
of the clumps to be less steep.
Finally, there is currently little theoretical evidence to support the existence of large $n=2$
neutral hydrogen populations at such short distances from the star. During transit, the clumps are
within $\sim$5-6 $R_*$ of the star, about 30-40\% closer than the planet's orbit, and receive a
factor of $\sim$2$\times$ more stellar radiation. Again, the clump trajectory could be less extreme
through interactions with the stellar wind. Almost all of the escaping planetary material is
photo-ionized by $\sim$5 $R_p$ from the planet \citep[e.g.,][]{murray09,tripathi15,salz16} and the stellar
wind is entirely composed of ions. Thus some highly non-equilibrium process is needed to maintain a
population of excited neutral hydrogen for at least the duration of the transit. It is not clear if
such a process (e.g., time-dependent recombination or charge-exchange of stellar wind protons with
the small population of planetary neutrals) is capable of producing the necessary neutral hydrogen
density. Modeling of the interaction between the planetary accretion flow and the stellar wind and
how this interaction affects the excited neutral hydrogen density would be useful to clarify this
outstanding question.
\section{Summary and conclusions}
\label{sec:summary}
We have presented new transit observations of HD 189733 b which cover a significant amount of
pre-transit phase and a small amount of post-transit phase. Our results can be summarized as the
following:
\begin{itemize}
\item We detect strong in-transit absorption from the planetary atmosphere in
H$\alpha$, H$\beta$, and H$\gamma$. We also detect weak \ion{Na}{1} 5896 \AA\ absorption and measure
a marginal amount of \ion{Mg}{1} 5184 \AA\ absorption. An upcoming paper will present an
interpretation of the in-transit H$\alpha$ line centroid velocities.
\item We confirm the presence of a pre-transit Balmer line signature at
a significant level for the 2015 data. However, the pre-transit signal observed in the 2015 transit
differs significantly from that observed in \citetalias{cauley15}. In particular, the 2015 pre-transit signal
does not match the prediction of the specific bow shock model presented in \citetalias{cauley15}.
\item We have reanalyzed the \ion{Ca}{2} H line from the 2013 transit and found a moderate
correlation between the $W_{H\alpha}$ values and residual core flux of the \ion{Ca}{2} lines.
A weaker correlation was found between
$W_{H\alpha}$ and $W_{CaII}$ for the 2015 data. We suggest that the residual core flux be used in
future studies as an activity proxy for short timescales rather than the $S_{HK}$ index.
\item Motivated by the $W_{H\alpha}$-$W_{CaII}$
correlations, we explored the potential of changing activity levels to produce the pre-transit
signature. We find that even small levels of intrinsic stellar variation in the line cores can
reproduce the observed transmission spectra for both pre-transit measurements. It is not clear,
however, if changes in the stellar activity level on these timescales are common.
\item We have modeled the pre-transit absorption as arising in a clumpy accretion flow. The clumpy
accretion model is able to reproduce both the timing and magnitude of the pre-transit events.
However, there are significant drawbacks to the clumpy accretion model, the
foremost of which is the lack of large red-shifted absorption velocities that should be produced by
ballistically in-falling material.
\end{itemize}
\added{Although we have modeled the pre-transit signature as arising in absorbing circumplanetary
material, we caution against a firm interpretation of the data. The magnitude of the pre-transit detections from both the 2013 and
2015 data can be explained by changing levels of stellar activity. Due to the lack of detailed information
concerning short-term, low-level variations in the H$\alpha$ core flux, no conclusive argument can be
made concerning one interpretation over the other. Both physical scenarios should be investigated for other hot planet
systems to give the current study context. We are currently pursuing a high-cadence H$\alpha$ monitoring
campaign of HD 189733, at all phases of the planet's orbit, with the intention of determining the frequency of
these low-level changes in the core of the Balmer lines.}
Pre-transit signatures around hot planets may provide unique information concerning the dynamics of
escaping planetary material and, potentially, the structure and strength of the planetary
magnetosphere. Hot planets around less active
stars than HD 189733 would make prime targets for future investigations into pre-transit signatures since
the observed signal can be attributed to the planet with a higher probability.
\bigskip
{\bf Acknowledgments:} We are grateful to the referee Joe Llama for comments that helped improve this manuscript.
The data presented herein were obtained at the W.M. Keck Observatory from
telescope time allocated to the National Aeronautics and Space Administration through the agency's
scientific partnership with the California Institute of Technology and the University of California.
This work was supported by a NASA Keck PI Data Award, administered by the NASA Exoplanet Science
Institute. The Observatory was made possible by the generous financial support of the W.M. Keck
Foundation. The authors wish to recognize and acknowledge the very significant cultural role and
reverence that the summit of Mauna Kea has always had within the indigenous Hawaiian community. We
are most fortunate to have the opportunity to conduct observations from this mountain. This work was
completed with support from the National Science Foundation through Astronomy and Astrophysics
Research Grant AST-1313268 (PI: S.R.). A. G. J. is supported by NASA Exoplanet Research Program
grant 14-XRP14\_2-0090 to the University of Nebraska-Kearney.
P.W.C. is grateful for useful exchanges with T. Matsakos
concerning the morphology and dynamics of hot Jupiter accretion flows. P.W.C. and S. R. acknowledge
helpful conversations with D. Christie and C. Huang regarding the interpretation of neutral hydrogen
absorption in the extended planetary atmosphere.
\clearpage
|
\section{INTRODUCTION}
\label{sec:intro}
Imaging whole brains at the cellular level without disturbing their underlying structure has always been challenging.
All cells are surrounded by a phospholipid bilayer which scatters light, rendering most biological tissues opaque to the naked eye.
Thus it is often necessary to physically slice brains in order to use light microscopy.
Sectioning tissue has two major drawbacks for researchers interested in building a whole brain connectome.
First, slicing can dislocate synapses and axons necessary for tracing neuronal circuitry.
Second, the inter-sectional resolution will always be much lower than the intra-sectional resolution, making neurite tracing difficult \cite{Kim}.
\subsection{CLARITY}
CLARITY avoids these problems by converting the brain into a translucent hydrogel-tissue hybrid.
In the procedure, the brain is first perfused with hydrogel monomers and formaldehyde.
When heated, the monomers and formaldehyde polymerize to form a molecular mesh which crosslinks amine groups of biological molecules.
Since phospholipids lack amine groups, they do not crosslink with the mesh and can be eluted away with a strong detergent.
The remaining hydrogel-brain hybrid is relatively translucent and permeable to fluorescent antibodies, making it amenable to labeling and interrogation by light-sheet microscopy. \cite{Chung, Tomer}
An axial slice through a CLARITY volume and magnified cutout are shown in Fig.~\ref{fig:clarity}.
\begin{figure}[t]
\begin{center}
\begin{tabular}{c c}
\includegraphics[height=2.5in]{clarityWithCutout} & \includegraphics[height=2.5in]{clarityCutout}
\end{tabular}
\end{center}
\caption{Axial slice through CLARITY image and magnified cutout acquired by light-sheet microscopy.}
\label{fig:clarity}
\end{figure}
\subsection{NeuroData Cluster}
CLARITY image volumes are often over 1 terabyte in size, far too large to be visualized and analyzed on a personal computer.
Furthermore, extracting meaningful information from ``big'' images can be both time-consuming and difficult.
Thus, the NeuroData cluster was created to address these challenges.
As part of the project, open source software for access, visualization, and analysis of terabyte-scale images was developed.
Images are stored in a multiresolution hierarchy with level 0 being the native resolution, and each subsequent level at half the previous level's resolution.
Thus the infrastructure is optimized for applying computer vision algorithms in parallel across multiple scales.\cite{Burns13}
Data can be accessed through a RESTful API, a stateless interface which allows end users to download image cutouts or upload data using specific URLs.
In the API, images are identified by a unique token, each of which can have one or more channels.\cite{Burns13}
The \emph{Connectome Annotation for Joint Analysis of Large data (CAJAL)} package provides access to this API through MATLAB \cite{GrayRoncal} while \emph{NeuroData Input/Output (ndio)} provides access through Python.
Images can also be visualized in a web browser using NeuroDataViz.
\subsection{Motivation}
By annotating an entire image volume, one can draw conclusions on the texture and shape of a given brain structure.
Since manual labeling is time-consuming, the most efficient annotation method is registration to a standard atlas.
Spatially normalizing several subject brains into an atlas space makes it easier to determine how one brain differs from another in any given structure or location.
Furthermore, this can also aide in visualization.
Raw CLARITY images are often acquired in an oblique plane, making it difficult for observers to identify structures on a 2D display.
Aligning the brains with an atlas solves this problem by allowing brain visualization in one of the three standard planes (axial, sagittal, and coronal).
Transforming atlas labels to a subject's space facilitates analysis of image features within brain structures. (Fig.~\ref{fig:workflow}).
\begin{figure}[t]
\begin{center}
\includegraphics[width=\textwidth]{workflow} \\
\end{center}
\caption{The registration pipeline transforms the atlas labels to the CLARITY image's space for image analysis and the CLARITY image to an atlas space for visualization.}
\label{fig:workflow}
\end{figure}
\subsection{Previous Work}
Registering CLARITY images to each other or to an atlas has recently become a topic of interest.
A preceding study described the development of a pipeline for registering the Allen Institute's Mouse Reference Atlas (ARA) to images of transsynaptic viral traced brains. \cite{Kutten}
The ARA is a widely used mouse brain atlas which includes Nissl-stained reference images and over 700 manually defined brain structures.\cite{ABAWhitePaper}
A test image was acquired using serial-two photon (STP) tomography, a technique which pairs a two-photon microscope with a vibratome for automatic tissue slicing. \cite{Ragan}
Since the intensity profiles of the Nissl-stained ARA and the test image differed greatly, the images were registered using their corresponding brain masks.
Masks were aligned first using affine registration, followed by deformable registration using Large Deformation Diffeomorphic Metric Mapping (LDDMM), an algorithm which computes smooth invertible transforms between image volumes.
Qualitative observation showed that the registration results were acceptable in most parts of the brain, although alignment of deeper structures were less accurate.\cite{Kutten}
In a different study, 25 CLARITY images were registered and averaged to create a single reference template using a mutual information metric with B-spline transforms. This template was then used to construct an atlas for experiments combining CLARITY with transsynaptic viral tracing. \cite{Menegas}
\subsection{ITK}
The \emph{Insight Segmentation and Registration Toolkit} (ITK) is an open source library funded by the National Library of Medicine.
It was developed by Kitware Inc. and has been used widely within the medical imaging community. \cite{ITK1}
ITK's registration framework is designed to be modular. To register a moving image to a fixed image, the user selects a metric (e.g. mean squared error) to compare the images, and the transform (e.g. affine) applied to the moving image. An optimizer (e.g. gradient descent) is used iteratively to improve the transform parameters.
Additionally, the user can select the type of interpolation (e.g. nearest neighbor) for resampling the moving image. \cite{ITK2}
ITK version 4 includes support for time-varying velocity field transforms often used in diffeomorphic registration algorithms such as LDDMM.
This greatly facilitates the implementation of these types of algorithms in ITK. \cite{Avants12}
This functionality can be used in Python through SimpleITK, an easy-to-use interface for ITK's algorithms. \cite{Lowekamp}
\subsection{Challenges}
In addition to their large size, CLARITY images present several unique challenges for image registration.
In the CLARITY images of this study, a neuron's brightness is proportional to its activity, which means that CLARITY images have a functional component.
Regions which appear bright in one CLARITY brain may appear dark in another (Fig.~\ref{fig:cerebralPeduncle}).
Registration is further complicated by brain deformation introduced in the clarifying process (Fig.~\ref{fig:deformed}) and missing data (Fig.~\ref{fig:missingData}).
\begin{figure}[ht]
\centering
\subfloat[]{\label{fig:cerebralPeduncle} \includegraphics[height=1.20in]{cerebralPeduncle} }
\subfloat[]{\label{fig:deformed}\includegraphics[height=1.20in]{deformed} }
\subfloat[]{\label{fig:missingData}\includegraphics[height=1.20in]{missingData} }
\caption{Arrows point out features which make registering CLARITY images challenging.
In (a) the cerebral peduncle is light in one CLARITY brain but dark in another.
In (b) the brain was greatly deformed in the clarifying process and in (c) the brain is missing data.}
\label{fig:challenges}
\end{figure}
\section{METHODS}
The registration pipeline from the preceding study \cite{Kutten} was reimplemented using SimpleITK.
This pipeline, known henceforth as the Mask-LDDMM pipeline, registered images using their masks.
Additionally, an Image-LDDMM pipeline which directly registered images using intensity was developed.
\subsection{Image Acquisition}
12 CLARITY mouse brains (5 wild type controls and 7 behaviorally challenged) were imaged using CLARITY-Optimized Light-sheet Microscopy (COLM) (whole brain COLM imaging and data stitching performed by R. Tomer, in preparation).
In brief, raw volumes were acquired in 0.585 $\mu$m x 0.585 $\mu$m resolution slices with a slice spacing of 5 to 8 $\mu$m.
Images were stored in the NeuroData cluster at 6 resolution levels, with level 0 being the full resolution and level 5 being the lowest resolution.
To avoid registration complications, four CLARITY brains which were not missing any data (Control239, Challenged178, Challenged199, and Challenged188) were selected to test the pipelines.
\subsection{Mask Generation}
Let $\Omega \subset \mathbb{R}^3$ be the background space.
Let $I_0: \Omega \rightarrow \mathbb{R}$ be the template (moving) image which will be deformed to match target (fixed) image $I_1: \Omega \rightarrow \mathbb{R}$.
Thus for CLARITY-ARA registrations $I_0$ is the the Nissl-stained ARA, $L_0: \Omega \rightarrow \mathbb{N}$ are the corresponding ARA annotations, and $I_1$ is a CLARITY image.
Since the resolution of ARA version 2 is 25 $\mu$m isotropic, CLARITY images were downloaded from the NeuroData cluster and resampled to the same resolution.
ARA mask $M_0: \Omega \rightarrow \{0,1\}$ was generated from $L_0$ by taking the union of all foreground labels.
For each CLARITY image $I_1$, mask $M_1: \Omega \rightarrow \{0,1\}$ was generated using the following procedure.
First, $I_1$ was binary-thresholded to remove the background.
Next, this rough mask was opened using a 50 $\mu$m radius ball-shaped kernel to remove foreground grains.
Finally, the mask was closed by a 125 $\mu$m radius kernel to remove large foreground holes.
\subsection{Preprocessing}
In the \emph{Image-LDDMM Pipeline}, preprocessing consisted of two steps.
In the first step, the brain masks $M_0$ and $M_1$ were applied to the images $I_0$ and $I_1$ respectively.
Next, the masked $I_0$ was histogram-matched to the masked $I_1$.
In the matching procedure, 32-bin histograms were calculated for both the template and target images.
The histograms were matched exactly at 8 quantile points, and by interpolation at all other intensities between these points.
In the \emph{Mask-LDDMM pipeline}, no preprocessing was done.
Instead $I_0$ and $I_1$ were replaced by their corresponding masks, $M_0$ and $M_1$, during registration.
\subsection{Registration}
Registration was then done in a three step process with rigid alignment, followed by affine alignment, and finally deformable registration.
In rigid registration, parameters for the rigid transformation augmented matrix $R \in \mathbb{R}^{4 \times 4}$ were optimized using gradient descent on the mean squared error between the transformed histogram-matched $I_0$ and $I_1$.
For affine alignment, the same optimization scheme and image metric were used to find affine matrix $A \in \mathbb{R}^{4 \times 4}$ between $I_0 \circ R^{-1}$ and $I_1$.
The results of these intermediate steps, the Mask-Rigid/Mask-Affine in the Mask-LDDMM pipeline and Image-Rigid/Image-Affine in the Image-LDDMM pipeline, were stored for quantitative evaluation.
Let $J_0 = I_0 \circ R^{-1}A^{-1}$ and $J_1 = I_1$. Deformable registration was done by the LDDMM algorithm which used gradient descent to minimize the objective function
\begin{equation*}
E(v) = \int_0^1 || L v(t) ||^2_{L_2} dt + \frac{1}{\sigma^2} ||J_0 \circ \phi_{10} - J_1||^2_{L_2}
\end{equation*}
where $v : [0,1] \times \Omega \rightarrow \mathbb{R}^3$ is the velocity of the flow from $I_0$ to $I_1$ and $L = (-\alpha \Delta + \gamma) \mathbf{I_3}$ is a kernel which ensures that $v$ is sufficiently smooth.
The greater $\alpha \in (0, \infty)$ is, the smoother the transform. \cite{Beg05}.
By convention $\gamma=1$ and $\sigma=1$ were used.
The algorithm was implemented in ITK by building upon the library's time-varying velocity field registration method.
To minimize computational cost, velocity field $v$ was discretized into only 4 time steps.
Displacement $\phi_{10} = \int_0^1 v(t,\phi_{10}) dt $ was found by integrating $v$ using a 4th order Runge-Kutta algorithm.
After LDDMM the final transform $\varphi = \phi_{01}AR$ and its inverse $\varphi^{-1} = R^{-1}A^{-1}\phi_{10}$ were found.
\subsection{Postprocessing}
Deformed labels $L_0 \circ \varphi^{-1}$ were resampled to a level 5 resolution and fed into the NeuroData cluster.
The infrastructure automatically propagated the annotations to higher resolution levels.
Thus the full resolution images with ARA labels overlaid could be visualized from a web browser in NeuroDataViz (Fig.~\ref{fig:ndviz}).
\begin{figure}[t]
\begin{center}
\begin{tabular}{c c}
\includegraphics[width=0.4\textwidth]{ndviz1} & \includegraphics[width=0.4\textwidth]{ndviz2}
\end{tabular}
\end{center}
\caption{Viewing CLARITY brain Challenged199 with overlay of ARA labels from the Mask-LDDMM pipeline in a web browser using NeuroDataViz.}
\label{fig:ndviz}
\end{figure}
\subsection{NeuroData Registration in Python}
Image processing and registration code from the pipelines were turned into a python module called \emph{NeuroData Registration (ndreg)}.
Functions from ndreg store results as NIfTI image files to leave a record of each step for easy debugging.
Internally, these functions called SimpleITK or custom binaries.
The module also includes convenience functions for downloading and uploading data to the NeuroData cluster through ndio.
The ndreg module is available as open source software at \url{http://NeuroData.io}.
\subsection{Quantitative Evaluation}
The registrations were quantitatively evaluated using aligned template-to-target mutual information, surface error, and landmark error.
Let $T$ and $U$ be random variables representing the intensities of deformed template image $I_0 \circ \varphi^{-1}$ and target image $I_1$ respectively.
The \emph{Mutual Information (MI)} between these images is
\begin{equation*}
MI(T,U) = \iint_{\mathbb{R}\times\mathbb{R}} p(t,u) \ln \left(\frac{p(t,u)}{p(t)p(u)}\right)dt du
\end{equation*}
where $p(t)$, $p(u)$, and $p(t,u)$ are the deformed template histogram, target histogram, and joint histogram respectively.
Computing MI directly may yield unstable results.
Therefore it was estimated using the \emph{Viola-Wells} method, as implemented in ITK.
As recomended in the doccumentation, a standard deviation of 0.4 was used to smooth the histograms after normalizing the template and target image intensities to a zero mean and unity standard deviation.
Densities $p(t)$, $p(u)$, and $p(t,u)$ were then estimated from 1000 foreground samples using a Gaussian distribution-based Parzen window.
Landmark-based methods were also used to evaluate the results.
Specifically, $N = 55$ landmarks were chosen and placed on the ARA and CLARITY images using the MRI Studio software suite's DiffeoMap program (\url{https://www.mristudio.org/}).
Of the landmarks, 28 were placed on the surface of the brain while the remaining 27 were placed on internal structures.
The \emph{error in position of the $k^{th}$ landmark} after registration is
\begin{equation*}
e_k = d(\varphi(x_k), y_k) = || \varphi(x_k) - y_k ||_{L_2}
\end{equation*}
where $x_k$ and $y_k$ are the positions of the $k^{th}$ template and target landmarks for $k \in \{1,..., N\}$.
The \emph{Hausdorff Distance (HD)} between the deformed template and target surfaces is defined as
\begin{equation*}
d_H(S_1,S(1)) = max \left\{ \sup_{x \in S(1)} \inf_{y \in S_1} d(x,y), \sup_{y \in S_1} \inf_{x \in S(1)} d(x,y) \right\}
\end{equation*}
where $S(1) = \partial(M_0 \circ \varphi^{-1})$ and $S_1 = \partial M_1$ denote the surfaces of the deformed template and target respectively.
Since the Hausdorff Distance is a maximum distance between surfaces, it was usually too sensitive to outlier surface points.
Therefore it was more convenient to compute the \emph{median of the surface distances between all points in $S_1$ and $S(1)$}
\begin{equation*}
d_M(S_1, S(1)) = median\left\{\inf_{y \in S_1} d(x,y) \} \cup \{\inf_{x \in S(1)} d(x,y)\right\}.
\end{equation*}
\section{RESULTS}
\begin{figure}[ht]
\centering
\subfloat[]{\includegraphics[width=0.33\textwidth]{Control239ToARAOverlap} \label{fig:clarityToAra}}
\subfloat[]{\includegraphics[width=0.33\textwidth]{Challenged199ToControl239Overlap} \label{fig:betweenConditions}}
\subfloat[]{\includegraphics[width=0.33\textwidth]{Challenged199ToChallenged188Overlap} \label{fig:withinConditions}}
\vspace{1cm}
\caption{ Selected alignments from the CLARITY-ARA experiment (a), the CLARITY-CLARITY experiment between Challenged and Control (b), and the CLARITY-CLARITY experiment within Challenged (c).
In each, the left column shows the template image and landmarks (in cyan). The right column shows the target image and landmarks (in red).
The middle column shows the aligned template image, surface, and landmarks overlaid on the target image, surface, and landmarks. }
\label{fig:expriments}
\end{figure}
Three experiments were performed.
In the first experiment, CLARITY images were registered directly to the ARA.
In the next experiment, CLARITY images were registered to other CLARITY images of different conditions (Challenged to Control).
In the final experiment, CLARITY images were registered to other CLARITY images of the same condition (Challenged to Challenged).
Registration results from the Image-LDDMM and Mask-LDDMM pipelines were compared to a MI-BSpline pipeline described in an upcomming work by Tomer \textit{et al.}
The pipeline registers images using their Mutual Information under a B-Spline spatial transformation.
The results of these experiments are summarized in Fig.~\ref{fig:expriments}-\ref{fig:summary}.
\subsection{CLARITY-ARA}
Fig.~\ref{fig:clarityToAra} and the green lines in Fig.~\ref{fig:summary} show the results of CLARITY to ARA registrations.
The Mask-LDDMM pipeline consistently outperformed the MI-BSpline and Image-LDDMM pipelines in surface and landmark distances.
Interestingly, the Image-LDDMM pipeline consistently yielded higher MI values than the Mask-LDDMM and MI-BSpline pipelines.
Despite this, visual inspection of Image-LDDMM results in Fig.~\ref{fig:clarityToAra} revealed that the alignments were relatively poor.
This can likely be attributed to the great difference in appearance between the Nissl-stained ARA and CLARITY images.
\subsection{CLARITY-CLARITY between conditions}
Fig.~\ref{fig:betweenConditions} and the blue lines in Fig.~\ref{fig:summary} show CLARITY to CLARITY registration results between conditions.
The MI-BSpline method did better than the Image and Mask-LDDMM pipelines in MI and median landmark error.
Mask-LDDMM, Image-LDDMM, and MI-BSpline gave similar surface distance results.
\subsection{CLARITY-CLARITY within a condition}
Fig.~\ref{fig:withinConditions} and the magenta lines in Fig.~\ref{fig:summary} show the registration results of CLARITY to CLARITY registrations within a condition.
As in the inter-condition registrations, Image-LDDMM yielded better MI values than Mask-LDDMM.
But unlike the inter-condition registrations, Image-LDDMM and Mask-LDDMM had lower median landmark errors than the MI-BSpline method.
Once again Mask-LDDMM, Image-LDDMM, and MI-BSpline gave comparable surface distance results.
\begin{figure}[th]
\includegraphics[width=\textwidth]{summary}
\caption{ A summary of metric values for all experiments over each registration method. }
\label{fig:summary}
\end{figure}
\section{CONCLUSION}
In most cases, the Mask-LDDMM pipeline outperformed both Image-LDDMM and MI-BSpline in aligning CLARITY brains with the ARA.
The MI-BSpline pipeline gave better results than Image-LDDMM in CLARITY-CLARITY transforms between conditions.
Image-LDDMM outperformed MI-BSpline in CLARITY-CLARITY transforms within a condition.
However, there are some limitations to these findings.
Human error in landmark placement may have been a factor in these results.
Furthermore, the pipelines were tested on brains without missing data.
When only partial data is available, complete brain masks cannot be constructed and Mask-LDDMM should not be used.
\acknowledgments
This work was graciously supported by the Defense Advanced Research Projects Agency (DARPA) SIMPLEX program through SPAWAR contract N66001-15-C-4041 and DARPA GRAPHS N66001-14-1-4028.
|
\section{Introduction}
\begin{flushright}
{\scriptsize \textit{%
Like a brush in hand\\
To paint a picture\\
Of what we would like to see}\\
C.M. Schuldiner}
\end{flushright}
In this Section we will motivate the approach of this paper by discussing how it is related to the structures and problems of K\"{a}hler geometric and C$^*$-algebraic approaches to quantum kinematics, as well as hamiltonian and path integral approaches to quantum dynamics. Our goal is to overview and explain the merits of the constructions that we pursue on the subsequent pages. The in-depth analysis of an approach to foundations of quantum theory\footnote{We distinguish between the quantum mechanics, understood as a framework defined in \cite{vonNeumann:1932:grundlagen}, and the quantum theory, understood as a (currently unknown) framework that should be capable of providing mathematically exact nonperturbative foundations for relativistic quantum field theory and nonequilibrium quantum statistical mechanics.} that we briefly postulate and use here is provided in another paper \cite{Kostecki:2016:microhauptwerk}.
\subsection{Quantum information geometric foundations: global postulates\label{qig.foundations.intro.section}}
The basic kinematic postulates of our framework are:
\begin{itemize}
\item \textbf{Postulate 1:} \textit{The underlying spaces of inquiry are W$^*$-algebras $\N$, instead of sample spaces and Hilbert spaces.}
\item \textbf{Postulate 2:} \textit{The state spaces of quantified knowledge are sets $\M(\N)$ of positive normal states on W$^*$-algebras, instead of probabilistic models and spaces of density matrices.}
\item \textbf{Postulate 3:} \textit{The observables are arbitrary real valued functions $f:\M(\N)\ra\RR$ of normal states, instead of arbitrary real valued functions on sample spaces and self-adjoint operators.}
\item \textbf{Postulate 4:} \textit{Given an experimental configuration space $\Theta$, a method of model construction defining a mapping $\Theta\ni\theta\mapsto\phi(\theta)\in\M(\N)$, and a choice of the set of functions $\tilde{f}:\Theta\ra\RR$ that one is interested in, the set of observables that are relevant for a given problem is given by $\{f:\M(\N)\ra\RR\mid f\circ\theta=\tilde{f}\}$.}
\end{itemize}
While C$^*$-algebras\footnote{A C$^*$-algebra is defined as an algebra $\C$ over $\CC$, equipped with: an operation $^*:\C\ra\C$ that algebraically abstracts the properties of a complex conjugation of complex numbers, and a norm $\n{\cdot}:\C\ra\RR^+$ that turns $\C$ into a Banach space satisfying $\n{x^*x}=\n{x}^2$ $\forall x\in\C$. W$^*$-algebras are characterised as such C$^*$-algebras for which there exists a Banach space, denoted $\C_\star$ and called a (Banach) predual, satisfying $(\C_\star)^\star\iso\C$. Given a Banach space $X$ over $\CC$, the operation $^\star:X\ra X^\star$ forms a Banach space of all continuous linear $\CC$-valued functionals, equipped with a supremum norm.} generalise algebras of complex continuous functions on compact topological spaces, W$^*$-algebras generalise $L_\infty$ spaces over localisable boolean algebras (or, equivalently, localisable measure spaces), so the problem of choice between them depends not only on the mathematical properties of a specific application but also on the general interpretation assigned to the quantum theoretic formalism. From the mathematical perspective of general integration theory (including integration on noncommutative W$^*$-algebras, on nonassociative Jordan algebras, and on spectral convex sets), it is completely natural to extend considerations from the sets of density matrices to the subsets of positive parts of Banach preduals of arbitrary W$^*$-algebras. The set of all (not necessarily normalised) density matrices is characterised as a positive part $\ell_1(\BH)^+$ of a noncommutative $\ell_1$ space associated to the W$^*$-algebra $\BH$ of bounded linear operators on a Hilbert space $\H$ (more precisely, $\ell_1(\BH)\iso\schatten_1(\H)$, where $\schatten_1(\H)$ is a Banach space of all trace class operators, equipped with a trace norm), while the Banach preduals $\N_\star$ of arbitrary W$^*$-algebras $\N$ are characterised as a noncommutative $L_1$ spaces associated to these algebras, and this association is functorial with respect to W$^*$-isomorphisms \cite{Falcone:Takesaki:2001}. Hence, if one considers quantum mechanics and probability theory as two instances of a more general class of information theories, then the use of W$^*$-algebras $\N$ and elements of $L_1(\N)^+\iso\N_\star^+$ for the mathematical foundations of quantum mechanics is a \textit{natural} and \textit{exact} generalisation of mathematical formulation of probability theory in terms of a normalised measure theory (as proposed by Steinhaus \cite{Steinhaus:1923} and developed by Kolmogorov \cite{Kolmogorov:1933}). This leads us: 1) to chose the framework of W$^*$-algebras $\N$ and normal positive states $\omega\in\N^+_\star$ instead the framework of C$^*$-algebras $\C$ and positive states $\omega\in\C^{\star+}$; 2) to consider subsets $\M(\N)\subseteq\N^+_\star$ as spaces of quantum states that are setting the arena for quantum kinematics. We view them as natural generalisation of probabilistic models $\M(\X,\mho(\X),\tmu)\subseteq L_1(\X,\mho(\X),\tmu)^+$. The observables in our framework are defined as \textit{arbitrary} real valued functions $f:\M(\N)\ra\RR$. The observables in the standard sense of quantum mechanics are precisely determined as an affine subset of the observables in our sense:
\begin{equation}
f_x(\phi):=\phi(x)\;\;\forall x\in\N^\sa\;\forall\phi\in\N^+_\star.
\end{equation}
Postulates 1-4 do not yet determine how we are going to define specific kinematic and dynamic models. Following Chencov's geometric approach to foundations of statistical inference theory \cite{Chencov:1964,Chencov:1965,Chencov:1966,Chencov:1967,Chencov:1968,Chencov:1969,Chencov:1972,Chencov:1978,Chencov:1980,Chencov:1987,Morozova:Chencov:1991} (developed later by Amari and others \cite{Amari:1968,Amari:1980,Amari:1982,Amari:1985,Amari:1987,Amari:Nagaoka:1993}), Jaynes information theoretic approach to foundations of statistical mechanics \cite{Jaynes:1957,Jaynes:1957:2,Jaynes:1963,Jaynes:Scalapino:1963,Jaynes:1965,Jaynes:1967,Jaynes:1979:where:do:we:stand,Jaynes:1985:macropred,Jaynes:1986,Jaynes:1993} (developed later in the geometric direction by Ingarden and others \cite{Ingarden:1963,Ingarden:Kossakowski:1975,Ingarden:1974:1976,Ingarden:1978,ISSK:1979,Ingarden:1981,Ingarden:Janyszek:1982,IJKK:1982,IKO:1997}), and the program of smooth geometrisation of quantum mechanics \cite{Strocchi:1966,Marsden:1968:generalized,Kibble:1979,Cirelli:Lanzavecchia:Mania:1983,Simon:1983,Cirelli:Lanzavecchia:1984,Heslot:1985,Page:1987,Aharonov:Anandan:1987,Anandan:Aharonov:1990,Cirelli:Mania:Pizzocchero:1990,Gibbons:1992,Cirelli:Mania:Pizzocchero:1994,Hughston:1995,Hughston:1996,Schilling:1996,Field:1996,Field:1997,Brody:Hughston:1998,Ashtekar:Schilling:1999,Cirelli:Gatti:Mania:1999,Brody:Hughston:2001,Chruscinski:Jamiolkowski:2004,Bengtsson:Zyczkowski:2006}, we propose the following
\begin{itemize}
\item \textbf{Postulate 5:} \textit{The construction of specific models of kinematics and dynamics is based upon the geometric structures over state spaces, provided by quantum relative entropies and Banach Lie algebras, instead of scalar product of Hilbert space.}
\end{itemize}
In order to investigate the possible generalisations of quantum mechanical prescriptions of dynamics we want first to understand the geometric structures on state spaces. For a given Hilbert space $\H$ equipped with a scalar product $\s{\cdot,\cdot}_\H:\H\times\H\ra\CC$, the projection $\PP:\H\setminus\{0\}\ni\xi\mapsto\frac{\xi}{\n{\xi}_\H}\in\PP\H$ induces a manifold structure on $\PP\H$, with tangent spaces given by $\H$, riemannian metric $\gbold^\H$ and symplectic form $\wbold^\H$ determined uniquely by a decomposition \cite{Abraham:Marsden:1978}
\begin{equation}
\s{\cdot,\cdot}_\H=\frac{1}{2}\gbold^\H(\cdot,\cdot)+\ii\frac{1}{2}\wbold^\H(\cdot,\cdot),
\end{equation}
and complex structure defined by $\s{\cdot,\jbold^\H(\cdot)}_\H=\ii\s{\cdot,\cdot}_\H$. The tangent bundle of $\H$ over $\PP\H$ can be viewed as a principal $U(1)$-bundle equipped with a $U(1)$ connection $1$-form $\nabla^\PP:\H\times\H\ni(\xi,\zeta)\mapsto\ii\s{\xi,\ddd\zeta}_\H\in\CC$. In our case, the lack of a unique global Hilbert space implies the lack of unique specification of riemannian metric and symplectic form derived from a scalar product. In order to facilitate a well-defined generalisation of riemannian and symplectic structure, the sets $\M(\N)$ can be equipped with two different smooth real Banach manifold structures. On one hand, an information manifold structure on the set $\N^+_{\star0}$ of all faithful (strictly positive) elements of $\N^+_\star$ is constructed by a choice of a quantum relative entropy functional on $\N^+_\star$, and has tangent spaces defined as noncommutative Orlicz spaces that provide suitable convergence behaviour of neighbourhoods of states as measured by constrained relative entropy minimisation (see Section \ref{relat.entr.MCP.bundle}). On the other hand, a Banach Lie--Poisson manifold structure is constructed by the choice of a Banach Lie algebra (such as the real Banach Lie algebra $\N^\sa$ of all self-adjoint elements of a W$^*$-algebra $\N$), and has tangent spaces defined as copies of a predual of this algebra (see Section \ref{W.star.BLP.section}). Given the information distance manifold structure on $\M(\N)$, the first two nonzero orders of Taylor expansion for a wide class of relative entropy functionals $D(\cdot,\cdot)$ on $\M(\N)$ give rise to the torsion-free smooth Norden--Sen geometries $(\M(\N),\gbold^D,\nabla^D,(\nabla^D)^\nsdual)$. The Norden--Sen geometry is defined by the condition \eqref{geometric.duality}, which directly generalises the condition characterising the Levi--Civita affine connection $\nabla^\gbold$ in riemannian geometries $(\M,\gbold)$. The Fubini--Study riemannian metric $\gbold^\H$ becomes recovered as an extension to the boundary of pure states for a wide class of geometries $(\M(\N),\gbold^D)$ \cite{Petz:Sudar:1999}. On the other hand, given the choice of a Banach--Lie algebra $\B$ such that $\M(\N)$ is a real Banach submanifold\footnote{For a discussion why an injective immersion of $\M(\N)$ into $\B_\star$ is not sufficient, see \cite{Bona:2004}.} of $\B_\star$ with $\T_\phi\B_\star\iso\B_\star$ $\forall \phi\in\B_\star$ (or, more generally, $\ad^\star_x(\B_\star)\subseteq\B_\star$ $\forall x\in\B$), the coadjoint action of $\B$ on $\B_\star$ induces a Poisson structure on Fr\'{e}chet smooth real valued functions on $\M(\N)$. If $\B$ is a Lie algebra of a group $G$, then the Banach Lie--Poisson manifolds $\M(\N)$ are symplectic if they are coadjoint orbits of $G$. In particular, if $\B\iso\N^\sa$, then $G\iso\N^\uni$ (a group of all unitary elements of $\N$), and $\B_\star\iso\N^\sa_\star:=\{\phi\in\N_\star\mid\phi(x^*)=\phi(x)^*\}\iso(\N^\sa)_\star$. The example of such case is given by the orbit of a group of unitary operators of density matrices with finite fixed number $n\in\NN$ of nonzero eigenvalues \cite{Bona:2000}. For $n=1$, one recovers precisely the symplectic structure $\wbold^\H$ on a projective Hilbert space $\PP\H$ that is induced from $\s{\cdot,\cdot}_\H$ on $\H$ \cite{Bona:1991}.
The resulting description of quantum geometry can be summarised as follows:
\begin{enumerate}
\item[(a)] In the general setting of W$^*$-algebras $\N$ and quantum states defined as elements of $\N_\star^+$ the smooth manifold structure required to implement the infinite-dimensional quantum generalisation of Poisson geometry does not match with the smooth manifold structure required to implement the infinite-dimensional quantum generalisation of riemannian geometry (this issue is discussed in more details in \cite{Kostecki:2016:microhauptwerk}). As a result, the geometry of a Hilbert space $\H$ (consisting of pure states), formulated in terms of riemannian metric $\gbold^\H$ and symplectic structure $\wbold^\H$ defined over the same real Hilbert smooth manifold $\PP\H$, becomes generalised to the geometry of spaces $\M(\N)$ equipped with two different real Banach smooth manifold structures: of a Banach Lie--Poisson manifold and of a quantum information geometric (relative entropic) manifold. The former is determined by the choice of a Banach Lie algebra $\B$, and in the special cases reduces to a symplectic space. The latter is determined by the choice of a relative entropy functional $D:\M(\N)\times\M(\N)\ra[0,\infty]$, and in the special cases reduces to a torsion free Norden--Sen manifold, or just a riemannian space.
\item[(b)] Apart from the above two alternative systems of tangent, cotangent, and higher jet bundles, one can also introduce a bundle of Hilbert spaces over $\M(\N)\subseteq\N^+_\star$, that can serve as as an ambient framework to represent different geometrical objects. A natural candidate is a Gel'fand--Na\u{\i}mark--Segal bundle $\H\M(\N)$ of Hilbert spaces \cite{Odzijewicz:Slizewska:2011}. Because the bundle $\H\M(\N)$ is defined by states of an underlying manifold (as opposed to a projection $\PP$), there is no $\nabla^{\PP}$ connection $U(1)$ action in fibers. However, for $\M(\N)\subseteq\N^+_{\star0}$, each fibre of $\H\M(\N)$ bundle is equipped with a strongly continuous $U(1)$ action of a modular (Tomita--Takesaki) automorphism $\sigma^\omega:\RR\ni t\mapsto\Ad(\Delta_\omega^{\ii t})\in\Aut(\pi_\omega(\N))$. We will study the role of this automorphism in Section \ref{algebraic.action.operator.section}, showing that it resembles some interesting similarities with $\nabla^{\PP}$ when considered over a trajectory $\RR\ra\M(\N)\subseteq\N^+_{\star0}$. Yet, in Section \ref{quant.info.geom.section} we will show that, for any $\M(\N)\subseteq\N^+_{\star0}$, the bundle $\H\M(\N)$ carries a natural connection structure, with parallel transport given by the standard unitary transition operators $V_{\phi,\omega}:=J_{\phi,\phi}J_{\phi,\omega}$, where $J_{\phi,\omega}$ is a relative modular conjugation between two faithful normal GNS representations, while $J_{\phi,\phi}$ is a Tomita modular conjugation. We observe that this connection is Levi-Civita with respect to the Wigner--Yanase riemannian metric, and its local geodesic free fall corresponds to constrained minimisation of the Hilbert space norm (projective measurement).
\item[(c)] Because two manifold structures mentioned in (a) do not coincide, there seems to be no obvious candidate for a `complex structure' on a general quantum model $\M(\N)$. However, we notice that the use of complex Hilbert spaces of dimension $n$ instead of real Hilbert spaces of dimension $2n$ is crucially associated with the requirement that the generators of unitary transformations of these spaces should be represented by self-adjoint operators (identified with observables). Observing further that the standard quantum mechanical method of defining relevant observables proceeds by representations of Lie algebra $\glie$ of Lie group $G$ on the complex Hilbert space, we introduce the structure of a principal $G$-bundle $E$ over $\M(\N)$, equipped with a family of representations of an associated Lie algebra in the fibers of the GNS Hilbert space bundle $\H\M(\N)$. This leads to a rise of a $\glie$-valued connection form on $E$, and a fiberwise family of its representations on the fibers of $\H\M(\N)$. This structure exhibits some interesting features of the relationship between Berry connection, complex structure of a Hilbert space, and construction of observables by means of representations of Lie algebras. The bundle structures of $E$ and $\H\M(\N)$ do not require $\M(\N)$ to be a Banach smooth manifold, but they require it to be a (Hausdorff and paracompact) topological space. We do not assume any \textit{a priori} relationship between the geometries of local causal dynamics, as provided by $\B$, and the transformations used to identify locally the relevant observables, as provided by $\glie$, because we consider it to be a model-dependent feature.
\item[(d)] Using the $\glie$-valued connection $\nabla^\glie$, we can define a kinematic propagator (in Prugove\v{c}ki sense) of the particles (in Wigner sense) as the holonomy of $\nabla^\glie$ along the geodesics of $\nabla^D$ (or $(\nabla^D)^\nsdual$ or $\nabla^{\gbold^D}$) affine connection.
\item[(e)] In Section \ref{MCP.bundle.section} we propose a construction of what we call the Morozova--Chencov--Petz Hilbert space bundle $\H^{\hhh}\M(\N)$. Its purpose is to use the riemannian metrics $\gbold^D$, associated with a class $D_\fff$ of quantum distances, in order to determine Hilbert spaces and corresponding representations that are different from the GNS construction, and include the information about the local riemannian geometry of a model (this is inspired by the ideas of \cite{Sukhanov:Rudoy:2005,Rudoy:Sukhanov:2005,Rudoy:2009}). In principle, it can be used as an alternative to $\H\M(\N)$ (especially for the purpose of the tasks (c)-(d)), however it is essentially harder to deal with mathematically. We consider this as an indication that the natural framework for a simultaneous implementation of geometric and algebraic tools used in this paper (entropic Norden--Sen geometries, Banach Lie--Poisson structure, perturbations of liouvilleans) are Banach dual pairs of noncommutative Orlicz spaces, used as tangent and cotangent spaces. However, the technical implementation of this idea requires one to develop a standard construction of Orlicz spaces for any (countably finite) W$^*$-algebra, and for a large class of quantum relative entropies, as well as to develop the theory of perturbations of ``Orlicz liouvilleans'' (by an analogy to $L_p$-liouvilleans of Jak\v{s}i\'{c} and Pillet \cite{JOPP:2012}). While these tasks are beyond the scope of this paper, it can be understood as a testing ground for them, so we will discuss more precisely the above ideas in Sections \ref{local.infogeometry.histories.intro.section} and \ref{effective.local.quantum.dynamics.section}.
\end{enumerate}
With all these tools on the stage, we can approach the problem of construction of quantum dynamics.
\begin{itemize}
\item \textbf{Postulate 6:} \textit{The elementary form of causal dynamics is given by a Poisson flow generated by a smooth observable on a state space, as opposed to a unitary evolution on a Hilbert space. The elementary form of inferential dynamics is given by a (nonlinear) constrained quantum relative entropy minimisation, as opposed to a (linear) projection in a Hilbert space or L\"{u}ders' rule.}
\end{itemize}
Unitary evolution is a special case of a hamiltonian evolution on the self-adjoint part of a predual of a W$^*$-algebra, understood as a BLP space \cite{Bona:1991,Bona:2000,Odzijewicz:Ratiu:2003}. Derivation of L\"{u}ders' rules (selective and nonselective) as a special case of constrained quantum relative entropy minimisation\footnote{Technically, $\PPP^D_\Q(\phi)$ is a nonlinear projection in a positive cone of a noncommutative $L_1(\N)$ space, constrained by the data represented as a convex closed set $\Q_\Upsilon$ in a noncommutative Orlicz space $L_\Upsilon(\N)$, associated by means of a nonlinear bijective mapping $\ell:L_1(\N)\ra L_\Upsilon(\N)$, so that $\Q=\ell^{-1}(\Q_\Upsilon)$ forms a codomain of a projection. While in general we are interested in entropic projections for a wide class of quantum distance functionals, the results cited in this paragraph were proven for the most recognised example: the Umegaki--Araki noncommutative generalisation of the Kullback--Leibler distance.} (in short: entropic projection) for $D$ given by the Umegaki--Araki distance \eqref{Araki.std.rep.distance} was provided in \cite{Hellmann:Kaminski:Kostecki:2016} and \cite{Kostecki:2014}.\footnote{Derivation of the nonselective L\"{u}ders' rule from minimisation of the Hilbert--Schmidt norm distance was provided much earlier in \cite{Herbut:1969}, while derivations of some special cases of the selective L\"{u}ders' rule using some symmetric quantum information distances were obtained in \cite{Hadjisavvas:1978,Hadjisavvas:1981,Dieks:Veltkamp:1983,Raggio:1984}. See \cite{Kostecki:2014} for more discussion.} Hence, the basic dynamical setting of quantum mechanical evolution of quantum states, which is a unitary evolution followed by a projective measurement can be completely recovered as a special case of a \textit{causal inference instrument} given by the map\footnote{The composition in the reverse order can also be studied.}
\begin{equation}
\M_1(\N)\ni\phi\mapsto\PPP^D_\Q\circ w^{\B,h}_t(\phi)\in\M_2(\N),
\label{global.hamiltonian.entropic.map}
\end{equation}
where
\begin{equation}
\PPP^D_\Q(\psi):=\arginff{\psi\in\Q}{D(\omega,\psi)}
\end{equation}
is an entropic projection onto constrained set $\Q$ for a quantum distance functional $D$, while $w^{\B,h}_t$ is a Poisson flow generated by a Banach Lie algebra $\B$ and a hamiltonian function $h$ for a time range $[0,t]$, corresponding uniquely to an integral line of a vector field
\begin{equation}
\XXX_h(\phi):=-\ad^\star_{\DF_\phi h}(\phi)\;\;\forall\phi\in\B_\star,
\end{equation}
where $\DF_\phi h$ is a Fr\'{e}chet derivative of $h$ at $\phi$, implementing the differential form $\ddd h(\phi)$.
In \cite{MunkNielsen:2015} it was shown that partial trace is also a special case of entropic projection. Hence, all linear completely positive maps can be considered as a special case of the maps formed by composition of tensor products, Poisson flows, and entropic projections \cite{Kostecki:MunkNielsen:2016,Kostecki:2016:microhauptwerk}. This way the kinematic and dynamic setting of quantum mechanics and nonrelativistic quantum information theory becomes fully recovered as a special case of the framework specified by Postulates 1-6 above.
This leads us to:
\begin{itemize}
\item \textbf{Open problem:} \textit{Reconstruct (some aspects of) dynamics of quantum field theory and nonequilibrium quantum statistical mechanics using the framework specified by Postulates 1-6.}
\end{itemize}
Unlike in quantum mechanics, the dynamics of both these theories is sensitive to local geometric features of the kinematic structure of a quantum model. In consequence, we are lead to investigate how, and to what extent, the above geometric structures, and the corresponding nonlinear dynamical maps, can give account of the local structures in QFT and NQSM.
\subsection{Local quantum information dynamics in algebraic and path integral approaches\label{local.quantum.dynamics.intro.section}}
The main questions underlying the constructions carried on in this paper are: what if the correct setting for bridging the gap between algebraic and path integral setting is to use quantum state spaces and their geometry:
\begin{enumerate}
\item[1)] to define local evolution in the algebraic approach by means of \textit{locally} defined and perturbed liouvilleans?
\item[2)] \textit{instead} of using phase space geometry in the continuous time coherent state path integral ``quantisation''?
\item[3)] to describe renormalisation as purely information theoretic procedure?
\end{enumerate}
The discussion below is intended to show that the proposal of the geometric framework for locally quantum information theories that we provide in Section \ref{locally.quantum.information.relativity.section} is remarkably grounded in the insights coming from \textit{three} very distinct theoretical frameworks: 1) a geometric extension of algebraic hamiltonian dynamics with local gauge and local sources by means of local perturbation of liouvilleans; 2) a generalisation of the Daubechies--Klauder path integration to an algebraic setting; 3) a geometric Jaynes--Mitchell--Favretti renormalisation, applicable in nonequilibrium quantum statistical mechanics.
The standard Haag--Kastler \cite{Haag:Kastler:1964,Haag:1992} setting of an algebraic approach to quantum field theory is widely considered as being unable to incorporate the local gauge principle\footnote{E.g. \cytat{The Lagrangean and the Feynman path integral are at present indispensable tools in the characterization and study of a specific theory. Together with the local gauge principle they pose questions which in the algebraic approach are not understood and should be tackled.} \cite{Haag:1992}.} (the global gauge principle has been partially incorporated to an algebraic approach by means of the Doplicher--Haag--Roberts theory \cite{Doplicher:Haag:Roberts:1969:1,Doplicher:Haag:Roberts:1969:2}). Apart from renormalisation techniques, this principle is a fundamental tool in the construction of the predictively sound models in quantum field theory. Its maintenance by the lagrangean/path-integral approach leads to an abandonment of the algebraic approach by most of the practitioners of QFT, but this is provided at the price of replacing mathematically well-defined objects by symbolic (and usually perturbative) techniques of calculations. This makes QFT very different from quantum mechanics, because the latter facilitates construction of predictive models without the expense of mathematical precision. In this paper we intend to show that the consideration of geometric structures on the spaces $\N^+_\star$ of normal states over W$^*$-algebras $\N$, as well as construction of effective local dynamics by means of local perturbations of liouvilleans, may provide an extension of an algebraic approach capable of dealing with the local gauge principle and the use of `external sources', typical in the path integral formalism. Our point of departure from the Haag--Kastler perspective is to consider locality in the Prugove\v{c}ki sense \cite{Prugovecki:1992}, being associated with the fiber at a given point of an underlying space (so the Lorentz or Poincar\'{e} covariance condition is to be applied fiberwisely), as opposed to a neighbourhood of this point (e.g. as given by the special relativistic diamond). This allows to think of local GNS Hilbert space associated to the manifold of quantum states as (a model of, or as a container of) local tangent space, corresponding to a local quantum mechanical description provided by a single (quantum bayesian) user. Further extension from the bundle of (self-dual) Hilbert spaces to the bundle of dual pairs of noncommutative Orlicz spaces is necessary to allow the geometry of local quantum inference to be governed by a wide class of conventions of estimation, beyond the Wigner--Yanase riemannian metric (so that the quantum nonequilibrium thermodynamic Kubo--Mori--Bogolyubov metric, as well as the quantum estimation theoretic Helstrom--Uhlmann--Bures metric can be included on the equivalent mathematical footing). The construction of an underlying manifold structure commits to the principle of equivalence between local inference by means of constrained maximisation of relative entropy, and the free fall along the geodesics of the dually flat local geometry, derived from this entropy.
On the path integral side, our approach is directly inspired by the Daubechies--Klauder \cite{Daubechies:Klauder:1985,Klauder:1988,Klauder:1995,Klauder:Maraner:1997,Bodmann:2003,Klauder:2011} continuous-time regularised coherent state phase space approach to path integration, and the closely related Ana\-sto\-pou\-los--Savvidou \cite{Anastopoulos:2001,Anastopoulos:Savvidou:2002,Anastopoulos:Savvidou:2003} analysis of decoherence functional in the Isham--Linden quantum histories approach. Both have shown that one can think of the underlying dynamical objects of respective theories (path integrals and decoherence functionals) as consisting of the hamiltonian evolution perturbed by the geometric structures on the space of quantum states. Our goal here is to follow Klauder's remark \cytat{If there is ever any hope to define path integrals rigorously as path integrals over a set of paths (functions of time), then it is \textit{essential} to give up the notion that the paths involved are sharp value paths and replace that with another interpretation of which the expectation value paths is a completely satisfactory example} \cite{Klauder:2003} by extending these approaches to the state spaces over W$^*$-algebras, and relating them with the local liouvillean approach to algebraic dynamics. The virtue of the Daubechies--Klauder approach is that it provides a mathematically rigourous continuous time regularisation of the functional integral in a way that gives the same results under arbitrary canonical transformations of the underlying phase space. This is not true for most of other approaches to quantisation, not only path integral based. The restrictions on the class of hamiltonians that are allowed in order to maintain this procedure to be well defined are quite mild. The key ingredient of this approach is introducing a regulariser that represents a riemannian metric on the phase space (corresponding to a Fubini--Study metric on coherent quantum states), and determines a pinned Wiener measure of the Brownian process on the phase space.
The heuristic ideas underlying our treatment of quantum dynamics are: 1) Quantum kinematics and dynamics should be defined without recourse to classical models and their quantisation; 2) Classical (phase space, but also space-time) geometry should be considered as a locally emergent feature describing particular properties of the multi-agent information processing systems and not as a fundamental structure (background); 3) Local spatial (phase space or space-time) variables should arise as epistemic (e.g. operational) parameters of information states (see e.g. \cite{Rodriguez:1999,Duch:Kostecki:2011}). This heuristics is in a disagreement with the perspectives of the orthodox algebraic and path integral approaches (yet, there are some exceptions\footnote{In particular, \cytat{the interpretation of the formal path integral (...) in terms of paths $\mathrm{p}(t)$ and $\mathrm{q}(t)$ for which the meaning of the variables is that of expectation values is far more acceptable than the one in which the meaning is that of both sharp position and sharp momentum (eigen)values. (...) One is almost tempted to assert that the usual interpretation in terms of sharp eigenvalues is ``wrong'', because it cannot be consistently maintained, while the interpretation in terms of [expectation] values is ``right'', because it can be consistently maintained} \cite{Klauder:1997}.}), however we consider this disagreement as a virtue, because it allows us to learn something new.\footnote{See \cite{Kostecki:2007:qht,Kostecki:2008:aqh,Kostecki:2010:aqh,Kostecki:2010:AIP} for some wi(l)der heuristic ideas that have lead to the current work.}
The main mathematical tools used in what follows are: the Hilbert space bundle over $\N^+_\star$ arising from the Gel'fand--Na\u{\i}mark--Segal representation, introduced recently by Odzijewicz and Sli\.zewska \cite{Odzijewicz:Slizewska:2011}, the Derezi\'{n}ski--Jak\v{s}i\'{c}--Pillet theory \cite{DJP:2003} of unbounded perturbation of standard liouvilleans, the B\'{o}na--Odzijewicz--Ratiu construction \cite{Bona:1991,Bona:1993,Bona:2000,Odzijewicz:Ratiu:2003} of the Banach--Lie--Poisson manifold structure on the self-adjoint part $\N^\sa_\star$ of the Banach predual of $\N$, Jen\v{c}ov\'{a}'s construction \cite{Jencova:2006,Jencova:2010} of real Banach smooth manifold over the spaces $\N^+_{\star01}$ of normalised faithful (strictly positive) elements of $\N^+_\star$, based on quantum relative entropic perturbations of states and noncommutative Orlicz spaces, the Daubechies--Klauder continuous time regularised coherent state path integrals \cite{Daubechies:Klauder:1985,Klauder:1988,Klauder:1997,Watson:Klauder:2002,Klauder:2011}, and Favretti's geometrisation \cite{Favretti:2007} of the Jaynes--Mitchell source theory \cite{Mitchell:1967,Jaynes:1985:scattering,Jaynes:1993,Grandy:1987,Grandy:2008}. Section \ref{geometric.structures.section} provides an introduction to most of these mathematical tools. More specifically, Section \ref{W.star.BLP.section} is intended as an introduction to the Banach--Lie--Poisson structure of preduals of W$^*$-algebras. This material is mostly based on papers \cite{Bona:2000,Odzijewicz:Ratiu:2003}. In Section \ref{standard.liouvilleans.section} we review the construction of a standard liouvillean (including standard representation and Haagerup's theorem), some relative modular theory, as well as the Odzijewicz--Sli\.zewska construction of the GNS bundle of Hilbert spaces. In Section \ref{relat.entr.MCP.bundle} we discuss some results from quantum information geometry which we will use in the subsequent sections, including Jen\v{c}ov\'{a}'s construction of a smooth quantum manifold structure. The Daubechies--Klauder approach is discussed in Section \ref{Hilbert.space.geometry.path.integrals}, while Favretti's approach is discussed in Section \ref{Favretti.section}.
\subsubsection{Locally perturbed liouvilleans\label{locally.perturbed.liouvilleans.intro.section}}
Section \ref{algebraic.hamiltonian} provides an elementary analysis of the relationship between W$^*$-dynamical systems, hamiltonian flows on BLP spaces, and standard liouvilleans of the W$^*$-dynamical systems $(\N,\RR,\alpha)$. In order to study the relationship of the BLP structure with the usual usage of standard liouvilleans\footnote{By the Haagerup theorem \cite{Haagerup:1975:standard:form} for standard representations of W$^*$-algebras, for every pair of a W$^*$-dynamical system $(\N,\RR,\alpha)$ and a standard representation $(\H,\pi,J,\stdcone)$ of a W$^*$-algebra $\N$, there exists a unique unitary evolution on $\H$ that represents $\alpha$ leaving $\stdcone$ unchanged. Its generator is an unbounded operator, called the \textit{standard liouvillean}.}, we begin with characterisation of the class of weakly-$\star$ continuous representations $\alpha:\RR\ra\Aut(\N)$ whose predualised actions on $\N_\star^{\sa}$ can be described as Poisson flows of some hamiltonian vector field.
Our main conclusion from this analysis is that the relationship between Poisson flows and standard liouvilleans should be \textit{localised}: instead of requiring a Poisson flow to globally agree with a family of norm continuous isometries arising from a predetermined W$^*$-dynamical system, we can start from a \textit{quantum Poisson system} (defined as a set $\M(\N)\subseteq\N^+_\star$ equipped with \textit{some} Banach Lie--Poisson manifold structure, not necessarily determined by the coadjoint action of the Lie algebra of self-adjoint elements of $\N$), and determine a fiberwise family of local W$^*$-dynamical systems generated by a $1$-form corresponding to a hamiltonian vector field on the state space. This way we consider the fiberwise family of local standard liouvilleans as a Hilbert space/algebraic counterpart of the smooth manifold/geometric hamiltonian vector field of the Poisson flow.
The main technique used in Section \ref{from.local.gauge.to.lagrangeans.section} is: 1) to represent a (possibly, nonlinear) local Poisson flow \textit{on the state space manifold} in each fibre of the GNS Hilbert space bundle by constructing a \textit{local} standard liouvillean, generating unitary evolution uniquely corresponding to the hamiltonian vector field of this flow, and then: 2) to perturb it using objects that represent additional geometric structures on the state space. The resulting structure is shown to determine a nonlinear instrument on $\N^+_\star$ (in the sense of \cite{Davies:Lewis:1970}), which we call a \textit{local liouvillean instrument}. It describes the temporal evolution of quantum states determined by the postulated `internal' dynamics (a W$^*$-dynamical system, a Poisson flow, or a globally defined vector field) perturbed by the \textit{geometric} structures on the quantum model. In other words, the local liouvillean instrument encodes the effective dynamics, that takes into account a nontrivial geometry of the space of quantum states. In addition, we discuss the possible expressions for time dependent $n$-point correlation functions that can be constructed using the above structures. Both local liouvillean instruments and correlation functions are understood as tools as quantification of the effective dynamics.
Noticing that both the GNS bundle and the tangent bundle of the manifold of quantum states can allow in principle for introduction of a nontrivial action of some Lie group $G$ on fibres, we propose to consider a specific relationship between local gauge (principal $G$-bundle connection) structure and the GNS bundle. We begin with incorporation of the (fiberwise representation of the) action of the nontrivial gauge connection $\mathbf{A}$ (one-form valued in the Lie algebra $\glie$ of $G$) into the perturbation of the local liouvillean, discussing also the possible relationship between affine connection on the tangent bundle of the quantum manifold and propagation of quantum particles (in Wigner sense) in the GNS fibre bundle.
Apart from the gauge connection, we study also the class of objects that, from the perspective of the BLP structure, could be considered as nonlinear quantum fields. These are introduced as the additional source/sink terms, representing the Lie algebra valued differential one-forms on the base quantum manifold. (The idea of using source-based approach is inspired by Schwinger's \cite{Schwinger:1969,Schwinger:1970} and the Mitchell--Jaynes \cite{Mitchell:1967,Jaynes:1993} approaches.)
As a result, we construct a setting that allows to define various nonlinear quantum models equipped with smooth geometric structures that can be represented directly in terms of the families of operators acting locally on the fibres of the GNS bundle of Hilbert spaces over the model. It seems that this framework covers quite well some of the components of the lagrangean framework (under nonorthodox assumption that space-time/phase space geometry is emergent from the geometry of quantum state spaces). The investigated correspondences between geometric and algebraic structures can be briefly summarised as:
\begin{center}
\begin{tabular}{c|c}
\textit{$C^\infty$-geometric} & \textit{GNS-bundle-algebraic}\\\hline
principal $G$-bundle sections & gauge propagators\\
one-forms & local quantum field source operators\\
Lie algebra valued one-forms & local gauge quantum fields\\
global charges & global source strengths\\\hline
\end{tabular}
\end{center}
Under some additional conditions we can establish also some relationships between structures of tangent and the GNS bundles:
\begin{center}
\begin{tabular}{c|c|c}
\textit{$C^\infty$-geometric} & \textit{GNS-bundle-algebraic} & \textit{extra condition}\\\hline
hamiltonian vector fields & standard liouvilleans & (PC$_2$)\\
geodesic trajectories & gauge geodesic propagations & (QP$_1$)\\\hline
\end{tabular}
\end{center}
There is also a correspondence between the local liouvillean instruments acting on $\M(\N)$ and local liouvillean operators acting on the fibres of the GNS bundle. These instruments might be nonsmooth. Two main ideas regarding the local quantum dynamics contained in Section \ref{from.local.gauge.to.lagrangeans.section} can be summarised as:
\begin{align}
\mbox{local gauge dynamics}&=\mbox{local Poisson dynamics}+\mathbf{A}\mbox{-propagation},\\
\mbox{local liouvillean dynamics}&=\mbox{local Poisson dynamics}+\mathbf{A}\mbox{-propagation}+\mbox{action of sources}.
\end{align}
An especially interesting possibility for introducing an affine connection $\nabla$ on a tangent bundle $\T\M(\N)$ is a third order Taylor expansion of a quantum relative entropy functional $D$. In such case the gauge geodesic propagation of quantum particles can be carried precisely along the lines of local information flow, defined by a constrained maximisation of a relative entropy, and equivalent to the local $\nabla^D$-geodesic free fall. This particular application shows a virtue of using the Hilbert space bundle combined with the technique of local perturbation of standard liouvilleans: it allows to accomodate different smooth manifold structures on the space of quantum states into a single fiber-wise operator formulation. See \cite{Kostecki:2016:microhauptwerk} for further discussion.
In Section \ref{algebraic.action.operator.section} we provide another example of application of this technique, constructing an algebraic generalisation of Savvidou's action operator. It is specified by perturbation of a standard liouvillean $L_\alpha$ of a weak-$\star$ continuous $*$-automorphism $\alpha:\RR\ra\Aut(\N)$ of a W$^*$-algebra $\N$ by the generator $K_\omega:=-\log\Delta_\omega$ of the Tomita--Takesaki modular automorphism $\sigma^\omega:\RR\ra\Aut(\N)$. This can be tentatively interpreted as incorporation of an action of a $U(1)$-connection on a fibre bundle of Hilbert spaces over a real line of a trajectory of $\alpha_\star$ on $\N^+_\star$. For any W$^*$-algebra $\N$, the Falcone--Takesaki theory \cite{Falcone:Takesaki:2001} functorially associates a `core' von Neumann algebra $\core$. If $\N$ is equipped with a faithful normal algebraic state $\omega$, then there exists a canonical unitary isomorphism $\core\iso\N\rtimes_{\sigma^\omega}\RR$ with the crossed product corresponding to a W$^*$-dynamical system $(\N,\RR,\sigma^\omega)$ formed by a modular $*$-automorphism $\sigma^\omega$ of $\N$. This crossed product is a von Neumann algebra generated by the operators $\pi_{\sigma^\omega}(x)$ and $u_\RR(t)=\ee^{-\ii t\tilde{V}}$ acting on the space $\H_\omega\otimes L_2(\RR,\dd t)$ by means of the equations \eqref{pi.sigma} and \eqref{lambda.tau}. The covariance equation \eqref{modular.histories.covariance}, where a self-adjoint linear operator $K_\omega$ is equal to the Tomita--Takesaki modular hamiltonian, turns the `Liouville' (in Savvidou's sense) action of $\ee^{-\ii t\tilde{V}}$ on $L_2(\RR,\dd t)$ into the action of $\ee^{-\ii tK_\omega}$ on the space $\H_\omega$. Hence, one can say that it `internalises the description of external unitary kinematics'. The perturbed operator $L_\alpha+K_\omega=L_\alpha-\log\Delta_\omega$ provides an algebraic replacement of the quantum histories description of action operator given by equations \eqref{s.h.auto} and \eqref{s.h.auto2}.
\subsubsection{Local information geometry in quantum histories\label{local.infogeometry.histories.intro.section}}
By analysis of the virtues and drawbacks of the above formulation of an algebraic action operator, we come to a conclusion that the proper candidate for a description of the geometric perturbation of a dynamics due to the local change of state in a projective measurement is not $K_\omega$, but a standard unitary transition operator $V_{\phi,\omega}$ (which is not easily incorporable into the local liouvillean framework). It is a parallel transport of the Levi-Civita connection $\nabla^{1/2}$ of the Wigner--Yanase metric $\gbold^{1/2}$, and projections along its ``free fall'' geodesics are equal to the linear projections in a (standard representation) Hilbert space. This observation leads us to revisit the use of a Fubini--Study metric $\gbold^{\mathrm{FS}}$ on the space of coherent states (which coincides, up to a multiplicative scalar factor $4$, with $\gbold^{1/2}$, when the latter is extended to the boundary of the pure states) for the purposes of regulation of the propagator of a quantum dynamics defined by means of functional integration. As a result, we propose a suitable generalisation of the Daubechies--Klauder expression \eqref{DK.Wiener.expression} for regularised continous time path integration, replacing the coherent states over phase space by all states in the given quantum model $\M(\N)\subseteq\N^+_{\star0}$. In what follows, we will discuss the conceptual aspects of the mathematical formulation that we provide.
Motivated by the Anastopoulos--Savvidou analysis of the term $\ee^{\ii\int p\dot{q}}$ in the Daubechies--Klauder formula \eqref{DK.Wiener.expression} as a holonomy of the Berry connection, we replace it by
\begin{equation}
\exp\left(\ii\int\dd t\s{\Omega_{\phi(t)},\ddd_{\nabla^{1/2}}(\phi(t))\Omega_{\phi(t)}}_{\H_{\phi(t)}}\right),
\label{one.form.nabla.onehalf.holonomy}
\end{equation}
where $\H_{\phi(t)}$ is the GNS Hilbert space associated with $\phi(t)\in\M(\N)$, $\Omega_{\phi(t)}$ is its representing vector, while $\ddd_{\nabla^{1/2}}$ is a $\nabla^{1/2}$ connection $1$-form. This is equivalent to a local integral of an infinitesimal entropic projection generated by a quantum Br\`{e}gman distance $D_{1/2}$ on $\M(\N)$. While the necessary mathematical background describing the relationship between entropic projections and geodesic free falls is discussed in Section \ref{distances.NS.geom.section}, let us briefly explain the conceptual perspective behind using it to define the dynamics in quantum theory.
In general, the entropic projections $\PPP^D_{\Q}$ can be used to generate the global temporal evolution of quantum models following the ideas of Jaynes \cite{Jaynes:Scalapino:1963,Mitchell:1967,Jaynes:1985:scattering,Jaynes:1993}, promoted from an absolute to relative entropy by Schl\"{o}gl \cite{Schloegl:1966,Schloegl:1967:produzierte,Schloegl:1967:foundation,Schloegl:1971:stability,Schloegl:1971:fluctuations,Schloegl:1971:near,Schloegl:1971:produced} and Hobson \cite{Hobson:1971,Hobson:Cheng:1973} (see \cite{Grandy:1987,Grandy:2008,Streater:2009:book,Caticha:2012} for the recent account on further developments of these approaches).\footnote{While the definition of $\int\mu p\log\frac{p}{q}$ as well as its conceptualisation as a measure of relative information gain is due to Kullback \cite{Kullback:Leibler:1951,Kullback:1959}, the use of this object for defining information dynamics can be credited to the above authors.} Given a time dependent set of constraints $\Q(s)$, the map
\begin{equation}
\phi_0\mapsto\PPP^D_{\Q(s)}(\phi_0)
\label{entropic.time.evolution}
\end{equation}
selects a unique trajectory of quantum states, if for each $s$ the set $\Q(s)$ is such that it gives a unique solution to the corresponding minimisation problem (in order to recover the typical formulation of dynamical problems, one may additionally require the map $s\mapsto\PPP^D_{\Q(s)}(\phi_0)$ to be continuous, and $\PPP^D_{\Q(0)}(\phi_0)=\phi_0$).
However, this construction is not the same as local re-updating of the state in time $s$ to the state in time $s+\delta s$ by the new data. While in principle it is nothing wrong with it (after all, the classical action principle $\delta S=0$ is an inherently nonlocal construction), it is interesting to see whether a \textit{local} entropic dynamics can be proposed. The equivalence of entropic projections with geodesic projections for the class of \textit{Br\`{e}gman distances} $D_\Psi$ provides such a possibility. In such case, instead of consideration of subsequent stages of a relative entropy driven evolution that is nonlocally determined by an initial state $\phi_0$, one can just follow the $\nabla^{D_\Psi}$-geodesics of the $\nabla^{D_\Psi}$-connection (derived from $D_\Psi$, as a third order Taylor expansion, by means of the Eguchi equations, see Section \ref{distances.NS.geom.section}), maintaining the condition of the $(\nabla^{D_\Psi})^\nsdual$-convexity and $(\nabla^{D_\Psi})^\nsdual$-affinity of the local constraints, as well as their $(\gbold^{D_\Psi},\nabla^{D_\Psi},(\nabla^{D_\Psi})^\nsdual)$-orthogonality with respect to the local $\nabla^{D_\Psi}$-geodesic trajectory. This way the local user's inference, based on smooth reminimisation of an information distance (locally $D_\Psi$-optimal learning process) becomes equivalent with a free fall along $\nabla^{D_\Psi}$-geodesics. One can call it a \textit{local equivalence principle} of an ``information gravity''. The restriction of an arbitrary $D_\Psi$ and $\nabla^{D_\Psi}$ to $D_{1/2}$ and $\nabla^{1/2}$, as expressed by \eqref{one.form.nabla.onehalf.holonomy}, is caused by two reasons: requirement of showing explicit backwards compatibility with the Daubechies--Klauder path integrals, and also the structural restrictions of the GNS bundle.\footnote{Using Hasegawa \cite{Hasegawa:1993} representation of a tangent space in terms of functions of density operator, we could generalise the use of the GNS bundle at least to the case of $\nabla^\gamma$ connections derived from $D_\gamma$. However, this would be restricted to the finite dimensional case. Moreover, we are interested here more in the search of an appropriate analytic setting for the general theory than in the explicit calculations of special finite dimensional cases.} In order to use other $\nabla^{D_\Psi}$ connections, we would have to systematically apply noncommutative Orlicz spaces, which is beyond the scope of this paper.
In addition, we replace an affine function $h(z(t))$ in the Daubechies--Klauder formula \eqref{Klauder.path.integral}, which is corresponding to a Killing hamiltonian vector field and is generated by a coherent state expectation value of a self-adjoint hamiltonian operator, by any smooth function on $\M(\N)$, understood as a hamiltonian function on a BLP manifold. This leads us to ask how one can relate local entropic and local hamiltonian dynamics in the histories context.
Combined with the causal inference Ansatz \eqref{global.hamiltonian.entropic.map}, the principle \eqref{entropic.time.evolution} leads to a \textit{global} evolution
\begin{equation}
\phi(t,s)=\PPP^D_{\Q(s)}\circ w^{\B,h}_t(\phi_0).
\label{two.times.global.causal.inference.instrument}
\end{equation}
While one can chose $t\in[r_0,r_1]$ and $s\in[r_1,r_2]$, $r_0,r_1,r_2\in\RR$, there is no obligation to do so. In general, $w^{\B,h}_t$ represents a causal evolution governed by the principle of a \textit{local conservation of absolute energy} (undestood as an element of the local space of quantitative effects, such as the self-adjoint observables in quantum mechanics), while $\PPP^D_{\Q(s)}$ represents an inferential evolution governed by the principle of a \textit{global growth of relative entropy} (understood as a function on the global space of states). It was first observed by K\k{e}pi\'{n}ski \cite{Kepinski:1972:Rytm,Kepinski:1972:Schizofrenia} that each of those dynamical processes carries its own notion of time. Our work grew out from consideration of this duality, and \textit{a priori} independence of two associated notions of time, as the fundamental principle of physical dynamics. In the context of the present paper, as we discuss it below, we postulate that the `energetic' time of causality and the `entropic' time of inference are equal, but only \textit{infinitesimally}.\footnote{Due to incoherence between the standard use of the word `local' in physics and in mathematics, it is hard to propose any universally optimal terminology for distinguishing between different regimes. In this paper we use the terms: \textit{global} to refer to objects acting on all space $\M$; \textit{local} and (equivalently) \textit{infinitesimal} to refer to objects acting at $\phi\in\M$; \textit{nonlocal} to refer to objects acting in some neighbourhood of $\phi\in\M$ (maybe \textit{quasi-local} would be a better term). Within our setting, the local regime corresponds to a single user system, defined by the states-and-effects kinematics equipped with the causal-inferential dynamics, global regime corresponds to a multi-user framework, while nonlocality corresponds to the issues of construction of effective multi-user kinematics and dynamics, based on the choice of specific criteria of synchronisation between individual user's systems, at the expense of some individual properties being no longer maintained at the effective level. See Sections \ref{curvature.intro.section} and \ref{locally.quantum.information.relativity.section} for more discussion.} This forms the first principle of local information dynamics. More precisely, we consider the local causal dynamics, governed by $\ddd h_\B$,\footnote{The notation $\ddd h_{\{\cdot,\cdot\}_\B}$ would be completely precise, and symmetric with the notation $\ddd_{\nabla^{D_\Psi}}$, but also quite expensive visually.} and the local inferential dynamics, governed by $\ddd_{\nabla^{D_{\Psi}}}$, as \textit{two} independent fundamental dynamical processes that should be treated on the equal footing as components generating jointly an infinitesimal temporal evolution in a \textit{single} time. In other words, we postulate that the complete description of the local dynamics should be governed by a differential $1$-form
\begin{equation}
{\cal F}_{D_\Psi,h_\B}:=\ddd h_\B+\ddd_{\nabla^{D_\Psi}}.
\label{local.causal.inferential.form}
\end{equation}
However, the existing constructions of a quantum information manifold use different tangent-cotangent space structure than the quantum Poisson spaces, so, as a result, the addition operation in \eqref{local.causal.inferential.form} is precisely as meaningful, as is adding the element of the noncommutative Orlicz space to the element of a Banach Lie subalgebra $\B$ of $\N^\sa$ (or of $\N$, if one wants to use some nonstandard possibilities). The best solution to this problem would be to represent the action of $\B$ on the cotangent space, given by the noncommutative Orlicz space $L_\Upsilon(\N)$. Yet, even in the simple case, when $\B$ is a Lie algebra of a group of unitaries $\N^\uni$, a corresponding functional analytic study requires to prove the analogue of the Haagerup theorem \cite{Haagerup:1975:standard:form} for the noncommutative Orlicz spaces, resulting in the $L_\Upsilon$-liouvillean (a generalisation of the Jak\v{s}i\'{c}--Pillet $L_{1/\gamma}$-liouvilleans \cite{JOPP:2012}). This is beyond the scope of this paper. As a result, the natural functional analytic representation of the above differential geometric framework is not yet available, and we need to carry our investigations by means of a bit less canonical tools, as reviewed in Section \ref{qig.foundations.intro.section}. The analysis of local perturbations of standard (i.e., $L_{1/2}$-) liouvilleans associated with the GNS Hilbert bundle, carried in Section \ref{from.local.gauge.to.lagrangeans.section} and briefly reviewed in Section \ref{locally.perturbed.liouvilleans.intro.section}, is intended to be a warm-up investigation preceding the (currently unavailable) theory of local perturbations of $L_\Upsilon$-liouvilleans, associated with the tangent and cotangent spaces of quantum information manifolds based on quantum Br\`{e}gman distances. In this sense, the current paper can be considered as an investigation of the mathematical framework for the local Ansatz \eqref{local.causal.inferential.form}, as well as a family of related dynamical problems.\footnote{In future work, we will also consider a replacement of $\ddd h_\B$ by the connection $1$-form of the quantum relative free energy, as generated by the Fenchel--Legendre conjugate of a relative entropy.}
\subsection{Effective local quantum dynamics\label{effective.local.quantum.dynamics.section}}
Comparing the path integral propagator \eqref{RPK.propagator} with the local liouvillean propagator \eqref{liouv.prop}, we can see that both consist essentially of the subtraction of the local free fall along $\nabla^{1/2}$ geodesics from the local hamiltonian flow. This free fall corresponds to the local $D_{1/2}$-projection, which can be in turn interpreted as a continuous projective measurement. This leads us to revisit the problem of analytical implementation of the postulate that the \textit{local} geometric dynamics should be generated by the $1$-form \eqref{two.times.global.causal.inference.instrument}. In particular, applying the infinitesimal approximation $\phi=\omega+\delta\omega$ in the local liouvillean propagator \eqref{liouv.prop} results in the generator $\pi_\omega(\DF_\omega h)-\log(J_{\omega+\delta\omega}J_{\omega+\delta\omega,\omega})$ , which can be considered as an implementation of \eqref{local.causal.inferential.form}.
As a warm-up, consider the case $\N\iso\BH$, $\B\iso\BH^\sa$, $\M(\N)=\schatten_1(\H)^+_{1}$. Combining the insights of B\'{o}na \cite{Bona:1993,Bona:2000} and Grandy \cite{Grandy:2004:1,Grandy:2008}, one can propose the following generalisation of the von Neumann equation,
\begin{equation}
\ii\frac{\dd}{\dd t}\rho(t)=
[\ddd h(\rho(t)),\rho(t)]-\frac{\partial}{\partial t}\rho(t),
\label{dissipative.generalisation.of.vonNeumann.eq}
\end{equation}
where
\begin{equation}
\frac{\partial}{\partial t}\rho(t)=\sum_{i\in I}P_i\frac{\partial}{\partial t}p_i(t)
\end{equation}
is determined by the nonhamiltonian change of probabilities $\{p_1(t),\ldots,p_n(t),\ldots\}$ that determine $\rho(t)$ by means of $\rho(t)=\sum_{i\in I}P_ip_i(t)$, given $\sum_{i\in I}P_i=\II$, $P_iP_j=\dirac_{ij}P_i=\dirac_{ij}P_j$, and $P_i\in\Proj(\BH)$ $\forall i\in I$. Grandy, following Jaynes \cite{Jaynes:Scalapino:1963,Jaynes:1979:where:do:we:stand,Jaynes:1986}, proposes to use maximum absolute relative entropy to construct the evolution $p(t)$. As compared to Jaynes' approach, we propose to replace the use of absolute entropy on probability densities by the use of relative quantum entropies on quantum states. The resulting geometrisation of a contribution $\frac{\partial}{\partial t}\rho(t)$ by means of a connection $1$-form $\ddd_{\nabla^{1/2}}(\phi(t))$ can be provided by a choice of a frame (ordered list of vector fields) $\xi(t):=(\xi_1(t),\ldots,\xi_n(t),\ldots)\in\H_{\rho(t)}$, such that $\ab{\xi_i(t)}^2=p_i(t)$ $\forall i\in I$, and evaluation
\begin{equation}
(\ddd_{\nabla^{1/2}})^j_{\;i}(\rho(t))=\sum_k({\Gamma^{\nabla^{1/2}}})^j_{\;\;ki}(\xi(t))P^k.
\label{d.gamma.two.sum.christo}
\end{equation}
Contracting the missing indices with $\xi^i(t)$, we derive the explicit representation of \eqref{one.form.nabla.onehalf.holonomy} as
\begin{equation}
\exp\left(\ii\int\dd t\,\xi_j(t)\sum_k({\Gamma^{\nabla^{1/2}}})^j_{\;\;ki}(\xi(t))P^k\xi^i(t)\right).
\end{equation}
Hence, \textit{if} one implements the principle \eqref{two.times.global.causal.inference.instrument} as a formal equation
\begin{equation}
\dot{\rho}(t)=-\ii\left[{\cal F}_{h,\nabla^{1/2}}(\rho(t)),\rho(t)\right],
\label{formal.effective.form.perturbation}
\end{equation}
\textit{then} the latter can be represented in the above situation as
\begin{equation}
\ii\frac{\dd}{\dd t}\rho=[\ddd h(\rho(t)),\rho(t)]-\sum_{k,i}({\Gamma^{\nabla^{1/2}}})^j_{\;\;ki}(\xi(t))P^k(P^i-J_{\rho(t)}P^iJ_{\rho(t)}),
\label{RPK.local.equation}
\end{equation}
where the modular conjugations $J_{\rho(t)}$ arise from a commutator of $\ddd_{\nabla^{1/2}}$ with $\rho(t)$.
Alternatively, taking into account our earlier observation that the standard transition unitary $V_{\omega,\phi}=J_{\phi}J_{\phi,\omega}$ is \textit{exactly} a parallel transport of $\nabla^{1/2}$, we can begin from the $\nabla^{1/2}$-parallel transport equation of a vector $v^a$ along the trajectory $\rho(t)$,
\begin{equation}
\frac{\dd}{\dd t}v^a(t)=-\sum_{b,c}({\Gamma^{\nabla^{1/2}}})^a_{\;\;bc}(\rho(t))v^b(t)\left(\frac{\dd}{\dd t}\rho(t)\right)^c.
\end{equation}
Substituting $v=\dot{\rho}(t)$, and integrating out, we get
\begin{equation}
\ii\frac{\dd}{\dd t}\rho(t)=
-\int_{-\infty}^t\dd t \sum_{b,c}({\Gamma^{\nabla^{1/2}}})^a_{\;\;bc}(\rho(t))\left(\frac{\dd}{\dd t}\rho(t)\right)^b\left(\frac{\dd}{\dd t}\rho(t)\right)^c.
\label{geodesic.free.fall}
\end{equation}
This equation, describes the equation of motion of the free fall along the $\nabla^{1/2}$-geodesic trajectory $\rho(t)$, when represented in the GNS Hilbert space bundle by means of $\H_{\rho(t)}\iso\schatten_2(\H)$.
An infinitesimal transformation $\rho\mapsto\rho+\der\rho$ can be decomposed as \cite{Hasegawa:1993}
\begin{equation}
\der\rho:=\widetilde{\der}\rho+[\rho,W]=\sum_{i=1}^n\left(\frac{\partial\rho(\theta)}{\partial\theta^i}+[\rho,W_i]\right)\ddd\theta^i,
\label{orthogonal.decomp.of.rho}
\end{equation}
where $\widetilde{\der}\rho=\sum_{i=1}^n\frac{\partial\rho(\theta)}{\partial\theta^i}\ddd\theta^i$ is defined by $[\widetilde{\der}\rho,\rho]=0$, and $W=\sum_{i=1}^nW_i\ddd\theta^i$ is an antiself-adjoint operator (hence, $k_i^*=k_i:=\ii W_i$). The mappings $\der$, $\widetilde{\der}$ and $[\,\cdot\,,W]$ are derivations on $\BH$. This determines a decomposition of tangent space at $\rho$ into the direct product of the corresponding subspaces. An explicit representation of the tangent space in terms of $\schatten_2(\H)$ space by means of finite dimensional coordinate parametrisation $\RR^n\supsetneq\Theta\ni\theta\mapsto\rho(\theta)\in\schatten_1(\H)^+$ reads \cite{Hasegawa:1993}
\begin{equation}
\T_\rho\ell_{1/2}(u)=\sum_{i=1}^nu^i\left(\sqrt{\rho}\frac{\partial\rho}{\partial\theta^i}+2[\sqrt{\rho},W_i]\right).
\end{equation}
As a result of these considerations, if we interpret the principle \eqref{two.times.global.causal.inference.instrument} as a statement that the effective local dynamics is generated by the sum of vectors $\dot{\rho}(t)$ arising independently from the hamiltonian flow and the geodesic free fall \eqref{geodesic.free.fall}, then we should use the equation
\begin{align}
\ii\frac{\dd}{\dd t}\rho(t)&=[\ddd h(\rho(t)),\rho(t)]\label{RPK.effective}\\
&-\int_{-\infty}^t\dd t \sum_{b,c}({\Gamma^{\nabla^{1/2}}})^a_{\;\;bc}(\rho(t))\left(\sum_{i}u^i\left(\sqrt{\rho}\frac{\partial\rho}{\partial\theta^i}+2\left[\sqrt{\rho},k_i\right]\right)\right)^b\left(\sum_{i}u^i\left(\sqrt{\rho}\frac{\partial\rho}{\partial\theta^i}+2\left[\sqrt{\rho},k_i\right]\right)\right)^c.\nonumber
\end{align}
This equation describes the \textit{effective local dynamics}, including causality and inference effects on the equal footing (thus, \textit{paralelly processing} them). Comparing the nonhamiltonian parts of the equations \eqref{RPK.local.equation} and \eqref{RPK.effective}, we see that the equation \eqref{RPK.local.equation} can be at best some sort of approximation of \eqref{RPK.effective}. We interpret it as an indication of the weakness of the implementation \eqref{formal.effective.form.perturbation}, as compared with \eqref{RPK.effective}.
The interpretation of \eqref{local.causal.inferential.form} an infinitesimal analogue (but not a generator) of the ``entropic inference following causal Poisson evolution'' global W$^*$-geometric dynamics \eqref{global.hamiltonian.entropic.map}, and the fact that the latter allows to reconstruct CPTP maps as a special case \cite{Kostecki:MunkNielsen:2016,Kostecki:2016:microhauptwerk}, suggests to intepret \eqref{RPK.effective} as a nonlinear geometric analogue of the Lindblad--Gorini--Kossakowski--Sudarshan equation \cite{Lindblad:1976,Gorini:Kossakowski:Sudarshan:1976}.
\subsection{Curvature measures desynchronisation in the multi-user inference\label{curvature.intro.section}}
In this Section we will consider the problem of the effective \textit{nonlocal} quantum information dynamics. As opposed to effective local dynamics, which provides the infinitesimal description of causality and inference from the perspective of a single user (thus, allowing for an immediate subjective bayesian interpretation), nonlocality is intended to describe the multi-user (intersubjective) conventions of relating causal and inferential dynamics of individual users. In our opinion the geometric space is an emergent property of the specific intersubjective conventions of information (causal and inferential) dynamics, which is, in turn, relative to a specific choice of (class of) users. In this text we are focused on an analysis of a specific example, rooted in quantum information geometry. More foundational discussion will be postponed to another paper \cite{Kostecki:2016:microhauptwerk}.
In principle, given two or more different users, each with his/her own method of providing inferences and evaluating causal evolution, nothing can be said about how their information dynamics is related. If only the inferential aspect of information dynamics is taken into account, then such situation can be understood as incommensurability of inferences provided by different subjective bayesians with their arbitrary choices of respective priors and of methods of updating. In order to relate these different instances of local information dynamics, one needs to introduce some method of translation between the respective evolutions, as well as their initial assumptions. Each such method represents a specific intersubjective convention, which allows to translate between individual instances of information dynamics at the expense of constraining its possible forms to such that are subjectible to a given convention. In particular, for inferential part of the information dynamics, there should be a way of identification of a given state of information as `the same' state for all users under consideration. Note that there is no need for a such identification being made globally for all \textit{possible} users: it is sufficient if one can do it for different sets of users that are under the scope of interest.
The general setting for these considerations can be defined as follows. Given an abstract set $\M$ of users, with a single user represented as a point $\phi\in\M$ equipped with a vector space of \textit{local states} $V(\phi)$ and a Banach dual vector space $V^\dual(\phi)$ of \textit{local effects} (one can think of them in terms somewhat similar to \cite{Ludwig:1985:1987}, but the duality does not have to be carried by Banach space structure, but e.g. by convenient vector space structure).\footnote{By the reasons discussed above and below, we postulate that \textit{if} $\M$ is given by $\M(\N)$, \textit{then} $V(\phi)$ should be specified as $L_{\Upsilon(\phi)}(\N)$. However, in order to distinguish the conceptual and the representational aspects of our approach, we state it in more general terms.} In order to model causality and inference, each user can chose locally his/her own `causal' Banach Lie algebra $\B$ acting on $V^\dual(\phi)$, as well as its own `inferential' \textit{Br\`{e}gman functional} $\tilde{D}_\Psi$ on $V(\phi)$, the latter determined by the duality between $V(\phi)$ and $V^\dual(\phi)$ and the choice of a function $\Psi:V(\phi)\ra\RR$. Given a set $U(\phi)\subseteq\M$ of users, such that $\phi\in U(\phi)$, a choice of a function $\ell_\phi:U(\phi)\ra V(\phi)$ allows to use the Br\`{e}gman \textit{functional} $\tilde{D}_\Psi:V(\phi)\times V(\phi)\ra[0,\infty]$ in order to construct the Br\`{e}gman \textit{distance} $D_\Psi:U(\phi)\times U(\phi)\ra[0,\infty]$ by means of
\begin{equation}
D_\Psi(\omega,\psi):=\tilde{D}_\Psi(\ell_\phi(\omega),\ell_\phi(\psi))
\end{equation}
Thus, the choice of the function $\ell_\phi$ defines how the local user interprets `inferentially' the subset $U(\phi)$ of the set $\M$, while the choice of a hamiltonian function $h(\phi)$ defines how he/she interprets it `causally'. The forms $\ddd h_\B(\phi)$ and $\ddd_{\nabla^{D_\Psi}}(\phi)$, constructed as elements of $V^\dual(\phi)$, are encoding the corresponding infinitesimal dynamics.\footnote{On the level of implementation, it is sufficient to model $V(\phi)$ and $V^\dual(\phi)$ as convenient vector spaces in order to guarantee that the infinitesimal calculus is well defined.} The translation between different users in the set $\Q\subseteq\M$ requires, within this model, to specify relationship between $\ell_\phi$ and $\ell_\psi$, as well as $h(\phi)$ and $h(\psi)$, for all elements $\phi,\psi\in\Q$.
The \textit{local} state-effect kinematics is completely described using the pair $(V(\phi),V^\dual(\phi))$ at a given point $\phi\in\M$, while the resulting \textit{local} causal-inferential dynamics is generated by \eqref{local.causal.inferential.form}. This leads to a question to what extent, and at what expense, this dynamics can be extended to a larger \textit{nonlocal} area of $\M$, for example allowing to interconnect the dual pairs of vector spaces of different users by means of the (not necessarily global) sheaf of tangent and cotangent spaces. On the conceptual level, this corresponds to the question how the local (individual) state-and-effect dual pairs of different users $\phi\in\M$ can be mutually related, how their respective causal and inferential local dynamics can be synchronised, and what is the price to pay for it? For example, even if all users $(\phi,V(\phi),V^\dual(\phi))$ agree to use the same generating objects for their respective local system of causality (e.g., a Banach Lie algebra $\B$) and local system of inference (e.g., a discrimination function $\Psi$), their individual choices of functions $h_\B(\phi)$ and $\ell_\phi$ may be not extendible to a set $\M$, resulting in sheaves of different effective local dynamics \eqref{local.causal.inferential.form}. The problem of conditions for integrability of \eqref{local.causal.inferential.form} is thus directly related to the issue of nonlocal (multi-user) synchronisation of local systems of causality and inference about local states and effects.
In general, the convention defining emergent multi-user inference can be arbitrarily different from the convention defining the emergent multi-user causality, and they both can differ from any of the local instances of inference and causality that are amalgamated into the emergent structure. As a result, the emergence of a nonlocal spatio-temporal causal-inferential dynamics is be provided at the expense of its departure from the local causal-inferential dynamics. In the context of the structures analysed in this paper, it is represented by the appearance of the system of local entropic priors associated with a specific integral line of a vector field on a model $\M$. (From the closely related point of view, one can notice that the postulate of Section \ref{local.infogeometry.histories.intro.section} identifies the local inferential time with local causal time for each individual user \textit{separately}, but it does not say anything about mutual relationships of those time parameters for different users. In principle, one could also study theories for which the local causal and inferential time of each user would \textit{not} be identified, yet the multi-user synchronisation of those two temporal structures would be considered. We find this perspective very attractive, but it is beyond the scope of the current paper.)
On the technical level, we observe that the Fubini--Study riemannian metric, used in the regularising term in the Daubechies--Klauder and the Anastopoulos--Savvidou approaches, can be replaced by the second order Taylor expansion $\gbold^{1/2}$ of a quantum relative entropy $D_{1/2}$. This leads us to postulate to use a quantum entropic prior \textit{localised} to a neighbourhood of a given state as a general geometric form of the regulariser (see Section \ref{entropic.priors.section} for a discussion of the notion of an entropic prior in the commutative case). Integration of a local entropic prior along a given trajectory on a state manifold constructs a regularised weight for this trajectory.
On the conceptual level, we interpret the appearance of local entropic prior as a measure of (impossibility of) synchronisation of the causal-inferential dynamics of the subsequent local users at a given spatio-temporal trajectory. More specifically, the choice of a \textit{single} nonlocal (multi-agent) time trajectory sets up the nonlocal vector field, corresponding to a specific system of synchronisation (=~nonlocal/noninertial observation frame) for the local forms ${\cal F}_{D_\Psi,h_\B}$. This choice is arbitrary, but with each such choice, the passage from point to point on the corresponding nonlocal trajectory of a \textit{single spatialised time} adds an additional term to an effective dynamics \textit{along this trajectory}. Only some specific conventions of the multi-user inference avoid the path-dependence of the synchronisation of the inference: in the example that we study here they are given by such models $(\M(\N),D_\Psi)$ for which the scalar curvature $\kappa(\nabla^{\gbold^\Psi})$ is constant over $\M(\N)$.
One can wonder why this phenomenon of ``breaking of symmetry'' between causality and inference results in an entropic (hence inferential), as opposed to a hamiltonian (hence causal) contribution. If all users along the trajectory (as well as in the neighbourhood of this trajectory, in order to have a situation that is more canonical than that of a $1$-dimensional manifold) agree to have the same choice of $\B$ and $\Psi$, then in principle they should share the same nontrivial inferential and causal geometry, so there is yet no reason for ``breaking of the symmetry''. However, the Carath\'{e}odory--Jacobi--Lie theorem (which generalises the Darboux theorem) implies that for any symplectic manifold with a hamiltonian function $h$, the function $h$ is a conserved quantity in an open neighbourhood for any $\phi$ for which $\ddd h(\phi)\neq0$. An analogous statement is not true in riemannian geometry, as measured by the curvature. The generalisation from symplectic to Poisson geometry and from riemannian to Norden--Sen geometry does not change qualitatively this difference. Thus, the passage from infinitesimal/local (jets) to neighbourhood/nonlocal (germs) is trivial for causality, but not for inference. So, while it is possible to nonlocally synchronise the local causal structure along a trajectory of users (whenever it is modelled by the Poisson geometry), it is impossible to do this with the local inferential structure (whenever it is modelled by the Norden--Sen geometry). It is a very interesting phenomenon, which may be interpreted by saying that while the global synchronisation of local systems of causality is a trivial consequence of choice of the same system of locally conserved quantities for each user, the global synchronisation of local systems of inference is impossible, with the system of local entropic priors measuring the scale of this impossibility for each of the local users along the trajectory. The local entropic priors could be in principle replaced by any other local priors. Each of such choices represents a different \textit{initial system of assumptions} that are used by local users for the purposes to measure the desynchronisation between their local/jet inferences, as provided by means of $\tilde{D}_\Psi$ on $V(\phi)$, and their nonlocal/germ inferences, as provided by means of $D_\Psi$ on $\M$. The local priors are just measures on the local neighbourhoods in the model $\M$, and they have to be \textit{postulated independently} from the choice of ${\cal F}_{D_\Psi,h_\B}$.\footnote{While a detailed discussion of this phenomenon is beyond the scope of the current paper, we want to note that a somewhat similar situation (roughly speaking, a necessity of using additional geometric structures on the pre-state spaces in order to select a specific sheaf of states) was discovered by Bostelmann \cite{Bostelmann:2000} in the Haag--Ojima approach \cite{Haag:Ojima:1996} to construction of germs of algebraic states corresponding to a choice of a specific predictive dynamical theory in the algebraic approach.}
Thus, while the equation ${\cal F}_{D_\Psi,h_\B}=0$ (whenever valid) can be understood as a causal-inferential analogue of Einstein's version of Newton's first law of motion (\textit{for a single user}, from his/her own perspective), the local priors are somewhat similar to the second law of thermodynamics or to the universality of the gravitation: the globalisation of inferences is provided at the expense of inevitability of making those inferences dependent on additional arbitrary assumptions, that are in principle different for each user, and \textit{are nonobservable in the infinitesimal causal-inferential reference frame}. The shadow of dependence of the multi-userr inferences, as well as of the effective \textit{nonlocal} causal-inferential dynamics, on arbitrary additional assumptions (\textit{attributed} to other users by the user residing at the end point of the integrated trajectory) is a price paid for a requirement of global spatialisation of the inferential dynamics. The lack of the similar effect in the case of causal dynamics is in essence a result of a local linearity of the Banach Lie--Poisson spaces. If other mathematical structure would be chosen to model causality, the similar dialectics may occur. For example, one could model causal dynamics by means of extremum of quantum relative free energies, defined as the Legendre--Fenchel conjugates of quantum information distances. The nonlocal trivialisation of the causal structure by means of the CJL theorem would not be applicable in such case. As a result, the dual, local relative free energy priors, should be also included, as an additional regularising term, into integration giving rise to an effective dynamics. In general, we propose to consider the priors on the model $\M$ as an information theoretic analogue of the notion of mass, so the local prior at the neighbourhood $U(\phi)$ of $\phi\in\M$ can be interpreted as an information theoretic analogue of the mass of the user $\phi$ distributed over the set $U(\phi)$ of users. The special case in which the integration against the local relative free energy priors would cancel the contribution arising from the local entropic priors can be then considered to be an analogue of Einstein's principle of equivalence of gravitational (inferential) and inertial (causal) mass.
As observed in \cite{Maraner:1992,Alicki:Klauder:Lewandowski:1993,Watson:Klauder:2002}, if the riemannian geometries of the phase space used for the Wiener measure regularisation have nonconstant scalar curvature, then the weighting of the phase space paths is nonuniform, corresponding to the phase space point dependency of the zero-point energy. On the other hand, in a completely different context, Jaynes has argued \cite{Jaynes:1957,Jaynes:1978:electrotoday,Jaynes:1990} that the zero-point energy should be interpreted not as an ontic feature of the system, but as observer's measure of uncertainty regarding his/her own prediction of the value of energy, as based on his/her own \textit{prior information}. Quite independently from these considerations, Rodr\'{\i}guez has developed the theory of entropic priors \cite{Rodriguez:1991} interpreting them as the \cytat{statistical representation of the vacuum of information in a given hypothesis space} \cite{Rodriguez:1999}. Our formulation recombines the above insights, integrating them into a single statement: zero-point energy's point-dependence is a manifestation of a dependence of nonlocal integrability of local inferential structure on the local geometry of user's prior knowledge (or, equivalently, prior ignorance).\footnote{This gives also some justice to an otherwise quite cryptic remark of Jaynes on the book \cite{Popov:1987} on functional integration: \cytat{A useful start on understanding of these phenomena, but still lacking any coherent theoretical basis -- which we think is supplied only by the principle of maximum entropy as a method of reasoning.} \cite{Jaynes:2003}.}
However, while our proposal is well founded on the geometric side, and has also an interesting information theoretic interpretation, it has to be regarded as heuristic from the perspective of stochastic process based foundations for path integration. More specifically, the Daubechies--Klauder formulation relies on the interpretation of $\int\dd t p\dot{q}=\int p \dd q$ as the Stratonovich integral (so $\dd q$ is considered as a Stratonovich differential $\dd_{\mathrm{S}}q$), and on the interpretation of (roughly) $\int\mathcal{D}p\mathcal{D}q\ee^{-\frac{1}{2\upsilon}\int\dd t(\dot{p}^2+\dot{q}^2)}$ as the pinned Wiener measure (see \cite{Bodmann:2003} for a systematic mathematical treatment of these objects in terms of the Berezin--Toeplitz operators). The generalisation to a wide class of connections $\nabla^{D_\Psi}$ and local entropic priors (even if kept at the second order riemannian level of $\gbold^D$) asks for a systematic development of a technique of stochastic integration of random walks $\mathbf{X}$ on $\RR^n$ associated with the Brownian motions on smooth manifolds $\M$, $\dim\M=n$, that could systematically address the functional integration of the above geometric structures beyond the level of heuristic treatment that is standard for physicists. In particular, following \cite{Ikeda:Watanabe:1981,Kendall:1987}, let us consider the system of stochastic differential equations
\begin{align}
\dd_\mathrm{S}\mathbf{X}&=\mathbf{e}\,\dd_\mathrm{S}\mathbf{B},\\
\dd_\mathrm{S}\mathbf{e}&=H_\mathbf{e}\dd_\mathrm{S}\mathbf{X},
\end{align}
where $\dd_\mathrm{S}$ are Stratonovich differentials, $\mathbf{B}$ is an euclidean brownian motion on $\RR^n$, frame $\mathbf{e}$ is a map from $\RR^n$ to the tangent space of $\M$, and $H$ is a horizontal lift of a tangent space at $\mathbf{X}$ to the tangent space at $\mathbf{e}$, dependent on the choice of an affine connection $\nabla$ on $\M$. If this connection is not Levi-Civita, then the process $\mathbf{X}$ will be not markovian. Furthermore, a suitable riemannian metric reproducing the Laplace--Beltrami operator can be uniquely constructed by an appropriate choice of an elliptic diffusion. Hence, the exact mathematical foundation for our generalisation of the Daubechies--Klauder formula is possible at least for the second order Taylor expansion of the entropic prior and for different connections $\nabla^{D_\Psi}$. While we consider the task of a systematic treatment of this topic to be of high importance, it will be left beyond the scope of this paper.
The issue of renormalisation cannot be omitted in any foundational discussion of local quantum dynamics. Section \ref{information.theoretic.renormalisation} is dedicated to the study how the tools of quantum information geometry can be used in order to deal with the tasks of renormalisation. In Sections \ref{section.JM.source.theory} and \ref{Favretti.section} we briefly review the Jaynes--Mitchell source theory \cite{Mitchell:1967,Jaynes:1985:scattering,Jaynes:1993,Grandy:1987,Grandy:2008} and its geometric generalisation by Favretti \cite{Favretti:2007}, respectively. This approach provides a geometric implementation of the idea of renormalisation of dynamics by reduction of dimensionality of the model by fixing the control parameter, which is specified as a constraint on the space of information states (as opposed to the space of functions or operators). We observe that the Jaynes--Mitchell--Favretti approach is canonically related to the use of Br\`{e}gman distances, and that it can be used to provide a \textit{local} description of entropic information dynamics and multiparameter nonlinear quantum control problems on an arbitrary quantum manifold. The key insight of this approach can be summarised as: renormalisation of the action of control parameters (sources) leads to departure of the geometry of a model $\M$ from a dually flat one. Hence, the appearance of nonzero curvature is an indicator of a nontrivial constraints for information dynamics. Moreover, the nonconstancy of this curvature indicated local dependence of these constraints. From the perspective of our approach to the Daubechies--Klauder path integrals, we postulate that the renormalised description of dynamics should use renormalised riemannian metric $\tilde{\gbold}$ instead of $\gbold^{D}$ in the regulariser. This corresponds to replacement of a ``vacuum of information'' by the ``vacuum of information storing the shadow of the knowledge about the sources that were renormalised out''.
We also introduce another type of information geometric renormalisation of inferential dynamics of quantum states, which describes situations where none of specific control (covariate) parameter is fixed, but the quantum model is subjected to the action of completely positive maps. This procedure is based on the use of $D_\fff$ distances as well as associated \textit{contraction coefficients}, introduced by Ruskai et al \cite{CIRRSZ:1993,Choi:Ruskai:Seneta:1993,Ruskai:1994,Lesniewski:Ruskai:1999}. For an alternative (and essentially more developed) approach to renormalisation based on $D_\fff$ on $\N_\star^+$, see \cite{Beny:Osborne:2012,Beny:Osborne:2013,Beny:Osborne:2014} (c.f. \cite{DeBrota:2015} for a pedagogical introduction).
\subsection{Locally quantum multi-user information relativity\label{locally.quantum.information.relativity.section}}
The discussion in this paper is aimed at the construction of the framework of the multi-user (intersubjective) information relativity with emergent spaces. While most of discussion is kept in the framework build upon the W$^*$-algebras, it is more a useful testing ground (and a verifying constraint for backwards compatibility), then a desired property of the framework. In other words, our intention is to get rid of the \textit{Postulate 1} of Section \ref{qig.foundations.intro.section} by a suitable reformulation of the \textit{Postulate 5} and \textit{6}. We propose to consider the following tentative structure:
\begin{enumerate}
\item[1)] Consider arbitrary set $\M$ of users.
\item[2)] each user $\phi\in\M$ is equipped with his/her own vector space $V(\phi)$ containing his/her descripton of `states' (`configurations', `preparations', `inputs') and a dual vector space $V^\dual(\phi)$, containing his/her description of `effects' (`registrations', `results', `outputs').
\item[3)] User's notion of inference (respectively, causality) is modelled by a choice of set of of endomorphisms of $V(\phi)$ (respectively, $V^\dual(\phi)$).
\item[4)] In particular, inferences on $V(\phi)$ can be provided by means of the Br\`{e}gman functional $\tilde{D}_{\Psi,\phi}:V(\phi)\times V(\phi)\ra[0,\infty]$, while the causality on $V^\dual(\phi)$ can be provided either by the Legendre--Fenchel conjugate of $\tilde{D}_{\Psi,\phi}$, or by a representation of some Lie algebra.
\item[5)] More specifically, if one wants to do statistical inference, it is necessary to make clear how the behaviour of finite data sets is associated with the specific idealisations used in the theoretic framework, as represented by the `ideal' theoretical states and effects. In order to assert such relation, it is necessary to admit some statistical tools that are representing the control over a `convergence to an ideal form of a data set', for any given nonideal form of a real data set. Large number estimation and asymptotic estimation are two typical tools (on the geometric level, they correspond, respectively to relative entropy maximisation, and the local linearity). If a user introduces a discrimination function (information potential, absolute entropy) $\Psi_\phi:V(\phi)\ra\,]-\infty,+\infty]$, then he/she is able to quantify the rates of convergence of sequences of data. The key property of the Br\`{e}gman functional is that it allows for a generalised pythagorean theorem \eqref{generalised.pythagore.Bregman}, which is a nonlinear generalisation of the fundamental property of euclidean and Hilbert spaces. Yet, it is doing it without necessity of assuming that $V(\phi)$ is normed, or even metrisable. As a result, user's local inferences on $V(\phi)$ based on entropic projections $\PPP^{\tilde{D}_{\Psi,\phi}}$ allow to decompose the information distance to an `ideal inference' ($\tilde{D}_{\Psi,\phi}(x,y)$) into a sum of a distance to an effective solution (satisfying given constraints and minimising the distance) and an uncertainty within the constrained space. (By the Legendre--Fenchel duality, completely parallel considerations are applicable to causality on $V^\dual(\phi)$ determined by the relative free energy $\tilde{D}_{\Psi^\lfdual,\phi}:V^\dual(\phi)\times V^\dual(\phi)\ra[0,\infty]$.)
\item[6)] The geometric structures on the set $\M$ are introduced as quantitative means of relating (synchronising) inferences and causality of different users.
\item[7)] Each user $\phi$ provides his/her own mappings from $\M$ into $V(\phi)$, given by bijective embeddings $\ell_\phi:U(\phi)\ra V(\phi)$, where $U(\phi)\subseteq\M$ is the subset of users in $\M$ that a user $\phi$ considers as representable in terms of his/her `configuration' space $V(\phi)$.
\item[8)] Using the embeddings $\ell_\phi$, each user can relate his/her individual inferences on $V(\phi)$ with other users using (a part of) the same set $U$. In particular, the relative Br\`{e}gman entropy $D_\Psi$ on $\M$ is induced by
\begin{equation}
\tilde{D}_{\Psi,\phi}(\ell_\phi(\omega_1),\ell_\phi(\omega_2))=D_{\Psi,U(\phi)\cap U(\psi)}(\omega_1,\omega_2)=\tilde{D}_{\Psi,\psi}(\ell_\psi(\omega_1),\ell_\psi(\omega_2))\;\;\forall \omega_1,\omega_2\in U(\phi)\cap U(\psi).
\end{equation}
This provides the means to relate the \textit{large number} inferential dynamics of different users (such as in Sanov's theorem).
\item[8)] The local smooth manifold structure induced by $D_\Psi$ on $\M$ provides the means to relate \textit{asymptotic} inferential dynamics of different users. In particular, the equivalence of local geodesic projection and local $D_\Psi$ projections can be understood as a method of locally linear synchronisation of inferences of different users. In such case the curvature of the manifold $\M$ measures impossibility of ideal synchronisation of inferences between different users. The relationships with the notions of causality, effective dynamics, and renormalisation were discussed in the previous Section. They are dependent only on the notion of Br\`{e}gman distance on an arbitrary set $\M$, and Eguchi equations on an arbirary smooth manifold $\M$, hence they hold in general, without assuming that $\M$ is a subset of $\N^+_\star$ for some W$^*$-algebra $\N$.
\item[9)] The reconstruction of the special case, when $\M$ is equal to the set $\M(\N)$ of states over a globally defined W$^*$-algebra $\N$ is an open problem. Our conjecture is that quantum mechanics can be characterised as a set $\M$ equipped with an induced structure of the riemannian manifold $(\M,\gbold^{D_\Psi})$ and a Poisson manifold such that the set of extremal points of the convex hull of $\M$ admits an induced metric and induced symplectic structure that determine a K\"{a}hler manifold satisfying the standard properties of the quantum mechanical K\"{a}hler manifolds.
\end{enumerate}
In principle, the above scheme is applicable to a wide class of postquantum information theoretical settings, such as general probabilistic theories. In what follows, having in mind the possible applications in nonequilibrium quantum statistical mechanics, we will focus on its implementation in the context of W$^*$-algebras and the associated functional analytic spaces.
While \eqref{SS} considers only the case of local $\nabla^{1/2}$/$D_{1/2}$-projections, \cite{Hellmann:Kaminski:Kostecki:2016,Kostecki:2014,Kostecki:MunkNielsen:2016} consider only the case of global $D_0$-projections. Yet, we think that the natural geometric objects for construction of local and nonlocal quantum kinematics and dynamics are arbitrary Banach Lie algebras (to describe causality) and arbitrary quantum Br\`{e}gman distances (to describe inference), connected together via dual pairs of noncommutative Orlicz spaces. More specifically, following a discussion in Section \ref{local.infogeometry.histories.intro.section} we think that the problems considered in this paper should be readdressed in a more general foundational framework, based on the following principles:
\begin{enumerate}
\item[1)] The use of GNS Hilbert bundle should be replaced by a suitable bundle of noncommutative Orlicz spaces $L_{\Upsilon(\phi)}(\N)$ playing the role of tangent spaces $\T_\phi\M(\N)$, understood as the spaces of local `configurations' (e.g. $\phi(\theta)\mapsto\theta\mapsto\left(\frac{\partial}{\partial\theta_i}\right)$), with their Banach duals $L_{\Upsilon(\phi)}(\N)^\star$ playing the role of cotangent spaces $\T^\ct_\phi\M(\N)$, understood as the spaces of local `effects' ($f(\phi)\mapsto\ddd f\mapsto(\ddd f_i)$). Somewhat similar ideas were considered earlier in \cite{Streater:2009:book,Majewski:Labuschagne:2014}, but only in a global context, restricted to a single `tangent' and `cotangent' space.
\item[2)] There should be provided a canonical construction of a quantum Br\`{e}gman distance $D_\Psi$ associated with a Banach dual pair $(L_{\Upsilon(\phi)}(\N),L_{\Upsilon(\phi)}(\N)^\star)$, a function $\Psi:L_{\Upsilon(\phi)}(\N)\ra\RR$, and a family of embeddings $\ell_\phi:\N^+_\star\ra L_{\Upsilon(\phi)}(\N)$, such that the second order G\^{a}teaux derivative of $D_\Psi$ determines a map
\begin{equation}
\gbold_\phi^{D_\Psi}:L_{\Upsilon(\phi)}(\N)\times L_{\Upsilon(\phi)}(\N)^\star\ra[0,\infty],
\end{equation}
while the third order G\^{a}teaux derivatives determine the connections $\nabla^{D_\Psi}$ and $(\nabla^{D_\Psi})^\nsdual$, with the respective parallel transports equal to the isometric transition operators
\begin{align}
\transport^{\nabla^{D_\Psi}}_{\phi,\omega}&:\T_\phi\M(\N)\ra\T_\omega\M(\N),\\
\transport^{(\nabla^{D_\Psi})^\nsdual}_{\phi,\omega}&:\T^\ct_\phi\M(\N)\ra\T^\ct_\omega\M(\N),
\end{align}
and satisfying the generalised Norden--Sen duality
\begin{equation}
\gbold^{D_\Psi}_\phi(\transport^{\nabla^{D_\Psi}}_{\phi,\omega}(x),\transport^{(\nabla^{D_\Psi})^\nsdual}_{\phi,\omega}(y))=\gbold^{D_\Psi}_\omega(x,y).
\end{equation}
This is intended to implement the principle discussed in Section \ref{Orlicz.Br\`{e}gman.section}: the local structure of an information manifold should implement the local equivalence of an entropic projection (user's inference) and a geodesic free fall (an absence of geometric `gravity').
\item[3)] Lie algebras $\glie$ and Banach Lie algebras $\B$ should be represented on $L_{\Upsilon(\phi)}(\N)^\star$, the latter giving rise to a Banach Lie--Poisson manifold structure on $\M(\N)$ determined by the action of $\B_\star$ on $L_{\Upsilon(\phi)}(\N)$.
\item[4)] As a result, both $\ddd_{\nabla^{D_\Psi}}(\phi)$ and $\ddd h_\B(\phi)$ become the elements of the same operator space $L_{\Upsilon(\phi)}(\N)^\star$, so the local effective dynamics can be formulated in terms of a $1$-form ${\cal F}=\ddd h_\B+\ddd_{\nabla^{D_\Psi}}$, treating local causality and local inference on an equal footing. Note that, in face of the discussion in Sections \ref{locally.perturbed.liouvilleans.intro.section}, \ref{local.infogeometry.histories.intro.section}, and \ref{Orlicz.Br\`{e}gman.section}, the condition ${\cal F}=0$ can be understood as an information-theoretic analogue of Einstein's version of Newton's first law of motion: a rest in a local causal-inference frame (provided by user's choice of generators $\B$ for causal dynamics in a cotangent space of effects, and a discrimination function $\Psi$ for inferential dynamics in a tangent space of states).
\item[5)] The local tangent BLP hamiltonian vector field, generating a local W$^*$-dynamical system associated with a local fiber of a GNS Hilbert bundle, should be represented in terms of $L_\Upsilon$-liouvillean, acting on the $L_{\Upsilon(\phi)}(\N)$ tangent space. The incorporation of the contribution of the local free fall along $\nabla^{D_\Psi}$-connection geodesics should be provided by the perturbation of this $L_\Upsilon$-liouvillean, defined in an analogy with the Jak\v{s}i\'{c}--Pillet $L_{1/\gamma}$-liouvillean \cite{JOPP:2012}.
\item[6)] The local entropic prior should be constructed using $D$ and $\gbold^D$ representing the \textit{effective geometry} of a model $\M(\N)$, after the renormalisation of all contributions from control sources. One can interpret the local entropic prior at a point $\phi\in\M$ as an information theoretic analogue of the (inferential) mass of the user at $\phi$. This interpretation gives a particularly neat meaning to the observation \cite{Maraner:1992,Alicki:Klauder:Lewandowski:1993}, discussed in Section \ref{local.quantum.dynamics.intro.section}, that the regularising term in the Daubechies--Klauder formula leads to a point-dependence of a zero-point energy if{}f the curvature of the Fubini--Study metric is nonconstant. From our point of view, it means: the curvature $\kappa^{\nabla^{D}}$ of a quantum model is a measure of desynchronisation of the ideal multi-user inference (as given locally by $\nabla^{D_\Psi}$-geodesics, or, equivalently, $\ddd_{\nabla^{D_\Psi}}$), reflected in the influence of a local mass on the local zero-point energy. The case of information model with a constant curvature is an inferential analogue of the Carath\'{e}odory--Jacobi--Lie theorem for symplectic manifold, allowing for a global trivial synchronisation of local causality systems.
\item[7)] Given these constructions, the dynamical Ansatz \eqref{SS} should be generalised by means of the replacement of the GNS Hilbert bundle by a corresponding dual bundle of noncommutative Orlicz spaces.
\item[8)] In principle, every quantum Br\`{e}gman functional $\tilde{D}_\Psi$ on $L_{\Upsilon(\phi)}(\N)$ determines the Legendre--Fenchel conjugate functional $\tilde{D}^\lfdual_{\Psi}$ on $L_{\Upsilon(\phi)}(\N)$, which can be naturally interpreted as a relative free energy. If one would replace the use of the (nonlocally trivially synchronisable) $1$-form $\ddd h_\B$ by the form $\ddd_{\nabla^{D^\lfdual_\Psi}}$, then it would be necessary to introduce a corresponding local entropic prior (causal mass), measuring the influence of desynchronisation of local systems of causality, along the spatial trajectory on $\M$, on the effective quantitative nonlocal evolution. The postulate of the local cancellation of effects of those two (inferential and causal) priors, when integrated together, would be an information theoretic analogue of the postulate of equality of gravitational and inertial mass.
\item[9)] In the JMF source theoretic approach the local departure of riemannian metric from the hessian geometry is described by the equation \eqref{JMF.renormalisation.of.metric}. Applying the change of geometry $\gbold^{D_\Psi}\mapsto\tilde{\gbold}$ to a definition of a prior $\exp\left(-k\frac{\epsilon^2}{2}\int_\pathgamma\dd t\,\gbold_{ab}(\phi)\dot{\phi}^a\dot{\phi}^b\right)\sqrt{\det(\gbold(\phi))}$, we can see that the change of the scalar curvature $\kappa(\phi)$ of the model is reflected in the redefinition of the local prior (information mass) $P$. This leads to a more general problem. Given any neighbourhood $U\subseteq\M$ of a user $\phi\in\M$, one can ask how to determine the departure of a prior $P$ on $U$ (from the originally postulated one) that is caused by the changes of curvature $\kappa$ of a model $\M$ associated with a given system of inference, when the transformation of geometry of a model $\M$ (e.g., due to the Jaynes--Mitchell renormalisation) is considered. The analogy with general relativity that we were pursuing in this paper, as well as an observation \cite{Watson:Klauder:2002} that the uniform weighting of the trajectories requires constant scalar curvature, leads us to conjecture that
\begin{equation}
\delta\int_UP=\delta\int_U\kappa.
\end{equation}
\end{enumerate}
We hope that this mathematical framework will allow for a unified treatment of the foundations of nonequilibrium quantum statistical mechanics, as discussed from different perspectives in \cite{Grandy:2008}, \cite{Streater:2009:book}, and \cite{JOPP:2012} (see also \cite{Caticha:2012}). More generally, we consider it to be a testing ground for a construction of a predictive dynamical theory that would unify several concepts of general relativity with quantum information theory into a single framework of multi-agent (post)quantum information relativity. Maybe this sounds a bit like a fountain of conjectures and high hopes, but we believe that the science fiction of today is just an advanced propagator of a scientific research of tomorrow.
\ifvargaugecompile
\section{Quantum geometry \& global dynamics\label{geometric.structures.section}}
\else
\subsection{Quantum geometry\label{geometric.structures.section}}
\fi
\ifvargaugecompile
Sections \ref{W.star.BLP.section}, \ref{standard.liouvilleans.section}, \ref{distances.NS.geom.section}, \ref{quant.info.geom.section} do not contain new results or constructions, except of the definition of the quantum Poisson system.
\subsection{Quantum Banach Lie--Poisson spaces\label{W.star.BLP.section}}
\else
\fi
\subsubsection{Banach--Lie--Poisson spaces}
Let $\KK\in\{\RR,\CC\}$. A vector space $X$ over $\KK$ is called a \df{Lie algebra} if{}f it is equipped with a function $[\cdot,\cdot]:X\times X\ra X$ such that for all $f_1,f_2,f_3\in X$ and for all $\lambda\in\KK$
\begin{enumerate}
\item[1)] $[f_1,f_2]=-[f_2,f_1]$ (antisymmetry),
\item[2)] $[f_1,f_2+\lambda f_3]=[f_1,f_2]+\lambda[f_1,f_3]$ (linearity),
\item[3)] $[f_1,[f_2,f_3]]+[f_3,[f_1,f_2]]+[f_2,[f_3,f_1]]=0$ (Jacobi identity).
\end{enumerate}
The function $[\cdot,\cdot]$ is called the \df{Lie bracket}. If $(X,[\cdot,\cdot])$ satisfies also the Leibniz's rule
\begin{enumerate}
\item[4)] $\{f_1,f_2f_3\}=\{f_1,f_2\}f_3+f_2\{f_1,f_3\}\;\;\forall f_1,f_2,f_3\in X$,
\end{enumerate}
then $[\cdot,\cdot]$ is called the \df{Lie--Poisson bracket} \cite{Lie:1890}, while $(X,[\cdot,\cdot])$ is called a \df{Lie--Poisson algebra}. If $G$ is a Lie group, then a Lie algebra of its generators will be denoted $\Lie(G)$. A vector space $X$ over $\KK$ is called a \df{Banach Lie algebra} if{}f it is a Banach space with a norm $\n{\cdot}$, a Lie algebra, and its Lie bracket $[\cdot,\cdot]:X\times X\ra X$ is bilinear and continuous in the topology of $\n{\cdot}$. If $\glie$ is a Banach Lie algebra, then the \df{adjoint map} $\rpktarget{ADMAP}\ad_x:\glie\ni y\mapsto[x,y]\in\glie$ and \df{coadjoint map} $\ad_x^\banach:\glie^\banach\ra\glie^\banach$,
\begin{equation}
\duality{y,\ad^\banach_x(z)}_{\glie\times\glie^\banach}:=\duality{\ad_x(y),z}_{\glie\times\glie^\banach}\;\;\forall z\in\glie^\banach,
\end{equation}
are norm continuous for each $x\in\glie$.
Let $M$ be a real Banach smooth manifold, and $\rpktarget{CIF}\CIF(M;\RR)$ denotes the space of all infinitely Fr\'{e}chet differentiable $\RR$-valued functions on $M$. Then a real \df{Poisson structure} on $M$ is defined as a function $\rpktarget{POISSON}\{\cdot,\cdot\}:\CIF(M;\RR)\times\CIF(M;\RR)\ra\CIF(M;\RR)$ such that $(\CIF(M;\RR),\{\cdot,\cdot\})$ is a Lie algebra \cite{Lie:1890}. %
\ifvargaugecompile
\else
\footnote{Note that, on an algebraic level, the conditions i.2) and 4) on $\{f,\cdot\}$ are the same as the conditions 1. and 2. on $\der(\cdot)$ in Section \ref{derivations.section}.}
\fi
If $M$ above is \textit{finite dimensional}, then $(M,\{\cdot,\cdot\})$ is called a real \df{Poisson manifold} \cite{Lichnerowicz:1977,Weinstein:1983,Vaisman:1994,Weinstein:1998}. If $M$ is a real Banach smooth manifold, then the cotangent space at $x$ can be defined by $\T_x^\ct M:=(\T_xM)^\banach$, and each element of $\T_x^\ct M$ has a form
\begin{equation}\rpktarget{DDDf}
\ddd f(x)\equiv\ddd_x f:\T_xM\ni v\mapsto\ddd_x f(v):=v(f)\in\RR
\end{equation}
for some $f\in\CIF(M;\RR)$. Let $\T^\ct{}^\ct M:=\bigcup_{x\in M}\T_x^\ct{}^\ct M$, where $\T_x^\ct{}^\ct M:=(\T^\ct_x M)^\banach=(\T_x M)^\banach{}^\banach$. If $(M,\{\cdot,\cdot\})$ is a real Poisson manifold, then every $k\in\CIF(M;\RR)$ determines a unique vector field $\XXX_k\in\T M$ by\rpktarget{HAMILVF}
\begin{equation}
\XXX_k(f):=\bigcup_{x\in M}\left\{x\mapsto\duality{\XXX_k(x),\ddd f(x)}_{\T_xM\times\T_x^\ct M}\right\}=\{f,k\}\;\forall f\in\CIF(M;\RR).
\label{hamiltonian.vector.field}
\end{equation}
Such $\XXX_k$ is called a \df{hamiltonian vector field}, while the corresponding $k$ is called the \df{Hamilton function}. If $M$ is infinite dimensional, then the value of $\{\cdot,\cdot\}$ at $x\in M$ depends only on the differentials $\ddd f(x),\ddd k(x)\in\T_x^\ct M$. Hence, there exists a section $\rpktarget{VARPI}\varpi$ of the vector bundle $\bigwedge^2\T^\ct{}^\ct M$ such that
\begin{equation}
\{f,k\}=\varpi(\ddd f,\ddd k).
\end{equation}
Let $\varpi$ be a \textit{smooth section}, meaning that for each $x\in M$ there exists a continuous bilinear antisymmetric function $\varpi_x:\T_x^\ct M\times\T_x^\ct M\ra\RR$ such that $x\mapsto\varpi_x$ is smooth. Let the function $\natural:\T^\ct M\ra\T^\ct{}^\ct M$ be the bundle map covering identity, isometric on fibres, and satisfying
\begin{equation}
\natural_x(\ddd f(x)):=\varpi_x(\cdot,\ddd f),
\end{equation}
which means that
\begin{equation}
(\natural_x(\ddd f(x)))(\ddd k(x))=\{k,f\}(x)\;\;\forall f,k\in\CIF(M;\RR).
\end{equation}
Then
\begin{equation}
\XXX_k:=\varpi(\cdot,\ddd k)=\natural(\ddd k)=\{\cdot,k\}
\label{smooth.section.ttwom}
\end{equation}
is a smooth section of $\T^\ct{}^\ct M$, but may not be a vector field on $M$, because $\T M\subseteq\T^\ct{}^\ct M$ and $\T M\not\iso\T^\ct{}^\ct M$ in general. In order to solve this problem, Odzijewicz and Ratiu \cite{Odzijewicz:Ratiu:2003} proposed (see also further discussion and results in \cite{Odzijewicz:Ratiu:2004,Beltita:Ratiu:2005,Beltita:Ratiu:Tumpach:2008,Odzijewicz:Ratiu:2008,Tumpach:2009,Ratiu:2011,Odzijewicz:2011}) to define a real \df{Banach Poisson manifold} as a pair $(M,\{\cdot,\cdot\})$ of a real Banach smooth manifold $M$ and a Poisson structure $\rpktarget{POISSON.TWO}\{\cdot,\cdot\}$ on it such that the function $\natural:\T^\ct M\ra\T^\ct{}^\ct M$ defined above satisfies $\natural(\T^\ct M)\subseteq\T M$.
If $(M,\{\cdot,\cdot\})$ is a real Banach Poisson manifold, then every $k\in\CIF(M;\RR)$ determines a unique vector field $\XXX_k\in\T M$ defined by \eqref{smooth.section.ttwom}, and called a \df{hamiltonian vector field}. Such $k$ is then called the \df{Hamilton function}. Every real Poisson manifold is a real Banach Poisson manifold, so this terminology is consistent. Odzijewicz and Ratiu \cite{Odzijewicz:Ratiu:2003} define also a \textit{holomorphic} Banach Poisson manifold which provides an analogous setting for $\KK=\CC$, but we will not use its specific properties here, so we omit its definition. If a Banach space $X$ over $\KK$ is equipped with a Poisson structure $\{\cdot,\cdot\}$ that turns it into a (real or holomorphic) Banach Poisson manifold, then $X^\banach\subseteq\CIF(X;\KK)$. Moreover, $\{\cdot,\cdot\}$ is linear on $\CIF(X;\KK)$ if $\{X^\star,X^\star\}\subseteq X^\star$.
A \df{Banach--Lie--Poisson space} is defined \cite{Odzijewicz:Ratiu:2003} as a pair $(X,\{\cdot,\cdot\})$ such that
\begin{enumerate}
\item[1)] $X$ is a Banach space over $\KK$,
\item[2)] $(X,\{\cdot,\cdot\})$ is a real (if $\KK=\RR$) or holomorphic (if $\KK=\CC$) Banach Poisson manifold,
\item[3)] $X^\banach\subseteq\CIF(X;\KK)$ is a Banach Lie algebra with respect to $\{\cdot,\cdot\}$.
\end{enumerate}
In such case, the restriction of $\{\cdot,\cdot\}$ on $\CIF(X;\KK)$ to $X^\banach$ will be denoted by $[\cdot,\cdot]$. As proved in \cite{Odzijewicz:Ratiu:2003}, the Banach space $X$ is a BLP space $(X,\{\cdot,\cdot\})$ if{}f $X^\banach$ is a Banach Lie algebra $(X^\banach,[\cdot,\cdot])$ satisfying
\begin{equation}
\ad_x^\banach(X)\subseteq X\subseteq X^\banach{}^\banach\;\;\forall x\in X^\banach,
\end{equation}
and in such case $\{\cdot,\cdot\}$ is given by
\begin{equation}
\{f,k\}(z)=\duality{z,[\DF_zf,\DF_zk]}_{X\times X^\banach}\;\;\forall f,k\in\CIF(X;\KK)\;\forall z\in X.
\label{BLP.three}
\end{equation}
Moreover, under those conditions the hamiltonian vector field associated to \textit{any} $k\in\CIF(X;\KK)$ reads\rpktarget{HAMILT.ZWEI}
\begin{equation}
\XXX_k(z)=-\ad^\banach_{\DF_zk}(z)\;\;\forall z\in X.
\label{BLP.four}
\end{equation}
If $(X,\{\cdot,\cdot\})$ is a BLP space, and if $X\iso\T_xX$ $\forall x\in X$, then $\T_x^\ct X\iso(\T_x X)^\banach\iso X^\banach$, so if $f\in X^\banach$, then one can identify $\ddd_x f\in\T_x^\ct X$ with $\DF_xf\in X^\banach$. As a result, for any $z\in X$ and $y\in X^\banach$ the linearity of $y$ gives $\DF_zy=y$, and for every $x\in X$ one has \cite{Bona:2000,Odzijewicz:Ratiu:2003}
\begin{align}
\duality{y,\ad^\banach_x(z)}_{X^\banach\times X^\banach{}^\banach}&=\duality{z,[x,y]}_{X\times X^\banach}=\{x,y\}(z)=-\{y,x\}(z)\\
&=(\XXX_x(y))(z)=\duality{\DF_zy,\XXX_x(z)}_{X^\banach\times X^\banach{}^\banach}=-\duality{y,\XXX_x(z)}_{X^\banach\times X^\banach{}^\banach}.
\end{align}
Hence,
\begin{equation}
\XXX_x(z)=-\ad^\banach_x(z)\;\;\;\forall z\in X\;\forall x\in X^\banach.
\end{equation}
The notion of the BLP space can be viewed as a suitable generalisation of the important properties of \textit{strong} symplectic manifold to infinite dimensional situation which need not admit decomposition into symplectic leaves. If $(M_1,\{\cdot,\cdot\}_1)$ and $(M_2,\{\cdot,\cdot\}_2)$ are BLP spaces, then a smooth function $w:M_1\ra M_2$ is called a \df{Poisson map} if{}f
\begin{equation}
\{f\circ w,k\circ w\}_1=\{f,k\}_2\circ w\;\;\forall f,k\in\CIF(M_2;\KK).
\label{BLP.five}
\end{equation}
As shown in \cite{Marsden:Ratiu:1994}, the condition iii) above makes \eqref{BLP.five} equivalent to
\begin{equation}
\XXX_k\circ w=\T w\circ\XXX_{k\circ w}\;\;\forall k\in\CIF(M_2;\KK).
\end{equation}
If $(M,\{\cdot,\cdot\})$ is a BLP space and $h\in\CIF(M;\KK)$, then the \df{Hamilton equation}
\begin{equation}
\frac{\dd}{\dd t}f(w^h_t(x))=\{h,f(w^h_t)\}(x)\;\;\forall f\in\CIF(M;\KK)\;\;\forall t\in\RR\;\;\forall x\in M
\label{hamilton.eqn}
\end{equation}
determines a unique \textit{local} map $\rpktarget{WHTM}w^h_t:M\ra M$, called a \df{hamiltonian flow} of $h$, which is a Poisson map. The solutions of the equation $x(t)=w^h_t(x)$ with $x(0)=x$ need not exist \textit{globally}, that is, for all $t\in\RR$ and all $x\in M$. If they exist globally, then the hamiltonian vector field $\{\cdot,h\}$ is called \df{complete}.
\ifvargaugecompile
\subsubsection{W$^*$-algebra predual as a BLP space}
\else
\subsubsection{W$^*$-algebra predual as a BLP space}
\fi
If $M$ is a Banach space, and a Banach smooth manifold modelled on itself by means of an identity mapping then for each $x\in M$ there is a Banach space isomorphism $\T_xM\iso M$. If $\N$ is a W$^*$-algebra, then the Banach Lie algebra structure of $(\N_\star)^\banach\iso\N$ is given by its commutator $[x,y]:=xy-yx$, while $\rpktarget{AD.SECOND}\ad_x:=[x,\,\cdot\,]=\LLL_x-\RRR_x$ and $\ad_x^\banach=\LLL^\banach_x-\RRR^\banach_x$ are defined by weakly-$\star$ continuous maps $\LLL_x:\N\ni y\mapsto xy\in\N$, $\RRR_x:\N\ni y\mapsto yx\in\N$. The condition $\ad_x^\banach(\N_\star)\subseteq\N_\star$ holds for all $x\in\N$, so $\N_\star$ is a BLP space that is a holomorphic Banach Poisson manifold (modelled on itself by the atlas consisting of one chart, an identity mapping $\id_{\N_\star}$) with the Poisson structure given by \eqref{BLP.three},\rpktarget{POISSON.DREI}
\begin{equation}
\{f,k\}(\phi)=\phi([\DF_\phi f,\DF_\phi k])\;\;\forall f,k\in\CIF(\N_\star;\CC)\;\forall\phi\in\N_\star.
\label{BLP.six}
\end{equation}
As a result, the hamiltonian vector field associated to every $k\in\CIF(\N_\star;\CC)$ by means of \eqref{BLP.four} takes a form\rpktarget{HAMILT.DREI}
\begin{equation}
\XXX_f(\phi)=-\ad^\banach_{\DF_\phi f}(\phi)=\LLL^\banach_{\DF_\phi f}(\phi)-\RRR^\banach_{\DF_\phi f}(\phi)\;\;\forall\phi\in\N_\star.
\label{BLP.seven}
\end{equation}
These results, including the BLP space structure of $\N_\star$, were discovered by B\'{o}na \cite{Bona:1991,Bona:1993,Bona:2000} in the $\N_\star=\schatten_1(\H)=\BH_\star$ case, and were generalised to arbitrary W$^*$-algebras by Odzijewicz and Ratiu \cite{Odzijewicz:Ratiu:2003}. We will call \eqref{BLP.seven} the \df{B\'{o}na--Odzijewicz--Ratiu equation}.
If $\N=\BH$ then $\N_\star\iso\schatten_1(\H)$ and for every $\rho\in\schatten_1(\H)$ and every $x,y\in\BH$
\begin{equation}
\duality{y,-\ad^\banach_x(\rho)}_{\BH\times\BH^\banach}=-\duality{[x,y],\rho}_{\BH\times\BH^\banach}=-\tr_\H([x,y]\rho)=\duality{[x,\rho],y}_{\schatten_1(\H)\times\BH}
\label{BLP.eight}
\end{equation}
which follows from the fact that $\schatten_1(\H)$ is an ideal in $\BH$. As a result,
\begin{equation}
-\ad_x^\banach(\rho)=[x,\rho],
\end{equation}
and the BOR equation \eqref{BLP.seven} turns to the Lax equation \cite{Lax:1968}
\begin{equation}
\XXX_f(\rho)=[\DF_\rho f,\rho]\;\;\forall\rho\in\schatten_1(\H).
\end{equation}
In particular, a choice of the Hamilton function $h(\rho):=\tr_\H(H\rho)$, where $H\in\BH$ but is not necessarily self-adjoint, turns \eqref{hamilton.eqn} to
\begin{equation}
\frac{\dd}{\dd t}\rho(t)=-\ad^\banach_{\DF_\rho h}(\rho)=[H,\rho].
\end{equation}
\subsubsection{Quantum Poisson systems}
The above equation is derived for a Poisson structure on $\N_\star=\schatten_1(\H)$ viewed as a holomorphic Banach Poisson manifold. However, the standard construction of unitary dynamics in quantum mechanics makes us to be more interested in the real Banach Poisson manifold $\N_\star^\sa$ (and submanifolds of it that are also subsets of $\N^+_\star$), equipped with the Poisson structure coinduced by the action of the Lie algebra $\N^\asa$ of anti-selfadjoint elements of $\N$. More precisely, the set $\N^\uni$ of all unitary elements of a W$^*$-algebra $\N$ is a real Banach Lie group and has a real Lie Banach algebra $\Lie(\N^\uni)=\N^\asa:=\{x\in\N\mid x=-x^*\}$ with $[x,y]:=xy-yx$ \cite{Bourbaki:1972}. The elements of the Banach Lie algebra $\N^\asa$ can be represented by $x\in\N^\sa=\ii\N^\asa$, using the Lie bracket $\N^\sa\times\N^\sa\ni(x,y)\mapsto\ii[x,y]\in\N^\sa$, which corresponds to the commutator $[\ii x,\ii y]=\ii z$ in $\N^\asa$. This algebra has a unique Banach predual, given by $\N_\star^\sa\iso L_1(\N)^\sa:=\{\phi\in L_1(\N)\mid\phi=\phi^*\}$, with an isomorphism $(\N_\star^\sa)^\banach\iso\N^\sa$ defined by duality
\begin{equation}
\N_\star^\sa\times\N^\sa\ni(\phi,x)\mapsto\duality{\phi,x}_{\N_\star^\sa\times\N^\sa}:=\phi(x)\in\RR.
\label{duality.sa.sa}
\end{equation}
The adjoint representation $\Ad(\N^\uni)$ of a Banach Lie group $\N^\uni$ on $\Lie(\N^\uni)=\ii\N^\sa=\N^\asa$,
\begin{equation}\rpktarget{AD.BIG}
\Ad(u)x:=uxu^*\;\;\forall u\in\N^\uni\;\forall x\in\N^\sa
\end{equation}
determine the coadjoint representation $\Ad^\banach(\N^\uni)$ on $(\N^\banach)^\sa$,
\begin{equation}
\duality{x,\Ad^\banach(u)\phi}_{\N^\sa\times(\N^\banach)^\sa}:=\duality{\Ad(u^{-1})x,\phi}_{\N^\sa\times(\N^\banach)^\sa}\;\;\forall x\in\N^\sa\;\forall\phi\in(\N^\banach)^\sa.
\end{equation}
Using these properties, one can show that \cite{Bona:2000,Odzijewicz:Ratiu:2003}
\begin{equation}
\ad^\banach_x(\N_\star^\sa)\subseteq\N_\star^\sa\subseteq(\N^\asa)^\banach\;\;\forall x\in\N^\asa.
\end{equation}
The space $\N_\star^\sa$ can be equipped with a real Banach smooth manifold structure modelled on itself by the atlas consisting of one chart, which is determined by the identity mapping on $\N_\star^\sa$. As a result, $\T_\phi(\N_\star^\sa)\iso\N_\star^\sa$ $\forall \phi\in\N_\star^\sa$. So, it is possible to use \eqref{BLP.three} and \eqref{duality.sa.sa} to \textit{define} the BLP structure on $\N_\star^\sa$ by
\begin{equation}
\{f,k\}(\phi):=\ii\phi([\DF_\phi f,\DF_\phi k])\;\;\forall f,k\in\CIF(\N_\star^\sa;\RR)\;\forall\phi\in\N_\star^\sa.
\label{BLP.on.N.sa}
\end{equation}
As a result, the Hamilton equation \eqref{hamilton.eqn} for $h\in\CIF(\N_\star^\sa;\RR)$ reads
\begin{equation}
\frac{\dd}{\dd t}f(\phi(t))=\{h,f\}(\phi(t))=\ii(\phi(t))\left([\DF_{\phi(t)}h,\DF_{\phi(t)}f]\right).
\label{BLP.nine}
\end{equation}
The spaces $\N^+_{\star1}$, $\N^+_\star$ and $(\N_\star^\sa,\{\cdot,\cdot\}_{\N_\star^\sa})$ are subsets of $(\N^\banach)^\sa$ that are invariant with respect to $\Ad^\banach(\N^\uni)$. As shown in \cite{Beltita:Ratiu:2005}, they decompose into union of orbits of $\Ad^\banach(\N^\uni)$, which in turn are weak symplectic manifolds, which provides the symplectic foliation of the BLP space $(\N_\star^\sa,\{\cdot,\cdot\}_{\N_\star^\sa})$. Similarly, $(\N_\star,\{\cdot,\cdot\}_{\N_\star})$ is invariant with respect to the action of the Banach Lie group $\N^\inv$ of all invertible elements of $\N$. If $\N=\BH$ and $\rho\in\schatten_1(\H)^\sa$, then the calculation analogous to \eqref{BLP.eight} gives
\begin{equation}
\ad_x^\banach(\rho)=[\rho,x]\;\;\forall x\in\BH^\asa\;\forall \rho\in\schatten_1(\H)^\sa.
\end{equation}
As a result, the BOR equation \eqref{BLP.seven} on $\N_\star^\sa=\schatten_1(\H)^\sa$ takes the form
\begin{equation}
\XXX_f(\rho)=[\rho,\DF_\rho f]\;\;\forall\rho\in\schatten_1(\H),
\end{equation}
while the Hamilton equation \eqref{BLP.nine} becomes \cite{Bona:2000}
\begin{equation}
\frac{\dd}{\dd t}f(\rho(t))=\ii\,\tr_\H\left([\rho(t),\DF_{\rho(t)}h]\DF_{\rho(t)}f\right).
\label{BLP.ten}
\end{equation}
Because of the identity
\begin{equation}
\frac{\dd}{\dd t}f(\rho(t))=\tr_\H\left((\DF_{\rho(t)}f)\frac{\dd}{\dd t}\rho(t)\right),
\end{equation}
the equation \eqref{BLP.ten} is equivalent to the \df{B\'{o}na equation} \cite{Bona:1991,Jordan:1993,Bona:2000},
\begin{equation}
\ii\frac{\dd}{\dd t}\rho(t)=[\DF_{\rho(t)}h,\rho(t)].
\label{gen.vN}
\end{equation}
The solutions of \eqref{gen.vN} are state-dependent unitary operators $U(\rho,t)$. They do not form a group, but satisfy a cocycle relationship:
\begin{equation}
U(\rho,t+s)=U((\Ad(U(\rho,t)))(\rho),s)U(\rho,t)\;\;\forall t,s\in\RR.
\end{equation}
In the special case, when $h(\rho)=\tr_\H(\rho H)$ for $H\in\BH^\sa=\ii\BH^\asa$, \eqref{gen.vN} turns to the \df{von Neumann equation}
\begin{equation}
\ii\frac{\dd}{\dd t}\rho(t)=[H,\rho(t)].
\label{vN.equation.from.BLP}
\end{equation}
So far we have followed the B\'{o}na--Odzijewicz--Ratiu approach, hence our main object of interest was the real Banach manifold $\N^\sa_\star$, equipped with the BLP space structure coinduced by the Banach--Lie algebra $\N^\asa$ via $\N^\sa$, corresponding to the group $\N^\uni$ of unitary elements of $\N$. However, in principle, a geometric setting for nonlinear dynamics of quantum models can be generated by an arbitrary Banach Lie algebra $\B$ over $\RR$ such that:
\begin{enumerate}
\item[(i)] its Banach predual space $\B_\star$ exists, is unique, and is a real Banach Poisson manifold,
\item[(ii)] $\ad^\star_x(\B_\star)\subseteq\B_\star\subseteq\B^\star$ $\forall x\in\B$,
\item[(iii)] there exists a nonempty set $\M(\N,\B)\subseteq\N^+_\star$ that is a real BLP submanifold of $\B_\star$.
\end{enumerate}
In what follows, we will call such Banach Lie algebras $\B$ to be \df{well-adapted}. For any choice of $h\in\CIF(\M(\N,\B);\RR)$, the pair $(\M(\N,\B),h)$ will be called a \df{quantum Poisson system} whenever
\begin{equation}
w^h_t(\phi)\in\M(\N,\B)\;\;\forall \phi\in\M(\N,\B)\;\;\forall t\in\RR.
\label{quantum.Poisson.system.condition}
\end{equation}
Hence, each quantum Poisson system is determined by a choice of: a space of quantum states $\M(\N)\subseteq\N^+_\star$, a Banach Lie algebra $\B$, a tangent bundle (real Banach manifold) structure on $\M(\N)$, and a real Fr\'{e}chet smooth function on $\M(\N)$, satisfying the conditions (i), (ii), (iii), and \eqref{quantum.Poisson.system.condition}. The assumptions $\B\iso\N^\sa$ and $\T_\phi\N^\sa_\star:=\N_\star^\sa$ $\forall\phi\in\N^\sa_\star$ recover the BOR setting completely. Note that a general quantum Poisson system $(\M(\N,\B),h)$ does not have to be related to any group, so in particular to a group of unitary operators $\N^\uni$. The only shared property (securing the backwards compatibility with quantum mechanical setting) is implementation of the Poisson flow on the predual by means of a coinduced action of a Banach--Lie algebra.
\ifvargaugecompile
\else
\subsubsection{Symplectic submanifolds}
[[write about symplectic orbits (and maybe also lagrangeans!)]]
[[write also about the solutions to B\'{o}na equation etc!]]
\fi
\ifvargaugecompile
\subsection{Relative modular operators, standard liouvilleans and the GNS bundle\label{standard.liouvilleans.section}}
A \df{weight} on a W$^*$-algebra $\N$ is defined as a function $\omega:\N^+\ra[0,+\infty]$ such that $\omega(0)=0$, $\omega(x+y)=\omega(x)+\omega(y)$, and $\lambda\geq0\limp\omega(\lambda x)=\lambda\omega(x)$, with the convention $0\cdot(+\infty)=0$. A weight is called: \df{faithful} if{}f $\omega(x)=0\limp x=0$; \df{finite} if{}f $\omega(\II)<\infty$; \df{semi-finite} if{}f a left ideal in $\N$ given by $\nnn_\phi:=\{x\in\N\mid\phi(x^*x)<\infty\}$ is weakly-$\star$ dense in $\N$; \df{normal} if{}f $\omega(\sup\{x_\iota\})=\sup\{\omega(x_\iota)\}$ for any uniformly bounded increasing net $\{x_\iota\}\subseteq\N^+$. A space of all normal semi-finite weights on a W$^*$-algebra $\N$ is denoted $\W(\N)$, while the subset of all faithful elements of $\W(\N)$ is denoted $\W_0(\N)$. Hence, $\N_\star^+\subset\W(\N)$ and $\N_{\star0}^+\subset\W_0(\N)$. For $\psi\in\W(\N)$, $\supp(\psi)=\II-\sup\{P\in\Proj(\N)\mid\psi(P)=0\}$. An element $\omega\in\N^+_\star$ is faithful if{}f $\supp(\omega)=\II$.
A \df{representation} of a W$^*$-algebra $\N$ is defined as a pair $(\H,\pi)$ of a Hilbert space $\H$ and a $*$-homomorphism $\rpktarget{pi}\pi:\N\ra\BH$. A representation $\pi:\N\ra\BH$ is called: \df{nondegenerate} if{}f $\{\pi(x)\xi\mid (x,\xi)\in\N\times\H\}$ is dense in $\H$; \df{normal} if{}f it is continuous with respect to the weak-$\star$ topologies of $\N$ and $\BH$; \df{faithful} if{}f $\ker(\pi)=\{0\}$. An element $\xi\in\H$ is called \df{cyclic} for a W$^*$-algebra $\N\subseteq\BH$ if{}f $\N\xi:=\bigcup_{x\in\N}\{x\xi\}$ is norm dense in $\BH$. A representation $\pi:\N\ra\BH$ is called \df{cyclic} if{}f there exists $\Omega\in\H$ that is cyclic for $\pi(\N)$. According to the Gel'fand--Na\u{\i}mark--Segal theorem \cite{Gelfand:Naimark:1943,Segal:1947:irreducible} for every pair $(\N,\omega)$ of a W$^*$-algebra $\N$ and $\omega\in\N^{\banach+}$ there exists a triple $(\H_\omega,\pi_\omega,\Omega_\omega)$ of a Hilbert space $\H_\omega$ and a cyclic representation $\pi_\omega:\N\ra\BH$ with a cyclic vector $\Omega_\omega\in\H_\omega$\rpktarget{h.omega}, and this triple is unique up to unitary equivalence. An analogue of this theorem for weights follows the similar construction, but lacks cyclicity. If $\omega$ is a weight on a W$^*$-algebra $\N$, then there exists the Hilbert space $\H_\omega$, defined as the completion of $\nnn_\omega/\ker(\omega)$ in the topology of a norm generated by the scalar product $\s{\cdot,\cdot}_{\omega}:\nnn_\omega\times\nnn_\omega\ni(x,y)\mapsto\omega(x^*y)\in\CC$,
\begin{equation}\rpktarget{h.omega.zwei}
\H_\omega:=\overline{\nnn_\omega/\ker(\omega)}=\overline{\{x\in\N\mid\omega(x^*x)<\infty\}/\{x\in\N\mid\omega(x^*x)=0\}}=\overline{\nnn_\omega/\I_\omega},
\end{equation}
and there exist the maps\rpktarget{rep.omega.zwei}\rpktarget{pi.omega.zwei}
\begin{align}
[\cdot]_\omega:\nnn_\omega\ni x&\mapsto [x]_\omega\in\H_\omega,
\label{GNS.class.weight}\\
\pi_\omega:\N\ni x&\mapsto([y]_\omega\mapsto[xy]_\omega)\in\BBB(\H_\omega),
\label{GNS.rep.weight}
\end{align}
such that $[\cdot]_\omega$ is linear, $\ran([\cdot]_\omega)$ is dense in $\H_\omega$, and $(\H_\omega,\pi_\omega)$ is a representation of $\N$. If $\omega\in\W(\N)$ then $(\H_\omega,\pi_\omega)$ is nondegenerate and normal. It is also faithful if $\omega\in\W_0(\N)$.
A standard representation \cite{Haagerup:1975:standard:form} of a W$^*$-algebra $\N$ is defined as a quadruple $(\H,\pi,J,\stdcone)$ of a Hilbert space $\H$, a nondegenerate faithful weakly-$\star$ continuous representation $\pi:\N\ra\BH$, a conjugation $J:\H\ra\H$, and a self-polar cone\footnote{A subspace $\D$ of a Hilbert space $\H$ is called a self-polar cone if{}f $\lambda\xi\in\D$ $\forall\xi\in\D$ $\forall\lambda\geq0$ and $\D=\{\zeta\in\H\mid\s{\xi,\zeta}_\H\geq0\;\forall\xi\in\D\}$.} $\stdcone\subseteq\H$, satisfying the conditions
\begin{equation}
J\pi(\N) J=\pi(\N)^\comm,\;\;
\xi\in\stdcone\;\limp\;J\xi=\xi,\;\;
\pi(x)J\pi(x)\stdcone\subseteq\stdcone,\;\;
y\in\zentr(\pi(\N))\;\limp\;JyJ=y^*.
\end{equation}
For any standard representation
\begin{equation}
\forall\phi\in\N^+_\star\;\;\exists!\stdembed_\pi(\phi)\in\stdcone\;\;\forall x\in\N\;\;\phi(x)=\s{\stdembed_\pi(\phi),\pi(x)\stdembed_\pi(\phi)}_\H.
\label{std.vector.representative}
\end{equation}
The map $\stdembed_\pi:\N^+_\star\ra\stdcone$ is order preserving. Moreover, $\stdembed^\natural_\pi:\stdcone\ra\N_\star^+$, defined by $(\stdembed^\natural_\pi(\xi))(x)=\s{\xi,\pi(x)\xi}_\H$ $\forall x\in\N$, is a bijective norm continuous homomorphism with $(\stdembed^\natural_\pi)^{-1}=\stdembed_\pi$. For any two standard representations $(\H_1,\pi_1,J_1,\H^\natural_1)$ and $(\H_2,\pi_2,J_2,\H^\natural_2)$ of a W$^*$-algebra $\N$ and a given $^*$-isomorphism $\varsigma:\pi_1(\N)\ra\pi_2(\N)$, there exists a unique unitary $u_\varsigma:\H_1\ra\H_2$ such that $\varsigma(x)=u_\varsigma xu_\varsigma^*$ $\forall x\in\pi_1(\N)$, $J_2=u_\varsigma J_1u_\varsigma^*$, $\H^\natural_2=u_\varsigma\H^\natural_1$. Such $u_\varsigma$ will be called a standard unitary equivalence. If $\phi\in\N^+_{\star0}$ then, by means of the Tomita--Takesaki theory \cite{Tomita:1967:b,Takesaki:1970}, the GNS representation associated with $\phi$ determines a unique conjugation $J_\phi$, and a weakly-$\star$ continuous group homomorphism $\sigma^\omega:\RR\ra\Aut(\N)$. An associated $\stdcone$ is given by \cite{Connes:1974,Araki:1974:modular:conjugation}
\begin{equation}
\H^\natural_\phi:=\overline{\bigcup_{x\in\nnn_\phi\cap\nnn_\phi^*}\{\pi_\phi(x)J_\phi\pi_\phi(x)J_\phi\Omega_\phi\}}^{\H_\phi}.
\label{Connes.Araki.natural.cone}
\end{equation}
If $\N\iso\BBB(\K)$ for some Hilbert space $\K$ and $\phi\in\W_0(\N)$ is given by $\tr_\K$, then the corresponding GNS Hilbert space $\H_\phi$ is given by the space of all Hilbert--Schmidt operators,
\begin{equation}
\schatten_2(\K):=\{x\in\BBB(\K)\mid(\tr_\K(x^*x))^{1/2}<\infty\}=\nnn_{\tr_\K},
\end{equation}
equipped with a scalar product $\s{x,y}_{\schatten_2(\K)}:=\tr_\K(x^*y)$, so that $\schatten_2(\K)\iso\K\otimes\K^\banach$ as Hilbert spaces. Moreover, $\H^\natural_\phi=\schatten_2(\K)^+$, $\pi_\phi(x)=\LLL_x$ (which denotes left multiplication by $x$), while $\stdembed_\pi:\schatten_1(\K)^+\ni\rho\mapsto\rho^{1/2}\in\schatten_2(\K)^+$.
\
For a given W$^*$-algebra $\N$, $\phi\in\W(\N)$, and $\omega\in\W_0(\N)$ the map
\begin{equation}
R_{\phi,\omega}:[x]_\omega\mapsto[x^*]_\phi\;\;\forall x\in\nnn_\omega\cap\nnn_\phi^*
\label{relative.modular.weights}
\end{equation}
is a densely defined, closable antilinear operator. Its closure admits a unique polar decomposition
\begin{equation}
\overline{R}_{\phi,\omega}=J_{\phi,\omega}\Delta^{1/2}_{\phi,\omega},
\end{equation}
where $\rpktarget{JREL}J_{\phi,\omega}$ is a conjugation operator, called \df{relative modular conjugation}, while $\rpktarget{DELTAREL}\Delta_{\phi,\omega}$ is a positive self-adjoint operator on $\dom(\Delta_{\phi,\omega})\subseteq\H_\omega$ with $\supp(\Delta_{\phi,\omega})=\supp(\phi)\H_\omega$, called a \df{relative modular operator} \cite{Araki:1973:relative:hamiltonian,Connes:1974,Digernes:1975}. Given $\omega\in\W_0(\N)$, $\Delta_{\omega,\omega}=:\Delta_\omega$ implements the action of the Tomita--Takesaki (modular) automorphism $\sigma^\omega$ by
\begin{equation}
\pi_\omega(\sigma^\omega_t(x))=\Delta_\omega^{\ii t}\pi_\omega(x)\Delta_\omega^{-\ii t}.
\end{equation}
If $\N\iso\BH$, $\phi=\tr_\H(\rho_\phi\,\cdot)$, $\omega=\tr_\H(\rho_\omega\,\cdot)$, and $\RRR_x$ denotes right multiplication by $x\in\BH$, then $\Delta_{\phi,\omega}=\LLL_{\rho_\phi}\RRR_{\rho_\omega^{-1}}$.
For every $\phi,\omega\in\W_0(\N)$ the relative modular conjugation $J_{\phi,\omega}$ determines a unique unitary operator $\rpktarget{STDUNITRANS}J_{\phi,\phi}J_{\phi,\omega}=:V_{\phi,\omega}:\H_\omega\ra\H_\phi$, such that
\begin{align}
\pi_\phi(x)&=V_{\phi,\omega}\pi_\omega(x)V_{\phi,\omega}^*,\label{std.uni.trans.eq.one}\\
V_{\phi,\omega}(\stdcone_\omega)&=\stdcone_\phi,\label{std.uni.trans.eq.two}\\
V_{\phi,\omega}J_{\omega,\omega}&=J_{\phi,\phi}V_{\phi,\omega}.\label{std.uni.trans.eq.three}
\end{align}
We will call $V_{\phi,\omega}$ \df{standard unitary transition} between $\H_\omega$ and $\H_\phi$. It is a standard unitary equivalence of a $^*$-isomorphism $\varsigma_{\phi,\omega}:\pi_\omega(\N)\ra\pi_\phi(\N)$ determined by the condition $\varsigma_{\phi,\omega}\circ\pi_\omega=\pi_\phi$. Thus, if $\phi,\omega\in\N^+_{\star0}$, then $V_{\phi,\omega}$ provides a default unitary mapping between the corresponding GNS Hilbert spaces and representations.
Given any group
$G$, a \df{representation} of $G$ in the group $\Aut(\N)$ of $*$-automorphisms of a W$^*$-algebra $\N$ is a map $\alpha:G\ni g\mapsto\alpha(g)=:\alpha_g\in\Aut(\N)$ which is a \df{group homomorphism}, that is,
\begin{enumerate}
\item[1)] $\alpha(e)=\id_\N$,
\item[2)] $\alpha(g_1)\circ\alpha(g_2)=\alpha(g_1\circ g_2)$ $\forall g_1,g_2\in G$,
\end{enumerate}
where $e$ denotes the neutral element of $G$. A group $G$ is called: \df{topological} if{}f it is also a topological space and a map $G\times G\ni(g_1,g_2)\mapsto g_1\circ g_2^{-1}\in G$ is continuous for all $g_1,g_2\in G$; \df{locally compact} if{}f it is topological and $e\in G$ has a compact
topological neighbourhood. For any W$^*$-algebra $\N$, $\Aut(\N)$ is a topological group with respect to \df{weak-$\star$ topology} on $\Aut(\N)$, defined by the collection of neighbourhoods \cite{Takesaki:1983}
\begin{equation}
N_{\{\omega_i\}}(\alpha):=
\{\varsigma\in\Aut(\N)\mid
\n{\omega_i\circ\alpha-\omega_i\circ\varsigma}_{\N_\star}<1,\;
\n{\omega_i\circ\alpha^{-1}-\omega_i\circ\varsigma^{-1}}_{\N_\star}<1\},
\end{equation}
where $\{\omega_i\}\subseteq\N_\star$, $i\in\{1,\ldots,n\}$, $n\in\NN$. A triple $(\N,G,\alpha)$ of a W$^*$-algebra, locally compact group $G$, and a representation $\alpha:G\ra\Aut(\N)$ is called a \df{W$^*$-dynamical system} (or a \df{W$^*$-covariant system}) if{}f $\alpha$ is continuous in the weak-$\star$ topology of $\Aut(\N)$. This condition is equivalent to the continuity of the map $G\ni g\mapsto\alpha_g(x)\in\N$ in the weak-$\star$ topology of $\N$ for any $x\in\N$, that is, to
\begin{equation}
G\ni g\mapsto\phi(\alpha_g(x))\in\CC\mbox{ is a continuous function}\;\;\forall x\in\N,
\label{weak.star.continuous.group}
\end{equation}
and such $\alpha$ is called a \df{weakly-$\star$ continuous} representation. Uniqueness of a predual of a W$^*$-algebra $\N$ allows to define isometries $\alpha_\star$ of $\N_\star$ that uniquely correspond to the elements $\alpha\in\Aut(\N)$, and to define the isometries of $\N_\star$ uniquely corresponding to representations $\alpha:G\ra\Aut(\N)$:
\begin{equation}
\duality{(\alpha_g)_\star(\phi),x}_{\N_\star\times\N}=\duality{\phi,\alpha_g(x)}_{\N_\star\times\N}=\phi(\alpha_g(x))\;\;\forall x\in\N\;\forall\phi\in\N_\star.
\end{equation}
The above equivalence can be shown (see e.g. \cite{Sakai:1991}) by proving that \eqref{weak.star.continuous.group} implies continuity of $\alpha_\star$ in the norm of $\N_\star$,
\begin{equation}
\lim_{g\ra e}\n{(\alpha_g)_\star(\phi)-\phi}_{\N_\star}=0\;\;\forall\phi\in\N_\star\;\forall g\in G.
\end{equation}
A \df{unitary implementation} of a representation $\alpha:G\ra\Aut(\N)$ in a given representation $\pi:\N\ra\BH$ is defined as a map $u:G\ni g\mapsto u(g)\in\BH^\uni$ that determines a family $\{u(g)\mid g\in G\}$ of unitary operators satisfying the \df{covariance equation}
\begin{equation}
\pi(\alpha_g(x))=u(g)\pi(x)u(g)^*\;\;\forall x\in\N\;\forall g\in G.
\label{covariance.equation.group.G}
\end{equation}
The condition \eqref{covariance.equation.group.G} alone does not determine $\{u(g)\mid g\in G\}$ uniquely. The setting of W$^*$-algebras admits a remarkable solution to this problem: every pair of a W$^*$-dynamical system $(\N,\RR,\alpha)$ and a standard representation $(\H,\pi,J,\stdcone)$ determines uniquely a corresponding unitary implementation \textit{together with} a unique self-adjoint generator of this family of unitaries. This generator is called a \textit{standard liouvillean}\footnote{It would be however more precise to call it \textit{quantum koopmanian}, because in the commutative setting (of statistical mechanics and probability measures) the `liouvillean operator' (defined by the Poisson bracket) acts on elements of $L_1(\X,\mho(\X),\tmu)$, while it is the `koopmanian operator' \cite{Koopman:1931,vonNeumann:1932:zur,vonNeumann:1932:zusaetze} that acts on the positive cone of $L_1(\X,\mho(\X),\tmu)$.}. It is not called `hamiltonian', because in general its spectrum may be not bounded from any side, while the notion of `hamiltonian' is usually understood as referring to a self-adjoint operator that generates a strongly continuous group of unitary operators \textit{and} has a nonnegative (or at least bounded from below) spectrum\footnote{E.g., \cytat{one of the most important principles of quantum field theory, ensuring the stability, demands that the energy should have a lower bound} \cite{Haag:1992}.}. Moreover, as opposed to hamiltonian, the construction of standard liouvillean for a given W$^*$-dynamical system does not require any additional analytic conditions that constrain derivation to an `integrable' infinitesimal generator. This way the W$^*$-algebraic approach makes the notion of a hamiltonian less relevant than the notion of a liouvillean.
For any W$^*$-algebra $\N$, the unique predualisation of action of $\alpha\in\Aut(\N)$ can be connected with the uniqueness property of representation of elements of $\N_\star^+$ in terms of a standard cone of a standard representation $(\H,\pi,J,\stdcone)$ of $\N$: any $\alpha\in\Aut(\N)$ defines a unique map $u:\stdcone\ra\stdcone$ by
\begin{equation}
u\stdembed_\pi(\phi):=\stdembed_\pi(\alpha_\star(\phi))\;\;\;\forall\phi\in\N_\star^+.
\end{equation}
This map is linear, can be extended to a unitary operator on all $\H$, and satisfies
\begin{equation}
u\pi(x)u^*=\pi(\alpha(x))\;\;\forall x\in\N.
\end{equation}
This leads to a question, whether it is possible to generate this way a \textit{standard} unitary implementation of a given representation $\alpha:G\ra\Aut(\N)$. The answer is in the affirmative, and was established by Haagerup \cite{Haagerup:1975:standard:form} (the special cases of this result were obtained earlier in \cite{Kadison:1965,Kallman:1971,Henle:1970,Halpern:1972,Pedersen:Takesaki:1973}). If $(\H,\pi,J,\stdcone)$ is a standard representation of a W$^*$-algebra $\N$, then there exists a unique strongly continuous unitary implementation $\rpktarget{V.ALPHA}V_\alpha(g)$ of $\alpha$ satisfying
\begin{align}
V_\alpha(g)\stdcone&=\stdcone,\\
JV_\alpha(g)&=V_\alpha(g)J.
\end{align}
Such family $\{V_\alpha(g)\mid g\in G\}$ is called a \df{standard} unitary implementation of $\alpha$.
Thus, if $(\N,\RR,\alpha)$ is a W$^*$-dynamical system with $\N$ in standard form $(\H,\pi(\N),J,\stdcone)$, then from the theorems of Haagerup and Stone \cite{Stone:1930,Stone:1932,vonNeumann:1932:Stone} it follows that there exists a unique strongly continuous group of unitaries $\{V_\alpha(t)\mid t\in\RR\}\subseteq\BH^\uni$, and a unique self-adjoint operator $\rpktarget{K.ALPHA}K^\alpha$ on $\H$, called \df{standard liouvillean}, such that $V_\alpha(t)$ is a strongly continuous unitary implementation of $\alpha$ and for every $t\in\RR$
\begin{enumerate}
\item[i)] $V_\alpha(t)=\ee^{-\ii tK^\alpha}$,
\item[ii)] $\ee^{-\ii tK^\alpha}\stdcone=\stdcone$,
\item[iii)] $JK^\alpha+K^\alpha J=0$.
\end{enumerate}
The definition of a standard liouvillean $K^\alpha$ does not depend on any choice of $\omega\in\N^+_\star$ or $\omega\in\W(\N)$: it depends only on a W$^*$-dynamical system and a standard representation of W$^*$-algebra. If $\N$ is semi-finite, $(\H,\pi,J,\stdcone)$ is its standard representation, $(\N,\RR,\alpha)$ is a W$^*$-dynamical system, $H\in\pi(\N)^\sa$, and $\{U(t):=\ee^{-\ii t H}\in\pi(\N)\mid t\in\RR\}$ is a strongly continuous group of unitary operators such that
\begin{equation}
\ee^{\ii tH}\pi(x)\ee^{-\ii tH}=\pi(\alpha_t(x))\;\;\forall x\in\N\;\forall t\in\RR,
\end{equation}
then the standard liouvillean reads \cite{Bratteli:Robinson:1979,JOPP:2012}
\begin{equation}
K^\alpha=H-JHJ=[H,\,\cdot\,].
\label{std.liouvillean.for.BH}
\end{equation}
\else
\fi
Following Odzijewicz and Sli\.zewska \cite{Odzijewicz:Slizewska:2011}, consider a bundle $\eee:V\ra\N_\star^+$, where
\begin{equation}
V:=\{(x,\omega)\in\N\times\N_\star^+\mid x\,\supp(\omega)=x\},
\end{equation}
and the bundle projection $\eee$ is given by a restriction of the cartesian product projection $\N\times\N_\star^+\ra\N_\star^+$ to $V$. Because $V_\omega:=\eee^{-1}(\omega)=\N\supp(\omega)$ $\forall\omega\in\N_\star^+$, the scalar product
\begin{equation}
V_\omega\times V_\omega\ni(x,y)\mapsto\s{x,y}_\omega:=\duality{\omega,x^*y}_{\N_\star\times\N}\in\CC
\end{equation}
is nondegenerate. Moreover, $\s{x,x}_\omega=0$ $\iff$ $x\in\N(\II-\supp(\omega))$. The completion $\bar{V}_\omega$ of $V_\omega$ under the norm generated by $\s{\cdot,\cdot}_\omega$ determines a bundle $\rpktarget{H.MN}\H\N_\star^+:=\bar{V}\ra\N_\star^+$ of Hilbert spaces, which the authors of \cite{Odzijewicz:Slizewska:2011} call the \df{Gel'fand--Na\u{\i}mark--Segal bundle} (the notion of the GNS bundle was earlier alluded in \cite{Jadczyk:1988,Chruscinski:Marmo:2009})
While \eqref{Connes.Araki.natural.cone} secures that the GNS representation is a standard representation whenever $\phi\in\N^+_{\star0}$, it doesn't have to be in more general case. Thus, in order to be sure that our use of GNS bundle coincides with the necessary conditions for Haagerup's theorem, we will restrict our discussion in multiple places of this paper to subsets and submanifolds of $\N^+_{\star0}$. We consider this restriction to be nonoptimal, but in order to work it out in larger generality, we would have to work with a different bundle of Hilbert spaces. Restriction to $\N^+_{\star0}$ allows us to use standard unitary transitions $V_{\phi,\omega}$ to map between Hilbert spaces, at the expense of consideration of unitarily equivalent representations only. Whenever the assumption of restriction to $\N^+_{\star0}$ is made, it implies restriction of considerations to countably finite W$^*$-algebras, because only for them $\N^+_{\star0}\neq\varnothing$.
\ifvargaugecompile
\subsection{Case study: Algebraic hamiltonian vector fields\label{algebraic.hamiltonian}
\else
\subsubsection{Algebraic hamiltonian vector fields\label{algebraic.hamiltonian}
\fi
The BLP structure of $\N_\star$ and $\N_\star^\sa$ allows to introduce and analyse the temporal evolution on $\N_\star^+$ by means of the hamiltonian vector field and the Hamilton equation. On the other hand, %
\ifvargaugecompile
\else
as discussed in Sections \ref{automorphisms.section} and \ref{standard.liouvilleans.section},
\fi
for any W$^*$-dynamical system $(\N,\RR,\alpha)$ one can predualise the representation $\alpha:
\RR\ni t\mapsto\alpha_t\in\Aut(\N)$ obtaining the family $\{\alpha_\star^t\mid t\in\RR\}$ of norm continuous isometries $\alpha_\star^t:=(\alpha_t)_\star:\N_\star\ra\N_\star$, which in turn can be analysed by means of a unique self-adjoint standard liouvillean operator $K^\alpha$ that generates a unitary evolution in $L_2(\N)$ leaving $L_2(\N)^+$ invariant. The virtue of a geometric description in terms of hamiltonian vector field is that it allows for an analysis of the local differential structure of temporal evolution in terms of a local Poisson flow and tangent space. However, it does not guarantee the existence of global flow. On the other hand, an algebraic description in terms of a predualised representation $\alpha_\star$ and an associated standard liouvillean $K^\alpha$ guarantees the existence of a global flow on $\N_\star^+$, but it is not necessarily a Poisson flow and it is not related to a tangent space, thus it does not allow (in general) for a refined smooth geometric description. This leads us to single out the class of evolutions on $\N_\star^\sa$ (and $\N_\star^+$) that satisfy both conditions.
The isometries $\alpha_\star^t$ of $\N_\star$ that are also the Poisson flows leaving $\N_\star^\sa$ invariant are characterised as solutions of the equation \eqref{BLP.five}
\begin{equation}
\{f\circ\alpha_\star^t,k\circ\alpha_\star^t\}_{\N_\star^\sa}=\{f,k\}_{\N_\star^\sa}\circ\alpha_\star^t\;\;\forall f,k\in\CIF(\N_\star^\sa;\RR)\;\;\forall t\in\RR,
\end{equation}
which gives, by \eqref{BLP.on.N.sa},
\begin{align}
\phi([\DF_\phi(f\circ\alpha_\star^t),\DF_\phi(k\circ\alpha_\star^t)])=(\alpha_\star^t(\phi))([\DF_{\alpha_\star^t(\phi)},\DF_{\alpha^t_\star(\phi)}k]),\\
0=\phi([\DF_\phi(f\circ\alpha_\star^t),\DF_\phi(k\circ\alpha_\star^t)]-\alpha_t([\DF_{\alpha_\star^t(\phi)}f,\DF_{\alpha_\star^t(\phi)}k])).
\end{align}
Hence, the predualisation $\alpha_\star$ of a weakly-$\star$ continuous representation $\alpha:\RR\ra\Aut(\N)$ is a Poisson flow on $(\N_\star^\sa,\{\cdot,\cdot\})$ if{}f $\alpha$ satisfies
\begin{equation}
\phi((\id_\N-\alpha_t)([\DF_{\alpha_\star^t(\phi)}f,\DF_{\alpha_\star^t(\phi)}k]))=0\;\;\;\forall\phi\in\N_\star^\sa\;\;\forall f,k\in\CIF(\N_\star^\sa;\RR)\;\;\forall t\in\RR.
\label{poisson.flow.compatibility}
\end{equation}
We will call \eqref{poisson.flow.compatibility} the \df{Poisson compatibility condition} (PC$_1$). Let $(\N,\RR,\alpha)$ be a W$^*$-dynamical system satisfying the Poisson compatibility condition. Then the Poisson flow $\alpha_\star|_{\N_\star^\sa}$ is generated by the Hamilton function $\rpktarget{HA.ALPH}h^\alpha\in\CIF(\N_\star^\sa;\RR)$ according to \eqref{hamilton.eqn},
\begin{equation}
\frac{\dd}{\dd t}f_t=\{h^\alpha,f_t\},\;\;f_t(x):=f(\alpha^t_\star(x))\;\;\forall t\in\RR\;\forall x\in\N_\star\;\forall f\in\CIF(\N_\star^\sa;\RR),
\end{equation}
which determines the corresponding unique hamiltonian vector field $\XXX_{h^\alpha}\in\T\N_\star^\sa$ by means of the BOR equation \eqref{BLP.seven},
\begin{equation}
\XXX_{h^\alpha}(\phi)=-\ad^\banach_{\DF_\phi h^\alpha}(\phi)=\LLL^\banach_{\DF_\phi h^\alpha}(\phi)-\RRR^\banach_{\DF_\phi h^\alpha}(\phi)\;\;\forall\phi\in\N_\star^\sa.
\end{equation}
We will call $\XXX_{h^\alpha}$ an \df{algebraic hamiltonian vector field}.
Now, let us recall from Section \ref{standard.liouvilleans.section} that, by the Haagerup theorem, each pair of a W$^*$-dynamical system $(\N,\RR,\alpha)$ and a standard representation $(\H,\pi,J,\stdcone)$ of $\N$ determines a unique self-adjoint generator $K^\alpha$ of a unitary implementation of $\alpha$ in $\BH$ satisfying the condition $\ee^{-\ii tK^\alpha}\stdcone\subseteq\stdcone$. By means of \eqref{std.vector.representative}, this condition expresses the requirement that $\alpha_\star(\N_\star^+)\subseteq\N_\star^+$ (if formulated in terms of Kosaki's canonical representation, for which $\stdcone= L_2(\N)^+$, this condition is just an $L_2(\N)$ version of $L_1(\N)^+=\N_\star^+$ invariance under $\alpha$). If $\omega\in\N^+_{\star0}$, then the GNS representation, $(\H_\omega,\pi_\omega,\Omega_\omega)$ is also a standard representation $(\H_\omega,\pi_\omega,J_\omega,\H^+_\omega)$, so we can apply Haagerup's theorem to the fibres of the GNS bundle $\H\N_\star^+$ restricted to the submanifold $\N^+_{\star0}$, $\eee:\H\N^+_{\star0}\ra\N^+_{\star0}$. Because $\supp(\omega)=\II$ for each $\omega\in\N^+_{\star0}$, the bundle projection $\eee$ reduces in this case to a cartesian product projection. Orbits of any Poisson flow leave $\N^+_{\star0}\subseteq\N_\star^\sa$ invariant \cite{Bona:2000,Beltita:Ratiu:2005,Ratiu:2011}, while $\alpha_\star^t$ is norm preserving, so the restrictions of Poisson compatible isometries $\alpha^t_\star$ to $\N^+_{\star0}$ are automorphisms of this space. As a result, we obtain a remarkable geometric correspondence: every weakly-$\star$ continuous representation $\alpha:\RR\ni t\mapsto\alpha_t\in\Aut(\N)$ satisfying the Poisson compatibility condition \eqref{poisson.flow.compatibility} determines a unique globally integrable hamiltonian vector field $\XXX_{h^\alpha}\in\T\N_\star^\sa$ \textit{and} a family of standard liouvillean operators $\rpktarget{K.OMEGA.ALPHA}\N^+_{\star0}\ni\omega\mapsto K_\omega^\alpha\in(\Lin(\H_\omega))^\sa$ acting pointwise on the GNS bundle of Hilbert spaces. In other words, the family $\{\alpha_t\in\Aut(\N)\mid t\in\RR\}$ of Poisson compatible, weakly-$\star$ continuous automorphisms of a W$^*$-algebra $\N$ is uniquely represented \textit{in} the tangent vector bundle $\T\N_\star^\sa\ra\N_\star^\sa$ (and, by linearity, also in $\T\N_\star^+\ra\N_\star^+$), as well as \textit{on} the GNS Hilbert bundle $\H\N^+_{\star0}\ra\N^+_{\star0}$.
Due to uniqueness property \eqref{std.vector.representative} of the embedding $\stdembed_\pi:\N_\star^+\ni\omega\mapsto\stdembed_\pi(\omega)\in\stdcone$ of any standard representation $(\H,\pi,J,\stdcone)$, this means that the embedding of a trajectory generated by $\alpha_\star^t$ on $\N^+_{\star0}$ to $\H_\phi$ for any $\phi\in\N^+_{\star0}$ coincides with the evolution in $\H_\phi^+$ generated by $\ee^{-\ii tK_\phi^\alpha}$. Hence, the hamiltonian flow $w_t^{h^\alpha}$ of $\XXX_{h^\alpha}$ on $\N^+_{\star0}$ from $\phi(0)$ to $\phi(t)$ can be always represented as liouvillean evolution
\begin{equation}
\ee^{-\ii tK_{\phi(t)}^\alpha}\stdembed_{\phi(t)}(\phi(0))=\stdembed_{\phi(t)}(\phi(t))=:\Omega_{\phi(t)},
\label{liovillean.movement}
\end{equation}
where by $\stdembed_\psi(\omega)$ we denote the standard representative of $\omega\in\N^+_\star$ in the positive cone $\H^+_\psi$ of the GNS representation Hilbert space of $\psi\in\N^+_{\star0}$.
If $\N=\BH$, $\rho\in\schatten_1(\H)^+_0$, and the weakly-$\star$ continuous representation $\alpha:\RR\ra\Aut(\BH)$ is unitary (that is, $\alpha_t=\Ad(u(t))$ with $u(t)\in\BH^\uni$ $\forall t\in\RR$), then the algebraic hamiltonian vector field can be expressed by the von Neumann equation \eqref{vN.equation.from.BLP}, while the corresponding liouvillean evolution $\rho(t)=\ee^{-\ii tK^\alpha}\rho(0)$ in $\schatten_2(\H)\iso\H\otimes\H^\banach$ is a solution of the equation
\begin{equation}
\frac{\dd}{\dd t}\rho(t)=-\ii K^\alpha\rho(t),
\end{equation}
which gives \eqref{std.liouvillean.for.BH}%
\ifvargaugecompile
.
\else
, in agreement with \eqref{hamiltonian.liouvillean.J.eqn}.
\fi
In this sense, the Poisson compatibility condition \eqref{poisson.flow.compatibility} extends the equivalence between the algebraic (liouvillean operator) and geometric (hamiltonian vector) descriptions of temporal evolution of quantum states to the general W$^*$-dynamical systems. In the next Section we will investigate how standard liouvilleans can be used to encode the perturbation of Poisson flow by additional geometric structures over state space, beyond the realms of W$^*$-dynamical systems.
\subsection{Relative entropy, Norden--Sen geometry, and noncommutative Orlicz spaces\label{relat.entr.MCP.bundle}}
\subsubsection{Distances, Norden--Sen geometries, and geodesic free falls as entropic projections\label{distances.NS.geom.section}}
A pair $(\nabla,\nabla^\nsdual)$ of two affine connections over a smooth manifold $\M$ will be called \df{Norden--Sen dual} with respect to a riemannian metric $\gbold$ on $\M$, if{}f \cite{Sen:1944,Sen:1945,Sen:1946,Norden:1945,Norden:1948,Norden:1949}\rpktarget{NORDEN.SEN}
\begin{equation}
\gbold(\nabla_uv,w)+\gbold(v,\nabla^\nsdual_uw)=u(\gbold(v,w))\;\forall u,v,w\in\T\M,
\label{geometric.duality}
\end{equation}
which is equivalent to
\begin{equation}
\gbold(\transport_c^\nabla u,\transport_c^{\nabla^\nsdual}v)=\gbold(u,v)
\label{geometric.duality.two}
\end{equation}
for all $u,v\in\T\M$ and for all curves $c:\RR\supset[r_1,r_2]\ra\M$ (the symbol $\transport_c^\nabla$ denotes a parallel transport along $c$ that is determined by an affine connection $\nabla$). The quadruple $(\M,\gbold,\nabla,\nabla^\nsdual)$ is called a \df{Norden--Sen geometry}. A riemannian geometry is characterised as a Norden--Sen geometry with $\nabla=\nabla^\nsdual$.
Given a set $M$, we define a \df{distance} on $M$ as a function $D:M\times M\ra[0,\infty]$ such that $D(\phi,\omega)=0$ $\iff$ $\omega=\phi$. Eguchi \cite{Eguchi:1983,Eguchi:1985,Eguchi:1992} showed that for any smooth manifold $\M$ and any smooth distance $D$ on $\M$ that satisfies
\begin{equation}
\DG_v|_p\DG_v|_pD(p,q)|_{q=p}\in\,]0,\infty[\;\;\;\forall p\in\M\;\forall v\in\T_p\M\setminus\{0\},
\label{Eguchi.condition}
\end{equation}
where $\DG_v|_p$ denotes here the G\^{a}teaux derivative at $p\in\M$ in the direction $v\in\T_p\M$, the distance $D$ determines a riemannian metric $\gbold$ and a pair of affine connections $(\nabla,\nabla^\nsdual)$ on $\M$, given by the \df{Eguchi equations}
\begin{align}
\gbold_\phi(u,v)&:=-\DG_u|_\phi\DG_v|_\omega D(\phi,\omega)|_{\omega=\phi},\label{Eguchi.metric}\\
\gbold_\phi((\nabla_u)_\phi v,w)&:=-\DG_u|_\phi\DG_v|_\phi\DG_w|_\omega D(\phi,\omega)|_{\omega=\phi},\label{Eguchi.connection}\\
\gbold_\phi(v,(\nabla_u^\nsdual)_\phi w)&:=-\DG_u|_\omega\DG_w|_\omega\DG_v|_\phi D(\phi,\omega)|_{\omega=\phi}\label{Eguchi.dualconnection}.
\end{align}
Every quadruple $(\M,\gbold,\nabla,\nabla^\nsdual)$ determined in this way is a Norden--Sen geometry such that both $\nabla$ and $\nabla^\nsdual$ are torsion-free. A torsion-free Norden--Sen geometry will be called an \df{Eguchi geometry}. While in riemannian geometry the affine connection is determined by the riemannian metric, in the Eguchi geometry the triple of riemannian metric and two Norden--Sen dual affine connections are determined by the distance. The Levi-Civita connection $\bar{\nabla}$ of an associated riemannian geometry $(\M,\gbold)$ satisfies $\bar{\nabla}=(\nabla+\nabla^\nsdual)/2$. In this sense, the Eguchi geometry (based on the nonsymmetric distance) provides a generalisation of a riemannian geometry and cartesian geometry, including all of their main notions: distance, length, parallelity and orthogonality. Generalisation of the cartesian distance is provided by the distance $D$, the induced riemannian metric $\gbold$ provides the generalisation of orthogonality and length, while the induced torsion-free Norden--Sen dual connections $(\nabla,\nabla^\nsdual)$ provide a generalisation of parallelity.\footnote{The idea that $D$ should be considered as generalisation of the cartesian distance, while the connection $\nabla$ associated to a projection by means of $D$ should be considered as a proper generalisation of parallelity (at least in the setting of statistical manifolds) is due to Chencov \cite{Chencov:1964,Chencov:1968}.} The invariance of length under parallel transport that characterises riemannian geometry is weakened to covariance in the sense of \eqref{geometric.duality.two}.
If both affine connections of a Norden--Sen geometry are flat and torsion-free, then it is called a \df{dually flat geometry} \cite{Nagaoka:Amari:1982,Amari:1985,Amari:Nagaoka:1993}. If $\dim\M=:n<\infty$, then every dually flat geometry $(\M,\gbold,\nabla,\nabla^\nsdual)$ determines a unique pair of affine immersions $\Psi:\M\ra\RR$ and $\Psi^\lfdual:\M\ra\RR$ such that
\begin{align}
\gbold_{ij}(\rho(\theta))&=\frac{\partial^2\Psi(\rho(\theta))}{\partial\theta^i\partial\theta^j},\\
\gbold_{ij}(\rho(\eta))&=\frac{\partial^2\Psi^\lfdual(\rho(\eta))}{\partial\eta^i\partial\eta^j},
\end{align}
where $\{\theta^i\}$ is a coordinate system such that $\Gamma^\nabla_{ijk}(\rho(\theta))=0$ $\forall\rho\in\M$ and $\Gamma^{\nabla^\nsdual}_{ijk}(\rho(\eta))=0$ $\forall\rho\in\M$ \cite{Dillen:Nomizu:Vracken:1990,Kurose:1990,Kurose:1993,Matsuzuoe:1998}. Conversely \cite{Amari:1985}, if there exists a convex function $\Psi$ such that its hessian (matrix of second derivatives) determines pointwise a riemannian metric, then there exists a pair of coordinate systems $\{\theta^i\}$ and $\{\eta^i\}$ and a convex function $\Psi^\lfdual:\M\ra\RR$ satisfying the above properties. The dual flatness of a pair $(\theta,\eta)$ of coordinate systems is equivalent to the orthogonality of their tangent vectors at $q$ with respect to the riemannian metric $\gbold$ at $q$,
\begin{equation}
\gbold_q\left((\T_q\theta)^{-1}\left(\frac{\partial}{\partial\theta^i}\right),(\T_q\eta)^{-1}\left(\frac{\partial}{\partial\eta_j}\right)\right)=\dirac_i^j\;\;\;\forall q\in\M.
\label{dual.metric.flatness}
\end{equation}
The transition between these two formulations in the real finite dimensional case is provided by means of bijective \df{Legendre transformation} $\rpktarget{LFTRAFO}\LFtrafo_\Psi:\Theta\ra\Xi$, which acts between suitable open subsets $\Theta\subset\RR^n$ and $\Xi\subset\RR^n$, and is given by the gradient,
\begin{equation}
\LFtrafo_\Psi:\Theta\ni\theta\mapsto\eta:=\grad\Psi(\theta)\in\Xi.
\label{Legendre.gradient}
\end{equation}
In the coordinate-dependent form this reads
\begin{align}
\eta_i&=(\LFtrafo_\Psi(\theta))_i:=\frac{\partial\Psi(\theta)}{\partial\theta^i},\label{LFgrad.one}\\
\theta^i&=(\LFtrafo^{-1}_\Psi(\eta))^i:=\frac{\partial\Psi^\lfdual(\eta)}{\partial\eta_i},\label{LFgrad.two}
\end{align}
whenever the duality pairing is given by
\begin{equation}
\duality{\cdot,\cdot}_{\RR^n\times\RR^n}:\RR^n\times\RR^n
\ni(\theta,\eta)
\mapsto
\theta\cdot\eta^\top:=\sum_{i=1}^n\theta^i\eta_i
\in\RR.
\label{selfduality.Rn}
\end{equation}
The Eguchi equations applied to the distance
\begin{equation}
D_\Psi(\rho,\sigma):=\Psi(\rho)+\Psi^\lfdual(\sigma)-\sum_{i=1}^n\theta^i(\rho)\eta^i(\sigma)
\end{equation}
yield $(\M,\gbold^{D_\Psi},\nabla^\Psi,{\nabla^{\Psi}}^\nsdual)=(\M,\gbold,\nabla,\nabla^\nsdual)$. We will call such $D_\Psi$ a canonical Br\`{e}gman distance of a dually flat geometry $(\M,\gbold,\nabla,\nabla^\nsdual)$. A riemannian metric $\gbold$ on an affine manifold $(\M,\nabla)$ with flat $\nabla$ is said to be \df{hessian}, and denoted $\gbold^\Psi$, if{}f there exists a smooth function $\Psi:\M\ra\RR$ such that \cite{Shima:1976,Shima:1977,Cheng:Yau:1982}
\begin{equation}
\gbold(u,v)=(\nabla_u\ddd\Psi)(v)\;\;\forall u,v\in\T\M.
\label{hessian.metric}
\end{equation}
Such triple $(\M,\gbold,\nabla)$ will be called a \df{hessian geometry} \cite{Shima:2007} (see also \cite{Koszul:1961,Vinberg:1963,Duistermaat:2001}). The function $\Psi$ in \eqref{hessian.metric} is the same as in the representation of $\gbold_{ij}(\rho(\theta))$ above, and
\begin{equation}
\gbold(u,v)=(\nabla^\nsdual_u\ddd\Psi^\lfdual)(v)\;\;\forall u,v\in\T\M.
\end{equation}
Hence, given a riemannian manifold $(\M,\gbold)$ and an affine connection $\nabla$ on $\M$ the following conditions are equivalent \cite{Shima:1986,Shima:Yagi:1997,Shima:2007}: (1) $(\M,\gbold,\nabla)$ is a hessian geometry; (2) $(\M,\gbold,\nabla,2\bar{\nabla}-\nabla)$ is a dually flat geometry.
Let $(\M,\gbold,\nabla,\nabla^\nsdual)$ be a dually flat geometry, and let $\Q\subseteq\M$ be $\nabla^\nsdual$-affine (i.e., there exists a coordinate system $\{\eta^i\}$ on $\Q$ such that $\Gamma^{\nabla^\nsdual}_{ijk}(\rho(\eta))=0$ $\forall\rho\in\Q$) and $\nabla^\nsdual$-convex (i.e. $\forall\rho_1,\rho_2\in\Q$ $\exists!$ $\nabla^\nsdual$-geodesics in $\Q$ connecting them). Then there exists a unique entropic projection
\begin{equation}
\M\in\rho\mapsto\PPP^{D_\Psi}_\Q(\rho):=\arginff{\sigma\in\Q}{D_\Psi(\sigma,\rho)}\in\Q,
\end{equation}
and it is equal to a unique projection $\rho_\Q$ of $\rho$ onto $\Q$ along a $\nabla$-geodesic that is $(\gbold,\nabla,\nabla^\nsdual)$-orthogonal at $\Q$ \cite{Amari:1985,Amari:Nagaoka:1993}. More precisely, the projection $\rho_\Q$ is defined as such element of $\Q$ that
\begin{equation}
\gbold_{\rho_\Q}(\dot{c}^\nabla(t),\dot{c}^{\nabla^\nsdual}(s))=0\;\;\forall c^{\nabla^\nsdual},
\end{equation}
where $c^\nabla(t)$ is a $\nabla$-geodesic connecting $\rho$ and $\rho_\Q$, while $c^{\nabla^\nsdual}$ varies over all $\nabla^\nsdual$-geodesics intersecting $\rho_\Q$ and entirely included in $\Q$.
\begin{center}
\includegraphics[width=8cm]{rzuty6.png}\\
{\small Picture 1. \textit{$\nabla$-geodesic projection onto $\nabla^\nsdual$-affine $\nabla^\nsdual$-convex set $\Q$.}}
\end{center}
Hence, for dually flat geometries, projections onto $\nabla^\nsdual$-affine $\nabla^\nsdual$-convex sets along $\nabla$-geodesics coincide with the entropic projections of the associated Br\`{e}gman relative entropies. In consequence, one can consider a local $\nabla$-geodesic flow to be an infinitesimal version of information dynamics defined by constrained relative entropy minimisation.
\subsubsection{Quantum information geometries\label{quant.info.geom.section}}
Given a standard representation $(\H,\pi,J,\stdcone)$ of a W$^*$-algebra $\N$, consider a family of distances $D_\fff:\N_\star^+\times\N_\star^+\ra[0,\infty]$ defined by \cite{Kosaki:1982:interpolation,Petz:1985:quasientropies}
\begin{equation}
D_\fff(\omega,\phi):=\s{\stdembed_{\pi}(\phi),\fff(\Delta_{\omega,\phi})\stdembed_{\pi}(\phi)}_{\H}
\label{Df.definition}
\end{equation}
if $\supp(\omega)\leq\supp(\phi)$ and $D_\fff(\omega,\phi):=+\infty$ otherwise, where $\fff:\RR^+\ra\RR$ is any operator convex function (i.e. $\fff(tx+(1-t)y)\leq t\fff(x)+(1-t)\fff(y)$ $\forall x,y\in\N^+$ $\forall t\in[0,1]$ \cite{Kraus:1936}) satisfying $\fff(0)\leq0$ and $\fff(1)=0$. As proved in \cite{Petz:1985:quasientropies,Tomamichel:Colbeck:Renner:2009}, all $D_\fff$ satisfy the condition
\begin{equation}
D(\omega,\phi)\geq D(T_\star(\omega),T_\star(\phi)) \forall\omega,\phi\in\M(\N)\;\;\forall T_\star:\N_\star^+\ra\N_\star^+,
\label{markovian.monotonicity.D}
\end{equation}
where $T_\star$ denotes a Banach predualisation of a weakly-$\star$ continuous unital completely positive map $T:\N\ra\N$, and $\M(\N)\subseteq\N^+_\star$ is arbitrary. Moreover, the equality is attained if{}f $T_\star$ is an isomorphism \cite{HMPB:2011}.
For
\begin{equation}
\fff_\gamma(t):=
\left\{
\begin{array}{ll}
\frac{1}{\gamma}+\frac{1}{1-\gamma}t-\frac{1}{\gamma(1-\gamma)}t^\gamma&:\gamma\in\RR\setminus\{0,1\}\\
t\log t-(t-1)&:\gamma=1\\
-\log t+(t-1)&:\gamma=0,
\end{array}
\right.
\label{f.gamma.function}
\end{equation}
the restriction to $\gamma=1$ and $\M(\N)\subseteq\N^+_{\star1}$ gives \cite{Araki:1976:relative:entropy:I,Araki:1977:relative:entropy:II}
\begin{equation}
D_1|_{\N^+_{\star1}}(\omega,\phi)=\s{\stdembed_\pi(\phi),\log(\Delta_{\omega,\phi})\stdembed_\pi(\phi)}_\H,
\label{Araki.std.rep.distance}
\end{equation}
which, for $\N\iso\BH$, turns to \cite{Umegaki:1961,Umegaki:1962}
\begin{equation}
D_1|_{\schatten_1(\H)^+_1}(\omega,\phi)=\tr_\H(\rho_\omega\log(\rho_\omega)-\rho_\omega\log(\rho_\phi)).
\end{equation}
Jen\v{c}ov\'{a} \cite{Jencova:2006,Jencova:2010} proposed to consider a (Young) function
\begin{equation}
\Orlicz_\phi:\N^\sa\ni h\mapsto\frac{1}{2}(\widetilde{\phi^h}(\II)+\widetilde{\phi^{-h}}(\II))-1\in\RR^+,
\end{equation}
where $\N$ is an arbitrary W$^*$-algebra, and
\begin{equation}
\widetilde{\phi^h}(\II):=\sup_{\omega\in\N^+_\star}\{-D_1|_{\N^+_{\star1}}(\omega,\phi)+\omega(h)+\omega(\II)\}
\end{equation}
With this function, she defined a noncommutative Orlicz\footnote{More precisely, it is a noncommutative analogue of a Morse--Transue--Krasnosel'ski\u{\i}--Ruticki\u{\i} space, see \cite{Kostecki:2013} for details.} space $L_{\Orlicz_\phi}(\N)$ as a completion of $\{x\in\N^\sa\mid\exists\lambda>0\;\;\Orlicz_\phi(\lambda x)<\infty\}$ in the norm $\n{x}_{\Orlicz_\phi}:=\inf\{\lambda>0\mid\Orlicz_\phi(\lambda^{-1}x)\leq1\}$. This space satisfies
\begin{equation}
\N^\sa\sqsubseteq L_{\Orlicz_\phi}(\N):=\overline{\N^\sa}^{\n{\cdot}_{\Orlicz_\phi}}.
\end{equation}
Given the choice of a Hilbert space $\H$ with $\dim\H=:n<\infty$ and a smooth bijective parametrisation $\Theta\ni\theta\mapsto\rho(\theta)\in\schatten_1(\H)^+_0$ by the elements of an open set $\Theta\subseteq\RR^m$, $m\in\NN$, the \textit{parametric quantum manifold} is defined as a quantum model
\begin{equation}
\M(\H)=\left\{\rho(\theta)\in\schatten_1(\H)^+_0\mid\theta\in\Theta\subseteq\RR^m\right\}\subseteq\schatten_1(\CC^n)^+_0\iso\MNC^+_0.
\label{manif.dens.matrix.hilb.sp}
\end{equation}
Usually, the additional condition $\tr_\H(\rho(\theta))=1$ is imposed on the elements of $\M(\H)$. A tangent space $\T_\rho\MNC^+_0$ is the real vector space of all Fr\'{e}chet derivatives in the directions of smooth curves in $\MNC^+_0$ that pass through $\rho$, so it can be identified with a restriction of $\MNC^\sa$. A restriction of $\rho$ to $\MNC^+_{01}$ implies a restriction of the tangent vectors to the space $\{x\in\MNC^\sa\mid\tr_{\CC^n}(x)=0\}$. A Banach smooth manifold structure on $\N_{\star01}^+$ for an arbitrary countably finite\footnote{The W$^*$-algebras $\N$ which are not countably finite do not allow faithful quantum states: $\N^+_{\star0}=\varnothing$.} W$^*$-algebra $\N$ was introduced by Jen\v{c}ov\'{a} \cite{Jencova:2006,Jencova:2010}. She proved that the quantum model $\N^+_{\star01}$ can be equipped with the smooth Banach manifold structure modeled on a family of Banach spaces
\begin{equation}
L^0_{\Orlicz_\phi}(\N):=\{x\in L_{\Orlicz_\phi}(\N)\mid\phi(x)=0\}=\overline{\{x\in\N^\sa\mid\phi(x)=0\}}^{\n{\cdot}_{\Orlicz_\phi}}.
\end{equation}
This structure is introduced by means of the smooth atlas $\{(w_\phi^{-1}(U(\phi)),w_\phi)\mid\phi\in\N^+_{\star01}\}$, where $U(\phi):=\{x\in L^0_{\Orlicz_\phi}(\N)\mid\n{x}_{\Orlicz_\phi}<1\}$ and
\begin{equation}
w_\phi^{-1}:L^0_{\Orlicz_\phi}(\N)\supseteq U(\phi)\ni h\mapsto\phi^h\in\N^+_{\star01}
\end{equation}
is a diffeomorphism. If $\N\iso L_\infty(\X,\mho(\X),\tmu)$, then this construction reduces to a smooth Banach manifold structure on $L_1(\X,\mho(\X),\tmu)^+_{01}$ introduced in \cite{Pistone:2001,Grasselli:2001:PhD}.\footnote{A closely related approach to construction of smooth information manifold, utilising Orlicz spaces of unbounded operators/functions instead of the MTKR spaces of bounded elements, was developed in \cite{Pistone:Sempi:1995} for $L_1(\X,\mho(\X),\tmu)^+_{01}$ and in \cite{Streater:2000:bounded,Streater:2004:Orlicz,Streater:2008,Streater:2009:book,Streater:2010:Banach} for (a subspace of) $\schatten_1(\H)_{01}$.} We conjecture that (analogously to the extension of this smooth manifold structure from $L_1(\X,\mho(\X),\tmu)^+_{\star01}$ to $L_1(\X,\mho(\X),\tmu)^+_{\star0}$, provided in \cite{AJLS:2012}) Jen\v{c}ov\'{a}'s construction can be extended to $\N^+_{\star0}$. Under this conjecture, we define a \df{nonparametric quantum manifold} as a quantum model $\M(\N)\subseteq\N^+_{\star0}$ equipped with a Banach smooth manifold structure induced from $\N^+_{\star0}$, by replacing $L_{\Orlicz_\phi}^0(\N)$ with $L_{\Orlicz_\phi}(\N)$.
Given any countably finite W$^*$-algebra $\N$, a finite dimensional quantum model $\M(\N)\subseteq\N^+_\star$ that is a Banach smooth submanifold of $\N^+_{\star0}$, and a quantum distance $D$ on $\M(\N)$ that is smooth and satisfies \eqref{Eguchi.condition}, one can derive the corresponding quantum Norden--Sen geometry $(\M(\N),\gbold^D,\nabla^D,{\nabla^D}^\nsdual)$. In particular, given any distance $D_\fff$, for $\N=\BH$ and $\M(\H)=\schatten_1(\H)^+_{01}$, the corresponding riemannian metric $\gbold^{D_\fff}$ takes the form \cite{Morozova:Chencov:1989,Petz:1996:monotone,Lesniewski:Ruskai:1999}
\begin{equation}
\gbold^{D_\fff}_\rho(u,v)=\s{u,\left(\hhh_\fff(\LLL_\rho\RRR_\rho^{-1})\RRR_\rho\right)^{-1}(v)}_{\schatten_2(\H)},
\label{g.D.f.eq}
\end{equation}
where $u,v\in\{x\in\BH\mid\tr_\H(x)=0\}$, while $\hhh_\fff:[0,\infty[\,\ra[0,\infty[$ is an operator monotone increasing function\footnote{A function $\hhh:\RR^+\ra\RR$ is called operator monotone increasing if{}f $0\leq x\leq y$ $\limp$ $\hhh(x)\leq\hhh(y)$ $\forall x,y\in\BH$ \cite{Loewner:1934}.} defined by
\begin{equation}
\hhh_\fff(\lambda):=\frac{(\lambda-1)^2}{\fff(\lambda)-\lambda\fff(\frac{1}{\lambda})}.
\label{hhh.fff.eqn}
\end{equation}
This implies that several different $\fff$ lead to the same $\hhh_\fff$. Hence, for any riemannian metric given by \eqref{g.D.f.eq} there is a family of distances $D_\fff$ that have it as its second order Taylor term.
Using an integral representation
\begin{equation}
\fff(\lambda)=c_1(\lambda-1)+c_2(\lambda-1)^2+c_3\frac{(\lambda-1)^2}{\lambda}+\int_0^\infty\tmu(t)(\lambda-1)^2\frac{1+t}{\lambda+t},
\end{equation}
where $c_2,c_3\geq0$, $c_1\in\RR$, and $\tmu:\,]0,\infty[\,\ra\RR^+$ is a measure satisfying $\int_0^\infty\tmu(t)\in\RR$ \cite{Lesniewski:Ruskai:1999}, Jen\v{c}ov\'{a} \cite{Jencova:2004:entropies} showed that the \df{$\fff$-connections}, defined by the Eguchi equation \eqref{Eguchi.connection} applied to $D_\fff$ distance, have the form
\begin{equation}
\gbold^{D_\fff}_\rho(\nabla^{D_\fff}_xy,z)=2\int_0^\infty\tmu(\lambda)\re(\widetilde{C}(\lambda,z,x,y))-2\int_0^\infty\tmu(\lambda^{-1})\left(\re(\widetilde{C}(\lambda,y,x,z)+\re(\widetilde{C}(\lambda,y,x,z))\right),
\end{equation}
where
\begin{equation}
\widetilde{C}(\lambda,x,y,z):=(1+\lambda)\tr\left(x\frac{1}{\lambda\RRR_\rho+\LLL_\rho}(y)\frac{1}{\RRR_\rho+\lambda\LLL_\rho}(z)\right).
\end{equation}
The connections $\nabla^{D_\fff}$ are torsion-free. Moreover, the family of quantum Norden--Sen smooth geometries $(\MNC^+_0,\gbold^{D_\gamma},\nabla^{D_\gamma},(\nabla^{D_{\gamma}})^\nsdual)$ for $\gamma\in[-1,2]$ is characterised as the dually flat Eguchi geometry arising from the $D_\fff$ distances \cite{Jencova:2003:flat,Jencova:2004:entropies}. This result corresponds to the class $D_{\fff_\gamma}$ of quantum distances determined by \eqref{f.gamma.function} belonging to both families: $D_\Psi$ and $D_\fff$ \cite{Jencova:2005,Kostecki:2013}. The relationships between various information geometric objects on quantum state spaces $\M(\N)$ can be summarised in the following diagram:
\begin{center}
\includegraphics[width=15cm]{infogeometrie2.png}\\
{\small Picture 2. \textit{Relationships between different quantum geometries.} $E$ denotes an application of the Eguchi equations. $c_\Psi$ denotes the construction of an associated canonical Br\`{e}gman distance.}
\end{center}
\subsubsection{Orlicz spaces and Br\`{e}gman projections\label{Orlicz.Br\`{e}gman.section}}
As a consequence of the above results, if $\M(\N)$ is a dually flat manifold with respect to the triple $(\gbold^{D_\gamma},\nabla^{D_\gamma},(\nabla^{D_\gamma})^\nsdual)$, then $D_\gamma$-entropic projections onto ${(\nabla^{D_\gamma})^\nsdual}$-affine-and-convex subsets are \textit{locally equivalent} to $\nabla^{D_\gamma}$-geodesic ``free fall''. The construction of families of $\nabla^{D_\gamma}$-connections in infinite dimensional noncommutative case was provided in \cite{Gibilisco:Isola:1999,Grasselli:2001:PhD,Streater:2004:duality,Jencova:2005,Jencova:2006} using the linear structure of noncommutative $L_{1/\gamma}(\N)$ spaces, and in such case this statement also holds \cite{Jencova:2005}.
More generally (going a bit beyond the scope of the current paper), the quantum Br\`{e}gman distance $D_\Psi$ is defined via nonlinear embeddings $(\ell_{L_\Upsilon(\N)(\phi)},\ell_{(L_\Upsilon(\N))^\star(\phi)})$ into noncommutative Orlicz spaces $L_\Upsilon(\N)$ and $(L_\Upsilon(\N))^\star$, respectively. These spaces play the role of a tangent space `of states' and the cotangent space `of effects', respectively (and similarly to the commutative case of \cite{Gibilisco:Pistone:1998,Gibilisco:Isola:1999}). The ${\nabla^{D_\Psi}}^\nsdual$-affinity and ${\nabla^{D_\Psi}}^\nsdual$-convexity are defined as linear affinity and linear convexity in the Orlicz space $(L_\Upsilon(\N))^\star$ of effects. Thus, the global flatness of the connection on $L_\Upsilon(\N)$ understood as a tangent space $\T_\phi\M(\N)$ corresponds to its parallel transport being given by a family of isomorphisms $U_{\phi,\omega}:\T_\phi\M(\N)\ra\T_\omega\M(\N)$ satisfying $U_{\phi,\phi}=\id_{\T_\phi\M(\N)}$ and $U_{\phi,\omega}U_{\omega,\psi}=U_{\phi,\psi}$ \cite{Jencova:2005}. From this point of view, the standard unitary transitions $V_{\phi,\omega}:\H_\omega\ra\H_\phi$ can be understood as parallel transports of the connection defined naturally by the linear structure of the fibers in the GNS Hilbert bundle. This allows us to understand $V_{\phi,\omega}$ as a ``free fall'' along the geodesics of the Levi-Civita connection of $\gbold^\gamma$ for $\gamma=\frac{1}{2}$, known as the Wigner--Yanase metric \cite{Wigner:Yanase:1963}. The riemannian distance of $\gbold^{1/2}$ for $\N\iso\BH$ reads \cite{Gibilisco:Isola:2003:WY}
\begin{equation}
d_{\gbold^{1/2}}(\rho_1,\rho_2)=2\arccos\left(\tr_\H(\sqrt{\rho_1}\sqrt{\rho_2})\right).
\label{Wigner.Yanase.distance}
\end{equation}
On the boundary of pure spaces $\gbold^{1/2}$ reduces to the Fubini--Study metric $\gbold^{\mathrm{FS}}$ \eqref{Fubini.Study.equation} multiplied by the scalar factor $4$ \cite{Petz:Sudar:1999}, hence \eqref{Wigner.Yanase.distance} divided by $2$ reduces to \eqref{Fubini.Study.distance}. A generalisation of $\gbold^{1/2}$ to countably additive W$^*$-algebras was provided by Connes and St{\o}rmer \cite{Connes:Stormer:1978}. For a given standard representation $(\H,\pi,J,\stdcone)$, it reads
\begin{equation}
\gbold^{1/2}_\phi(x,y)=\frac{1}{2}\n{(J\pi(x^*)J-\pi(y))\stdembed_\pi(\phi)}_\H^2.
\label{CS.WY.eq}
\end{equation}
For any $\M(\N)\subseteq\N^+_{\star0}$, the GNS construction equipped with the Tomita--Takesaki theory defines a corresponding standard representation $(\H_\phi,\pi_\phi,J_\phi,\stdcone_\phi)$. In such case, a direct calculation based on the properties \eqref{std.uni.trans.eq.one}-\eqref{std.uni.trans.eq.three}, using $V_{\phi,\omega}$ in the role of $\transport^{\nabla^{1/2}}_{\phi,\omega}$ applied with respect to \eqref{CS.WY.eq}, shows that these objects satisfy the Levi-Civita version of the equation \eqref{geometric.duality.two}. The corresponding relative entropy is \cite{Jencova:2005}
\begin{equation}
D_{1/2}(\phi,\psi)=2\n{u_\phi\stdembed_\pi(\phi)-u_\psi\stdembed_\pi(\psi)}^2_\H,
\end{equation}
where $u_\phi$ and $u_\psi$ are unique unitary operators arising from the polar decomposition of relative modular operators $\Delta_{\phi,\omega}$ and $\Delta_{\psi,\omega}$, respectively, where $\omega\in\N^+_{\star0}$ is arbitrary. When expressed as a Br\`{e}gman distance on the standard representation Hilbert space $\H$, this relative entropy takes the form $D_{1/2}(x,y)=\frac{1}{2}\n{x-y}^2_\H$ \cite{Jencova:2005}. Hence, the local (infinitesimal) action of the operators $V_{\phi,\omega}$ can be understood as a geodesic free fall that is locally equivalent to the minimisation of the Hilbert space norm, which in turn corresponds to a continuous linear projection operator onto a convex closed subset. In what follows, we will use the GNS Hilbert bundle having in mind the above observations.
In face of presence of other approaches to construction of smooth manifold structure on the space $L_1(\X,\mho(\X),\tmu)^+_{01}$ \cite{Burdet:Combe:Nencka:2001,AJLS:2012,Newton:2012}, one may ask for the specific motivation of the Orlicz space based approach. The main reason is to guarantee that the local neighbourhoods of an information state (which are identified with the tangent space) are accessible by means of entropic projection. More precisely, each tangent vector is identified as an equivalence class of one dimensional exponential models (i.e., $p\exp(\lambda f-\log Z(p,\lambda f))$ in the neighbourhood of $p\in L_1(\X,\mho(\X),\tmu)^+_{01}$ and $\phi^{\lambda h}$ in the neighbourhood of $\phi\in\N^+_{\star01}$). This can be viewed as a localised version of Jaynes' maximum entropy principle \cite{Jaynes:1957,Jaynes:1979:where:do:we:stand} (cf. \cite{Streater:2011}) of model construction. We consider it as a step that is conceptually similar to localisation of Minkowski space in the passage from special to general relativity theory: instead of working with information models that are globally exponential, we assume only that they are locally (infinitesimally) exponential. This leads to a question whether one can postulate local approximation by means of some other models (corresponding to minimisation of some different information distance functional), and construct smooth information manifold structure out of this postulate. In the commutative case this question has been answered in the affirmative by the recent work \cite{Vigelis:Cavalcante:2013,Vigelis:Cavalcante:2013:delta,Vigelis:Cavalcante:2014}, who have generalised the construction of \cite{Pistone:Sempi:1995} to a large family of Orlicz (more precisely, Musielak--Orlicz) spaces. The Young functions that define these spaces define the corresponding Br\`{e}gman distances. We conjecture that the similar construction can be carried out in the noncommutative case. In the case when the information distance used for the construction of the smooth information manifold belongs to the Br\`{e}gman class, the resulting information manifold is no longer locally exponential, but it is locally dually flat (locally hessian). Because dually flat manifolds can be thought of as a generalisation of cartesian space, this construction strengthens analogy to relationship between Minkowski space-time and general lorentzian manifold. From the perspective of applications of entropic projections, one can say that the generalisation of Jen\v{c}ov\'{a}'s construction of a manifold structure from one based on $D_1$ to one based on Br\`{e}gman distances $D_\Psi$ would allow for local representations of entropic $D$-projections in terms of projections along $\nabla^D$-geodesics. If entropic projections are regarded as a form of information dynamics, then one can say that such construction of the smooth information manifold facilitates the possibility of introducing local (infinitesimal) representation of entropic information dynamics. Turning it to a slogan: \textit{introducing the structure of quantum information manifold based on a Young function $\Psi$ and a corresponding Br\`{e}gman distance $D_\Psi$ amounts to postulating that information flows locally along $\nabla^{D_\Psi}$-geodesics}.
In principle, one can construct smooth information manifold structure of $\M(\N)$ using some distance $D$, and then consider the geometric structures and information dynamics on $\M(\N)$ using some other distance $\tilde{D}$, or even using some class of distances, $\{\tilde{D}^i\mid i\in I\}$. However, using the same distance on both levels allows for stronger optimality results (concerning, for example, asymptotic estimation). More specifically, the same asymptotic results (up to third order) will be obtained for any $\widetilde{D}$ that locally generates a dually flat geometry that agrees with a dually flat geometry of a Br\`{e}gman distance $D_\Psi$. Hence, the above slogan can be equipped with a user's notice: \textit{a local Norden--Sen geometry and information dynamics of such manifold can be described by an arbitrary distance $D$ that has the same Taylor expansion, up to third order, as $D_\Psi$}. Given an arbitrary quantum model $\M(\N)$ and a distance $D$ on $\M(\N)$, the pair $(\M(\N),D)$ can be called \textit{dually flat localisable} quantum geometry if{}f $\M(\N)$ can be equipped with a smooth manifold structure based on some Br\`{e}gman distance $D_\Psi$ that agrees with $D$ up to third order. Such quantum geometry can be considered as a proper information geometric analogue of a lorentzian manifold: while global geometry (and dynamics) of $\M(\N)$ is described in terms of $D$, locally it is equivalent with the description in terms of $D_\Psi$, which is equivalent with the description in terms of a dually flat geometry. See Section \ref{section.dynamics.renormalisation} for an application of these considerations for the problem of geometric nonperturbative renormalisation in quantum nonequilibrium statistical mechanics.
Let W$^*$-algebra $\N$ admit a trace $\tau\in\W_0(\N)$. A closed densely defined linear operator $x:\dom(x)\ra\H$ is called \df{$\tau$-measurable} \cite{Segal:1953,Nelson:1974} if{}f
\begin{equation}
\exists\lambda>0\;\;\tau(P^{\ab{x}}(]\lambda,+\infty]))<\infty.
\end{equation}
Let $\MMM(\N,\tau)$ denote the space of all $\tau$-measurable operators affiliated with $\pi_\tau(\N)$. Let $\Upsilon:[0,\infty[\,\ra[0,\infty]$ be an \df{Orlicz function}, i.e. convex, continuous, nondecreasing, $\Upsilon(0)=0$, $\lambda>0$ $\limp$ $\Upsilon(\lambda)>0$, and $\lim_{\lambda\ra+\infty}\Upsilon(0)=+\infty$. A noncommutative Orlicz space over $\N$ associated with $\Upsilon$ is defined as \cite{Kunze:1990}
\begin{equation}
L_\Upsilon(\N,\tau):=\Span_\CC\{x\in\MMM(\N,\tau)\mid\tau(\Upsilon(\ab{x}))\leq1\},
\end{equation}
equipped with a norm
\begin{equation}
\n{\cdot}_\Upsilon:\MMM(\N,\tau)\ni x\mapsto\inf\{\lambda>0\mid\tau(\Upsilon(\lambda^{-1}\ab{x}))\leq1\}
\end{equation}
which turn it into a Banach space. It follows that
\begin{equation}
L_\Upsilon(\N,\tau)=\{x\in\MMM(\N,\tau)\mid\exists\lambda>0\;\;\tau(\Upsilon(\lambda\ab{x}))<\infty\}.
\end{equation}
An issue of canonical generalisation of the notion of a noncommutative Orlicz space to an arbitrary W$^*$-algebra is a matter of a current research, see \cite{Kostecki:2014:Orlicz} for a discussion.
The construction of a general notion of a quantum Br\`{e}gman distance for arbitrary spaces $\N^+_\star$ is an open problem. Let $X$ be a reflexive Banach space, let $\Psi:X\ra\,]-\infty,+\infty]$ be convex, lower semi-continuous, and Legendre (see \cite{Bauschke:Borwein:Combettes:2001} for a definition). Let $C\subseteq X$ be nonempty and convex, $C\cap\INT(\efd(\Psi))\neq\varnothing$, where $\efd(\Psi):=\{x\in X\mid\Psi(x)\neq+\infty\}$, and let $y\in\INT(\efd(\Psi))$. Then the \df{Br\`{e}gman functional} on $X$, defined by
\begin{equation}
\widetilde{D}_\Psi(x,y):=\Psi(x)-\Psi(y)-\duality{x-y,\DG_y\Psi(y)}_{X\times X^\star}
\end{equation}
for $y\in\INT(\efd(\Psi))$, and $\widetilde{D}_\Psi(x,y)=+\infty$ otherwise, satisfies \cite{Bauschke:Borwein:Combettes:2001}:
\begin{enumerate}
\item[1)] $\widetilde{D}_\Psi(\cdot,y)$ is convex and lower semi-continuous,
\item[2)] $\efd(\widetilde{D}_\Psi(\cdot,y))=\efd(\Psi)$,
\item[3)] $\widetilde{D}_\Psi(x,y)=0$ $\iff$ $x=y$,
\item[4)] $\PPP^{\widetilde{D}_\Psi}_C(y)=\{*\}\in\INT(\efd(\Psi))$,
\item[5)] $\PPP^{\widetilde{D}_\Psi}_C\circ\PPP^{\widetilde{D}_\Psi}_C(y)=\PPP^{\widetilde{D}_\Psi}_C(y)$,
\item[6)] if $K$ is a vector subspace of $X$, then Chencov's generalised pythagorean theorem \cite{Chencov:1968,Chencov:1972} holds:
\begin{equation}
\widetilde{D}_\Psi(x,y)=\widetilde{D}_\Psi(x,\PPP^{\widetilde{D}_\Psi}_K(y))+\widetilde{D}_\Psi(\PPP^{\widetilde{D}_\Psi}_K(y),y)\;\;\forall(x,y)\in K\times X.
\label{generalised.pythagore.Bregman}
\end{equation}
\end{enumerate}
Let $X\iso L_\Upsilon(\N)$, and consider a map $\ell_{L_\Upsilon(\N)}:\N_\star\ra L_\Upsilon(\N)$ satisfying $\ell_{L_\Upsilon(\N)}(\N_\star^+)\subseteq(L_\Upsilon(\N))^+$, $\ell_{L_\Upsilon(\N)}(\N_\star)\subseteq\INT(\efd(\Psi))$, and bijective on its codomain. Then the \df{quantum Br\`{e}gman distance} is defined as \cite{Kostecki:2014:qig}
\begin{equation}
D_\Psi(\phi,\psi):=\widetilde{D}_\Psi(\ell_{L_\Upsilon(\N)}(\phi),\ell_{L_\Upsilon(\N)}(\psi))\;\;\forall(\phi,\psi)\in\N^+_\star\times\N^+_\star.
\end{equation}
The nontrivial open problem consists of finding the minimal additional conditions on $\Psi$ and $\ell_{L_\Upsilon(\N)}$ that are necessary and sufficient to prove that $D_\Psi(\phi,\psi)$ is smooth (or at least triple G\^{a}teaux differentiable) and satisfies the infinite dimensional analogue of the property \eqref{geometric.duality.two} as well as an equivalence of $D_\Psi$-projections onto linear convex closed subspaces $\Q$ of $L_\Upsilon(\N)$ with $(\gbold^{D_\Psi},\nabla^{D_\Psi},(\nabla^{D_\Psi})^\nsdual)$-orthogonal projections along $\nabla^{D_\Psi}$-geodesics onto $\Q$. These conditions will necessarily intertwine the properties of $\Psi$, $\Upsilon$, and $\ell_{L_\Upsilon(\N)}$. See \cite{Kostecki:2014:qig} for an additional discussion.
\subsubsection{Conjecture: a Morozova--Chencov--Petz bundle\label{MCP.bundle.section}}
Let $(\H_\phi,\pi_\phi,\Omega_\phi)$ be a GNS representation of a W$^*$-algebra $\N$ for $\phi\in\N^+_{\star0}$. Consider a scalar product on $\H_\omega$, defined by
\begin{align}
\s{[x]_\phi,[y]_\phi}_{\hhh,\phi}&:=\s{[x^*]_\phi,\JJJ^\hhh_{\phi,\phi}([y]_\phi)}_\phi,\label{MCP.scalar.product}\\
\JJJ^\hhh_{\phi,\psi}&:=\frac{1}{\hhh(\Delta_{\phi,\psi})}\RRR(\psi)^{-1},
\end{align}
where $\RRR(\psi)$ is a right multiplication by $\psi\in\N^+_{\star0}$ \cite{Sherman:2001}, while $\hhh:\RR^+\ra\RR^+$ is an operator monotone increasing function satisfying \cite{Petz:1996:monotone} $\hhh(\lambda)=\lambda\hhh(\lambda^{-1})$ $\forall\lambda>0$ (all functions $\hhh_\fff$ given by \eqref{hhh.fff.eqn} satisfy this property). We conjecture that:
\begin{itemize}
\item[1)] the completion of the vector space $\pi_\phi(\N)/\ker(\JJJ^\hhh_\phi)$ in the scalar product \eqref{MCP.scalar.product} is a Hilbert space (denoted below as $\H_{\hhh,\phi}$, with elements denoted by $[x]_{\hhh,\phi}$ for any $x\in\nnn_\phi$),
\item[2)] $\pi_{\hhh,\phi}(x):[y]_{\hhh,\phi}\mapsto[xy]_{\hhh,\phi}$ defines a nondegenerate faithful normal representation $\N\ra\BBB(\H_{\hhh,\phi})$,
\item[3)] a unique extension of the antilinear isometry $J_{\hhh,\phi}:[x]_{\hhh,\phi}\mapsto[x^*]_{\hhh,\phi}$ $\forall x\in\nnn_\phi$ determines a conjugation on $\H_{\hhh,\phi}$, turning a quadruple
\begin{equation}
\left(\H_{\hhh,\phi},\pi_{\hhh,\phi},J_{\hhh,\phi},\bigcup_{x\in\nnn_\phi\cap\nnn_\phi^*}\overline{\left\{\pi_{\hhh,\phi}(x)J_{\hhh,\phi}[x]_{\hhh,\phi}\right\}}^{\H_{\hhh,\phi}}\right)
\end{equation}
to a standard representation.
\end{itemize}
If this conjecture holds, then the above construction determines a bundle of Hilbert spaces over any topological space $\M(\N)\subseteq\N^+_{\star0}$. We will refer to it as a Morozova--Chencov--Petz bundle (it could be defined equivalently using the Morozova--Chencov functions $c$ instead of Petz's functions $\hhh$, see e.g. \cite{Kostecki:2014:qig}). Clearly, it provides an alternative to the GNS bundle. A virtue of the MCP bundle is that it encodes the local riemannian geometry of the state space in the variability of changes of the scalar product. Yet, it remains an open problem whether the above conjecture is true.
\ifvargaugecompile
\subsection{Local gauge and geodesic propagation\label{local.geodesic.propagation}}
\else
\subsubsection{Local gauge and geodesic propagation\label{local.gauge.propagation}}
\fi
If $\glie$ is a Banach Lie algebra with a Lie bracket $[\cdot,\cdot]$, then a \df{representation} of $\glie$ on a dense subset $\D\subseteq\H$ of a Hilbert space $\H$ is defined as a linear function $\aaa$ mapping each $x\in\glie$ to an anti-selfadjoint operator $\rpktarget{AAA}\aaa(x):\D\ra\H$ such that
\begin{equation}
\aaa([x,y])=\aaa(x)\aaa(y)-\aaa(y)\aaa(x)\;\;\forall x,y\in\glie.
\end{equation}
Hence, for a given representation $\aaa$ of $\glie$ on $\D\subseteq\H$, every $x\in\glie$ determines a unique self-adjoint, and generally unbounded, operator $\ii\aaa(x)$. By definition, $\D=\dom(\ii\aaa(x))$.
Let $G$ be a Lie group, and let $\M(\N)\subseteq\N^+_\star$ be equipped with a principal $G$-bundle $E\ra\M(\N)$, and a $\glie$-valued connection one-form $\rpktarget{AAAA}\mathbf{A}$ on $E$, where $\glie$ is a Lie algebra of $G$. Moreover, assume that the GNS bundle $\H\M(\N)$ is equipped with the family $\aaa$ of the representations of the Lie algebra $\glie$,\footnote{This definition covers also the representations of $\glie$ in the well-adapted Banach--Lie subalgebras $\B\subseteq\eee^{-1}(\omega)$ (thus, within the fibers of $\H\M(\N)$) as the special case.}
\begin{equation}
\aaa:=\{\aaa_\omega:\glie\ra(\Lin(\H_\omega))^\asa\mid\omega\in\M(\N)\}.
\end{equation}
The triple $(\M(\N),\mathbf{A},\aaa)$ satisfying the above conditions will be called a \df{local gauge model}, while the pair $(\mathbf{A},\aaa)$ will be called a \df{local gauge structure} on $\M(\N)$.\footnote{The term \df{gauge} means the section of a principal $G$-bundle \cite{Weyl:1918,Weyl:1929}. The \df{local gauge} means the local section, while the \df{global gauge} means the global section. A particularly interesting example of a local gauge structure is provided by the choice of a locally compact and connected Lie group $G=\mathrm{SO}^\uparrow(1,3)\ltimes\RR^4$, known as \df{ortochronous Poincar\'{e} group}.} In principle, a given manifold $\M(\N)$ can admit various different local gauge structures.
If the model $\M(\N)$ is equipped with the local gauge structure, then any curve $c:\RR\ni t\mapsto\phi(t)\in\M(\N)$ corresponds also to a specific choice of a section of the principal $G$-bundle $E$ along this trajectory, which can be expressed by means of integral of a $\glie$-valued connection 1-form $\mathbf{A}$.
If $\aaa$ is determined by setting $\ii\aaa_\omega(\glie)$ to be equal to the generators of the irreducible unitary representation of an action of $G$ on $\H_\omega$, then one can apply Wigner's theorem \cite{Wigner:1939} to each fibre of $\H\M(\N)$ \textit{separately}, classifying the elements of $\H_\omega$ into subsets by means of their transformation properties. According to Wigner's interpretation of this mathematical property (which became widely accepted afterwards), an element of $\H_\omega$ transforming under the above representation of $G$ shall be understood as a pure state of a `quantum particle', where pure state means a vector in a Hilbert space.
Our framework allows to enrich this interpretation by considering a propagation of a `quantum particle' state over the trajectory on the manifold $\M(\N)$, using the $\glie$-valued connection one-form $\mathbf{A}$. As discussed in Section
\ifvargaugecompile
\ref{standard.liouvilleans.section}%
\else
\ref{relative.modular.theory.section}%
\fi
, if $\M(\N)\subseteq\N^+_{\star0}$, then every two standard representations determine a unique standard unitary transition between them that preserves the standard cone. Hence, one can map uniquely between the elements of the fibres $\H_{\phi_1}$ and $\H_{\phi_2}$, whenever $\phi_1,\phi_2\in\N^+_{\star0}$, by means of the standard unitary transition operator $V_{\phi_1,\phi_2}$. Let $\M(\N)\subseteq\N_\star^+$, let $c:[0,t]\ra\M(\N)$ be a curve with $c(0)=\omega$ and $c(t)=\phi$, let $\xi\in\H_\omega$ and $\zeta\in\H_\phi$. Then $\zeta$ will be called an \df{$\mathbf{A}$-propagation} of $\xi$ along $c$ if{}f $\ii\aaa_\phi(\int_c\mathbf{A})-\ii J_\phi\aaa_\phi(\int_c\mathbf{A})J_\phi$ is essentially self-adjoint on $\dom(\ii\aaa_\phi(\int_c\mathbf{A}))\cap\dom(\ii J_\phi\aaa_\phi(\int_c\mathbf{A})J_\phi)$ and
\begin{equation}
\zeta=U^{\mathbf{A}}_{c,\phi}(t)V_{\phi,\omega}\xi
:=\ee^{-\ii t(\ii\aaa_\phi(\int_{c(t)}\mathbf{A})+\ii J_\phi\aaa_\phi(\int_{c(t)}\mathbf{A})J_\phi)}V_{\phi,\omega}\xi
=\ee^{t([(\aaa_\phi(\int_{c(t)}\mathbf{A})),\,\cdot\,])}V_{\phi,\omega}\xi.
\label{parallel.zeta}
\end{equation}
The operator $V_{\phi,\omega}$ is a parallel transport associated with the natural connection in the GNS Hilbert bundle determined by the linear structure of the Hilbert space. Hence, the equation \eqref{parallel.zeta} can be understood as an updating map $\xi\mapsto\zeta$ along the trajectory $c(t)$ that takes into account both $\mathbf{A}$ and $\nabla^{\gbold^{1/2}}$ connections.
The above construction suggests introducing more tight relationship between the connection structures of $\T\M(\N)$ and $\H\M(\N)$ for \textit{any} local gauge model $(\M(\N),\mathbf{A},\aaa)$ such that $\M(\N)$ can be equipped with a smooth manifold structure, and with $\M(\N)\subseteq\N^+_{\star0}$. Two quite intriguing possibilities are:
\begin{enumerate}
\item[(QP$_1$)] introduce an affine connection $\nabla$ on $\T\M(\N)$ and define \df{gauge geodesic propagation} as an $\mathbf{A}$-propagation along $\nabla$-geodesic in $\M(\N)$;
\item[(QP$_2$)] introduce: i) an action of $G$ on $\T\M(\N)$ turning it to a tangent $G$-bundle, ii) a $\glie$-valued connection 1-form $\mathbf{A}^\glie_\T$ on $\T\M(\N)$, and iii) a $\glie$-valued connection 1-form $\mathbf{A}_\H^\glie$ on $\H\M(\N)$, such that the $\mathbf{A}_\H^\glie$-parallel transports along $\mathbf{A}_\T^\glie$-geodesics in $\M(\N)$ are equal to $U^\mathbf{A}_{c,\phi}(t)V_{\phi,\omega}$ or $V_{\phi,\omega}U^\mathbf{A}_{c,\omega}(t)$ (where $t$ is an affine parameter of an $\mathbf{A}_\T^\glie$-geodesic $c(t)\in\M(\N)$), and define \df{gauge geodesic propagation} as a horizontal lift of an $\mathbf{A}_\T^\glie$-geodesic in $\M(\N)$ with respect to $\mathbf{A}_\H^\glie$. These gauge geodesic propagations are precisely the $\mathbf{A}$-propagations along $\mathbf{A}_\T^\glie$-geodesics in $\M(\N)$. So, this what we gain by such definition is an additional structure on $\H\M(\N)$ that allows for further study of a relationship between $(G,\glie,\mathbf{A}^\glie_\cdot)$-structures of $\T\M(\N)$ and $\H\M(\N)$. On the other hand, the price paid is the requirement that the $\mathbf{A}_\H^\glie$-parallel transport along $\mathbf{A}_\T^\glie$-geodesic depends only on its endpoint, which holds if $\M(\N)$ is simply connected and $\mathbf{A}_\H^\glie$ is flat.
\end{enumerate}
If the definition (QP$_1$) is used, and $G$ and $\aaa$ are chosen as for Wigner's `quantum particle' classification discussed above, then the gauge geodesic propagation has a direct interpretation as an $\mathbf{A}$-propagation of a `quantum particle' due to ``free fall'' along $\nabla$-geodesic. Note that (despite `dynamical' feeling associated with the word `fall') this propagation has no `dynamical' (causal) content: it is an extension of the description of gauge transformation properties of a `quantum particle' state from a single Hilbert space to a Hilbert space fibre bundle over a quantum model equipped with a local gauge structure. In such approach, a `quantum particle' becomes identified with a (not necessarily global) section of a fibre $G$-bundle represented in terms of the fibre bundle of the GNS Hilbert spaces, so one can discuss its quantum propagation between some `source model' and some `sink model', defined as suitable submanifolds of $\M(\N)$. Under the choice of a definition (QP$_1$), we will define a \df{gauge geodesic propagation model} as a local gauge model $(\M(\N),\mathbf{A},\aaa)$ equipped with an affine connection $\nabla$.
The motivation and interpretation of these definitions echoes Einstein's postulate \cite{Einstein:Grossmann:1913,Einstein:1916} of identification of geodesic lines with the \cytat{world-lines of freely moving point-particles}, but generalised from the pseudo-riemannian geometry, for which the geodesics of the Levi-Civita connection coincide with the curves of extreme distance of a pseudo-riemannian metric, to the setting of general affine connections \cite{Schouten:Haantjes:1934} (see e.g. \cite{Sa:1997,Manoff:2000} and references therein for a comparative discussion of these two approaches). The above construction of geodesic propagation of `quantum particles' is partially influenced by the works of Drechsler \cite{Drechsler:1984,Drechsler:1992,Drechsler:Tuckey:1996} and Prugove\v{c}ki \cite{Prugovecki:1985,Prugovecki:1987,Prugovecki:1991,Prugovecki:1992,Prugovecki:1994,Prugovecki:1995,Prugovecki:1996:I,Prugovecki:1996:II} (see also \cite{Graudenz:1994,Graudenz:1996}). As opposed to them, we do not require any pseudo-riemannian metric on the base manifold, so we do not introduce soldered Poincar\'{e} frame bundles, and we also consider the GNS Hilbert spaces (which may be unitarily inequivalent, if $\omega\notin\N^+_{\star0}$) varying over the base manifold instead of pasting fibre bundle from identical copies of a single Hilbert space. Moreover, our manifold is a space of quantum states over W$^*$-algebras, as opposed to \textit{a priori} postulated background space-time. On the other hand, similarly to Prugove\v{c}ki (see \cite{Prugovecki:1992,Prugovecki:1995,Prugovecki:1996:II}), and as opposed to the approaches of Wightman \cite{Wightman:1956,Wightman:1957,Wightman:1964,Wightman:Gaerding:1964} and Haag--Kastler \cite{Haag:1959,Haag:Kastler:1964}, we impose the requirement of Poincar\'{e} covariance not on the \textit{topological} subsets of the base manifold and on the (presheaves of) algebras of operators associated (functorially \cite{Brunetti:Fredenhagen:Verch:2003}) to those subsets, but on the fibres of \textit{geometric} fibre bundle and on the fibre bundle of Hilbert spaces over this manifold. (Note that, in addition, one can also introduce an independent group covariance requirement on the elements of an underlying W$^*$-algebra, determining this way their transformation properties at each fibre by means of the GNS representation.)
If $\M(\N)$ is equipped with the structure of quantum information manifold, and with a quantum information distance $D_\fff$, then (at least in the finite dimensional case) $\nabla$ can be chosen to be given by $\nabla^{D_\fff}$, $(\nabla^{D_\fff})^\nsdual$, or $(\nabla^{D_\fff}+(\nabla^{D_\fff})^\nsdual)/2$ (the latter is a Levi--Civita of the riemannian geometry $(\M(\N),\gbold^{D_\fff})$). Moreover, if $\M(\N)$ is flat with respect to $\nabla^{D_\fff}$ and $(\nabla^{D_\fff})^\nsdual$, then one can identify the entropic projection of the associated Br\`{e}gman distance $D_\Psi$ with the geodesic projection by $\nabla^{D_\fff}$. If $\fff=\fff_\gamma$, then $D_\fff$ is also a Br\`{e}gman distance. In such case, an assumption (QP$_1$) and identification of $\nabla$ with $\nabla^{D_\gamma}$, allows to state that the propagation of a `quantum particles' is due to the geodesic ``free fall'' along the (local) information dynamics, determined by a constrained minimisation of a relative entropy.
\ifvargaugecompile
\section{Locally perturbed liouvilleans\label{from.local.gauge.to.lagrangeans.section}}
\else
\subsection{Locally perturbed liouvilleans\label{from.local.gauge.to.lagrangeans.section}}
\fi
In this Section we will use the approach to unbounded `perturbations' of liouvilleans, developed by the Derezi\'{n}ski, Jak\v{s}i\'{c} and Pillet in \cite{DJP:2003}, and based on earlier results of Araki \cite{Araki:1973:relative:hamiltonian,Araki:1973:golden:thompson}. According to \cite{DJP:2003}, if
\begin{enumerate}
\item $\C\subseteq\BH$ is a von Neumann algebra,
\item $(\C,\RR,\varsigma)$ is a W$^*$-dynamical system with a standard liouvillean $K^\varsigma$ associated with a standard form $(\H,\C,J,\stdcone)$ of $\C$,
\item $Q$ is a self-adjoint linear operator affiliated
to $\C$,
\item $K^\varsigma+Q$ is an essentially self-adjoint linear operator on $\dom(K^\varsigma)\cap\dom(Q)\subseteq\H$,
\end{enumerate}
then for
\begin{equation}\rpktarget{DJP.PERT.ALPH}
\varsigma_t^Q(x):=\ee^{\ii t(K^\varsigma+Q)}x\ee^{-\ii t(K^\varsigma+Q)}\;\;\forall x\in\C\;\forall t\in\RR
\end{equation}
the following statements are true:
\begin{enumerate}
\item $(\C,\RR,\varsigma^Q)$ is a W$^*$-dynamical system.
\item the operator
\begin{equation}
E_{\varsigma,Q}(t):=\ee^{\ii t(K^\varsigma+Q)}\ee^{-\ii tK^\varsigma}\in\C,
\end{equation}
called an \df{expansional}, is unitary and for all $t,t_1,t_2\in\RR$ and all $x\in\C$ it satisfies the following cocycle conditions:
\begin{align} \varsigma_t^Q(x)&=E_{\varsigma,Q}(t)\varsigma_t(x)E_{\varsigma,Q}(t)^{-1},\\
E_{\varsigma,Q}(t)^{-1}&=E_{\varsigma,Q}(t)^*=\varsigma_t(E_{\varsigma,Q}(-t)),\\
E_{\varsigma,Q}(t_1+t_2)&=E_{\varsigma,Q}(t_1)\varsigma_{t_1}(E_{\varsigma,Q}(t_2)).
\label{EQcocycle}
\end{align}
\item if\rpktarget{DJP.PERT}
\begin{equation}
K^{\varsigma,Q}:=K^\varsigma+Q-JQJ
\label{pert.Q}
\end{equation}
is an essentially self-adjoint linear operator on $\dom(K^\varsigma)\cap\dom(Q)\cap\dom(JQJ)$, then a unique self-adjoint extension of $K^{\varsigma,Q}$, denoted (with an abuse of notation) by the same symbol, is a standard liouvillean of $\varsigma^Q$ in $(\H,\C,J,\stdcone)$. We will call $K^{\varsigma,Q}$ and $\varsigma^Q$ a \df{Derezi\'{n}ski--Jak\v{s}i\'{c}--Pillet perturbation} of $K^\varsigma$ and $\varsigma$, respectively.
\item if $Q$ is bounded, then, for any $x\in\C$ and $t\in\RR$, the Dyson--Feynman--Fujiwara--Araki perturbative expansions \cite{Dyson:1949,Feynman:1951,Fujiwara:1952,Araki:1973:expansionals},
\begin{align}
\varsigma_t^Q(x)&=\sum\limits_{n=0}^\infty \ii^n\int\limits_{0\leq t_n\leq\cdots\leq t_1\leq t}\dd t_1\cdots\dd t_n[\varsigma_{t_n}(Q),[\ldots, {[}\varsigma_{t_1}(Q),\varsigma_t(x)]\ldots]],\\
E_{\varsigma,Q}(t)&=\sum\limits_{n=0}^\infty\ii^n\int\limits_{0\leq t_n\leq\cdots\leq t_1\leq t}\dd t_1\cdots \dd t_n\varsigma_{t_n}(Q)\cdots\varsigma_{t_1}(Q).
\end{align}
are convergent in weak-$\star$ topology and define a norm convergent series of bounded operators. Moreover, the associated generators $\der$ of $\varsigma$ and $\der_Q$ of $\varsigma^Q$ have in such case the same domain, and are related by
\begin{equation}
\der_{\varsigma^Q}(x)=\der_\varsigma(x)+\ii[Q,x]\;\;\forall x\in\dom(\der_\varsigma).
\label{bounded.perturbation.of.generator}
\end{equation}
\end{enumerate}
\ifvargaugecompile
\subsection{Local gauge liouvilleans\label{local.gauge}}
\else
\subsubsection{Local gauge liouvilleans\label{local.gauge}}
\fi
In Section \ref{algebraic.hamiltonian} we have observed that the Poisson flow of an algebraic hamiltonian vector field on $\N^+_{\star0}$ can be always represented by a unitary evolution in a fibre of the GNS bundle $\H\N_{\star0}^+\ra\N^+_{\star0}$, generated by a standard liouvillean operator on this fibre. In what follows, we will abstract this relationship, replacing tangent bundle by a principal $G$-bundle and replacing the procedure of \textit{restriction} of a \textit{global} algebraic evolution, by the procedure of \textit{extension} of a \textit{local} algebraic evolution in order to incorporate geometric structure as an additional component of an effective dynamics.
If one \textit{assumes} that the principal $G$-bundle structure participates in an effective form of a temporal evolution, then this evolution shall be represented not by the standard liouvillean on $\H_{\phi(t)}$ alone, but by the standard liouvillean perturbed by the `gauge connection' operator $\ii\aaa_{\phi(t)}(\mathbf{A})$, which represents the change of vectors in the fibres $\H_{\phi(t)}$ caused by the fibrewise action of the group $G$ and the choice of an $\mathbf{A}$-section of a principal $G$-bundle $E$. (When the trajectory along the curve $c:[0,t]\ra\M(\N)$ with $c(0)=\phi(0)$ and $c(t)=\phi(t)$ is investigated as a source of memory effects, then $\mathbf{A}$ should be replaced by $\int_{\phi(0)}^{\phi(t)}\mathbf{A}$.)
For this purpose, we will use DJP perturbation approach, setting $\omega\in\M(\N)\subseteq\N^+_{\star0}$, $\H:=\H_\omega$, $\C:=\pi_\omega(\N)$, $Q:=\ii\aaa_\omega(\mathbf{A})$. We can define $\varsigma$ in two different ways. If $\N$ is equipped with a W$^*$-dynamical system structure $(\N,\RR,\alpha)$, then one can define $\varsigma$ globally in each fiber of the GNS bundle by means of
\begin{equation}
\varsigma_t(\pi_\omega(x)):=\pi_\omega(\alpha_t(x))\;\;\forall x\in\N\;\;\forall t\in\RR\;\;\forall\omega\in\M(\N).
\label{boring.covariance.cond}
\end{equation}
Alternatively, if $\N$ is equipped with a quantum Poisson system $(\M(\N,\B),h)$ such that $\DF_\omega h\in\N^\sa$ $\forall\omega\in\M(\N,\B)$ and $\M(\N,\B)=\M(\N)$, then one can define $\varsigma$ pointwisely in each fiber by means of
\begin{equation}
\varsigma_t(\pi_\omega(x)):=\ee^{\ii t\pi_\omega(\DF_\omega h)}\pi_\omega(x)\ee^{-\ii t\pi_\omega(\DF_\omega h)}\;\;\forall x\in\N\;\;\forall t\in\RR\;\;\forall\omega\in\M(\N).
\label{gen.Poisson.compatibility.cond}
\end{equation}
We will call these assuptions a \df{generalised Poisson compatibility condition} (PC$_2$). If $\B=\N^\sa$, $\M(\N,\N^\sa)$ is a submanifold of $\N^\sa_\star$, and the pair $(h,\alpha)$ satisfies \eqref{poisson.flow.compatibility}, then the above two definitions of $\varsigma$ agree. The difference between \eqref{gen.Poisson.compatibility.cond} and \eqref{boring.covariance.cond} (the latter corresponding to the covariance equation \eqref{covariance.equation.group.G}) indicates our approach to quantum dynamics, as being defined locally by the differential geometric properties of state space, instead of a global automorphism of an underlying algebraic structure. Alternatively, if $\M(\N)\subseteq\N^+_{\star0}$ is equipped with a quantum information manifold structure and a global vector field $\XXX_h\in\T\M(\N)$ such that $\XXX_h(\phi)\in\N^\sa$ $\forall\phi\in\M(\N)$, then $\varsigma$ can be defined by
\begin{equation}
\varsigma_t(\pi_\omega(x)):=\ee^{\ii t\pi_\omega(\XXX_h(\omega))}\pi_\omega(x)\ee^{-\ii t\pi_\omega(\XXX_h(\omega))}\;\;\forall x\in\N\;\forall t\in\RR\;\forall\omega\in\M(\N).
\end{equation}
Thus, under some relatively weak conditions (affiliation of $\ii\aaa_\omega(\mathbf{A})$ with $\pi_\omega(\N)\subseteq\BBB(\H_\omega)$ and essential self-adjointness of sums $K_\omega^\varsigma+\ii\aaa_\omega(\mathbf{A})$ and $K_\omega^\varsigma+\ii\aaa_\omega(\mathbf{A})-\ii J_\omega\aaa_\omega(\mathbf{A})J_\omega$ on the intersection of domains of their components), the local gauge structure can be incorporated in the redefinition of the standard liouvillean. If
\begin{equation}
K_\omega^{\varsigma,\ii\aaa(\mathbf{A})}=K_\omega^\varsigma+\ii\aaa_\omega\left(\mathbf{A}\right)-\ii J_\omega\aaa_\omega\left(\mathbf{A}\right)J_\omega
\label{lagrangean.eq}
\end{equation}
satisfies the above conditions, then we will call it a \df{local gauge liouvillean} at $\omega$.
Let $(\M(\N),\mathbf{A},\aaa)$ be a local gauge model with $\M(\N)\subseteq\N^+_{\star0}$ and let $\varsigma$ be defined as above, either by a quantum Poisson system $(\M(\N,\B),h)$ with $\DF_\omega h\in\N^\sa$ $\forall\omega\in\M(\N,\B)$ and $\M(\N,\B)=\M(\N)$, or by a W$^*$-dynamical system $(\N,\RR,\alpha)$ with $\alpha_\star^t(\M(\N))\subseteq\M(\N)$ $\forall t\in\RR$. Let $\rho_\star$ denote $w^h$ or $\alpha_\star$, respectively.
If for every $\omega\in\M(\N)$ there exists a family of local gauge liouvilleans
\begin{equation}
K_\omega^\varsigma+\ii\aaa_\omega\left
\mathbf{A}\right)-\ii J_\omega\aaa_\omega\left
\mathbf{A}\right)J_\omega,
\label{local.gauge.liouvillean}
\end{equation}
parametrised by $t\in\RR$, then the quadruple $(\M(\N),\varsigma,\mathbf{A},\aaa)$ will be called a \df{local gauge liouvillean model}.
From the above construction we see that the W$^*$-dynamical system $(\pi_\omega(\N),\RR,\varsigma^{\ii\aaa(\mathbf{A})})$ may not correspond to any W$^*$-dynamical system on $\N$. The description of temporal evolution in terms of $\varsigma^{\ii\aaa(\mathbf{A})}$ is `local' in the sense that it is provided inside of each fibre of the GNS bundle independently.
\ifvargaugecompile
\else
\footnote{If $G_1$ and $G_2$ are groups, and $G_1$ acts on $G_2$ by means of $\cdot:G_1\times G_2\ni(g_1,g_2)\mapsto g_1\cdot g_2\in\G_2$, then a \df{semi-direct product} $G_1\ltimes G_2$ is defined as a group with elements given by pairs $(g_1,g_2)\in G_1\times G_2$, and composition given by $(g_1,g_2)(g_3,g_4)=(g_1g_3,g_2(g_1\cdot g_4))$.}
\fi
\ifvargaugecompile
\subsection{Local source liouvilleans\label{local.source.liouv.section}}
\else
\subsubsection{Local source liouvilleans\label{local.source.liouv.section}}
\fi
In principle, apart from the `internal' dynamics (implemented by the evolution $\varsigma^t$) and the kinematic local gauge structure, the effective dynamics can also depend on some controlled `external' constraints. We will assume that these constraints can be specified in terms of external `sources', which can generally be represented by the variations $\delta(\phi(x))$ of expectation values. These variations can be decomposed into two parts: the variations $(\delta\phi)(x)$ of states, and the variations $\phi(\delta x)$ of operators. The constraints on changes $\delta\phi$ can be handled by restricting the form of the model $\M(\N)$ (for a geometric approach, see \cite{Mitchell:1967,Jaynes:1993,Favretti:2007,Kostecki:2013:PhD}). On the other hand, note that $\delta x$ can be in principle arbitrary, so it can also depend on $\phi$, and it may not arise as an infinitesimal change generated by a global automorphism of $\N$ (thus, it cannot be described by the setting of derivations of C$^*$-algebras). We will implement the perturbations $\delta x$ of elements $x$ of (a local GNS representation of) a W$^*$-algebra $\N$ by means of state dependent perturbations of liouvilleans. In this sense, the constraints on changes of operators will be handled by local (state dependent) additional terms modifying liouvillean evolution.
For this purpose, apart from local gauge structure on $\M(\N)\subseteq\N_\star^+$, we introduce also \df{local source term}, defined as a fibrewise family of operators
\begin{equation}
(\lambda,H):\M(\N)\ni\omega\mapsto\lambda(\omega)H(\omega)\in(\Lin(\H_\omega))^\sa,
\end{equation}
with $\lambda(\omega)\in\RR$ called \df{local source strength} and $H(\omega)\in(\Lin(\H_\omega))^\sa$ called \df{local source operator}. If $\lambda(\omega)$ is independent of $\omega$, then it will be called \df{global source strength}. The $(2n+1)$-tuple
\begin{equation}
(\M(\N),(\lambda_1,H_1),\ldots,(\lambda_n,H_n))
\label{local.source.model}
\end{equation}
will be called \df{local source model} if{}f $(\lambda_i,H_i)$ is a local source term for each $i\in\{1,\ldots,n\}$. If \eqref{local.source.model} is a local source model, $\M(\N)\subseteq\N^+_{\star0}$, $\varsigma$ is defined as in Section \ref{local.gauge}, and for $\omega\in\M(\N)$ the DJP perturbation of $K_\omega^\varsigma$ by $\lambda_1(\omega)H_1(\omega)+\ldots+\lambda_n(\omega)H_n(\omega)$ exists, then a unique self-adjoint extension of an essentially self-adjoint operator
\begin{equation}
K_\omega^{\varsigma,\lambda_1H_1,\ldots,\lambda_nH_n}=K_\omega^\varsigma+\lambda_1(\omega)H_1(\omega)+\ldots+\lambda_n(\omega)H_n(\omega)-J_\omega(\lambda_1(\omega)H_1(\omega)+\ldots+\lambda_n(\omega)H_n(\omega))J_\omega
\end{equation}
will be called \df{local source liouvillean} at $\omega$. If a local source liouvillean exists for each $\omega\in\M(\N)$, then the $2(n+1)$-tuple $(\M(\N),\varsigma,\lambda_1,H_1,\ldots,\lambda_n,H_n)$ will be called \df{local source liouvillean model}.
Let us note that a local source $\lambda_i(\omega)H_i(\omega)$ at $\omega$ should be understood not as the $i$-th type ``interaction source'' \textit{localised} at $\omega$, but as a strength-and-action of the $i$-th type ``interaction source'' \textit{perceived} at location $\omega$.
Let $(\M(\N),\mathbf{A},\aaa)$ be local gauge model. Let $(\M(\N),\lambda_1,H_1,\ldots,\lambda_n,H_n)$ be local source model. Let $\varsigma$ be such as defined in Section \ref{local.gauge}. If, for a given $t\in\RR$ and $\omega\in\M(\N)$, there exists a DJP perturbation of a standard liouvillean $K_\omega^\varsigma$, given by the unique self-adjoint extension of an essentially self-adjoint operator
\begin{equation}
\Lcal(\omega,t):=K_\omega^\varsigma+\ii\aaa_\omega\left
\mathbf{A}\right)-\ii J_\omega\aaa_\omega\left
\mathbf{A}\right)J_\omega+\sum_{i=1}^n\left(\lambda_i(\omega)H_i(\omega)-J_\omega\lambda_i(\omega)H_i(\omega)J_\omega\right),
\end{equation}
then $\rpktarget{EXT.LIOUV}\Lcal(\omega,t)$ will be called a \df{local liouvillean operator} at $(\omega,t)$.
If $\Lcal(\omega,t)$ exists for all $t\in\RR$ and all $\omega\in\M(\N)$, then the $(4+2n)$-tuple
\begin{equation}
(\M(\N),\varsigma,\mathbf{A},\aaa,\lambda_1,H_1,\ldots,\lambda_n,H_n)
\end{equation}
will be called an \df{local liouvillean model}. In the special case, all of operators $H_1,\ldots,H_n$ can be determined by the elements $h_1,\ldots,h_n$ of a W$^*$-algebra $\N$, with
\begin{equation}
H_i(h_i):\M(\N)\ni\omega\mapsto H_i(\omega):=\pi_\omega(h_i)\in\BH^\sa.
\label{linear.sources.in.GNS.bundle}
\end{equation}
This allows, in particular, for a fibrewise representation of a `global gauge' action $G_0\ra\Aut(\N)$ of some Lie group $G_0$ (not necessarily related to $G$), whenever $\{h_i\}\subseteq\N$ are the generators of the representation of $G_0$ in $\Aut(\N)$. This observation can be generalised to $l$ subsets of $\{h_1,\ldots,h_n\}$ playing the role of generators of $l$ representations of $l$ Lie groups $G_l\ra\Aut(\N)$.
\ifvargaugecompile
\subsection{Case study: The BLP perspective on nonlinear quantum fields}
\else
\subsubsection{Case study: The BLP perspective on nonlinear quantum fields}
\fi
Restriction of considerations from topological spaces $\M(\N)$ to BLP manifolds $\M(\N,\B)$ allows us to equip the operator algebraic approach with an additional differential geometric content,
using Fr\'{e}chet derivatives of smooth functions on $\B_\star$ in the role of differential forms. In this Section we will investigate the possibility of interpretation of these forms as nonlinear quantum fields (understood in quite formal sense).
If $f\in\CIF(\B_\star;\RR)$ and $\phi\in\M(\N,\B)$, then $\DF_\phi f=\ddd_\phi f\equiv\ddd f(\phi)\in\T^\ct_\phi\M(\N,\B)$. Thus, if $\M(\N,\B)\subseteq\N^+_{\star0}$, $f\in\CIF(\M(\N,\B);\RR)$ satisfies $\DF_\omega f\in\N^\sa$ $\forall\omega\in\M(\N,\B)$, and $\varsigma$ is defined as in the previous two Sections, then one can consider local source liouvilleans determined by the perturbation
\begin{equation}
K_\omega^\varsigma+\lambda(\omega)\left(\pi_\omega(\DF_\omega f)-J_\omega\pi_\omega(\DF_\omega f)J_\omega\right),
\end{equation}
for some family $\lambda(\omega)\in\RR$ $\forall\omega\in\M(\N,\B)$. As opposed to a function used for definition of $\varsigma$ and $K^\varsigma_\omega$, $f$ is not considered as a generator of a Poisson flow, and it is allowed to be arbitrary rescaled by $\lambda(\omega)$ at each point.
Every $x\in\B$ can be represented again as a smooth function on $\B_\star$ by means of $\B_\star\ni\omega\mapsto\omega(x)\in\RR$, allowing to consider elements of $\B$ arising from multiple Fr\'{e}chet differentiation,
\begin{equation}
\DF_{\omega_n}(\phi_{n-1}(\DF_{\omega_{n-1}}(\cdots(\phi_1(\DF_{\omega_1}f)))))\in\B,
\label{multiple.point.frechet}
\end{equation}
for $\omega_1,\ldots,\omega_n,\phi_1,\ldots,\phi_{n-1}\in\B$. These derivatives can be added and multiplied as elements of $\B$, and any of the resulting elements of $\B$ can be subjected to a representation in the GNS bundle as a local source operator. However, despite multiple application of Fr\'{e}chet differentiation, objects of type \eqref{multiple.point.frechet} are (just) elements of $\T^\ct_{\omega_n}\B_\star$. This leads us to ask whether it is possible to introduce higher order tensors on $\B_\star$, which could be used as source terms acting on the GNS bundle. The natural candidates for this purpose are \df{$(n,m)$-tensor fields} over $\B_\star$, defined pointwisely as
\begin{equation}
\XXX_{k_1}(\phi)\boxtimes\cdots\boxtimes\XXX_{k_n}(\phi)\boxtimes\ddd f_1(\phi)\boxtimes\cdots\boxtimes\ddd f_m(\phi),
\end{equation}
which are the elements of
\begin{equation}
\left(\bigboxtimes^n\T_\phi\B_\star\right)\boxtimes\left(\bigboxtimes^m\T^\ct_\phi\B\right)
\iso\left(\bigboxtimes^n\B_\star\right)\boxtimes\left(\bigboxtimes^m\B\right),
\end{equation}
where $\boxtimes$ denotes the tensor product considered in an algebraic sense (that is, without taking topological completion) and the dependence on $\phi$ is assumed to be smooth. The contraction at $\phi$ of an $(n,m-1)$-tensor field with an $(n,m)$-tensor field by means of the componentwise application of the duality $\duality{\cdot,\cdot}_{\B_\star\times\B}$ gives a one-form at $\phi$, which belongs to $\B$:
\begin{align}
&\duality{\ddd l_1(\phi)\boxtimes\cdots\boxtimes\ddd l_n(\phi)\boxtimes\XXX_{h_1}(\phi)\boxtimes\cdots\boxtimes(\cdot)_i\boxtimes\cdots\XXX_{h_m}(\phi),\XXX_{k_1}(\phi)\boxtimes\cdots\boxtimes\XXX_{k_n}(\phi)\boxtimes\ddd f_1(\phi)\boxtimes\cdots\boxtimes\ddd f_m(\phi)}\nonumber\\
&=\duality{\XXX_{k_1}(\phi),\ddd l_1(\phi)}\cdots
\duality{\XXX_{k_n}(\phi),\ddd l_n(\phi)}
\duality{\XXX_{h_1}(\phi),\ddd f_1(\phi)}
\duality{\XXX_{h_{i-1}}(\phi),\ddd f_{i-1}(\phi)}
\duality{\XXX_{h_{i+1}}(\phi),\ddd f_{i+1}(\phi)}\nonumber\\
&\cdots\duality{\XXX_{h_m}(\phi),\ddd f_m(\phi)}\cdot\ddd f_i(\phi)=:\lambda(\phi)\ddd f_i(\phi).
\end{align}
When subjected to representation as a source term, $\lambda(\phi)$ is a natural candidate for a local source strength of a local source operator $\pi_\phi(\ddd f_i(\phi))$. We will (tentatively) call the local source operators of this type \df{quantum fields}. One can also introduce the antisymmetric wedge product on vectors and covectors (one forms) and define the corresponding contraction to 1-form and its source term representation in an analogous way. In particular, for $\B\iso\N^\sa$, a vector field $\XXX_k\in\T\M(\N,\B)$ can be represented in terms of a GNS fibre bundle $\H\M(\N,\B)$ as a family
\begin{equation}
\M(\N,\B)\ni\phi\mapsto\omega_{\XXX_k}(\phi)\in\BBB(\H_\phi)_\star^\sa\iso\schatten_1(\H_\phi)^\sa
\end{equation}
determined by
\begin{equation}
\duality{\XXX_k(\phi),\ddd f(\phi)}_{\T_\phi\N_\star^\sa\times\T^\ct_\phi\N_\star^\sa}=:\duality{\omega_{\XXX_k}(\phi),\pi_\phi(\ddd f(\phi))}_{\BBB(\H_\phi)_\star^\sa\times\BBB(\H_\phi)^\sa}\;\;\forall\phi\in\M(\N)\;\forall f\in\CIF(\N_\star^\sa;\RR).
\end{equation}
The GNS representation of $(0,n)$-tensor fields with $n>1$ is also possible, however it can be provided in different ways. For example, given a $(0,2)$-tensor field $(\ddd f\boxtimes\ddd k)(\phi)$, it is possible to represent it as:
\begin{align}
\pi_\phi(\ddd f(\phi))\otimes\pi_\phi(\ddd k(\phi))&\in\BBB(\H_\phi)\otimes\BBB(\H_\phi),\\
\pi_\phi(\ddd f(\phi))\otimes\pi_\phi(\ddd k(\phi))&\in\BBB(\H_\phi\otimes\H_\phi),\\
\pi_{\phi\otimes\phi}((\ddd f\boxtimes\ddd k)(\phi))&\in\BBB(\H_{\phi\otimes\phi}),\label{two.form.represent}
\end{align}
where $(\H_{\phi\otimes\phi},\pi_{\phi\otimes\phi},\Omega_{\phi\otimes\phi})$ is the GNS representation of $\N\otimes\N$ in the state $\phi\otimes\phi\in(\N\otimes\N)_\star^+$. It seems that the representation \eqref{two.form.represent} preserves most precisely the geometric content of $(0,2)$-tensor field, so we feel tempted to consider it as a preferred construction. However, this leads us to construction of a whole family of fibre bundles of Hilbert spaces over the manifold $\M(\N,\B)$. It is unclear at this stage whether this phenomenon should be considered as a virtue or as a failure. For $n$-ary tensor product $\phi\otimes\cdots\otimes\phi$ the corresponding fibre bundle of $\H_{\phi\otimes\cdots\otimes\phi}$ spaces, with $\phi$ varying over $\M(\N,\B)$, will be called a \df{$\bigotimes^n$-GNS bundle}, and denoted $\rpktarget{GNS.n}(\bigotimes^n\H)\M(\N,\B)$. For any $(0,n)$-tensor field $\ddd f_1\boxtimes\cdots\boxtimes\ddd f_n$ on $\M(\N,\B)$ with $n\leq\dim\M(\N,\B)$ there exists a unique representation
\begin{equation}
\pi_{\phi\otimes\cdots\otimes\phi}\left((\ddd f_1\boxtimes\cdots\boxtimes\ddd f_n)(\phi)\right)\in\BBB(\H_{\phi\otimes\cdots\otimes\phi}).
\end{equation}
The same is true for any \df{$n$-form field}, defined as a section of the fibre bundle of an antisymmetric wedge product $\bigwedge^n\T^\ct\B_\star$, because $\bigwedge^n\T_\phi^\ct\B_\star\subset\bigboxtimes\T^\ct_\phi\B_\star$. As a result, each smooth section of $\bigwedge^n\T^\ct\M(\N,\B)$ can be represented as a family of bounded operators,
\begin{equation}
\pi_{\phi\otimes\cdots\otimes\phi}:\bigwedge^n\T^\ct_\phi\B_\star\ni x\mapsto\pi_{\phi\otimes\cdots\otimes\phi}(x)\in\BBB(\H_{\phi\otimes\cdots\otimes\phi}),
\end{equation}
acting fibrewise over the $\bigotimes^n$-GNS bundle $(\bigotimes^n\H)\M(\N,\B)$. We will call such family a \df{quantum $n$-form field} over $\M(\N,\B)$.
The constant function on $\N_\star^\sa$, $\hat{\lambda}:\N_\star^\sa\ni\omega\mapsto\hat{\lambda}(\omega):=\lambda\in\RR$, is a geometric representation of an algebraic element of a center of $\N$, $\lambda\II\in\zentr_\N\subseteq\N$, in terms of an element of a smooth algebra, $\hat{\lambda}\in\CIF(\N_\star^\sa;\RR)$. Such function on $\N_\star^\sa$ will be called a \df{global charge}. From this it follows that, provided $\B\iso\N^\sa$, each globally constrained source strength is a global charge.
The set of quantum field one-forms in $\T^\ct\B_\star$, considered under its restriction to some quantum model $\M(\N,\B)\subseteq\N_\star^+$, can be equipped with the additional structure of a Lie algebra $\mathfrak{h}$, determined by the structure constants $\epsilon_\mathfrak{h}^{abc}$ of its adjoint representation by means of
\begin{equation}
[(\ddd_\phi f)^a,(\ddd_\phi k)^b]_\mathfrak{h}=\sum_c\epsilon_\mathfrak{h}^{abc}(\phi)(\ddd_\phi h)^c\;\;\;\forall\phi\in\M(\N,\B).
\label{structure.Lie.forms}
\end{equation}
Here $(\ddd_\phi f)^a$ denotes a Lie algebra representation map $\mathfrak{h}\ra(\T^\ct\B_\star)$ at $\phi\in\M(\N,\B)$. Representation of these forms on the fibres of the GNS bundle gives
\begin{equation}
[\pi_\phi((\ddd_\phi f)^a),\pi_\phi((\ddd_\phi k)^b)]_\mathfrak{h}=\sum_c\epsilon_\mathfrak{h}^{abc}(\phi)\pi_\phi((\ddd_\phi h)^c)\;\;\;\forall\phi\in\M(\N,\B).
\label{structure.Lie.operators}
\end{equation}
Note that the Lie algebra structure given by \eqref{structure.Lie.operators} is a priori independent of any possible principal $G$-bundle structure of $\H\M(\N,\B)$ or a principal $G$-bundle $E\ra\M(\N,\B)$ represented in terms on $\H\M(\N,\B)$ by means of a local gauge liouvillean. In order to keep the same relationship between source terms $\lambda_i(\phi)\pi_\phi((\ddd_\phi f_i)^a)$ in each fibre $\H_\phi$, one has to set $\lambda_i$ to be given by a global charge. We will call the source terms $\pi_\phi((\ddd_\phi f_i)^a)$ \df{local gauge quantum fields}, while the corresponding local source models with global source strengths $\lambda_i$ will be called \df{local gauge quantum field models}. If an extended liouvillean model $(\M(\N,\B),\mathbf{A},\aaa,\lambda_1,H_1,\ldots,\lambda_n,H_n)$ is equipped with an affine connection $\nabla$ such that $(\M(\N,\B),\mathbf{A},\aaa,\nabla)$ is a gauge geodesic propagation model, while $(\M(\N,\B),\lambda_1,H_1,\ldots,\lambda_n,H_n)$ is a quantum field model with $H_i(\phi)=\pi_\phi((\ddd_\phi f_i)^{a_i})$ $\forall i\in\{1,\ldots,n\}$, then the $(4+2n)$-tuple
\begin{equation}
(\M(\N),\mathbf{A},\aaa,\nabla,\lambda_1,(\ddd f_1)^{a_1},\ldots,\lambda_n,(\ddd f_n)^{a_n})
\end{equation}
can be called a `quantum field model with gauge geodesic propagation'. If $G$ and $\aaa$ are chosen to agree with the Wigner classification theorem, then such model describes a family of quantum fields together with a quantum particle geodesic propagation. However, while the availability of these constructions is a quite remarkable fact, it is yet unclear how they could be translated to the usual objects of quantum field theory.
\ifvargaugecompile
\subsection{Local liouvillean instruments and correlation functions}
\else
\subsubsection{Local liouvillean instruments and correlation functions}
\fi
The temporal evolution of a local liouvillean model is completely described by the fibrewise evolution
\begin{equation}
\xi(\omega,t)=\ee^{-\ii t\Lcal(\omega,t)}\Omega_\omega,
\label{lagrangean.movement}
\end{equation}
which is a generalisation of \eqref{liovillean.movement} taking local gauge and local source structures on $\M(\N)\subseteq\N^+_{\star0}$ into account. The corresponding propagator (transition amplitude) between initial state $\omega$ and the final state $\phi$ reads
\begin{equation}
\s{\Omega_\phi,V_{\phi,\omega}\xi(\omega,t)}_{\H_\phi}=\s{\Omega_\phi,V_{\phi,\omega}\ee^{-\ii t \Lcal(\omega,t)}\Omega_\omega}_{\H_\phi}.
\label{liouville.propagator}
\end{equation}
If $\Lcal$ is determined only by a given Poisson system $(\M(\N,\B),h)$ with $\DF_\omega h\in\N^\sa$ $\forall\omega\in\M(\N,\B)$, then the propagator \eqref{liouville.propagator} reads
\begin{equation}
\s{\Omega_\phi,V_{\phi,\omega}\ee^{-\ii t\pi_\omega(\DF_\omega(h))}}_{\H_\phi}
=\s{\Omega_\phi,\ee^{\ii t\log(J_\phi J_{\phi,\omega})}\ee^{-\ii t\pi_\omega(\DF_\omega(h))}\Omega_\omega}_{\H_\phi}.
\label{liouv.prop}
\end{equation}
We will now show that this evolution can be expressed as an instrument acting on $\N_\star^+$ and parametrised by $t\in\RR$. Let us choose some $\psi\in\N^+_{\star0}$. Then the elements of each fibre of the GNS bundle over $\N^+_{\star0}$ can be uniquely mapped to $\H_\psi$ by means of the standard unitary transition operator $V_{\psi,\omega}=J_\psi J_{\psi,\omega}=J_{\psi,\omega}J_\omega$, which preserves the positive cones
\ifvargaugecompile
(see Section \ref{standard.liouvilleans.section}).
\else
(see Section \ref{relative.modular.theory.section}).
\fi
Hence, at each value of $t\in\RR$ and at each $\psi\in\N^+_{\star0}$, the set
\begin{equation}
\bigcup_{\omega\in\M(\N)}\{V_{\psi,\omega}\xi(\omega,t)\}\subseteq\H_\psi^+
\end{equation}
represents completely the evolution in a fibre bundle $\H\M(\N)$ that is defined by means of a local liouvillean operator. Using the bijective norm continuous homomorphism $\stdembed^\natural_\psi:\H_\psi^+\ra\N_\star^+$ (defined as $\stdembed^\natural_\pi$ for $\pi=\pi_\psi$), we can represent the mapping
\begin{equation}
\RR\ni t\mapsto\bigcup_{\omega\in\M(\N)}\left\{V_{\psi,\omega}\xi(\omega,t)\right\}\subseteq\H_\psi^+
\end{equation}
as a temporal evolution of subsets of $\N_\star^+$,
\begin{equation}
t\mapsto\stdembed^\natural_\psi\left(\bigcup_{\omega\in\M(\N)}\left\{V_{\psi,\omega}\xi(\omega,t)\right\}\right)\subseteq\N_\star^+.
\end{equation}
\ifvargaugecompile
The
\else
It follows that the
\fi
family of mappings
\begin{equation}
\RR\ni t\mapsto\left\{\III_{\Lcal,\psi}(t):\M(\N)\ni\omega\mapsto(\III_{\Lcal,\psi}(t))(\omega):=\stdembed^\natural_\psi(V_{\psi,\omega}\ee^{-\ii t\Lcal(\omega,t)}\Omega_\omega)\in\N_\star^+\right\}
\end{equation}
\ifvargaugecompile
\else
is an instrument, which
\fi
will be called a \df{local liouvillean instrument} (this name may be a bit deceiving, because of nonlocality inherent in the $V_{\phi,\omega}$ operation. The uniqueness of standard unitary transition for each pair of elements of $\N^+_{\star0}$ together with bijectivity of $\stdembed^\natural_\psi$ implies
\begin{equation}
\stdembed_\psi^\natural V_{\psi,\omega}\xi=\stdembed_\varphi^\natural V_{\varphi,\omega}\xi\;\;\;\forall\xi\in\H_\omega\;\forall\varphi,\omega\in\N^+_{\star0},
\end{equation}
hence an extended liouvillean instrument does not depend on the choice of $\psi$. In what follows we will denote it by $\rpktarget{LIOUV.INSTR}\III_{\Lcal(t)}$. Due to bijectivity of $\stdembed^\natural_\psi$ there is an equivalence between the evolution generated on the GNS bundle by the family of extended liouvillean operators and the evolution generated on $\N_\star^+$ by the extended liouvillean instrument. Thus, one can consider the extended liouvillean operator structure over a given model (including local gauge and local source structures, as well as the isometry $\alpha_\star$ or a Poisson flow $w^h$) as \textit{auxiliary} tools allowing to define suitable extended liouvillean instrument, but otherwise devoid of any foundational meaning.
\ifvargaugecompile
\else
\fi
The GNS bundle allows to construct the $n$-point correlation functions, whenever all quantum states under consideration are faithful. Let $\N$ be a W$^*$-algebra, let $\phi_0,\phi_1,\ldots,\phi_n\in\N^+_{\star0}$ and let $x_1,\ldots,x_n\in\N$. Then we can define the \df{time independent $n$-point correlation function} as
\begin{equation}
\e{x_1(\phi_1)\cdot\ldots\cdot x_n(\phi_n)}_{\phi_0}:=\s{\Omega_{\phi_0},V_{\phi_0,\phi_1}\pi_{\phi_1}(x_1)\cdots V_{\phi_{n-1},\phi_n}\pi_{\phi_n}(x_n)V_{\phi_n,\phi_0}\Omega_{\phi_0}}_{\H_{\phi_0}}.
\label{time.independent.corr.fun}
\end{equation}
If $\M(\N)\subseteq\N^+_{\star0}$ is $m$-dimensional, $\phi_0,\phi_1,\ldots,\phi_n\in\M(\N)$ and $\theta:\M(\N)\ra\RR^m$ is a coordinate system on $\M(\N)$, then \eqref{time.independent.corr.fun} can be expressed in terms of $\theta$ as
\begin{align}
\e{x_1(\theta_1)\cdot\ldots\cdot x_n(\theta_n)}_{\phi(\theta_0)}=&\left\langle\Omega_{\phi_0},V_{\phi(\theta_0),\phi(\theta_1)}\pi_{\phi(\theta_1)}(x_1)\cdots\right.\nonumber\\&\left.\cdots V_{\phi(\theta_{n-1}),\phi(\theta_n)}\pi_{\phi(\theta_n)}(x_n)V_{\phi(\theta_n),\phi(\theta_0)}\Omega_{\phi(\theta_0)}\right\rangle_{\H_{\phi(\theta_0)}},
\end{align}
where $\theta_0,\theta_1,\ldots,\theta_n\in\RR^m$ and $\phi_i=\phi(\theta_i):=\theta^{-1}(\theta_i)$. Constructions provided in this
\ifvargaugecompile
paper
\else
section
\fi
allow us to define also the \df{time dependent $n$-point correlation functions} as
\begin{align}
\e{x_1(\phi_1,t_1)\cdot\ldots\cdot x_n(\phi_n,t_n)}_{\phi_0}:=&\left\langle\Omega_{\phi_0},V_{\phi_0,\phi_1}\ee^{+\ii t\Lcal(\phi_1,t_1)}\pi_{\phi_1}(x_1)\ee^{-\ii t\Lcal(\phi_1,t_1)}\cdots\right.\nonumber\\&\left.\cdots V_{\phi_{n-1},\phi_n}\ee^{+\ii t\Lcal(\phi_n,t_n)}\pi_{\phi_n}(x_n)\ee^{-\ii t\Lcal(\phi_n,t_n)}V_{\phi_n,\phi_0}\Omega_{\phi_0}\right\rangle_{\H_{\phi_0}},
\label{time.dependent.corr.fun}
\end{align}
where $t_1,\ldots,t_n\in\RR$ and $\Lcal(\phi,t)$ is an extended liouvillean. If a reformulation in terms of a coordinate system $\theta$ (defined above) is possible, then \eqref{time.dependent.corr.fun} can be expressed as
\begin{align}
\e{x_1(\theta_1,t_1)\cdot\ldots\cdot x_n(\theta_n,t_n)}_{\phi(\theta_0)}
=\left\langle\Omega_{\phi(\theta_0)},V_{\phi(\theta_0),\phi(\theta_1)}
\ee^{+\ii t\Lcal(\phi(\theta_1),t_1)}\pi_{\phi(\theta_1)}(x_1)\ee^{-\ii t\Lcal(\phi(\theta_1),t_1)}\cdots\right.\nonumber\\\left. \cdots V_{\phi(\theta_{n-1}),\phi(\theta_n)}\ee^{+\ii t\Lcal(\phi(\theta_n),t_n)}\pi_{\phi(\theta_n)}(x_n)\ee^{-\ii t\Lcal(\phi(\theta_n),t_n)}V_{\phi(\theta_n),\phi(\theta_0)}\Omega_{\phi(\theta_0)}\right\rangle_{\H_{\phi(\theta_0)}}
\end{align}
Due to different values taken by the components of $\Lcal$ operator at different points $(\phi,t)$, the evolution determined by \eqref{lagrangean.movement} and \eqref{time.dependent.corr.fun} does not have to be unitary. It is so only when the dynamics and perturbations in all fibres of the Hilbert bundle are the same.
Equation \eqref{time.dependent.corr.fun} describes how the predictive time dependent content of a quantum model $\M(\N)$ can be determined using the representation of geometric structures on $\M(\N)$ in terms of the algebraic structures on the GNS bundle. However, let us note that this equation is only an example of the variety of possible definitions of the time dependent correlation functions that could be constructed with the help of local liouvilleans and the GNS bundle. Moreover, one could carry the above constructions also for MCP bundle, obtaining different quantitative results. The identification of the proper construction should be based on a more detailed analysis of backwards compatibility with other approaches. In next Section we will approach the derivation of the path integral analogue of the propagator \eqref{liouville.propagator}.
\ifvargaugecompile
\else
\subsubsection{Quantum lagrangeans}
\fi
\section{Quantum histories\label{quantum.histories.section}}
In order to solve the problems of `measurement' and `time' in quantum theory Griffiths \cite{Griffiths:1984}, Omn\`{e}s \cite{Omnes:1988:I,Omnes:1988:II,Omnes:1988:III,Omnes:1989:IV,Omnes:1990,Omnes:1992,Omnes:1994}, and Gell-Mann and Hartle \cite{GellMann:Hartle:1990:I,GellMann:Hartle:1990:II,GellMann:Hartle:1990:III,Hartle:1991:I,Hartle:1991:II,GellMann:Hartle:1993,Hartle:1995} have developed `consistent histories' approach to quantum theory. Isham and Linden \cite{Isham:1994,Isham:Linden:1994,Isham:Linden:1995,Isham:Linden:SS:1997,Isham:Linden:Schreckenberg:1994} have proposed a modification of this approach, called the (continuous-time) `history projection operator' approach, which was developed later by Savvidou and Anastopoulos \cite{Savvidou:1999,Savvidou:1999:PhD,Anastopoulos:Savvidou:2002,Anastopoulos:2001,Anastopoulos:Savvidou:2003}.
In Section \ref{histories.propositions.section} we will recall the elementary mathematical structure of the Isham--Linden approach, in Section \ref{Savvidou.action.operator} we will discuss Savvidou's construction \cite{Savvidou:1999,Savvidou:1999:PhD} of an action operator within this setting, while in Sections \ref{geometric.phase.histories} and \ref{Hilbert.space.geometry.path.integrals} we will follow the Anastopoulous--Savvidou analysis of the relationship of this framework with the geometric structures on the spaces of pure quantum states, and the Daubechies--Klauder \cite{Daubechies:Klauder:1985,Klauder:1988,Watson:Klauder:2002} continous-time regularised coherent states phase space path integration. Sections \ref{histories.propositions.section}-\ref{Hilbert.space.geometry.path.integrals} do not contain new results. Their aim is to lead us to a refined geometric perspective on the relationship between the Daubechies--Klauder formula and the local liouvilleans. In Section \ref{algebraic.action.operator.section} we will apply the local liouvillean approach to the Falcone--Takesaki construction of noncommutative flow of weights in order to construct the algebraic analogue of Savvidou's histories action operator. The discussion of limitations of this construction in the face of the results of Section \ref{from.local.gauge.to.lagrangeans.section} will lead us to construction of W$^*$-geometric generalisation of the Daubechies--Klauder path integration in Section \ref{algebraic.quantum.histories}.
\subsection{Propositions and evolution\label{histories.propositions.section}}
The starting point of the history projection operator approach is consideration of a \textit{history} $\varpi$ of abstract `propositions' $(P_{t_1},{P}_{t_2},\ldots,{P}_{t_n})$ about a quantum theoretic model, assigned to an ordered finite sequence $(t_1,t_2,\ldots,t_n)$, where $t_1<t_2<\ldots<t_n$, and $t$ is interpreted as `time' parameter. Following the ideas of Mittelstaedt \cite{Mittelstaedt:1977} and Stachow \cite{Stachow:1980,Stachow:1981}, Isham \cite{Isham:1994} has proposed to specify this entity by the projection operator
\begin{equation}
P_\varpi:=P_{t_1}\otimes P_{t_2}\otimes\ldots\otimes P_{t_n},
\end{equation}
acting on the Hilbert space
\begin{equation}
{\cal{V}}_n:=\bigotimes_{i=1}^n\H_{t_i}:=\H_{t_1}\otimes\H_{t_2}\otimes\ldots\otimes\H_{t_n},
\end{equation}
where ${P}_{t_i}$ is a projection operator in the $n$-th copy of the Hilbert space $\H_{t_i}:=\H$ of a given quantum model. The history $\varpi$ of nonunitary propositions as well as the description of the unitary dynamics are contained in the \textit{class operator} on ${\cal{V}}_n$, defined as \cite{GellMann:Hartle:1990:I}
\begin{equation}
C_\varpi:=U(t_0,t_1){P}_{t_1}U(t_1,t_2){P}_{t_2}\cdots U(t_{n-1},t_n){P}_{t_n}U(t_n,t_0),
\end{equation}
where $U(t_i,t_{i+1})=\ee^{-\ii\int_{t_{i}}^{t_{i+1}}\dd tH}$ are unitary evolution operators between times $t_i$ and $t_{i+1}$, acting on the Hilbert space ${\H}_{t_{i+1}}$ and generated by a self-adjoint hamiltonian operator $H$. For a dynamics generated by the hamiltonian $H$ with an initial state described by the density operator $\rho$, \textit{the probability of a history $\varpi$} is defined as
\begin{equation}
p(\varpi;\rho,H):=\tr_{{\cal{V}}_n}(C^*_\varpi\rho C_\varpi).
\end{equation}
Using this equation, for two given histories $\varpi$ and $\vartheta$, one defines the \textit{histories functional}\footnote{For historical reasons, this object is usually called `decoherence functional'. However, such name suggests that the quantum histories formalism necessary involves the `decoherence approach to quantum measurement' semantics, which is not true. For this reason we choose to change the name of this mathematical object to much more neutral with respect to possible semantics.} \cite{Griffiths:1984}
\begin{equation}
\hf_{\rho,H}:\BBB({\cal{V}}_n)\times\BBB({\cal{V}}_n)\ni(P_\varpi,P_\vartheta)\mapsto\tr_{{\cal{V}}_n}(C^*_\varpi\rho C_\vartheta)\in\CC,
\label{decoherencefunctional}
\end{equation}
which, by definition, depends on $\rho$ and $H$. It satisfies, for $P_\varpi\leq\II-P_{\varkappa}$,
\begin{align}
\hf(P_\varpi,P_\varpi)&\geq0,\\
\hf(P_\varpi,P_\vartheta)&=\hf(P_\vartheta,P_\varpi)^*,\\
\hf(0,P_\varpi)&=0,\\
\hf(\II,\II)&=1,\\
\hf(P_\varpi+P_{\varkappa},P_\vartheta)&=\hf(P_\varpi,P_\vartheta)+\hf(P_{\varkappa},P_\vartheta).
\label{decoherence.rep.prop}
\end{align}
Description of a quantum theoretic model provided in terms of the class operator and histories functional is intended to serve as a single replacement for dualistic description of temporal behaviour of model in terms of Schr\"{o}dinger's
and von Neumann--L\"{u}ders'
equations.
The \textit{Born--Jordan--Dirac--Heisenberg} (BJDH) \textit{algebra} \cite{Born:Jordan:1925,Born:Heisenberg:Jordan:1926,Dirac:1925}, generated by the \textit{canonical commutation relations}
\begin{equation}
[\qq,\qq]=0,\;\;\;
[\pp,\pp]=0,\;\;\;
[\qq,\pp]=\ii\II,
\label{BJDH.CCR}
\end{equation}
is extended in the quantum histories framework to an algebra generated by the relations:
\begin{equation}
[\qq_{t_i},\qq_{t_j}]=0,\;\;\;
[\pp_{t_i},\pp_{t_j}]=0,\;\;\;
[\qq_{t_j},\pp_{t_k}]=\ii\delta_{jk}\II,
\label{heisenberg}
\end{equation}
where the operators $\qq_{t_i}$ and $\pp_{t_i}$ are considered as operators defined in the Schr\"{o}dinger picture for different moments $t_i$ of time, acting on the Hilbert space $\H_{t_i}$, while $\II$ is a unit element of that algebra. In order to formulate an extension of this formalism to the case of a continuous time $t\in\RR$, Isham and Linden \cite{Isham:Linden:1995} have changed the above relations to the form of the so-called \textit{history algebra}:
\begin{equation}
\begin{array}{c}
[\qq_f,\qq_g]=0,\;\;
[\pp_f,\pp_g]=0,\;\;
[\qq_f,\pp_g]=\ii\int_{-\infty}^{+\infty}\dd t f(t)g(t)\II,
\end{array}
\label{histories.cr}
\end{equation}
where $\qq$ and $\pp$ are operator valued distributions, $f,g\in L_2(\RR,\dd t)$, $\pp_f:=\pp(f)$, $\qq_f:=\qq(f)$.\footnote{Correspondingly, in the histories approach to quantum field theory one considers the `field' operator-valued distributions $\qq$ and $\pp$ which act on a subspace of $L_2(\RR^3,\dd^3x)$, where the parameter $\vec{x}\in\RR^3$ is interpreted as representing a three-dimensional `space', and extends the canonical commutation relations at single point of time with the additional dependence on time dimension handled by the histories algebra \cite{Isham:Linden:1995,Isham:Linden:SS:1997,Savvidou:1999}. When the dependence on the functions in $L_2(\RR^3,\dd^3x)$ is made implicit, these relations read
\begin{equation}
[\qq_{t_1}(\vec{x}_1),\qq_{t_2}(\vec{x}_2)]=0,\;\;[\pp_{t_1}(\vec{x}_1),\pp_{t_2}(\vec{x}_2)]=0,\;\;
[\qq_{t_1}(\vec{x}_1),\pp_{t_2}(\vec{x}_2)]=\ii\delta(t_1-t_2)\delta^3(\vec{x}_1-\vec{x}_2)\II.
\end{equation}
This means that the three-dimensional histories commutation relations are actually three-plus-one-dimensional canonical commutation relations.}
Isham and Linden \cite{Isham:Linden:1994,Isham:Linden:SS:1997} (see also \cite{Savvidou:1999}), have shown that these commutation relations may be represented on the Hilbert \textit{continuous history space}
\begin{equation}
{\cal{V}}:=\bigotimes_{t\in\RR}\H_t:=\bigotimes_{t\in\RR}({L_2}\left(\RR,\dd x)\right)_t.
\end{equation}
The `continuous tensor product' space $\bigotimes_{t\in\RR}\H_t$ is defined to be the symmetric Fock--Cook space \cite{Fok:1932,Cook:1953}
\begin{equation}
\FFF[\H]=\FFF[L_2(\RR,\dd x)]=\bigoplus_{n=0}^\infty\sym_n\left(\bigotimes_{k=0}^{n}L_2(\RR,\dd x)\right),
\end{equation}
where $\bigotimes_0L_2(\RR,\dd x):=\CC$,
\begin{equation}
\sym_n:\xi_1\otimes\ldots\otimes\xi_n\mapsto\frac{1}{\sqrt{n!}}\sum_{s\in\SS(n)}\xi_{s(1)}\otimes\ldots\otimes\xi_{s(n)}
\end{equation}
and $\SS(n)$ is the group of permutations of the set $\{1,\ldots,n\}$. The history of $(t_1,\ldots,t_n)$ is represented by the vector of $\FFF[\H]$ generated by the action of $n$ `creation operators'
\begin{equation}
\begin{array}{c} b_{t_i}^*(f):\FFF[\H]\ni\sym_n(\xi_1\otimes\ldots\otimes\xi_n)
\mapsto\sqrt{n+1}\sym_{n+1}(f\otimes\xi_1\otimes\ldots\otimes\xi_n)\in\FFF[\H],
\end{array}
\end{equation}
which, together with the `annihilation operators'
\begin{equation}
\begin{array}{c} b_{t_i}(f):\FFF[\H]\ni\sym_n(\xi_1\otimes\ldots\otimes\xi_n)
\mapsto\frac{1}{\sqrt{n}}\sum^n_{k=1}f\cdot\sym_{n-1}(\xi_1\otimes\ldots\otimes\xi_{k-1}\otimes\xi_{k+1}
\otimes\ldots\otimes\xi_n)\in\FFF[\H],
\end{array}
\end{equation}
satisfy
\begin{equation}
\begin{array}{c}
{}[b_{t_i}(f),b_{t_j}(g)]=0=[b^*_{t_i}(f),b^*_{t_j}(g)],\;\;
{}[b_{t_j}(f),b^*_{t_k}(g)]=\int_{-\infty}^{+\infty}\dd tf(t)g(t)\II
\end{array}
\end{equation}
on the common domain $\D\subset\H$. Moreover, it is assumed that there exists a unit vector $\Omega\in\D:=\Span\{b^*_t(f_1)\cdots b^*_t(f_n)\Omega\mid\forall f_1,\ldots,f_n\in\H\}$ such that
\begin{equation}
b_t(f)\Omega=0\;\forall t\in\RR\;\forall f\in\H.
\end{equation}
These assumptions define the Isham--Linden representation of the history algebra (\ref{histories.cr}) to be the Fock--Cook representation. The spectral projectors of this representation of histories algebra are interpreted \cite{Isham:Linden:1995} as propositions about the temporal histories of a given quantum theoretic model.
For a given hamiltonian $H_t$, the self-adjoint histories hamiltonian operator in the Schr\"{o}dinger picture is defined as
\begin{equation}
H_\kappa:=\int_{-\infty}^{+\infty}\dd t\kappa(t) H_t,
\label{Hkappa}
\end{equation}
where $\kappa(t)\in L_2(\RR,\dd t)$ is a function which `smears' $H_t$ in time. The histories algebra generates the commutation relations with this hamiltonian.
The Araki theorem \cite{Araki:1960} states the existence and uniqueness of hamiltonian operator in the Fock--Cook representation if this operator (if unsmeared) has a form
\begin{equation}
H_t(\pp_t,\qq_t)=\frac{1}{2}\int\dd x \pp_t^2(x)+\tilde{H}_t(\qq_t).
\end{equation}
For example, for a given `harmonic oscillator' hamiltonian operator $H_t=\frac{\pp_t^2}{2m}+\frac{m\mathrm{w}^2}{2}\qq_t^2$ acting on the Fock--Cook space ${\cal{V}}$, the representation of the history algebra is constructed through the `annihilation operator' $b_t$, which takes the form
\begin{equation}
b_t=\sqrt{\frac{m\mathrm{w}}{2}}\qq_t+\ii\sqrt{\frac{1}{2m\mathrm{w}}}\pp_t,
\end{equation}
with the commutation relations
\begin{equation}
[b_{t_i},b_{t_j}]=0,\;\;[b_{t_i},b_{t_j}^*]=\delta(t_i-t_j)\II.
\end{equation}
By the Araki theorem, the Fock--Cook representation of the history algebra in ${\cal{V}}$ is uniquely selected by the requirement that the operator $H_\kappa$ (\ref{Hkappa}) exists in this representation. In nonsmeared version, $H_t=\mathrm{w}b_t^*b_t$. This leads to the following commutation relations \cite{Savvidou:1999}:
\begin{equation}
[H_\kappa,\qq_f]=-\frac{\ii}{m}\pp_{\kappa f},\;\;\;[H_\kappa,\pp_f]=\ii\mathrm{w}^2\qq_{\kappa f},\;\;\;[H_\kappa,H_{\kappa'}]=0.
\end{equation}
Anastopoulos \cite{Anastopoulos:2001} has shown that the construction of the continuous history Hilbert space ${\cal{V}}$ and the representation of the history algebra in ${\cal{V}}$ can be provided also for nonquadratic hamiltonians using the coherent states representation \cite{Schroedinger:1926:Naturwiss,Glauber:1963}. However, such representation lacks any characterisation of its uniqueness.
\subsection{Savvidou's action operator\label{Savvidou.action.operator}}
The important property of the histories approach, discovered by Savvidou \cite{Savvidou:1999}, is the existence of the self-adjoint \textit{quantum action operator}, acting on ${\cal{V}}$ and defined as
\begin{equation}
S_{\lambda,\kappa}:=\int_{-\infty}^{+\infty}\dd t(\lambda(t)\pp_t\dot{\qq}_t-\kappa(t)H_t),
\label{quantumactionoperator}
\end{equation}
by an analogy to a Hamilton--Jacobi action functional in classical mechanics theory
\begin{equation}
S_{\mathrm{HJ}}=\int_{t_2}^{t_1}dt(\pp_\Gamma(t)\dot{\qq}_\Gamma(t)-H_\Gamma(t)),
\label{Hamilton.Jacobi.action}
\end{equation}
where $\pp_\Gamma,\qq_\Gamma,H_\Gamma\in C^\infty(\Gamma)$, and $\Gamma$ is a classical mechanics phase space. In both these equations, the dot symbol denotes the differentiation $\frac{\dd}{\dd s}$ with respect to to time parameter $s$ of the evolution generated by hamiltonian, that is (in the quantum theory)
\begin{equation}
\qq_t:=\qq_{t}(s):=\ee^{\ii sH_t}\qq_t\ee^{-\ii sH_t},
\end{equation}
\begin{equation}
\dot{\qq}_t:=\frac{\dd}{\dd s}\qq_{t}(s).
\end{equation}
Savvidou has shown that there also exists the \textit{Liouville operator}
\begin{equation}
V:=\int_{-\infty}^{+\infty}\dd t\widetilde{V}_t:=\int_{-\infty}^{+\infty}\dd t(\pp_t\dot{\qq}_t)
\end{equation}
which is self-adjoint on ${\cal{V}}$. Hence, for $\lambda(t)\equiv 1$, one may express the action operator as
\begin{equation}
S_\kappa=V-H_\kappa=\int_{-\infty}^{+\infty}dt\pp_t\dot{\qq}_t-\int_{-\infty}^{+\infty}\dd t\kappa(t)H_t
\label{action}
\end{equation}
with the following commutation relations:
\begin{equation}
[S_\kappa,H_{\kappa'}]=\ii H_{\dot{\kappa}'},\;\;[S_\kappa,V]=-\ii H_{\dot{\kappa}},\;\;[V,H_\kappa]=-\ii H_\kappa.
\end{equation}
For $\kappa(t)\equiv 1$ the histories quantum theory reduces to ordinary quantum theory, which (for $H:=\int_{-\infty}^{+\infty}\dd t H_t$) is reflected in the commutators $[V,H]=0$ and $[V,S]=0$.
The operator $V$ acts on $b_t$ in the following way \cite{Savvidou:1999}:
\begin{equation}
\ee^{\ii rV}b_{f(t)}\ee^{-\ii rV}=b_{f(t+r)}.
\end{equation}
Moving to the Heisenberg picture, one can compare the action of $V$, $H_t$ and $S$:
\begin{align}
\ee^{\ii rV}b_{t}(s)\ee^{-\ii rV}&=b_{t+r}(s),\label{V.hist.auto}\\
\ee^{\ii rH_t}b_{t}(s)\ee^{-\ii rH_t}&=b_{t}(s+r),\label{H.hist.auto}\\
\ee^{\ii r S}b_{t}(s)\ee^{-\ii r S}&=b_{t+r}(s+r),\label{S.hist.auto}
\end{align}
where the `smeared' operator $S_\kappa$ acts by an automorphism
\begin{equation}
\ee^{\ii rS_\kappa}b_{f(t)}\ee^{-\ii rS_\kappa}=b_{\Sigma_r(f)},
\end{equation}
where $\Sigma_r$ is an unitary operator acting on $\zeta\in L_2(\RR,\dd t)$ by
\begin{equation}
(\Sigma_r\zeta)(t):=\ee^{-\ii\mathrm{w}\int_{t}^{t+r}\dd r'\kappa(t+r')}\zeta(t+r).
\end{equation}
For not smeared $b_t$ this can be formally written as
\begin{equation}
\ee^{\ii rS_\kappa}b_t\ee^{-\ii rS_\kappa}=\ee^{-\ii\mathrm{w}\int_t^{t+r}\dd r'\kappa(t+r')+r\frac{\dd}{\dd t}}b_t,
\label{s.h.auto2}
\end{equation}
where the self-adjoint generator $S_\kappa$ on $\FFF[L_2(\RR,\dd t)]$ corresponds to an action of self-adjoint $\sigma_\kappa$ on $L_2(\RR,\dd t)$ given by
\begin{equation}
\sigma_\kappa\zeta(t):=-\left(-\mathrm{w}\kappa(t)-\ii\frac{\dd}{\dd t}\right)\zeta(t).
\label{s.h.auto}
\end{equation}
The map $\RR\ni r\mapsto \ee^{\ii rS_\kappa}$ is a weakly continuous representation of a one-parameter family of unitary operators. In the same way the action of the automorphism generated by the Liouville operator $V$
\begin{equation}
\ee^{\ii rV}b_{f(t)}\ee^{-\ii rV}=b_{f(t+r)}
\end{equation}
corresponds to the action of
\begin{equation}
(v_r\zeta)(t):=\ee^{r\frac{\dd}{\dd t}}\zeta(t)=\zeta(t+r)
\label{l.h.auto}
\end{equation}
on $\zeta\in L_2(\RR,\dd t)$.
Hence, $V$ transforms $b_t$ from time $t$, related with the Hilbert space $\H_t$, to time $t+r$, related with the Hilbert space $\H_{t+r}$ (strictly speaking, $V$ transforms the support of the operator valued distribution). This is, by definition, purely kinematical operation, which does not depend on the hamiltonian $H_t$. On the other hand, $H_t$ generates the unitary evolution of the system in the single space $\H_t\subset{\cal{V}}$. The action operator (\ref{action}) joins together these two types of transformations. This is in some sense analogous to the Hamilton--Jacobi formulation of classical mechanics, in which the Hamilton--Jacobi action functional (\ref{Hamilton.Jacobi.action}) is the generator of a canonical transformation of the classical mechanical model from one time to another.
Savvidou \cite{Savvidou:1999} suggests that these two operators ($V$ and $H_\kappa$) are related respectively with two different types of time evolution: the nonunitary `reduction' (L\"{u}ders' rule) related with subsequent propositions $P_t$, and the ordinary hamiltonian evolution (Schr\"{o}dinger's equation) given by operators $\ee^{\ii sH_t}$. However, the projection operators $P_t$, class operators $C_\varpi$, and histories functional $\hf_{\rho,H}$ were used neither in derivation of the kinematical evolution related with the Liouville operator $V$, nor in derivation of the quantum action operator $S_\kappa$. This is reflected in the apparent unitary character of the corresponding temporal evolutions generated by $V$ and $S_\kappa$. Hence, so far this suggestion has been ungrounded. In the next two subsections we will follow Anastopoulos and Savvidou on their way of reintroduction of nonunitary elements in quantum histories formalism.
\subsection{Geometric phase as a trace of a history\label{geometric.phase.histories}}
There exists a class of quantities in quantum theory, which do not correspond to any element of an algebra of operators, but are nevertheless very closely related to quantitative results of experimental procedures. One of the important representatives of this class is the \textit{geometric phase}. It reflects the geometric structure of the Hilbert space. Mathematically, it is defined as a holonomy of the Berry connection on the Hopf bundle \cite{Simon:1983}. The Hopf bundle is a $U(1)$ principal bundle of the Hilbert space $\H$ over the projective Hilbert space $\PH\subset\H$. In other words, it is a principal bundle of the subspaces $\H_\PP$ of vectors in $\H$ over the space of generating rays of $\H$. The \textit{Berry connection} $\nabla^\PP$ is defined as a $U(1)$ connection $1$-form on the Hopf bundle, induced naturally by an inner product on the Hilbert space $\H$, and given in the coordinate-free form by
\begin{equation}
\nabla^\PP:=\s{\cdot,\ddd\cdot}_\H:\H\times\H\ni(\zeta,\xi)\mapsto\ii\s{\zeta,\ddd\xi}_\H\in\CC.
\end{equation}
The geometric phase (called also the Pancharatnam--Berry phase \cite{Pancharatnam:1956,Berry:1984}) is then defined as
\begin{equation}
\ee^{\ii\theta[\gamma]}:=\ee^{-\int_{\zeta\in\pathgamma}\s{\zeta,\ddd\zeta}_\H},
\label{geom.phase.def}
\end{equation}
where $\pathgamma$ is a closed path in $\PH$. We will denote a path generated by family of vectors $\RR\ni t\mapsto\zeta(t)\in\H$ by $\zeta(\cdot)$. In case of open paths $\zeta(\cdot)$ in $\PH$ it was shown in \cite{Aharonov:Anandan:1987} and \cite{Samuel:Bhandari:1988} that the geometric phase (called also the Aharonov--Anandan phase) is given by
\begin{equation}
\ee^{\ii\theta[\pathgamma]}=\ee^{-\int_{t_0}^{t_1}\dd t\s{\zeta(t),\frac{\dd}{\dd t}\zeta(t)}_\H}\s{\zeta_{t_0},\zeta_{t_1}}_\H.
\label{generalberryphase}
\end{equation}
Consider now an initial vector $\zeta(t=0)$, the final vector $\zeta(t=r)$, equal to the initial one up to phase, and the unitary time evolution $U(s)$ on $\H$, described by the solution of the Schr\"{o}dinger equation with the hamiltonian $H$,
\begin{equation}
U(r):\zeta(t=0)\mapsto\zeta(t=r):=\ee^{-\ii rH}\zeta(t=0),
\end{equation}
which acts along a loop $\pathgamma(t)$, $t\in[0,r]$, on the space $\PH$. The phase on the Hopf bundle is then transformed into
\begin{equation}
\begin{array}{l}
\ee^{\int_0^r\dd t\s{\zeta(t),(-\frac{\dd}{\dd t}-\ii H)\zeta(t)}_\H}=
\ee^{-\int_0^r \dd t\s{\zeta,\frac{\dd}{\dd t}\zeta}_\H}\ee^{-\ii\int_0^r\dd t\s{\zeta(t),H\zeta(t)}_\H}=
\ee^{\ii\theta[\pathgamma]}\ee^{-\ii\int_0^r\dd t\s{\zeta(t),H\zeta(t)}_\H}.
\end{array}
\label{hopfbundlephase}
\end{equation}
The first term, given by the geometric phase, does not refer to dynamics generated by the hamiltonian $H$, and reflects purely geometric structure of the kinematical Hilbert space $\H$.
Following Anastopoulos and Savvidou \cite{Anastopoulos:Savvidou:2002}, one can analyse the geometric phase from the perspective of quantum histories. Consider first a quantum model with a hamiltonian $H=0$. For a given history $\varpi$ of the projections $({P}_{t_0},{P}_{t_1},\ldots,{P}_{t_n})$, where ${P}_{t_i}$ is a projection onto one dimensional vector subspace of $\H$ spanned by $\zeta_{t_i}$, the trace of the class operator is
\begin{equation}
\tr_{{\cal{V}}_n}(C_\varpi)=\s{\zeta_{t_0},\zeta_{t_n}}_\H\s{\zeta_{t_1},\zeta_{t_0}}_\H\s{\zeta_{t_2},\zeta_{t_1}}_\H\cdots\s{\zeta_{t_n},\zeta_{t_{n-1}}}_\H,
\label{trace.class.operator.omega}
\end{equation}
while the corresponding histories functional
\begin{equation}
\begin{array}{c}
\hf_{|\zeta_{t_0}\rangle\langle\zeta_{t_0}|,H=0}(P_{\zeta(\cdot)},P_{\xi(\cdot)})=\s{\zeta_{t_n},\zeta_{t_{n-1}}}_\H\cdots\s{\zeta_{t_1},\zeta_{t_0}}_\H
\cdot\s{\zeta_{t_0},\xi_{\tilde{t}_1}}_\H\cdots\s{\xi_{\tilde{t}_{m-1}},\xi_{\tilde{t}_m}}_\H
\end{array}
\label{eq.Bargmann}
\end{equation}
is a $(n+m+1)$ Bargmann invariant \cite{Bargmann:1964,Simon:Mukunda:1993,Anastopoulos:Savvidou:2003}. Assuming that $\delta t:=\sup_i\{|t_i-t_{i-1}|\}\approx\O(\frac{1}{n})$, one can approximate $\{\zeta_{t_i}\}_{i=1}^n$ by the path $\zeta(t)$ on $\PH$, and for large $n$ this gives
\begin{align}
\log\tr_{{\cal{V}}_n}(C_\varpi)&=\log\s{\zeta_{t_0},\zeta_{t_n}}_\H-\sum_{i=1}^{n}\log\s{\zeta(t_i),\zeta(t_{i-1})}_\H\\
&=\log\s{\zeta_{t_0},\zeta_{t_n}}_\H-\sum_{i=1}^{n}\s{\zeta(t_i),\zeta(t_i)-\zeta(t_{i-1})}_\H+\O(n^{-2}),
\end{align}
hence
\begin{equation}
\lim_{\delta t\ra 0}\log\left(\tr_{{\cal{V}}_n}(C_\varpi)\right)
= \log\s{\zeta_{t_0},\zeta_{t_n}}_\H-\int_{\zeta_{t_0}}^{\zeta_{t_n}}\s{\zeta(t),\dd\zeta(t)}_\H,
\end{equation}
where the last term is the Stieltjes integral. Comparing this result with equation \eqref{generalberryphase}, one can see that for any path which allows for the definition of the Stieltjes integral, the trace of a class operator is equal to a geometric phase \eqref{generalberryphase}:
\begin{equation}
\tr_{{\cal{V}}}(C_\varpi) = \ee^{\ii\theta[\zeta(\cdot)]}.
\label{traceless.berry}
\end{equation}
Hence, for a given history $\varpi$, its corresponding geometric phase is defined by the trace of a class operator. Observing that $C_\varpi$ is used in the definition \eqref{decoherencefunctional} of the histories functional, one can rewrite the latter in terms of the geometric phase:
\begin{equation}
\hf_{\rho_{t_0},H=0}(P_{\zeta(\cdot)},P_{\xi(\cdot)})=\s{\zeta(t_0),\rho_{t_0}\xi(t_0)}_\H\s{\zeta(t_n),\xi(t_n)}_\H\ee^{-\int\limits_{t_0}^{t_n}\dd t\s{\zeta(t),\frac{\dd\zeta(t)}{\dd t}}_\H-\int\limits_{t_0}^{t_n}\dd t\s{\xi(t),\frac{\dd}{\dd t}\xi(t)}_\H}.
\label{geometric.phase.in.decoherence.functional}
\end{equation}
For a quantum theoretic model with a nonzero hamiltonian $H$ the histories functional is equal to \cite{Anastopoulos:Savvidou:2002}:
\begin{equation}
\begin{array}{c}
\hf_{\rho_{t_0},H}(P_{\zeta(\cdot)},P_{\xi(\cdot)})=\s{\zeta(t_0),\rho_{t_0}\xi(t_0)}_\H\s{\zeta(t_n),\xi(t_n)}_\H\ee^{\ii\e{S^*[\zeta(\cdot)]}+\ii\e{S^*[\xi(\cdot)]}},
\end{array}
\label{dd1}
\end{equation}
where
\begin{equation}
\begin{array}{c}
\e{ S^*[\zeta(\cdot)]}:=\int_{t_0}^{t_n}\dd t\s{\zeta(t),\left(\ii\frac{\dd}{\dd t}-H\right)\zeta(t)}_\H
= \ii\int_{t_0}^{t_n}\dd t\s{\zeta(t),\frac{\dd\zeta(t)}{\dd t}}_\H-\int_{t_0}^{t_n}\dd t\s{\zeta(t),H\zeta(t)}_\H.
\end{array}
\label{act}
\end{equation}
This agrees with the earlier result of Isham and Linden \cite{Isham:Linden:1995}, who have constructed a special case of histories functional $\hf_{\rho,H}$. Using the continuous time projection operator on ${\cal{V}}$ corresponding to coherent states and using the technical assumption of $t_0\ra-\infty$ and $t_n\ra+\infty$, they have obtained
\begin{equation}
\begin{array}{c}
\hf_{\rho,H}(P_{\zeta(\cdot)},P_{\xi(\cdot)})=\s{\zeta(t_0),\rho_{t_0}\xi(t_0)}_\H\ee^{\int_{t_0}^{t_n}\left(\s{\zeta(t),\frac{\dd\zeta(t)}{\dd t}}_\H-\s{\xi(t),\frac{\dd\xi(t)}{\dd t}}_\H\right)}\ee^{\ii\int_{t_0}^{t_n}(\s{\xi(t),H\xi(t)}_\H-\s{\zeta(t),H\zeta(t)}_\H)},
\end{array}
\end{equation}
where $\int_{t_0}^{t_n}\s{\zeta(t),\dd\zeta(t)}$ is a Stieltjes integral. In order to compare the equation (\ref{dd1}) with the action equation (\ref{action}), consider the Schr\"{o}dinger representation of the histories algebra (\ref{heisenberg}), provided by the operators $\pp_t=-\ii\frac{\partial}{\partial x_t}$ and $\qq_t=x_t$, acting on the space $L_2(\RR,\dd x_t)$. Then $\widetilde{V}_t=\pp_t\dot{\qq}_t=-\ii\frac{\dd}{\dd s}$, and the equation (\ref{act}) can be written in the form
\begin{equation}
\e{S^*[\zeta(\cdot)]}=-\int_{t_0}^{t_n}\dd t\s{\zeta(t),(\pp_t\dot{\qq}_t+H)\zeta(t)}_\H
\label{complexaction}
\end{equation}
\textit{if} it is assumed that $t=s$ and $\frac{\dd}{\dd t}=\frac{\dd}{\dd s}$. This equation shows that the nonhamiltonian part of the action $S$ is reflected in the geometric phase.
According to Anastopoulos and Savvidou, the complete form of histories functional can be reconstructed by summing over all paths $\zeta(\cdot)$ and $\xi(\cdot)$ that are compatible with the given histories $\varpi$ and $\vartheta$ respectively:
\begin{equation}
\hf_{\rho,H}(P_\varpi,P_\vartheta)=\sum_{\zeta(\cdot)\subset\varpi}\sum_{\xi(\cdot)\subset\vartheta}\hf_{\rho,H}\left(P_{\zeta(\cdot)},P_{\xi(\cdot)}\right).
\label{dd2}
\end{equation}
In face of the above results, they conclude, that \cytat{the knowledge of the geometric phase---for a set of histories and of the automorphism that implements the dynamics---is sufficient to fully reconstruct the decoherence [histories] functional---and hence all the probabilistic content of the [histories approach to quantum] theory} \cite{Anastopoulos:Savvidou:2002}. In other words, the histories approach provides the complete description of temporal behaviour of quantum theoretic models using two levels of description: the unitary action automorphism and the histories functional, which incorporates the nonunitary changes of geometry of the Hilbert space related with the sequences of projection operators taken into consideration.
However, this result is not completely clear. The functional $\e{S^*[\zeta(\cdot)]}$, despite the suggestive notation, is not the expectation value of the adjoint of the action operator $S_\kappa$ (\ref{quantumactionoperator}), unless $\int_{-\infty}^{+\infty}\dd t$ is interchangeable with $\s{\zeta(t),\cdot\;\zeta(t)}_\H$, and the smearing function $\kappa$ is introduced consistently at some stage of derivation of (\ref{complexaction}). Moreover, the assumption $t=s$ is not justified by any reasons other than \textit{ad hoc} decision. It is unsatisfactory that in order to derive the relationship of two different temporal evolutions with the geometric phase one has to set the values of corresponding time parameters to be identical. There also remains the question to what extent several different technical assumptions used in the construction of $S_\kappa$ and $\e{S^*[\zeta(\cdot)]}$ are essential for the final results and conclusions. Moreover, it is problematic to what extent the operator $V$ can be related with the `external time' \textit{without} introducing the family of projections $P_{t_i}$ and the object $\e{S^*[\zeta(\cdot)]}$. By definition, the operators $V$ and $S_\kappa$ refer only to continuous number of copies of the same Hilbert space, but without any reference to projections or `measurements'.
Finally, these results are based on the arbitrary choice of the particular Fock--Cook (or coherent states) representation of histories version of the BJDH commutation relations. If the number of degrees of freedom of the algebra is finite, then the Stone--von Neumann \cite{Stone:1930,vonNeumann:1931} theorem guarantees that the Schr\"{o}dinger representation of the Weyl form of the BJDH algebra of canonical commutation relations is a unique, up to unitary equivalence, irreducible representation. However, in the infinite-dimensional case there exists uncountable many different unitarily inequivalent irreducible representations of this algebra \cite{Garding:Wightman:1954:ccr,Wightman:Schweber:1955}. Hence, the choice of a particular representation provides a nontrivial decision problem and should be justified by some argument, but there is no such argument at sight. In order to resolve these problems we have to move to a more general approach to quantum theory, the algebraic approach.
\subsection{Hilbert space geometry and coherent state path integrals\label{Hilbert.space.geometry.path.integrals}}
In this subsection we will discuss the basic aspects of the geometric approach to the formalism of the Hilbert space based quantum theory \cite{Strocchi:1966,Marsden:1968:generalized,Kibble:1979,Cirelli:Lanzavecchia:Mania:1983,Simon:1983,Cirelli:Lanzavecchia:1984,Heslot:1985,Page:1987,Aharonov:Anandan:1987,Anandan:Aharonov:1990,Cirelli:Mania:Pizzocchero:1990,Gibbons:1992,Cirelli:Mania:Pizzocchero:1994,Hughston:1995,Hughston:1996,Schilling:1996,Field:1996,Field:1997,Brody:Hughston:1998,Ashtekar:Schilling:1999,Cirelli:Gatti:Mania:1999,Brody:Hughston:2001,Chruscinski:Jamiolkowski:2004,Bengtsson:Zyczkowski:2006} and its relationship with the description of temporal behaviour of quantum theoretic models in the Hilbert space based quantum histories approach.
Every complex Hilbert space $\H$ can be considered as the real Hilbert space of double dimension, equipped with a complex structure operator $\jbold^\H:\H\ra\H$ such that $(\jbold^\H)^2=-\II$ and $\s{\xi,\jbold^\H\zeta}_\H=\ii\s{\xi,\zeta}_\H$ \cite{Stueckelberg:1960}. The decomposition of the inner product on $\H$ into real and imaginary parts,
\begin{align}
\re\s{\xi,\zeta}_\H&=:\frac{1}{2}\gbold^\H(\xi,\zeta),\\
\im\s{\xi,\zeta}_\H&=:\frac{1}{2}\wbold^\H(\xi,\zeta),
\end{align}
equips the real Hilbert space $\H$ with the structure of the nondegenerate positive definite real inner product $\gbold^\H$ and nondegenerate closed two-form $\wbold^\H$. They turn, respectively, to a riemannian and a symplectic structure on the manifold $\PP\H$, with $\H$ understood as a tangent space over $\PP\H$. The complex structure $\jbold^\H$ imposes the relationships
\begin{align}
\gbold^\H(\xi,\zeta)&=\wbold^\H(\xi,\jbold^\H\zeta),\\
\nabla^{\gbold^\H}\jbold^\H&=0,
\end{align}
where $\nabla^{\gbold^\H}$ is a covariant derivative on $\H$ associated with $\gbold^\H$. These two equations imply that the triple $(\gbold^\H,\wbold^\H,\jbold^\H)$ equips $\H$ in the structure of the K\"{a}hler manifold. If $\H\iso\CC^{n+1}$ and some orthonormal basis $\{e_a\}$ in $\H$ is chosen, $a\in\{0,\ldots,n\}$, then the inner product on $\H$ can be denoted (using abstract index notation)
\begin{equation}
\s{\xi,\zeta}_\H=\xi_a^*\zeta^a,
\end{equation}
while the infinitesimal equations for riemannian metric $\gbold^\H$ and symplectic form $\wbold^\H$ read
\begin{align}
\ddd\tilde{s}^2&=\gbold_{ab}\ddd{\zeta^a}^*\otimes\ddd\zeta^b,\\
\tilde{w}^\H&=\wbold^\H_{ab}\ddd\zeta^a\wedge\ddd\zeta^b.
\end{align}
The projection of these structures to the projective space $\PH$, provided in finite-dimensional case by $z^a:=\zeta^a/\zeta^0$, $a\in\{1,\ldots,n\}$, induces a metric $\ddd s_{\PH}^2$ and a $U(1)$ connection one-form $A_{\PH}$,
\begin{align}
\ddd s^2_{\PH}&:=\frac{1}{1+z_a^*z^a},\\
A_{\PH}&:=\ii z_a^*dz^a.
\end{align}
The space $\PH$ has the structure of the compact K\"{a}hler manifold, the metric $\ddd s^2_{\PH}$ is the Fubini--Study metric, while the connection one-form $A_{\PH}$ is the Berry connection. In finite dimensional case,
\begin{equation}
\H\iso\CC^{n+1}\limp\PH\iso\CC\PP^n\iso S^{2n+1}/U(1),
\end{equation}
while the Fubini--Study metric on $\PH$ is given explicitly by
\begin{equation}
\gbold^{\mathrm{FS}}_{ab}=\frac{\s{\zeta,\zeta}_\H\delta_{ab}-\zeta_{(a}\zeta^*_{b)}}{\ab{\s{\zeta,\zeta}_\H}^2},
\label{Fubini.Study.equation}
\end{equation}
where the round brackets denote the symmetrisation of indices. The space $\CC\PP^n$ has a symmetry group of dimension $n(n+2)$, which is generated by a family of $n(n+2)$ Killing vector fields.
This framework allows a geometric description and reconsideration of the structure of the Hilbert space based framework of quantum theory. In particular, the self-adjoint operators on $\H$, generating the unitary Schr\"{o}dinger evolutions on $\H$, correspond to such smooth functions on the K\"{a}hler manifolds $\PH$ that preserve the K\"{a}hler structure (that is, their hamiltonian vector fields are also the Killing vector fields). These hamiltonian functions on $\PH$ are given by the normalised expectations $\s{\zeta,H\zeta}_\H/\s{\zeta,\zeta}_\H$ of the corresponding self-adjoint operators $H$ on $\H$. Moreover, the geodesic distance $d_{\gbold^{\mathrm{FS}}}$ with respect to the Fubini--Study metric determines the transition probability between two vectors,
\begin{equation}
p(\zeta|\xi)=\ab{\s{\zeta,\xi}_\H}^2=\cos^2(d_{\gbold^{\mathrm{FS}}}(\zeta,\xi)),
\end{equation}
thus
\begin{equation}
d_{\gbold^{\mathrm{FS}}}(\zeta,\xi)=\arccos(\tr_\H(P_\zeta P_\xi)),
\label{Fubini.Study.distance}
\end{equation}
where $P_\xi$ and $P_\zeta$ are projection operators on the $1$-dimensional subspaces of $\H$ that are linearly spanned by $\xi$ and $\zeta$, respectively.
An interesting additional geometric structure can be introduced using the coherent vectors representation. Let there be given a group $G$ together with its irreducible unitary representation $G\ni g\mapsto U(g)\in\BH$. Then, for a given choice of a normalised reference vector $\zeta\in\H$ (specified, for example, as the vector invariant under the maximal compact subgroup of $G$), one can define the Hilbert space vectors $U(g)\zeta\in\H$, introduce the equivalence relation
\begin{equation}
g_1\sim g_2\iff\exists \ee^{\ii\lambda}\in\CC\;\;\;U(g_1)\zeta=\ee^{\ii\lambda}U(g_2)\zeta,
\end{equation}
and define the homogeneus quotient space $\Gamma:=G/\sim$. The space $\Gamma$ is a parameter space that defines and labels the \textit{coherent vectors} of $\H$ by \cite{Schroedinger:1926:Naturwiss,Glauber:1963,Perelomov:1972,Perelomov:1986}
\begin{equation}
\iota_\Gamma:\Gamma\ni z\mapsto U(z)\zeta\in\PH.
\end{equation}
Using $\iota_\Gamma$, one can pullback the geometric objects from $\PH$ to $\Gamma$, equipping $\Gamma$ with the symplectic, riemannian and affine structure:
\begin{align}
\ddd s^2_\Gamma&:=\n{\ddd z}^2-\ab{\s{z,\ddd z}_\H}^2=\ab{\s{\ddd z,\ddd z}_\H}^2-\ab{\s{z,\ddd z}_\H}^2,\\
A_\Gamma&:=\ii\s{z,\ddd z}_\H,\\
w_\Gamma&:=\ddd A_\Gamma,
\end{align}
where $\dd$ denotes the exterior derivative on $\Gamma$, and $w_\Gamma$ is a symplectic structure on $\Gamma$ if it is nondegenerate. If the space $\Gamma$ is interpreted as the `phase space', then $\iota_\Gamma$ is intepreted as a map from `phase space' to `space of rays'.
Anastopoulos and Savvidou \cite{Anastopoulos:Savvidou:2003} have used these results in order to uncover the relationship between the quantum histories and the metric structure on the projective Hilbert space. Using coherent vectors $z\in\H$, they derive
\begin{align}
\s{z,z+\delta z}&=1+\s{z,\partial_az^a}_\H\delta z^a+\frac{1}{2}\s{z,\partial_a\partial_bz}_\H\delta z^a\delta z^b+\O(\delta z^3)\\
&=\exp\left(\ii A_a(z+\frac{1}{2}\delta z)\delta z^a-\frac{1}{2}\gbold^{\mathrm{FS}}_{ab}\delta z^a\delta z^b\right)+\O(\delta z^3).
\end{align}
For $\delta z_k=z_{k+1}-z_k$, the equation (\ref{eq.Bargmann}), written in the form
\begin{equation}
\hf_{H=0}(P_\varpi,P_\vartheta)=\prod_k\s{z_k,z_k+\delta z_k}_\H,
\end{equation}
leads to second-order approximation
\begin{equation}
\hf_{H=0}(P_\varpi,P_\vartheta)=\exp\left(\ii\sum_k(z_k+\frac{1}{2}\delta z_k)\delta z_k-\frac{1}{2}\sum_k\delta s^2_k\right).
\end{equation}
If the paths $z(\cdot)$ are continuous and the variations $\delta z_k$ are bounded ($\ab{\delta z_k^a}<\epsilon$ and $\epsilon\ra 0$), then this equation converges to the expression (\ref{geom.phase.def}) on the geometric phase. However, if the paths $z(\cdot)$ cannot be considered as continuous (or differentiable) functions of $t$, then the approximation of the histories functional for the cut-off of the scale of $t$ given by $\frac{1}{\upsilon}$ leads to \cite{Anastopoulos:Savvidou:2003}
\begin{equation}
\hf_{H=0}(P_\varpi,P_\vartheta)=\exp\left(\ii\int_\pathgamma A_\Gamma-\frac{1}{2\upsilon}\int_\pathgamma \dd t\gbold^{\mathrm{FS}}_{ab}(z(t))\dot{z}^a\dot{z}^b\right).
\label{histories.functional.with.metric}
\end{equation}
This result is very closely related to the Daubechies--Klauder approach \cite{Daubechies:Klauder:1985,Klauder:1988,Klauder:1995,Klauder:2011}, who introduced exact continuous-time regularised coherent vectors propagator for the phase space path integral, and proved that under mild assumptions on hamiltonian (square and quadric integrability, see e.g. \cite{Klauder:2003} for a brief statement of those) one has
\begin{align}
\s{z(t=s),\ee^{-\ii Hs}z(t=0)}_\H&=\lim\limits_{\upsilon\ra+\infty}\int \DD z(\cdot)
\ee^{\left(\ii\int_\pathgamma A_\Gamma\right)}
\ee^{\left(-\ii\int_0^s\dd t\;h(z(t))\right)}
\ee^{-\frac{1}{2\upsilon}\left(\int_0^s\dd t\;g_{ab}(z(t))\dot{z}^a\dot{z}^b\right)}\label{Klauder.path.integral}\\
&=2\pipi\lim\limits_{\upsilon\ra+\infty}\ee^{\upsilon s/2}\int\tmu_{\mathrm{W}}^\upsilon(\pp_\Gamma,\qq_\Gamma)\ee^{\ii\int(\pp_\Gamma\dd\qq_\Gamma-H(\pp_\Gamma,\qq_\Gamma)\dd t)}\label{DK.Wiener.expression},
\end{align}
where $h(z(t))$ is a hamiltonian function\footnote{In the context of our paper, we consider it to be defined by $h(z(t)):=\s{z(t),Hz(t)}_\H$ for a given self-adjoint hamiltonian operator $H$ on $\H$.} on $\Gamma$, while $\tmu_{\mathrm{W}}^\upsilon(\pp_\Gamma,\qq_\Gamma)$ is a pinned Wiener measure on a phase space $\Gamma$. Moreover, this formulation is covariant under canonical transformations of phase space coordinates, what is not the case for most of other approaches to quantisation, including the Schr\"{o}dinger quantisation
\begin{equation}
\pp_\Gamma,\qq_\Gamma\in C^{\infty}(\Gamma)\rightsquigarrow-\ii\frac{\partial}{\partial x},x:L_2(\RR,\dd\lambda)\ra L_2(\RR,\dd\lambda),
\end{equation}
the Born--Jordan--Dirac--Heisenberg quantisation
\begin{equation}
\pp_\Gamma,\qq_\Gamma\in C^{\infty}(\Gamma)\rightsquigarrow \pp,\qq:\ell_2(\NN)\ra\ell_2(\NN),\;[\qq,\pp]=\ii\II,
\end{equation}
as well as the lattice formulation of phase space version of Feynman path integral (see a discussion in \cite{Klauder:2003}). All these methods of quantisation depend on the particular choice of the phase space coordinates, what makes these prescriptions incomplete, because descriptions which are considered to be canonically equivalent on the level of phase space become unitarily inequivalent on the level of the Hilbert space, and there is provided no procedure solving the problem of choice of unique description among inequivalent ones. The above result shows that the metric structure on the Hilbert space (and the corresponding metric structure on $\Gamma$) provides an important conceptual and mathematical element of the quantum theory.
It is also interesting to note that for finite value of $\upsilon$ the propagator (\ref{Klauder.path.integral}) is not longer unitary \cite{Klauder:1995}. From the perspective of histories approach to quantum theory, this means that the metric structure on the Hilbert space allows (some sort of) quantification of the nonunitary (and noncontinuous) temporal behaviour. This observation should be furnished by an additional result of Klauder and Maraner \cite{Klauder:Maraner:1997}, who showed that the usual definition of dynamics on phase space by means of Hamilton's variational principle,
\begin{equation}
\delta\int \dd t(\theta_a\dot{\xi}^a-h(\xi(t)))=0,
\label{hamilton.variational}
\end{equation}
where $\omega_{ab}=\partial_a\theta_b-\partial_b\theta_a$ is a symplectic form on the phase space, while $\xi$ are arbitrary phase space coordinates, is equivalent to the variational principle
\begin{equation}
\delta\int\dd t(\theta_a\dot{\xi}^a+\frac{1}{2}\lambda\gbold^{\mathrm{FS}}_{ab}(\xi(t))\dot{\xi}^a\dot{\xi}^b)=0
\label{Klauder.Maraner.metric.variation}
\end{equation}
under constraint
\begin{equation}
\det\left(\gbold^{\mathrm{FS}}_{ab}(\xi)\right)=h^{-2n}(\xi)
\label{Klauder.Maraner.prior}
\end{equation}
and in the limit $\lambda\ra0$, where $\gbold^{\mathrm{FS}}_{ab}(\xi)$ is a riemannian metric on the phase space, $2n$ is the dimension of this space, while $\lambda\in\RR$ is an arbitrary scale factor. This result was derived in the context of phase space $\Gamma$, but nothing forbids us from applying it to $\PH$, with the hamiltonian function provided by normalised expectation of hamiltonian operator and with the riemannian metric provided by the Fubini--Study metric of $\PH$. The equation (\ref{hamilton.variational}) takes then the form of variation of the equation (\ref{act}),
\begin{equation}
\delta\e{S^*[\zeta(\cdot)]}=\delta\int_{t_0}^{t_n}\dd t\s{\zeta(t),\left(\ii\frac{\dd}{\dd t}-H\right)\zeta(t)}_\H=0,
\end{equation}
which gives the Schr\"{o}dinger equation. This leads to a question whether some modification of the variational principle (\ref{Klauder.Maraner.metric.variation}) on $\PH$ could result in an interesting form of the temporal behaviour of quantum theoretic models? In particular, it seems that for not vanishing metric term, provided by the finite values of $\lambda$ corresponding to finite values of $\upsilon$ in (\ref{histories.functional.with.metric}) and (\ref{Klauder.path.integral}), the resulting temporal behaviour would be nonunitary
These observations will play an important guiding role in generalisation of the elements of histories approach to an algebraic context.
\subsection{Case study: Algebraic action operator and the limits of unitarity\label{algebraic.action.operator.section}}
Given a W$^*$-algebra $\N$ and $\psi\in\W_0(\N)$, consider a crossed product $\N\rtimes_{\sigma^\psi}\RR$, defined as the von Neumann algebra acting on the Hilbert space $L_2(\RR,\dd t;\H)\iso\H\otimes L_2(\RR,\dd t)$ and generated by the operators $\pi_{\sigma^\psi}(x)$ and $u_\RR(t)$, which are defined by\rpktarget{PI.SIGMA} \rpktarget{URR}
\begin{align}
(\pi_{\sigma^\psi}(x)\xi)(t)&:=\sigma_{-t}^\psi(x)\xi(t),
\label{pi.sigma}\\
(u_\RR(t_2)\xi)(t_1)&:=\xi(t_1-t_2),
\label{lambda.tau}
\end{align}
for all $x\in\N$, $t,t_1,t_2\in\RR$, $\xi\in L_2(\RR,\dd t;\H)$, see e.g. \cite{vanDaele:1978}. These two operators satisfy the covariance equation
\begin{equation}
u_\RR(t)\pi_{\sigma^\psi}(x)u_\RR^*(t)=\pi_{\sigma^\psi}(\sigma_t^\psi(x)).
\label{covariance.eqn}
\end{equation}
The equation \eqref{pi.sigma} can be written as
\begin{equation}
(\pi_{\sigma^\psi}(x)\xi)(t)=\Delta_\psi^{\ii t}x\Delta_\psi^{-\ii t}\xi(t)=\ee^{-\ii K_\psi t}x\ee^{\ii K_\psi t}\xi(t),
\end{equation}
where $K_\psi$ is a modular hamiltonian of the modular operator $\Delta_\psi$. So, the covariance equation \eqref{covariance.eqn} translates between the family of unitaries that partially generate the crossed product algebra $\N\rtimes_{\sigma^\psi}\RR$ and the modular automorphism of the underlying von Neumann algebra $\N$:
\begin{equation}
u_\RR(t)\pi_{\sigma^\psi}(x)u_\RR(t)^*=
\pi_{\sigma^\psi}(\ee^{-\ii t K_\psi}x\ee^{\ii t K_\psi}).
\label{modular.histories.covariance}
\end{equation}
Using the uniqueness of the standard representation up to unitary equivalence, Falcone and Takesaki \cite{Falcone:Takesaki:2001} (see \cite{Kostecki:2013} for a pedagogical introduction) proved that the map $\N\mapsto\N\rtimes_{\sigma^\psi}\RR$ extends to a functor $\VNCore$ from the category $\VNIso$ of von Neumann algebras with $*$-isomorphisms to its own subcategory $\VNsfIso$ of semi-finite von Neumann algebras with $*$-isomorphisms. The functoriality $\CanVN:\WsIso\ra\VNIso$ of Kosaki's construction \cite{Kosaki:1980:PhD} of canonical representation $\pi_\C$ of any W$^*$-algebra $\C$ turns the assignment
\begin{equation}
\C\mapsto\pi_\C(\C)=:\N\mapsto\core=\widetilde{\pi_\C(\C)}
\end{equation}
to a functor
\begin{equation}
\WstarCore:\WsIso\ra\VNsfIso,
\end{equation}
where $\WstarCore:=\VNCore\,\circ\,\CanVN$, while $\WsIso$ consists of W$^*$-algebras and $*$-isomorphisms. For any W$^*$-algebra $\N$, the object $\WstarCore(\N)\in\Ob(\VNsfIso)$ will be called \df{canonical core} of $\N$ and denoted $\core$. By equipping the canonical core von Neumann algebra $\core$ of the countably finite W$^*$-algebra $\N$ with the choice of some $\omega\in\N^+_{\star0}$, we obtain a unitary isomorphism $\core\iso\N\rtimes_{\sigma^\omega}\RR$.
The operator $u_\RR(t)$, when considered as an operator on $L_2(\RR,\dd t)$, takes the form
\begin{equation}
u_\RR(r)=\ee^{-\ii r\tilde{V}},\;\;\;
\tilde{V}:=-\ii\frac{\dd}{\dd t}.
\label{explicit.form}
\end{equation}
So, the analogue of a quantum `histories liouvillean' automorphism \eqref{l.h.auto} is naturally present in the structure of a unitary representation of the canonical core algebra. From the covariance equation \eqref{covariance.eqn} it follows that this automorphism of $\N\rtimes_{\sigma^\omega}\RR$ uniquely corresponds to the modular automorphism of $\N$. Hence, the pair $(\N,\omega)$ uniquely determine a W$^*$-dynamical system $(\N,\RR,\sigma^\omega)$. But there might be also given another description of a temporal behaviour related with the same algebra $\N$, provided by some group of $*$-automorphisms $\alpha:\RR\ra\Aut(\N)$. If $\alpha$ is continuous in the weak-$\star$ topology, then one has to consider the coexistence of two W$^*$-dynamical systems: $(\N,\RR,\sigma^\omega)$ and $(\N,\RR,\alpha)$. While $\sigma^\omega$ is completely determined by the properties of $\N$ and $\omega$, the $*$-automorphism $\alpha$ is arbitrary. Using familiar terminology, one can say that $\sigma^\omega$ is a `kinematic' automorphism, while $\alpha$ is a `causal' automorphism. These characteristics of $\sigma^\omega$ and $\alpha$, together with an equation \eqref{explicit.form} lead us to propose to consider $\sigma^\omega$ as an algebraic replacement of the `histories liouvillean' automorphism \eqref{V.hist.auto} and to consider $\alpha$ as an algebraic replacement of the `histories hamiltonian' automorphism \eqref{H.hist.auto}. We will join these two separated automorphisms into one automorphism, representing the `complete temporal behaviour' of the quantum theoretic model $(\N,\alpha,\omega)$, and forming an algebraic replacement for the `histories action' automorphism \eqref{S.hist.auto}.
We do not require the model $(\N,\alpha,\omega)$ to be a `quantum dynamical system' (in the sense of \cite{Pillet:2006}) with respect to $\alpha:\RR\ra\Aut(\N)$, because we do not need to (and do not want to) assume the invariance of $\omega$ with respect to $\alpha$. In fact, instead of declaring invariance of $\omega$ with respect to $\alpha$, we will use $\alpha$ in order to construct a new algebraic state $\phi$. Consideration of the derivations of these $*$-automorphisms together with the corresponding hamiltonians is also not useful here, because of the lack of a unique characterisation of unbounded generators of $*$-automorphisms in terms of corresponding self-adjoint hamiltonians (see \cite{Kostecki:2013} for more discussion and further references on this).
The standard liouvillean of $\sigma^\omega$ is given by its modular hamiltonian $K_\omega=-\log\Delta_\omega$. We will denote by $L_\alpha$ the standard liouvillean of $\alpha$ in the GNS representation $(\H_\omega,\pi_\omega,\Omega_\omega)$. If
\begin{equation}
L_{\alpha,\omega}:=L_\alpha+K_\omega+J_\omega K_\omega J_\omega
\label{updated.action.liouvillean}
\end{equation}
is essentially self-adjoint on $\dom(K_\omega)\cap\dom(L_\alpha)\cap\dom(J_\omega K_\omega J_\omega)$, and if $K_\omega+L_\alpha$ is essentially self-adjoint on $\dom(K_\omega)\cap\dom(L_\alpha)$, then
\begin{equation}
u_{\alpha,\omega}(x)(t):=\ee^{\ii t(K_\omega+L_\alpha)}x\ee^{-\ii t(K_\omega+L_\alpha)} \;\;\forall x\in\pi_\omega(\N),
\label{updated.action.automorphism}
\end{equation}
is a weak-$\star$ continuous $*$-automorphism of $\pi_\omega(\N)$, and $L_{\alpha,\omega}$ is its standard liouvillean. Hence, $(\pi_\omega(\N),\RR,u_{\alpha,\omega})$ is a W$^*$-dynamical system with a corresponding crossed product $\pi_\omega(\N)\rtimes_{u_{\alpha,\omega}}\RR$. Moreover, if $L_\alpha$ is bounded, then the DFFA convergent perturbation expansions hold:
\begin{align}
u_{\alpha,\omega}(x)&=
\sum_{n=0}^\infty \ii^n
\int_{0\leq t_n\leq\cdots\leq t_1\leq t}
\dd t_1\cdots \dd t_n
{}[\alpha_{t_n}(K_\omega),
{}[\ldots,
{}[\alpha_{t_1}(K_\omega),
\alpha_t(x){}]
\ldots{}]
{}],\\
E_{\alpha,K_\omega}(t):=
\ee^{\ii t(K_\omega+L_\alpha)}\ee^{-\ii tL_\alpha}&
=\sum_{n=0}^\infty \ii^n
\int_{0\leq t_n\leq\cdots\leq t_1\leq t}
\dd t_1\cdots \dd t_n
\alpha_{t_n}(K_\omega)\cdots\alpha_{t_1}(K_\omega).
\end{align}
Hence, under some relatively weak conditions, the modular automorphism $\sigma^\omega$ and the additional $*$-automorphism $\alpha$ form together a unique automorphism $u_{\alpha,\omega}$ with its own self-adjoint liouvillean $L_{\alpha,\omega}$. If one thinks of $\alpha$ as an algebraic analogue of an automorphism generated by the `interaction hamiltonian', then the automorphism $u_{\alpha,\omega}$ can be considered as a `correction' of $u_\alpha$ by means of an associated $U(1)$ connection $1$-form $[K_\omega,\,\cdot\,]$ on the Hilbert bundle of Hilbert spaces $\H_\omega$ over the image of real line $\RR$ in $\M(\N)$. Given $L_\alpha$ and $K_\omega$, we can define also the operator
\begin{equation}
L_{\omega,\alpha}:=K_\omega+L_\alpha+J_\omega L_\alpha J_\omega,
\label{dual.std.action.liouvillean}
\end{equation}
with the conditions on essential self-adjointness analogous to the case of $L_{\alpha,\omega}$. This operator is not interpretable as a standard liouvillean of $\alpha$ perturbed by a $U(1)$ connection form. However, as we will see below, it also encodes some very interesting information.
By an analogy with the Hilbert space based histories approach to quantum theory we will call $L_{\alpha,\omega}$ an (\textit{algebraic}) \df{action operator} and will call $u_{\alpha,\omega}$ an (\textit{algebraic}) \df{action automorphism}. We will call $L_{\omega,\alpha}$ a \df{dual action operator}.
Recall that in the Hilbert space based histories approach the action operator was a generator of a complete unitary temporal behaviour of a given quantum theoretic model, including not only the `internal' temporal unitary changes related to the fixed Hilbert space, but also the `external' temporal unitary changes between two different Hilbert spaces (in fact, that formalism was limited to the continuous one-parameter family of identical copies of the same Hilbert space). In the algebraic approach, the change of a Hilbert space corresponds to the change of an algebraic state, and it implies the corresponding change of a representation of an underlying W$^*$-algebra. In order to strengthen the relationship between the Hilbert space based histories and our algebraic approach, we will show how the action operator is related to the change between two representations or between two different algebraic states.
In order to do this, we need to use one more result of Derezi\'{n}ski, Jak\v{s}i\'{c} and Pillet \cite{DJP:2003}. They show that if $\omega$ is a faithful Kubo--Martin--Schwinger state with respect to a $*$-automorphism $u$ with a parameter $\beta$, then under the assumptions used previously for derivation of $*$-automorphism $u_t^Q$ and its standard liouvillean $L_Q$, and assuming additionally that $\n{e^{-\beta Q/2}\Omega_\omega}_{\H_\omega}<\infty$,
\begin{align}
\Omega_Q&:=\ee^{-\beta(L_u+Q)/2}\Omega_\omega,\\
\omega_Q(\cdot)&:=\s{\Omega_Q,\cdot\Omega_Q}/\n{\Omega_Q}^2_{\H_\omega}
\end{align}
satisfy
\begin{enumerate}
\item[0)] $\Omega_\omega\in\dom(\ee^{-\beta(L_u+Q)/2})$,
\item[1)] $\Omega_Q\in\stdcone_\omega$ is cyclic and separating for $\pi_\omega(\N)$,
\item[2)] $\omega_Q$ is KMS with respect to $u^Q$ and $\beta$,
\item[3)] $\log\Delta_{\Omega_Q}=-\beta L_Q$ and
\begin{equation}
\log\Delta_{\Omega_Q,\Omega_\omega}=-\beta L_u-\beta Q.
\label{log.updated.rel.mod.op}
\end{equation}
\end{enumerate}
By the Takesaki theorem, the faithful state $\omega$ on a W$^*$-algebra is always KMS with respect to $\sigma^\omega$ with $\beta=1$. Hence, under the assumptions allowing for the construction of the dual algebraic action operator $L_{\omega,\alpha}$, and assuming also that $\n{e^{-(K_\omega+L_\alpha)/2}\Omega_\omega}<\infty$, it holds that
\begin{equation}
\phi(\cdot):=\frac{\s{\ee^{-(K_\omega+L_\alpha)/2}\Omega_\omega,(\,\cdot\,)\ee^{-(K_\omega+L_\alpha)/2}\Omega_\omega}_{\H_\omega}}{\n{\ee^{-(K_\omega+L_\alpha)/2}\Omega_\omega}^2_{\H_\omega}}
\label{updated.action.omega}
\end{equation}
is KMS with respect to $u_{\omega,\alpha}$ with $\beta=1$. Hence, $u_{\omega,\alpha}$ is a modular automorphism of $\pi_\omega(\N)$ with respect to $\phi$. This is a very interesting result, because it means that while $(u_{\alpha,\omega},L_{\alpha,\omega})$ play the role of an action automorphism and an action operator with respect to the pair $(\N,\omega)$, $(u_{\omega,\alpha},L_{\omega,\alpha})$ play the role of a modular automorphism and a modular hamiltonian with respect to the pair $(\N,\phi)$. Hence, under the assumptions
\begin{enumerate}
\item $\omega$ is a faithful normal algebraic state on a W$^*$-algebra $\N$,
\item $L_\alpha$ is a standard liouvillean of $*$-automorphism $\alpha$ of $\N$ affiliated with $\pi_\omega(\N)$,
\item $K_\omega+L_\alpha$ is essentially self-adjoint on $\dom(K_\omega)\cap\dom(L_\alpha)$,
\item $K_\omega+L_\alpha-J_\omega L_\alpha J_\omega$ is essentially self-adjoint on $\dom(K_\omega)\cap\dom(L_\alpha)\cap\dom(J_\omega L_\alpha J_\omega)$,
\item $\n{\ee^{-(K_\omega+L_\alpha)/2}\Omega_\omega}_{\H_\omega}<\infty$,
\end{enumerate}
the $*$-automorphism $\alpha$ can be always assimilated as a part of the modular automorphism $\sigma^{\phi}$ that is uniquely specified by an `updated' algebraic state $\phi$ \eqref{updated.action.omega}. In other words, the $*$-automorphism forming a `causal' part of an algebraic quantum action automorphism can be considered as a constitutive element of a `kinematic' temporal behaviour, just with respect to another algebraic state.
This way Savvidou's construction of the Liouville and action operators acting on the symmetric Fock--Cook Hilbert space $\FFF[\H]$ and generating two corresponding types of unitary temporal evolution becomes replaced by the algebraic construction of liouvillean and action operators generating the $*$-automorphisms of corresponding representations of canonical core W$^*$-algebra $\core$. The main structures of each approach, the Fock--Cook Hilbert space $\FFF[\H]$ and the Falcone--Takesaki W$^*$-algebra $\core$, are constructed in a functorial way from the corresponding underlying ingredients of the given quantum theoretic model: the Hilbert space $\H$ and the W$^*$-algebra $\N$, respectively (for a functorial description of Fock space construction see e.g. \cite{Blute:Panangaden:Seely:1993}). Both approaches show that every quantum theoretic model is generically equipped with two different types of \textit{unitary} temporal evolution: the `kinematic' automorphism and the `causal' automorphism. Moreover, while the particular quantitative form of the latter evolution can be arbitrarily postulated, the quantitative form of the former is determined by the particular (quantitative) representation of an abstract algebra that is used in the given model. Both approaches enable to incorporate these two different unitary temporal evolutions into a single unifying unitary `action' evolution. Both approaches enable also to describe the generators of the `kinematic' and `action' evolutions in terms of operators acting on the `temporal' space $L_2(\RR,\dd t)$. In the case of the Hilbert space based approach, all these automorphisms are generated by the corresponding hamiltonian operators, while in the case of algebraic approach the quantitative representations of all these automorphisms are generated by the corresponding standard liouvillean operators.
However, apart from these similarities, there are also important differences between those two approaches. In particular, the representation of the histories algebra on the Fock--Cook space is unique, up to unitary equivalence, only for hamiltonians which have a form specified by the Araki theorem \cite{Araki:1960}. For a general hamiltonian there is no possibility to guarantee the uniqueness (up to unitary equivalence) of the Fock--Cook representation of the histories algebra of the Fock--Cook Hilbert space $\FFF[\H]$. In contrast to this, the representation of a core algebra $\core$ in terms of a crossed product algebra $\N\rtimes_{\sigma^\omega}\RR$ acting on $\H_\omega\otimes L_2(\RR,\dd t)$ is uniquely determined, up to unitary equivalence, by any particular choice of a state $\omega\in\N^+_{\star0}$, \textit{which is considered as part of the definition of the model}. Moreover, while in both approaches the initial `dynamic' unitary automorphism can be postulated as an arbitrary additional component of the model, only in the algebraic approach can the resulting (dual) `action' automorphism be considered as purely `kinematic' (modular) automorphism, related to the change of the algebraic state. The change of unitary description of the temporal behaviour of the quantum theoretic model $(\N,\omega)$ equipped with an additional `unitary' $*$-automorphism $\alpha$ is completely determined by the quantum theoretic model $(\N,\omega)$ and the map $\omega\mapsto\phi$, \textit{which can be considered as a part of the definition of the model}. There is no corresponding result of such type in the Hilbert space based approach to quantum histories. We consider these two results as an important suggestion in favour of the change of perspective on the role of unitary temporal behaviour of quantum theoretic models. Stating it briefly, instead of \textit{postulating the hamiltonian} as an independent component of quantum theoretic model and later perturbing it (what seems to be the only method within the frames of the Hilbert space based approach to mathematical foundations of quantum theory), an algebraic approach allows to \textit{derive the liouvillean} that characterises the unitary temporal behaviour, given the information about change of state. The change between two identical quantitative Hilbert spaces equipped with the same quantitative representation of the operator algebra becomes replaced by the change between two different (but faithful) algebraic states which correspond to two different (but unitarily equivalent) quantitative representations.
In the similar way as in Savvidou's Hilbert space based formulation: 1) when the generators of `kinematic' and `causal' automorphisms are joined into the new `action' generator $K_\omega+L_\alpha$, the reference to two \textit{different} temporal parametrisations of $\alpha$ and $\sigma^\omega$ disappears (the choice of rescaling of the time parameter between $\alpha_t$ and $\sigma^\omega_s$ was implicitly set above to be $t=s$, however any scalar relationship $t=\lambda s$, $\lambda\in\RR$, will work, and any of such choices corresponds to the choice of a specific section of a $U(1)$ bundle for the $-[\log\Delta_\omega,\,\cdot\,]$ connection); 2) the resulting description of temporal behaviour is an unitary automorphism which does not possess any explicit relationship with the von Neumann--L\"{u}ders nonunitary type of temporal behaviour. In consequence, the above construction is insufficient to deal with the problem of algebraic reformulation of the Anastopoulos--Savvidou histories description of the geometric phase. It seems that the idea of construction of localised unitary evolution without taking into account more specific information about the changes of local geometry of quantum state spaces is just not enough.
In particular, the restriction of description of temporal behaviour of quantum models to $*$-au\-to\-mor\-ph\-isms implies the preservation of spectrum: if $\alpha$ is a $*$-automorphism of a C$^*$-algebra $\C$, then
\begin{equation}
\alpha((z\II-x)^{-1})=(z\II-\alpha(x))^{-1}\;\forall x\in\C\;\forall z\in\CC.
\label{automorphisms.preserve.spectrum}
\end{equation}
The \textit{decision} that description of temporal behaviour of quantum theoretic models should be provided in terms of the $*$-automorphisms removes \textit{a priori} the possibility to describe the changes of the eigenvalues in time. This restriction is imposed by the `spectral principle' which is a part of an idealistic ontological interpretation of a quantum theoretic formalism. However, it is too strong for many practical purposes.\footnote{In consequence, the range of applicability of the `unitary' framework is usually extended by the use of additional mathematical tools and techniques, like parameter fitting or renormalisation, which are explicitly nonunitary, but are not considered as part of the content of the quantum theoretic model.} We do not see any reasons for accepting this situation other than wish of securing the validity of some very particular interpretation.\footnote{According to this interpretation, the eigenvalues of operators can be specified with infinite precision (at least in principle) by the quantitative results of experimental procedures, hence they have ontological meaning, and correspond to the `possessed properties' of `ontological quantum systems', as opposed to `postulated properties' of `quantum theoretic models'. Unfortunately, this idealistic ontological interpretation does not apply to any actual experimental situation without \textit{additional} techniques of processing of the quantitative results of experimental procedures which render the fundamental assumption of this interpretation false (or at least meaningless).} In order to develop the framework which bypasses the double standards of dealing with description of experimental information and temporal behaviour, one has to consider the nonunitary description of temporal behaviour as a valid constituent of the structure of quantum theoretic models. The nonunitary changes of quantitative representation can be determined in an algebraic approach by nonunitary changes of algebraic states. Hence, in order to provide such nonunitary description, one has to introduce some method of `updating' the algebraic state that corresponds to a specified information.
Note that in the current Section we can replace the use of a standard liouvillean of a \textit{global} W$^*$-dynamical system $(\N,\RR,\alpha)$ by a \textit{local} quantum Poisson system $(\M(\N,\B),h)$, using the perturbations of a local liouvillean $\pi_\omega(\DF_\omega h)$ by $K_\omega$ (and, dually, $K_\omega$ by $\pi_\omega(\DF_\omega h)$). This localises the linearity of a flow to a tangent space, allowing for a nonlinear generating function for the `causal' part of the dynamics. As a result, the above discussion can be applied to \df{local action operator} and its dual. Yet, the localisation does not change the qualitative conclusions drawn from the above discussion, so we have chosen to keep the presentation in maybe a bit more familiar global language. Because $\B$ is a Banach Lie algebra and $\M(\N,\B)$ is constructed as a Banach Lie--Poisson submanifold, locally the generators of causal dynamics will be always linear, so will be the flow determined by the Lie--Poisson bracket $\{\cdot,\cdot\}$.
Hence, in order to get nonlinear contributions to the effective dynamics, some other geometric structure, beyond $\pi_\omega(\DF_\omega h)$ and $K_\omega$, has to be used. In particular, from the discussion in Section \ref{Orlicz.Br\`{e}gman.section} it follows that, given a bundle of the GNS Hilbert spaces over a trajectory of faithful normal states, a natural parallel transport operator is given by the standard unitary equivalence $V_{\phi,\psi}$. The corresponding connection $\nabla^{1/2}$ is a Levi-Civita connection of the Wigner--Yanase riemannian metric, and the local geodesic `free fall' along $\nabla^{1/2}$ corresponds to a norm projection in the (standard representation) Hilbert space, associated to a local continuous-time projective measurement. In this sense, the connection $\nabla^{1/2}$ \textit{locally} implements this what was an original intention of the nonhamiltonian part of the histories functional, as exposed by the equations \eqref{trace.class.operator.omega}, \eqref{eq.Bargmann}, \eqref{traceless.berry}, and \eqref{geometric.phase.in.decoherence.functional}. In a discussion of Savvidou's action operator in Section \ref{geometric.phase.histories} we have noticed that it does not restore this aspect of histories functional. Because the above algebraic action operator provides an exact algebraic generalisation of Savvidou's formulation, it shares the same feature. One can think of Savvidou's `Liouville' operator $V$ and modular hamiltonian $K_\omega$ as generators of `intrinsic' kinematic automorphisms of, respectively, a single Hilbert space $\H$ or a single W$^*$-algebra $\N$. These should be taken into account when one provides a \textit{spatial representation} of the `intrinsic' causal automorphism of $\H$ or $\N$, respectively, in terms of a bundle of copies of $\H$ or $\N$ over a real line $\RR$. However, neither $\ee^{\ii s V}$ or $\sigma^\omega_s$ can be understood as representing the \textit{changes} between \textit{quantitatively distinct} Hilbert spaces, corresponding to different measurements. In case of the Hilbert space based histories, this change requires to use the Berry connection, while in the algebraic framework (as implemented systematically in Sections \ref{local.geodesic.propagation} and \ref{from.local.gauge.to.lagrangeans.section}) this requires to use the connection $\nabla^{1/2}$, corresponding to the parallel transport operator $V_{\phi,\psi}$.\footnote{As opposed to $K_\omega$, $V_{\phi,\psi}$ cannot be used in the perturbation of the standard liouvillean, because it is a mapping between two different standard representations, not an operator acting on a single Hilbert space.}
\subsection{W$^*$-geometric quantum histories\label{algebraic.quantum.histories}}
In \cite{Kostecki:2014,Hellmann:Kaminski:Kostecki:2016} we showed that the nonunitary change of quantum states due to L\"{u}ders' rule (and other rules, see also \cite{MunkNielsen:2015}) is a special case of the constrained minimisation of the quantum relative entropy functional $D_0$. Moreover, the local smooth geometry of the quantum models can be derived (under mild conditions) as the subsequent terms of the Taylor expansion of any smooth information distance $D$. This leads to the idea \cite{Kostecki:2007:qht,Kostecki:2008:aqh,Kostecki:2010:AIP} to use quantum relative entropy as a general tool of generating nonunitary evolution of quantum states that takes into account the geometric structure of the quantum model.
Taking into account the above discussion, we consider the connection $\nabla^{D_\Psi}$ derived from a Br\`{e}gman distance $D_\Psi$ to be the appropriate replacement for the Berry connection used in the Anastopoulos--Savvidou analysis, as well as for our own use of $K_\omega$ above. However, we are unfortunately lacking the mathematical structure that would allow us to practically use other connection then $\nabla^{1/2}$, thus below we will consider only this possibility. On the other hand, the affine Killing hamiltonian vector field used in Sections \ref{geometric.phase.histories} and \ref{Hilbert.space.geometry.path.integrals} can be replaced by an arbitrary hamiltonian function $h$ on $\M(\N)$, provided the latter is equipped with a BLP manifold structure. Those two substitutions allow us to state the W$^*$-geometric versions of the formulas \eqref{hopfbundlephase} and \eqref{act}. On the differential geometric level (and ignoring for a moment a functional analytic incompatibility between BLP, GNS, and quantum information geometric manifold structures), the effective dynamics is described by the $1$-form
\begin{equation}
{\cal F}=\ddd h(\phi)-\ddd_{\nabla^{1/2}}(\phi),
\label{causal.inferential.form}
\end{equation}
where $\ddd_{\nabla^{1/2}}$ is a connection form of the Levi-Civita connection $\nabla^{1/2}$. This formula states that local causal dynamics and local inferential dynamics participate to the same extent in the effective local dynamics. Hence, neither inference nor causality is considered as more fundamental. The form ${\cal F}$ can be considered as a localisation of the causal inference instrument \eqref{global.hamiltonian.entropic.map} that does not impose the ordering on composition of causal and inferential dynamics.
In order to generalise the additional regularising riemannian term in \eqref{histories.functional.with.metric} and \eqref{Klauder.path.integral}, let us consider the expansions
\begin{equation}
D(\phi+\varepsilon v,\phi)=\frac{\varepsilon^2}{2}\gbold^D_{ab}(\phi)v^av^b+\mathcal{O}(\varepsilon^3)
\end{equation}
and \cite{Rodriguez:2003}
\begin{equation}
D_\gamma(\phi+\varepsilon v,\phi)=\frac{\varepsilon^2}{2}\gbold^{D_\gamma}_{ab}(\phi)v^av^b+\frac{\varepsilon^3}{6}(\Gamma^{\nabla^0}_{abc}+\Gamma^{\nabla^\gamma}_{abc}+\Gamma^{\nabla^1}_{abc})v^av^bv^c+\mathcal{O}(\varepsilon^4),
\end{equation}
where $\Gamma^\nabla_{abc}$ are the Christoffel symbols of the corresponding connections. Setting $\varepsilon ^2=\frac{1}{\upsilon}$ suggests us to use the quantity
\begin{equation}
P^{\pathgamma,\epsilon}_{k,1,1}:=\ee^{-k\int_\pathgamma \dd t D\left(\phi(t)+\varepsilon\frac{\dd\phi(t)}{\dd t},\phi(t)\right)}\sqrt{\det(\gbold^D)}
\label{local.entropic.prior}
\end{equation}
as a generalised regulariser, where $k\in\RR^+$ is a constant. We interpret this object as a \textit{local} quantum entropic prior: an expression for a local prior measure representing user's ignorance about the choice of propagation between neighbouring states along a specific trajectory $\pathgamma:[0,s]\ra\M(\N)$. See Section \ref{entropic.priors.section} for a discussion of entropic priors in the commutative case. More specifically, \eqref{local.entropic.prior} is a localised quantum version of the $P_{k,1,1}$ prior. For $D_{1/2}(\sigma,\rho)=\frac{1}{2}\n{\sqrt{\sigma}-\sqrt{\rho}}^2_\H$, this corresponds to integrating against a local gaussian measure. The global Jeffreys prior $\sqrt{\det(\gbold^{D_{1/2}})}$ appears already in the Klauder--Maraner formula \eqref{Klauder.Maraner.metric.variation}, as a constant \eqref{Klauder.Maraner.prior}, which sets a relationship between local measure of uncertainty of inference and local generator of causal dynamics.
Thus, we propose to generalise the formula \eqref{Klauder.path.integral} to
\begin{align}
\lim\limits_{\varepsilon\ra+0}
\int\mathcal{D}\phi(\cdot)
&\ee^{\ii\int_\pathgamma\dd t\s{\Omega_{\phi(t)},\ddd_{\nabla^{1/2}}(\phi)\Omega_{\phi(t)}}_{\H_{\phi(t)}}}
\ee^{-\ii\int_\pathgamma\dd t \s{\Omega_{\phi(t)},\pi_{\phi(t)}(\DF_{\phi(t)}h)\Omega_{\phi(t)}}_{\H_{\phi(t)}}}\nonumber\\
&\cdot\ee^{-k\int_\pathgamma \dd t D\left(\phi(t)+\varepsilon\frac{\dd\phi(t)}{\dd t},\phi(t)\right)}\sqrt{\det(\gbold^D)}.
\label{RPK.propagator}
\end{align}
As discussed in Section \ref{local.quantum.dynamics.intro.section}, there are some legitimate reasons to believe that at least at the level of the second order approximation of an entropic prior, the above formula can receive an exact foundation by means of stochastic integration process. Yet, without a proof of this conjecture, the formula \eqref{RPK.propagator} has now a status of a heuristic proposal. However, most of the applications of path integrals in theoretical physics have precisely the same status (the exactness of the Daubechies--Klauder formula \eqref{Klauder.path.integral} is more an exception than a rule).
The differences between the formulas \eqref{RPK.propagator} and \eqref{liouv.prop} correspond to the standard differences between algebraic and path integral formulations. Both formulations admit introducing additional local gauge and source terms, so they can be used to study various applied models. Taking a closer look at the Daubechies--Klauder formula \eqref{Klauder.path.integral}, one may note that the left side of this equation is formulated without taking into consideration the possible changes of the GNS representation along the states, because the coherent vector states are considered to be defined in a single Hilbert space. If such changes would be considered (as we do it here), then an operator $\ee^{-\ii Hs}$ in \eqref{Klauder.path.integral} should be multiplied from left by a corresponding standard unitary transition operator. This leads us to the conjecture:
\begin{equation}
\eqref{liouv.prop} = \eqref{RPK.propagator},
\label{SS}
\end{equation}
\textit{if} the left hand side of this equation is evaluated in terms of the MCP, instead of the GNS, Hilbert space. While this conjecture is quite heuristic, it seems to be a legitimate candidate for a W$^*$-geometric analogue of the Daubechies--Klauder propagator formula \eqref{Klauder.path.integral}. A development of a suitable stochastic calculus allowing for an exact mathematical treatment of \eqref{RPK.propagator}, as well as the proof that the proposed construction of MCP Hilbert space is well defined, are the necessary conditions to approach the problem of proving this conjecture. Yet, in Section \ref{new.local.quantum.foundations}, based on the discussions in Sections \ref{local.infogeometry.histories.intro.section}, \ref
{Orlicz.Br\`{e}gman.section}, and \ref{section.dynamics.renormalisation}, we will propose another, more geometric approach to the equivalence intended behind the formula \eqref{SS}, without requiring equality on the level of Hilbert bundles.
\subsection{Appendix: Entropic priors on statistical models\label{entropic.priors.section}}
To simplify the notation, whenever we will use the coordinate-dependent formulas in this Section, we will assume that the statistical model $\M:=\M(\X,\mho(\X),\tmu)\subseteq L_1(\X,\mho(\X),\tmu)^+$ is equipped with a global coordinate system $\theta:\Theta\ra\M$, where $\Theta\subseteq\RR^n$ is open.
The entropic prior $P_{k,\alpha,\beta}$ is defined in order to provide the general \cytat{statistical representation of the [notion of the] vacuum of information in a given hypothesis space} \cite{Rodriguez:1999}. Every probability measure encodes some knowledge, hence the notion of the `vacuum of information' has also to refer to some given knowledge which defines it. The `vacuum of information' is relative to the given information manifold, and as such it is defined to depend on the invariant volume measure on the information manifold, the Jeffreys prior \cite{Jeffreys:1946}
\begin{equation}
\J(\theta)=\sqrt{\ab{\det\gbold(\theta)}}\dd\theta_1\land\ldots\land\theta_n,
\end{equation}
which distributes prior probability over all hypothesis space, as well as on the initial reference density $p_0(x,\theta)=p_0(x|\theta)P(\theta)$ on $\X\times\Theta$ which sums up all additional reference knowledge (e.g., the quantitative results of previous experimental procedures) which will be encoded into the structure of the vacuum of information.
In particular, when the reference knowledge consists only of model-independent information encoded in the density $m(x)$, then the reference density factorises to $p_0(x,\theta)=m(x)P(\theta)$. The entropic prior build with respect to such factorisation encodes the `vacuum of information' regarding the dependence between the parameters $\Theta$ of the model $\M$ and the initial knowledge $m(x)$ about the data ${\X}$.
In the nonparametric formulation based on $D_\gamma$ entropies for $\gamma\in[0,1]$, the \df{entropic prior} is defined as such $P_{k,\alpha,\beta}$ which minimises the functional \cite{Rodriguez:2003}
\begin{equation}
\inf_P(k\int P(p)D_\alpha(p,p_0)+D_\beta(P,\J)),
\label{xep}
\end{equation}
where $(\alpha,\beta)\in[0,1]\times[0,1]$, the distance $D_\alpha$ is calculated over $\X\times\Theta$, the distance $D_\beta$ is calculated over $\Theta$, and the scalar $k\geq0$ parametrises the preference of $P$ over $\J$ with respect to the reference density $p_0$. In the parametric formulation, the functional minimised in \eqref{xep} reads
\begin{equation}
k D_\alpha(p(x|\theta)P(\theta),p_0(x|\theta)P(\theta))+D_\beta(P(\theta),\sqrt{\ab{\det\gbold(\theta)}}\dd\theta).
\end{equation}
By definition, the entropic priors are minimisers of the estimation by expected loss (decision) functional $D_\alpha$ under the constraint that the $D_\beta$-distance of entropic prior $P$ from volume measure $\J$ does not exceed some constant value. In other words, they express the degree of confidence in the reference distribution, relatively to degree of confidence in volume measure invariance of $P$.
The general solution of the above minimisation problem takes the form
\begin{equation}
P_{k,\alpha,\beta}=\tilde{P}_{k,\alpha,\beta}\J,
\end{equation}
where $\tilde{P}_{k,\alpha,\beta}(\theta)$ is a scalar density which, up to normalisation, is equal to \cite{Rodriguez:2003,Snoussi:2005,Snoussi:2007}
\begin{equation}
\tilde{P}_{k,\alpha,\beta}(\theta):=
\left\{
\begin{array}{ll}
1+k(1-\beta)D_\alpha(p_\theta,p_0)^{-\frac{2}{1+\beta}}
&:\;\beta\neq1,\\
\exp(-k D_\alpha(p_\theta,p_0))
&:\;\beta=1,
\end{array}
\right.
\label{entropic.priors}
\end{equation}
under the condition that
\begin{equation}
k_{min}:=\inf\{k\geq 0\mid\textstyle\int P_{k,\alpha,\beta}<\infty\}
\end{equation}
exists and $k\geq k_{min}$. If $k_{min}=0$, then ${P}_{0,\alpha,\beta}$ is Jeffreys prior. On the other hand, ${P}_{\infty,\alpha,\beta}$ is the Dirac delta concentrated on $p_0$. Moreover, $\tilde{{P}}_{k,\alpha\in\{0,1\},\beta\in[0,1[}$ is a multivariate Student $t$ distribution, $\tilde{{P}}_{k,\alpha,0}$ is a generalised multivariate Cauchy distribution, $\tilde{{P}}_{k,\alpha,1}$ is a minimum $D_\alpha$ prior density, and
\begin{equation}
\tilde{{P}}_{k,1,1}(p)=\ee^{-kD_1(p,p_0)}
\end{equation}
is a maximum relative entropy (= minimum $D_1$ distance) prior density. If $p_0$ is taken to be the Bernoulli--Laplace uniform prior, then $\tilde{{P}}_{1,1,1}(p)=\ee^{-\entropy_{GS}(p)}$ is a maximum Gibbs--Shannon entropy density (Jaynes prior \cite{Jaynes:1957}). A maximum Gibbs--Shannon entropy density can be also recovered as $\tilde{{P}}_{\infty,1,1}(p)$ if $p_0$ is an element of the exponential family. When no reference distribution (no background information) is specified, then ${P}_{k,\alpha,\beta}$ reduces trivially to Jeffreys' prior. If the reference measure $p_0$ does not belong to a manifold $\Q$ on which the minimisation procedure generating the entropic prior is evaluated, then $k\left(1-\beta\right)$ factor for $\beta\neq1$ is replaced by a more general scalar quantity $\tilde{k}$, dependent on the projection of $p_0$ on $\Q$ (for details, see \cite{Snoussi:2005}).
If $p_0$ is a maximum entropy distribution obtained under some given constraints, then the entropic prior quantifies to what extent the densities other than $p_0$ (within a given model) are less probable or less reliable. The reliability of $p(x|\theta)$ other than $p_0(x|\theta)$ decreases exponentially with the deviation of $p(x|\theta)$ from $p_0(x|\theta)$, and the sensitivity for this exponential decrease is controlled by the constant $k$. In general, the larger $k$ is, the stronger is the impact of reference distribution (assumed background information) on the inference provided with respect to the `vacuum of information' ${P}(\theta)$. If the reference hypothesis is built up from knowledge independent of the model (encoded in $m(x)$), then the larger $k$ is, the more preference is given to this independent knowledge. On the other hand, the smaller $k$ is, the more inference based on ${P}(\theta)$ will depend on distributions other than $p_0(x|\theta)$, so it becomes easier for the eventual `noise' in constraints to be taken by inference to be a `signal'. So, while $k\ra 0$ smoothens the prior, $k\ra\infty$ sharpens it. Jeffreys' and Jaynes' (maximum Gibbs--Shannon entropy) priors are just two extreme points of this scale.
The entropic priors $P_{k,1,1}$ are the only entropic priors on probabilistic manifold which are based on the measure of distance which is coordinate invariant, local, consistent for independent subsystems and additive. This is a characterisation of $D_1$ as a unique distance functional on the space of normalised probability densitites used for the purpose of probability updating, as provided in \cite{Shore:Johnson:1980,Shore:Johnson:1981,Johnson:Shore:1983}. Hence, they are the unique priors which encode the notion of coordinate invariant, local, additive, and independent subsystem consistent `vacuum of information'. For more discussion on the topic of entropic priors, see \cite{Skilling:1989,Caticha:2001:a,Caticha:Preuss:2003,Caticha:Preuss:2004,Rodriguez:1991,Rodriguez:2003,Snoussi:MohhamadDjafari:2003,Snoussi:2005}.
\section{Information theoretic renormalisation\label{information.theoretic.renormalisation}
In this section we will analyse another application of quantum information geometry for the purpoes of inference over quantum models. We start from discussion of the Jaynes--Mitchell source theory \cite{Mitchell:1967,Jaynes:1985:scattering,Jaynes:1993,Grandy:1987,Grandy:2008}, which describes the general continuous changes of information states of an exponential model driven by the sources of information. Next, we discuss Favretti's \cite{Favretti:2007} information geometric generalisation of this theory to the setting of dually flat information manifolds. It allows for a strictly geometric implementation of the idea of renormalisation of dynamics by reduction of dimensionality of the model by fixing the control parameter, which is provided on the space of information states (as opposed to the space of functions or operators). Our original contribution amounts to an observation that the Jaynes--Mitchell--Favretti approach is canonically related to the use of Br\`{e}gman distances, so it can be used to \textit{locally} approximate information dynamics on an arbitrary manifold of quantum states. We discuss how this setting allows to use the departure of local geometry from the dually flat smooth geometry (generated by quantum Br\`{e}gman distances) as the geometric description of multiparameter nonlinear quantum control and renormalisation problems. We also introduce another type of geometric renormalisation of inferential dynamics of quantum states, which describes situations where none of specific control (covariate) parameter is fixed, but the quantum model is subjected to the action of completely positive maps. This procedure is based on the use of $D_\fff$ distances as well as associated \textit{contraction coefficients}, introduced by Ruskai et al \cite{CIRRSZ:1993,Choi:Ruskai:Seneta:1993,Ruskai:1994,Lesniewski:Ruskai:1999}.
\subsection{Jaynes--Mitchell source theory\label{section.JM.source.theory}}
Consider first an arbitrary statistical model $\M(\boole)$ over finite boolean algebra\footnote{For the reasons of mathematical fanciness, we occasionally consider the sets $\M(\boole)\subseteq L_1(\boole)^+$ of finite positive measures over localisable boolean algebras $\boole$, but this is completely equivalent to consideration of localisable measure spaces $\M(\X,\mho(\X),\tmu)\subseteq L_1(\X,\mho(\X),\tmu)^+$.} $\boole$ (with $m\in\NN$ denoting the number of elements of $\boole$), a set $\{f_k\}_{k=1}^{n}$ of functions $f_k:\boole\ra\RR$ with $n\in\NN$, and a change in the expectation $\e{f_k}_p$, caused by the independent changes in both $f_k(x_i)=:f_k^i$ and $p(x_i)=:p_i$,
\begin{equation}
\delta\e{f_k}=\sum_{i=1}^mp_i\delta f_k^i+\sum_{i=1}^mf_k^i\delta p_i.
\label{change.of.expectation}
\end{equation}
If $f_k$ depends on some additional parameters $r=(r_1,\ldots,r_l)$, such that
\begin{equation}
\delta f_k(x_i,r)=\sum_{j=1}^{l}\frac{\partial f_k(x_i,r)}{\partial r_j}\delta r_j,
\label{add.param}
\end{equation}
then the first term of \eqref{change.of.expectation} reads
\begin{equation}
\sum_{i=1}^mp_i\delta f_k^i=\e{\delta f_k}_p=\e{\sum_{j=1}^l\frac{\partial f_k}{\partial r_j}\delta r_j}_p=:\delta W_k.
\label{work.def}
\end{equation}
We denote the second term of \eqref{change.of.expectation} by $\delta Q_k$, so
\begin{equation}
\delta Q_k:=\sum_{i=1}^mf_k^i\delta p_i=\delta\e{f_k}_p-\e{\delta f_k}_p,
\label{Q.def}
\end{equation}
which gives
\begin{equation}
\delta\e{f_k}_p=\delta W_k+\delta Q_k.
\label{gen.1st.law.1}
\end{equation}
Consider now an \df{exponential family} defined as an $n$-dimensional parametric probabilistic manifold \cite{Darmois:1935,Koopman:1936,Pitman:1936}\rpktarget{MEXP}
\begin{equation}
\M_{\mathrm{exp}}(\X,\mho(\X),\tmu):=
\{
p(\xx,\theta):=
\exp(
-\log Z(\theta)-\textstyle\sum_{i=1}^n\theta^if_i(\xx)
)\mid
\theta:=(\theta^1,\ldots,\theta^n)\in\Theta\subseteq\RR^n
\},
\label{commutative.exponential.model}
\end{equation}
where $f_i:\X\ra\RR$ are assumed to be arbitrary functions, linearly independent of each other and of the constant function $1$ (this guarantees that $\theta\mapsto p(\theta)$ is one-to-one and that the matrix $\gbold_{ij}$ is invertible \cite{Wehrl:1978}),
\begin{equation}
\log Z(\theta):=\log\int_\X \tmu(\xx)\exp\left(-\sum_{i=1}^n\theta^if_i(\xx)\right)
\label{partition.function.eq}
\end{equation}
is a factor arising from normalisation condition $\int_\X\tmu(\xx)p(\xx,\theta)=1$, called a \df{Massieu functional} \cite{Massieu:1869,Massieu:1876}, while $\Theta\subseteq\RR^n$ is supposed to be such open set that the integral in \eqref{partition.function.eq} converges. The study of geometric properties of this family provided an original stimulus for development of information geometry \cite{Chencov:1966,Chencov:1972,Efron:1978,Amari:1982}. In particular, Chencov found \cite{Chencov:1966,Chencov:1968} that the finite dimensional exponential families are geodesic surfaces of $\nabla^0$-connections and admit the generalised pythagorean equation \eqref{generalised.pythagore.Bregman} for the Kullback--Leibler distance.
If $\dim\X=:m<\infty$, then $\int_\X\tmu(\xx)k(\xx)=\sum_{j=1}^mk(\xx_j)$ for any $k:\X\ra\RR$. In such case $\M_{\mathrm{exp}}(\X,\mho(\X),\tmu)$ can be characterised in terms of the Gibbs--Jaynes \cite{Gibbs:1902,Jaynes:1957} procedure of maximisation of the \df{Gibbs--Shannon entropy} \cite{Gibbs:1902,Shannon:1948,Shannon:Weaver:1949}
\begin{equation}
\entropy_\GS(p):=-\sum_{j=1}^mp(\xx_j)\log p(\xx_j)
\rpktarget{SGS}
\end{equation}
subject to constraints $F(p)$ given by
\begin{equation}
\left\{
\begin{array}{l}
\sum_{j=1}^mp(\xx_j)1=1,\\
\sum_{j=1}^mp(\xx_j)f_i(\xx_j)=\eta_i,
\end{array}
\right.
\label{exp.family.constraints}
\end{equation}
with $\eta:=(\eta_i)\in\Xi\subseteq\RR^n$. The maximum value attained by $\entropy_\GS$ for a given $(\eta_i)$ (or, equivalently, for a given $(\theta^i)$), reads
\begin{equation}
\entropy_\GS(p(\theta))=\log Z(\theta)+\sum_{i=1}^n\theta^i\eta_i.
\label{max.ent.equation}
\end{equation}
If $p$ belongs to an exponential family with $\lambda_k:=\theta^k$, $k\in\{1,\ldots,n\}$, then the corresponding change in entropy reads
\begin{align}
\delta\entropy_\GS&=\delta\log Z(\lambda)+\delta\left(\sum_{k=1}^n\lambda_k\e{f_k}_p\right)\notag\\
&=-\frac{1}{Z}\left(\sum_{k=1}^n\delta\lambda_k f_k^i+\sum_{k=1}^n\lambda_k\delta f_k^i\right)\ee^{-\sum_{k=1}^n\lambda_k f_k^i}+\sum_{k=1}^n\delta\lambda_k\e{f_k}_p+\sum_{k=1}^n\lambda_k\delta\e{f_k}_p\notag\\
&=\sum_{k=1}^n\lambda_k(\delta\e{f_k}_p-\e{\delta f_k}_p)=\sum_{k=1}^n\lambda_k\delta Q_k.
\label{entropy.and.work}
\end{align}
Due to \eqref{add.param}, $\sum_{k=1}^n\lambda_k\delta Q_k(\e{f_{\widetilde{k}}}_p,r)$ is an exact differential of $\entropy_\GS(\e{f_{\widetilde{k}}}_p,r)$, even if $\delta Q_k(\e{f_{\widetilde{k}}}_p,r)$ is not an exact differential of any function. Thus, \eqref{entropy.and.work} is equivalent to
\begin{equation}
\delta\entropy_\GS=\sum_{k=1}^n\lambda_k\delta\e{f_k}_p-\sum_{k=1}^n\sum_{j=1}^l\lambda_k\e{\frac{\partial f_k}{\partial r_j}}_p\delta r_j
\label{gen.1st.law.2}
\end{equation}
and
\begin{equation}
\left.\frac{\partial\entropy_\GS(\e{f_k}_p,r)}{\partial\e{f_k}_p}\right|_{r=\mathrm{const}}=\lambda_k.
\label{gen.1st.law.3}
\end{equation}
These results are completely analogous to the first law of equilibrium thermodynamics. The change of `information work' $W_k$ is dependent only on the changes of quantity $f_k$. The change of `information heat' $Q_k$ and of absolute entropy $\entropy_\GS$ depends only on the change of information state provided by probability $p$. While having the same mathematical form, the above results are completely independent of thermodynamics and hold for any exponential family. The first law of equilibrium thermodynamics is just a \textit{special case} of the above result.
Let us now consider a three dimensional exponential family
\begin{equation}
\M_{\mathrm{exp}}(\X,\mho(\X),\tmu;\Theta):=\left\{p(\lambda_A,\lambda_B,\lambda_C)=\frac{1}{Z}\ee^{-\lambda_Af_A(x)-\lambda_Bf_B(x)-\lambda_Cf_C(x)}\mid(\lambda_A,\lambda_B,\lambda_C)\in\Theta\right\},
\end{equation}
where $f_A,f_B,f_C\in L_\infty(\X,\mho(\X),\tmu)$, and $\Theta\subseteq\RR^3$ is some fixed open set. Let the change of information be described by $\e{f_A}_p\ra\e{f_A}_p+\delta\e{f_A}_p$ with the additional conditions that the possible changes of $\e{f_B}_p$ are left unconstrained ($\delta\lambda_B=0$ but we allow $\delta\e{f_B}_p\neq0$), and it is known that $\e{f_C}_p$ does not change ($\delta\e{f_C}_p=0$ but we allow $\delta\lambda_C\neq0$). The quantity $\e{f_A}_p$ is called a `driving variable'. Thus, we consider a \textit{source-and-response problem} with an \textit{additional control variable}:
\begin{align*}
\delta\e{f_A}&=0,\;\;\delta\lambda_A\neq0\;\;\mbox{`driving variable' (source parameter)}\\
\delta\e{f_B}&\neq0,\;\;\delta\lambda_B=0\;\;\mbox{`information heat bath' (response parameter)}\\
\delta\e{f_C}&=0,\;\;\delta\lambda_C\neq0\;\;\mbox{`control variable' (additional source)}
\end{align*}
Following Mitchell and Jaynes, will now provide an answer to a question: how the presence of the second source affects the relationship between first source and the response parameter?
Given \textit{some} finite dimensional statistical model $\M(\boole)$ parametrised by a coordinate system $\lambda:\M(\boole)\ra U\subseteq\RR^n$ with $n:=\dim(\M(\boole))$, then the general form of the variation of an expectation functional $\e{f}_p$ for some element $p(\lambda_0)\in\M(\boole)$ reads
\begin{equation}
\delta\e{f}_p
:=\e{f}_{p(\lambda)}-\e{f}_{p(\lambda_0)}
=\sum_{i=1}^\infty
\frac{1}{i!}\sum_{(j_1,\ldots,j_i)}
\left.
\frac{\partial^i\e{f}_{p(\lambda)}}%
{\partial\lambda_{j_1}\cdots\partial\lambda_{j_i}}
\right|_{\lambda=\lambda_0}
\delta\lambda_{j_1}\cdots\delta\lambda_{j_i}.
\label{infinite.perturbation.covariance.expansion}
\end{equation}
The first order term of \eqref{infinite.perturbation.covariance.expansion} (corresponding to the linear character of variation) reads
\begin{equation}
\delta\e{f}_p=
\sum_{j=1}^n\frac{\partial\e{f}_p}{\partial\lambda_j}
\delta \lambda_j.
\label{first.order.covariance.expansion}
\end{equation}
In the case of exponential model $\M_{\mathrm{exp}}(\boole;\Theta)$, from the equations \eqref{first.order.covariance.expansion} and \begin{equation}
K_{ij}=\frac{\partial^2\log Z(\lambda)}{\partial\theta^i\partial\theta^j}=-\frac{\partial\e{f_i}_p}{\partial\lambda^j}=-\frac{\partial\e{f_j}_p}{\partial\lambda^i},
\end{equation}
it follows that the relationship between `fluxes of information' $\delta\e{f_k}_p$ and `forces of information' $(-\delta\lambda_k)$ can be determined in the first (linear) order by the covariance matrix\footnote{In this terminology $\lambda_k$ play the role of the `potentials of information', but this should not be confused with the `scalar potentials' $\Psi$ and $\Psi^\lfdual$ on hessian manifolds, such as $-\log Z(p)$ and $\entropy_{\mathrm{GS}}(p)$ (which play the role of information discrimination functionals).}
\begin{equation}
\left(\begin{array}{c}\delta\e{f_A}_p\\\delta\e{f_B}_p\\\delta\e{f_C}_p\end{array}\right)=-\left(\begin{array}{ccc}
K_{AA}&K_{AB}&K_{AC}\\
K_{BA}&K_{BB}&K_{BC}\\
K_{CA}&K_{CB}&K_{CC} \end{array}\right)\left(\begin{array}{c}\delta\lambda_A\\\delta\lambda_B\\\delta\lambda_C\end{array}\right).
\label{fff}
\end{equation}
Hence,
\begin{equation}
\delta\lambda_C=-\frac{K_{CA}}{K_{CC}}\delta\lambda_A,
\end{equation}
which means that $\delta\lambda_C$ and $\delta\lambda_A$ are not independent of each other. This is also reflected in the second equation following from \eqref{fff}, namely
\begin{equation}
\delta\e{f_A}_p=-\delta\lambda_A\left(K_{AA}-\frac{K^2_{AC}}{K_{CC}}\right),
\end{equation}
which is equivalent to
\begin{equation}
\frac{\delta\e{f_A}_p}{\delta\lambda_A}=\frac{\partial\e{f_A}_p}{\partial\lambda_A}-\frac{K^2_{AC}}{K_{CC}}.
\end{equation}
If we \textit{decide} to consider only the variables $\e{f_A}_p$ and $\e{f_B}_p$ (removing $\e{f_C}_p$ from the \textit{definition} of the problem), then the covariance matrix of the problem takes the form
\begin{equation}
\left(\begin{array}{c}\delta\e{f_A}_p\\\delta\e{f_B}_p\end{array}\right)=-\left(\begin{array}{cc}
K_{AA}&K_{AB}\\
K_{BA}&K_{BB}
\end{array}\right)\left(\begin{array}{c}\delta\lambda_A\\\delta\lambda_B\end{array}\right).
\label{kv1}
\end{equation}
From the assumption that there are no additional parameters $r$ of control associated with the element $f_A$ of the abstract algebra (that is, $\e{\delta f_A}_p=0$), it follows that $\delta\e{f_A}_p=\delta Q_A$. In such case the above equation turns into
\begin{equation}
\delta\e{f_B}_p=\frac{K_{BA}}{K_{AA}}\delta Q_A.
\end{equation}
Hence, the changes of $\e{f_B}_p$ are driven by the `source of information' $\delta Q_A$. We will call the corresponding evolution of probability distribution $p(\lambda_A,\lambda_B,\lambda_C)\in\M_{\mathrm{exp}}(\boole;\Theta)$ an `information driving'. The number of different variables is not limited to three, but three variables are sufficient to describe all possible \textit{types} of constraints. Mitchell \cite{Mitchell:1967} has shown that the readjustment of expectation values of functions $\{f_k\}$ under driving caused by sources of information can be described the following \textit{equivalent} principles:
\begin{itemize}
\item[i)] expectations uncorrelated with driven variables remain unchanged,
\item[ii)] Lagrange multipliers of unconstrained variables remain unchanged,
\item[iii)] $\entropy_\GS$ is re-maximised under new values of constraints.
\end{itemize}
Now we move to the problem of renormalisation of sources, which amounts to removing the variable $C$ from the definition of the model, while keeping it as a constraint in the allowed transformations of variables (information flows). Consider again the covariance matrix \eqref{fff}, with the constraint $\delta\e{f_C}_p=0$. A direct calculation shows that the relationships between `fluxes' and `forces' of information related with $A$ and $B$ can be completely described by the covariance matrix
\begin{equation}
\left(\begin{array}{c}\delta\e{f_A}_p\\\delta\e{f_B}_p\end{array}\right)=-\left(\begin{array}{cc}
\tilde{K}_{AA}&\tilde{K}_{AB}\\
\tilde{K}_{BA}&\tilde{K}_{BB}
\end{array}\right)\left(\begin{array}{c}\delta\lambda_A\\\delta\lambda_B\end{array}\right),
\label{kv3}
\end{equation}
where
\begin{equation}
\left\{
\begin{array}{l}
\tilde{K}_{AA}:=K_{AA}-K_{AC}K_{CC}^{-1}K_{CA}\\
\tilde{K}_{AB}:=K_{AB}-K_{AC}K_{CC}^{-1}K_{CB}\\
\tilde{K}_{BA}:=K_{BA}-K_{BC}K_{CC}^{-1}K_{CA}\\
\tilde{K}_{BB}:=K_{BB}-K_{BC}K_{CC}^{-1}K_{CB}.
\end{array}
\right.
\end{equation}
The covariance matrix \eqref{kv3} can be thought of as a `renormalised' version of the covariance matrix \eqref{kv1}, where the dependence on an additional correlated information related to variable $C$ is taken into account. Assuming again that $\delta\lambda_B=0$ and $\e{\delta f_A}_p=0$, the predicted change of $\e{f_B}_p$ due to the action of the source $\delta Q_A$ takes the form
\begin{equation}
\delta\e{f_B}_p=\frac{\tilde{K}_{BA}}{\tilde{K}_{AA}}\delta Q_A=\left(\frac{K_{BA}}{K_{AA}}-\frac{K_{BC}}{K_{CC}}\frac{K_{CA}}{K_{AA}}\right)\delta \tilde{Q}_A,
\label{renormalized.equation}
\end{equation}
where
\begin{equation}
\delta \tilde{Q}_A:=\frac{\delta Q_A}{(1-R_{AC}^2)}
\label{reqf}
\end{equation}
is the `renormalised information source strength', while
\begin{equation}
R_{AC}:=\frac{K_{AC}}{(K_{AA}K_{CC})^{\frac{1}{2}}}
\end{equation}
is the correlation coefficient. In other words, the additional constraint ($\delta\e{f_C}_p=0$) imposed on the information related to an additional variable that is correlated with the driving variable ($\e{f_A}_p$) is observed in `renormalisation' of the action of the driving source ($\delta Q_A$) on the dimensionally `reduced' system of variables (without $\e{f_C}_p$):
\begin{equation}
\delta\e{f_B}_p=\frac{\tilde{K}_{BA}}{\tilde{K}_{AA}}\frac{\delta Q_A}{1-R_{AC}^2}.
\label{reduced.system}
\end{equation}
Now, if $R_{AC}$ has a spectral radius smaller than $1$, one can expand the renormalisation factor in \eqref{reqf} and \eqref{reduced.system},
\begin{equation}
(1-R_{AC}^2)^{-1}=\sum_{n=0}^\infty(R_{AC}^2)^n=1+R^2_{AC}+R^4_{AC}+\ldots\;.
\label{Jaynes.expansion}
\end{equation}
Defining the `propagators' $\mathcal{G}_{ij}:=-K_{ij}K_{jj}^{-1}$, one can expand \eqref{renormalized.equation} in the form
\begin{equation}
\begin{array}{c}
\delta\e{f_B}_p=(\mathcal{G}_{BA}-\mathcal{G}_{BC}\mathcal{G}_{CA}+\mathcal{G}_{BA}\mathcal{G}_{AC}\mathcal{G}_{CA}-\mathcal{G}_{BC}\mathcal{G}_{CA}\mathcal{G}_{AC}\mathcal{G}_{CA}+\ldots)\delta Q_A.
\end{array}
\label{expanded.change}
\end{equation}
Thus, the dimensional reduction of the information model which removes from the scope the correlated constrained variable changes the description of information flow, which can be recasted in terms of perturbative series of propagators between the `sources' of driving variables and `information fluxes' of driven variables (`sinks') that are mediated by the ``virtual'' (removed) variable. Comparison of \eqref{first.order.covariance.expansion} with \eqref{infinite.perturbation.covariance.expansion} leads us to note, following Jaynes, that the above effects appear at the \textit{first} level of perturbative expansion in powers of information source strength. In consequence, the corresponding classification of approximated results is provided by the degree of fine tuning of the available information. This brings a clear meaning to the perturbative expansion and renormalisation as the process of classification of approximated description of \textit{the quantitative effects of change of information} with respect to the degree of quantitative refinement of this information (which is given by information source strength). This approximation does not refer to any additional `theoretical' or `physical' dimensional constant parameters and keeps the values and meaning of the constants defining experimental response scales, etc., to be fixed by definition and not entering the scene. Thus, there is also no need for `renormalisation' of these constants, avoiding the conceptual problems which are always caused by such procedure.
\subsection{Favretti's dually flat geometrisation\label{Favretti.section}}
Now we turn to reformulation and generalisation of the Jaynes--Mitchell source theory provided by Favretti \cite{Favretti:2007}. Suppose that $\M(\boole)$ is a probability manifold with $\dim\M(\boole)=n\in\NN$, equipped with the pair of coordinate systems $(\theta,\eta):\M(\boole)\ra\Theta\times\Xi\subseteq\RR^n\times\RR^n$. Let the information about trajectory $p(t)\in\M(\boole)$ be specified as the constraints expressed in terms of both coordinate systems:
\begin{equation}
\left\{
\begin{array}{l}
F_1(\theta(p),t)=0,\\
F_2(\eta(p),t)=0.
\end{array}
\right.
\label{constraints.JF}
\end{equation}
Favretti shows that under additional assumption that $\M(\boole)$ is equipped also with a riemannian metric $\gbold$ and a pair of affine connections $(\nabla^\theta,\nabla^\eta)$ such that $(\M(\boole),\gbold,\nabla^\theta,\nabla^\eta)$ is a dually flat manifold with a dually flat coordinate system given by $(\theta,\eta)$ (see Section \ref{distances.NS.geom.section}), the implicit function theorem allows one to describe geometrically the evolution $p(t)$ quantitatively, in terms of one of these coordinate systems.
Let the scalar potential functions determined by the above dually flat geometry be denoted by $\Psi(\theta):=\Psi\circ\theta$ and $\Psi^\lfdual(\eta):=\Psi^\lfdual\circ\eta$, where $\Psi^\lfdual$ is a Fenchel dual of $\Psi$ with respect to \eqref{selfduality.Rn}. Consider the diagram
\begin{equation}
\xymatrix{&\ar[dl]_{\eta}\M(\boole)\ar[dr]^{\theta}&\\
\Xi\ar@<0.5ex>[rr]^{\LFtrafo^{-1}_\Psi}\ar@<0.5ex>[d]^{\pi_A^\Xi}&&\Theta\ar@<0.5ex>[ll]^{\LFtrafo_\Psi}\ar@<0.5ex>[d]^{\pi_B^\Theta}\\
\Xi_A\ar@<0.5ex>[u]^{(\pi^\Xi_A)^{-1}}&&\Theta_B\ar@<0.5ex>[u]^{(\pi^\Theta_B)^{-1}},}
\label{favretti.diag}
\end{equation}
where $\LFtrafo^{-1}_\Psi:\Xi\ra\Theta$ and $\LFtrafo_\Psi:\Theta\ra\Xi$ are the Legendre transforms given by smooth diffeomorphisms, which are expressed in coordinate-dependent way as
\begin{align}
\theta^i=(\LFtrafo^{-1}_\Psi(\eta))^i&=\frac{\partial}{\partial\eta_i}\Psi^\lfdual(\eta)=:\partial^i\Psi^\lfdual(\eta),\label{LFT.MJF.Phi}\\
\eta_i=(\LFtrafo_\Psi(\theta))_i&=\frac{\partial}{\partial\theta^i}\Psi(\theta)=:\partial_i\Psi(\theta),
\label{LFT.MJF.Psi}
\end{align}
while
\begin{align}
\pi_A^\Xi:\Xi\ni\eta&\mapsto\eta_A\in A\subseteq\RR^k,\label{first.project.fv}\\
\pi_B^\Theta:\Theta\ni\theta&\mapsto\theta^B\in B\subseteq\RR^{n-k},
\end{align}
are projections with
\begin{align}
\eta&=(\eta_A,\eta_B)\in\RR^k\times\RR^{n-k},\\
\theta&=(\theta^A,\theta^B)\in\RR^k\times\RR^{n-k}.\label{last.project.fv}
\end{align}
The maps $\pi_A^\Xi$ and $\pi_B^\Theta$, when equipped with particular values at their codomain (denoted here, respectively, by $\bar{\eta}_A\in\Xi_A\subseteq\RR^k$ and $\bar{\theta}^B\in\Theta_B\subseteq\RR^{n-k}$), provide an example of the constraints \eqref{constraints.JF}:
\begin{equation}
\left\{\begin{array}{l}
\eta_A(p(t))=\pi^\Xi_A(\eta(p(t)))=\bar{\eta}_A,\\
\theta^B(p(t))=\pi^\Theta_B(\theta(p(t)))=\bar{\theta}^B.
\end{array}\right.
\end{equation}
The subspaces $\Xi_A$ and $\Theta_B$ denote, respectively, the range of the values $\bar{\eta}_A$ and $\bar{\theta}^B$ of the constraints $\pi_A^\Xi$ and $\pi_B^\Theta$. The fibres corresponding to these projections are given by
\begin{align}
\M_\Xi(\bar{\eta}_A)&:=(\pi_A^\Xi)^{-1}(\bar{\eta}_A)=\{\eta\in\Xi\mid\eta_A=\bar{\eta}_A\}\subseteq\Xi,\\
\M_\Theta(\bar{\theta}^B)&:=(\pi^\theta_B)^{-1}(\bar{\theta}^B)=\{\theta\in\Theta\mid\theta^B=\bar{\theta}^B\}\subseteq\Theta,
\end{align}
and they induce the corresponding leaves of a pair of foliations of $\M(\boole)$ by
\begin{align}
\M(\bar{\eta}_A)&:=\{p\in\M(\boole)\mid(\pi_A^\Xi\circ \eta)(p)=\bar{\eta}_A\},\\
\M(\bar{\theta}^B)&:=\{p\in\M(\boole)\mid(\pi_B^\Theta\circ \theta)(p)=\bar{\theta}^B\},
\end{align}
with
\begin{equation}
\bigcup_{\bar{\theta}^B\in\Theta_B}\M(\bar{\theta}^B)=\M(\boole)=\bigcup_{\bar{\eta}_A\in\Xi_A}\M(\bar{\eta}_A).
\end{equation}
Using the orthogonality \eqref{dual.metric.flatness} of the coordinate systems $\theta^j$ and $\eta_i$, Favretti shows that the tangent space at the point $p\in\M(\bar{\theta}^B)\cap\M(\bar{\eta}_A)$ has the following orthogonal decomposition
\begin{align}
\T_p\M(\boole)&=\T_p\M(\bar{\theta}^B)\oplus\T_p\M(\bar{\eta}_A)=\Span\{\partial_1,\ldots,\partial_k\}\oplus \Span\{\partial^{k+1},\ldots,\partial^n\}.
\end{align}
For any $a\in\Xi_A$ and $b\in\Theta_B$, the leaves $\M(a)$ and $\M(b)$ are, respectively, $\nabla^\eta$- and $\nabla^\theta$- autoparallel submanifolds of $\M(\boole)$, hence they are called \df{mutually dual foliations}. Favretti observes that this allows to consider the evolution $t\mapsto p(t)$ geometrically, as a horizontal lift with respect to an integrable Ehresmann connection.
Let us now assume that $\M(\boole)$ is an $(n+m)$-dimensional dually flat probability manifold equipped with the projections \eqref{first.project.fv}-\eqref{last.project.fv}, as well as with an additional projection generated by
\begin{equation}
\pi^{\Xi}_C:\Xi\ni\eta\mapsto\eta_C\in \Xi_C\subseteq\RR^m,
\end{equation}
where $\eta=(\eta_A,\eta_B,\eta_C)\in\RR^k\times\RR^{n-k}\times\RR^m$. One can introduce the foliation $\M(\bar{\eta}_C)$, corresponding to the constraint $\eta_A-\bar{\eta}_A=0$, in the same way as before. In such case the pairs of mutually dual foliations are given by $\M(\bar{\eta}_A,\bar{\eta}_C)$, $\M(\bar{\theta}^B)$ and $\M(\bar{\eta}_C)$, $\M(\bar{\theta}^A,\bar{\theta}^B)$. However, $\M(\bar{\eta}_C)$, $\M(\bar{\theta}^B)$ are not mutually dual.
This setting allows for the geometric generalisation of the source renormalisation procedure in the following form. Let temporal evolution $t\mapsto p(t)$ satisfy the constraints $\theta^B(t)=\bar{\theta}^B$ and $\eta_C(t)=\bar{\eta}_C$, that is,
\begin{equation}
\left\{\begin{array}{l}
\theta^B(p(t))=(\pi_B^\Theta\circ \theta)(p(t))-\bar{\theta}^B=0,\\
\eta_C(p(t))=(\pi_C^\Xi\circ \eta)(p(t))-\bar{\eta}_C=0,
\end{array}\right.
\label{constr.F}
\end{equation}
then
\begin{equation}
p(t)\in\M(\bar{\theta}^B)\cap\M(\bar{\eta}_C).
\end{equation}
These conditions rephrase the Jaynes--Mitchell conditions $\delta\lambda_B=0$ and $\delta\e{f_C}=0$ in the information geometric terms. Now one can find what is the form of evolution determined by these constraints, if changes of information are specified by the temporally driven `response' parameters $\eta_A=\eta_A(t)$ (which corresponds to the `driving variable' $\e{f_A}$). The constraints \eqref{constr.F} on the evolution can be restated using \eqref{LFT.MJF.Phi} in the form
\begin{equation}
\partial^B\Psi^\lfdual(\eta_A(t),\eta_B(t),\bar{\eta}_C)-\bar{\theta}^B=0.
\label{linear.evolution.favretti}
\end{equation}
Favretti has shown that the implicit function theorem applied to \eqref{linear.evolution.favretti} implies the existence of a smooth map
\begin{equation}
h:\Xi_A\times\Xi_C\ni(\eta_A,\eta_C)\mapsto\eta_B\in\RR^{n-k}
\end{equation}
such that
\begin{align}
\eta_B(t)&=h(\eta_A(t),\bar{\eta}_C),\\
\dot{\eta}_B(t)&=\partial^A h(\eta_A(t),\bar{\eta}_C)\dot{\eta}_A(t),\label{eta.B.dot}\\
h(\eta_A(t),\bar{\eta}_C)&=
\left.
-\left(
(\partial^B\partial^B\Psi^\lfdual)^{-1}
\partial^A\partial^B\Psi^\lfdual
\right)
\right|
_{\eta_B=h(\eta_A(t),\bar{\eta}_C)}.
\end{align}
By the assumption of dual flatness, this gives also
\begin{equation}
\dot{\eta}_B(t)=
(\tilde{\Psi}_{,BA}(\tilde{\Psi}_{,AA})^{-1})
|_{\theta=\bar{\theta}}\dot{\eta}_A(t)
=:\tilde{\mathcal{G}}_{BA}(\theta)\,\dot{\eta}_A(t),
\label{inferred.favr.evol}
\end{equation}
where
\begin{align}
(\Psi^\lfdual)^{,ij}&:=
\partial^i\partial^j\Psi^\lfdual\equiv
\frac{
\partial^2\Psi^\lfdual(\eta)
}{
\partial\eta_i\partial\eta_j
}=
\gbold^{ij}(\eta),\label{g.phi.hessian}\\
\Psi_{,ij}&:=
\partial_i\partial_j\Psi\equiv
\frac{
\partial^2\Psi(\theta)
}{
\partial\theta^i\partial\theta^j
}=
\gbold_{ij}(\theta),\label{g.psi.hessian}
\end{align}
and
\begin{align}
\bar{\theta}&:=
\LFtrafo^{-1}_\Psi(
\eta_A,
h(
\eta_A(t),
\bar{\eta}_C
),
\bar{\eta}_C
),
\\
\tilde{\Psi}_{,BA}&:=
\Psi_{,BA}-
\Psi_{,BC}
(\Psi_{,CC})^{-1}
\Psi_{,CA},
\\
\tilde{\Psi}_{,AA}&:=
\Psi_{,AA}
(\II-R^2_{AC}),\label{Psi.prim.AA}
\\
R_{AC}^2&:=
(\Psi_{,AA})^{-1}
\Psi_{,AC}
(\Psi_{,CC})^{-1}
\Psi_{,CA}.
\end{align}
The equation \eqref{inferred.favr.evol} can be written more explicitly as
\begin{equation}
\dd\eta_B(p(t))=\gbold_{BA}(p(t))\frac{1}{\II-R^2_{AC}(p(t))}\left(\gbold_{AA}(p(t))\right)^{-1}\dd\eta_A(p(t)).
\label{JMF.renormalisation.of.metric}
\end{equation}
From \eqref{Psi.prim.AA} it follows that
\begin{equation}
(\tilde{\Psi}_{,AA})^{-1}=(\II-R_{AC}^2)^{-1}(\Psi_{,AA})^{-1},
\end{equation}
hence, if $R_{AC}$ has a spectral radius smaller than $1$, one can use \eqref{Jaynes.expansion}, which leads to the perturbative expansion in terms of corrections that come from interaction with the additional source,
\begin{align}
\tilde{\mathcal{G}}_{BA}&=\tilde{\Psi}_{,BA}(\II-R_{AC}^2)^{-1}(\Psi_{,AA})^{-1}=\mathcal{G}_{BA}-\mathcal{G}_{BC}\mathcal{G}_{CA}+\mathcal{G}_{BA}\mathcal{G}_{AC}\mathcal{G}_{CA}+\ldots,
\end{align}
where
\begin{equation}
\mathcal{G}_{ij}:=\Psi_{,ij}(\Psi_{,jj})^{-1}
\end{equation}
For $\dd\eta_B=\dot{\eta}_B(t)\dd t$, $\dd\eta_A=\dot{\eta}_A(t)\dd t$, the above expression takes the form
\begin{equation}
\dd\eta_B=(\mathcal{G}_{BA}-\mathcal{G}_{BC}\mathcal{G}_{CA}+\mathcal{G}_{BA}\mathcal{G}_{AC}\mathcal{G}_{CA}-\ldots)\dd\eta_A,
\label{diff.renorm.eq}
\end{equation}
which is a generalisation of \eqref{expanded.change}. Hence, the additional constraint $\dd\eta_C=0$ acts as a source of information, which imposes nontrivial corrections in the relationship between the evolution of $\dd\eta_A$ and $\dd\eta_B$, that are perturbatively described by equation \eqref{diff.renorm.eq}. Note that the implicit function theorem does not provide an explicit form of the function $h$. Thus, one may need to integrate the equation \eqref{eta.B.dot}. The equation \eqref{diff.renorm.eq} provides a perturbative approximation of \eqref{eta.B.dot}, which can be subjected to integration. This might be called a `perturbative renormalisation' or `inferential scattering' of $\mathcal{G}_{BA}$.
\subsection{Br\`{e}gman distance and nonlinear quantum control\label{section.dynamics.renormalisation}}
An important feature of the Jaynes--Mitchell theory is that it allows to consider not only the `source' (`input', `configuration') and `response' (`output', `registration') variables, but also the `control' (`covariate') variables, defined as fixed parameters of the model. The constraints imposed by these fixed variables can be factored out from the relationship between causes and effects, but at the price of `renormalisation' of the source terms. It amounts to reduction of the dimensionality of the model (removing the dimensions described by control parameters) and subsequent rescaling of the remaining source terms by the `renormalisation factors'. Thus, one can eliminate control parameter from the \textit{model construction} at the price of renormalisation of the source terms that determine the \textit{information dynamics}. This procedure is \textit{nonperturbative} and \textit{geometric}, but under certain conditions it can be expanded in the perturbative series of corrections. Quite remarkably, the renormalisation factor that appears at the \textit{first order} of expansion in powers of source strength can be perturbatively expanded in an infinite series of corrections, which contain \textit{all orders} of interaction effects with the `virtual' source terms that can be associated with the factored-out `control' variables.\footnote{This phenomenon was first observed in the Heims--Jaynes analysis of the gyromagnetic effect \cite{Heims:Jaynes:1962}, and appeared later also in Jaynes' analysis of the Rayleigh acoustic scattering \cite{Jaynes:1985:scattering,Jaynes:1993} and (independently) in Schwinger's source theory \cite{Schwinger:1969,Schwinger:1970}.}
Besides generalisation from exponential families to dually flat manifolds, the information geometric framework introduces important conceptual change: the `source', `response', and `control' variables are no longer associated with particular functions on the sample space, but rather with the particular coordinate variables on the information model. As discussed in Section \ref{qig.foundations.intro.section}, all these variables form specific examples of \textit{observables} in our approach to quantum foundations. As a consequence, renormalisation of sources can be considered \textit{exactly} as a transformation of information models that amounts to `coarse graining' and subsequent rescaling. The coarse graining provides a a reduction of dimension of the model that is preserving the operational definitions of the coordinate variables on a submodel, but at the price of redefinition of their functional relationship by means of change of the local geometry of the model from dually flat to curved one.
In principle, the change of an information state $\phi\in\M(\boole)$ or $\phi\in\M(\N)$, associated with an integrable real function $f$ over $\boole$ or an operator $f\in\N^\sa$, respectively, can be specified in three different ways: by means of $\delta(\phi(f))$, by means of $\phi(\delta f)$, or by means of a \df{source term} $(\delta\phi)(f)=:\delta Q_f$. The main insight of the source theory is that the changes specified by source terms have the direct operational meaning whenever the model $\M(\boole)$ is equipped with a pair of dually flat coordinate systems. In such case, the changes $(\delta\phi)(\cdot)$ can be reexpressed in terms of corresponding changes of source-and-response parameters. However, the change of information driven by the change of one of source-and-response variables leads to change of other variables that are correlated with it.
This can be interpreted as a cause-and-effect relationship, but under the condition that `causes' and `effects' are understood as inputs and outputs of correlation relationships, respectively. This is \textit{different} from the meaning assigned to these terms in other sections of this work. In general, there are possible at least two clearly distinct perspectives on what the `causes' and `effects' are. From the purely operational perspective, any reproducible relationship between configuration and response parameters of description of experimental situation deserves to be called a causal relationship, and any predictively verifiable inferential procedure relating them is considered as a satisfying method of the theoretical modelling of causality (see e.g. \cite{Hinkelmann:Kempthorne:2005}). On the other hand, from the ontologically flavoured perspective, the `causes', `effects', and their relationships are theoretical notions, which may indirectly correspond to epistemic parameters and predictively verifiable relationships between them, derived from some inductive procedure (see e.g. \cite{Pearl:2000}). In this Section we chose the former terminology (speaking of `epistemic causality', because the term ``inferential causality'' would probably cause, nomen omen, more confusion), while in the rest of this work we consider causality and inference as a priori independent theoretical constructs, but without attribution of any ontological claims. From the inferential perspective, `causes' are just the same as `configurations'.
If one of the parameter spaces $\{\Theta,\Xi\}$ can be considered as a space of `causes' (`configurations'), the other becoming a space of `effects' (`registrations'). The role of the Legendre transform $\LFtrafo_\Psi:\Theta\ra\Xi$ is to associate effects with causes (and vice versa).\footnote{Note that this notion of `causality' belongs strictly to a theoretical layer of scientific inquiry. Without specification of some particular epistemic semantics it is not related in any specific way with the experimental `effects' and `causes'.} This allows to use $\LFtrafo_\Psi$ (and ``epistemic'' `cause-and-effect' interpretation) in order to analyse changes of effects following (correlatively, inferentially) from the changes of causes, as well as changes of causes following from the changes of effects. These two issues are known, respectively, as \textit{forward} and \textit{backward induction problems}. Let us also note that the dually flat geometry always satisfies the relationships \eqref{LFT.MJF.Phi} and \eqref{LFT.MJF.Psi} \textit{as well as} \eqref{g.phi.hessian} and \eqref{g.psi.hessian}. Hence, it also satisfies
\begin{equation}
\gbold_{ij}^\Psi(\theta)=\frac{\partial\eta_j}{\partial\theta^i}.
\label{local.causality.metric}
\end{equation}
The equations \eqref{LFT.MJF.Psi} and \eqref{local.causality.metric} assert that
\begin{itemize}
\item a Legendre transform $\LFtrafo_\Psi$ governs the relationship between causes and effects,
\item a riemannian metric $\gbold^\Psi$ governs the relationships between \textit{changes} of causes and \textit{changes} of effects.
\end{itemize}
From the perspective discussed in the Section \ref{local.infogeometry.histories.intro.section}, $\LFtrafo_\Psi$ defines the `system of epistemic causality' of an individual user (the perspective of a fixed measurement frame), relating the `configurations' and `registrations' in such way that the discrimination function on the space of configurations defines the discrimination function on the space of registration. On the other hand, $\gbold^\Psi$ defines the `local system of epistemic causality' \textit{on} the information manifold, allowing to translate between different local users. Hence, the source theoretic renormalisation can be interpreted as a perturbation of the local system of \textit{epistemic} causality due to presence of the nonzero sources of information.
Now let us observe that, as discussed in Section \ref{distances.NS.geom.section}, every dually flat manifold determines an associated Br\`{e}gman distance $D_\Psi$. While our notation in Section \ref{Favretti.section} indicates this fact, it was left unnoticed by previous authors. Also, we note, following the discussion in Section \ref{Orlicz.Br\`{e}gman.section}, that the framework of dually flat manifolds and associated Br\`{e}gman distances is applicable \textit{locally} to \textit{any} quantum manifolds, as long as one defines the manifold structure using a specific Br\`{e}gman function. The extension from commutative to quantum dually flat geometries is straightforward. The notation applied by us in Section \ref{Favretti.section} keeps the correct order of multiplications, so the quantities used in this Section can be interpreted as operators as well.
The equation \eqref{local.causality.metric} is a geometric equivalent of the linear case \eqref{first.order.covariance.expansion} of the expansion \eqref{infinite.perturbation.covariance.expansion}. In order to obtain higher-order terms of \eqref{infinite.perturbation.covariance.expansion}, one needs to consider information models that are not dually flat. Hence, one can in principle \textit{begin} with an \textit{arbitrary} quantum information model $\M(\N)$, equipped with a Br\`{e}gman distance $D_\Psi$ which defines the local `ideal' dually flat manifold structure $(\M(\N),\gbold^{D_\Psi},\nabla^{D_\Psi},(\nabla^{D_\Psi})^\nsdual)$, and consider the emergence of the nonzero curvature of the effective riemannian geometry $(\M(\N),\tilde{\gbold})$ as a result of presence of the additional source (control) terms that are `renormalised out' by the transition $\gbold\mapsto\tilde{\gbold}$. It is quite interesting that the departure from the dually flat geometry and constant curvature of a model implies the presence of additional information sources operating at different points. From the perspective of analogy with general relativity, we can say that \textit{sources curve the geometry of a quantum information manifold}. Combining this with our discussion of the role of quantum riemannian metric in the Daubechies--Klauder formula (see Sections \eqref{local.infogeometry.histories.intro.section} and \eqref{algebraic.quantum.histories}), we can conclude that the JMF renormalisation leads to the redefinition of the \textit{local} prior measure used for the path integration. Interpreting the local prior measure as an information theoretic analogue of mass, we can say that this process encodes dependence on additional sources by the change of the geometry of a model, which is in turn reflected in the renormalisation of an information theoretic local mass, and a corresponding point-dependence of the zero-point energy.
We will use the term \df{br\`{e}gmanian renormalisation} to refer to a local renormalisation of $(\M,D_\Psi)$ using JMF source theory. More generally, let us observe (following Lauritzen \cite{Lauritzen:1987:statistical:manifolds}), that the Norden--Sen geometry captures the description of information geometry only up to third order of Taylor series expansion of information distance, which in principle allows to develop higher--order differential tensor theories of information geometry, more general than the Norden--Sen geometry. Thus, it is plausible that the higher order source renormalisation terms may also possess a complete geometric representation, but requiring to use higher order tensor geometries arising from the Taylor expansion of the Br\`{e}gman distance as the referential object subjected to renormalisation.
The `source term' defined as above has different meaning than the `source term' introduced in Section \ref{local.source.liouv.section}. Yet they are complementary. The former corresponds to a perturbation $\delta\theta$ of a coordinate system $\theta:\phi\mapsto\theta(\phi)=\phi(x)$ under constraint $\phi(\delta x)=0$. This description rests on the assumption that all relevant local information which has to be taken under consideration is completely specified by means of the variations $\delta(\phi(x))$ and $(\delta\phi)(x)$. The latter corresponds to perturbation $\delta x$ of an element $x$ of (a local GNS representation of) a $W^*$-algebra $\N$ by means of state dependent perturbation of liouvillean. Our approach allows $\delta x$ to be arbitrary, so it can also depend on $\phi$, and may not arise as an infinitesimal change generated by a global automorphism of $\N$. These two different uses of a single notion are compatible and complementary in the sense provided by the equations \eqref{work.def}-\eqref{gen.1st.law.1}: while the `sources' of Sections \ref{section.JM.source.theory} and \ref{Favretti.section} generalise the notion of `heat sources', the `sources' of Section \ref{local.source.liouv.section} generalise the notion of `work sources'. In our work we view `work sources' (respectively, `heat sources') as the geometric perturbation of the geometry of causal evolution (respectively, inferential evolution).
\subsection{Contraction coefficients}
The Br\`{e}gmannian renormalisation answers the question about the behaviour of the constraints of inference (and resulting information dynamics) under \textit{dimensional} reduction of information model due to the presence of constant control parameters. However, given any information model and constraints of inference, there appears also another renormalisation-type question: what is the behaviour of these objects under coarse grainings? Because the constraints may involve information geometric quantities (for example, the `two point correlation function' $K_{xy}(\rho)=\int_0^1\dd\lambda\tr_\H(\rho^\lambda x\rho^{1-\lambda}y)$ is an evaluation of the quantum Bogolyubov--Kubo--Mori riemannian metric $\gbold^{D_1}_\rho$ on the pair of tangent vectors $x,y\in\T_\rho\M(\N)$), this is related to the question about behaviour of information geometric quantities under completely positive maps. Restriction to quantum $D_\fff$-geometries, where $\fff$ is an operator convex function defining the $D_\fff$ distance, secures the Markov monotononicity of $\gbold^{D_\fff}$ and $\nabla^{D_\fff}$, but this does not extend naturally to every geometric quantity on $\M(\N)$ that can be built using these objects and their derivatives.
In the commutative case Chencov \cite{Chencov:1969,Chencov:1972} has defined the Markov monotone connections as such affine connections $\nabla^{D_\fff}$ that for any Markov map $T$ the image of a $\nabla^{D_\fff}$-geodesic line on $\M(\boole)$ belongs to a $\nabla^{D_\fff}$-geodesic line on $T_\star(\M(\boole))$ as its interval or its point, while an affine parameter of this line remains, up to rescaling, an affine parameter of the $\nabla^{D_\fff}$-geodesic line in the image \cite{Chencov:1972}. Thus, the behaviour of any trajectory along a given $\nabla^{D_\fff}$-geodesic under the action of Markov maps is characterised by their invariance properties under \textit{coarse graining} by preduals of Markov maps and \textit{rescaling} by an affine parameter. This leads to a question whether it is possible to find a suitable analogue of an affine parameter for arbitrary quantum information model $\M(\N)$ which would allow for some sort of control over the mutual behaviour of $\M(\N)$, its information geometry, and information dynamics under coarse grainings. More specifically, we need to find some scalar \df{contraction coefficient} $\rpktarget{CONTR}\contrcoeff$, which globally characterises the geometry of $\M(\N)$ \textit{and} is Markov monotone, $\contrcoeff(\M(\N))\geq\contrcoeff(T_\star(\M(\N))$, and then use it in order to rescale the constraints of inference on $\M(\N)$.
Some examples of contraction coefficients $\contrcoeff(T_\star)$ were provided in the case when $\dim\M(\N)<\infty$, with semi-finite $\N$ by Lesniewski and Ruskai \cite{Lesniewski:Ruskai:1999}, following earlier works \cite{CIRRSZ:1993,Choi:Ruskai:Seneta:1993,Ruskai:1994}:
\begin{align}
\contrcoeff_{D_\fff}(T_\star)&:=
\sup_{
\omega,\phi\in\M(\N),
}
\left\{
\frac{
D_\fff(
T_\star(\omega),
T_\star(\phi)
)
}{
D_\fff(
\omega,
\phi
)
}
\mid
\omega\neq\phi
\right\}
,\\
\contrcoeff_{\gbold^{D_\fff}}(T_\star)&:=
\sup_{
\phi\in\M(\N)
}
\left\{
\sup_{
u\in\T_\phi\M(\N)
}
\left\{
\frac{
\gbold^{D_\fff}_{
T_\star(\phi)
}(
T_\star(u),T_\star(u)
)
}{
\gbold^{D_\fff}_{\phi}(u,u)
}
\right\}
\right\},
\\
\contrcoeff_{d_{\gbold^{D_\fff}}}(T_\star)&:=
\sup_{\omega,\phi\in\M(\N)}
\left\{
\frac{(d_{\gbold^{D_\fff}}(T_\star(\omega),T_\star(\phi)))^2}{(d_{\gbold^{D_\fff}}(\omega,\phi))^2}
\mid
\omega\neq\phi
\right\}
,
\end{align}
where
\begin{equation}
d_{\gbold^{D_\fff}}(\omega,\phi):=\inf_{c\in C}\left\{\int_0^1\dd t\sqrt{\gbold^{D_\fff}_{c(t)}(\dot{c}(t),\dot{c}(t))}\right\},
\end{equation}
and $C$ is defined as a class of all smooth curves $c:[0,1]\ni t\mapsto c(t)\in\M(\N)$ such that $c(0)=\omega$ and $c(1)=\phi$. Apart from Markov monotonicity of the above coefficients, Lesniewski and Ruskai proved that these coefficients are convex in $T_\star$, and satisfy
\begin{equation}
1\geq\contrcoeff_{D_\fff}(T_\star)\geq\contrcoeff_{\gbold^{D_\fff}}(T_\star)\geq\contrcoeff_{d_{\gbold^{D_\fff}}}(T_\star).
\end{equation}
Now, let the inferential quantum dynamics be given by $D_\fff$ entropic projection on $\M(\N)$, with the constraints $\Q\subseteq\M(\N)$ specified in terms of lower semi-continuous convex function $F:\M(\N)\ra\,]-\infty,+\infty]$,
\begin{equation}
\M(\N)\ni\omega\mapsto\arginf_{\phi\in\M(\N)}\left\{D_\fff(\omega,\phi)+F(\phi)\right\}\in\M(\N).
\label{renormalise.this.dynamics}
\end{equation}
With the nontrivial examples of contraction coefficients at hand, we can propose to control the behaviour of constraints $\Q$ under coarse grainings $T_\star$ by means of \df{markovian renormalisation semi-group} transformation
\begin{equation}
F(\phi)\mapsto\frac{1}{\contrcoeff(T_\star)}F(T_\star(\phi)),
\label{renorm.eq}
\end{equation}
which amounts to subsequent coarse graining and rescaling of constraints. The choice of a particular form \eqref{renorm.eq} of transformation of constraints can be justified either by appealing to arguments and insights based on ordinary renormalisation semi-group theory or by recalling another result of Lesniewski and Ruskai:
\begin{equation}
\contrcoeff_{D_\fff}(T_\star)\neq\contrcoeff_{\gbold^{D_\fff}}(T_\star)\iff\exists \omega\neq\phi\;\;\mbox{such that}\;\; \frac{1}{\contrcoeff_{D_\fff}(T_\star)}D_\fff(T_\star(\omega),T_\star(\phi))=D(\omega,\phi).
\label{renorm.Dfff}
\end{equation}
In such case, the invariant $\contrcoeff_{D_\fff}(T_\star)$ contains a complete information about the behaviour of distance $D_\fff(\omega,\phi)$ under rescaling by coarse grainings $T_\star$. In view of \eqref{renorm.Dfff}, the aim of rescaling \eqref{renorm.eq} is to obtain \textit{the same form of information dynamics} (for a given initial state) independently of the coarse graining. In consequence, we will say that the quantum dynamics \eqref{renormalise.this.dynamics} is in a \df{fixed point} of markovian renormalisation semi-group transformation with respect to a contraction coefficient $\contrcoeff$ if{}f, given an initial state $\omega\in\M(\N)$, $\omega\neq\phi$, the equations
\begin{equation}
\left\{
\begin{array}{l}
\contrcoeff(T_\star)F(\phi)=F(T_\star(\phi)),\\
\contrcoeff(T_\star)D(\omega,\phi)=D_\fff(T_\star(\omega),T_\star(\phi))
\end{array}\right.
\label{renorm.sg.fixed.point}
\end{equation}
hold for any $T_\star$ on $\M(\N)$. If \eqref{renorm.sg.fixed.point} is not satisfied, then the action of \eqref{renorm.eq} generates the `flow' of forms of dynamics along the `trajectory' of semi-group of markovian morphisms.
As these examples show, quantum information geometry provides quantitative tools allowing to develop various renormalisation procedures for quantum inference, which possess explicit conceptual and quantitative meaning. In particular, the br\`{e}gmannian and markovian renormalisation procedures reflect, respectively, two different problems: coarse graining and rescaling of the \textit{solution} of dynamical (inferential) problem and coarse graining and rescaling of the \textit{definition} of dynamical (inferential) problem.\footnote{Consider $y=f(x)$, where $f$ is an arbitrary function. It is clear that the procedure used to control the quality of approximation of the initial data $x$ does not need to correspond to the procedure used to control the quality of approximation on the space of solutions of this equation.} We refer to \cite{Beny:Osborne:2012,Beny:Osborne:2013,Beny:Osborne:2014} for another (and more developed) approach to quantum information geometric renormalisation based on the use of markovian morphisms $T_\star$ (see also \cite{DeBrota:2015} for a pedagogical overview).
\pagebreak
\section*{Acknowledgments}
\begin{flushright}
{\begin{spacing}{0.7}
\foreignlanguage{russian}{\scriptsize \textit{Пока жива,\\
Я могу стараться на лету не упасть,\\
Не разучиться мечтать... любить...}\\
О.В. Яковлева}
\end{spacing}}
\end{flushright}
{\small This research was supported in part by Perimeter Institute for Theoretical Physics. Research at Perimeter Institute is supported by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Research and Innovation. This research was also partially financed by the National Science Center of the Republic of Poland (Narodowe Centrum Nauki) through the grant number N N202 343640. Early versions of this work were developed as a part of research supported by the grants: \textit{Mistrz} 2007 grant of Foundation for Polish Science for Jerzy Lewandowski and his students, Quantum Geometry and Quantum Gravity grants 1955 and 2706 of European Science Foundation, and 182/N QGG/2008/0 grant of Ministerstwo Szkolnictwa Wyższego i Nauki. I acknowledge substantial reliance on the knowledge resources provided for free by Library Genesis (\href{http://gen.lib.rus.ec}{http://gen.lib.rus.ec}) and Sci-Hub (\href{http://sci-hub.cc}{http://sci-hub.cc}). I believe that providing unrestricted free access to the results of scientific research is a matter of an elementary respect to other human beings. This paper grew out of the heuristic considerations in the preprints \cite{Kostecki:2007:qht,Kostecki:2008:aqh,Kostecki:2010:aqh}, motivated in turn by the ideas of Stanis{\l}aw Lem \cite{Lem:1974,Lem:1981} (information can be exchanged into energy and create a black hole), Jean Baudrillard \cite{Baudrillard:1980,Baudrillard:1981} (symbolic systems of information gravitate), and Antoni K\k{e}pi\'{n}ski \cite{Kepinski:1972:Rytm,Kepinski:1972:Schizofrenia} (time of entropic processing and time of energy flow are \textit{a priori} independent). I am indebted to Stanis{\l}aw Woronowicz for introducing me to the theory of noncommutative flow of weights, which allowed me to make some of these ideas mathematically sound. I want to warmly thank: C\'{e}dric B\'{e}ny, Pavel B\'{o}na, Dylan Butson, John DeBrota, Jan Derezi\'{n}ski, Paolo Gibilisco, Frank Hellmann, Carlos Guedes, Anna Jen\v{c}ov\'{a}, Wojtek Kami\'{n}ski, Jerzy Kijowski, Jacek Kope\'{c}, Władysław A. Majewski, W{\l}odek Natorf, Dmitri\u{\i} Pavlov, Claude-Alain Pillet, Carlos C. Rodr\'{\i}guez, David Sherman, Ray Streater, and Karol \.Zyczkowski for discussions, Marco Favretti for sharing his unpublished preprint \cite{Favretti:2007}, Jerzy Lewandowski for his trust in, and support of, my nonorthodox research, Chris Isham for his interest in, and comments on, the early version of this paper, Fotini Markopoulou, Olaf Dreyer, Jamie Vicary, Cecilia Flori, and Tomasz Kołodziejski for their generous hospitality in London, Nottingham, Oxford, and Berlin during the periods of intense work on this project, Basia Baranowska, Janka Spicha-Konarzewska, and Aneta P\k{e}ka{\l}a for sharing their life, love and support, and my flatmates Teresa Szczepi\'{n}ska, Rafa{\l} Gr\"{o}ger, and Agnieszka Gronowicz, as well as medical teams of \textit{Szpital Kliniczny Dzieci{\k{a}}tka Jezus} and \textit{Centralny Szpital Kliniczny MSW}, for rescuing me from the carbon monoxide brain death.}
\section*{References}
\addcontentsline{toc}{section}{References}
{%
\scriptsize
All Cyrillic titles and names were transliterated from the original papers and books. For the Latin transliteration of the Cyrillic script (in references and surnames) we use the following modification of the system GOST 7.79-2000B: {\foreignlanguage{russian}{ц}} = c, \foreignlanguage{russian}{{ч}} = ch, \foreignlanguage{russian}{х} = kh, \foreignlanguage{russian}{ж} = zh, \foreignlanguage{russian}{ш} = sh, \foreignlanguage{russian}{щ} = shh, {\foreignlanguage{russian}{ю}} = yu, {\foreignlanguage{russian}{я}} = ya, {\fontencoding{OT2}\selectfont\textcyrup{\"{e}}} = \"{e}, {\foreignlanguage{russian}{ъ}} = `, {\foreignlanguage{russian}{ь}} = ', {\foreignlanguage{russian}{э}} = \`{e}, {\foreignlanguage{russian}{й}} = \u{\i}, with an exception that names beginning with {\foreignlanguage{russian}{Х}} are transliterated to H. For Russian texts: {\fontencoding{OT2}\selectfont\textcyrup{y}} = y, {\fontencoding{OT2}\selectfont\textcyrup{i}} = i; for Ukrainian: {\fontencoding{OT2}\selectfont\textcyrup{i}} = y, i = i, \"{\i} = \"{\i}. Note: All links provided in references link to the free access files. Files and digital copies that are subject to any sort of restricted access were not linked. See \href{http://michaelnielsen.org/polymath1/index.php?title=Journal_publishing_reform}{michaelnielsen.org/polymath1/index.php?title=Journal\_publishing\_reform} for the reasons why.
\begingroup
\raggedright
\bibliographystyle{rpkbib}
\renewcommand\refname{\vskip -1cm}
|
\section{Introduction} \label{sec:intro}
In the last decade, noticeable advances were made in the characterization of atmospheric properties and chemical compositions of cool, sun-like stars.
The achieved high precision in chemical abundances allowed observers to study the chemical abundances of stars with and without planets in detail
\citep[e.g.,][]{Gilli-06, Robinson-06, Delgado-14, Delgado-15, Kang-11, Adibekyan-12a, Adibekyan-12b, Adibekyan-15a}. These studies are very important
since the chemical abundances of stars with planets provide a huge wealth of information about the planet formation process
\citep[e.g.,][]{Bond-10, Delgado-10}, composition of planets \citep[e.g.,][]{Thiabaud-14, Dorn-15, Santos-15}, and even about the habitability of the
planets \citep[e.g.,][]{Adibekyan-16}.
The achieved extremely high precision in chemical abundances also facilitated the study of very subtle chemical peculiarities in stars that had initially appeared to be Sun-like.
The quintessential example, the so-called $T_{\mathrm{c}}$ \ trend, is the trend of individual chemical abundances with the condensation temperature ($T_{\mathrm{c}}$) of the elements.
Many studies, beginning with \citet{Gonzalez-97} and \citet{Smith-01}, attempted to search for planet accretion and/or formation signatures in the photospheric
chemical compositions of cool stars by looking at the $T_{\mathrm{c}}$\ trends \citep[e.g.,][]{Takeda-01, Ecuvillon-06, Sozzetti-06, Melendez-09, Ramirez-09, GH-10,
Schuler-11, GH-13, Maldonado-15, Nissen-15, Biazzo-15, Saffe-15}.
\citet{Melendez-09} claimed that the Sun shows a deficiency in refractory elements with respect to other solar twins because
these elements were trapped in the terrestrial planets in our solar system. The same conclusion was also reached by \citet{Ramirez-09}, who analyzed 64 solar twins and
analogs with and without detected planets. However, later results by \citet{GH-10} and \citet{GH-13} strongly challenge the relation between the presence of
planets and the abundance peculiarities of the stars.
Rocky material accretion is by far not the only explanation for the $T_{\mathrm{c}}$ \ trend. Recently, \citet{Adibekyan-14} showed that the $T_{\mathrm{c}}$\ trend strongly
depends on the stellar age\footnote{We note that \citet{Ramirez-14} also observed a correlation between the $T_{\mathrm{c}}$ \ slope and stellar age
for metal-rich solar analogs, but apparently with opposite sign. In contrast to our results, they found that
most refractory element depleted stars are younger than the those with the highest refractory element abundances.} %
and they found a tentative dependence on the galactocentric distances of the stars. The authors concluded that the $T_{\mathrm{c}}$\ trend depends on the Galactic
chemical evolution (GCE) and suggested that the difference in $T_{\mathrm{c}}$ \ slopes, observed between planet hosting solar analogs and solar analogs without detected
planets, may reflect their age difference. \citet{Maldonado-15} and \citet{Maldonado-16} also confirmed the correlation of the $T_{\mathrm{c}}$\ trend with age, and further suggested a
significant correlation with the stellar radius and mass. \citet{Onehag-14} in turn showed that while the Sun shows a different $T_{\mathrm{c}}$\ trend when compared to the
solar-field twins, it shows a very similar abundance trend with $T_{\mathrm{c}}$ \ when compared to the stars from the open cluster M67. These authors suggested that the Sun,
unlike most stars, was formed in a dense stellar environment where the protostellar disk was already depleted in refractory elements before the star formation.
\citet{Gaidos-15} also suggested that gas-dust segregation in the disk can be responsible for the $T_{\mathrm{c}}$\ trends.
While above mentioned mechanisms and processes (e.g., GCE or age effect) can be responsible for the general $T_{\mathrm{c}}$\ trends for field-star samples,
they cannot easily explain trends observed between companions of binary systems. Several authors studied the $T_{\mathrm{c}}$\ trend in binary stars with and without planetary
companions \citep[e.g.,][]{Liu-14, Saffe-15, Mack-16} or binary stars where both components host planets \citep[e.g.,][]{Biazzo-15, Teske-15, Ramirez-15, Teske-16}.
The results and conclusions of these studies show that there are no systematic differences in the chemical abundances of
stars with and without planets in binary systems. Moreover, there is no consensus on the results even for the same, individual systems \citep[e.g., 16 Cyg AB;][]{Laws-01, Takeda-05,
Schuler-11a, TucciMaia-14}.
Recently, \citet[][]{Maldonado-15} studied a large sample of stars with and without debris disks to search for chemical anomalies related to the formation of planets.
They found no significant differences in chemical abundances or in the $T_{\mathrm{c}}$\ trends between the two samples. However, very recently, \citet[][ hereafter S16]{Saffe-16}
reported a positive $T_{\mathrm{c}}$ \ trend in the binary system, $\zeta^1$ Ret \ -- $\zeta^2$ Ret, \ where one of the stars ($\zeta^2$ Ret) hosts a debris disk.
The authors explained the deficit of the refractory elements relative to volatiles in $\zeta^2$ Ret\ as caused by the depletion of about $\sim$3 M$_{\oplus}$ rocky material.
The two stars have very similar atmospheric parameters, which in principle allows for a very high-precision relative abundance characterization of the stars.
Taking full advantage of this, S16 carried out a detailed and careful analysis of the system. However, the authors chose to use only the spectra
of the stars observed during the same night with the same instrument (High Accuracy Radial velocity Planet Searcher -- HARPS). The final signal-to-noise ratio (S/N) of the spectra that they used was $\sim$ 300
(S16), while we noticed that much higher S/N (see Table\,\ref{tab:sample}) can be achieved
if all of the spectra from the ESO (European Southern Observatory) HARPS archive\footnote{http://archive.eso.org/wdb/wdb/adp/phase3\_spectral/form} are combined. Moreover, some of the individual
spectra found in the archive have S/N of more than 350, which allows us to carry out a differential abundance
analysis between different spectra of the same star, which in turn provides an independent estimation of the precision of our measurements.
This case motivated us to explore whether the observed differences in the chemical abundances of the stars could be simply due to some
systematics in the spectra.
In this work, we use very high-quality spectra of the $\zeta^1$ Ret \ and $\zeta^2$ Ret \ binary star system to re-evaluate the presence and variability of the $T_{\mathrm{c}}$ \
trend in this system and, as such, better understand the impact of the presence of the debris disk on a star. We organized this paper
as follows. In Sect.\,\ref{sec:data} we present the data, in Sect.\,\ref{sec:parameters}
we present the methodology used to derive the stellar parameters and chemical abundances, and in Sect.\,\ref{sec:tc_slope} we explain how we
calculate and evaluate the significance of the $T_{\mathrm{c}}$ \ trends. After presenting the main results in Sect.\,\ref{sec:results}, we
conclude in Sect.\,\ref{sec:conclusion}.
\begin{table*}[t!]
\setlength{\tabcolsep}{3pt}
\caption{\label{tab:sample} Stellar parameters, S/N, and observation dates of $\zeta^1$ Ret \ and $\zeta^2$ Ret \ derived from individual and combined spectra.}
\centering
\begin{tabular}{lcccccc}
\hline\hline
Star(spectrum) & $T_{\mathrm{eff}}$ \ & $\log g$ & [Fe/H] & $V_{\mathrm{tur}}$ & S/N & DATE-OBS\\
\hline
sun\_vesta & 5777$\pm$10 & 4.43$\pm$0.02 & 0.020$\pm$0.010 & 0.95$\pm$0.02 & 1340 & -- \tabularnewline
$\zeta^2$ Ret\_a & 5856$\pm$15 & 4.49$\pm$0.02 & -0.205$\pm$0.012 & 0.95$\pm$0.02 & 470 & 2006-11-08T04:02:51.439 \tabularnewline
$\zeta^2$ Ret\_b & 5837$\pm$12 & 4.47$\pm$0.02 & -0.215$\pm$0.010 & 0.96$\pm$0.02 & 450 & 2009-08-27T08:42:10.196 \tabularnewline
$\zeta^2$ Ret\_c & 5835$\pm$13 & 4.48$\pm$0.03 & -0.216$\pm$0.011 & 0.94$\pm$0.02 & 420 & 2010-01-31T01:32:05.689 \tabularnewline
$\zeta^2$ Ret\_comb & 5861$\pm$12 & 4.53$\pm$0.02 & -0.215$\pm$0.010 & 1.00$\pm$0.02 & 3000 & -- \tabularnewline
$\zeta^2$ Ret\_S16 & 5833$\pm$15 & 4.48$\pm$0.03 & -0.218$\pm$0.012 & 0.95$\pm$0.03 & 230 & 2004-02-04 \tabularnewline
$\zeta^1$ Ret\_a & 5713$\pm$14 & 4.48$\pm$0.02 & -0.197$\pm$0.011 & 0.88$\pm$0.02 & 375 & 2005-11-15T03:41:34.591 \tabularnewline
$\zeta^1$ Ret\_b & 5696$\pm$14 & 4.47$\pm$0.02 & -0.197$\pm$0.011 & 0.81$\pm$0.02 & 370 & 2005-11-15T03:52:37.899 \tabularnewline
$\zeta^1$ Ret\_c & 5702$\pm$13 & 4.47$\pm$0.03 & -0.203$\pm$0.011 & 0.85$\pm$0.03 & 360 & 2005-11-15T03:47:07.475 \tabularnewline
$\zeta^1$ Ret\_comb & 5720$\pm$13 & 4.52$\pm$0.03 & -0.206$\pm$0.010 & 0.89$\pm$0.02 & 1300 & -- \tabularnewline
$\zeta^1$ Ret\_S16 & 5722$\pm$15 & 4.50$\pm$0.03 & -0.191$\pm$0.011 & 0.86$\pm$0.02 & 215 & 2004-02-04 \tabularnewline
\hline
\end{tabular}
\end{table*}
\section{Data} \label{sec:data}
Both stars of this binary system were extensively observed with the HARPS high-resolution spectrograph \citep{Mayor-03} at the 3.6 m telescope (La Silla Paranal Observatory, ESO).
The HARPS archive contains $\sim$70 and $\sim$170 spectra for $\zeta^1$ Ret \ and $\zeta^2$ Ret, respectively. We use the combined spectra
of the stars (referred as 'starname\_comb' in Table\,\ref{tab:sample}), spectra that were used in S16 (referred as 'starname\_S16' in Table\,\ref{tab:sample}),
and three individual spectra with the highest S/N for each star (referred as 'starname\_a', 'starname\_b', and 'starname\_c' in Table\,\ref{tab:sample}).
The highest S/N spectra are used to provide the final characteristics of this system. The individual high S/N spectra are selected to estimate the internal
precision of the methods, while the spectra used by S16 are selected for external comparison.
A few of the spectra of these stars were obtained under poor atmospheric conditions (the headers of the fits files report -1.00 value for seeing) ,
and we considered these spectra unsuitable for this work and, therefore, chose not to use them. These
spectra had very low S/N and combining them would increase the total S/N of the combined spectra (listed in Table\,\ref{tab:sample})
by 0.03 and 1.6\% for $\zeta^1$ Ret \ and $\zeta^2$ Ret, respectively.
The S/N at the 40th fiber order (central wavelength $\sim$ 5060\AA{}) for each of the HARPS spectra (retrieved from the header of the spectra) are presented in Table\,\ref{tab:sample}\footnote{
The S/N for the combined spectra are calculated as a quadrature sum of the individual S/Ns.}. As one can see, even the individual spectra of the stars
have a much higher S/N than that of the spectra used in S16 and, in the combined spectra, the S/N are higher by a factor of about six and $\sim$13 for $\zeta^1$ Ret \ and $\zeta^2$ Ret, \ respectively.
We note again that S16 chose to use the spectra of the stars obtained during the same night, probably to minimize the time-dependent
systematics.
For the Sun, we used a combined HARPS reflected spectrum from Vesta (extracted from the same public archive, S/N $\sim$1300).
\section{Stellar parameters and chemical abundances} \label{sec:parameters}
We derived the stellar parameters ($T_{\mathrm{eff}}$, [Fe/H], $\log g$, and $V_{\mathrm{tur}}$) for the stars from the various spectra and for the Sun with the procedure described in
\citet[][]{Sousa-14}. In short, we first automatically measured the equivalent widths (EWs) of iron lines ($\sim$250 \ion{Fe}{i} and $\sim$40 \ion{Fe}{ii} lines)
using the ARES v2 code\footnote{The last version of ARES code (ARES v2) can be downloaded at http://www.astro.up.pt/$\sim$sousasag/ares} \citep{Sousa-15}.
Then the spectroscopic parameters were derived by imposing excitation and ionization balance assuming LTE (see Fig.~\ref{plot_fe_param}).
We used the grid of ATLAS9 plane-parallel model
of atmospheres \citep{Kurucz-93} and the 2014 version of MOOG\footnote{The source code of MOOG can be downloaded at
http://www.as.utexas.edu/$\sim$chris/moog.html} radiative transfer code \citep{Sneden-73}.
We derived the parameters of the stars with a classical rather than a line-by-line differential approach.
In general, when stars with similar stellar parameters are compared, the two methods give very similar results and the estimated errors in case of differential
line-by-line analysis are slightly smaller. The uncertainties of the parameters are derived as in our previous works
and they are well described in \citet[][]{Neuforge-97}. The derived stellar parameters are listed in Table\,\ref{tab:sample}.
The values in Table\,\ref{tab:sample} show that the stellar parameters of the same stars derived from different spectra agree very well within the
estimated errors. Furthermore, when the same spectra are used, our atmospheric parameters agree very well with those derived in S16.
In particular, we obtained $\Delta$$T_{\mathrm{eff}}$ \ = 12 K, $\Delta$[Fe/H] = 0.004 dex, $\Delta\log g$ = -0.03 dex,
and $\Delta$$V_{\mathrm{tur}}$ \ = 0.06 km s$^{-1}$ for $\zeta^1$ Ret, \ where the differences are defined as our values minus those of S16.
For $\zeta^2$ Ret, \, the differences are $\Delta$$T_{\mathrm{eff}}$ \ = -21 K, $\Delta$[Fe/H] = -0.003 dex, $\Delta\log g$ = -0.06 dex, and
$\Delta$$V_{\mathrm{tur}}$ \ = 0.00 km s$^{-1}$.
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[angle=0,width=1.0\linewidth]{HD20766_20807_v2.pdf}
\end{tabular}
\end{center}
\vspace{-0.5cm}
\caption{Iron abundance vs. excitation potential (upper panels) and iron abundance vs. reduced EW (lower panels) for $\zeta^1$ Ret\_comb and $\zeta^2$ Ret\_comb.
The blue and red symbols correspond to the neutral and ionized iron lines. The black solid line indicates the linear fit of the data.}
\label{plot_fe_param}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[angle=0,width=1.0\linewidth]{oxygen_lines.pdf}
\end{tabular}
\end{center}
\vspace{-0.5cm}
\caption{Spectral regions containing oxygen lines in $\zeta^1$ Ret\_comb, $\zeta^2$ Ret\_comb, $\zeta^1$ Ret\_S16, and $\zeta^2$ Ret\_S16 spectra.}
\label{plot_oxygen}
\end{figure}
\begin{figure*}
\begin{center}
\begin{tabular}{c}
\includegraphics[angle=0,width=1.0\linewidth]{Tc_trend_rel_sun.pdf}
\end{tabular}
\end{center}
\vspace{-1.0cm}
\caption{Differential abundances [X/H] against condensation temperature for $\zeta^1$ Ret \ and $\zeta^2$ Ret . The abundances are derived relative to the Sun with the combined spectra and spectra from S16. The black dashed line represents the trend when all of the elements are used for the linear regression and
the red line is the result of the linear regression when only elements with $T_{\mathrm{c}}$ \ $>$ 900K are used.}
\label{fig_rel_sun}
\end{figure*}
Similar to S16, we derived 0.027 dex higher metallicity for $\zeta^1$ Ret \ compared to $\zeta^2$ Ret \ with the same spectra. However, the difference
decreases to 0.009 dex when the spectra with the highest S/N ratio are considered. Also, when different pairs of spectra of the stars are compared,
the maximum difference is obtained when the $\zeta^1$ Ret\_S16 and $\zeta^2$ Ret\_S16 are used, and when $\zeta^1$ Ret\_comb and $\zeta^2$ Ret\_a pairs are compared, the $\zeta^2$ Ret \
appears even slightly more metal rich than $\zeta^1$ Ret \ by -0.001 dex. These results suggest that one should be cautious when
reaching such an extremely small difference in stellar parameters.
Elemental abundances for the stars were also determined using an LTE analysis and the same tools and codes as for stellar parameters determination.
We adopted the initial line list for Na, Mg, Al, Si, Ca, Ti, Cr, Ni, Co, Sc, Mn, and V from \citet{Adibekyan-12c},
but several lines (two \ion{Si}{i} lines $\lambda$5701.11, $\lambda$6244.48, two \ion{Ca}{i} lines $\lambda$5261.71, $\lambda$5352.05,
three \ion{Ti}{i} lines $\lambda$4722.61, $\lambda$5039.96, $\lambda$5965.84, one \ion{Ti}{ii} line $\lambda$ 5381.03,
two \ion{Cr}{i} lines $\lambda$5122.12, $\lambda$6882.52, and two \ion{Ni}{i} lines $\lambda$5081.11, $\lambda$6767.78)
were excluded because of large [X/Fe] star-to-star scatter at solar metallicities \citep{Adibekyan-15b}. Finally, \ion{Ni}{i} line at $\lambda$6360.81
was excluded because of unreliable EW measurements (because of a depressed continuum) and the \ion{Sc}{i} line at $\lambda$5520.50 was excluded because
there is no data about the hyperfine structure splitting (HFS) of the lines in
Kurucz atomic line database\footnote{http://kurucz.harvard.edu/linelists.html} nor in \citet{Prochaska-00}, which are our two sources for
the relative $\log gf$ values.
We determined oxygen abundances using two weak lines at 6158.2\AA{} and 6300.3\AA{} following the work of
\citet{Bertrandelis-15}, although it was only possible to measure the 6158.2\AA{} lines for the combined spectra of $\zeta^1$ Ret \ and $\zeta^2$ Ret\ and
for the three individual spectra of $\zeta^2$ Ret.
Sulfur abundances were calculated by performing spectral synthesis with MOOG around
the lines 6046.0\AA{}, 6052.5\AA{}, 6743.5\AA,{} and 6757.1\AA{}. The atomic data of those lines and the
surrounding lines were taken from VALD3 database\footnote{http://vald.astro.univie.ac.at/~vald3/php/vald.php?newsitem=0}.
The lines at 6046.0\AA{} and 6052.5\AA{} (used in S16) suffer from non-negligible blends of CN bands, thus, our initial
abundances derived from EWs yielded higher abundances than the other two lines.
To fit the lines in the solar spectra, we had to calibrate the $\log gf$s of the S lines (the line list and atomic data are available at the CDS), to match a solar sulfur abundance A(S)=7.16 \citep{Asplund-09}.
Cerium abundances were derived using the three spectral lines presented in \citet{Reddy-03} and the line at 5274.2\AA{} for which the
atomic data was extracted from VALD3.
Carbon abundances were based on the two well-known CI optical lines at 5052\AA{} and 5380\AA{}.
The atomic data for carbon and elements heavier than Ni were also extracted from VALD3
(a more detailed line list will be shown in Delgado Mena et al. in prep).
When available, the Barklem damping van der Waals constants were used
for all of the elements.
Otherwise, the Uns\"{o}ld approximation multiplied by a factor (1.0+0.67$\times$E.P., where E.P. is the excitation potential of a line) suggested by the Blackwell
group was used (option 2 in the damping parameter inside MOOG).
For the Sc, V, Mn, Co, Cu, and Ba lines, HFS was considered. We adopted the atomic parameters and isotopic ratios from \citet{Prochaska-00} for Ba, and the relative $\log gf$ values and isotopic ratios for the lines were taken from Kurucz database for
the remaining elements.
The EWs of the lines were derived with ARES v2 code, but with careful visual inspection. In a few cases, when the ARES measurements were not
satisfactory (this can happen, e.g., as a result of a presence of cosmic rays or bad pixels), we measured the EWs using the task \emph{splot}
in IRAF\footnote{IRAF is distributed by National Optical Astronomy Observatories,
operated by the Association of Universities for Research in Astronomy, Inc., under contract with the National Science Foundation, USA.}.
We calculated the final abundance of the elements (when several spectral lines were available) as a weighted mean of all
of the abundances, where the distance from the median abundance was considered as a weight. As demonstrated in \citet{Adibekyan-15b}, this method can be effectively
used without removing suspected outlier lines.
We performed differential line-by-line analysis relative to the Sun for the combined and S16 spectra. We also performed differential
abundance analysis of $\zeta^2$ Ret\_comb relative to $\zeta^1$ Ret\_comb, and $\zeta^2$ Ret\_S16 relative to $\zeta^1$ Ret\_S16.
Finally, for both stars we derived differential abundances from each of the individual three spectra
('starname\_b' and 'starname\_c' relative to 'starname\_a') relative to each other.
\begin{table*}[t!]
\setlength{\tabcolsep}{3pt}
\caption{\label{tab:slopes} Slopes of the [X/H] vs. $T_{\mathrm{c}}$ \ for different pairs of spectra. Both frequentist and Bayesian approaches are chosen to derive
the slopes and their uncertainties.}
\tiny
\begin{center}
\begin{tabular}{lccccc|ccccc}
\hline\hline
Star -- Reference &\multicolumn{3}{c}{WLS (all elements)} & \multicolumn{2}{c}{Bayesian (all elements)} & \multicolumn{3}{c}{WLS (refractories)} & \multicolumn{2}{c}{Bayesian (refractories)}\\ \cmidrule(l){2-4}\cmidrule(l){5-6}\cmidrule(l){7-9}\cmidrule(l){10-11}
& \multicolumn{1}{c}{Slope$\pm \sigma$} & 95\% CI & P(F-stat) & Slope & \multicolumn{1}{c}{95\% HPD} & \multicolumn{1}{c}{Slope$\pm \sigma$} & 95\% CI & P(F-stat) & Slope & \multicolumn{1}{c}{95\% HPD}\\
\hline
$\zeta^1$ Ret\_comb -- Sun & -1.32$\pm$2.28 & [-6.04, 3.40] & 0.567 & -1.32 & [-2.11, -0.46] & 5.40$\pm$ 4.37 & [-3.78, 14.58] & 0.232 & 5.37 & [3.75, 6.99] \tabularnewline
$\zeta^1$ Ret\_S16 -- Sun & -3.83$\pm$3.92 & [-11.96, 4.30] & 0.339 & -3.83 & [-5.67, -1.75] & -0.99$\pm$ 5.84 & [-13.25, 11.28] & 0.868 & -1.01 & [-3.56, 1.75] \tabularnewline
$\zeta^2$ Ret\_comb -- Sun & -5.95$\pm$2.66 & [-11.46, -0.44] & 0.036 & -5.95 & [-6.93, -4.84] & -1.81$\pm$ 5.07 & [-12.45, 8.83] & 0.725 & -1.81 & [-3.78, -0.02] \tabularnewline
$\zeta^2$ Ret\_S16 -- Sun & -7.68$\pm$3.19 & [-14.30, -1.07] & 0.025 & -7.68 & [-9.22, -6.12] & -7.16$\pm$ 6.89 & [-21.63, 7.31] & 0.312 & -7.19 & [-10.22, -4.26] \tabularnewline
$\zeta^1$ Ret\_comb -- $\zeta^2$ Ret\_comb & 3.41$\pm$1.23 & [0.86, 5.96] & 0.011 & -3.41 & [-4.27, -2.52] & 4.88$\pm$ 2.16 & [0.35, 9.41] & 0.036 & 4.88 & [3.45, 6.29] \tabularnewline
$\zeta^1$ Ret\_S16 -- $\zeta^2$ Ret\_S16 & 3.77$\pm$2.16 & [-0.71, 8.25] & 0.095 & -3.79 & [-5.37, -2.00] & 4.83$\pm$ 3.42 & [-2.34, 12.01] & 0.174 & 4.86 & [2.52, 7.31] \tabularnewline
$\zeta^1$ Ret\_b -- $\zeta^1$ Ret\_a & 2.35$\pm$2.21 & [-2.24, 6.94] & 0.300 & 2.34 & [1.17, 3.62] & 4.83$\pm$ 2.83 & [-1.12, 10.77] & 0.105 & 4.83 & [3.34, 6.44] \tabularnewline
$\zeta^1$ Ret\_c -- $\zeta^1$ Ret\_a & -0.14$\pm$0.84 & [-1.89, 1.61] & 0.867 & -0.14 & [-1.30, 0.99] & -0.22$\pm$ 1.40 & [-3.16, 2.73] & 0.880 & -0.21 & [-1.91, 1.59] \tabularnewline
$\zeta^1$ Ret\_c -- $\zeta^1$ Ret\_b & 1.13$\pm$0.76 & [-0.45, 2.70] & 0.153 & 1.13 & [0.24, 2.00] & -0.28$\pm$ 0.90 & [-2.18, 1.61] & 0.759 & -0.28 & [-1.45, 0.80] \tabularnewline
$\zeta^2$ Ret\_b -- $\zeta^2$ Ret\_a & -1.60$\pm$0.79 & [-3.24, 0.04] & 0.055 & -1.61 & [-3.09, -0.18] & -1.67$\pm$ 1.04 & [-3.84, 0.51] & 0.125 & -1.68 & [-3.43, 0.12] \tabularnewline
$\zeta^2$ Ret\_c -- $\zeta^2$ Ret\_a & -1.36$\pm$0.76 & [-2.94, 0.22] & 0.087 & -1.35 & [-2.59, -0.17] & -0.35$\pm$ 0.94 & [-2.33, 1.63] & 0.714 & -0.37 & [-2.16, 1.32] \tabularnewline
$\zeta^2$ Ret\_c -- $\zeta^2$ Ret\_b & -1.86$\pm$1.06 & [-4.05, 0.33] & 0.092 & -1.86 & [-3.24, -0.44] & 0.72$\pm$ 1.37 & [-2.15, 3.59] & 0.603 & 0.70 & [-1.52, 3.09] \tabularnewline
\hline
\end{tabular}
\end{center}
Note:$^{*}$ The units of the slopes are in 10$^{-5}$ dex K$^{-1}$.
\end{table*}
The errors of the [X/H] abundances are calculated as a quadrature sum of the errors due to EW measurements and errors due to uncertainties in the atmospheric
parameters. When three or more lines were available, the EW measurement error is estimated as $\sigma /\sqrt{(n-1)}$, where $\sigma$ is
the line-to-line abundance scatter and $n$ is the number of the observed lines. The errors arising from uncertainties in the stellar parameters
are calculated as a quadratic sum of the abundance sensitivities on the variation of the stellar parameters by their one $\sigma$ uncertainties.
These errors are usually much smaller (because of very small uncertainties in stellar parameters) than the errors due to EW measurements.
The error estimation method described in the previous paragraph is most frequently used, even when only two lines are observed. However,
the dispersion ($\sigma_{sample}$) calculated using only two lines can be very different from the dispersion rooted in the uncertainty on the abundances,
a value that can only be efficiently estimated if large number of lines are available.
A simple calculation shows that if one assumes that the errors follow a normal distribution and that $\sigma_{sample}$ is the standard deviation of that
distribution, then the $\sigma_{two-lines}$ calculated from two random lines from that distribution peaks close to zero, underestimating $\sigma_{sample}$.
The $\sigma_{two-lines}$ is smaller than 0.5$\sigma_{real}$ in $\sim$38\% cases and $\sigma_{two-lines}$ is five times
smaller than $\sigma_{real}$ in $\sim$16\% cases. Moreover, in the cases when two lines of an element give exactly the same abundance, then the calculated dispersion is zero
and the final error is equal to the error due to uncertainties in the atmospheric parameters. Since the latter is usually very small, the weight given to
this element when calculating the $T_{\mathrm{c}}$ \ slope is extremely high (see Sect.\,\ref{sec:tc_slope}). Such underestimated errors can play a crucial role
in the incorrect determination of the slopes.
We calculated the errors on EWs following \citet[][]{Cayrel-98} to provide more realistic errors for the abundances of elements that
only have two observed spectral lines in our spectra (except oxygen) . The calculated uncertainty takes into account the statistical photometric error due to the noise in each pixel and
the error related to the continuum placement, which is the dominant contribution to the error \citep[][]{Cayrel-98, Bertrandelis-15}. Then, these
errors are propagated to derive the abundance uncertainties for each line. The final uncertainties for the average abundance are propagated from these individual errors.
When only one line for a given element is available, as is the case for O (for some spectra) and Sr, we determine the error by
measuring a second EW with the position of the continuum displaced within the root mean square \textit{(rms)}
due to the noise of the spectra, and calculating the difference in abundance with respect to the original value. In this manner, the elements with only a weak line
(as for oxygen in some cases) have a large error in their abundance measurements. This is in contrast to elements with strong lines, such as Sr, where the fact
of having a single line does not affect the final error too much; see, for comparison, the error bars in Fig.~\ref{fig_rel_sun} for O and Sr and how they vary
between the highest and lowest S/N spectra). We stress here the case of oxygen for which we find a very different value than S16.
Their differential abundance ($\zeta^1$ Ret \ -- $\zeta^2$ Ret) is almost 0.1 dex lower than ours and their reported error is much smaller.
In Fig.~\ref{plot_oxygen} we show a part of the spectra used by S16 around the 6300\AA{} region; it is strongly affected by noise.
The EW values are very similar for both stars, but our abundance determinations have a larger error because of the strong
effect of a small change in the continuum in this very weak line.
As mentioned by S16, this line is blended with a Ni line, and if we consider that the two stars have similar
Ni abundances, it seems difficult
to reconcile such a large difference in abundance values even if our atomic parameters for the Ni blend are different. However, for the combined, very high S/N spectra we can also
measure with better precision the oxygen line at 6158.2\AA,{} which suggests very similar oxygen abundance to that derived from the
6300.3\AA{} line. This example shows the difficulty in measuring
oxygen abundances. Such an issue is very important in the context of $T_{\mathrm{c}}$ \ trend analysis, as the low condensation
temperature of oxygen leads to a strong leverage on the final slope value.
Since both oxygen lines are very weak and deserve special attention \citep[see, e.g.,][]{Bertrandelis-15}, even when both lines were
observed for a spectrum we calculated individual errors for each line (as described in the previous paragraph) and then propagated the error of the average
oxygen abundance.
The EWs of all of the measured lines and the abundances of the elements derived from all of the spectra are available on the CDS.
\section{$T_{\mathrm{c}}$ \ slope} \label{sec:tc_slope}
Once the differential abundances and corresponding errors are derived we can search for abundance trends with the condensation temperature of the elements.
The 50\% $T_{\mathrm{c}}$ \ equilibrium condensation temperatures for a gas of solar system composition are taken from \citet{Lodders-03}. It is common practice to plot
[X/Fe] (and not just [X/H]) versus $T_{\mathrm{c}}$ \ , which allows us to remove the trends related to Galactic chemical evolution \citep[e.g.,][]{GH-13, Saffe-16}. Although
this procedure can be justified when stars with different metallicities are compared, it can produce a bias in the derived slope.
Subtracting the iron abundance (a constant value) from each elemental
abundance should only shift the trend but should not change the slope. However, since iron abundance has an error itself, the propagated errors of the [X/Fe] abundance ratios
should be adjusted. This change in turn can have an impact on the original slope. In an extreme (although not very realistic) case that the error of iron abundance is much larger
than the errors of individual [X/H] abundances, then the propagated errors of [X/Fe] abundance ratios would be dominated by the error of [Fe/H]. In this case the
weighted least-squares regression would give a result similar to an ordinary least-squares regression, i.e., the weights of all of the elements would be similar. This discussion
suggests that if the abundances are normalized to another element (not iron), the result would be different depending on the error of the element used
for normalization. Moreover, if the abundances are not normalized to the iron abundance, the [Fe/H] value can also be used in the calculations of the slopes.
For these reasons we use [X/H] in our analysis when deriving the $T_{\mathrm{c}}$ \ trend.
In Table\,\ref{tab:slopes} we present the $T_{\mathrm{c}}$ \ slopes, corresponding standard errors, 95\% confidence intervals (CI) and p-values (at $\alpha$=0.05 significance level).
The $T_{\mathrm{c}}$ \ slopes refer to the slope of the linear dependence between [X/H] and $T_{\mathrm{c}}$ . We calculated the slopes with the weighted least-squares (WLS)
technique, whereas we calculated the weights as the inverse of the variance ($\sigma^2$) of the abundance. The p-values come from
the F-statistics that tests the null hypothesis that the data can be modeled accurately by setting the regression coefficients to zero. The 95\% confidence intervals
are calculated using the standard error of the slopes and the p-value from T-statistics that tests the null hypothesis that the
coefficient of a predictor variable is zero, i.e., slope is zero.
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[angle=0,width=1\linewidth]{Tc_trend_rel_z1.pdf}
\end{tabular}
\end{center}
\vspace{-1.cm}
\caption{Differential abundances ($\zeta^2$ Ret \ -- $\zeta^1$ Ret ) against condensation temperature. The abundances are derived for the combined spectra and spectra that were used in S16. The black dashed line and red line are the same as in Fig.~\ref{fig_rel_sun}.}
\label{fig_rel_z1}
\end{figure}
\begin{figure*}
\begin{center}
\begin{tabular}{c}
\includegraphics[angle=0,width=1.0\linewidth]{Tc_trend_zet1abc_zet2abs.pdf}
\end{tabular}
\end{center}
\vspace{-1.0cm}
\caption{Differential abundances against condensation temperature for $\zeta^1$ Ret \ and $\zeta^2$ Ret , derived from three highest S/N individual spectra.
The black dashed line and red line are the same as
in Fig.~\ref{fig_rel_sun}.}
\label{fig_zet1abc_zet2abs}
\end{figure*}
We also use a Bayesian approach to assess the presence of a dependence of [X/H] on $T_{\mathrm{c}}$ \, and to calculate
the slope of the dependence.
It has been shown that p-value analysis, even if very widespread, is not as robust as expected; the assessment of the significance of a
correlation is further complicated when the p-value is close to the significance level.
In \citet{Figueira-16} we provided a straightforward alternative Bayesian approach to the assessment and characterization of a
correlation in a dataset. However, relying on the Pearson correlation coefficient and Sperman's rank, \citet{Figueira-16} did not consider the
impact of error bars on the measurements. We consider the case of a linear regression in which the value of each ordinate $Y_i$
follows a Gaussian distribution of center dictated by the evaluation of a linear slope on the abscissa $X_i$ and variance given by $\sigma^2$.
Using the MCMC (Markov Chain Monte Carlo) algorithm implemented in PyMC \citep{Patil-10}, and assuming uninformative Gaussian distributions of center 0 and $\sigma$\,=\,1000 as priors
for the slope parameters and intersect, we estimated the best fit to the data.
In Table\,\ref{tab:slopes} we present the $T_{\mathrm{c}}$ \ slopes and the associated 95\% credible intervals (highest posterior density; HPD) calculated by
applying this Bayesian approach.
The Bayesian and frequentist analyses differ in important ways. Since both approaches are addressing and solving exactly
the same problem, the slopes calculated by the two methods are almost identical. However, since the CI (in frequentist analysis) and the HPD (in Bayesian approach)
have different meanings, their values can be different and should be interpreted differently \citep[e.g.,][]{Morey-15}.
A 95\% CI is an interval that in repeated sampling has a 0.95 probability of containing the true value of the parameter,
i.e, if a large number of samples are used, the true value of the slope
will fall within the CI in 95\% of the cases. A 95\% HPD is the interval that contains the true value of slope with a probability of 0.95 given the current sample (data).
This said, it should not come as a surprise that a 95\% HPD is usually much smaller than the 95\% CI, which makes most of the correlations significant
(at 95\% level).
In Table\,\ref{tab:slopes} we present the $T_{\mathrm{c}}$ \ slopes that were derived for different pairs of stars and spectra. The slopes
are derived by considering all of the elements and also by considering only refractory elements that have $T_{\mathrm{c}}$ \ $>$ 900 K.
To understand better how the calculated $T_{\mathrm{c}}$ \ slopes are sensitive to the abundance uncertainties, we also calculated the slopes without considering
the errors. In the frequentist approach giving the same error (which means the same weight) to all of the elements or not considering errors yields the same
results, since the calculated slope value is only sensitive to the relative weights of the points. In the Bayesian approach, the absolute error value of each point is
important when evaluating the significance of the correlations, i.e., assuming 0.01 dex error for all of the elements produces
a result that is very different if an error of 0.1 dex was assumed for all of the elements.
In the Appendix we present all of the $T_{\mathrm{c}}$ \ trend calculations discussed in the main text,
without considering the errors for the individual abundances, i.e., applying an ordinary linear regression (OLS).
\section{Results} \label{sec:results}
In Fig.~\ref{fig_rel_sun} we show the dependence of differential [X/H] abundances of $\zeta^1$ Ret \ and $\zeta^2$ Ret \ relative to the Sun on the corresponding $T_{\mathrm{c}}$. The abundances
are derived using the combined spectra and spectra that were used in S16. Table\,\ref{tab:slopes} and the corresponding plot shows
that there is no trend with $T_{\mathrm{c}}$ \ for $\zeta^1$ Ret, \ while $\zeta^2$ Ret \ shows a significant trend when the frequentist p-values are considered
for null-hypothesis rejection testing. This trend, seen both from the combined and S16 spectra, is heavily driven
by SI and ZnI (due to their small errors and hence large weight) and disappears when only refractory elements with $T_{\mathrm{c}}$ \ $> 900$ K
are considered. We note that S16 did not find a significant $T_{\mathrm{c}}$ \ trend for the two stars when compared with the Sun.
However, direct comparison of the slopes is not straightforward since we use different elements than S16 and, furthermore, those investigators corrected the abundances for the Galactic chemical evolution effect,
while we did not.
Since the differential abundances of the stars relative to the Sun can be affected by several factors, such as GCE, age, or the relative effect of NLTE
on stars with similar, but not exactly the same stellar parameters, and we are interested in testing whether the two stars of the binary system show different
refractory-to-volatile element ratios, in Fig.~\ref{fig_rel_z1} we plot the differential abundances of $\zeta^1$ Ret \ relative to $\zeta^2$ Ret.
The plot and corresponding
Table\,\ref{tab:slopes} shows that there is a strong and significant correlation with the $T_{\mathrm{c}}$ \ in both cases when all of the elements
and only refractory elements are considered for the linear regression. The slopes derived from the combined and S16 spectra
are very similar, but the errors of the slopes for the case of combined spectra are smaller and the significance is higher.
Following the interpretation of \citet{Melendez-09}, in S16 the authors
based on this result proposed that refractory elements depleted in $\zeta^2$ Ret \ are locked up in the debris disk that the star hosts.
To test the hypothesis that the observed $T_{\mathrm{c}}$ \ trend is due to the presence of a debris around $\zeta^2$ Ret ,\ we performed the following test. We used
the three highest S/N
individual spectra for each star ($\zeta^1$ Ret \ and $\zeta^2$ Ret) and derived differential abundances for each star using different spectra as reference. The results
are plotted in Fig.~\ref{fig_zet1abc_zet2abs} and presented in Table\,\ref{tab:slopes}. This test clearly shows that the slope
of the $T_{\mathrm{c}}$ \ trend significantly changes depending on which pair of spectra are compared.
It is interesting to note that some of the observed trends are relatively significant (e.g., $\zeta^2$ Ret\_b -- $\zeta^2$ Ret\_a).
These results, in turn, mean that if different spectra of the two stars
are used to derive the $T_{\mathrm{c}}$ \ trend between the two stars, very different results can be obtained, ranging in significance from nonsignificant
(i.e., having a slope compatible with zero) to very significant.
It is difficult to identify the main reason(s) of the observed $T_{\mathrm{c}}$ \ trend between different spectra of the same star.
All of the spectra are observed with the same, and most stable spectrograph, HARPS, which minimizes the possible effects of spectral
resolution and instrumental profile. While the three spectra of $\zeta^2$ Ret \
are taken at different dates (on 11 August 2006, 27 August 2009, and 31 January 2010, for $\zeta^2$ Ret\_a, $\zeta^2$ Ret\_b, and $\zeta^2$ Ret\_c, respectively), the selected three
individual high S/N spectra of $\zeta^1$ Ret \ are taken on the same night (15 November 2005). This means that time-dependent systematics
are much less likely to play an important role for the observed trend for $\zeta^1$ Ret. This could be the reason that the $T_{\mathrm{c}}$ \ slopes
for this star are less significant compared to those of its companion.
The two stars exhibit different levels of activity;
$\zeta^1$ Ret \ ($<\log R_{HK}^{'}> \ = -4.662$) is more active than $\zeta^2$ Ret \ ($<\log R_{HK}^{'}> \ = -4.892$) \citep{Zechmeister-13}.
Although $\zeta^1$ Ret \ is more active than its companion, the stellar variability should not play a significant role in the observed difference in the chemical abundances
between the different spectra of the star since, again, the observations are carried out on the same night.
Recently, \citet{Bedell-14} analyzed solar spectra observed with different instruments, from different asteroids, and at different times, i.e, conditions.
The authors reached a conclusion that the major effect on differential relative abundances is caused by the use of different instruments. They also found no significant (more than 2$\sigma$) $T_{\mathrm{c}}$ \ trend between
the different spectra they used.
One of the main factors that can produce (or affect) the $T_{\mathrm{c}}$ \ trend is the correct derivation of stellar parameters and their errors.
Different elements (and corresponding spectral lines) show different sensitivities to atmospheric temperature
(depending on the excitation potential), surface gravity (depending on the ionization state), and stellar metallicity
\citep[see, e.g.,][]{Adibekyan-12c}.
Another very important factor that can determine the $T_{\mathrm{c}}$ \ slope is the error estimation for the individual elements. Since we use
the errors as weights for each abundance, the elements with the smallest errors have the largest weights and may determine the
slope and its confidence interval. As discussed in Sect.\,\ref{sec:parameters}, the errors can be underestimated only if two
lines are used and the line-to-line dispersion is used to estimate the errors on the EWs.
It is also important to mention the high weights of elements with a large number of lines in the derivation of the $T_{\mathrm{c}}$ \ slopes.
Since the error of the abundances is
calculated as a $\sigma /\sqrt{(n-1)}$, the elements such as Ni and Ti, which have many spectral lines, show the smallest errors
and, hence, have a key role in the determination of the slope value.
The importance of the errors of individual chemical abundances is well illustrated in
Figs.~\ref{fig_rel_sun_without_errors},\ref{fig_rel_z1_without_errors},\ref{fig_zet1abc_zet2abs_without_errors}, and
in Table\,\ref{tab:slopes_without_errors}. In these plots and table, we present the derivation of the $T_{\mathrm{c}}$ \ slopes for the
above discussed stars and spectra, but applying OLS, i.e., no errors (or equal errors) are considered. We can clearly see that for the same star or spectra the OLS and WLS give very different results. The Table\,\ref{tab:slopes_without_errors}
also shows that the most significant trends are observed for ($\zeta^1$ Ret\_comb -- $\zeta^2$ Ret\_comb), ($\zeta^2$ Ret\_b -- $\zeta^2$ Ret\_a), and
($\zeta^2$ Ret\_c -- $\zeta^2$ Ret\_a). Moreover, the only significant trend that remains when considering only the refractory elements
with $T_{\mathrm{c}}$ \ $>$ 900K is for the ($\zeta^2$ Ret\_c -- $\zeta^2$ Ret\_a) pair.
\section{Conclusion} \label{sec:conclusion}
The $\zeta$ Reticuli binary system, composed of two solar analogs, is a very interesting and well-known system not only in the scientific literature,
but also in science fiction literature\footnote{See \url{https://en.wikipedia.org/wiki/Stars\_and\_planetary\_systems\_in\_fiction\#Zeta\_Reticuli}} %
and movies,\footnote{\url{https://en.wikipedia.org/wiki/The_UFO_Incident}} %
as a word of Zeta Reticulans\footnote{\url{https://en.wikipedia.org/wiki/Grey_alien}}%
.
Both stars of this binary system are in planet search programs and the presence of short-period small (or massive) planets or
larger period massive planets can be excluded (see S16). However, there is indirect evidence that the $\zeta^2$ Ret \ may have a massive eccentric
companion that perturbed its eccentric debris disk \citep{Faramaz-14}.
We used several high-quality (S/N $>$ 360) individual and combined spectra (S/N of 1300 and 3000) of $\zeta^1$ Ret \ and $\zeta^2$ Ret \ to revisit the results
obtained in S16, namely that $\zeta^2$ Ret \ shows a deficit of refractory elements (relative to volatiles) when compared to $\zeta^1$ Ret \ probably because of the
presence of the debris disk that $\zeta^2$ Ret \ hosts. We first derived the stellar parameters using the classical (nondifferential) method,
then applied careful line-by-line differential abundance analysis of the stars relative to the Sun and relative to each other.
When we consider the combined, highest S/N spectra of the two stars we obtain similar $T_{\mathrm{c}}$ \ trend as was reported in S16.
The trend exists when all of the elements and only refractory elements are considered in the derivation of the $T_{\mathrm{c}}$ \ slope.
We also show that when comparing the chemical abundances of the same individual stars ($\zeta^1$ Ret \ with $\zeta^1$ Ret \ and $\zeta^2$ Ret \ with $\zeta^2$ Ret)
derived from different individual spectra, we observe different and sometimes significant $T_{\mathrm{c}}$ \ trends. The trends observed between different individual
spectra of $\zeta^1$ Ret \ that are observed during the same night, are less significant than those observed for the $\zeta^2$ Ret \ that
are observed on different nights.
In this context, there is no consensus on the results of the $T_{\mathrm{c}}$ \ trend
for some individual systems \citep[e.g., 16 Cyg AB;][]{Laws-01, Takeda-05, Schuler-11a, TucciMaia-14}. The reported differences
can be due to the same sort of effect that we discussed in this study.
Our results show that when studying tiny chemical abundance trends with condensation temperature, it is very important to use very high-quality combined spectra.
The combination of several spectra increase the S/N and may minimize possible time-dependent effects.
The results also show that there are other nonastrophysical factors (such as over- or underestimation of the errors of the individual elements)
that may be responsible for the observed $T_{\mathrm{c}}$ \ trends and indicate that
one should be very careful when analyzing very subtle differences in chemical abundances between stars.
However, we should stress that this result does not imply that all of the observed $T_{\mathrm{c}}$ \ trends do not have an astrophysical origin.
\begin{acknowledgements}
{This work was supported by Funda\c{c}\~ao para a Ci\^encia e Tecnologia (FCT) through national funds (project ref. PTDC/FIS-AST/7073/2014)
and by FEDER through COMPETE2020 (project ref. POCI-01-0145-FEDER-007672). This work was also supported by FCT through the research grants (ref. PTDC/FIS-AST/7073/2014
and ref. PTDC/FIS-AST/1526/2014) through national funds and by FEDER through COMPETE2020 (ref. POCI-01-0145-FEDER-016880 and ref. POCI-01-0145-FEDER-016886).
This work results within the collaboration of the COST Action TD1308.
P.F., N.C.S., and S.G.S. also acknowledge the support from FCT through Investigador FCT contracts of reference IF/01037/2013, IF/00169/2012,
and IF/00028/2014, respectively, and POPH/FSE (EC) by FEDER funding through the program ``Programa Operacional de Factores de Competitividade - COMPETE''.
PF further acknowledges support from FCT in the form of an exploratory project of reference IF/01037/2013CP1191/CT0001.
V.A. and E.D.M acknowledge the support from the FCT in the form of the grants SFRH/BPD/70574/2010 and SFRH/BPD/76606/2011, respectively.
V.A also acknowledges the support from COST Action TD1308 through STSM grant with reference Number: COST-STSM-TD1308-32051.
G.I. acknowledges financial support from the Spanish Ministry project MINECO AYA2011-29060.
JIGH acknowledges financial support from the Spanish Ministry of Economy and Competitiveness (MINECO)
under the 2013 Ram\'{o}n y Cajal program MINECO RYC-2013-14875, and the Spanish ministry project MINECO AYA2014-56359-P.}
\end{acknowledgements}
|
\section{Introduction}\label{1}
This is the second part in a series of three articles devoted to the
study of a class of geometric structures and of differential complexes
naturally associated to these structures. The main motivation for this
series are the examples of differential complexes on complex
projective space, which were constructed and applied to questions of
integral geometry in \cite{E-G}. An attempt to put these complexes into
the context of geometric structures was made in the first version of
the preprint \cite{E-S}, which introduced the concept of a conformally
Fedosov structure. The aim of providing a framework analogous to
Bernstein--Gelfand--Gelfand resolutions (or BGG resolutions)
associated to parabolic geometries was realized (for conformally
Fedosov structures) in the second version of \cite{E-S}, which has
appeared recently.
Our article builds on \cite{PCS1}, where we introduced a family of
first order structures, which all have an underlying almost
conformally symplectic structure. There is one such structure for each
contact grading of a simple Lie algebra, which is not of type
$C_n$. These gradings are related to certain parabolic subalgebras and
to parabolic contact structures as discussed below, which motivates
the name \textit{parabolic almost conformally symplectic structures}
(or PACS--structures for short). The main result of \cite{PCS1} is
that each such structure determines a canonical connection on the
tangent bundle, which is characterized by a normalization condition on
its torsion. The torsion of this connection is a basic invariant of
the structure, which naturally splits into two parts. One of these
parts is the obstruction to the almost conformally symplectic
structure being conformally symplectic, and requiring this part of the
torsion to vanish, one arrives at the subclass of PCS--structures or
parabolic conformally symplectic structures.
In this second part, we only consider PCS--structures and we relate
them to another class of geometric structures, called parabolic
contact structures. There is one such structure for each contact
grading of a simple Lie algebra, and they all have an underlying
contact structure. The most prominent example of a parabolic contact
structure is provided by (partially integrable almost) CR--structures
of hypersurface type. Via the relation to contact gradings, each type
of PACS--structure determines a corresponding type of parabolic contact
structure in one higher dimension.
Now for parabolic contact structures, there is the concept of
transversal infinitesimal automorphisms, which in particular are
transversal infinitesimal contactomorphisms for the underlying contact
structure. Any transversal infinitesimal contactomorphism defines a
foliation with one--dimensional leaves. In \cite{Cap-Salac}, we have
shown that any local leaf--space for such a foliation naturally
inherits a conformally symplectic structure. Further we have shown
that locally any conformally symplectic structure can be realized in
this way and that these realizations are locally unique up to
contactomorphism.
In the current article, we extend all these results to the setting of
PCS--structures and parabolic contact structures, see Theorems
\ref{2.5}, \ref{2.7}, and \ref{2.8}. Moreover, we discuss the relation
between distinguished connections associated to parabolic contact
structures and the canonical connections of PCS--structures. Together
with Section 4.7 of \cite{PCS1}, this allows us to complete the
discussion of the relation of PCS--structures to special symplectic
connections in the sense of \cite{Cahen-Schwachhoefer} and thus to
exceptional holonomies, see Theorem \ref{2.10}.
The contact gradings of Lie algebras of type $C_n$ do not give rise to
a PACS--structure. However, there is a type of parabolic contact
structures associated to these gradings, the so--called contact
projective structures. For the applications in \cite{E-G} it was only
necessary to construct the initial parts of certain differential
complexes. However, the constructions can be extended, exhibiting that
the results look similar to the BGG resolutions associated to locally
flat contact projective structures. Also, the tractor bundle
associated to a conformally Fedosov structure in \cite{E-S} looks
similar to the standard tractor bundle of a contact projective
manifold in one higher dimension. So all that suggests that
conformally Fedosov structures might be related to parabolic contact
structures. In Section \ref{3} of this article, we show that this is
indeed the case and extend the above mentioned results on structures
on quotients and contactifications from \cite{Cap-Salac} also to this
setting. This will allow us to include the $C_n$--types into a uniform
treatment in the last part of this series.
There is also a class of special symplectic connections associated to
Lie algebras of type $C_n$, the so--called symplectic connections of
Ricci type, see \cite{Cahen-Schwachhoefer} and \cite{BC}. We also get
a characterization of these connections as conformally Fedosov
structures with locally flat parabolic contactification, see Corollary
\ref{3.3}. Together with Theorem \ref{2.10} this recovers the fact
that all special symplectic connections can be obtained from local
quotients of the homogeneous models of parabolic contact structures,
which is the crucial result of \cite{Cahen-Schwachhoefer}.
A third important aim of this article is to give examples of global
leaf spaces of compact contact manifolds which inherit PCS--structures
respectively conformally Fedosov structures. From the point of view of
differential complexes, such global quotients are interesting, since
they lead to information on the cohomology of the resulting sequences,
which are important for the applications in \cite{E-G}. As one of
these examples, we show that the Hopf--fibration $S^{2n+1}\to\mathbb
CP^n$ can be considered as a global parabolic contactification in two
ways. On the one hand, this concerns the CR--structure on $S^{2n+1}$
coming from the embedding as a hypersurface in $\mathbb C^{n+1}$, which
induces the K\"ahler structure on $\mathbb CP^n$ as the underlying
PCS--structure. On the other hand, viewing $S^{2n+1}$ as the space of
real rays in $\mathbb C^{n+1}$ (viewed as a real symplectic vector
space), it inherits a contact projective structure. This gives rise to
the conformally Fedosov structure on $\mathbb CP^n$ studied in
\cite{E-S}.
The second example of such a global space of leaves we discuss is a
bit more exotic. In Theorem \ref{2.6}, we show that the Grassmannian
of complex planes in $\mathbb C^{n+1}$ can be realized as a global space
of leaves for a transversal infinitesimal automorphism of the space of
those quaternionic lines in $\mathbb H^{n+1}$, which are isotropic for a
quaternionic skew--Hermitian form. This induces on the Grassmannian
the PCS--structure of quaternionic type from Example 3.4 of
\cite{PCS1}, which is homogeneous under $SU(n+1)$ and can be viewed as
a counterpart of the quaternion--K\"ahler metric on the Grassmannian
of planes.
In the last part \cite{PCS3} of the series, we show how invariant
differential operators for parabolic contact structures induce natural
differential operators on PCS--quotients. There is a well developed
theory which in particular produces invariant differential complexes
on locally flat parabolic contact structures and in certain curved
situations. These can then be used to obtain invariant differential
complexes associated to special symplectic connections as well as
certain more general PCS--structures.
\section{Contactification of PCS--structures}\label{2}
We start this section by briefly recalling the concept of
PCS--structures as introduced in \cite{PCS1} and the class of
parabolic contact structures. Then we describe the relation between
the two kinds of structures, generalizing the analogous results from
\cite{Cap-Salac}, which relate conformally symplectic structures to
contact structures.
\subsection{Contact gradings and associated groups}\label{2.1}
A \textit{contact grading} on a simple Lie algebra $\mathfrak g$ over
$\mathbb K=\mathbb R$ or $\mathbb C$ is a decomposition $\mathfrak g=\mathfrak
g_{-2}\oplus\mathfrak g_{-1}\oplus\mathfrak g_0\oplus\mathfrak g_1\oplus\mathfrak g_2$
which is compatible with the Lie bracket in the sense that $[\mathfrak
g_i,\mathfrak g_i]\subset\mathfrak g_{i+j}$ (putting $\mathfrak g_\ell=\{0\}$ if
$|\ell|>2$), such that $\dim(\mathfrak g_{-2})=1$ and such that the bracket
$\mathfrak g_{-1}\times\mathfrak g_{-1}\to\mathfrak g_{-2}$ is a non--degenerate
bilinear form. It turns out that then the analogous statements hold
for $\mathfrak g_2$ and the bracket $\mathfrak g_1\times\mathfrak g_1\to\mathfrak g_2$.
Such gradings are well known from the theory of quaternionic symmetric
spaces as well as from the theory of parabolic contact
structures. They exist on any complex simple Lie algebra and on most
non--compact real simple Lie algebras (namely on those which contain a
highest root vector) and if they exist, they are always unique up to
isomorphism. The complete list of contact gradings can be found in
Example 3.2.10 of \cite{book}.
Via $\mathfrak g_0\subset\mathfrak g$, which is a Lie subalgebra by the
grading property, contact gradings give rise to geometric structures
in two ways. On the one hand, consider the graded Lie algebra $\mathfrak
g_-:=\mathfrak g_{-2}\oplus\mathfrak g_{-1}$ which, by definition of a contact
grading, is a Heisenberg algebra. By the grading property and the
Jacobi--identity, the adjoint action of any element $A\in\mathfrak g_0$
can be restricted to $\mathfrak g_-$ and defines a derivation on this Lie
algebra which preserves the grading. It is easy to see that the
resulting Lie algebra homomorphism $\mathfrak g_0\to\mathfrak{der}(\mathfrak g_-)$
is always injective.
It is easy to describe $\mathfrak{der}(\mathfrak g_-)$ explicitly. Viewing the
bracket as a linear map $\Lambda^2\mathfrak g_{-1}\to\mathfrak g_{-2}$, its kernel
defines a codimension one--subspace $\Lambda^2_0\mathfrak
g_{-1}\subset\Lambda^2\mathfrak g_{-1}$. Now one easily shows that a linear
map $\phi:\mathfrak g_{-1}\to\mathfrak g_{-1}$ can be extended to a derivation
of $\mathfrak g_{-1}$ if and only if the induced endomorphism of
$\Lambda^2\mathfrak g_{-1}$ preserves the hyperplane $\Lambda^2_0\mathfrak g_{-1}$ and
thus factors to an endomorphism of the quotient which is isomorphic to
$\mathfrak g_{-2}$. The bracket defines a non--degenerate bilinear form on
$\mathfrak g_{-1}$ up to scale and thus a line in $\Lambda^2(\mathfrak g_{-1})^*$,
which is exactly the annihilator of $\Lambda^2_0\mathfrak g_{-1}$. The
stabilizer of this line in $L(\mathfrak g_{-1},\mathfrak g_{-1})$ will be
denoted by $\mathfrak{csp}(\mathfrak g_{-1})$ and called the
\textit{conformally symplectic Lie algebra} of $\mathfrak g_{-1}$. The
above discussion implies that the adjoint action of $\mathfrak g_0$ on
$\mathfrak g_{-1}$ defines an injective homomorphism $\mathfrak
g_0\hookrightarrow\mathfrak{csp}(\mathfrak g_{-1})$.
The case of algebras of type $C_n$ is special with respect to contact
gradings. In this case $\mathfrak g$ itself is a symplectic Lie algebra
and it turns out that the adjoint action actually gives rise to an
isomorphism $\mathfrak g_0\cong \mathfrak{csp}(\mathfrak
g_{-1})\cong\mathfrak{der}_{gr}(\mathfrak g_-)$, so one just obtains the full
conformally symplectic algebra in this case. Therefore, we will have
to exclude the algebras of type $C_n$ from the considerations in this
section. Conformally Fedosov structures and their contactifications,
which will be discussed in Section \ref{3} below, can be thought of as
a $C_n$--analog of the situation discussed in the rest of this
section.
To proceed towards geometric structures determined by a contact
grading, we also need a choice of group associated to the Lie algebra
$\mathfrak g_0$. While the theory of PACS--structures can be developed for
more general Lie groups with Lie algebra $\mathfrak g_0$, we will restrict
the choice of group here to get a complete correspondence to parabolic
contact structures. Namely, let us start with a Lie group $G$ with Lie
algebra $\mathfrak g$. Then it is well known that the Lie subalgebra
$\mathfrak p:=\mathfrak g_0\oplus\mathfrak g_1\oplus\mathfrak g_2\subset\mathfrak g$ is a
parabolic subalgebra, so in particular, the normalizer of $\mathfrak p$ in
$G$ is a Lie subgroup with Lie algebra $\mathfrak p$. Let $P\subset G$ be
a \textit{parabolic subgroup} of $G$ corresponding to $\mathfrak p$,
i.e.~a subgroup lying between this normalizer and its connected
component of the identity. Then $P$ has Lie algebra $\mathfrak p$ and it
is well known that the adjoint action of any element of $P$ maps each
$\mathfrak g_i$ to $\mathfrak g_i\oplus\dots\oplus\mathfrak g_2$, so it preserves
the filtration of $\mathfrak g$ induced by the grading.
Finally, one defines a Lie subgroup $G_0\subset P$ as consisting of
those elements whose adjoint action preserves the grading on $\mathfrak
g$. Let us denote by $CSp(\mathfrak g_{-1})\subset GL(\mathfrak g_{-1})$ the
subgroup of those linear isomorphisms for which the induced
automorphism of $\Lambda^2(\mathfrak g_{-1})^*$ preserves the line determined
by the bracket. Then this can be identified with the group
$\operatorname{Aut}_{gr}(\mathfrak g_-)$ of automorphisms of the graded Lie algebra
$\mathfrak g_-$ and the adjoint action defines an infinitesimally
injective homomorphism $G_0\to CSp(\mathfrak g_{-1})$.
\subsection{PCS--structures}\label{2.2}
Having constructed the infinitesimally injective homomorphism $G_0\to
CSp(\mathfrak g_{-1})\subset GL(\mathfrak g_{-1})$ there is the natural notion
of a first order structure with structure group $G_0$ on smooth
manifolds of dimension $\dim(\mathfrak g_{-1})$. These are the
PACS--structures as introduced in \cite{PCS1}. The simplest way to
describe such a structure is as a principal bundle $p:\mathcal G_0\to M$
with structure group $G_0$, which is endowed with a strictly
horizontal, $G_0$--equivariant one--form $\th\in\Omega^1(\mathcal G_0,\mathfrak
g_{-1})$. Here equivariancy means that for the principal right action
$r^g$ of $g\in G_0$, one has $(r^g)^*\th=\operatorname{Ad}(g^{-1})\o\th$ while
strict horizontality means that in each point $u\in\mathcal G_0$, the
kernel of $\th(u):T_u\mathcal G_0\to\mathfrak g_{-1}$ coincides with the
vertical subbundle of $\mathcal G_0\to M$. In particular, $\th(u)$ descends
to a linear isomorphism $T_{p(u)}M\to\mathfrak g_{-1}$ which identifies
$\mathcal G_0$ as a reduction of the linear frame bundle of $TM$.
Given $(\mathcal G_0\to M,\th)$ as above, a representation of $G_0$ on a
vector space $V$ gives rise to the natural vector bundle $\mathcal
G_0\times_{G_0}V\to M$. Via $\th$, the bundle $\mathcal G_0\times_{G_0}\mathfrak
g_{-1}$ is identified with $TM$. Further, the $G_0$--invariant line in
$\Lambda^2(\mathfrak g_{-1})^*$ gives rise to a line subbundle
$\ell\subset\Lambda^2T^*M$ such that each non--zero element of $\ell$ is
non--degenerate as a bilinear form on the corresponding tangent
space. This is the almost conformally symplectic structure underlying
a PACS--structure. While the theory of PACS--structures is developed
in this general setting in \cite{PCS1}, we will impose a restriction
on such structures throughout this article. Namely, we will only deal
with PCS--structures, i.e.~assume that the underlying structure is
conformally symplectic. This means that the line subbundle $\ell$
admits local sections which are closed as two--forms on $M$.
A central result of \cite{PCS1} is that any PACS--structure on $M$
determines a canonical linear connection on $TM$, whose torsion
satisfies a certain normalization condition. We formulate this result
only for PCS--structures, where the normalization condition on the
torsion is simpler. To formulate this condition, we need the fact that
there a natural $G_0$--invariant subspace
$\ker(\square)\subset\Lambda^2(\mathfrak g_{-1})^*\otimes\mathfrak g_{-1}$, called
the \textit{harmonic subspace}. Here $\square:\Lambda^2(\mathfrak
g_-)^*\otimes\mathfrak g_-\to \Lambda^2(\mathfrak g_-)^*\otimes\mathfrak g_-$ is the
Kostant--Laplacian, see Section 4.3 of \cite{PCS1} for the definition
and Sections 4.5 and 4.6 of this reference as well as Section 4.2 of
\cite{book} for an explicit description of this subspace in several
cases. Via associated bundles, this gives rise to a subbundle
$\ker(\square)\subset\Lambda^2T^*M\otimes TM$, whose elements are called
\textit{algebraically harmonic}. Specialized to the PCS--case,
Corollary 4.4 of \cite{PCS1} shows
\begin{thm*}
Let $(p:\mathcal G_0\to M,\th)$ be a PCS--structure on a smooth manifold
$M$. Then there is a unique connection compatible with this
structure, such that the induced linear connection on $TM$ has
algebraically harmonic torsion.
\end{thm*}
\subsection{Parabolic contact structures}\label{2.3}
There is a second way to obtain a geometric structure from a contact
grading of a real simple Lie algebra $\mathfrak g$. Consider a smooth
manifold $M^\#$ of dimension $\dim(\mathfrak g_-)=\dim(\mathfrak g_{-1})+1$
and suppose that $H\subset TM^\#$ is a smooth distribution of corank
one. Denoting by $Q$ the quotient bundle $TM^\#/H$ (which is a real
line bundle), the Lie bracket of vector fields induces a skew
symmetric bilinear bundle map $\mathcal L:H\times H\to Q$ called the
Levi--bracket. Now $H$ is a contact structure if and only if this
Levi--bracket is non--degenerate in each point $x\in
M^\#$. Equivalently, for each point $x$, the associated graded
$H_x\oplus Q_x$ of the tangent space $T_xM^\#$ endowed with $\mathcal L_x$
as a Lie bracket is isomorphic to $\mathfrak g_-$. Supposing that
$H\subset TM^\#$ is a contact structure, it is then easy to obtain an
adapted frame bundle for the associated graded vector bundle
$\operatorname{gr}(TM^\#)$ to the filtered vector bundle $H\subset TM^\#$ with
structure group $\operatorname{Aut}_{gr}(\mathfrak g_-)\cong CSp(\mathfrak g_{-1})$.
If $\mathfrak g$ is not of type $C_n$, then a parabolic contact structure
of the corresponding type can be simply described as a reduction of
structure group of this adapted frame bundle corresponding to the
infinitesimally injective homomorphism $G_0\to \operatorname{Aut}_{gr}(\mathfrak g_-)$
defined by the adjoint action. Similarly to the case of standard first
order structures, such a reduction can be described as an abstract
principal $G_0$--bundle $p_0^\#:\mathcal G_0^\#\to M^\#$ endowed with an analog
of a soldering form as follows.
Define $T^{-1}\mathcal G_0^\#\subset T\mathcal G_0^\#$ as the preimage of
$H\subset TM^\#$ under the natural projection. Of course, this
contains the vertical subbundle $V\mathcal G_0^\#=\ker(Tp_0^\#)$. Then the
soldering form consists of two components, namely
$\th_{-2}^\#\in\Omega^1(\mathcal G_0^\#,\mathfrak g_{-2})$ and
$\th_{-1}^\#\in\Gamma(L(T^{-1}\mathcal G_0^\#,\mathfrak g_{-1}))$. So while
$\th_{-2}^\#$ is an ordinary differential form, $\th_{-1}^\#$ is only
defined on the subbundle $T^{-1}\mathcal G_0^\#$. Both components are
required to be $G_0$--equivariant (with respect to the adjoint
action), and they should be strictly horizontal in the sense that for
each $u\in\mathcal G_0^\#$ one has $\ker(\th_{-2}^\#(u))=T^{-1}_u\mathcal
G_0^\#$ and $\ker(\th_{-1}^\#(u))=V_u\mathcal G_0^\#$. This implies that
$\th_{-2}^\#(u)$ descends to a linear isomorphism
$Q_{p_0^\#(u)}\to\mathfrak g_{-2}$ while $\th_{-1}^\#(u)$ descends to a
linear isomorphism $H_{p_0^\#(u)}\to\mathfrak g_{-1}$. These isomorphisms
together define an isomorphism of Lie algebras, i.e.~they intertwine
between the Levi--bracket and the Lie bracket on $\mathfrak g_-$.
Parabolic contact structures admit an equivalent uniform description,
which also works in the $C_n$--case. Namely, given a smooth manifold
$M^\#$ as above, one considers a principal bundle $p^\#:\mathcal G^\#\to
M^\#$ with structure group $P$, which is endowed with a Cartan
connection $\omega\in\Omega^1(\mathcal G^\#,\mathfrak g)$. This by definition means
that $\omega$ is $P$--equivariant and reproduces the generators of
fundamental vector fields, and that for each $u\in\mathcal G^\#$ the map
$\omega(u):T_u\mathcal G^\#\to\mathfrak g$ is a linear isomorphism. Moreover, one
has to assume the conditions of regularity and normality on the
curvature of the Cartan connection $\omega$, see Section 3.1 of
\cite{book} for the precise definitions.
The two descriptions are related as follows. Given the Cartan
geometry $(\mathcal G^\#,\omega)$ and a point $u\in\mathcal G^\#$ with
$x=p^\#(u)\in M^\#$, $\omega(u)$ descends to a linear isomorphism
$T_xM^\#\to \mathfrak g/\mathfrak p$. The subset $\mathfrak g_{-1}\oplus\mathfrak
p\subset \mathfrak g$ is $P$--invariant so its preimage gives rise to a well
defined subspace $H_x\subset T_xM^\#$. These subspaces fit together to
a corank one subbundle $H\subset TM^\#$ which defines a contact structure
by regularity of $\omega$. Moreover, the exponential mapping restricts to
a diffeomorphism from $\mathfrak p_+:=\mathfrak g_1\oplus\mathfrak g_2$ onto a
closed normal subgroup $P_+\subset P$, such that $G_0\cong P/P_+$ via
the restriction of the canonical projection. The group $P_+$ acts
freely on $\mathcal G^\#$ via the principal right action, and the quotient
$\mathcal G_0^\#:=\mathcal G^\#/P_+$ naturally is a principal bundle over
$M^\#$ with structure group $P/P_+=G_0$. Finally, one shows that the
component of $\omega$ in $\mathfrak g_{-2}$ descends to
$\th_{-2}^\#\in\Omega^1(\mathcal G_0^\#,\mathfrak g_{-2})$ while an appropriate
restriction of the $\mathfrak g_{-1}$--component of $\omega$ descends to
$\th_{-1}^\#\in\Gamma(L(T^{-1}\mathcal G_0^\#,\mathfrak g_{-1}))$.
In the $C_n$--case, the underlying structure constructed in this way
is just the adapted frame bundle of a contact structure on $M^\#$ and
does not contain any additional information. In all other cases,
however, the Cartan geometry $(\mathcal G^\#,\omega)$ can be reconstructed
from the underlying structure by a prolongation procedure as described
in Section 3.1 of \cite{book}. One forms the obvious extension $\mathcal
G_0^\#\times_{G_0}P$ of structure group and then shows that there is a
normal Cartan connection on this principal bundle inducing the given
soldering form. Further, one proves that any isomorphism of underlying
structures lifts to an isomorphism of normal Cartan geometries, which
establishes an equivalence (in the categorical sense) between the two
pictures.
\subsection{PCS--quotients and parabolic contactifications}\label{2.4}
Recall that for a contact manifold $(M^\#,H)$ an \textit{infinitesimal
contactomorphism} is a vector field $\xi\in\mathfrak X(M^\#)$ such that
for all $\eta\in\Gamma(H)$ we get $[\xi,\eta]\in\Gamma(H)$. An infinitesimal
contactomorphism is called \textit{transversal} if $\xi(x)\notin H_x$
for all $x\in M^\#$. This in particular implies that $\xi$ is nowhere
vanishing and thus defines a foliation of $M^\#$ with one--dimensional
leaves. By a \textit{quotient} of $M^\#$ by $\xi$ one then means a
global space of leaves, i.e.~a surjective submersion $q:M^\#\to M$
with connected fibers such that for each $x\in M^\#$ the vertical
subspace $\ker(T_xq)\subset T_xM^\#$ coincides with the line spanned
by $\xi(x)$.
In this situation, Proposition 2.2 of \cite{Cap-Salac} shows that the
contact structure on $M^\#$ induces a conformally symplectic structure
$\ell\subset\Lambda^2 T^*M$. Denoting by $\alpha$ the unique contact form on
$M^\#$ such that $\alpha(\xi)=1$, there even is a global symplectic form
$\omega$ on $M$, which is a section of $\ell$ and such that
$q^*\omega=d\alpha$. Conversely, given a manifold $M$ endowed with a
conformally symplectic structure $\ell$, one can try to (locally)
realize it as a quotient of a contact manifold $(M^\#,H)$, which then
is called a (local) \textit{contactification} of $(M,\ell)$. Existence
and uniqueness of such contactifications is also studied in
\cite{Cap-Salac}.
We want to develop refinements of these concepts which relate parabolic
contact structures to PCS--structures. Let us start with the regular
normal parabolic geometry $(p:\mathcal G^\#\to M^\#,\omega)$ of type $(G,P)$
corresponding to a parabolic contact structure with contact
distribution $H\subset TM^\#$. An \textit{infinitesimal symmetry} of
the geometry then is a vector field $\tilde\xi\in\mathfrak X(\mathcal G)$
which is $P$--invariant and satisfies $L_{\tilde\xi}\omega=0$, where $L$
denotes the Lie derivative. A $P$--invariant vector field is
automatically projectable, so there is a corresponding vector field
$\xi\in\mathfrak X(M^\#)$ as well as an intermediate $G_0$--invariant
vector field $\xi_0\in\mathfrak X(\mathcal G_0^\#)$. It is easy to verify
directly that $\xi$ is an infinitesimal contactomorphism. As before,
an infinitesimal symmetry is called \textit{transverse} if
$\xi(x)\notin H_x$ for all $x\in M^\#$.
In the case that $\mathfrak g$ is not of type $C_n$, infinitesimal
symmetries can be equivalently described in terms of the underlying
structure $(p_0^\#:\mathcal G_0^\#\to M^\#,\th^\#)$. For a
$G_0$--invariant vector field $\xi_0\in\mathfrak X(\mathcal G_0^\#)$ to be an
infinitesimal symmetry, one first has to require that
$[\xi_0,\eta]\in\Gamma(T^{-1}\mathcal G_0^\#)$ for all $\eta\in\Gamma(T^{-1}\mathcal
G_0^\#)$ or equivalently that the projection $\xi\in\mathfrak
X(M^\#)$ of $\xi_0$ is an infinitesimal contactomorphism. If this is
the case, there is a well defined Lie derivative
$L_{\xi_0}\th^\#_{-1}$ of the partially defined differential form
$\th_{-1}^\#$, and $\xi_0$ is an infinitesimal symmetry if
$L_{\xi_0}\th^\#_i=0$ for $i=-1,-2$.
\begin{definition*}
Consider a parabolic contact structure of type $(G,P)$ with
underlying structure $(p_0^\#:\mathcal G_0^\#\to M^\#,\th^\#)$ and a
transversal infinitesimal symmetry $\xi_0\in\mathfrak X(\mathcal G_0^\#)$ of
this structure. Then a \textit{PCS--quotient} of the parabolic
contact structure is a PCS--structure $(p:\mathcal G_0\to M,\th)$ with
structure group $G_0$ together with a morphism $q_0:\mathcal
G_0^\#\to\mathcal G_0$ of principal bundles such that
\begin{itemize}
\item $q_0$ is a surjective submersion with connected fibers.
\item For each $u^\#\in\mathcal G_0^\#$ the kernel of $T_{u^\#}q_0$ is
spanned by $\xi_0(u^\#)$.
\item The restriction of $q_0^*\th\in\Omega^1(\mathcal G_0^\#,\mathfrak
g_{-1})$ to elements of $T^{-1}\mathcal G_0^\#$ coincides with
$\th^\#_{-1}$.
\end{itemize}
In this situation, $(p_0^\#:\mathcal G_0^\#\to M^\#,\th^\#,\xi_0)$ is
also referred to as a \textit{parabolic contactification} of the
PCS--structure $(p:\mathcal G_0\to M,\th)$.
\end{definition*}
To study these concepts, we need some preliminary
observations. Consider the underlying structure $(p_0^\#:\mathcal
G_0^\#\to M^\#,\th^\#)$ of a parabolic contact structure of type
$(G,P)$, a transverse symmetry $\xi_0\in\mathfrak X(\mathcal G_0^\#)$ of this
structure and the projected vector field $\xi\in\mathfrak X(M^\#)$. Then
both $\xi_0$ and $\xi$ are nowhere vanishing and hence define
foliations with one--dimensional leaves.
\begin{lemma*}
Let $(p_0^\#:\mathcal G_0^\#\to M^\#,\th^\#)$ define a parabolic contact
structure of type $(G,P)$ with $G$ not of type $C_n$, let
$\xi_0\in\mathfrak X(\mathcal G_0^\#)$ be a transverse infinitesimal
symmetry of this structure and $\xi\in\mathfrak X(M^\#)$ the projected
vector field.
(1) If $N\subset\mathcal G_0^\#$ is a leaf of the foliation defined by
$\xi_0$, then $p_0^\#(N)\subset M^\#$ is a leaf of the foliation
defined by $\xi$ and the restriction of $p_0^\#$ to $N$ is a
covering map $N\to p_0^\#(N)$.
(2) Suppose that there is a PCS--structure $(p:\mathcal G_0\to M,\th)$
and a PCS--quotient $q_0:\mathcal G_0^\#\to\mathcal G_0$ by $\xi_0$. Then
the base map $q:M^\#\to M$ of $q_0$ is a quotient by the transverse
infinitesimal contactomorphism $\xi$. Moreover in this case the
coverings from (1) actually have to be diffeomorphisms.
\end{lemma*}
\begin{proof}
(1) Since $\xi_0$ projects to $\xi$ and is transversal to the
vertical subbundle of $p_0^\#$, it is clear that $p_0^\#$ maps small
open subsets of $N$ to integral submanifolds of the distribution
spanned by $\xi$. Hence $p_0^\#(N)$ is a connected immersed integral
manifold for the foliation defined by $\xi$ and thus contained in
some leaf $\underline{N}$ of this foliation. Again by construction,
$p_0^\#|_N:N\to\underline{N}$ has bijective tangent maps in all
points, so it is a local diffeomorphism. In particular,
$p_0^\#(N)\subset\underline{N}$ is open, and we can complete the
proof of (1) by showing that it is also closed, since then
$p_0^\#|_N$ is surjective and thus a covering map.
To see this, observe that $\xi_0$ is $G_0$--equivariant. This means
that whenever the flow of $\xi_0$ in a point is defined up to some
time, then the same is true for all points in the same fiber. Now
suppose that $x\in\underline{N}$ is a point such that each open
neighborhood of $x$ in $\underline{N}$ has non--trivial intersection
with $p_0^\#(N)$. Choose $u\in\mathcal G_0^\#$ lying over $x$, and
consider a local integral curve $c$ for $\xi_0$ through $u$. Then
$p_0^\#\o c$ fills an open subset of $\underline{N}$ and hence
intersects $p_0^\#(N)$. Shifting $c$ by the principal right action
of an appropriate element of $G_0$, we may assume that $c$
intersects $N$, and hence is contained in $N$. Hence $p_0^\#(N)$ is
closed in $\underline{N}$, so $p_0^\#(N)=\underline{N}$.
\medskip
(2) By assumption, we have $p\o q_0=q\o p_0^\#$. Since both $p$ and
$q_0$ are surjective submersions, we see that $q$ is a surjective
submersion. Since $\xi_0$ projects to $\xi$, it is also clear that the
values of $\xi$ lie in $\ker(Tq)$ and since each of these kernels has
to be one--dimensional, it is spanned by the value of $\xi$. Now
consider a fiber $N$ of $q_0$. Then $p_0^\#(N)$ lies in one fiber of
$q$ and using $G_0$--equivariancy of $q_0$, one easily verifies that
$p_0^\#(N)$ is all of this fiber. Hence the fibers of $q$ are
connected, so the first statement is proved. But now the fibers of
$q_0$ and $q$ are closed connected integral submanifolds for the
distributions defined by $\xi_0$ and $\xi$, and hence have to coincide
with the leaves. But since $q_0$ is a morphism of principal bundles,
its restriction to each fiber of $p_0^\#$ is injective, so $p_0^\#$ is
injective on fibers of $q_0$, and the last claim follows.
\end{proof}
\subsection{Existence of PCS--quotients}\label{2.5}
Part (2) of Lemma \ref{2.4} tells us how to naturally phrase the
question of existence of PCS--quotients. We assume that we start with
a parabolic contact structure on $M^\#$ and the underlying vector
field $\xi\in\mathfrak X(M^\#)$ of an infinitesimal symmetry of this
structure. Then the natural question to ask is when a quotient
$q:M^\#\to M$ of $M^\#$ by $\xi$ can be made into a PCS--quotient. We
can now show that the necessary condition from part (2) of Lemma
\ref{2.4} is also sufficient.
\begin{thm*}
Consider a parabolic contact structure of type $(G,P)$ with $G$ not
of type $C_n$ on a smooth manifold $M^\#$ with underlying
$G_0$--bundle $p_0^\#:\mathcal G_0^\#\to M^\#$. Let $\xi_0\in\mathfrak
X(\mathcal G^\#_0)$ be a transverse infinitesimal symmetry of this
structure and let $\xi\in\mathfrak X(M^\#)$ be the underlying
infinitesimal contactomorphism. Suppose that for each leaf
$N\subset\mathcal G_0^\#$ of the foliation defined by $\xi_0$, the
restriction $p_0^\#|_N:N\to p_0^\#(N)$ is a diffeomorphism.
Then for any quotient $q:M^\#\to M$ by $\xi$, there is a canonical
PCS--structure $\mathcal G_0\to M$ on $M$ such that $q$ becomes a
PCS--quotient.
\end{thm*}
\begin{proof}
Let $\sim$ be the equivalence relation on $\mathcal G_0^\#$ defined by
the foliation induced by $\xi_0$, i.e.~two points are equivalent if
they lie in the same leaf. Define $\mathcal G_0:=\mathcal G_0^\#/\sim$, the
set of equivalence classes, and let $q_0:\mathcal G_0^\#\to\mathcal G_0$ be
the canonical map. From the proof of Lemma \ref{2.4}, we know that
$p_0^\#$ maps leaves to leaves, and the leaves in $M^\#$ are the fibers
of $q$, so there is a set map $p_0:\mathcal G_0\to M$ such that $p_0\o
q_0=q\o p_0^\#$.
We claim that $p_0:\mathcal G_0\to M$ is a $G_0$--principal bundle and
$q_0$ is a morphism of principal bundles. To prove this, observe
first that $q\o p_0^\#$ is a surjective submersion. Hence for each
$x\in M$, there is a neighborhood $U$ of $x$ in $M$ and a smooth map
$\tau:U\to\mathcal G_0^\#$ such that $q\o p_0^\#\o\tau=\operatorname{id}_U$. Using
this, we define a map $\psi:U\times G_0\to\mathcal G_0$ by
$\psi(y,g):=q_0(\tau(y)\cdot g)$. This evidently satisfies
$p_0\o\psi=\operatorname{pr}_1$, so it has values in $p_0^{-1}(U)$. If
$\psi(y,g)=\psi(\tilde y,\tilde g)$, then applying $p_0$ we get
$y=\tilde y$. Moreover, $\tau(y)\cdot g$ and $\tau(y)\cdot\tilde g$
lie in the same leaf in $\mathcal G_0^\#$. Since we have assumed that
$p_0^\#$ restricts to an injection on each leaf, we conclude that
$\tau(y)\cdot g=\tau(y)\cdot\tilde g$ and thus $g=\tilde g$. On the
other hand, for $y\in U$, a point $z$ in $p_0^{-1}(U)$ corresponds
to a leaf $N\subset\mathcal G_0^\#$ such that
$p_0^\#(N)=q^{-1}(\{y\})$. Hence there is a point $z^\#\in N$ such
that $p_0^\#(z^\#)=p_0^\#(\tau(y))\in q^{-1}(\{y\})$. But this
implies $z^\#=\tau(y)\cdot g$ for some $g\in G_0$ and hence
$z=\psi(y,g)$, so $\psi:U\times G_0\to p_0^{-1}(U)$ is bijective.
Now suppose that for some open subset $U\subset M$ we have two
sections $\tau$ and $\hat\tau$ as above, and let us consider the
``chart change'' $\phi:U\times G_0\to G_0$, i.e.~the map characterized by
$q_0(\tau(x)\cdot g)=q_0(\hat\tau(x)\cdot\phi(x,g))$ (which evidently
exists). It suffices to do this locally around a fixed point $x$. By
construction $p_0^\#(\tau(x))$ and $p_0^\#(\hat\tau(x))$ both lie in
the leaf $q^{-1}(x)$, so there is a time $t_0\in\mathbb R$ such that
$p_0^\#(\tau(x))=\operatorname{Fl}^{\xi}_{t_0}(p_0^\#(\hat\tau(x)))$. It is easy
to see that the flow $\operatorname{Fl}^{\xi_0}_{t_0}$ is defined in the point
$\hat\tau(x)$. Thus, we may shrink $U$ in such a way that for some
$\epsilon>0$, $\operatorname{Fl}^{\xi_0}_{t_0+t}(\hat\tau(y))$ is defined for all $y\in
U$ if $|t|<\epsilon$, and moreover $(y,t)\mapsto
p_0^\#(\operatorname{Fl}^{\xi_0}_{t_0+t}(\hat\tau(y)))$ is a diffeomorphism from
$U\times (-\epsilon,\epsilon)$ onto an open neighborhood $U^\#$ of
$p_0^\#(\tau(x))$ in $M^\#$. Then there is an open neighborhood $V$
of $x$ in $M$ such that $p_0^\#(\tau(V))\subset U^\#$, and we obtain
a smooth function $\beta:V\to (-\epsilon,\epsilon)$ such that
$p_0^\#(\tau(y))=p_0^\#(\operatorname{Fl}^{\xi_0}_{t_0+\beta(y)}\hat\tau(y))$.
Now observe that
$\tilde\tau(y):=\operatorname{Fl}^{\xi_0}_{t_0+\beta(y)}(\hat\tau(y))$ is a smooth
section of $q\o p_0^\#$ on $V$ which produces the same map $\psi$ as
$\hat\tau$. But now $p_0^\#\o\tilde\tau=p_0^\#\o\tau$, so there is a
smooth map $\gamma:V\to G_0$ such that
$\tilde\tau(y)=\tau(y)\cdot\gamma(y)$, and hence
$q_0(\tilde\tau(y)\cdot g)=q_0(\tau(y)\cdot (\gamma(y)g))$. This shows
that we can endow $\mathcal G_0$ with a topology by requiring that the
maps $\psi$ are homeomorphisms, and then there inverses define
principal bundle charts, so the claim is proved.
In the proof of Proposition 2.2 of \cite{Cap-Salac} it is shown that
the quotient $q:M^\#\to M$ has the property that for each $x\in
M^\#$, the tangent map $T_xq$ restricts to a linear isomorphism
$H_x\to T_{q(x)}M$. Hence for each $u\in\mathcal G_0^*$ the tangent map
$T_uq_0$ restricts to a linear isomorphism $T^{-1}_u\mathcal G_0^\#\to
T_{q_0(u)}\mathcal G_0$, which in addition respects the vertical
subbundles. Composing $\th^\#_{-1}(u)$ with the inverse of this
isomorphism, we obtain a surjective linear map $T_{q_0(u)}\mathcal
G_0\to \mathfrak g_{-1}$ whose kernel is the horizontal subspace
$V_{q_0(u)}\mathcal G_0$. Since $\th_{-1}^\#$ is preserved by the flow
of $\xi_0$, it follows that locally around $u$ points in the same
fiber lead to the same map on $T_{q_0(u)}\mathcal G_0$. Since the fibers
of $q_0$ are connected, we obtain a well defined map
$\th(q_0(u)):T_{q_0(u)}\mathcal G_0\to\mathfrak g_{-1}$ with kernel the
vertical subspace.
To see that this defines a smooth $\mathfrak g_{-1}$--valued one--form
on $\mathcal G_0$, observe that $q_0$ admits local smooth sections
$\sigma$. Moreover, $\ker(Tq_0)\subset T\mathcal G_0^\#$ is a smooth
line--subbundle which is complementary to $T^{-1}\mathcal G_0^\#$, so
there is a corresponding smooth bundle map $\Pi$ projecting $T\mathcal
G_0^\#$ onto the subbundle $T^{-1}\mathcal G_0^\#$. Now for a smooth
vector field $\xi$ on $\mathcal G_0$, and a local section $\sigma$, we can
by construction compute $\th(\xi)$ as
$\th_{-1}^\#(\Pi(T\sigma\o\xi))$, which shows that
$\th\in\Omega^1(\mathcal G_0,\mathfrak g_{-1})$. Finally $\th$ is
strictly horizontal by construction and $q_0^*\th$ restricts to
$\th^\#_{-1}$ on $T^{-1}\mathcal G_0^\#$. Since equivariancy of
$\th$ follows easily from equivariancy of $\th_{-1}^\#$, this
completes the proof.
\end{proof}
\subsection{Examples}\label{2.6}
It is an easy consequence of the Frobenius theorem that given a
contact manifold $M^\#$ and a transversal infinitesimal
contactomorphism $\xi\in\mathfrak X(M^\#)$, there locally exist quotients
by $\xi$ for which the fibers are intervals. Hence Theorem \ref{2.5}
shows that locally any transversal infinitesimal automorphism of a
parabolic contact structure admits PCS--quotients. Likewise, Theorem
\ref{2.5} shows that any global quotient $q:M^\#\to M$ can be made
into a PCS--structure in case that all fibers of $q$ are simply
connected. We now discuss examples which show that in the case of
fibers which are circles, the situation is more subtle. Nonetheless,
we obtain several interesting examples of global contactifications in
which all fibers are circles.
These examples start from the homogeneous models of parabolic contact
structures. So we start with a simple Lie group $G$ and a parabolic
subgroup $P\subset G$ corresponding to a contact grading of the Lie
algebra $\mathfrak g$ of $G$, and consider the homogeneous space $G/P$
with the parabolic contact structure coming from the Maurer--Cartan
form on $G$. The automorphism group of this geometry is $G$, so
looking for quotients, it is natural to consider the actions of
$1$--parameter subgroups of $G$ on $G/P$. We are particularly
interested in finding such actions for which the infinitesimal
generator is transversal everywhere, since these may lead to compact
quotients.
The first example concerns complex projective space $\mathbb CP^n$. The
Fubini--Study metric on $\mathbb CP^n$ is a K\"ahler metric, thus
defining a PCS--structure by Proposition 3.2 of \cite{PCS1}.
\begin{prop*}
Consider $S^{2n+1}$ as the unit sphere in $\mathbb C^{n+1}$ and let
$q:S^{2n+1}\to\mathbb CP^n$ be the Hopf fibration. Then endowing
$S^{2n+1}$ with its usual CR--structure and $\mathbb CP^n$ with the
PCS--structure of type $(PSU(n+1,1),P)$ defined the Fubini--Study
metric, the map $q$ is a global parabolic contactification with
circles as fibers.
\end{prop*}
\begin{proof}
Recall the description of $S^{2n+1}$ as the homogeneous model of
strictly pseudoconvex CR--structures. Consider $V:=\mathbb C^{n+1}\times\mathbb
C$ endowed with the standard Hermitian form of signature $(n+1,1)$,
i.e.~the difference of the standard positive definite forms on the
two factors. Then any non--zero isotropic vector has to have
non--zero first component. Mapping a point $z$ in the unit sphere
$S^{2n+1}\subset\mathbb C^{n+1}$ to the line spanned by $(z,1)$
identifies $S^{2n+1}$ with the space of isotropic complex lines in
$V$. This gives rise to a transitive action of $G:=SU(n+1,1)$ on
$S^{2n+1}$, thus identifying it with $G/P$, where $P$ is the
stabilizer of an isotropic line. It is well known that the resulting
diffeomorphisms of $S^{2n+1}$ are exactly those which preserve the
strictly pseudoconvex CR--structure induced by the embedding
$S^{2n+1}\hookrightarrow\mathbb C^{n+1}$.
Next, the obvious action of $U(1)$ on $S^{2n+1}$ by complex
multiplication can be realized by the action of a $1$--parameter
subgroup in $SU(n+1,1)$. Namely, for $t\in\mathbb R$, one considers
multiplication by $e^{it/(n+2)}$ in the first factor and
multiplication by $e^{-i(n+1)t/(n+2)}$ in the second factor. This
evidently defines a unitary map of determinant one and the point
$(e^{it/(n+2)}z,e^{-i(n+1)t/(n+2)})$ determines the same complex line
as $(e^{it}z,1)$. From this it is clear, that the infinitesimal
generator of this group is simply multiplication by $i$, which maps
any point into its real orthocomplement. However, $iz$ never lies in
the complex orthocomplement of $z$, which defines the CR
subspace. Hence we conclude that the infinitesimal automorphism
generating the one--parameter group is transversal everywhere, and
by definition the Hopf fibration $q$ is a global space of leaves for
this foliation.
By construction, the $U(1)$--orbits of the action on $SU(n+1,1)$ are
$(n+2)$--fold coverings of the $U(1)$ orbits in $S^{2n+1}$, since the
projection sends $\lambda$ to multiplication by $\lambda^{n+2}$. The
underlying $G_0$--bundle in this case is simply $G/P_+\to G/P$, where
$P_+\subset P$ is the subgroup introduced in Section \ref{2.3}. It is
well known that its Lie algebra $\mathfrak p_+$ consists of maps vanishing
on the distinguished complex line in $\mathbb C^{n+2}$, so any element of
$P_+$ acts as the identity on this line. This shows that $P_+$
intersects our one--parameter subgroup only in the identity, so the
projection $G\to G/P_+$ restricts to a diffeomorphism on each
orbit. Thus Lemma \ref{2.4} shows that $q:S^{2n+1}\to\mathbb CP^n$ cannot
be globally made into a PCS--quotient for the parabolic contact
structure of type $(G,P)$.
This is related to the fact that $G$ does not act effectively on
$G/P$, and correspondingly $G_0$ does not act effectively on $\mathbb
C^n$ (it is an $(n+2)$--fold covering of the conformal unitary group
$CU(n)$). This problem can be resolved by replacing $G$ by the
projective group $\underline{G}:=PSU(n+1,1)$ and $P$ by its image
$\underline{P}$ in $\underline{G}$. Then
$\underline{G}/\underline{P}=G/P$ and now the action becomes
effective. Moreover, $\underline{G}$ is the quotient of $G$ by its
center, which consists of the $(n+2)$nd roots of unity times the
identity map. This shows that the projection $G\to\underline{G}$
restricts to an $(n+2)$--fold covering on $U(1)$--orbits which exactly
identifies the different preimages of points in $S^{2n+1}$. Hence
Theorem \ref{2.5} implies that $q:S^{2n+1}\to\mathbb CP^n$ is globally a
PCS--quotient of geometries of type $(\underline{G},\underline{P})$,
for which $\underline{G}_0=CU(n+1)$.
\end{proof}
Our second example involves a type of parabolic contact structures
which has not been studied intensively in the literature, namely the
one associated to the groups $SO^*(2n)$. To formulate the necessary
background, consider a finite--dimensional right quaternionic vector
space $V$. Then a \textit{quaternionically skew Hermitian form} is a
map $\tau:V\times V\to\mathbb H$, which is bilinear over $\mathbb R$ and
satisfies $\tau(v,wq)=\tau(v,w)q$ and
$\tau(w,v)=-\overline{\tau(v,w)}$ for all $v,w\in V$ and $q\in\mathbb
H$. The following result is well known, we include a proof for
completeness.
\begin{lemma*}
On any finite dimensional quaternionic vector space, there exists a
non--degenerate quaternionically skew Hermitian form, which is unique
up to isomorphism.
\end{lemma*}
\begin{proof}
For a quaternionically skew Hermitian form $\tau$ on $\mathbb H$, one
must have $\tau(p,q)=\overline{p}\tau(1,1)q$ and
$\tau(1,1)\in\operatorname{im}(\mathbb H)$ by skew symmetry. Conversely for each
choice of $\tau(1,1)\in\operatorname{im}(\mathbb H)$ the above formula is immediately
seen to define such a form. Now $GL(1,\mathbb H)$ can be identified
with $\mathbb H\setminus\{0\}$ acting by multiplication from the
left. Acting by $0\neq a\in\mathbb H$ on $\tau$, we obtain
$(a\cdot\tau)(1,1)=\tau(a,a)=\overline{a}\tau(1,1)a$. It is well
known that the action of $\mathbb H\setminus\{0\}$ by conjugation is
transitive on $\operatorname{im}(H)\setminus\{0\}$, which completes the proof for
$\dim(V)=1$.
Similarly, we can define a quaternionically skew Hermitian form on
$\mathbb H^n$ by $\tau((p_1,\dots,p_n),(q_1,\dots,q_n)):=\sum_\ell
\overline{p_\ell}jq_\ell$. Conversely, given any form $\tau$,
non--de\-ge\-ne\-ra\-cy easily implies that there is a vector $v\in V$ such
that $\tau(v,v)\neq 0$, and multiplying by an appropriate scalar, we
can assume $\tau(v,v)=j$. But then the space $\{w:\tau(v,w)=0\}$ is
immediately seen to be a quaternionic subspace of codimension one, and
inductively we conclude that we can construct a basis
$\{v_1,\dots,v_n\}$ of $V$ over $\mathbb H$ such that
$\tau(v_a,v_b)=j\delta_{ab}$, which completes the proof.
\end{proof}
The group $SO^*(2n)$ is defined as the group of all
quaternionically linear automorphisms of $\mathbb H^n$ which preserve a
non--degenerate quaternionically skew Hermitian form. It is easy to
see that the form $\tau$ can be recovered from its real part
$\tau_{\mathbb R}$, so preserving $\tau$ is equivalent to preserving
$\tau_{\mathbb R}$. Of course, $\tau_{\mathbb R}$ is just a skew symmetric
bilinear form $\mathbb H^n\times\mathbb H^n\to\mathbb R$ which satisfies
$\tau_{\mathbb R}(va,wa)=\tau_{\mathbb R}(v,w)$ for any $a=i,j,k$ or
equivalently for any unit quaternion $a$.
For $n\geq 3$ consider the quaternionic projective space $\mathbb HP^n$
of one--dimensional quaternionic subspaces in $\mathbb H^{n+1}$. The
standard representation of $GL(n+1,\mathbb H)$ induces a transitive
action on $\mathbb HP^n$, which can be restricted to
$G:=SO^*(2n+2)$. However, this subgroup does not act transitively,
since the restriction of $\tau$ to a one--dimensional subspace can
either be zero or non--degenerate. It is easy to see that $G$ acts
transitively on the spaces of non--degenerate lines, which form an
open subspace in $\mathbb HP^n$, and on the space $\mathcal N$ of isotropic
lines, which is a closed subspace of $\mathbb HP^n$. This identifies
$\mathcal N$ with $G/P$, where $P\subset G$ is the stabilizer of an
isotropic line in $\mathbb H^{n+1}$.
For a quaternionic line $\ell\subset\mathbb H^{n+1}$, we can realize
$T_\ell\mathbb HP^n$ as $\mathbb H^{n+1}/\ell$. Differentiating the defining
equation $\tau(v,v)=0$, we see that $\mathcal N\subset\mathbb HP^n$ is a
smooth submanifold and that $T_\ell\mathcal N=\{w\in
V:\operatorname{im}(\tau(v,w))=0\}/\ell$, so in particular this submanifold has real
codimension three. Moreover, there is a natural subspace
$H_\ell:=\{w:\tau(v,w)=0\}/\ell\subset T_\ell\mathcal N$, which has real
codimension one. Taking an explicit realization of $G$, it is easy to
verify that the codimension subbundle $H\subset T\mathcal N$ constructed
above defines a parabolic contact structure, consisting of a
quaternionic structure on $H$ such that the Levi--bracket is the real
part of a quaternionically skew Hermitian form. One further verifies
directly that the resulting subgroup $G_0$ is a two--fold covering of
the group $CSO^*(2n)$ generated by $SO^*(2n)$ and real multiples of the
identity.
On the other hand, consider the complex Grassmannian $Gr(2,\mathbb
C^{n+1})$ of two planes as a homogeneous space of $SU(n+1)$. Then it
is well known that this space admits an invariant complex structure as
well as an invariant quaternionic structure, and an invariant
Riemannian metric which is K\"ahler respectively quaternion--K\"ahler
with respect to these structures. In Corollary 3.4 of \cite{PCS1} we
have shown that the K\"ahler form of the K\"ahler metric together with
the quaternionic structure defines an $SU(n+1)$--invariant
PCS--structure of quaternionic type on $Gr(2,\mathbb C^{n+1})$.
\begin{thm*}
Let $\mathcal N\subset\mathbb HP^n$ be the space of quaternionic lines
which are isotropic for a quaternionically skew Hermitian form. Then
there is a projection $q:\mathcal N\to Gr(2,\mathbb C^{n+1})$, which
defines a global parabolic contactification of the
$SU(n+1)$--invariant PCS--structure of type $(PSO^*(2n),P)$ on the
Grassmannian with circles as fibers.
\end{thm*}
\begin{proof}
Let $\langle\ ,\ \rangle$ be the standard positive definite
quaternionically Hermitian form on $V:=\mathbb H^{n+1}$. Fix a
quaternionically linear map $\mathcal J:V\to V$ such that $\mathcal J\o\mathcal
J=-\operatorname{id}$ and such that $\langle \mathcal J(v),\mathcal J(w)\rangle=\langle
v,w\rangle$. (For example, one can take multiplication by $i$ from
the left.) Then it follows immediately that $\tau(v,w):=\langle
v,\mathcal J(w)\rangle$ is a non--degenerate quaternionically skew
Hermitian form, so we can use this form to realize $SO^*(2n+2)$. A
quaternionic line $\ell\subset V$ is isotropic for $\tau$ if and
only if $\mathcal J(\ell)\subset \ell^\perp$, the orthocomplement of $\ell$
with respect to $\langle\ ,\ \rangle$.
Scalar multiplication with respect to the complex structure on $V$
defined by $\mathcal J$ defines an action of $U(1)$ on $V$ by
quaternionically linear maps. From the construction of $\tau$, one
readily verifies that this action is orthogonal for $\tau$, so we have
found a subgroup of $G$ isomorphic to $U(1)$. By definition,
$\tau(v,v)=0$ means that $\langle v,\mathcal J(v)\rangle=0$, so $\mathcal
J(v)$ does not lie in the quaternionic span of $v$. This shows that
only $\pm 1\in U(1)$ act as the identity on $G/P$. Moreover, for
$v\in\ell$ we have $\tau(v,\mathcal J(v))=-\langle v,v\rangle\neq 0$, so
the infinitesimal automorphism generating this subgroup is transversal
on all of $G/P$.
Given an isotropic line $\ell\in\mathcal N$, we conclude from $\ell\cap
\mathcal J(\ell)=\{0\}$ that $\ell\oplus\mathcal J(\ell)$ is quaternionic
subspace in $V$ of quaternionic dimension two, which in addition is
invariant under $\mathcal J$. The $U(1)$--orbit of $\ell$ by construction
consists of null lines which are contained in this plane. From above,
we know that the space of null lines in a quaternionic plane is a
codimension three subspace in $\mathbb HP^1\cong S^4$, so it has
dimension one, and it is easily verified to be connected. This shows
that the $U(1)$--orbit of $\ell$ coincides with the space of those
isotropic lines which are contained in $\ell\oplus\mathcal J(\ell)$.
Hence we see that mapping $\ell$ to $\ell\oplus\mathcal J(\ell)$ defines a
(evidently smooth) map $q$ from $G/P$ to the space of $\mathcal
J$--invariant quaternionic planes in $V$, whose fibers are exactly the
orbits of the $U(1)$--action constructed above. On the other hand,
since $\tau(v,\mathcal J(v))\neq 0$ for any $v\in V$, it follows that the
restriction of $\tau$ to any $\mathcal J$--invariant quaternionic plane
$W\subset V$ is non--degenerate. In particular, any such plane
contains a family of quaternionic null--lines isomorphic to $U(1)$, so
$q$ is surjective.
So it remains to show that the space of $\mathcal J$--invariant
quaternionic planes in $V$ is isomorphic to the complex
Grassmannian. Choose two anti--commuting imaginary unit quaternions,
for example $i$ and $j$. Then right multiplication by $i$ defines a
linear map $V\to V$ which squares to $-\operatorname{id}$ and commutes with $\mathcal
J$. Hence we obtain a decomposition $V=V^{(1,0)}\oplus V^{(0,1)}$ into
two $\mathcal J$--invariant summands characterized by $v\cdot i=\mathcal J(v)$
for $v\in V^{(1,0)}$ and $v\cdot i=-\mathcal J(v)$ for $v\in
V^{(0,1)}$. Moreover, one immediately verifies that multiplication by
$j$ from the right maps $V^{(1,0)}$ to $V^{(0,1)}$ and vice versa, so
both spaces must have complex dimension $n+1$. The projections onto
$V^{(1,0)}$ and $V^{(0,1)}$ can be constructed from $\mathcal J$ and right
multiplication by $i$. A $\mathcal J$--invariant quaternionic subspace
$W\subset V$ thus is invariant under the projections, and hence
$W=(W\cap V^{(1,0)})\oplus (W\cap V^{(1,0)})\cdot j$. Hence we see
that we can identify the space of $\mathcal J$--invariant quaternionic
planes in $V$ with the space of complex planes in $V^{(1,0)}$ and
hence with the Grassmannian $Gr(2,\mathbb C^{n+1})$.
Hence we see that there is a natural projection $q:\mathcal N\to Gr(2,\mathbb
C^{n+1})$, which is a quotient by a transversal infinitesimal
contactomorphism. The action of $G$ on $\mathcal N$ does not descend to
the Grassmannian, but one can consider the stabilizer of $\mathcal J$ in
$G$. It is easy to see that this stabilizer is isomorphic with
$SU(n+1)$ via the action on $V^{(1,0)}$. The conformally symplectic
structure on the Grassmannian is $SU(n+1)$ invariant and coincides
with the one used in Example 3.4 and Corollary 3.4 of \cite{PCS1}.
Now as in the first example, the projection $G\to G/P$ induces a
covering on each $U(1)$--orbit but this time, the covering is just
$2$--fold. Hence $q$ can be made into a PCS--quotient by passing to
the quotient of $G$ by $\mathbb Z_2=\{\pm\mathbb I\}$, and the resulting
PCS--structure of quaternionic type on the Grassmannian is exactly the
one from Corollary 3.4 of \cite{PCS1}.
\end{proof}
\subsection{Contactification of PCS--structures}\label{2.7}
For the applications to BGG sequences, we will mainly need the
following counterpart to Theorem \ref{2.5}. We show that any local
contactification of the conformally symplectic structure underlying a
PCS--structure can be canonically made into a PCS--quotient. Together
with the result in Lemma 3.1 of \cite{Cap-Salac}, this shows that any
PCS--structure can be locally realized as a PCS--quotient.
\begin{thm*}
Suppose that $(M^\#,H)$ is a contact manifold, $(M,\ell)$ is a
conformally symplectic manifold and $q:M^\#\to M$ is a reduction by
a transverse infinitesimal contactomorphism $\xi\in\mathfrak X(M^\#)$.
Then any PCS--structure on $M$, which has $\ell$ as its underlying
conformally symplectic structure canonically lifts to a parabolic
contact structure on $M^\#$ for which $\xi$ is an infinitesimal
automorphism.
\end{thm*}
\begin{proof}
As discussed in \ref{2.2}, a PCS--structure on $M$ corresponding to
the group $G_0\subset\ CSp(\mathfrak g_{-1})$ is given by a principal
$G_0$--bundle $p:\mathcal G_0\to M$ together with a one--form
$\th\in\Omega^1(\mathcal G_0,\mathfrak g_{-1})$ which is $G_0$--equivariant and
strictly horizontal. Now we can simply form the pullback $\mathcal
G_0^\#:=q^*\mathcal G_0\to M^\#$, which is a principal $G_0$--bundle
over $M^\#$. Explicitly, $\mathcal G_0^\#=\{(u,x^\#)\in\mathcal G_0\times
M^\#:p(u)=q(x^\#)\}$, and the principal right action on this bundle
is given by the principal right action of $\mathcal G_0$ acting on the
first factor. Pulling back $\th$ along the projection onto the first
factor, we obtain a smooth $\mathfrak g_{-1}$--valued one--form
$\th^\#_{-1}$ on $\mathcal G_0^\#$, which evidently is
$G_0$--equivariant. From the construction it is clear that, for each
$x^\#\in M^\#$, $\th_{-1}^\#(u,x^\#)$ descends to a linear map
$T_{x^\#}M^\#\to\mathfrak g_{-1}$, which restricts to a linear
isomorphism on the contact subspace $H_{x^\#}$.
Next, there is a unique contact form $\alpha$ on $M^\#$ such that
$\alpha(\xi)\equiv 1$ and by Proposition 2.2 of \cite{Cap-Salac} there
is a unique symplectic form $\omega$ on $M$ which is a section of
$\ell$ and satisfies $q^*\omega=d\alpha$. By assumption, the conformally
symplectic structure induced by the reduction $p:\mathcal G_0\to M$ of
structure group is $\ell\subset\Lambda^2T^*M$. Given a point $u\in\mathcal
G_0$ with $p(u)=x$, we have $\omega(x)\in\ell_x$ so there is a linear
isomorphism $\psi(u):\mathfrak g_{-2}\to\mathbb R$ such that, viewing
$\th(u)$ as a map $T_xM\to\mathfrak g_{-1}$, we get
$$
\omega(x)(\eta_1,\eta_2)=\psi(u)([\th(u)(\eta_1),\th(u)(\eta_2)])
\quad\forall \eta_1,\eta_2\in T_xM,
$$ where the Lie bracket is in $\mathfrak g_{-}$. This defines a smooth
map $\psi:\mathcal G_0\to L(\mathfrak g_{-2},\mathbb R)$ and equivariancy of $\th$
implies that $\th(u\cdot g)(\eta)=\operatorname{Ad}(g^{-1})(\th(u)(\eta))$, and
hence $\psi(u\cdot g)=\psi(u)\o\operatorname{Ad}(g)$.
Using this, we now define $\th_{-2}^\#\in\Omega^1(\mathcal G_0^\#,\mathfrak
g_{-2})$ by $\th_{-2}^\#(u,x^\#)=\psi(u)^{-1}\o (p^\#)^*\alpha$, where
$p^\#:\mathcal G_0^\#\to M^\#$ is the canonical projection. This evidently
is a smooth one--form, and since $(p^\#)^*\alpha$ is invariant under the
principal right action, we see that $\th^\#_{-2}$ is
$G_0$--equivariant. Next, the kernel of $\th^\#_{-2}$ in a point
coincides with the kernel of $(p^\#)^*\alpha$ in that point and thus with
the pre--image $(Tp^\#)^{-1}(H)$. Finally, if $\tilde\eta_1$ and
$\tilde\eta_2$ are two sections of this subbundle, then we compute
\begin{align*}
d\th_{-2}^\#(u,x^\#)(\tilde\eta_1,\tilde\eta_2)&=
-\th_{-2}^\#(u,x^\#)([\tilde\eta_1,\tilde\eta_2])=
-\psi(u)^{-1}((p^\#)^*\alpha(u,x^\#)([\tilde\eta_1,\tilde\eta_2])\\
&=\psi(u)^{-1}(d(p^\#)^*\alpha(u,x^\#)(\tilde\eta_1,\tilde\eta_2)).
\end{align*}
Now since $d(p^\#)^*\alpha=(p^\#)^*d\alpha=(p^\#)^*q^*\omega$, we get, putting
$x=q(x^\#)$ and denoting by $\eta_i\in T_xM$ the value of
$\tilde\eta_i(u,x^\#)$ under the natural
projection for $i=1,2$:
\begin{align*}
d\th_{-2}^\#(u,x^\#)(\tilde\eta_1,\tilde\eta_2)&=
\psi(u)^{-1}(\omega(x)(\eta_1,\eta_2))=
[\th(u)(\eta_1),\th(u)(\eta_2)]\\
&=[\th^\#_{-1}(u,x^\#)(\tilde\eta_1),\th^\#_{-1}(u,x^\#)(\tilde\eta_2)],
\end{align*}
where the Lie bracket is in $\mathfrak g_-$. Together with the above
properties, this shows that $\th^\#:=(\th^\#_{-2},\th^\#_{-1})$ makes
$p^\#:\mathcal G_0^\#\to M^\#$ into a regular infinitesimal flag structure
on $M^\#$, see Section \ref{2.3}. Since $\mathfrak g$ is not of type
$C_n$, this is equivalent to a parabolic contact structure of type
$(G,P)$.
So it remains to show that $\xi$ is an infinitesimal automorphism of
this parabolic contact structure or equivalently of the infinitesimal
flag structure. Observe first that $T_{(u,x^\#)}\mathcal
G_0^\#=\{(\eta_1,\eta_2)\in T_u\mathcal G_0\times
T_{x^\#}M^\#:T_up\cdot\eta_1=T_{x^\#}q\cdot\eta_2\}$, and hence
$\tilde\xi:=(0,\xi)$ is a well--defined vector field on $\mathcal
G_0^\#$. The flow of this vector field is clearly given by
$\operatorname{Fl}^{\tilde\xi}_t(u,x^\#)=(u,\operatorname{Fl}^\xi_t(x^\#))$ and it suffices to
prove that this flow preserves $\th^\#$.
Take a point $(u,x^\#)\in\mathcal G_0^\#$ and a tangent vector
$(\eta_1,\eta_2)\in T_{(u,x^\#)}\mathcal G_0^\#$. Then
$$
T_{(u,x^\#)}\operatorname{Fl}^{\tilde\xi}_t\cdot
(\eta_1,\eta_2)=(\eta_1,T_{x^\#}\operatorname{Fl}^\xi_t\cdot\eta_2),
$$
wherever the right hand side is defined. But by definition
$$
\th^\#_{-1}(u,x^\#)(\eta_1,\eta_2)=\th(u)(\eta_1),
$$
which immediately implies
$(\operatorname{Fl}^{\tilde\xi}_t)^*\th^\#_{-1}=\th^\#_{-1}$, whenever the flow is
defined. Again by definition, we get
$$
\th_{-2}^\#(u,x^\#)(\eta_1,\eta_2)=\psi(u)^{-1}(\alpha(x^\#)(\eta_2)),
$$
and hence $(\operatorname{Fl}^{\tilde\xi}_t)^*\th^\#_{-2}=\th^\#_{-2}$ follows
immediately from $(\operatorname{Fl}^{\xi}_t)^*\alpha=\alpha$.
\end{proof}
\subsection{Local uniqueness of PCS--contactifications}\label{2.8}
To complete the picture, we have to prove that different realizations
of PCS--structures as PCS--quotients as constructed in Theorem
\ref{2.7} are locally compatible. The corresponding result for
conformally symplectic structures has been proved in Proposition 3.1
of \cite{Cap-Salac}, so we only have to prove compatibility with the
additional structures.
\begin{thm*}
Suppose that $M$ and $\tilde M$ carry a PCS--structure of some fixed
type with underlying conformally symplectic structures $\ell$ and
$\tilde\ell$. Suppose further that $(M^\#,H)$ and $(\tilde
M^\#,\tilde H)$ are contact manifolds and that $q:M^\#\to M$ and
$\tilde q:\tilde M^\#\to\tilde M$ are reductions with respect to
transverse symmetries $\xi\in\mathfrak X(M^\#)$ and $\tilde\xi\in\mathfrak
X(\tilde M^\#)$. Suppose finally that $\phi:M\to\tilde M$ is a
PCS--diffeomorphism and that $\phi^\#:M^\#\to\tilde M^\#$ is a
contactomorphism which lifts $\phi$.
Then $(\phi^\#)^*\tilde\xi=\lambda\xi$ for a nowhere--vanishing, locally
constant function $\lambda$ and $\phi^\#$ is an automorphism of the lifted
parabolic contact structures from Theorem \ref{2.7}.
\end{thm*}
\begin{proof}
Since $\phi^\#$ is a contactomorphism, $(\phi^\#)^*\tilde\xi\in\mathfrak
X(M^\#)$ is a transverse infintesimal automorphism of the contact
structure $H\subset TM^\#$. On the other hand, since $\phi^\#$ lifts $\phi$ it
has to map fibers of $q$ to fibers of $\tilde q$, which shows that
$(\phi^\#)^*\tilde\xi=\lambda\xi$ for some smooth function $\lambda$ on
$M^\#$. But then for $\eta\in\mathfrak X(M^\#)$ we get $[\lambda
\xi,\eta]=\lambda[\xi,\eta]-(\eta\cdot\lambda)\xi$, and if $\eta\in\Gamma(H)$,
then by assumption both the left hand side and the first summand in
the right hand side are sections of $H$. Since $\xi$ is transverse,
this implies that $\eta\cdot \lambda=0$ for any $\eta\in\Gamma(H)$ and since
$H$ is bracket--generating this shows that $\lambda$ is locally constant.
Now let $(p:\mathcal G_0\to M,\th)$ and $(\tilde p:{\tilde{\mathcal G}}_0\to\tilde
M,\tilde\th)$ be the bundles defining the PCS--structures. Then by
assumption there is an isomorphism $\Phi:\mathcal G_0\to{\tilde{\mathcal G}}_0$ of
principal $G_0$--bundles such that $\tilde p\o\Phi=\phi\o p$ and such
that $\Phi^*\tilde\th=\th$. Forming the pullback bundles as in the
proof of Theorem \ref{2.7}, we see that
$$
(u,x^\#)\mapsto (\Phi(u),\phi^\#(x^\#))
$$
defines an isomorphism $(\Phi,\phi^\#):\mathcal G_0^\#\to{\tilde{\mathcal G}}_0^\#$ lifting
$\phi^\#$, so it remains to prove that this is compatible with the
frame forms on the two bundles constructed in the proof of Theorem
\ref{2.7}. Inserting the definitions, one immediately verifies that
$$
(\Phi,\phi^\#)^*\tilde\th_{-1}^\#(u,x^\#)(\eta_1,\eta_2)=
\Phi^*\tilde\th(u)(\eta_1)=\th(u)(\eta_1),
$$
and thus $(\Phi,\phi^\#)^*\tilde\th_{-1}^\#=\th_{-1}^\#$. On the other
hand, since $(\phi^\#)^*\tilde\xi=\lambda\xi$, we see that the contact
forms corresponding to $\xi$ and $\tilde\xi$ are related by
$(\phi^\#)^*\tilde\alpha=\frac1{\lambda}\alpha$. Since $\frac1\lambda$ is locally
constant, the corresponding symplectic forms are related by
$\phi^*\tilde\omega=\frac1{\lambda}\omega$. Viewing the values of $\th$ and
$\tilde\th$ as linear isomorphisms on tangent spaces of $M$ and
$\tilde M$, the fact that $\Phi^*\tilde\th=\th$ reads as
$$
\tilde\th(\Phi(u))(T_{p(u)}\phi\cdot\eta)=\th(u)(\eta)
$$
for all $u\in\mathcal G_0$ and $\eta\in T_{p(u)}M$. Together with
$\phi^*\tilde\omega=\frac1{\lambda}\omega$, this implies that the isomorphisms
$\psi$ and $\tilde\psi$ as constructed in the proof of Theorem \ref{2.7}
satisfy $\tilde \psi(\Phi(u))=\tfrac1\lambda\psi(u)$. Using this, we compute
\begin{align*}
(\Phi,\phi^\#)^*\tilde\th_{-2}^\#&(u,x^\#)(\eta_1,\eta_2)=
\tilde\th_{-2}^\#(\Phi(u),\phi^\#(x^\#))(T_u\Phi\cdot\eta_1,
T_{x^\#}\phi^\#\cdot\eta_2)\\ &=\tilde\psi(\Phi(u))^{-1}((\phi^\#)^*\tilde\alpha(x^\#)(\eta_2))=\lambda\psi(u)^{-1}(\tfrac1\lambda\alpha(x^\#)(\eta_2)),
\end{align*}
so $(\Phi,\phi^\#)^*\tilde\th_{-2}^\#=\th_{-2}^\#$.
\end{proof}
\subsection{Distinguished connections and
contactifications}\label{2.9}
As shown in \cite{PCS1}, any PACS--structure on $M$ determines a
canonical compatible linear connection on $TM$. In the case of
PCS--structures, these are closely related to special symplectic
connections as studied in \cite{Cahen-Schwachhoefer}, see Section 4.7
of \cite{PCS1} for a discussion of the relation. We now conclude the
discussion of parabolic contactifications by relating the
distinguished connection associated to a PCS--quotient of a parabolic
contact structure to distinguished connections ``upstairs''. On the
one hand, this provides alternative proofs for and extensions of some
results from \cite{Cahen-Schwachhoefer}, on the other hand, the result
will be useful for the discussion of conformally Fedosov structures
in Section \ref{3} below as well as for the third part of this series.
A family of distinguished connections associated to a parabolic
contact structure is obtained via so--called Weyl structures. Any Weyl
structure determines a Weyl connection, which is a principal
connection on the bundle $\mathcal G_0^\#\to M^\#$. Equivalently, this can
be described as a linear connection on the vector bundle $H\to M^\#$
which is compatible with the reduction of structure group defined by
$\mathcal G_0^\#$. In addition, a Weyl structure defines an isomorphism
between $TM^\#$ and its associated graded vector bundle $H\oplus
T^*M/H$.
For the purpose of the current article, we only need a few facts on
Weyl structures, which are contained in Section 5.2.11 of
\cite{book}. First, the line bundle $Q:=TM^\#/H$ can be used as a
so--called bundle of scales. In particular, this means that a globally
non--vanishing section of this line bundle determines a unique
Weyl--structure, which is characterized by the fact that this section
is parallel for the induced connection.
In particular, in the situation of a reduction $q:M^\#\to M$ by a
transverse symmetry $\xi\in\mathfrak X(M^\#)$ we can project $\xi$ to a
(by construction nowhere--vanishing) section of $Q$ and use this to
determine a unique Weyl structure. This will be called the
\textit{Weyl structure associated to $\xi$}.
\begin{thm*}
Let $q_0:\mathcal G_0^\#\to\mathcal G_0$ be a PCS--quotient with respect to
a transverse infinitesimal automorphism $\xi_0\in\mathfrak X(\mathcal
G_0^\#)$ covering $q:M^\#\to M$.
Then there is a unique principal connection form $\gamma^\#$ on $\mathcal
G_0^\#\to M^\#$ for which $\xi_0\in\mathfrak X(\mathcal G_0^\#)$ is
horizontal and which coincides with the principal connection form
determined by the Weyl structure associated to $\xi$ on $T^{-1}\mathcal
G_0^\#\subset T\mathcal G_0^\#$. This connection form descends to a
principal connection form on $\mathcal G_0\to M$ which corresponds to
the canonical connection associated to the PCS structure from
Theorem \ref{2.2}.
\end{thm*}
\begin{proof}
Observe first that $\xi_0$ spans $\ker(Tq_0)$, so any one--form on
$\mathcal G_0^\#$ is determined by its value on $\xi_0$ and its
restriction to $T^{-1}\mathcal G_0^\#$. Now let $\tilde\gamma^\#$ be the
principal connection on $\mathcal G_0^\#$ determined by the Weyl
structure associated to $\xi$. Since $\xi_0$ is $G_0$--invariant,
the function $\tilde\gamma^\#(\xi_0):\mathcal G_0^\#\to\mathfrak g_0$ is
$G_0$--equivariant. Next let $\alpha\in\Omega^1(M^\#)$ be the unique
contact form with $\alpha(\xi)=1$ and put
$\gamma^\#:=\tilde\gamma^\#-\tilde\gamma^\#(\xi_0)(p_0^\#)^*\alpha$. This is
evidently a $G_0$--equivariant one--form with values in $\mathfrak
g_0$. Since $(p_0^\#)^*\alpha$ vanishes on $T^{-1}\mathcal G_0^\#$, it is a
principal connection form, it coincides with $\tilde\gamma^\#$ on
$T^{-1}\mathcal G_0^\#$, and $\gamma^\#(\xi_0)=0$.
By definition, the fibers of $q_0:\mathcal G_0^\#\to\mathcal G_0$ are
connected and can locally be represented as flow lines of
$\xi_0$. Hence by Corollary 2.3 in \cite{BCG3} we can prove that
$\gamma^\#$ descends to $\mathcal G_0$ by showing that it is invariant
under the flow of $\xi_0$.
Let $(\mathcal G^\#\to M^\#,\omega)$ be the Cartan geometry of type $(G,P)$
determined by the parabolic contact structure. Then a Weyl structure
is a $G_0$--equivariant section $\sigma$ of the natural projection
$\mathcal G^\#\to\mathcal G_0^\#$. The corresponding Weyl connection is
$\tilde\gamma^\#=\sigma^*\omega_0$, where $\omega_0\in\Omega^1(\mathcal G,\mathfrak g_0)$
is the component in $\mathfrak g_0\subset\mathfrak g$ of the Cartan
connection $\omega$. As discussed in Section \ref{2.3}, the
infinitesimal automorphism $\xi_0$ uniquely lifts to a
$P$--invariant vector field $\tilde\xi\in\mathfrak X(\mathcal G)$ such that
$L_{\tilde\xi}\omega=0$, with $L$ denoting the Lie derivative. Hence
there is a natural action of the local flow of $\tilde\xi$ on the
space of Weyl--structures via $\sigma\mapsto
\operatorname{Fl}^{\tilde\xi}_{-t}\o\sigma\o\operatorname{Fl}^{\xi_0}_t$. Since
$(\operatorname{Fl}^{\tilde\xi}_t)^*\omega=\omega$ we conclude that the connection form
of the Weyl connection associated to a pulled back Weyl structure
coincides with $(\operatorname{Fl}^{\xi_0}_t)^*\tilde\gamma^\#$. Thus we can prove
invariance of $\tilde\gamma^\#$ under the flow of $\xi_0$ by showing
that $\tilde\xi$ preserves the Weyl structure associated to $\xi$.
Denoting by $\tau\in\Gamma(Q)$ the section obtained by projecting
$\xi$, it is evident that the Weyl structure obtained by pulling
back $\sigma$ as above, leaves $(\operatorname{Fl}^\xi_t)^*\tau$ invariant and is
uniquely characterized by this property. But since $\tau$ is a
projection of $\xi$ and the Lie derivative of $\xi$ along $\xi$
equals $[\xi,\xi]=0$, we conclude $(\operatorname{Fl}^\xi_t)^*\tau=\tau$, so $\sigma$
is invariant under the flow of $\tilde\xi$.
Knowing that $\tilde\gamma^\#$ is invariant under the flow of $\xi_0$,
it follows readily that the same is true for the function
$\tilde\gamma^\#(\xi_0)$. Also, the pullback of $\alpha$ clearly is
invariant under the flow of $\xi_0$, since $\alpha$ is invariant under
the flow of $\xi$. Thus we conclude that $\gamma^\#$ is invariant under
the flow of $\xi_0$ and hence descends to a $\mathfrak g_0$--valued
one--form, which then clearly is a principal connection form on
$\mathcal G_0\to M$.
To complete the proof it thus suffices to prove that the
corresponding linear connection on $TM$ has algebraically harmonic
torsion. Now the torsion of a linear connection on $TM$ can be
computed from the corresponding principal connection by evaluating
the exterior derivative of the soldering form on horizontal lifts of
the vector fields. Now the soldering form on $\mathcal G_0$ pulls back
to the form $\th_{-1}^\#$ on $\mathcal G_0^\#$. This can be extended to
a one--form on $\mathcal G_0^\#$ by requiring that it vanishes on
$\xi_0$ (which corresponds to the isomorphism between $TM^\#$ and
$H\oplus Q$ defined by the Weyl structure determined by
$\xi$). Evaluating the exterior derivative of this one--form on the
horizontal lifts of two sections $\eta_1$ and $\eta_2$ of $H\to
M^\#$, one obtains the function corresponding to
$$
(\eta_1,\eta_2)\mapsto
\nabla_{\eta_1}\eta_2-\nabla_{\eta_2}\eta_1-\pi([\eta_1,\eta_2]),
$$
where $\nabla$ is the Weyl connection on $H$ and $\pi:TM^\#\to H$ is
the projection determined by the Weyl structure. The resulting bundle
map $H\times H\to H$ is well known to have values in $\ker(\square)$,
compare with Theorem 5.2.11 of \cite{book}. Taking $\eta_1$ and
$\eta_2$ to be local lifts of vector fields on $M$, $[\eta_1,\eta_2]$
is a local lift of their Lie bracket. But this differs from
$\pi([\eta_1,\eta_2])$ by a multiple of $\xi$, so
$\pi([\eta_1,\eta_2])$ is the section of $H$ lifting the Lie bracket
of the downstairs vector fields. This immediately implies that this
expresses the torsion of the linear connection associated to $\gamma$,
which completes the proof.
\end{proof}
\subsection{Contactifications and special symplectic
connections}\label{2.10}
Using Theorem \ref{2.9}, we can now complete the discussion of the
relation between PCS--structures and special symplectic connections in
the sense of \cite{Cahen-Schwachhoefer}. Consider a contact grading of
a simple Lie algebra $\mathfrak g$, which is not of type $C_n$, and let
$\mathfrak g_0\subset\mathfrak{csp}(\mathfrak g_{-1})$ be the corresponding
reductive subalgebra. As discussed in Section 4.7 of \cite{PCS1}, the
intersection $\mathfrak g^0_0$ of $\mathfrak g_0$ with $\mathfrak{sp}(\mathfrak
g_{-1})$ is called a special symplectic subalgebra and there is a
corresponding concept of special symplectic connections. In particular,
that class contains all linear connections with exceptional holonomy
that preserve a symplectic form.
In Theorem 4.7 of \cite{PCS1}, we have shown that any special
symplectic connection of type $\mathfrak g$ is the distinguished
connection of some PCS--structure of type $(G,P)$ for a group $G$ with
Lie algebra $\mathfrak g$. Now we can characterize the special symplectic
connections among these distinguished connections via local flatness
of contactifications.
\begin{thm*}
Consider a PCS--structure of type $(G,P)$ on a smooth manifold $M$
and let $\nabla$ be the distinguished connection on $TM$ determined
by this structure. Then $\nabla$ is a special symplectic connection
if and only if any local parabolic contactification of $M$ is
locally flat (as a parabolic contact structure).
\end{thm*}
\begin{proof}
We can apply Theorem \ref{2.9} to a local contactification $M^\#$ of
$M$ and we denote by $\xi$ the corresponding infinitesimal
automorphism of $M^\#$. As we have noted in the proof of that
theorem, the Weyl structure determined by $\xi$ defines a projection
$\pi:TM^\#\to H$, as well as a partial linear connection $\nabla^\#$
on $H$. Further, taking local lifts $\eta_1,\eta_2\in\Gamma(H)$ of two
vector fields on $M$, $\pi([\eta_1,\eta_2])\in\Gamma(H)$ is the lift of
the Lie bracket. Using this, we showed that the component in
$\Lambda^2H^*\otimes H$ of the torsion of $\nabla^\#$ is simply the lift
of the torsion of the distinguished connection $\nabla$ on $M$.
Since $\nabla$ has algebraically harmonic torsion by definition, we
concluded that $\nabla$ is torsion--free if and only if the harmonic
part of the above torsion component of $\nabla^\#$ vanishes
identically. Since this torsion component has homogeneity one, it
coincides with the component of the harmonic curvature in
$\Lambda^2H^*\otimes H$. But by Lemma 4.2.2 of \cite{book}, all the
harmonic torsion of a parabolic contact structure is concentrated in
$\Lambda^2H^*\otimes H$. Hence torsion--freeness of $\nabla$ is
equivalent to vanishing of the harmonic torsion of the parabolic
geometry on $M^\#$.
From the discussion of parabolic contact structures in Section 4.2
of \cite{book} it follows that for $\mathfrak g$ not of type $A_n$,
vanishing of the harmonic torsion is equivalent to local flatness of
the parabolic geometry. On the other hand, by Theorem 4.7 of
\cite{PCS1}, for $\mathfrak g$ not of type $A_n$, torsion--freeness of
$\nabla$ is equivalent to $\nabla$ being special symplectic, so the
proof is complete in this case.
For $\mathfrak g$ of type $A_n$, Proposition 4.2.3 and the discussion in
Section 4.2.4 in \cite{book} show that vanishing of the harmonic
torsion of the parabolic geometry on $M^\#$ is equivalent to
torsion--freeness of this geometry. Moreover, we can apply part (3)
of Theorem 4.2.2 of \cite{book} to $\nabla^\#$ and conclude that the
harmonic curvature component in homogeneity two of this geometry
(which is the only remaining component in dimensions $\geq 5$) can
be recovered as a component of curvature of $\nabla^\#$. This
curvature is a two--form with values in $\text{End}(H)$ and one has to
restrict this two--form to $\Lambda^2_0H^*\subset\Lambda^2 T^*M$, the kernel
of the Levi--bracket within $\Lambda^2H$. This defines a section of the
bundle associated to $\Lambda^2_0\mathfrak g_1\otimes\mathfrak g_0$ and one has
to further restrict to the kernel of the Kostant--Laplacian
$\square$ on this representation. But on $\Lambda^2_0H^*$, one can also
compute the curvature via
$$
R(\eta_1,\eta_2)(\xi)=\nabla^\#_{\eta_1}\nabla^\#_{\eta_2}\xi-
\nabla^\#_{\eta_2}\nabla^\#_{\eta_1}\xi-\nabla^\#_{\pi([\eta_1,\eta_2])}\xi.
$$
This shows that the relevant component of the curvature of
$\nabla^\#$ is just the pullback of the curvature of $\nabla$. Hence
we see that the parabolic contact structure on $M^\#$ is locally flat
if and only if the component of its curvature in the subbundle
corresponding to $\ker(\square)\subset \Lambda^2\mathfrak g_1\otimes\mathfrak g_0$
vanishes identically. This subspace is well known to be the $\mathfrak
g_0$--irreducible component of maximal highest weight, which
immediately implies that this condition is equivalent to vanishing of
the Bochner curvature of $\nabla$ and thus the result.
\end{proof}
\section{Conformally Fedosov structures and their
contactifications}\label{3}
In the previous discussions, we have excluded type $C_n$, since in
this case $\mathfrak g_0=\mathfrak{csp}(\mathfrak g_{-1})$. This corresponds to
the fact that the parabolic contact structure determined by this
grading is not defined by a regular infinitesimal flag structure in
the sense of Section \ref{2.3}. The contact grading of
$\mathfrak{sp}(2n,\mathbb R)$ gives rise to so--called contact projective
structures, an analog of classical projective structures. In this
section, we discuss a geometric structure refining a conformally
symplectic structure, which can be thought of as the PCS--structure
associated to the contact grading of $\mathfrak{sp}(2n,\mathbb R)$. (It seems
to be possible to define a similar concept related to almost
conformally symplectic structures, but since contactifications are
only available in the conformally symplectic case, we do not pursue
this direction here.)
Initially, this structure can be described as a projective structure
compatible with a conformally symplectic structure. However, there is
a counterpart to the vanishing of the first prolongation for
PACS--structures, which implies that such a geometry is given by a
conformally symplectic structure and a torsion--free affine connection
preserving this structure. Locally, such a structure is
even equivalent to a symplectic structure together with a
torsion--free symplectic connection. These geometries were originally
introduced by M.G.~Eastwood and J.~Slov\'ak (in a slightly different
presentation) in the first version of \cite{E-S} under the name
``conformally Fedosov structures'', and we will keep this name here.
\subsection{Projective structures and conformally Fedosov
structures}\label{3.1}
Recall that two torsion--free linear connections $\nabla$ and
$\hat\nabla$ on the tangent bundle of a smooth manifold $M$ are called
\textit{projectively equivalent} if and only if they have the same
geodesics up to reparametrization. This can be equivalently
characterized as the existence of a one--form $\Upsilon$ on $M$ such that
$$
\hat\nabla_\xi\eta=\nabla_\xi\eta+\Upsilon(\xi)\eta+\Upsilon(\eta)\xi,
$$
for all $\xi,\eta\in\mathfrak X(M)$. A projective structure on a smooth
manifold $M$ is then defined as a projective equivalence class of
torsion--free connections on $M$. Equivalently, it can be viewed as
being given by the family of one--dimensional submanifolds formed by
the geodesic paths of the connections in the class.
\begin{definition*}
A \textit{conformally Fedosov structure} on a conformally symplectic
manifold $(M,\ell\subset\Lambda^2T^*M)$ is a projective structure
$[\nabla]$ on $M$, which contains a connection that preserves $\ell$.
\end{definition*}
The following characterization is essentially contained in \cite{E-S},
we reproduce it for convenience of the reader. We use abstract index
notation as in \cite{E-S}.
\begin{prop*}
For a projective equivalence class $\Phi$ of torsion--free affine
connections on a conformally symplectic manifold $(M,\ell)$, the
following conditions are equivalent:
(i) $\Phi$ defines a conformally Fedosov structure on $(M,\ell)$.
(ii) For one (or equivalently any) connection $\nabla=\nabla_a$ in
the class $\Phi$ and one non--vanishing (or equivalently any) local
section $\omega=\omega_{bc}$ of $\ell$, we have
$\nabla_{(a}\omega_{b)c}=\phi_{(a}\omega_{b)c}$ for some one--form
$\phi=\phi_a$.
(iii) There is a unique connection $\nabla_a$ in $\Phi$ which
preserves $\ell$.
\smallskip
\noindent
Moreover, the connection on $\ell$ induced by the connection $\nabla$
from (iii) is flat with any local closed section being parallel. In
particular, locally $\nabla$ is a torsion--free symplectic connection
with respect to each of these closed sections.
\end{prop*}
\begin{proof}
The formula for the change of the induced connection on two--forms
from \cite{E-S} says that changing from $\nabla_a$ to $\hat\nabla_a$
corresponding to a one--form $\Upsilon_a$ one gets
$$
\hat\nabla_a\omega_{bc}=\nabla_a\omega_{bc}-2\Upsilon_a\omega_{bc}-\Upsilon_b\omega_{ac}-\Upsilon_c\omega_{ba}.
$$
If we symmetrize over $(a,b)$,then the last term does not contribute
and we get
$$
\hat\nabla_{(a}\omega_{b)c}=\nabla_{(a}\omega_{b)c}-3\Upsilon_{(a}\omega_{b)c}.
$$
This shows that if the condition in (ii) is satisfied for one
connection in the projective class $\Phi$, then it is satisfied for all
such connections. Likewise, one easily verifies that if the condition
in (ii) is satisfied for one non--vanishing section, then it is
satisfied for any section of $\ell$.
Condition (i) by definition means that there is a connection
$\nabla_a$ in $\Phi$ such that for any local section $\omega_{bc}$ of
$\ell$, we have $\nabla_a\omega_{bc}=\phi_a\omega_{bc}$ for some one--form
$\phi=\phi_a$. Thus (i) implies (ii) and evidently (iii) implies (i), so
we can complete the proof by showing that (ii) implies (iii).
To do this, take a local non--vanishing section $\omega_{bc}$ of $\ell$,
which is closed as a two--form on $M$, and any connection $\nabla_a$
in the class $\Phi$. Then by (ii) there is a one--form $\phi_a$ such
that $\nabla_{(a}\omega_{b)c}=\phi_{(a}\omega_{b)c}$. Passing to the
projectively equivalent connection $\hat\nabla$ corresponding to
$\Upsilon_a=\tfrac13\phi_a$ we see from above that
$\hat\nabla_{(a}\omega_{b)c}=0$. This means that apart from the apparent
skew symmetry in $b$ and $c$, $\hat\nabla_a\omega_{bc}$ is also skew
symmetric in $a$ and $b$. But this easily implies that the complete
alternation of $\hat\nabla_a\omega_{bc}$ is a multiple of
$\hat\nabla_a\omega_{bc}$, and since this alternation coincides with
$d\omega=0$, we see that $\omega$ is parallel for $\hat\nabla$. This implies
the existence part of (iii) as well as the last claim.
To prove the uniqueness part of (iii) observe that by part (iii) of
Corollary 4.4 of \cite{PCS1} every connection compatible with $\ell$
has the property that any local closed section of $\ell$ is parallel
for the induced connection. But if $\omega$ is any such section, then by
non--degeneracy $\omega^n=\omega\wedge\dots\wedge\omega$ is a volume form on
$M$, which clearly is parallel for the induced connection. But it is a
well known fact in the theory of projective structures that a
projective class can contain at most one connection which leaves some
fixed volume form parallel.
\end{proof}
\subsection{Contact projective structures}\label{3.2}
A parabolic contact structure corresponding to the contact grading of
a simple Lie algebra of type $C_n$ is a so--called \textit{contact
projective structure}. As discussed above, this geometry is
exceptional among parabolic contact structures, since it is not
equivalent to an underlying regular infinitesimal flag
structure. Rather than that, it can be described in terms of an
equivalence class of compatible (partial) connections.
To recall the necessary notions, let $M^\#$ be a smooth manifold
endowed with a contact structure $H\subset TM^\#$. A \textit{partial
connection} on $H$ is a bilinear operator
$\nabla:\Gamma(H)\times\Gamma(H)\to\Gamma(H)$, which is linear over smooth
functions in the first variable and satisfies the usual Leibniz rule
in the second variable. So this is like a covariant derivative, but it
is only possible to differentiate in directions lying in the contact
subbundle. As for usual linear connections, one obtains an induced
partial connection on $\Lambda^2H$, and $\nabla$ is called a
\textit{partial contact connection} if this induced connection
preserves the kernel of the Levi--bracket $\mathcal L:\Lambda^2H\to Q$. In
this case, there is an induced partial connection $\nabla^Q$ on the
line bundle $Q$.
Using this, one can next associate a projection $\pi:TM^\#\to H$ to a
partial contact connection $\nabla$, see Section 4.2.2 and 4.2.6 of
\cite{book}. This is uniquely characterized by the fact that for
$\zeta\in\mathfrak X(M^\#)$ and $\eta\in\Gamma(H)$ we have
$$
\mathcal L(\pi(\zeta),\eta)=\nabla^Q_\eta(\zeta+H)-([\eta,\zeta]+H).
$$
here $\zeta+H$ denotes the section of $Q$ determined by $\zeta$ and
likewise for the Lie bracket. Using this projection, we can define the
\textit{contact torsion} of the partial contact connection $\nabla$
via
$$
T(\eta_1,\eta_2)=\nabla_{\eta_1}\eta_2-\nabla_{\eta_2}\eta_1-\pi([\eta_1,\eta_2])
$$
for sections $\eta_1,\eta_2\in\Gamma(H)$. This expression is immediately
seen to be bilinear over smooth functions and thus defines a bundle
map $T:\Lambda^2H\to H$.
Finally, there is a notion of contact projective equivalence for
partial contact connections. Consider a smooth section $\Upsilon$ of the
bundle $H^*$ of linear functionals on the contact subbundle. Then we
define $\tilde\Upsilon:Q\to H$ by $\mathcal
L(\tilde\Upsilon(\tau),\eta)=\Upsilon(\eta)\tau$. Given a partial contact
connection $\nabla$ on $H$, one defines a partial connection
$\hat\nabla$ on $H$ by
$$
\hat\nabla_{\eta_1}\eta_2=\hat\nabla_{\eta_1}\eta_2+\Upsilon(\eta_1)\eta_2+
\Upsilon(\eta_2)\eta_1+\tilde\Upsilon(\mathcal L(\xi,\eta)).
$$
One easily verifies directly that this is again a partial contact
connection on $H$. Further, one shows that the projection $\hat\pi$
associated to $\hat\nabla$ is given by
$\hat\pi(\zeta)=\pi(\zeta)+2\tilde\Upsilon(\zeta+H)$ and this easily
implies that $\hat\nabla$ and $\nabla$ have the same contact torsion.
One calls two partial contact connections \textit{contact projectively
equivalent} if they are related in the above way. A (torsion--free)
contact projective structure on $M^\#$ is then given by a contact
structure together with a class of contact projectively equivalent
partial contact connections with vanishing contact torsion. It turns
out that such a structure can be equivalently described by a Cartan
geometry as discussed in Section \ref{2.3}, which corresponds to the
contact grading of the Lie algebra $\mathfrak{sp}(2n+2,\mathbb R)$. As a
group, one can take $G=Sp(2n+2,\mathbb R)$ and $P\subset G$ the
stabilizer of a ray in the standard representation $\mathbb R^{2n+2}$ of
$G$. Then it turns out that $G_0$ is the conformally symplectic group
$CSp(2n,\mathbb R)$, see Section 4.2.6 of \cite{book}.
\begin{remark*}
(1) Similarly to classical projective structures, there is also an
interpretation of contact projective equivalence in terms of
geodesics in directions tangent to the contact distribution, and a
corresponding description of contact projective structures, see
\cite{Fox} and Section 4.2.7 in \cite{book}.
(2) In contrast to the case of usual affine connections, it is in
general not possible to remove the contact torsion of a partial
contact connection without changing its contact geodesics. Hence it
is natural to extend the notion of a contact projective structure to
the case of non--vanishing contact torsion, see \cite{Fox}. It turns
out that also in this more general setting, one can associate a
canonical Cartan geometry to such a structure. The resulting Cartan
geometry does not fit into the general scheme of regular normal
parabolic geometries, however.
\end{remark*}
\subsection{Contactifications of conformally Fedosov
structures}\label{3.3}
Our final task in this article is to establish analogs of Theorems
\ref{2.5}, \ref{2.7} and \ref{2.8} for contact projective structures
and conformally Fedosov structures. This turns out to be rather easy
since, as indicated in \ref{3.1}, the projective freedom in a
conformally Fedosov structure does not really show up. Likewise, in
the setting of contact projective structures, we will always have a
preferred infinitesimal automorphism around. As discussed in
\ref{2.9}, this determines a preferred Weyl structure and thus a
distinguished representative in the contact projective class. Hence we
can always related connections rather than projective equivalence
classes, and this works similarly as in the proof of Theorem
\ref{2.9}.
\begin{prop*}
Let $q:M^\#\to M$ be a quotient of a contact manifold by a
transverse infinitesimal contactomorphism $\xi\in\mathfrak X(M^\#)$, and
let $\ell\subset\Lambda^2T^*M$ be the induced conformally symplectic
structure.
(1) A contact projective structure on $M^\#$, for which $\xi$ is an
infinitesimal automorphism, induces a conformally Fedosov structure
on $M$ with underlying conformally symplectic structure $\ell$.
(2) Conversely, a conformally Fedosov structure on $M$ with
underlying conformally symplectic structure $\ell$ canonically lifts
to a contact projective structure for which $\xi$ is an
infinitesimal automorphism.
(3) Suppose that we have two conformally Fedosov structures realized
as quotients of contact projective structures as in (2). Then any
local lift of a diffeomorphism respecting the conformally Fedosov
structures to a contactomorphism is automatically compatible with
the infinitesimal automorphisms as in Theorem \ref{2.8} and a local
isomorphism of contact projective structures.
\end{prop*}
\begin{proof}
Let $\mathcal G_0^\#\to M^\#$ be the natural frame bundle for the
contact subbundle $H$ with structure group $G_0=CSp(2n,\mathbb R)$ and
let $\mathcal G_0\to M$ be the natural frame bundle for $TM$ induced by
$\ell$, which also has structure group $G_0$, see Section 2.3 of
\cite{PCS1}. By construction, there is a tautological soldering form
$\th\in\Omega^1(\mathcal G_0,\mathbb R^{2n})$. Denoting as before by
$T^{-1}\mathcal G_0^\#$ the preimage of $H\subset TM^\#$ in $T\mathcal
G_0^\#$, we likewise get a tautological soldering form
$\th_{-1}^\#\in\Gamma(L(T^{-1}\mathcal G_0^\#,\mathbb R^{2n}))$.
By definition, the local flows of $\xi$ are contactomorphisms and
hence lift to local principal bundle automorphisms of $\mathcal G_0^\#$.
Differentiating these local flows, we obtain a $G_0$--invariant lift
$\xi_0\in\mathfrak X(\mathcal G_0^\#)$ of $\xi$. As observed in the proof of
Theorem \ref{2.5} the tangent maps of $q$ restrict to linear
isomorphisms on the contact subspaces, which provides a
$G_0$--equivariant lift $q_0:\mathcal G_0^\#\to\mathcal G_0$ of $q:M^\#\to
M$. Moreover, by construction the restriction of $q_0$ to each fiber
of $\mathcal G_0^\#$ is injective. As in the proofs of Lemma \ref{2.4}
and Theorem \ref{2.5}, this shows that $q_0$ is a surjective
submersion and that the restriction of $p_0^\#:\mathcal G_0^\#\to M^\#$
to any fiber of $q_0$ is a diffeomorphism onto a fiber of $q$, so in
particular $q_0$ has connected fibers. Finally, it is easy to see
that the tangent spaces to these fibers are spanned by $\xi_0$.
(1) We can pass to the Cartan geometry determined by the given
contact projective structure on $M^\#$, which has $\mathcal G_0^\#$ as
its underlying $G_0$--bundle. The general theory of Weyl structures
applies to contact projective structures. Hence as in the proof of
Theorem \ref{2.9}, we see that the Weyl structure determined by
$\xi$ determines a principal connection form $\tilde\gamma^\#$ on $\mathcal
G_0^\#$. Here the Weyl connection is an extension of a partial
contact connection with vanishing contact torsion, which lies in the
contact projective class. Continuing as in the proof of Theorem
\ref{2.9}, we can change this to a principal connection form
$\gamma^\#$ for which $\xi_0$ is horizontal, and show that this
descends to $\mathcal G_0$. Still as in that proof, we can verify that
the resulting connection on $M$ is torsion--free and by construction
it preserves $\ell$. So we can use the projective class of this
connection to define a conformally Fedosov structure on $M$.
(2) Starting with a conformally Fedosov structure on $M$, we know from
Proposition \ref{3.1} that there is a unique connection $\nabla$ in
the projective class, which preserves the conformally symplectic
structure $\ell$. Then $\nabla$ defines a principal connection
$\gamma\in\Omega^1(\mathcal G_0,\mathfrak g_0)$. Now we can pull back this
connection form to $\mathcal G_0^\#$ and then restrict it to $T^{-1}\mathcal
G_0^\#$. The result is a $G_0$--equivariant section of $L(T^{-1}\mathcal
G_0^\#,\mathfrak g_0)$ which by construction reproduces the generators
of fundamental vector fields. Analogously to the standard
construction of induced connections, this gives rise to a partial
contact connection $\nabla^\#$ on $H\subset TM^\#$.
By construction, the local flows $\operatorname{Fl}^{\xi_0}_t$ satisfy
$q_0\o\operatorname{Fl}^{\xi_0}_t=q_0$ and they preserve the contact distribution
$H\subset TM^\#$. This immediately implies that these follows
preserve the pullback of $\gamma$ and its restriction to the contact
distribution and hence the partial contact connection
$\nabla^\#$. Now let $\alpha$ be the unique contact form on $M^\#$ such
that $\alpha(\xi)=1$ and $\omega$ the unique symplectic form on $M$ in the
given conformal symplectic class such that $q^*\omega=d\alpha$. Consider
sections $\eta_1$, $\eta_2$ and $\zeta$ of $H\subset TM^\#$ which
project onto vector fields $\underline{\eta}_1$,
$\underline{\eta}_2$ and $\underline{\zeta}$ on $M$. Then by
definition, the Levi bracket $\mathcal L(\eta_1,\eta_2)$ is equal to
$\alpha([\eta_1,\eta_2])(\xi+H)=-(\omega(\underline{\eta}_1,\underline{\eta}_2)\o
q)(\xi+H)$. Now by definition of the partial connection on $Q$
induced by $\nabla^\#$, we get
$$
\nabla^Q_\zeta\mathcal L(\eta_1,\eta_2)=\mathcal L(\nabla^\#_\zeta
\eta_1,\eta_2)+\mathcal L(\eta_1,\nabla^\#_\zeta \eta_2).
$$
By construction $\nabla^\#_\zeta \eta_i$ is a lift of
$\nabla_{\underline{\zeta}}\underline{\eta}_i$ for $i=1,2$. From
Proposition \ref{3.1}, we then know that $\nabla\omega=0$ and
inserting this, we easily conclude that $\nabla^Q(\xi+H)=0$. Since
$\xi$ is an infinitesimal contactomorphism, this implies that the
projection $\pi:TM^\#\to H$ induced by $\nabla^\#$ satisfies
$\pi(\xi)=0$, which of course determines $\pi$ completely.
Using these facts, we can now continue completely parallel to the
proof of Theorem \ref{2.9}. We see that
$\pi([\eta_1,\eta_2])\in\Gamma(H)$ is the unique section of $H$ lifting
$[\underline{\eta}_1,\underline{\eta}_2]$. Using this,
torsion--freeness of $\nabla$ implies that $\nabla^\#$ has vanishing
contact torsion. Hence we can use its contact projective equivalence
class to define a contact projective structure on $M$, and since the
flows of $\xi$ even preserve the partial contact connection
$\nabla^\#$, it is an infinitesimal automorphism of this structure.
(3) Let $\phi:M\to\tilde M$ be the morphism of conformally Fedosov
structures and let $\phi^\#:M^\#\to\tilde M^\#$ be a contactomorphism
lifting $\phi$. Then the first step of the proof of Theorem \ref{2.8}
shows that $(\phi^\#)^*\tilde\xi=\lambda\xi$ for a nowhere vanishing,
locally constant function $\lambda$. Now there are natural lifts
$\Phi:\mathcal G_0\to{\tilde{\mathcal G}}_0$ and $\Phi^\#:\mathcal G_0^\#\to{\tilde{\mathcal G}}_0^\#$ to the
frame bundles and $(\Phi^\#)^*\tilde\xi_0=\lambda\xi_0$. In view of the
constructions of $\Phi^\#$ and $q_0$ and $\tilde q_0$, we see that
$\tilde q\o\phi^\#=\phi\o q$ implies $\tilde q_0\o\Phi^\#=\Phi\o q_0$.
Now a morphism of conformally Fedosov structures preserves both the
conformally symplectic structure and the projective class. Hence the
uniqueness statement in statement (iii) in Proposition \ref{3.1}
implies that it is also compatible with the (unique) connections in
the projective class which preserve the conformally symplectic
structures. Hence for their connection forms, we get
$\Phi^*\tilde\gamma=\gamma$, which together with the above implies that
$(\Phi^\#)^*\tilde q_0^*\tilde\gamma=q_0^*\Phi^*\tilde\gamma=q_0^*\gamma$. Together
with the fact that $(\Phi^\#)^*\tilde\xi_0=\lambda\xi_0$, this implies
that $\Phi^\#$ is compatible with the partial contact connections
constructed in (2) and hence in particular with the contact
projective structures they generate.
\end{proof}
Parallel to Theorem \ref{2.10}, there also is a relation between
conformally Fedosov structure and a class of special symplectic
connections. The relevant connections here are called
\textit{symplectic connections of Ricci type}, see \cite{BC} and
\cite{Cahen-Schwachhoefer}. The module of formal curvatures of
symplectic connections splits over $\mathfrak{sp}(2n)$ as a direct sum of
the kernel of the Ricci--type contractions and a complementary
submodule. A symplectic connection of Ricci type is a torsion--free
symplectic connection for which the curvature is concentrated in that
complementary submodule.
Proposition \ref{3.1} immediately implies that for a symplectic
manifold $(M,\omega)$ and a torsion--free symplectic connection $\nabla$,
the underlying conformally symplectic and projective structures define
a conformally Fedosov structure. On the other hand, for a general
conformally Fedosov structure, the connection $\nabla$ from part (iii)
of Proposition \ref{3.1} has curvature in
$\Lambda^2T^*M\otimes\mathfrak{sp}(TM)$ by the last part of the
proposition. Hence the concept of being of Ricci type makes sense for
this connection.
\begin{cor*}
The distinguished connection $\nabla$ of a conformally Fedosov
structure from part (iii) of Proposition \ref{3.1} is of Ricci type
if and only if any local parabolic contactification of the structure
is locally flat as a contact projective structure.
\end{cor*}
\begin{proof}
Consider a local parabolic contactification $q:M^\#\to M$. In the
construction of the conformally Fedosov structure on $M^\#$
in the proof of part (2) of Theorem \ref{3.3}, we used $\nabla$ to
construct a partial contact connection $\nabla^\#$ on $M^\#$. The
verifications in that proof actually show that $\nabla^\#$ is the
partial contact connection coming from the Weyl structure determined
by the infinitesimal automorphism $\xi$ giving rise to $q$.
The parabolic geometry determined by a contact projective structure
with vanishing contact torsion is well known to be torsion--free, see
Section 4.2.6 of \cite{book}. There it is also verified that the only
harmonic curvature of a contact projective structure has homogeneity 2
and is a section of the bundle induces by the highest weight subspace
in $\Lambda^2\mathfrak g_{-1}^*\otimes\mathfrak g_0$, which is exactly the kernel
of the Ricci--type contraction. Now using Theorem 4.2.2 of
\cite{book}, one concludes as in the proof of Theorem \ref{2.10} that
this harmonic curvature can be computed from $\nabla^\#$. Using the
information from the proof of part (2) of Theorem \ref{3.3} one
deduces as in the proof of Theorem \ref{2.10} that vanishing of the
component is equivalent to vanishing of the corresponding component
of the curvature of $\nabla$, which implies the result.
\end{proof}
\subsection{Example}\label{3.4}
There is an analog of the global contactifications discussed in
Section \ref{2.6} in the context of conformally Fedosov
structures. This actually is the example relevant for the results on
integral geometry in \cite{E-G} which motivate the developments in this
series of articles.
Let us realize the standard symplectic form in dimension $2n+2$ as the
imaginary part of the standard Hermitian form on $\mathbb C^{n+1}$. This
gives rise to an action of $G=Sp(2n+2,\mathbb R)$ on $\mathbb C^{n+1}$
(which does not preserve the complex structure). Viewing $S^{2n+1}$ as
the space of real rays in $\mathbb C^{n+1}$ we get a transitive action of
$G$, which makes $S^{2n+1}$ into the homogeneous model for contact
projective structures, compare with Section 4.2.6 of \cite{book}. Now
the standard action of $U(1)$ on $\mathbb C^{n+1}$ by complex
multiplication defines a subgroup in $G$ and hence an action of $U(1)$
on $S^{2n+1}$ by automorphisms of the contact projective
structure. The tangent space of $S^{2n+1}$ in a point corresponding to
a ray in $\mathbb C^{n+1}$ can be viewed as the quotient of the real
vector space $\mathbb C^{n+1}$ by the line spanned by the ray. The
contact subspaces then correspond to the quotient of the symplectic
orthocomplement of the ray by this line. This readily implies that the
infinitesimal generator of our $U(1)$--action is transversal
everywhere. Of course the $U(1)$--action on $S^{2n+1}$ admits a global
quotient, namely the canonical projection $q:S^{2n+1}\to\mathbb CP^n$
mapping a real ray in $\mathbb C^{n+1}$ to the complex line it spans.
Hence we obtain exactly the same quotient map as in example (1) from
\ref{2.6}, where we started from the CR--structure on $S^{2n+1}$,
which is also preserved by the $U(1)$--action. It is actually easy to
see that the Weyl connections associated to the infinitesimal
generator of the $U(1)$--action via the CR--structure and via the
projective contact structure coincide. But this implies that the
underlying conformally Fedosov structure on $\mathbb CP^n$ actually is
the one determined by the Levi--Civita connection of its K\"ahler
metric.
\begin{remark*}
The last bit of this example actually exhibits a general
phenomenon. Suppose that we have a torsion--free PCS--structure (not
of type $C_n$) on a smooth manifold $M$. Then via Proposition
\ref{3.1}, the (torsion--free) canonical connection actually
determines an underlying conformally Fedosov structure. This may
look surprising, since there is no corresponding construction for an
``underlying contact projective structure'' associated to a
(torsion--free) parabolic contact structure of different
type. However, such an underlying structure does become available in
the presence of a transversal infinitesimal automorphism. Indeed,
one can then consider the Weyl connection determined by this
transversal infinitesimal automorphism and restrict it to a partial
contact connection. General results then ensure that in the case of
a torsion--free parabolic contact structure, this partial contact
connection has vanishing contact torsion and hence determines a
torsion--free contact projective structure, which is preserved by the
given infinitesimal automorphism. Hence a contactification of the
initial PCS--structure at the same time defines a contactification
of the underlying conformally Fedosov structure.
\end{remark*}
\begin{bibdiv}
\begin{biblist}
\bib{BC}{article}{
author={Bourgeois, F.},
author={Cahen, M.},
title={A variational principle for symplectic connections},
journal={J. Geom. Phys.},
volume={30},
date={1999},
number={3},
pages={233--265},
issn={0393-0440},
review={\MR{1692232 (2000d:53121)}},
doi={10.1016/S0393-0440(98)00059-X},
}
\bib{BCG3}{book}{
author={Bryant, R. L.},
author={Chern, S. S.},
author={Gardner, R. B.},
author={Goldschmidt, H. L.},
author={Griffiths, P. A.},
title={Exterior differential systems},
series={Mathematical Sciences Research Institute Publications},
volume={18},
publisher={Springer-Verlag},
place={New York},
date={1991},
pages={viii+475},
isbn={0-387-97411-3},
review={\MR{1083148 (92h:58007)}},
}
\bib{Cahen-Schwachhoefer}{article}{
author={Cahen, Michel},
author={Schwachh{\"o}fer, Lorenz J.},
title={Special symplectic connections},
journal={J. Differential Geom.},
volume={83},
date={2009},
number={2},
pages={229--271},
issn={0022-040X},
review={\MR{2577468 (2011b:53045)}},
}
\bib{Cap-Salac}{article}{
author={{\v{C}}ap, Andreas},
author={Sala{\v{c}}, Tom{\'a}{\v{s}}},
title={Pushing down the Rumin complex to conformally symplectic
quotients},
journal={Differential Geom. Appl.},
volume={35},
date={2014},
number={suppl.},
pages={255--265},
issn={0926-2245},
review={\MR{3254307}},
doi={10.1016/j.difgeo.2014.05.004},
}
\bib{PCS1}{article}{
author={{\v{C}}ap, Andreas},
author={Sala\v c, Tom\'a\v s},
title={Parabolic conformally symplectic structures I; definition
and distinguished connections},
eprint={arxiv:1605.01161},
}
\bib{PCS3}{article}{
author={{\v{C}}ap, Andreas},
author={Sala\v c, Tom\'a\v s},
title={Parabolic conformally symplectic structures III; Invariant
differential operators and complexes},
eprint={arXiv:1701.01306},
}
\bib{book}{book}{
author={{\v{C}}ap, Andreas},
author={Slov{\'a}k, Jan},
title={Parabolic geometries. I},
series={Mathematical Surveys and Monographs},
volume={154},
note={Background and general theory},
publisher={American Mathematical Society},
place={Providence, RI},
date={2009},
pages={x+628},
isbn={978-0-8218-2681-2},
review={\MR{2532439 (2010j:53037)}},
}
\bib{E-G}{article}{
author={Eastwood, Michael},
author={Goldschmidt, Hubert},
title={Zero-energy fields on complex projective space},
journal={J. Differential Geom.},
volume={94},
date={2013},
number={1},
pages={129--157},
issn={0022-040X},
review={\MR{3031862}},
}
\bib{E-S}{article}{
author={Eastwood, Michael G.},
author={Slov{\'a}k, Jan},
title={Conformally Fedosov manifolds},
eprint={arXiv:1210.5597},
}
\bib{Fox}{article}{
author={Fox, Daniel J. F.},
title={Contact projective structures},
journal={Indiana Univ. Math. J.},
volume={54},
date={2005},
number={6},
pages={1547--1598},
issn={0022-2518},
review={\MR{2189678 (2007b:53163)}},
doi={10.1512/iumj.2005.54.2603},
}
\bib{Kostant}{article}{
author={Kostant, Bertram},
title={Lie algebra cohomology and the generalized Borel-Weil theorem},
journal={Ann. of Math. (2)},
volume={74},
date={1961},
pages={329--387},
issn={0003-486X},
review={\MR{0142696 (26 \#265)}},
}
\bib{Sternberg}{book}{
author={Sternberg, Shlomo},
title={Lectures on differential geometry},
publisher={Prentice-Hall, Inc., Englewood Cliffs, N.J.},
date={1964},
pages={xv+390},
review={\MR{0193578 (33 \#1797)}},
}
\end{biblist}
\end{bibdiv}
\end{document}
in \cite{Cahen-Schwachhoefer} and it determines two spaces $\mathcal
R_{\mathfrak g^0_0}\subset K(\mathfrak g^0_0)$ of curvature tensors. Here
$K(\mathfrak g^0_0)$ is the usual space of formal curvature tensors with
values in $\mathfrak g^0_0$, while the existence of $\mathcal R_{\mathfrak g^0_0}$
depends on the fact that we are dealing with a special symplectic
subalgebra. For our purposes, it suffices to know $\mathcal R_{\mathfrak
g^0_0}\cong\mathfrak g^0_0$ and that the Ricci--type contraction is
injective on $\mathcal R_{\mathfrak g^0_0}$. Then a special symplectic
connection of type $\mathfrak g$ on a smooth manifold $M$ of dimension
$\dim(\mathfrak g_{-1})$ as a symplectic linear connection whose curvature
has values in the subbundle corresponding to $\mathcal R_{\mathfrak
g^0_0}\subset K(\mathfrak g^0_0)\subset K(\mathfrak{sp}(\mathfrak g_{-1})$.
|
\section{Introduction}
\label{intro}
The evolution of circumstellar disks around young stars is symbiotic with the formation of planetary systems. In the generally accepted paradigm of star formation, about 80\% of the stars younger than $1\,$Myr are surrounded by a thick circumstellar disk, composed of dust and gas in a typical mass ratio of 1:100 and emitting in the near-infrared. This fraction decreases to $\sim55\%$ at $3\,$Myrs and $\sim15\%$ at $\sim5\,$Myrs \citep{HaischLL2001, HernandezHMG2007, Mamajek2009}. Disks are generally dissipated, then, in less than $10\,$Myrs, although this timescale has been brought into question by recent studies \citep{BellNMJ2013, DeMarchiPGB2013}. This timescale is long enough to allow the early steps of planet formation to occur, such as settling of dust grains on the disk midplane, aggregation of dust grains into larger solid bodies and the formation of planetesimals (i.e. \citealp{Zuckerman2001, DullemondDominik2004}).{\bf Consequently, stars that have dissipated their disks quickly (i.e. in less than 2 or 3 Myrs) may be less likely to form planets than disks which evolved unperturbed}. \par
The evolutionary timescale of circumstellar disks is dictated by several processes. Viscosity in the disk drives a radial flow of gas that finally accretes onto the central star (see, e.g., \citealp{Pringle1981}). In the very inner part of the disk, the accretion is controlled by the stellar magnetic field, which funnels the plasma in accretion columns \citep{GullbringHBC1998,RomanovaLLK2011}. The accreting material hits the stellar surface almost in free-fall at speeds of a few hundred km/s, heating the accretion spots up to temperatures of approximately ten thousand degrees \citep{CalvetGullbring1998}. This process is a source of energetic ultraviolet and soft X-ray radiation. This energetic radiation (together with the intense coronal X-ray emission which characterizes young stars; \citealp{Montmerle1996}) has an important impact on the evolution of the circumstellar disk itself. Far ultraviolet ($FUV$) photons (with energy between $6\,$eV$\,<\,$h$\nu\,<13.6\,$eV) dissociate H$_2$ molecules, while the extreme ultraviolet ($EUV$) and X-ray photons (with energy h$\nu\,>13.6\,$eV) are capable of ionizing hydrogen atoms. This input of energy raises the gas temperature in the outer layers in the disk up to some thousand degrees, and sometime even more than ten thousand, creating an intense thermal pressure that drives a flow of gas away from the disk. This process, called ``{\it disk photoevaporation}'' (e.g. \citealp{JohnstoneHB1998}), is thought to be responsible for the clearing of an intermediate region in the disk in few Myrs and the formation of pre-transition disks \citep{ClarkeGS2001,AlexanderCP2006,ErcolanoDRC2008}. Once the inner disk is decoupled from the outer disk, the inner part accretes onto the central star in a viscous timescale of few $10^5\,$yrs, leading to the creation of transition disks \citep{CalvetDWF2005}. \par
Disk photoevaporation is then a crucial process in normal disk evolution, occurring in less than $10\,$Myrs, when it is induced by the central star itself, and it may also impact planets formation and migration across the disk. However, a significant fraction of stars in our Galaxy form in the proximity of OB stars, which are intense sources of UV radiation and may induce photoevaporation in nearby disks. For instance, in the Trapezium Cluster in Orion disk photoevaporation externally induced by the O6V star $\Theta^1$ Ori has been invoked to explain the protoplanetary disks embedded in an evaporating envelope observed by the Hubble Space Telescope \citep{ODellW1994}. Similar structures have been observed in other clusters, such as the proplyd-like objects in Cygnus~OB2 \citep{WrightDDG2012}, some of which were shown to be evaporating disks by \citet{GuarcelloDWG2014}. In the intermediately massive clusters (i.e. hosting tens of massive stars) NGC~6611 \citep{GuarcelloPMD2007, GuarcelloMDP2009, GuarcelloMPP2010} and NGC~2244 \citep{BalogMRS2007}, the fraction of members with disks has been observed to decline in the proximity of the massive cluster members, as a consequence of a fast disk dissipation due to the intense local UV field. Such externally induced photoevaporation can dissipate disks in timescales as short as $\sim1\,$Myr \citep[i.e.][]{StorzerHollenbach1999}. In low mass clusters hosting only a few massive stars, disks are expected to be affected by the UV radiation field only in proximity (i.e. $\ll 1 \,$pc) of OB stars. \par
Externally induced photoevaporation is not the only environmental feedback that can affect disk evolution. During the dynamical evolution of clusters stars can occasionally encounter other members at small distances. In those cases when the impact parameter is smaller than a few hundred AU and one or both of the interacting stars has a disk, the close encounter can have dramatic effects on the morphology and evolution of the disk, resulting in: significant mass loss from the disk, with part of the material being dispersed in the surrounding medium and part captured by the other star \citep{ClarkePringle1993,PfalznerUH2005,ThiesKGS2010}; enrichment of one circumstellar environment as a result of the mass exchange with more evolved systems \citep{AdamsS2005}; expulsion of forming planets from the disk \citep{AdamsLaughlin2001,SpurzemGHL2009}, that may result in floating planets, or perturbation of the orbits of forming planetesimals and planets \citep[i.e. ][]{ZapateroOsorioBMR2000}. The mass loss from the disk during a close encounter depends on several factors: masses and velocities of the interacting stars, the impact parameter, the direction of motion of the interacting stars with respect to the orbital motion in the disk, the angle between the plane of the disk and the direction of the stellar interaction, etc... \par
Several theoretical studies have examined the impact of close encounters on the disk population in clusters with different stellar density. For instance, \citet{AdamsPFM2006} found that in small clusters with a population of a few hundred members, stars have a probability between 0.1\% to 1\% of having a close encounter with an impact parameter of $100-200\,$AU in one Myr, and concluded that in low-mass environments disks are rarely dispersed by the gravitational interaction between stars. \citet{PfalznerOE2006} have simulated the gravitational interactions in intermediately massive environments such as the Trapezium in Orion. They found that, on average, in $10\,$Myrs disks around massive stars ($M_{stars}>10\,$M$_{\odot}$) lose $\sim80\%$ of their initial mass, disk in stars with $1\,M_{\odot}<M_{stars}<10\,$M$_{\odot}$ about 30\%, while less massive stars lose about 20\%. In this environment, the chances for very close encounters (i.e. $b<90\,$AU) for a solar mass star has been calculated by \citet{Adams2010} to range between 1\%-10\% in $10\,$Myrs. The hybrid N-body/SPH (smoothed particle hydrodynamic) simulation presented in \citet{RosottiDDH2014} have shown that close encounters in clusters mainly affect disks size than disks mass. In the most extreme environments, such as the Arches cluster where the central 1$\,$pc cubic region contains about 125 O stars, one third of the disks are destroyed by gravitational interactions in less than $3\,$Myrs \citep{OlczakKHP2012}. There is a wide range of situations, then, suggested by these theoretical studies, but which still lack observational support. \par
As we discuss in Section \ref{cygob2_sect}, Cyg~OB2 is dynamically not evolved, rich of O stars and low-mass stars both with and without disks, thus particularly well-suited to the study of disk photoevaporation, also because the radiation exposure of the low mass stars can be estimated reliably. This is not the case in most of the known massive clusters, which are more distant, more dynamically evolved and mixed. In this paper we study the evolutionary timescale for disk dissipation in Cygnus~OB2, and we compare our results to previous studies focused on the same topic but in different star forming regions. In Sect \ref{cygob2_sect} we describe Cygnus~OB2 and the data from the Cygnus~OB2 Chandra Legacy Project; the evidence that disks are dissipated by the environment feedback in Cyg~OB2 are presented in Sect. \ref{df_sect}, and in Sect \ref{map_sec} we compare the effectiveness of externally induced photoevaporation and stellar close encounters on disk dissipation. Finally, in Sect. \ref{disc_sec} we discuss the importance of the environment feedback on disk evolution and planet formation, first in Cygnus~OB2 and then in other star-forming environments of our Galaxy.\par
\section{Cygnus~OB2 and the Cyg~OB2 Chandra Legacy Survey}
\label{cygob2_sect}
The massive association Cygnus~OB2 in the Cygnus~X complex provides a laboratory to test the effects of environmental feedback on the evolution of circumstellar disks and planet formation. At a distance of $1400\,$pc \citep{RyglBSM2012} it is the closest massive young association to the Sun. The massive population of Cyg~OB2 has been the subject of several studies \citep{ReddishLP1967, Knodlseder2000, ComeronPRS2002, Hanson2003, DrewGIS2008, WrightDM2015}. In particular, this association hosts two of the few known O3 stars in our Galaxy \citep{Walborn1973,WalbornHLM2002}, together with an incredibly luminous B supergiant, Cyg~OB2~\#12 \citep{MasseyThompson1991,NegueruelaMHC2008}. \par
Together with this massive population, Cyg~OB2 is also rich in young low mass members. Censuses of low mass members based on $Chandra$/ACIS-I observations have found from 1000 to 1500 stars in the central region (down to $\sim 0.5\,$M$_{\odot}$, but complete at $\sim1\,$M$_{\odot}$; \citealp{AlbaceteColomboFMS2007} and \citealp{WrightDrake2009}). The age of these stars has been estimated to range between $3\,$Myrs and $5\,$Myrs by \citet{WrightDDV2010}, although a complex star formation history emerges from other studies: \citet{Hanson2003} dated some massive stars as younger than $2\,$Myrs; \citet{DrewGIS2008} identified a $5-7\,$Myrs old population of A stars southward the association center; stars with disks with intense accretion have been identified by \citet{VinkDSW2008}, and very young embedded protostars and active star forming regions are observed in the periphery of Cyg~OB2 \citep{WrightDDG2012, GuarcelloDWD2013}. In particular, the morphology of some of these structures, such as DR18 \citep{SchneiderBSJ2006}, with ongoing embedded star formation, and the observed orientation toward the center of Cyg~OB2 of several UV illuminated features \citep{Nicolainprep1}, clearly indicates a high level of feedback from the massive members of Cyg~OB2 \citep{GuarcelloDWD2013}. \par
Despite its proximity to the Sun, compared to other massive associations, Cygnus~OB2 is affected by high extinction mainly due to the dust associated with the Cygnus Rift in the foreground. Evidence that the Rift is responsible for an extinction of few magnitudes along this line of sight was first raised by \citet{DickelWendker1978}. \citet{SchneiderSBC2007} found an upper limit for the extinction due to the Rift equal to $A_V=5^m$. This was confirmed by \citet{SaleDUI2009}, who found an increase of extinction along this line of sight from $A_V=2^m$ to $A_V=5^m$ caused by dusts present in a region between 1 and $2\,kpc$ from the Sun. A similar estimate has been done by \citet{DrewGIS2008}, who found that the extinction affecting the stars associated with Cyg~OB2 ranges from $2.5^m<A_V<7^m$, and by \citet{WrightDDV2010}, who found a median $A_V=7.5^m$ in the central region and $A_V=5.5^m$ northward. A slightly lower extinction has been estimated by \citet{GuarcelloWDG2012} from $r-i$ vs. $i-z$ colors, finding a main range of extinction of $2.6^m<A_V<5.6^m$ with a median value $A_V=4.3^m$. The sample of candidate members of the association defined by \citet{Kashyapinprep} have larger extinctions, more similar to what has been found in previous studies: the 10\% and 90\% quantiles of their A$_V$ distribution are in fact $4.4^m$ and $8.5^m$, respectively, with a median value of $6.4^m$. \par
A region of 1 square degree centered on Cyg~OB2 has recently been surveyed with $Chandra$/ACIS-I for the $Chandra$ Cygnus~OB2 Legacy Project \citep{Drakeinprep}. The survey is composed of 36 ACIS-I fields overlapping each other in order to have a almost constant sensitivity in the central $40^{\prime} \times 40^{\prime}$ area. The resulting catalog of 7924 X-ray sources is described in \citet{WrightDGA2014}. Since optical and infrared data are crucial for most of the scientific aims of the survey, the X-ray catalog has been combined with a large set of photometric data: optical data in $riz$ bands from specific observations with OSIRIS@GTC \citep{GuarcelloWDG2012} down to $r=25^m$ ; data in $ugriz$ bands from the SDSS/DR9 public catalog \citep{AiharaAAA2011} down to $r=16^m$; data in $r^{\prime}i^{\prime}H_{\alpha}$ bands from the second data release of the IPHAS public catalog \citep{DrewGIA2005,BarentsenFDG2014} down to $r^{\prime}=21.5^m$; NIR data in $JHK$ from 2MASS/PSC \citep{CutriSDB2003} and UKIDSS/GPS \citep{LucasHLS2008, KingNBG2013}, down to $J=18.5^m$ and $J=21^m$, respectively; Spitzer and MIPS data from the {\it Spitzer Legacy Survey of the Cygnus X region} \citep{BeererKHG2010}. With the exception of the OSIRIS catalog, covering the central $40^{\prime}\times40^{\prime}$ area, all these catalogs cover the entire survey area. This large set of photometric data has been combined in an unique optical-infrared-X-ray catalog containing 328540 sources (\citealp{GuarcelloDWD2013}, and 2014). Fig. \ref{field_img} shows a $8.0\mu$m image of the region of Cyg~OB2, with indicated the area observed with $Chandra$/ACIS-I, together with the positions of known O and B stars. \par
In this paper we study how the disk fraction (i.e.: the fraction of stars associated with Cyg~OB2 bearing a circumstellar disk) changes across the $Chandra$ Cygnus~OB2 Legacy Survey area as a function of the local UV (FUV and EUV) field and the local surface stellar density. For this aim, we use the list of 1843 members with disks selected by \citet{GuarcelloDWD2013}, which has been purged from contamination by foreground field stars, background giants, and extragalactic sources; and the list of 5022 candidate members obtained by \citet{Kashyapinprep} from the catalog of X-ray sources described in \citet{WrightDGA2014}. Among these X-ray sources, 441 are disk-bearing stars and 102 are known OB stars. The remaining 4479 (with 154 multiple matches between the X-ray and optical-infrared catalogs, \citealp{GuarcelloDWN2015}) are good candidate class~III objects of the association. Fig. \ref{spadis_img} shows the spatial distribution of the selected members with disks (left panel) and without disks (right panel), overplotted with the position of the O stars. The contours mark the emission levels at $8.0\mu m$ from Spitzer observations, revealing the dense nebular structures. Some of these are sites of ongoing star formation, such as DR18. In the spatial distribution of the disk-bearing sources, it is possible to distinguish the central cluster and the surrounding annular stellar overdensity discussed in \citet{GuarcelloDWD2013}. Also the spatial distribution of the members without disks shows a clear overdensity corresponding to the central cluster, as well as a rich population in the outer regions. \par
For the purposes of this paper, it is necessary to estimate the UV flux emitted by the O stars associated with Cyg~OB2. We will omit from the calculation the flux emitted by the early-B stars, whose census is still incomplete. Table \ref{ostars_tb} shows the positions and spectral types of the known O stars, together with their expected FUV and EUV luminosity. Their FUV luminosity have been obtained using the interpolated relations from \citet{ParravanoHM2003}. Stellar masses and spectral types, necessary to calculate the FUV luminosity, are taken from \citet{WrightDM2015}. The EUV luminosity in Table \ref{ostars_tb} are in units of number of ionizing photons with $\lambda < 912\,$\AA{} per second (the $Q_0$ value in \citealp{MartinSH2005}). We included in the list the three WR stars of the association and the O stars which are outside the field of our survey but within a few arcminutes. \par
\section{Spatial variation of the disk fraction in Cygnus~OB2}
\label{df_sect}
To study the effect that O stars have on disk evolution in Cyg~OB2, we need to calculate how the disk fraction varies across the association as a function of the local EUV and FUV radiation fields. For this aim, using the technique adopted in \citet{GuarcelloPMD2007}, we calculate the FUV and EUV fluxes emitted by each O star and incident at the position of each star associated with Cyg~OB2, both with and without a disk, using the projected distances (i.e. the 2D projection of the real distances). In this way it is possible to calculate the disk fraction in given ranges of incident UV flux. The results are shown in Fig. \ref{dfuv_img}. The upper and lower left panels show the disk fraction variation vs. the incident FUV and EUV fluxes, respectively. The FUV fluxes are measured in terms of the Habing flux $G_0$, equal to $1.6\times10^{-3}\,$erg/cm$^{2}$/s; 1.7$\,$G$_0$ corresponds to the average UV flux in the $912-2000\,$\AA{} spectral range in the Solar neighborhood \citep{Habing1968}. The EUV fluxes are described in units of photons/s/cm$^{2}$. The size of the bins is fixed in order to include the same number of disk-less members in each of them, so that the observed variation is given by the decrease of the number of stars with disks. The upper limit of incident fluxes used in Fig. \ref{dfuv_img} (i.e. log$\left( F_{FUV} \right)=4.7$ and log$\left(F_{EUV}\right)=13.5$) are chosen only to improve the visualization of the figure and they are much smaller than the actual upper limits of the flux incident on the stars. The right panels show the spatial distribution of the candidate members (both with and without disks) and the O stars. The different colors are used to mark the position of the stars falling in each of the bins in the right panels. \par
Fig. \ref{dfuv_img} clearly shows a smooth decline of the disk fraction from $\sim40\%$ in the regions with low UV fluxes to $\sim18\%$ in those close to the O stars. Hereafter, the bins will be labeled with recurring numbers starting from those at high fluxes. In this way, the first bin is the one corresponding to the highest UV fluxes, the sixth bin the one corresponding to the lowest fluxes. \par
In this calculation, we are not accounting for any evolution of the UV radiation field in Cyg~OB2 due to the evolution of the O stars. This approximation should not have a strong impact on our calculation given the age of Cyg~OB2 members and the expected evolutionary timescale of these massive stars. Star formation in Cyg~OB2 is expected to have started about 5$\,$Myrs ago \citep{WrightDDV2010}, with most of the members formed about 2-3$\,$Myrs ago \citep{MasseyDW2001}. Since: 1) The lifetime of O stars is expected to range between 4$\,$Myrs and 7$\,$Myrs for the early \citep{LangerHLN1994} and late \citep{MeynetMSS1994,SchaererDeKoter1997} O stars, respectively; 2) star formation is ongoing only in the periphery of the association \citep{GuarcelloDWD2013}; the use of present-day UV flux is a good approximation of the overall UV field experienced by the disks during the evolution of the association. It must also be noted that the study of \citet{WrightPGD2014} showed that in the past Cyg~OB2 never had a highly clustered stellar configuration. {\bf We are also ignoring for the moment the absorption of the UV radiation by intracluster dust particles that may be still present in the association. Given that even reasonably small concentrations of dust particles may result in significant attenuation of the UV radiation, this may play an important role in shaping the disk fraction vs. incident UV flux relation we observe, especially in the case of disks at large distance from the center of the association. The effects of this absorption are described in more detail in Sect. \ref{safe_sec}.}\par
The spatial variation of disk fraction shown in Fig. \ref{dfuv_img} is compatible with a scenario where disks closer to O stars are dissipated faster by externally induced photoevaporation than those in the outer part of the association. However, there are other effects that can produce similar results: a sequence of star formation from the center of the association outward such that the central stars are significantly older, a 2D projection effect and a non uniform sensitivity of the data sets employed to identify Cyg OB2 members with and without disks. \par
\subsection{2D projection effect}
\label{2d_sec}
{\bf In Sect. \ref{df_sect} we describe how we calculate the FUV and EUV fluxes at the position of the candidate members using the projected distances between each member to the O stars. The projected distances are lower limits to the real distances between stars, and they lead to overestimating the UV fluxes experienced by the candidate members. Since we are studying the effects on the disk fraction, i.e. the ratio of the number of stars with disks over the total number of members, it is not trivial to deduce how the 2D projection of the real stellar distribution affects the correlation observed between disk fraction and incident UV flux. One way to infer how the trend in Fig. \ref{dfuv_img} would appear using the real 3D distances between stars is by simulating a large number of realistic morphologies of the association and calculating the disk fraction vs. FUV flux for each simulation. We want to verify whether possible 3D configurations of the association exist where the decrease of the disk fraction as a function of the incident UV flux would not be observed. \par
In order to simulate the 3D morphology of the association, we need to assign to each member an {\it elevation} ({\it z}, in parsec) from the 2D projection plane. The elevation can be either positive or negative. If Cyg~OB2 were a cluster and not an association, it would have been appropriate to adopt as an upper limit of $z$ a multiple (such as 5 times) of the cluster radius. Cyg~OB2 is, however, an association containing several subclusters and structures (e.g. \citealp{GuarcelloDWD2013}), and Cyg OB2 has been shown to have a substructured and fractal structure and is not completely mixed \citep{WrightPGD2014}, so that stars that are close together on the plane of the sky will also be close together along the line of sight. More realistic simulated 3D configurations can be obtained by considering its clumpy structure. \par
We first calculate the {\it Minimum Spanning Tree} (MST) of the members \citep{BarrowBS1985} using the $R$ statistics package \emph{nnclust}\footnote{http://cran.r-project.org/web/packages/nnclust/nnclust.pdf}. The MST is defined as the unique set of branches connecting all the points in a given data set with the minimum total length and not producing closed loops. This technique is typically used to select and extract clustered groups of points, and it has been used in several different studies related to stellar clusters (e.g. \citealt{GutermuthMMA2008}). In this case we do not attempt any selection of subclusters, but instead we simulate an isotropic distribution for the orientation of each branch of the MST with respect to the projection plane adopting a uniform probability distribution. We set an upper limit to the simulated elevation of each star equal to $z_{max}=\pm 20\,$pc, in order to avoid elevations ranging from 0 to $\pm \infty$. \par
We run 5000 simulations. Fig. \ref{3d_img} shows one of the 3D distributions that we obtain. A significant stellar population at large elevations apparent in Fig. \ref{3d_img} is found to be a common feature of all the simulated 3D configurations. However, the fall-off of the distribution towards the upper and lower box edges shows that out result is not unduly sensitive to the value of z$_{max}$ chosen. \par
For each simulation, we calculate the disk fraction variation as a function of the incident FUV flux, using the simulated 3D distances between the low mass members and the O stars. The decline of the disk fraction toward the O stars is always observed. In Fig. \ref{3d_histo} we show the median and 25\% and 75\% quartiles of the disk fractions as a function of UV flux for all the realizations. The difference between the 25\% and 75\% quantiles of the disk fraction shown in Fig. \ref{3d_histo} is similar to its error bars as shown in Fig. \ref{dfuv_img}, meaning that the 2D approximation has a very small impact on our study. In conclusion, there is no evidence that the variation of the disk fraction as a function of the incident UV flux observed in Cyg~OB2 is a consequence of the 2D projection.} \par
\subsection{The sequence of star formation}
\label{trig_sec}
\citet{Mamajek2009} describes the decline of disk fraction with cluster age as an exponential decay with an e-folding timescale of $2.5\,$Myrs. As these authors state, however, the observed decay strongly depends on the diagnostics adopted to select members with and without disks and on the type of disks included in the selection (i.e., active, passive, transitional, etc...). The disk selection in Cyg~OB2 has been performed with $JHK$, Spitzer/IRAC, and Spitzer/MIPS 24$\,\mu$m data, and it involves both stars with thick disks, candidate stars with transition, pre-transition, and low-mass (anemic) disks \citep{GuarcelloDWD2013}, classified according to their spectral energy distributions and infrared colors. \par
A decrease of disk fraction from $\sim40\%$ to $\sim18\%$ as we observe in Fig. \ref{dfuv_img} can be explained if the population close to the O stars is about $5\,$Myrs old, while that in the outer part of the association is younger than $2\,$Myrs. Such a chronology can be a consequence of a triggered star formation in the periphery of the association by the OB stars in the center, which is a plausible hypothesis in Cyg~OB2. In order to examine this possibility, it is necessary to estimate the age of the stars in the different bins in Fig. \ref{dfuv_img}. \par
A rigorous estimate of stellar age would require an accurate spectral classification of all the members, for which spectroscopic data are still not available. A less accurate method, based on multi-band photometry that is available for most of the stars in Cyg~OB2, consists in interpolating the position of each member in dereddened optical color-magnitude diagrams with the pre-main sequence isochrones from \citet{SiessDF2000} and the PARSEC stellar evolutionary model \citep{BressanMGS2012}\footnote{Downloaded from the CMD 2.7 web interface}. Since the presence of a circumstellar disk can alter the optical colors of the central star by adding a blue-excess due to the accretion process \citep{CalvetGullbring1998} or scattering of photospheric optical radiation into the line of sight \citep{GuarcelloDMP2010}, this method for evaluating stellar age can not be applied to the disk-bearing members but only to candidate class~III objects.\par
The optical $r$ vs. $r-i$ and $r-i$ vs. $i-z$ diagrams of all the X-ray sources with optical counterparts are shown in \citet{GuarcelloWDG2012}. Most of these sources are classified as members of Cyg~OB2, so these diagrams also show the loci of the disk-less members. Since the extinction in Cyg~OB2 changes drastically with position, in order to obtain a reliable estimate of the age of the disk-less members it is necessary to deredden their optical colors using an estimate of the individual extinction of each star, rather than some approximated average value. \citet{GuarcelloWDG2012} used the displacement along the reddening vector from the $A_V=0^m$ $2.5\,$Myrs isochrone in the $r-i$ vs. $i-z$ diagram to calculate the individual extinctions of the optical+X-ray sources in the central region of the association. In the $Chandra$ Cygnus~OB2 Legacy Survey we adopt the same approach but using the main sequence locus defined by \citep{CoveyISF07} rather than those from \citet{SiessDF2000} and the source classification provided by \citet{Kashyapinprep}, calculating ages and masses only for the association members. \par
Fig. \ref{age_img} shows the age distributions calculated with \citet{SiessDF2000} isochrones of the disk-less stars falling in each of the six bins of Fig. \ref{dfuv_img}. The main result is that the distribution of stellar ages for the stars falling in the first and last bin is similar, with a difference of median age of only 0.8$\,$Myrs. This age difference, as well as the difference observed among the stars falling the other bins, can not account for the observed difference of disk fraction.
A similar result is found using the PARSEC models (in this case the difference of median age is $0.6\,$Myrs). In both cases, only the stellar population of the sixth bin shows a wider distribution, with a population of younger stars falling in the active star forming regions identified by \citet{GuarcelloDWD2013}. This finding is in agreement with the lack of significant spatial variation of stellar age as found by \citet{WrightDM2015} for the OB population. \par
As a further test, we fit the $r$ and $r-i$ distributions of the populations of the six bins with a set of isochrones using the maximum likelihood $\tau^2$-minimization method \citep{NaylorJeffries2006,Naylor2009}. This method employs fitting of two-dimensional photometric data in a color-magnitude space, using a wide set of isochrones and also taking into account the effects of the binary population as well as observed photometric errors. Since the method does not calculate the extinction separately, we use the deredden colors as described above. We fit the $r_0$ vs. $r_0-i_0$ distribution of the stellar population of the six bins using all the available isochrones \citep{DAntonaMAzzitelli1997,BaraffeCAH1998,PallaStahler1999,SiessDF2000,Dotter2008,TognelliDP2012}, adopting a binary fraction of 0.5 and solar metallicity. Since we correct for the individual reddening, we do not apply any further systematic color error. Only when using the isochrones from \citet{Dotter2008} we observe a likely inside-out age gradient with a difference of 3$\,$Myrs between the inner and outer stellar populations. With the remaining six sets of isochrones we use, no such a trend is observed, with the age difference between the stellar populations of adjacent bins being typically smaller than 1$\,$Myr. Assuming that the disk-bearing and disk-less population are almost coeval, we can conclude that the observed decline of disk fraction toward the O stars in Cyg~OB2 is not a consequence of an inside-out sequence of star formation.
\subsection{Completeness}
\label{comp_sec}
Any not uniform sensitivity of our survey across the field might impact the spatial variation of disk fraction that we observe. For instance, the sensitivity across the ACIS-I detector is known to decrease with the displacement from the center. Since pre-main sequence stars are in general in the saturated activity regime \citep{PreibischFeigelson2005}, which means that the X-ray luminosity in these stars is independent from stellar rotation but it scales with bolometric luminosity, which itself is dependent on stellar mass and age, single ACIS-I observations of young clusters will be deeper in stellar mass at small off-axis angles than in the outer part of the field of view. {\bf Since there are indications that less massive stars dissipate their disks more slowly than more massive stars \citep{CarpenterMHM2006ApJ,CarpenterMHM2009ApJ,LuhmanMamajek2012ApJ}}, and also accounting for the fact that disk-bearing stars are in general less X-ray bright than disk-less stars \citep{PreibischKFF2005,FlaccomioMS2003}, this can affect the observed spatial variation of disk fraction. However, this effect would mainly tend to increase the observed disk fraction toward the center of a single field. Even if the $Chandra$ Cygnus OB2 Legacy Survey was designed in order to reduce the impact of not uniform sensitivity \citep{Drakeinprep,WrightDG2015}, it is important to verify whether this issue may impact our results.\par
Another important effect can be the loss of sensitivity in the optical and infrared images close to very bright stars, such as most of the O stars in Cyg~OB2. The bright wings of the PSF of such stars can result in brighter background and larger difficulty in detecting the emission from nearby faint stars. It is not easy to predict the impact that this problem has in our study, since observations at different wavelength are affected in different ways. \par
About 95\% of the candidate cluster members have $JHK$ counterparts, in general from UKIDSS. The simplest way to verify whether a not uniform sensitivity of the data has affected the result shown in Fig. \ref{dfuv_img} is to derive the variation of the disk fraction as a function of the FUV flux with different cuts in $J$ magnitude. Considering only members with $J\leq 17^m$, $J\leq 16^m$, or $J\leq 15^m$ the disk fraction declines from 35\% to 19\% in the former two cases and from 44\% to 24\% in the latter. We run further tests to verify whether other completeness issues may affect our results. We observe a smooth decline of the disk fraction toward the center even considering only candidate class~III sources brighter in X-rays than the 25\%, 33\%, and 50\% quantile of the X-ray fluxes distribution, with a total decline of the disk fraction from the 46\% to 23\%, 49\% to 25\%, and 62\% to 33\% respectively. The masses of the candidate cluster members detected in X-rays are estimated from the dereddened optical and infrared color-magnitude diagrams \citep{Kashyapinprep}. Even if the presence of the disk can affect the optical colors observed from the class~II sources, and thus the estimate of the individual extinction and mass, we adopt the same procedure to estimate the masses of the disk-bearing stars not detected in X-rays, and calculate the disk fraction vs. the incident FUV flux adopting two cuts in stellar mass: 0.4$\,$M$_{\odot}$ and 0.7$\,$M$_{\odot}$. The resulting spatial variation of the disk fraction is still characterized by a significant decrease toward the center of the association, even if with a smaller total disk fraction given the small number of stars for which masses can be calculated: from 21\% to 15\% and from 25\% to 17\% respectively. \par
{\bf As a last test, disk lifetimes may depend on the mass of the central star \citep[e.g. ][]{CarpenterMHM2006ApJ,CarpenterMHM2009ApJ,LuhmanMamajek2012ApJ}}. If the mass content of the central cluster is different than that of the population in the outskirt of our field, this would affect the spatial variation of the disk fraction. However, the disk fraction calculated only for members with M$\leq 1\,$M$_{\odot}$ declines from 27\% to 13\%, while for members with M$\leq 0.6\,$M$_{\odot}$ from 28\% to 15\%. Analogously, the disk fraction for members with $J\geq15.4/m$ or $J\geq16.1^m$ (i.e. the expected $J$ magnitudes of a 1$\,$M$_{\odot}$ and 0.6$\,$M$_{\odot}$ stars, respectively, at 1400$\,$pc of distance and extinguished by A$_V$=4.5$^m$) declines from 39\% to 18.5\% and from 44\% to 21\%, respectively.
\section{Photoevaporation vs. disruption due to close encounters}
\label{map_sec}
In this section we compare the effects on disk evolution due to UV radiation and close encounters adopting two different approaches: the analysis of disk fraction maps with a higher spatial resolution than employed in Fig. \ref{dfuv_img}; and estimating the local stellar density around the position of each member simulated in the 15000 simulations of the 3D configuration of the association as discussed in Sect. \ref{2d_sec}.
\subsection{Disk fraction maps}
\label{dfmaps_sec}
Fig. \ref{dfmapirimg} shows a map of the disk fraction calculated in a grid of $15\times 15$ bins and encoded with red tones. The value of the disk fraction in those bins where $\sigma_{DF}/DF<0.25$ is overplotted ($DF$ indicates the disk fraction). \par
In Fig. \ref{dfmapirimg} the contours mark the emission intensity at $8.0\mu m$ as inferred from Spitzer data. This map allows us to identify those bins where local peaks of the disk fraction correspond with nebular structures that have been identified as active star forming regions in \citet{GuarcelloDWD2013}, such as the DR18 cloud at $\alpha \sim 308.8,\, \delta \sim 41.2$, the {\it Globule 1} at $\alpha \sim 308.0,\, \delta \sim 41.2$ \citep{SchneiderBSJ2006} and the {\it bright rimmed clouds} in the north-east from the central cluster\footnote{Approximately at $\alpha \sim 308.6,\, \delta \sim 41.5$, $\alpha \sim 308.7,\, \delta \sim 41.55$, and $\alpha \sim 308.8,\, \delta \sim 41.6$} \citep{GuarcelloDWD2013}. On average, the disk fraction in these structures ranges from $\sim10\%$ to $\sim20\%$ larger than that in some of the surrounding bins. This is compatible with the embedded stellar population being $1.5-2\,$Myrs younger than the surrounding stars \citep{Mamajek2009}. \par
In the left panel of Fig. \ref{dfmapdensimg} the spatial variation of disk fraction is compared with the stellar surface density. Contours mark the area where the local stellar surface density is equal to typical values of in clusters within 2$\,$kpc from the Sun. The adopted limits are the 25\% and 95\% quantiles of the distribution of surface densities of the clusters within $2\,$kpc from the Sun, as compiled by \citet{LadaLada2003} and \citet{PorrasCAF2003}: $18\,$N$\,$pc$^{-2}$ and $22\,$N$\,$pc$^{-2}$, respectively. The limit corresponding to the densest region in Fig. \ref{dfmapdensimg} is $33\,$N$\,$pc$^{-2}$, which is the slope of the $N$ vs. $R$ relation for these clusters (with $R$ being the cluster radius) as found by \citet{AdamsPFM2006}. These low densities are in agreement with the finding of \citet{WrightPGD2014} who have demonstrated that Cyg~OB2 has always been an association, characterized by modest stellar density such as that we observe. The right panel shows clear decrease of the disk fraction as a function of the stellar surface density measured in those bins with low relative error in the disk fraction. \par
In the left panels of Figg. \ref{dfmapfuvimg} and \ref{dfmapeuvimg} the contours mark the local FUV and EUV fluxes, with the values fixed to the limits of the bins in Fig. \ref{dfuv_img}. By a simple visual inspection of these images, the correlation between the local disk fraction and the local UV field (shown in the right panels of both figures for those bins with low relative error in the disk fraction) is evident. In most of the bins containing one or two O stars, which are characterized by $F_{FUV} > 10^4\,$G$_0$ or log$\left( F_{EUV} \right) > 11.5$ in units of photons/s/cm$^{2}$, the disk fraction is about $15\%\, -\, 20\%$ smaller than that in surrounding bins not containing O stars and characterized by less intense UV fields. Considering that the stellar population of these regions is almost coeval, this gives an estimate of the fraction of disks which have been destroyed by the UV feedback in Cyg~OB2 to date. \par
The center of the association, hosting the two central groups of O stars and the elongated group in the north-west direction, is characterized by an almost constant disk fraction ($15\%-18\%$), high surface stellar density ($>33\,$N/pc$^{2}$), and intense UV field ($F_{FUV}>10^4\,$G$_0$ and log$\left( F_{EUV} \right) > 11.5$). This does not allow us to use the data from the central region to compare the efficiency of photoevaporation and collisional destruction. For this aim, the values of the disk fraction in bins hosting isolated O stars are more useful. For instance, in the north-east the region around the group of 7 O stars (marked with a black box in the disk fraction maps) is characterized by intense UV fluxes, but low stellar surface density ranging between $5\,$N/pc$^{2}$ and $17\,$N/pc$^{2}$, and the disk fraction of the whole area is about $15\%$. This is similar to the center of the association in our grid, and significantly larger than the disk fraction observed southward and eastward (also northward, but in this direction the presence of the cloud cavity front and a trunk with ongoing star formation surely affects the disk fraction). Analogously, the disk fractions in the two bins hosting the three O stars at approximately $\alpha\sim307.9$ $\delta\sim41.2$ and $\delta\sim41.3$ (spectral classes O8.5III, O8.5I, and O9.5I), marked with a black box in the disk fraction maps, are 18\% and 26\%. These values are smaller than those in most of the surrounding bins even if the surface stellar density is similar. \par
The analysis of the disk fraction maps may suggest a greater impact on disk dissipation timescale from the feedback from O stars rather than from close encounters. In fact, this result is only marginally significant and far to be final. However, the spatial variation of disk fraction correlates in a similar way with the local UV field and the stellar surface density, as shown in the Figg. \ref{dfmapdensimg}, \ref{dfmapfuvimg}, and \ref{dfmapeuvimg}, which is not surprising given the strong correlation observed between the local stellar density and the median value of the UV field in these bins (Fig. \ref{corr_img}). Given the difficulty in isolating and comparing regions with different stellar density and similar UV radiation field with regions with similar density and different UV radiation field with good statistic, the analysis described in this section provides only a marginal evidence that disks photoevaporation is a more important mechanism than close encounters in the evolution of disks. \par
\subsection{Simulated local stellar density}
\label{density_sec}
In Sect. \ref{2d_sec} we describe the 5000 simulated 3D configurations of the association realized with three different approaches. These simulations allow us to estimate the local stellar density per parsec$^{-3}$ at the position of each member, and to use these estimates to infer the expected rate of stellar encounters. \par
Given that Cygnus~OB2 is an association by some degree of subclustering, the stellar density in the 3D simulations derived in 3D binning or at increasing radial distances from a nominal center is not useful to characterize the local stellar density experienced by each member. A better estimate can be obtained from the method of \citet{WhitworthBTW1995} and devised for SPH simulations. Considering the $j^{th}$ member of the association whose position is given by its celestial coordinates and the simulated elevations $z$, the local stellar density can be given by $\delta_j=h_j^{-3}\times \sum_{i} W_{ij}$, with $i$ running over the other members, $h_j$ being the smoothing length associated with the $j^{th}$ member, and $W$ an appropriate kernel function. The smoothing length is defined as the radius of the sphere containing 50 members and centered on the $j^{th}$ member, while $W_{ij}$ is defined as:
\begin{displaymath}
W_{ij}(s) = \frac{1}{\pi}\left\{
\begin{array}{lr}
1-\frac{3s^2}{2}+\frac{3s^3}{4} & 0\leq s < 1\\
\frac{\left(2-s \right)^3}{4} & 1\leq s < 2\\
0 & s > 2
\end{array}
\right.
\end{displaymath}
{\bf with $s$ being $|r_i-r_j|/h_j$. The average distributions of local stellar densities obtained with the simulations described in Sect. \ref{2d_sec} are shown in Fig. \ref{meddens_img}. The number of members experiencing a local stellar density larger than 200 stars per parsec$^{-3}$ is shown in the last bin. In the simulated 3D stellar configurations there is a significant number (i.e. about 300) of stars with such a large surrounding stellar density. Only for very few members in the simulated configurations do the local stellar densities exceed the values typically observed in open clusters (i.e. from tens to a few hundreds of stars per parsec$^{-3}$).} \par
The simulated local stellar densities experienced by each association member allow us to make estimates of the chances of close encounters in a given time. For instance, using the calculations presented in \citet{ClarkePringle1991} and adopting a typical disk size of 100$\,$AU, we estimate that densities of $\sim$100 stars per cubic parsec correspond to about 1\% probability of an encounter in 1 Myr. {\bf In the simulated 3D configurations there is a significant fraction of members (14.6\%) surrounded by such a high local stellar density, but even in those cases the 1\% chance of close encounters in 1~Myr is negligible in the context of the disk sample in Cyg~OB2 as a whole.}
{\bf As a further test, we define a fixed grid in the simulated 3D space, with 10 bins per axis, for a total of 1000 cells. For each of the 5000 simulated 3D configurations, we calculate the disk fraction and total stellar density. To infer the total stellar density (i.e. across the entire mass spectrum), we adopt the following procedure. In our X-ray catalog of cluster members, we have 439 stars more massive than 1.5$\,$M$_{\odot}$, 1292 between 1.5$\,$M$_{\odot}$ and 0.7$\,$M$_{\odot}$, and 3098 less massive than 0.7$\,$M$_{\odot}$ \citep{WrightDGA2014,Kashyapinprep}. Using the completeness as a function of stellar mass calculated by \citet{WrightDG2015}, we can predict a total number of members of 488, 1615, and 8551 stars in these three mass intervals, respectively. In order to estimate the number of stars at about 0.1$\,$M$_{\odot}$, we observe that following the IMF defined by \citet{Kroupa01}, the expected population with such a mass is about ten times larger than that with about 0.5$\,$M$_{\odot}$, which in our case means about 85510 stars. Using these numbers, we predict a total population of 96164 stars in Cyg~OB2 down to 0.1$\,$M$_{\odot}$. Incidentally, we note that since the average stellar mass adopting the \citet{Kroupa01} IMF is 0.2$\,$M$_{\odot}$, the total stellar population of Cyg~OB2 we estimate implies a total mass of the association which is very similar to that predicted in existing studies \citep[e.g.][]{WrightDM2015}. To estimate the total stellar density in each cell, we make the assumption that the fraction of members falling in each spatial cell does not depend on stellar mass, i.e. we multiply the fraction of the members falling in each spatial cell by the estimated total number of stars. With this calculation, the estimated stellar density in the 3D grid never exceeds 90 stars per cubic parsec. Such a stellar density is too low to result in significant destructive feedback on disks evolution from the gravitational interaction between members. For instance, \citet{SteinhausenPfalzner2014AA} estimated the number of disks destroyed by close encounters in 2$\,$Myrs in clusters with different stellar density, and no significant effects are observed below stellar densities of 3000 stars per parsec$^3$. Similarly, in the simulations by \citet{VinckeBP2015}, after 5$\,$Myrs in environments with stellar densities of about 90 stars per parsec$^3$ no disks have been destroyed by close encounters below 10$\,$AU, while the fraction of disks affected at radii $\geq100\,$AU from the central star (which is a region of disks that we can not probe with our data) goes from about 10\% to about 17\% with the stellar density increasing from 15 to 90 stars per cubic parsec. In conclusion, the analysis of the simulated 3D stellar distributions of Cyg~OB2 and the comparison between the effects of the local UV field and stellar density indicates that the evolution of protoplanetary disks is impacted by photoevaporation, and that erosion through collisions plays only a minor role.} \par
\section{Discussion}
\label{disc_sec}
\subsection{Is Cygnus~OB2 a safe environment for disk survival?}
\label{safe_sec}
Externally induced disk photoevaporation has been proven to expedite the dissipation of circumstellar disks. The extraordinary HST optical images of the evaporating proplyds (i.e. protoplanetary disks) in the Trapezium Cluster in Orion \citep{BallySDJ1998} clearly show the impact that photoevaporation induced by nearby O stars (in this case the O6V star $\Theta^1$ Orionis) has on the evolution and morphology of nearby disks. The cometary shape of these evaporating proplyds has been studied by \citet{JohnstoneHB1998} in terms of a neutral flow of gas evaporating from the disk under the influence of the FUV radiation. This gas is then ionized by the EUV radiation forming an ionization front which lies at a distance from the disk surface that depends on several factors, mainly the intensity of the incident UV flux and the optical depth of the evaporating column. The disk mass loss rate of these proplyds has been predicted to range between $\sim10^{-7},\, \sim10^{-8}\,$M$_{\odot}$/yr \citep{StorzerHollenbach1999}, which has been later confirmed by spectroscopic observations and analysis of optical emission lines \citep{HenneyODell1999}. \par
In the Trapezium Cluster photoevaporation is driven by the FUV radiation. When the ionization front lie on the disk surface, however, the EUV radiation can directly ionize the gas in the disk, inducing a more intense mass loss. \citet{StorzerHollenbach1999} have studied the range of intensities of the incident FUV flux produced by the O stars in the Trapezium which results in a photoevaporative flow which is dense enough to absorb the incident EUV photons, i.e. in which the photoevaporation is FUV dominated. With the typical EUV/FUV flux ratio emitted by O stars, they found that the FUV-dominated region occurs for FUV incident fluxes in the range $10^5\,$G$_0 \lesssim F_{FUV} \lesssim 5\times 10^7\,$G$_0$. The lower limit is dictated by the fact that at less intense FUV fluxes the wind is transparent to incident EUV photons that can directly ionize the disk surface, while the upper limit corresponds to distances from the ionizing source with such intense EUV fields that the ionization front coincides with the disk surface. Outside this interval, photoevaporation is EUV-dominated. \par
In Cyg~OB2, the disk fraction in the region irradiated by an FUV flux within the FUV-dominated range is 18.3\%, typical of the very inner area of the association as shown in the disk fraction maps. The distance of these stars from the closest O or WR star (independently from its spectral type) is about 0.4$\,$pc. This can be compared with the decline of the disk fraction with the distance from the closest O star shown in Fig. \ref{histodist_img}. On average, the regions characterized by a disk fraction $\sim20\%$ are less than 1 parsec away from the closest O star. \par
Considering only the photoevaporation induced by FUV photons, this could lead to the conclusion that even in massive associations such as Cygnus~OB2, the externally induced disk photoevaporation is important only nearby the O stars. However, we observe a smooth decline of the disk fraction over the entire region toward the center of the association. To explain this,{\bf we note that a number of simulations \citep[e.g.][]{Clarke2007MNRAS,AndresonAC2013ApJ,FacchiniCB2016MNRAS} suggest that even with incident FUV radiation in the range $3000\,$G$_0 \lesssim F_{FUV} \lesssim 30000\,$G$_0$ the lifetime of protoplanetary disks can be reduced to less than 2$\,$Myrs, unless their viscosity is exceptionally low}. We can also verify whether in regions with $F_{FUV}<10^5\,G_0$ the EUV radiation can drive the photoevaporation of the disks. It is possible to estimate the mass loss rate induced by EUV radiation using the equation derived by \citet{Adams2010}:
\begin{dmath}
\dot{M}\approx 9\times 10^{-8}\left(\frac{L_{EUV}}{10^{49}\,s^{-1}}\right)^{1/2}\left(\frac{10^{17}\,cm}{d}\right) \times \left(\frac{r_d}{30\,AU}\right)^{3/2}
\label{mdot_eq}
\end{dmath}
where $L_{EUV}$ is the EUV luminosity of the ionizing sources (in photons/s), $d$ is their distance from the disk, and $r_d$ is the disk radius. We calculate the mass-loss rates due to the incident EUV flux emitted by the O stars of Cyg~OB2 adopting $r_d=100\,$AU. The result is shown in Fig. \ref{mdot_img}, where we plot over the disk fraction map the contours of the expected mass loss rate induced by the local EUV radiation field. The contour labels show the log$\left(\dot{M} \right)$ values in units of M$_{\odot}$/yrs. The mass loss rates induced by the extreme O population in Cyg~OB2 ranges from $1.5\times10^{-8}\,$M$_{\odot}$/yr to $3.9\times10^{-7}\,$M$_{\odot}$/yr across the entire field. Such values are capable of dissipating a disk with a typical mass of $0.05\,$M$_{\odot }$ in 3.3 and 0.1 Myrs, respectively. For comparison, a mass loss rate larger than $1.5\times10^{-8}\,$M$_{\odot}$/yr can be induced by $\Theta^1$ Ori only within $0.44\,$pc. This distance is just 0.03 times the projected radius of our survey area. The question that must be addressed now is: why we still observe stars with disks in Cyg~OB2 given this intense mass loss rate induced by the EUV radiation across the entire field? \par
One possibility is that we are overestimating the incident EUV flux because of the use of the projected distances rather than the real ones. However, in order to dissipate the disks in Cyg~OB2 in a timescale comparable with the association lifetime, the mass loss rate must be about 5 times smaller. This corresponds to an increase in distance $d$ of a factor 5 for the same emitted EUV flux, resulting to unreasonable distances between the outer part of the association and its center. \par
Such intense mass loss rates as those we calculate can also be a consequence of the adopted stationary value for $r_d$. Given that the surface density in circumstellar disks decreases with increasing distance from the central star, externally induced photoevaporation, which results in an almost uniform mass loss rate across the disk, dissipates the disk from outside inward \citep{JohnstoneHB1998}. With time, the disk radius shrinks down to the gravitational radius\footnote{The gravitational radius is the maximum distance from the central star where the evaporating gas is gravitationally bound to the system.}. After reducing the disk outer radius to, for instance, 35$\,$AU, the mass loss rate has been reduced by a factor of $\sim$5, following Eq. \ref{mdot_eq}. It must be noted that a such small disk can still produce detectable NIR excesses. \par
Another likely explanation is that Equation \ref{mdot_eq} does not account for the attenuation of the EUV radiation. Two contributions of such absorption are likely to play an important role. First, the calculation of the mass-loss rate in Eq. \ref{mdot_eq} accounts only for the wind produced by the externally induced photoevaporation, ignoring any contribution from the photoevaporation induced by the central star itself (by both X-ray and UV photons). As discussed, this process is thought to be important in the evolution of disks and to be the main process leading to the formation of pre-transition and transition disks \citep[e.g.:][]{ClarkeGS2001,AlexanderCP2004}. However, this contribution is expected to be not dependent on quantities which decrease with the distance from the O stars, such as disks and central stars properties. It is more likely that the presence of a large population of stars with disks in the outer part of the association is instead possible thanks to the absorption of the UV radiation by intracluster material. There is evidence that the inner cluster of Cyg~OB2 is almost clear of gas \citep{SchneiderBSJ2006}. This is not surprising given that massive stars in clusters usually create cavities in the parental cloud in about $3\, -\, 5\,$Myr (e.g. \citealp{AllenMGM2007}). However, H$\alpha$ images of the center of Cyg~OB2 \citep{DrewGIA2005,GuarcelloDWD2013} shows diffuse H$_{\alpha}$ emission around the O stars that must be emitted by intracluster gas. There is also evidence for the presence of diffuse dust emission and dense clumps mainly in the outer part of the association (see Fig. \ref{field_img} and \citealp{GuarcelloDWD2013}). There is, then, a column of absorbing material that partially shields the stars with disks in the outskirt of the association from the UV radiation emitted by the O members, and this extinction is expected to increase with the distance between stars with disks and the O stars. For instance, a decrease of the mass loss rate by a factor of 5 is enough for having disks dissipation timescale comparable with the age of the association. Since, in Eq. \ref{mdot_eq}, $\dot{M}\propto \sqrt{F_{EUV}}$, this requires a decrease by a factor of 25 of the incident UV flux. Dust absorption can be accounted using the extinction law of \citet{CardelliCM1989}: For the EUV radiation ($\lambda<0.125\,\mu m$) a similar absorption can be achieved with an extinction $A_V=1^m$, while in FUV ($\lambda<0.18\,\mu m$) with $A_V=1.4^m$. Using the relation $N_H/A_V=1.8\times 10^{21}\,$atoms/cm$^{2}$/mag, and assuming a constant column density along a distance of 2 or 3 pc, this extinction corresponds to a particle density of $\sim300\,$cm$^{-3}$, which is still typical of giant molecular clouds \citep{SolomonRBY1987}, and it looks realistic given the presence of intracluster material mainly in the outskirt of the association. Gas absorption is even more efficient. We calculate that residual intracluster gas with column density of $\sim10^{19}\,$atoms/cm$^{2}$ is opaque to EUV radiation with $\lambda<912\,$\AA{}: a column density of $1.8\times10^{18}\,$atoms/cm$^{2}$ corresponds to a transmittance (the fraction of incident flux which is not absorbed by the gas) of $\sim0.004$, becoming about two orders of magnitude smaller by increasing the gas column density of a factor of two. Considering that the hydrogen (H+H$_2$) column density that we can calculate from Herschel data is larger ($\sim10^{21}\,$atoms/cm$^{2}$), even if part of this comes from the Rift in the foreground, the absorption of the EUV radiation by residual intracluster material is then the most reliable hypothesis to explain the presence of a larger fraction of members with disk in the outer part of the association with respect to the center. \par
\subsection{Cygnus~OB2 in context}
\label{comp_sec}
In the previous sections, we have found evidence that in Cygnus~OB2 disk evolution has been seriously affected by the surrounding environment. \par
Similar effects have been studied in other clusters with different properties. In their series of papers, \citet{GuarcelloPMD2007, GuarcelloMDP2009, GuarcelloMPP2010} have found that the spatial variation of disk fraction in NGC~6611 is constant across the cluster and equal to $\sim40\%$, except for a sudden decrease down to $26\%$ within $1\,$pc from the massive stars (O plus early B). NGC~6611 is an intermediately massive cluster hosting 54 OB stars, among which are 13 O stars mainly concentrated in a cavity $2.2\,$pc in diameter \citep{HillenbrandMSM1993}. The most massive star in this cluster is W205 with a mass of $75-80\, M_{\odot}$ and a spectral class O3-O5V \citep{EvansSLL2005}. NGC~6611 is younger than Cyg~OB2, having a median age of $\sim1\,$Myr \citep{GuarcelloPMD2007}. In Fig. \ref{6611_img} we recalculate the variation of the disk fraction in NGC~6611 as a function of the incident FUV flux (in terms of G$_0$) as we did in this paper for Cygnus~OB2. The difference between the disk fraction in the first and fourth bins is the only significant difference we observe, with the disk fraction going from $\sim41\%$ to $\sim31\%$. Compared to Cyg~OB2, the disk fractions in the outer population of the two regions are similar, while the main difference is observed in the central parts, close to the O stars. The differences between the two regions are likely a consequence of the different ages of the two stellar populations, with a combination of normal disk evolution and induced photoevaporation\footnote{The hypothesis of an inside-out star formation chronology in NGC~6611 has been discarded by \citet{GuarcelloMPP2010}}. However, in the recent paper of \citet{RichertFGK2015} no evidence of rapid disk dissipation nearby the massive stars in NGC~6611 is found. These authors claim that the different result is due to the different membership selection, since their selection of stars with excesses in the NIR bands is based on UKIDSS data while that in \citet{GuarcelloMDP2009} used 2MASS data, allowing them to observe less massive stars and disks. Since the decrease of the disk fraction in the center of NGC~6611 was observed also by \citet{GuarcelloMPP2010}, in which stars with excesses in $JHK$ bands were also selected using UKIDSS data, it is very likely that the disparate results are due to the different approaches both in selecting cluster members and in the analysis of the spatial variation of disk fraction. This simply suggests that the external feedback in disk evolution in clusters such as NGC~6611 is still an open question which requires further analysis. \par
A similar situation has been found in the coeval cluster Pismis~24 by \citet{FangBKH2012}. This $\sim1\,$Myr old cluster hosts dozens of OB stars, with some very massive stars such as Pis24-1 (O3I) and Pis24-17 (O3.5III) \citep{MasseyDW2001}, together with a rich low mass population. \citet{FangBKH2012} found that in this cluster the disk fraction is constant at $\sim36\%$ across the field, decreasing down to $19\%$ at a distance of $0.5\,$pc from the most massive members. The stars within 0.5$\,$pc from these two massive stars are irradiated by a FUV flux more intense than 21000$\,$G$_0$, which is similar to the FUV flux incident on the closest stars to the O stars in Cyg~OB2 and NGC~6611. The scenario resulting from these studies is similar: in intermediately massive clusters after $1\,$Myr the environment feedback on disk evolution is important only in the immediate proximity of the O and B stars, and generally in the cluster core, while on this timescale no effects are experienced by the disks in the outer region. \par
Even for slightly older intermediately massive clusters the situation is similar, as demonstrated by the spatial variation of disk fraction in NGC~2244. This cluster is $2\, -\, 3\,$Myrs old \citep{ParkSung2002} and it hosts 7 O stars and a conspicuous low mass population. In NGC~2244 the average disk fraction is $\sim40\%$, and constant across the cluster but with a rapid decrease down to $\sim23\%$ at distances from the O stars $<0.5\,$pc \citep{BalogMRS2007}. \par
Another important contribute in this context are the $N$-body simulations of the evolution of protoplanetary disks in the Orion Nebula Cluster by \citet{ScallyClarke2001}. They followed the evolution of the disks with radii of 100$\,$AU and 10$\,$AU for $10^7\,$years in a cluster with 4000 members, considering the feedback provided by both the UV radiation and stellar encounters separately. They have found that photoevaporation removes between 0.01$\,$M$_{\odot}$ and 1$\,$M$_{\odot}$ from the larger (100$\,$AU) disks, while the 10$\,$AU disks are not seriously affected. Similarly, they estimated that only 4\% of the stars have encounters closer than 1000$\,$AU. These studies confirms then that important effects on disks evolution in such environments can be experienced only in the dense cluster core, characterized by intense UV field, even for timescales larger than $1\,$Myr. \par
In clusters with small populations of massive stars and low stellar density no important feedback on the evolution of circumstellar disks is expected. The local UV field can reach critical intensities only at very small distances from the few massive members. Besides, as noted in Sect. \ref{intro}, in clusters hosting a few hundred members the chances of close encounters with small ($\sim$100$\, -\, 200\,$AU) impact parameters is $\leq1\%$ in $1\,$Myr. This is confirmed by studies of the spatial variation of the disk fraction in IC~1795 \citep{RoccatagliataBHG2011}. This $\sim3\,$Myr old cluster hosts two O stars (O6.5V and O9.7I), and no variation of disk fraction has been observed toward the position of these ionizing sources. This can be understood in terms of the FUV flux produced by these sources. For instance, the median FUV flux experienced by the stars in Cyg~OB2 lying in the first bin in Fig. \ref{dfuv_img} (i.e. the one with the highest UV fluxes, containing more than 700 candidate members) is $\sim54100\,$G$_0$. The two O stars in IC~1795 only produce such intense FUV flux within 0.2$\,$pc, where 9/525 candidate members lie. In their search for accreting objects in the young cluster IC~1396, around the O6.5V star HD~206267, \citet{BarentsenVDG2011} found that mass accretion rates, number of accretors, and intensity of the infrared excesses increase marginally with the distance from this massive star, interpreting this result as a consequence of triggered star formation rather than induced photoevaporation. \par
All these results taken together indicate that most of the environments where stars formation occurs in our Galaxy are safe for disk evolution and planet formation: in sparse clouds and low mass clusters, as well as in the outer regions of intermediately massive clusters, the local FUV and EUV fluxes and the local stellar densities never exceed values that would cause a rapid erosion of circumstellar disks, potentially halting or disrupting the planet formation process. Important feedback is expected and observed only in the core of intermediately massive clusters, at distances of $\ll 1\,$pc from the massive stars that are usually found in the cluster center. \par
There is strong evidence showing that even our Sun and Solar System formed in the outer part of an intermediately massive cluster. \citet{Adams2010} noted that the orbits of the planets in the Solar System require no close encounters with $b\leq90\,$AU, indicating that the parental cluster of the Sun had less than $10^4$ members; at the same time, the orbits of some trans-neptunian objects such as Sedna require at least one encounter with $200\,$AU$\leq b \leq 300\,$AU; the paucity of gas in the trans-Neptunian objects and the presence of giant planets require the incident FUV flux to have been in the range $2000\,$G$_0 \leq F_{FUV} \leq 10^4\,$G$_0$; the presence in the Solar System today of short-lived radio nuclei with half-life $<$ few Myrs (such as $^{26}$Al and $^{30}$Fe) requires the presence of O stars. All these factors indicate that the parental cluster of the Sun was an intermediately massive cluster hosting a few thousand members, and that the Sun formed in its periphery, and eventually moved toward its center \citep{Adams2010}. The outer regions of clusters like Pismis~24 and NGC~6611 are good examples of such star forming regions, which are quite common in our Galaxy. \par
In this paper we complete this picture by studying the feedback provided in very massive associations that, even if rare in our Galaxy, harbor tens of thousands of stars. Our study indicates that in environments such as Cygnus~OB2 the local values of the EUV and FUV radiation field are intense enough to externally induce the photoevaporation of disks even at large distances from the O stars ($\leq 10\,$pc), and likely exceeding the influence of close encounters, given the moderate stellar density observed across the region. Using the estimate from existing and detailed models of the photoevaporation process induced by nearby O stars, photoevaporation is driven by the FUV radiation in regions characterized by FUV fluxes $\ge 10^5\,$G$_0$, i.e. at distances $\leq 0.5\,$pc from the O stars, and then by the EUV photons in the rest of the association. The induced mass loss rate from the EUV flux produced by the massive population in Cyg~OB2 would be intense enough to completely dissipate the disks across the entire association very quickly. The presence to date of a large disks population is likely a consequence of the column of extinguishing material between the low-mass and the O members of Cyg~OB2, that absorbs the EUV radiation and then shields the disks. This likely plays also a role in shaping the observed smooth decline of the disk fraction with the distance from the center of the association. Our data suggest that the overall effect on the timescale corresponding to the age of Cyg~OB2 members (i.e. between $3\,$Myrs and $5\,$Myrs) of the environmental feedback on disks evolution is the decrease of the disk fraction by about $20\%$. \par
While disk erosion is likely to play a role in the planet formation process, the effect it has is not yet known. The observation of photoevaporating disks at $8.0\mu m$ in NGC~2244 \citep{BalogRSM2006} indicates that small dust grains remain trapped in the photoevaporating flow, reducing the reservoir of solid material available for the formation of planetesimals in the disk. On the other hand \citet{OwenEC2011} show that only a small fraction of the existing small dust remain trapped in the photoevaporative flow, and the simulations presented by \citet{ThroopBally2005} indicate that in the midplane of gas-depleted disks gravitational instabilities occur more easily, allowing a rapid formation of planetesimals with dimensions from centimeters to kilometers. More work is necessary to better assess the true impact of rapid photoevaporation on planet formation. \par
\section{Summary and conclusions}
\label{conclusions}
In this paper we study how the environment in Cygnus~OB2 affects the dissipation timescale of protoplanetary disks, by analyzing the spatial variation of the disk fraction across the association. We use the selection of members with and without disks provided by other publications related to the $Chandra$ Cygnus~OB2 Legacy Survey. We correlate the local values of disk fraction across the association with the local values of EUV and FUV fluxes, and stellar surface density, observing a smooth decline of the disk fraction from $\sim40\%$ to $\sim18\%$ with increasing UV fluxes and stellar density. \par
We rule out the hypothesis that the observed decline of disk fraction across the association is a consequence of an inside-out triggering of star formation, a 2D projection effect, or non uniform sensitivity of our data. We also briefly discuss the existing evidence supporting the hypothesis that the association is not dynamically evolved and never been much denser than we observe today. \par
We interpret the result as a consequence of the rapid erosion of the disks in Cyg~OB2 by the incident UV radiation. In particular, using models of externally induced photoevaporation, we conclude that FUV radiation dominates the process in the region within $\sim 0.5\,$pc from the O stars. However, the EUV field is intense enough to induce the dissipation of the disks in few Myrs across the entire association. The presence of a significant fraction of stars with disks in the outer regions and the observed smooth decline of the disk fraction with the intensity of the incident UV flux is explained in terms of an absorption of the EUV radiation by the material still associated with the cloud whose efficiency increases at increasing distances from the O stars. We find the destruction of disks by close stellar encounters to be rare, such that only of the order of 1\% or fewer disks in the regions of the association characterized by a stellar density larger than 100 stars/pc$^3$ per Myr. \par
Finally, we consider similar studies published so far on clusters with different size and age. Only the core of intermediately massive clusters are characterized by UV fluxes and stellar densities that have a strong impact on disks evolution, while the outer regions are relatively safe environments for disk evolution. Analogously, massive associations are harsh environments only if intracluster material can not provide efficient shielding from the ionizing radiation emitted by the massive stars. These results suggest that a large variety of star forming environments in our Galaxy are safe environments for disk evolution and planet formation (i.e.: sparse clouds, low mass clusters and the outer regions of intermediately massive clusters).
\acknowledgments
We thank the referee for his/hers suggestions and help improving our manuscript. MGG acknowledges the grant PRIN-INAF 2012 (P.I. E. Flaccomio). NJW acknowledges a Royal Astronomical Society Research Fellowship. They also have been supported by the Chandra grant GO0-11040X during the course of this work. JJD and VK were supported by NASA contract NAS8-03060 to the Chandra X-ray Center, and thank the Director, B. Wilkes, and the CXC science team for advice and support.
\addcontentsline{toc}{section}{\bf Bibliografia}
\bibliographystyle{apj}
|
\section{Introduction}
\label{s:intro}
To face the emerging needs for service flexibility, network efficiency, traffic diversification, and network security and reliability, today's network nodes are called to a more flexible and richer packet processing. The original Internet nodes, historically limited to switches and routers providing ``just'' the plain forwarding services, have been massively complemented with a variety of heterogeneous middlebox-type functions \cite{Wan11,She12,Qaz13,Gem14} such as network address translation, tunneling, load balancing, monitoring, intrusion detection, and so on. The diversification of network equipment and technologies has definitely provided an increased availability of network functionalities, but at the cost of a significant extra complexity in the control and management of large scale multi-vendor networks.
Software-defined Networking (SDN) emerged as an attempt to address such problem. Coined in 2009 \cite{Gre09} as a direct follow-up of the OpenFlow proposal \cite{OF08}, SDN has broadly evolved since then \cite{frz14} and does not in principle restricts to OpenFlow (a {\em ``minor piece in the SDN architecture''}, according to the OpenFlow inventors themselves \cite{She11}) as device-level abstraction. Nevertheless, most of the high level network programming abstractions proposed in the last half a dozen years \cite{Gud08,Nay09,Fos11,Voe12,Mon13,Nel14,Kim15} still rely on OpenFlow as southbound (using RFC 7426's terminology) programming interface. Indeed, OpenFlow was designed with the desire for rapid adoption, opposed to first principles \cite{frz14}; i.e., as a pragmatic attempt to address the dichotomy between i) flexibility and ability to support a broad range of innovation, and ii) compatibility with commodity hardware and vendors' need for closed platforms \cite{OF08}.
The aftermath is that most of the above mentioned network programming frameworks circumvent OpenFlow's limitations by promoting a ``two-tiered'' \cite{Ara15} programming model: {\em any} stateful processing intelligence of the network applications is delegated to the network controller, whereas OpenFlow switches limit to install and enforce stateless packet forwarding rules delivered by the remote controller. Centralization of the network applications' intelligence may not be a problem (and actually turns out to be an advantage) whenever changes in the forwarding states do not have strict real time requirements, and depend upon global network states. But for applications which rely only on local flow/port states, the latency toll imposed by the reliance on an external controller rules away the possibility to enforce software-implemented control plane tasks at wire speed, i.e. while remaining on the fast path\footnote{
A 64 bytes packet takes about 5 ns on a 100 gbps speed, roughly the time needed for a signal to reach a control entity placed one meter away. And the execution of an albeit simple software-implemented control task may take way more time than this. Thus, even the physical, capillary, distribution of control agents (as proxies of the remote SDN controller for low latency tasks) on each network device would hardly meet fast path requirements. }.
One might argue that we {\em do not even need} such ultra-fast processing and packet-by-packet manipulation and control capabilities. However, not only the large real-world deployment of proprietary hardware network appliances (e.g. for traffic classification, control/balancing, monitoring, etc), but also the evolution of the OpenFlow specification itself shows that this may not be the case. As a matter of fact, since the creation of the Open Networking Foundation (ONF) in 2011, and up to the latest (version 1.5) specification, we have witnessed an hectic evolution of the OpenFlow standard, with several OpenFlow extensions devised to fix punctual shortcomings and accommodate specific needs, by incorporating {\em extremely specific} stateful primitives (such as meters for rate control, group ports for fast failover support or dynamic selection of one among many action buckets at each time - e.g. for load balancing -, synchronized tables for supporting learning-type functionalities, etc).
Indeed, in the last couple of years, a new research trend has started to challenge improved programmability of the data plane, beyond the elementary ``match/action'' abstraction provided by OpenFlow, and (even more recently) initial work on higher level network programming frameworks devised to exploit such newer and more capable lower-level primitives are starting to emerge \cite{Sha15,Ara15}. Proposals such as POF \cite{Son13,Son15}, although not yet targeting stateful flow processing, do significantly improve header matching flexibility and programmability, freeing it from any specific structure of the packet header. Programmability of the packet scheduler inside the switch has been recently addressed in \cite{Siv15}. Works such as OpenState \cite{ccr14,Pon15} and FAST \cite{Mos14} explicitly add support for per-flow state handling inside OpenFlow switches, although the abstractions therein defined are still simplistic and severely limit the type of applications that can be deployed (for instance, OpenState supports only a special type of Finite State Machines, namely Mealy Machines, which do not provide the programmer with the possibility to declare and use own memory or registries). The P4 programming language \cite{Bos14,Jos15} leverages more advanced hardware technology, namely dedicated processing architectures \cite{flexpipe} or Reconfigurable Match Tables \cite{Bos13} as an extension of TCAMs (Ternary Content Addressable Memories) to permit a significantly improved programmability in the packet processing pipeline. In its latest 1.0.2 language specification \cite{P4spec}, P4 has made a further crucial step in improving stateful processing, by introducing registers defined as {\em ``stateful memories} [which] {\em can be used in a more general way to keep state''}. However, the P4 language does not specify how registries should be scalably supported and managed by the underlying HW.
\vspace{-3pt}
\subsubsection*{Contribution}
This work is an attempt to revisit fast-path programmability, by (i) proposing a programming abstraction which retains the platform independent features of the original ``match/action'' OpenFlow abstraction, and by (ii) showing how our abstraction can be directly ``executed'' over an HW architecture (whose feasibility is concretely shown via an HW FPGA prototype). In analogy with the OpenFlow's ``match/action'' abstraction, which exposes a network node's TCAM to third party programmability, also our abstraction directly refers to the HW interface, and as such it can be directly exposed to the programmer as a machine-level ``configuration'' interface, hence without any intermediary compilation or adaptation to the target (i.e., unlike the case of P4).
In conceiving our abstraction, we have been largely inspired by \cite{ccr14}, where eXtended Finite State Machines \cite{Che93} (therein referred to as ``full'' XFSM) were conjectured as a possible forward-looking abstraction. Our key difference with respect to \cite{ccr14} is that we not limit to postulate that such ``full'' XFSMs may ultimately be a suitable abstraction, but we concretely show their viability and their ``executability'' over an HW architecture leveraging commodity HW (standard TCAMs, hash tables, ALUs, and somewhat trivial additional circuitry), and with a strictly bounded number of clock ticks.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{fig1.png}
\vspace{-2em}
\caption{a) A typical OpenFlow pipeline architecture. b) the OPP enabled pipeline. OPP ``stages'' can be pipelined with other OPP stages or ordinary OpenFlow Match/Action stages.}
\vspace{-1em}
\label{f:pipe}
\end{figure}
A limitation in this paper is our focus on a ``single'' packet processing stage, opposed to a more general packet processing pipeline comprising multiple match/action tables. In essence, our work shows the viability of an XFSM-based abstraction as a (significant) generalization of the original single-table OpenFlow's match/action. While multiple pipelined instances of our atomic Open Packet Processor stages are clearly possible, exactly as multiple match/action tables can be pipelined since OpenFlow version 1.1 (see figure \ref{f:pipe}), our present work does not yet take advantage of HW pipeline optimizations such as Reconfigurable Match Tables \cite{Bos13}.
Finally, and similarly to the OpenFlow's original design philosophy, even if our proposed architecture is pragmatically limited by the specific set of primitives implemented by the HW (supported packet processing and forwarding actions, matching facilities, arithmetic and logic operations on registry values, etc), it nevertheless remains extensible (by adding new actions or instructions) and largely expressive in terms of how the programmer shall use and combine such primitives within a desired stateful operation. As it will be hopefully apparent later on, a ``full'' XFSM permits to formally describe a wide variety of programmable packet processing and control tasks, which our architecture permits to directly convey and deploy {\em inside} the switch. And even if probably not of any practical interest, the fact that not only the OpenFlow legacy statistics, but also further tailored stateful extensions today integrated in the OpenFlow standard (hence hardcoded in the switch) might be {\em externally} programmed using an apparently viable platform agnostic abstraction merits further considerations (see discussion in section \ref{s:disc}).
\section{Concept}
\label{s:concept}
As anticipated in the previous section, our work focuses on the design of a single Open Packet Processor (OPP) stage, as a significant generalization of the traditional OpenFlow's Match/Action abstraction. More specifically, our goal is to provide a packet processing stage which holds the following properties.
\vspace{3pt} \noindent {\bf Ability to process packets directly on the fast path}, i.e., {\em while} the packet is traveling in the pipeline (nanoseconds time scale). The requirement of performing packet processing tasks in a {\em deterministic} and {\em small} (bounded) number of HW clock cycles hardly copes with the possibility to employ a standard CPU (and the relevant programming language), and requires us to implement a domain-specific (traffic/network) computing architecture from scratch.
\vspace{3pt} \noindent {\bf Efficient storage and management of per-flow stateful information}. Other than parsing packet header information and exposing such fields to a match/action stage (or a pipeline of match action stages\cite{Bos13,flexpipe}), we also aim at permitting the programmer to further use the {\em past} flow history for defining a desired per-packet processing behaviour. As shown in section \ref{ss:arch}, this can easily accomplished by {\em pre-pending} a dedicated storage table (concretely, an hash table) that permits to retrieve, in O(1) time, stateful flow information. We name this structure as {\em Flow Context Table}, as, in somewhat analogy with context switching in ordinary operating systems, it permits to retrieve stateful information associated to the flow to which an arriving packet belongs, and store back an updated context the end of the packet processing pipeline. Such (flow) context switching will operate at wire speed, on a packet-by-packet basis.
\vspace{3pt} \noindent {\bf Ability to specify and compute a wide (and programmable) class of stateful information}, thus including counters, running averages, and in most generality stateful features useful in traffic control applications. It readily follows that the packet processing pipeline, which in standard OpenFlow is limited to match/action primitives, must be enriched with means to describe and (on the fly) enforce conditions on stateful quantities (e.g. the flow rate is above a threshold, or the time elapsed since the last seen packet is greater than the average inter-arrival time), as well as provide arithmetic/logic operations so as to update such stateful features in a bounded number of clock cycles (ideally one).
\vspace{3pt} \noindent {\bf Platform independence}. A key pragmatic insight in the original OpenFlow abstraction was the decision of restricting the OpenFlow switch programmer's ability to just {\em select} actions among a finite set of supported ones (opposed to permitting the programmer to develop own custom actions), and associate a desired action set (bundle) to a specific packet header match. We conceptually follow a similar approach, but we cast it into a more elaborate eXtended Finite State Machine (XFSM) model. As described in section \ref{ss:xfsm}, an XFSM abstraction permits us to formalize complex behavioral models, involving custom per-flow states, custom per-flow registers, conditions, state transitions, and arithmetic and logic operations. Still, an XFSM model does not require us to know {\em how} such primitives are concretely implemented in the hardware platform, but ``just'' permits us to combine them together so as to formalize a desired behaviour. Hence, it can be {\em ported} across platforms which support a same set of primitives.
\subsection{XFSM abstraction}
\label{ss:xfsm}
The OpenFlow's ``Match-action'' abstraction has been widely extended throughout the various standardization steps, with the extension of the match fields (including the possibility to perform matches on meta-data), with new actions (and instructions), and with the ability to associate a set of actions to a given match. Nevertheless, the basic abstraction conceptually remains the same. It is instructive to formally re-interpret the (basic) OpenFlow match/action abstraction as a ``map'' $T : I \rightarrow O$, where $I=\{i_1, \ldots, i_M\}$ is a finite set of {\em Input Symbols}, namely all the possible matches which are technically supported by an OpenFlow specification (being irrelevant, at least for this discussion, to know how such Input Symbols' set $I$ is established, and that each input symbol is a Cartesian combination of all possible header field matches), and $O=\{o_1, \ldots, o_K\}$ is a finite set of {\em Output Symbols}, i.e. all the possible actions supported by an OpenFlow switch. The obvious limit of this abstraction is that the match/action mapping is statically configured, and can change only upon controller's intervention (e.g. via flow-mod OpenFlow commands). Finally, note that the ``engine'' which performs the actual mapping $T : I \rightarrow O$ is a standard TCAM.
As observed in \cite{ccr14}, an OpenFlow switch can be trivially extended to support a more general abstraction which takes the form of a {\em Mealy Machine}, i.e. a Finite State Machine with output,
and which permits to formally model {\em dynamic} forwarding behaviors, i.e. permit to change in time the specific action(s) associated to a same match. It suffices to add a further finite set $S=\{s_1, s_2, …, s_N\}$ of {\em programmer-specific states}, and use the TCAM to perform the mapping $T : S \times I \rightarrow S \times O$. While remaining feasible on ordinary OpenFlow hardware, such Mealy Machine abstraction brings about two key differences with respect to the original OpenFlow abstraction. First, the (output) action associated to a very same (input) match may now differ depending on an (input) {\em state} $s_i \in S$, i.e., the state in which the flow is found when a packet is being processed. Second, the Mealy Machine permits to specify in which, possibly different, (output) state $s_o \in S$ the flow shall enter once the packet will be processed. While quite interesting, this generalization appears still insufficient to permit the programmer to implement meaningful applications, as it lacks the ability to run-time compute and exploit in the forwarding decisions per-flow features commonly used in traffic control algorithms.
\begin{table}[t]
\centering
{\footnotesize
\begin{tabular}{|c|p{2.9cm}|p{4.2cm}|} \hline
\multicolumn{2}{|c|} {\bf XFSM formal notation} &
{\bf Meaning} \\ \hline
I & input symbols & all possible matches on packet header fields \\ \hline
O & output symbols & OpenFlow-type actions \\ \hline
S & custom states & application specific states, defined by programmer \\ \hline
D & n-dimensional linear space $ D_1 \times \cdots \times D_n $ &
all possible settings of $n$ memory registers;
include both custom per-flow and global switch registers (see text) \\ \hline
F & set of enabling functions $f_i: D \rightarrow \{0,1\}$ &
Conditions (boolean predicates) on registers \\ \hline
U & set of update functions $u_i: D \rightarrow D$ &
Applicable operations for updating registers' content \\ \hline
T & transition relation $T: S \times F \times I \rightarrow
S \times U \times O $ &
Target state, actions and register update
commands associated to each transition \\ \hline
\end{tabular}}
\caption{eXtended Finite State Machine model}
\vspace{-1.2em}
\label{t:xfsm}
\end{table}
The goal of this paper is to show that a switch architecture can be further easily extended (section \ref{ss:arch}) to support an even more general Finite State Machine model, namely the eXtended Finite State Machine (XFSM) model introduced in \cite{Che93}. As summarized in table \ref{t:xfsm}, this model is formally specified by means of a 7-tuple $M=(I,O,S,D,F,U,T)$. Input symbols $I$ (OpenFlow-type matches) and Output Symbols $O$ (actions) are the same as in OpenFlow. Per-application states $S$ are inherited from the Mealy Machine abstraction \cite{ccr14}, and permit the programmer to freely specify the possible states in which a flow can be, in relation to her desired custom application (technically, a state label is handled as a bit string). For instance, in an heavy hitter detection application, a programmer can specify states such as \texttt{NORMAL}, \texttt{MILD}, or \texttt{HEAVY}, whereas in a load balancing application, the state can be the actual switch output port number (or the destination IP address) an already seen flow has been pinned to, or \texttt{DEFAULT} for newly arriving flows or flows that can be rerouted. With respect to a Mealy Machine, the key advantage of the XFSM model resides in the additional programming flexibility in three fundamental aspects.
\vspace{3pt} \noindent {\bf (1) Custom (per-flow) registers and global (switch-level) parameters.} The XFSM model permits the programmer to explicitly define her own registers, by providing an array of per-flow variables whose content (time stamps, counters, average values, last TCP/ACK sequence number seen, etc) shall be decided by the programmer herself. Additionally, it is useful to expose to the programmer (as further registers) also switch-level states (such as the switch queues' status) or ``global'' shared variables which all flows can access. Albeit practically very important, a detailed distinction into different register types is not foundational in terms of abstraction, and therefore all registers that the programmer can access (and eventually update) are summarized in the XFSM model presented in Table \ref{t:xfsm} via the {\em array D} of {\em memory registers}.
\vspace{3pt} \noindent {\bf (2) Custom conditions on registers and switch parameters.} The sheer majority of traffic control applications rely on {\em comparisons}, which permit to determine whether a counter exceeded some threshold, or whether some amount of time has elapsed since the last seen packet of a flow (or the first packet of the flow, i.e., the flow duration). The {\em enabling functions} $f_i:D\rightarrow\{0,1\}$ serve exactly for this purpose, by implementing a set of (programmable) boolean comparators, namely conditions whose input can be decided by the programmer, and whose output is 1 or 0, depending on whether the condition is true or false. In turns, the outcome of such comparisons can be exploited in the transition relation, i.e. a state transition can be triggered only if a programmer-specific condition is satisfied.
\vspace{3pt} \noindent {\bf (3) Register's updates.} Along with the state transition, the XFSM models also permits the programmer to update the content of the deployed registers. As we will show later on, registers' updates require the HW to implement a set of {\em update functions} $u_i:D\rightarrow D$, namely arithmetic and logic primitives which must be provided in the HW pipeline, and whose input and output data shall be configured by the programmer.
\vspace{3pt} Finally, we stress that the actual computational step in an XFSM is the transition relation $T : S \times F \times I \rightarrow S \times U \times O$, which is nothing else than a ``map'' (albeit with more complex inputs and outputs than the basic OpenFlow map), and hence is naturally implemented by the switch TCAM, as shown in the next section \ref{ss:arch}.
\begin{figure*}[t]
\centering
\includegraphics[width=1\textwidth, height=5cm]{fig2.png}
\vspace{-2em}
\caption{OPP architecture}
\vspace{-1em}
\label{f:hl_arch}
\end{figure*}
\subsection{OPP architecture}
\label{ss:arch}
To our view, what makes the previously described XFSM abstraction compelling is the fact that it can be {\em directly executed on the switch's fast path} using off the shelf HW, as we will prove in section \ref{s:hardware} with a concrete HW prototype. As discussed in the next section, practical restrictions of course emerge in terms of memory deployed for the registers, as well as capability of the ALUs used for register updates, but such restrictions are mostly related to an actual {\em implementation}, rather than to the design which remains at least in principle very general and flexible. A sketch of the proposed Open Packet Processor architecture is illustrated in Figure \ref{f:hl_arch}. The packet processing workflow is best explained by means of the following {\em stages}.
\vspace{3pt} \noindent {\bf Stage 1: flow context lookup}. Once a packet enters an OPP processing block, the first task is to extract, from the packet, a {\em Flow Identification Key} (FK), which identifies the entity to which a state may be assigned. The flow is identified by an unique key composed of a subset of the information stored in the packet header. The desired FK is configured by the programmer (an IP address, a source/destination MAC pair, a 5-tuple flow identifier, etc) and depends on the specific application. The FK is used as index to lookup a {\em Flow Context Table}, which stores the {\em flow context}, expressed in terms of (i) the state label $s_i$ currently associated to the flow, and (ii) an array $\vec{R} = \{R_0, R_1, ..., R_k\}$ of (up to) $k+1$ registers defined by the programmer. The retrieved flow context is then appended as metadata and the packet is forwarded to the next stage.
\vspace{3pt} \noindent {\bf Stage 2: conditions' evaluation}. Goal of the {\em Condition Block} illustrated in Figure \ref{f:hl_arch} (and implemented using ordinary boolean circuitry, see section \ref{s:hardware}) is to compute programmer-specific conditions, which can take as input either the per flow register values (the array $\vec{R}$), as well as global registers delivered to this block as an array $\vec{G} = \{G_0, G_1, ..., G_h\}$ of (up to) $h+1$ global variables and/or global switch states. Formally, this block is therefore in charge to implement the {\em enabling functions} specified by the XFSM abstraction. In practice, it is trivial to extend the assessment of conditions also to packet header fields (for instance, port number greater than a given global variable or custom per-flow register). The output of this block is a boolean vector $\vec{C} = \{c_0, c_1, ..., c_m\}$ which summarizes whether the $i$-th condition is true ($c_i=1$) or false ($c_i=0$).
\vspace{3pt} \noindent {\bf Stage 3: XFSM execution step}. Since boolean conditions have been transformed into 0/1 bits, they can be provided as input to the TCAM, along with the state label and the necessary packet header fields, to perform a wildcard matching (different conditions may apply in different states, i.e. a bit representing a condition can be set to ``don't care'' for some specific states). Each TCAM row models one transition in the XFSM, and returns a 3-tuple: (i) the next state in which the flow shall be set (which could coincide with the input state in the case of no state transition, i.e., a self-transition in the XFSM), (ii) the actions associated the transition (usual OpenFlow-type forwarding actions, such as \texttt{drop}, \texttt{push\_label}, \texttt{set\_tos} etc...), and (iii) the information needed to update the registers as described below.
\vspace{3pt} \noindent {\bf Stage 4: register updates}. Most applications require arithmetic processing when updating a stateful variable. Operations can be as simple as integer sums (to update counters or byte statistics) or can require tailored floating point processing (averages, exponential decays, etc). The role of the {\em Update logic block} component highlighted in Figure \ref{f:hl_arch} is to implement an array of Arithmetic and Logic Units (ALUs) which support a selected set of computation primitives which permit the programmer to update (re-compute) the value of the registers, using as input the information available at this stage (previous values of the registers, information extracted from the packet, etc). Section \ref{ss:alu} will describe the specific instruction set implemented in our HW prototype, where (with no pretence of completeness, nor willingness to impose our own set of operations) we implement a set of operations which appear to be either useful to the specific network programmer's needs, as well as computationally effective in terms of implementation (ideally, executable in a single clock tick). It is worth to mention that the problem of extending the set of supported ALU instructions is merely a technical one, and does not affect the OPP architecture.
\vspace{3pt} \noindent {\bf Extension: Cross-flow context handling}. As noted in \cite{ccr14}, there are many useful stateful control tasks, in which states for a given flow are updated by events occurring on {\em different} flows. A simple but prominent example is MAC learning: packets are forwarded using the {\em destination} MAC address, but the forwarding database is updated using the {\em source} MAC address. Thus, it may be useful to further generalize the XFSM abstraction as suggested in \cite{ccr14}, i.e. by permitting the programmer to use a Flow Key during lookup (e.g. read information associated to a MAC destination address) and employ a possibly different Flow Key (e.g. associated to the MAC source) for updating a state or a register.
\subsection{Programming the OPP}
Is is useful to conclude this section with at least a sketch of which types of applications (and programs) may be deployed.
A first trivial example of dynamic forwarding actions is that of a simple mechanism which distinguishes {\bf short-lived flows from long-lived flows} by considering ``long'' any flow that has transmitted at least N packets, and applies different DSCP tags. The OPP programmer would simply need to define two states (\texttt{DEFAULT} also associated to every new flow, and \texttt{LONG}), one per-flow register $R_0$ (a packet counter), one global register $G_0$ (storing the constant threshold $N$), a condition $R_0>G_0$ applicable when in state \texttt{DEFAULT}, and an update function ${\rm ADD}(R_0,1) \rightarrow R_0$. Note that we {\em did not} assume any pre-implemented counters or meters in the switch, but the counter and the relevant threshold check has been programmed using the OPP abstraction.
The usage of packet inter-arrival times and timers is exemplified by a {\bf dynamic intra-flow load balancing} application, which can reroute a flow {\em while} it is in progress. As suggested in \cite{Kan07}, rerouting should {\em not} occur during packet bursts, to avoid out of ordering and relevant performance impairments. Support in OPP just requires, for each packet being transmitted, to update a per-flow register $R$ with the quantity $t+\Delta$, being $t$ the actual packet timestamp and $\Delta$ a suitable threshold. When the next packet arrives (time $t_1$), we check the condition $t_1>R$. If this is false, we route the packet to the assigned path (indicated by the state label); conversely we route it to an alternative path, and we change the state accordingly. Again, note that we have not assumed any pre-implemented support from the switch (e.g. timeouts or soft states), besides the ability to provide time information (e.g., via a global register, or timestamps as packet metadata).
Finally, the integration in the ALU design of monitoring-specific update instructions (averages, variances, smoothing filters, see section \ref{ss:alu}), several features frequently used in traffic control and classification applications can be computed on the fly during the pipeline. We defer relevant examples to section \ref{s:usecases}.
\section{Hardware feasibility}
\label{s:hardware}
Despite the current trend in softwarization of network functions and the widespread deployment of software switches, we believe that the viability of switch-level programming abstractions which challenge OpenFlow limitations, hence including this work, {\em must} still be proven in terms of hardware feasibility and ability to run in a strictly bounded number of clock cycles.
\begin{figure}[t]
\centering
\includegraphics[width=0.52\textwidth]{FPGA2.pdf}
\vspace{-2em}
\caption{Scheme of an OPP stage}
\vspace{-1.3em}
\label{F:HW_core}
\end{figure}
Figure \ref{F:HW_core} provides a block-level overview of a candidate hardware implementation of a single OPP stage, which we prove feasible with an FPGA prototype. Pipelining of an OPP stage with other OPP stages or ordinary match/action tables does not affect the single stage design (although it does not permit us to benefit from hardware extensions and TCAM optimizations such as those introduced in \cite{Bos13}, which we leave to future work). Figure \ref{F:HW_core} also illustrates the necessary auxiliary blocks devised to handle packet input capture and output delivery, in the assumption of a $4 \times 4$ port switch.
\vspace{3pt} \noindent {\bf Packet reception and header field extraction}.
Packets received on the input queues are collected and serialized by a mixer block, so that the OPP block receives one packet per clock cycle. Such packet is then processed by a {\em Packet Fields Extractor}, configured to provide, together with the header fields (8 in our prototype), the blocks required in the next processing stages - specifically: i) the Flow Key used to query the Flow Context Table, ii) the header fields used by the Condition block, iii) the header fields used by the Update Logic Block, and iv) the (eventually different) Flow Key used for updating the Flow Context. The Packet Fields Extractor is easily implemented in HW as a parallel array of elementary Shift and Mask (SaM) blocks where each SaM block selects the beginning of the targeted header field (the shift function), and performs a bit-wise mask operation. This operation closely resembles that proposed in POF \cite{Son13}: we also use offsets, but instead of lengths we use bit masks.
\vspace{3pt} \noindent {\bf Flow Context Table}. This data structure is in charge to store both state as well as registries associated to Flow Keys. It consists of an hash table (we implemented a {\em d-left hash table} with $d=4$) to handle exact matches, plus a TCAM to handle wildcard matches. Unlike the hash table, which must be arguably large to store per-flow states, a very small TCAM can be deployed, as it is required to handle the very few special cases where wildcard matches are needed (mainly default states, where the TCAM priority permits to differentiate default states for different protocols or packet formats). Our implementation uses 128 bit Flow Keys, and returns a 146 bit value which is sufficient to support a 16 bit state label, four 32 bit per-flow registries, and two auxiliary bits per entry used by the microcontroller for housekeeping (see below).
\vspace{3pt} \noindent {\bf Condition Logic Block}. This block permits to configure conditions on input pairs (per-flow registries, global registries, header fields), and evaluate them so as to return as output a boolean 0/1 vector. This block, shown in figure \ref{F:cond}, comprises multiple (8 in our implementation) parallel configurable comparators, each of which takes as input two operands selected among all the flow registries $R_i$, all the global registries $G_i$ and the header fields $H_i$ coming from the packet field extractor. The selection operation is provided by two multiplexers (one for each operand). Each comparator supports five arithmetic comparison functions: $>$, $\ge$, $=$, $\le$, $<$.
\vspace{3pt} \noindent {\bf XFSM Table}. While, conceptually, this is a key ``computational'' stage in our proposed architecture (it performs a state transition step), in practice its implementation is straightforward: it just relies on an ordinary TCAM. Although a-posteriori it may seem obvious, such a simple support for a ``full'' XFSM transition step was enabled by the clear distinction between the configuration and evaluation of conditions (by the Condition Logic Block) and their usage as boolean outcomes, hence one bit per condition which can be directly used as TCAM input, along with the state label and the usual packet header fields used in OpenFlow matches. As in standard implementations, the TCAM provides as output the row associated to the matching rule with higher priority, and is followed by a companion RAM which stores the associated output. In our specific case, this consists in i) the next state label (16 bits) used to update the flow context table, ii) the action to perform on the packet (16 bits) and the ALU instructions which shall be applied to update registries (our prototype supports up to 5 instructions of 32 bits each).
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth]{comp.png}
\vspace{-2em}
\caption{Condition logic block array element}
\vspace{-1.8em}
\label{F:cond}
\end{figure}
\vspace{3pt} \noindent {\bf Update Logic Block}. This is the second ``computational'' stage in our architecture. This block deploys an array of ALUs (Arithmetic and Logic Units) which support a specific set of (micro)instructions useful for traffic processing tasks, and which {\em execute} in parallel the instructions provided as output of the XFSM Table. The updated registry values are then stored in the relevant memory locations (flow registries and/or global registries). Technical details are provided in section \ref{ss:alu}.
\vspace{3pt} \noindent {\bf Miscellaneous Blocks and Microcontroller}. To complete the HW architecture, a few necessary extra blocks have been implemented. The {\em Action Block} applies the selected actions to the packet and is perfectly analogous to an OpenFlow implementation. Being just a proof-of-concept, our prototype implements only basic ``sample'' actions (drop, forward, flood). {\em Global registries} are implemented as a standard register file unit for concurrent access. The {\em Metadata block} is in charge to provide additional information associated to an arriving packet, i.e., input port and timestamp.
Finally, our prototype has been complemented with a {\em microcontroller} providing a communication interface (UART) to configure the various programmable components inside the OPP (configuration registers, TCAM and RAM memories, etc.), i.e. to deploy in the switch an externally programmed application. Each configurable quantity is memory mapped in the microcontroller address space, which can directly read/write the content of these components. The microcontroller further implements management functions, among which slow-time-scale flow context table management (housekeeping): the microcontroller periodically scans the entries in the flow context table to detect and clean stale entries. To this purpose, two activity flag bits are stored in each flow entry and permit to label entries as \texttt{ACTIVE}, \texttt{INACTIVE} (no accesses have occurred in a configurable management cycle, e.g., order of seconds), and \texttt{DELETED}. It is worth to note that this is the only operation performed by the OFP that is not triggered by a packet.
\subsection{Update logic block}
\label{ss:alu}
Besides the support for state transition, a further key motivation behind this work was the attempt to cleanly design {\em inside the abstraction} (and concretely support in the OPP architecture) computational primitives involving arithmetic processing, as this is frequently needed in many application. The Update Logic Block is the OPP component in charge to provide such facility. It comprises a number (5 in our implementation) of small parallel ALUs (Arithmetic Logic Units) able to perform a set of elementary instructions which frequently occur in traffic control applications. The ones specifically implemented in our prototype are listed in table \ref{T:ALU} and \ref{T:specistr}. Some of these instructions are those of a typical RISC architecture, while others are specific for packet processing tasks (last row in the table).
At each step, the specific computations that the Update Logic Block must perform are provided by the output of the XFSM transition, and are expressed in the form of a tuple of instructions (32 bit instructions in our prototype). Each instruction comprises an 8 bits $OPCODE$, followed by a variable number of operands that depend on the specific instruction. Input operands ($INi$) can be any among the available per flow registries $R_i$, the global variables $G_i$, or the header fields $H_i$ provided by the Extractor. Output operands ($OUTi$) indicate where the result of the instruction must be written (e.g. in a given per-flow register, or in a global variable). In some instructions, one or more of the operands ($IOi$) are both used as input and output. Our implementation supports 4 per-flow registries, 4 global registries and 8 header fields. Therefore, it may in principle support up to $24/\log_2(16)=6$ operands. In practice, we envision at most 4 operands (e.g., for the variance or for the ewma smoothing instructions) and thus our implementation may readily support up to 64 among registries and header fields. In the case of logic/arithmetic/shift operations, which only require at most two operands plus a third output, we have also considered the case in which one of the operands is an actual value (immediate value) which can hence use 16 bits.
\begin{table}[t]
\vspace{-.6em}
\centering
\small
\begin{tabular}{|c|c|c|}
\hline
Type & Instructions & definition \\
\hline
Logic & NOP & do nothing\\
& NOT & $OUT1 \leftarrow NOT(IN1)$\\
& XOR, AND, OR & $OUT1 \leftarrow IN1\; op\; IN2$\\
\hline
Arit. & ADD,SUB,MUL,DIV & $OUT1 \leftarrow IN1\; op\; IN2$\\
& ADDI,SUBI,MULI,DIVI & $OUT1 \leftarrow IN1\; op\; IMM$\\
\hline
Shift/ & LSL (Logical Shift Left) & $OUT1 \leftarrow IN1 << IMM$ \\
Rotate & LSR (Logical Shift Right) & $OUT1 \leftarrow IN1 >> IMM$\\
& ROR (Rotate Right) & $OUT1 \leftarrow IN1\; ror\; IMM$ \\
\hline
\end{tabular}
\vspace{-1em}
\caption{ALU basic instruction set}
\vspace{-1em}
\label{T:ALU}
\end{table}
The packet/flow specific instructions supported in our prototype do implement, as a dedicated HW primitives running at the system clock frequency and with a maximum latency of two clock cycles\footnote{As they involve a division, which we had to limit to 16 bits for dividend and divisor to target a 2 clock cycles latency.}, domain-specific operations which we deem useful in traffic control applications, and which would normally require multiple clock cycles if implemented using more elementary operations. Such domain specific operations include the online computation of running averages ($avg$) and variances ($var$), and the computation of exponentially decaying moving averages ($ewma$) which can serve the purpose of a moving average, but which can be incrementally computed and do not require to maintain a window of samples.
Usage and implementation details about packet/flow specific instructions are provided in Table \ref{T:specistr}. The $avg$ operation stores the number of samples in $IO1$, and includes a new sample $IN1$ in the running average $IO2$. Similarly, the $var$ operation stores the number of samples in $IO1$, the average of the value $IN1$ in $IO2$ and the variance in $IO3$.
The $ewma$ operation\footnote{Being $t_k$ the last sample time, and $x_{k'}$ a new sample occurring at time $t_{k'}$, for simplicity of HW implementation we {\em approximate} the exponentially weighted moving average as $m(t_{k'})=m(t_k) \alpha^{t_{k'}-t_k} + x_{k'}$, and we use $\alpha=1/2$ to compute powers as shift operations. The intermediate $decay$ quantity in the second line is used just for clarity of presentation.}
was included to permit smoothing. It stores the last timestamp ($IN1$) of a packet in the register identified by $IO1$, computes the exponentially weighted moving average of the value $IN2$ using the equation in Table \ref{T:specistr} and stores the result in $IO2$.
As a final remark, similar to the action set in standard OpenFlow, we stress that the specific instruction set provided by the Update Logic Block is independent of our proposed OPP abstraction, i.e., its extension or improvement (e.g. with further dedicated domain-specific instructions) does not affect the overall OPP design.
\begin{table}[t]
\vspace{-.6em}
\centering
\small
\begin{tabular}{|c|c|}
\hline
Instruction & definition \\
\hline
avg() & $IO1 \leftarrow IO1+1$ \\
& $IO2 \leftarrow IO2+(IN1-IO2)/(IO1+1)$\\
\hline
& $IO1 \leftarrow IO1+1$\\
var() & $IO2 \leftarrow IO2+(IN1-IO2)/(IO1+1)$\\
& $IO3 \leftarrow IO3+((IN1-IO2)^2 -IO3)/(IO1+1)$\\
\hline
& $IO1 \leftarrow IN1$ \\
ewma() & $decay = 1<<(IN1-IO1)$ \\
& $IO2 \leftarrow IO2/decay +IN2$ \\
\hline
\end{tabular}
\vspace{-1em}
\caption{ALU packet/flow specific instructions}
\vspace{-1em}
\label{T:specistr}
\end{table}
\subsection{FPGA prototype}
\label{s:proto}
The OFP HW prototype has been designed using as target development board the NetFPGA SUME \cite{sume}, an x8 Gen3 PCIe adapter card incorporating a Xilinx Virtex-7 690T FPGA \cite{V7}, four SFP+ transceivers providing four 10GbE links, three 72 Mbits QDR II SRAM and two 4GB DDR3 memories.
The FPGA is clocked at 156.25 MHz, with a 64 bits data path from the Ethernet ports, corresponding to a 10 gbps throughput per port. The aggregated bus output of the mixer is 320 bits wide and is able to provide an overall throughput of 50 Gbps. The d-left hash table implementing the flow context table is sized for 4K entries. In order to support the target throughput, the RAMs composing the d-left table are realized as dual port RAM, so as to provide a read and a write operation for each clock cycle.
The prototype implements very small TCAMs. The TCAM associated to the hash table in the flow context table has 32 entries of 128 bits, whereas the XFSM TCAM has 128 entries of 160 bits. Indeed, TCAM implementation over FPGAs is very inefficient and is currently a widely open research issue \cite{TCAM1,TCAM2,TCAM4}, especially since the priority resolution hardware limits the maximum operating frequency when the number of TCAM entries increase. It is thus more interesting to {\em understand the performance that would be achievable with an ASIC implementation}. Following the same technology assumptions of \cite{Bos13}, an OPP ASIC design would be able to work at 1GHz operating frequency. This corresponds to an aggregate throughput of 960M packets/s, that is the maximum achievable by a 64 ports 10 Gb/s switch chip. However, the most important scaling provided by the ASIC implementation is given by the number of entries that can be stored in the OPP tables. The size of the SRAM that can be instantiated on a last generation chip is up to 32 MB, corresponding to around 1 millions of entries in the d-left hash for the context flow table. The size of a TCAM can be up to 40 Mb, corresponding to 256K XFSM table entries.
The system latency, i.e. the time interval from the first table lookup to the last context update is 6 clock cycles. The FPGA prototype is able to sustain the full throughput of 40 Gbits/sec provided by the 4 switch ports. If we suppose a minimum packet size of 40 bytes (320 bits), the system is able to process 1 packet for each clock cycle, and thus up to 6 packets could be pipelined. However, the feedback loop (not present in the forward-only OpenFlow pipelines \cite{OF1.4}) raises a concern: the state update performed for a packet at the sixth clock cycle would be missed by pipelined packets. This could be an issue for packets {\em belonging to a same flow} arriving back-to-back (consecutive clock cycles); in practice, as long as the system is configured to work by aggregating $N \geq 6$ different links, the mixer's round robin policy will separate two packets coming from the same link of $N$ clock cycles, thus solving the problem. Note that the 6 clock cycles latency is fixed by the hardware blocks used in the FPGA (the TCAM and the Block RAMs) and basically does not change scaling up the number of ingress ports or moving to an ASIC.
\begin{table}[t]
\centering
\small
\begin{tabular}{|c|c|c|}
\hline
resource type & Reference switch & OPP switch \\
\hline
\# Slice LUTs & 49436 (11\%) & 71712 (16\%) \\
\hline
\# Block RAMs & 194 (13\%) & 393 (26\%) \\
\hline
\end{tabular}
\vspace{-0.6em}
\caption{Hardware cost of OPP compared with the reference NetFPGA SUME switch.}
\vspace{-1.5em}
\label{t:synth}
\end{table}
The whole system has been synthesized using the standard Xilinx design flow. Table \ref{t:synth} reports the logic and memory resources (in terms of absolute numbers and fraction of available FPGA resources) used by the OPP FPGA implementation, and compare these results with those required for the NetFPGA SUME single-stage reference switch. As expected, the logic uses a small fraction of the total area (the increase with respect the reference switch is 5\% of the available FPGA logic resources), that is dominated by memory (that doubles with respect the reference switch). The synthesis results hence confirm the trend already shown by \cite{Bos13}: the HW area is dominated by memory, while adding intelligence/features in the logic require a small silicon overhead. The performance in terms of latency of an OPP stage and throughput of deployed FPGA prototype has been measured sending several synthetic traces of packets of different size. The results are presented in Fig. \ref{F:perfo}. As expected, the FPGA is able to sustain the expected throughput\footnote{Due to the limitation of our hardware measurement set-up, we were unable to actually send to the FPGA more than 24 Gbits/sec, so the data referred to 64B packet size could not be measured. The expected theoretical value is reported.}
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth]{fig_hw_perf.pdf}
\vspace{-1em}
\caption{Performance of the FPGA prototype}
\vspace{-1.2em}
\label{F:perfo}
\end{figure}
\section{Programming examples}
\label{s:usecases}
To functionally test the ability of OPP to support stateful applications, we have developed a complete OPP virtual software environment. For both the switch and controller implementation we have extended the CPqD OpenFlow 1.3 software virtual switch \cite{ofsoftswitch13} and the widely adopted OpenFlow controller Ryu \cite{ryu}. The sotware implementation serves just for testing purpose, hence it closely mimics the described OPP hardware operation, including the relevant limitations. To configure an OPP switch via the controller, we developed an OPP-specific extension of the OpenFlow protocol. Due to space limitations (the interested reader can find configuration files in the repository) we just mention that the configuration of the XFSM in the OPP architecture is a straightforward extension of an OpenFlow configuration: it just requires to populate the XFSM table entries and to configure of conditions, functions, key extractors and initial global register values. All software components required to test the proposed applications are bundled in a mininet \cite{mininet} based virtual machine avilable at a {\em dedicated OPP repository} \cite{repository}, along with our prototype's VHDL HW code.
To understand how an application can be programmed using OPP, let's walk through a simplistic example of a (quite inefficient, but at least trivial to follow) TCP port scan detection and mitigation application. Since the target is to detect IP address which behave as scanners, we use as Flow Key the IP address. Figure \ref{F:portscan} represents our desired application's behavior, expressed in the form of an XFSM ``program'', whereas figure \label{F:pscan_table} provides a corresponding tabular configuration delivered to the switch. For every IP packet, we check in the Flow Context Table whether the IP source has an associated context; if this is not the case, a \texttt{DEFAULT} state is conventionally returned. the XFSM table now checks whether the packet is a TCP SYN, and only in this case we will allocate a Flow Context Table entry for the considered IP source, and we will set it in \texttt{MONITOR} state. In this state, we measure the rate of new TCP SYN arrivals toward hosts behind the switch port 1. Such rate (computed with the EWMA update function) is stored and updated in the flow registers $R_0$.
\begin{figure}[t]
\centering
\includegraphics[width=0.49\textwidth]{portscan.pdf}
\vspace{-2.2em}
\caption{Port scan detection XFSM}
\label{F:portscan}
\end{figure}
While in \texttt{MONITOR} state, the value of $R_0$ is verified for each new TCP flow. If a given threshold (say 20 SYN/s, a value stored in the global register $G_0$) is exceeded, the state associated to this flow is set to \texttt{DROP} and all packets from this IP addresses are discarded. Suppose now that the programmer wants to block the scanner for 5 seconds. Lacking explicit timers (a non trivial HW extension), such mechanism is realised by the following procedure: (i) when the flow state transits from \texttt{MONITOR} to \texttt{DROP} the register $R_1$ is set to the packet time stamp value plus 5 sec. (a value stored in the global register $G_1$); (ii) in \texttt{MONITOR} state the $R_1$ value is checked for every received packet; (iii) If $R_1 <= pkt.ts$ the flow state is reverted to \texttt{MONITOR}. Moreover, the application needs to store the last packet timestamp in the flow register $R_2$. This time stamp will be used by the EWMA update function.
\begin{figure}[t]
\centering
\includegraphics[width=0.49\textwidth]{portscantable.pdf}
\vspace{-2em}
\caption{Port scan detection XFSM table}
\vspace{-1.5em}
\label{F:pscan_table}
\end{figure}
To implement this application, the OPP switch is configured by the controller using the OPP protocol that performs the following elementary operations: i) the \texttt{DEFAULT}, \texttt{MONITOR} and \texttt{DROP} states are encoded with 0, 1 and 2 respectively; ii) the lookup and update scopes are set to \texttt{ip.src}; iii) the global registers are defined as follows: $G_0 = 20$ and $G_1 = 5$[s]; iv) the condition table is configured to include two conditions: $C_0: R_0 \ge G_0$ and $C_1: R_1 > pkt.ts$ (packet arrival time); v) the XFSM table is configured as in fig. \ref{F:pscan_table}.
\subsection{Decision tree based traffic classification}
Machine Learning (ML) tools are widely adopted by the networking community for detecting anomalies and classifying traffic patterns \cite{nguyen2008survey}. We have tested the feasibility of using OPP to support this kind of traffic monitoring schemes by implementing a decision tree supervised classifier based on the C4.5 algorithm \cite{quinlan2014c4} that has been exploited by different work on ML based network traffic classification \cite{williams2006preliminary,pan2003hybrid,ma2008study,zhang2010method,li2007machine,alshammari2009machine}.
Any ML based classification mechanisms has two phases: a {\em training phase} and a {\em test phase}. The training phase is off line and used to create the classification model by feeding the algorithm with labeled data that associate a measured traffic feature vector to one of {\em n} decision classes. In the case of decision tree based ML algorithms, the output of such phase is the {\em binary classification tree}. The training phase must obviously be performed outside the switch. For our use case implementation the decision rules have been created using the Orange data mining framework \cite{JMLR:demsar13a}, and a feature set proposed in \cite{li2007accurate}. We considered a simple scenario where it is necessary to discriminate between WEB and P2P (control) traffic. The selected features for each flow are: packet size average/variance and total number of received bytes. These features are mapped directly to the per flow memory registers R1, R2, R3. Moreover the application XFSM requires two additional registers: R0 (packet counter) and R4 (measurement window expiration time). The input feature vectors are evaluated over a time window of 10 seconds.
The test phase, which is performed online, consists of two operations: (1) for each flow the feature set described above is computed; (2) after 10 seconds a decision is made according to the decision tree. This testing mechanism is implemented in OPP according to the XFSM shown in Figure \ref{fig:tree}. The XFSM flows states are encoded as follows: State 0 $\rightarrow$ default; State 1 $\rightarrow$ measurement and decision; State 2 $\rightarrow$ WEB traffic (DSCP class AF11); State 3 $\rightarrow$ P2P traffic (DSCP class best effort). The condition set is: $C0: (now > R4)$; $C1: (R2 > G2)$, $C2: R3 > G3$, $C3: R1 <= G1$. According to our simplified training phase, the (ceiled) thresholds values are: $G1 = 306$; $G2= 1575$ and $G3 = 203$.
\begin{figure}[t]
\centering
\includegraphics[width=.45\textwidth]{tree_xfsm.pdf} \vspace{-1em}
\caption{XFSM table for the traffic classifier}
\vspace{-1.5em}
\label{fig:tree}
\end{figure}
\subsection{Traffic policing with token buckets}
In this second use case we have implemented in OPP a single rate token bucket with burst size $B$ and token rate $1/Q$, where $Q$ is the token inter arrival time. Since in the OPP architecture the update functions are performed after the condition verification, we cannot update the number of tokens in the bucket based on packet arrival time before evaluating the condition (token availability) for packet forwarding. For this reason we have implemented an alternative and equivalent algorithm based on a time window. For each flow a time window W ($T_{min} - T_{max}$) of length $BQ$ is maintained to represent the availability times of the tokens in the bucket. At each packet arrival, if arrival time $T0$ is within W (Case 1), at least one token is available and the bucket is not full, so we shift W by Q to the right and forward packet. If the arrival time is after $T_{max}$ (Case 2), the bucket is full, so packet is forwarded and W is moved to the right to reflect that $B-1$ tokens are now available ($T_{min}=T0-(B-1)Q$ and $T_{max}=T0+Q$). Finally, if the packet is received before $T_{min}$ (Case 3), no token is available, therefore W is left unchanged and the packet is dropped.
In the OPP implementation, upon receipt of the first flow packet, we make a state transition in which we initialize the two registers: $Tmin=T0-(B-1)*Q$ and $Tmax=T0+Q$ (initialization with full bucket).
At each subsequent packet arrival we verify two conditions: $C0: Tnow>=Tmin$; $C1: Tnow<=Tmax$.
The three cases defined by the algorithm can be easily identified with these two conditions: case 1) C0 == True AND C1 == True; case 2) C0 == True AND C1 == False; case 3) C0 == False.
The XFSM is shown in Figure \ref{fig:tok_xfsm}.
\begin{figure}[t]
\centering
\includegraphics[width=.45\textwidth]{tok_xfsm.pdf}
\vspace{-1.2em}
\caption{Token bucket XFSM. The flow registers $R0$, $R1$ are used to store respectively Tmin, Tmax. The global registers $G0, G1$ are used to store $B*Q$ and $Q$. The extractors are $ip.src$}
\vspace{-1.5em}
\label{fig:tok_xfsm}
\end{figure}
\section{Discussion and extensions}
\label{s:disc}
\noindent \textbf{Hard-coded vs. programmable features}. \\
The reader might have noted that, in the proposal of the XFSM abstraction and in the technical design of OPP, we have attempted to {\em not rely} on {\em any} stateful feature today available in OpenFlow switches. This might seem odd, as, for a very basic example, per-flow statistics are assumed since \cite{OF08} to be collected and implemented by the switch hardware itself, and thus would be readily available in virtually any baseline reference OpenFlow switch architecture. However, at least for the purpose of this work, we have voluntary avoided to {\em expose}, via our abstraction, features (such as OpenFlow-type statistics) that could at least in principle be programmed from the outside, by {\em using} the abstraction. Thus, in the attempt to keep our proposed abstraction as clean as possible, we avoided focusing on possible optimizations, such as in-switch efficient implementation of frequently used features (such as per flow count/byte statistics, soft states and timers, etc.) which one would arguably expect from an advanced implementation, so as to concentrate programming efforts on one own application's logic.
\vspace{3pt} \noindent \textbf{Proposed abstraction: at which ``level''?}. \\
Our abstraction resides at a very low level, as in practice it specifies a machine language which directly configures the hardware. In this, albeit technically very different, we are quite similar in spirit to the original OpenFlow abstraction, envisioned as an abstraction of the switch's Flow table component. Independence of the underlying platform is therefore {\em not} accomplished via a compiler, such as in the case of emerging data plane programming languages such as P4 \cite{Jos15}, but it is accomplished by {\em decoupling} the actual identification and (XFSM-based) combination of the hardware primitives from how they are implemented inside the switch (the proposed OPP hardware design being a possible implementation, but not nearly the unique). As long as two platforms expose a same set of (header) matching facilities, forwarding actions, enabling functions for evaluating conditions, and a same instruction set for the update functions, and as long as they do support the XFSM state transition execution logic, the application's description in terms of XFSM can be ported across platforms. Loosely speaking, our abstraction is more closer to a ``bytecode'' or to an assembly language rather than to a practical programming language. While we believe this is a strength of our proposal, we also clearly recognize the importance of offering to the applications' developers an higher level and more user-friendly programming language (P4 being an obvious candidate, see next discussion) and relevant compilers which transform an higher level description into the proposed OPP XFSM-based machine language.
\vspace{3pt} \noindent \textbf{Use of P4 for the description of OPP stages.} \\
Even if the details of the reference P4 hardware architecture is not publicly available, the language has the goal of being expressive enough to describe any configurable architecture for packet forwarding. Therefore, even if out of the scopes of our paper, and for just the purpose to stimulate possible discussion (i.e. we don't claim this paper to provide any specific contribution in this direction), we nevertheless made our own preliminary attempt to understand how the proposed OPP architecture could be described using P4, i.e. if OPP could be used as a further platform's target for a P4 program. To this purpose, we have defined (and uploaded over the anonymized OPP's repository) a reusable OPP.p4 library. From such experience, we gathered a twofold impression. On one side, the current version of the language already permits to describe/support key OPP functionalities. Moreover, the ability to store data in persistent registries permits to properly describe, using P4, the ``computational loop'' characterizing OPP.
On the other side, we suffered from the lack, among the P4 constructs, of an explicit state/context table and a relevant clean way to store and access per-flow data. In our OPP.p4 library, a table functionally equivalent to our Context table was actually constructed by combining arrays of registers with hash keys generators which are provided as P4 language primitives. However, besides the obvious stretch (P4 registers are generic, and not specifically meant to be deployed on a per-flow basis), this construction also suffers from hash collisions, a non trivial problem if constrained to be addressed {\em while} the packet is flying through the pipeline. The availability of a tailored context/state table structure in P4 would greatly simplify the support of an OPP target platform.
\vspace{3pt} \noindent \textbf{Structural limitations and possible extensions} \\
While (we believe) very promising, our proposed approach is not free of {\em structural} concerns. If, on one side, limitations in the set of supported enabling functions and ALU functions for registry updates may be easily addressed with suitable extensions, and integration of more flexible packet header parsing (following \cite{Bos14}) is not expected to bring significant changes in the architecture, there are at least three pragmatic compromises which we took in the design, and which suggest future research directions. The first, and major, one resides in the fact that state transitions are ``clocked'' by packet arrivals: one and only one state transition associated to a flow can be triggered only if a packet of {\em that} flow arrives; asynchronous events, such as timers' expiration, are not supported. So far we have partially addressed this limitation with, on one side, the decoupling between lookup and update functions (the cross-flow state handling feature), and on the other side with programming tricks such as the handling of time performed while implementing the token bucket example. But further flexibility in this direction is a priority in our future research work. A second shortcoming is the deployment of ALU processing only in the Update Logic Block. This decision was done in favour of a cleaner abstraction and a simpler implementation. However (programmable) arithmetic and logic operations would be beneficial also {\em while} evaluate conditions ($e.g.$, $A-B > C$) which, in the most general case, may require to be postponed to the next packet (the update function can store $A-B$ in a registry, and the next condition can use such registry). A third, minor, shortcoming relates to the fact that all updates occur in parallel. This prevents the programmer to pipeline operations, i.e. use (in the same transition step) the output of an instruction as input to a next one. While this issue is easily addressed by deploying multiple Update Logic Blocks in series, this would increase the latency of the OPP loop.
\section{Related work}
\label{s:related}
This work focuses on data plane programming architectures and abstractions, a relatively recent trend. In such field, so far the mostly influential work is arguably P4 \cite{Bos14} a programming language specifically focusing on data path packet processing. In turns, such initial work has stimulated the creation of a consortium (p4.org) which has so far produced a release 1.0.2 of the language specification \cite{P4spec}. Our OPP work is at a different (lower) level than P4: it describes an hardware programming interface and a relevant architecture which could be in future adapted to be used as compilation target \cite{Jos15} for P4. Furthermore, our work deals with stateful processing {\em across different packets of a flow}, and as such it appears perfectly complementary to the original P4 proposal \cite{Bos14} which initially focused mainly on programming flexibility of the packet pipeline (P4 registers have been introduced in \cite{P4spec}).
Concerning stateful processing, the work closer to ours is OpenState \cite{ccr14}, and in part FAST \cite{Mos14}. With respect to OpenState, OPP makes a very significant step forward, as we support full eXtended Finite State Machines (XFSM, as defined in \cite{Che93}) opposed to the much simpler OpenState's Mealy Machines, and hence we significantly broaden the variety of applications that can be programmed on the switch. Such step requires additional new specialized hardware blocks with respect to OpenState which instead requires only marginal extensions to an OpenFlow hardware design \cite{Pon15}.
Finally, OPP shares some technical similarities with \cite{Bos13} and with the Intel Flexpipe architecture \cite{flexpipe}, especially for what concerns the handling of ALUs in the packet processing pipeline. However, both OPP focus (on stateful processing) and architecture design remain extremely different from both \cite{Bos13} and \cite{flexpipe}. Indeed, an advised extension consists in extending OPP to handle multiple pipelined stages and hence exploit the TCAM reconfigurability concepts introduced in \cite{Bos13}.
\section{Conclusions}
\label{s:conclusions}
OPP is an attempt to find a pragmatic and viable balance between platform-independent HW configurability and data plane (packet-level) programming flexibility. While permitting programmers to deploy more sophisticated stateful forwarding tasks with respect to the basic OpenFlow's static match/action abstraction, we believe that an asset of our configuration interface resides in the fact that it does not significantly depart from OpenFlow-type configurations - our extended finite state machine model is indeed conveyed to the switch in the usual form of a TCAM's Flow Table. We thus hope that our work might stimulate further debate in the research community on how to incrementally deploy programmable traffic processing inside the network nodes, e.g. via gradual OpenFlow extensions.
\newpage
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{intro}
Comets, as the most primitive bodies from the outer region of the solar system, are known to be very abundant in volatile ices and refractory dust grains. For a comet with orbital parameters inside 3-4 au, solar radiation raises the surface temperature to such an extent that ice sublimation initiates, as indicated by the appearance of a coma containing expanding gas and small dust particles. Therefore, the outgassing activity increases as the comet approaches perihelion. The behavior of comet 67P Churymov-Gerasimenko (67P hereafter) follows this pattern closely (\cite{2011A&A...525A..36L}; \cite{2011A&A...531A..54T}; \cite{2013A&A...549A.121V}). In addition to the nearly spherically symmetric coma, anisotropic structures in the form of collimated jets have been identified (\cite{2011A&A...525A..36L}; \cite{2013A&A...549A.121V}). The identification of the source mechanism and acceleration process of these dust jets are one of the main focus of the Rosetta mission.
The close-up observations of the OSIRIS scientific camera on the Rosetta spacecraft (\cite{2007SSRv..128..433K}) have provided an unprecedented view of the morphology of the near-nucleus coma (\cite{Sierks2015}; \cite{Thomas2015}). The jet feature were visible in the early phases of the rendezvous mission (\cite{2015A&A...583A..A11}; \cite{2015A&A...583A..A9}), and their development was closely monitored since then. Due to the orientation of its rotational axis and its complex shape, and a rotation period of 12.4 hours (Mottola et al. 2015), significant diurnal variations and seasonal effect of the gas flow and dust coma structure were detected (\cite{Gulkis2015}; \cite{2015Sci...347a0276H}).
\begin{figure}
\includegraphics[width=0.5\textwidth]{JetvsTime3e.eps}
\caption{Gradual growth of the dust coma and dust jets of comet 67P: (a) image taken on 5 August, 2014, when the heliocentric distance r$_h$= 3.60 au, only a small dust jet can be seen emanating from the Hapi region; (b) 9 February, 2015 for r$_h$ = 2.35 au, the formation of a main jet is accompanied by a few fainter jets originating from other regions; (c) 10 May, 2015 for r$_h$ = 1.67 au, a system of bright jets appearing on the sunward side of the coma. The image contrast level is adjusted to log scale, ranging from -4 to -7.}
\label{JetsTime}
\end{figure}
Figure 1 shows the time development of the dust coma and jets as 67P approaches the Sun. From August 2014 ($r_h$ $\sim$3.60 au) to May 2015 ($r_h$ $\sim$1.67 au), the dust coma became more dense with the same image contrast level. At the beginning, the Hapi region located in the neck between the two lobes appeared to be the main source of the water gas flow (\cite{Gulkis2015}; \cite{2015A&A...583A..A3}; \cite{2015Sci...347a0276H}) and dust jets (\cite{Sierks2015}; \cite{2015A&A...583A..A11}; \cite{2015A&A...583A..A9}; \cite{Vincent2015a}). All the dust jets have very straight configurations perpendicular to the surface, suggesting efficient acceleration of the embedded solid grains to radial speeds far exceeding the angular velocity ( $V_r$$\sim$1-2 m s$^{-1}$) of the nucleus due to its rotation. It was therefore surprising that a jet structure with large curvature appeared in late May (Fig. 2). This is the first time that a spiral structure was seen in the near-coma region of a comet made possible because of the close distance of the Rosetta spacecraft to the comet nucleus. Nevertheless, from ground-based observations (\cite{1991Icar...93..194S}, \cite{2007A&A...469..771L}, \cite{2012A&A...537A.101L}, \cite{2013AJ....146....4L}), we did see the repeatability of a curved appearance of the gaseous and dust jets related to the rotation of the nucleus or of the dust jets forming to dust tail due to the solar gravity and radiation pressure. However, these ground-based observations have much larger scales than Rosetta observations and the important physical processes might not be the same as the curved jets observed by ROSETTA. The curved jets persisted about two months and disappeared in early-August, 2015. In this work, we will examine the observed properties and dynamics of the curved jet as well as the localization of its possible source region.
The paper is organized as follows. Section 2 describes the morphology and time variation of the curved jet in comparison to other collimated dust jet features. The results of a set of computations making use of the gravity field model of the comet nucleus and of the Direct Simulation Monte Carlo (DSMC) simulation are presented in Section 3. A discussion on the theoretical results and the OSIRIS imaging data is given in Section 4.
\section{Observations and data analysis}\label{obs}
\subsection{Observations on May 30-31}
Figure 2 shows the time sequence of the dust coma of 67P in approximately one nucleus rotation on May 30-31, 2015. The sun is towards the top and that is the reason why a clear shadow was cast behind the nucleus. The spin axis was pointing away from the projection plane with the rotation in the clockwise direction. According to the shape model (\cite{Sierks2015}; \cite{2015A&A...583A..A33}; \cite{Jorda2015}), the sub-solar point during this observation was close to the equator of comet 67P. We note that three different components can be identified in Fig.2.a. The brightest one (J1) can be traced to the middle of the Hapi region. On its left side an array of jets (e.g., J2) became clearly visible in Fig. 2.b because of the sun-lit effect on Imhotep. Last but not least, on the right-hand side a jet (J3) in the form of a spiral arm came into view in Fig. 2.b. It is less discernable in subsequent images as a result of the orientation of the spacecraft relative to the rotating nucleus and/or decrease of the dust production. The spiral jet reappeared again in Fig. 2.h just one rotation period later.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{CurvedJet_May_All.eps}
\end{center}
\caption{Jet structures obtained with the wide-angle camera from 13:03 UT on May 30 to 07:28 UT on May 31, 2015. Sub-panels (a) to (h) are separated by about two hours between two frames. The spatial scales and field of view range from 6.72 m/px,13.75 km (fig. a) to 7.01 m/px, 14.36 km (fig. h). }
\label{CurvedJetAll}
\end{figure}
\subsection{Source region of the Curved jet}
In order to find the source region of the curved jet, we both used the method described in Lin et al (2015) and the jet inversion method tracing the orientation of the colliminated beams back to their emission points (\cite{Vincent2015b}, \cite{2015A&A...583A..A9}). The shape model constructed by Jorda et al (2015) and current SPICE kernels have been used for this purpose. Nonetheless, it is difficult to obtain a precise location of the source region from the OSIRIS images taken from late-May to early-June, 2015 because of the diffusive structure of the jet. The most probable source region is located between Nut and Serget (see Figure 3 for a context image of where these regions are located on the nucleus). At closer scrutiny, it can be seen that this region is covered by smooth deposits of fine materials which are likely the result of airfall of low velocity particles, not being able to escape from the nucleus surface (\cite{Thomas2015}, \cite{2015A&A...583A..A41} and reference in). However when looking closely at Nut region there is a more granular deposit including boulders with diameters up to a few tens of meters. A part of this area contains "pit-like" features with sub-meter to 4-5 meter diameter (see fig. 7 in \cite{2015A&A...583A..A41}) that might be remnants formed by wind erosion or sublimation of the volatile-rich blocks. The curved jet appeared for two more months since the first detection. As a consequence, the features of the source region surface might have changed. A study of possible changes in localized areas of the identified source region would however be possible only in the later phase of the Rosetta mission, when high resolution images at spatial scales as small as 0.5 m/px will be available again.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{SR_curvedJet2.eps}
\end{center}
\caption{Geological map of 67P observed from the top of the head lobe and nomenclature of the geological regions (modified from Fig. 2 in \cite{2015A&A...583A..A26} and the possible source regions, red-circles, of the curved jet. The latitude of the possible source region is about 6 degree at the northern hemisphere. The subsolar point at the time of the present observation (31, May, 2015) is 9 degree at the southern hemisphere and moves toward to higher latitude until the beginning of September, 2015. The right-panel is the NAC image obtained on September 19.5, 2014 with a resolution about 0.53 m/px.}
\label{SR_CurvedJet}
\end{figure}
\section{Numerical Simulation}
To examine the origin and dynamical evolution of the curved jet, we need to consider the trajectories of different dust particles with different sizes under the influence of the gravitational attraction of the nucleus. Because of its highly irregular shape, the gravitational field have been computed by dividing the whole object into 33681 elements - according to the shape model with homogeneous structure and a bulk density of 532 kg m$^{-3}$ (\cite{Jorda2015}). Figure 4.a shows the contour plot of the gravitational field in the vicinity of the nucleus. At distances larger than three nucleus radii ($\sim$6 km) the gravitational field can be reasonably approximated by that from a point mass. However, close to the nucleus surface, the field distribution is far from spherical symmetry.
Figure 4.b shows the flow field of the coma gas expanding from the nucleus surface. To treat the transition from a collisional region close to the central nucleus to the collisionless coma at large distance, the Direct Simulation Monte Carlo (DSMC) method is required (\cite{Bird1994}). The basic structure of the DSMC code used for obtaining this result has been described in detail in \cite{Wu2004166}, \cite{Su20101136}, \cite{Lai2016} et al, and \cite{Liao2015} and will not be repeated here. For this simulation, we assume the water production rate is 10$^{27}$ molecules s$^{-1}$ and the sunlit portion of the nucleus surface at the time of consideration has been assumed to be all active in outgassing. Thus, a uniform gas production rate in sunlit side is Z = 4.8$\times$10$^{19}$ H$_2$O molecules m$^{-2}$ s$^{-1}$. The initial velocity distribution of the gas is described by a half-Maxwellian distribution with a thermal temperature of 228 K. The sunlit side is assumed to be free of gas outflow even though we know this is not necessarily true according to both the Rosina measurements (\cite{2015Sci...347a0276H}) and the recent discovery of the so-called night-side outbursts or Sunset jet activities by the OSIRIS camera team (\cite{Knollenberg2015} submitted and \cite{Xian2016}). It is clear that a certain level of weak outgassing activity existed on the nightside of the nucleus. Also, there can be activity driven by other gases (CO$_2$) - as may be the case for both the sunrise jets and for the sunset jets, and as was seen by Haessig (CO$_2$/H$_2$O = 4). However, even under such circumstances, the number distribution of the gas molecules should be highly non-isotropic within 5-10 nucleus radii covered by the simulation box. In a different study (\cite{Lai2016} et al), it has been demonstrated that the global gas flows tend to follow streamlines perpendicular to the surface of their source regions if the gas emission rate (Z) on the illuminated side is assumed to be proportional to square root $\cos \theta$ where $\theta$ is the solar zenith angle where $\theta$ $>$0. For the whole surface, Z = 0.1$\times$ Z$_0$ where Z$_0$ is the peak sublimation rate at $\theta$= 0. It is interesting to note that, even for localized outburst events, the dust jets appeared to be highly collimated (J-B.Vincent, private communication, 2016).
\begin{figure}
\begin{center}
\includegraphics[width=0.4\textwidth]{Gravity_GasFlow1.eps}
\end{center}
\caption{ (a) A contour plot of the gravitational field of comet 67P in the XY plane containing the long and short axes. The rotational axis is in the perpendicular direction. The assumed bulk density is $\rho$= 532 kg m$^{-3}$. (b) The flow field of the expanding gas outflow. }
\label{CurvedJet_GG}
\end{figure}
In the simulation code, the motion of a dust particle of mass m with initial zero velocity is determined by the viscous drag effect of the expanding gas flow and the gravitational attraction of the nucleus. Note that the effects of solar radiation pressure on curved jet are not considerable in our simulation but will be involved in the future. The used equation of motion is shown below. (\cite{Gombosi1986}; \cite{1999Icar..140..173S}; \cite{2008EM&P..102..521M}). \\
\begin{equation}\label{color_measure}
m\frac{dv}{dt} = m\vec{g}+\frac{1}{2}A\rho_gC_dv_r^2
\end{equation}
where m is the mass of dust particle, and g is the "local" gravity. The second term on the right-hand side represents the gas drag effect. The dust particle of cross section A is assumed to be spherical and its density ($\rho$) is taken to be 1000 kg m$^{-3}$ (\cite{2015ApJ...802L..12F}). In Equation 1, v$_r$ is the relative velocity between the gas molecules and the dust particle, $\rho$$_g$ is the mass density of gas flow, and C$_d$ is the drag coefficient. Note that C$_d$ = 2 (\cite{Wallis1982}).
To demonstrate the combined effect of the gas drag and nucleus gravity, a source region is tentatively chosen in the Nut region (see Figure 3) with zero initial velocity. The general idea is to examine the dependence of particle size, gas production rate and emission location on the jet dynamics.
\section{Result and conclusions}
Figure 5.a illustrates the obtained velocity profiles of dust grains with radius ranging between 1 $\mu$m and 1 mm for the gas sublimation rate of Z = 4.8$\times$10$^{19}$ H$_2$O molecules m$^{-2}$ s$^{-1}$. The bigger dust grains will be accelerated to lower radial velocity as compared with the particles of smaller sizes. This effect can be understood in terms of the size dependence of the gas drag force under the influence of the gravitational attraction of the comet nucleus. In addition, the dust grains are accelerated to their terminal speeds within a distance of the order of 2 km or slightly more in all the considered cases. The small micron-sized dust could reach an outflow speed as high as 30 m s$^{-1}$ while those of mm-size have terminal speed of the order of 0.5 m s$^{-1}$ (or less) which is comparable to the rotational speed of the nucleus at its surface. This immediately suggests that the grains in the curved jet must be relatively large (i.e, d $\sim$0.1-1 mm). The result is consistent with previous findings (\cite{2015Sci...347a3905R}) that the optically dominant particles in 67P coma are exactly those of size of 0.1-1 mm. In additional to the curved jet, the straight jets (Fig. 2 J1 and J2) might consist of relative smaller particles. The identification of the exact sizes would depend on the characteristics of the surface material and the effective outgassing rate which controls the gas drag acceleration.
\begin{figure}
\begin{center}
\includegraphics[width=0.4\textwidth]{velocity4.eps}
\end{center}
\caption{(a)The velocity profiles of dust grains of different sizes with sublimation rate Z = 4.8$\times$10$^{19}$ H$_2$O molecules m$^{-2}$ s$^{-1}$. (b) Trajectories of dust grains of different radii (from 1$\mu$m to 1mm) as the nucleus rotates. }
\label{CurvedJet_V}
\end{figure}
It is perhaps not an accident that, the curved jet was observed to be emitted near the equatorial region of the head of 67P (Nut, Serqet and Ma'at ), i.e. where the nucleus rotation speed is the largest. As mentioned before in Section 2, the collimated jets of linear configuration seemed to have been all emanated from the Hapi region that is more or less along the spin axis, where the centrifugal force is minimal.
From our data analysis and preliminary numerical simulation, it is now understood that the appearance of a curved jet in May and June 2015 is caused by a combination of the ejection of mm-sized dust grains from the equatorial source region in the vicinity of Nut, Serqet and Ma'at. This unique set of OSIRIS observations provides important information on the physical properties of the dust grains and on the acceleration process.
\begin{acknowledgements}
\tiny
OSIRIS was built by a consortium led by the Max-Planck-Institut f\"{u}r Sonnensystemforschung, G\"{o}ttingen, Germany, in collaboration with CISAS, University of Padova, Italy, the Laboratoire d'Astrophysique de Marseille, France, the Instituto de Astrof\'{\i}sica de Andaluc\'{\i}a, CSIC, Granada, Spain, the Scientific Support Office of the European Space Agency, Noordwijk, Netherlands, the Instituto Nacional de T\'{e}cnica Aeroespacial, Madrid, Spain, the Universidad Polit\'{e}chnica de Madrid, Spain, the Department of Physics and Astronomy of Uppsala University, Sweden, and the Institut f\"{u}r Datentechnik und Kommunikationsnetze der Technischen Universit\"{a}t Braunschweig, Germany. The support of the national funding agencies of Germany (Deutschen Zentrums f\"{u}r Luft- und Raumfahrt), France (Centre National d'Etudes Spatales), Italy (Agenzia Spaziale Italiana), Spain (Ministerio de Educaci\'{o}n, Cultura y Deporte), Sweden (Swedish National Space Board; grant no. 74/10:2), and the ESA Technical Directorate is gratefully acknowledged. This work was also supported by grant number NSC 102-2112-M-008-013-MY3 and NSC 101-2111-M-008-016 from the Ministry of Science and Technology of Taiwan. We are indebted to the whole Rosetta mission team, Science Ground Segment, and Rosetta Mission Operation Control for their hard work making this mission possible.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{#1}\setcounter{equation}{0}}
\newcommand{\LM}[1]{\hbox{\vrule width.2pt \vbox to#1pt{\vfill \hrule
width#1pt
height.2pt}}}
\newcommand{{\mathchoice {\>\LM7\>}{\>\LM7\>}{\,\LM5\,}{\,\LM{3.35}\,}}}{{\mathchoice {\>\LM7\>}{\>\LM7\>}{\,\LM5\,}{\,\LM{3.35}\,}}}
\iffalse
\def \trait (#1) (#2) (#3){\vrule width #1pt height #2pt depth #3pt}
\def \qed{\hfill
\trait (0.1) (6) (0)
\trait (6) (0.1) (0)
\kern-6pt
\trait (6) (6) (-5.9)
\trait (0.1) (6) (0)
\medskip}
\fi
\renewcommand{\theenumi}{\arabic{enumi}.}
\renewcommand{\labelenumi}{\arabic{enumi}.}
\renewcommand{\theenumii}{(\roman{enumii})}
\renewcommand{\labelenumii}{(\roman{enumii})}
\newcommand{\,dx}{\,dx}
\newcommand{\,dt}{\,dt}
\newcommand{\,dy}{\,dy}
\newcommand{\rightharpoonup}{\rightharpoonup}
\newcommand{; {\it i.e., }}{; {\it i.e., }}
\newcommand{\varepsilon}{\varepsilon}
\newcommand{\overline x}{\overline x}
\newcommand{{L}^p}{{L}^p}
\newcommand{\lp(\Omega)}{{L}^p(\Omega)}
\newcommand{\lQ}{{L}^q}
\newcommand{\lQ(\Omega)}{\lQ(\Omega)}
\newcommand{{L}^1}{{L}^1}
\newcommand{\lu(\Omega)}{{L}^1(\Omega)}
\newcommand{{L}^\infty}{{L}^\infty}
\newcommand{\li(\Omega)}{{L}^\infty(\Omega)}
\newcommand{{L}^{p'}}{{L}^{p'}}
\newcommand{\lpp(\Omega)}{{L}^{p'}(\Omega)}
\newcommand{{W}^{1,q}}{{W}^{1,q}}
\newcommand{{W}^{1,p}}{{W}^{1,p}}
\newcommand{{W}^{1,n}}{{W}^{1,n}}
\newcommand{{W}^{1,2}}{{W}^{1,2}}
\newcommand{\wup(\Omega)}{{W}^{1,p}(\Omega)}
\newcommand{{W}^{1,1}}{{W}^{1,1}}
\newcommand{\wuu(\Omega)}{{W}^{1,1}(\Omega)}
\newcommand{{W}^{1,p}_0}{{W}^{1,p}_0}
\newcommand{\wupz(\Omega)}{{W}^{1,p}_0(\Omega)}
\newcommand{{W}^{1,\infty}}{{W}^{1,\infty}}
\newcommand{\wui(\Omega)}{{W}^{1,\infty}(\Omega)}
\newcommand{C^\infty_0}{C^\infty_0}
\newcommand{(\Omega;\Bbb R^m)}{(\Omega;\Bbb R^m)}
\newcommand{{\rm loc}}{{\rm loc}}
\newcommand{\divv}{{\rm div}\,}
\newcommand{{\rm dist}\,}{{\rm dist}\,}
\newcommand{{\rm spt}\,}{{\rm spt}\,}
\newcommand{{\rm rank}\,}{{\rm rank}\,}
\newcommand{{\rm sign}\,}{{\rm sign}\,}
\newcommand{{\overline u}}{{\overline u}}
\newcommand{\Hom}{{\rm hom}}
\newcommand{{\cal F}}{{\cal F}}
\newcommand{{\cal F}_{\rm hom}}{{\cal F}_{\rm hom}}
\newcommand{{x\over \e}}{{x\over \varepsilon}}
\newcommand{\|}{\|}
\newcommand{-\hskip -.38cm\int}{-\hskip -.38cm\int}
\newcommand{-\hskip -.318cm\int}{-\hskip -.318cm\int}
\newcommand{{\scriptstyle {-\hskip-.29cm\int}}}{{\scriptstyle {-\hskip-.29cm\int}}}
\newcommand{u.a.p.~}{u.a.p.~}
\newcommand{T_\eta}{T_\eta}
\newcommand{{(0,t)^n}}{{(0,t)^n}}
\newcommand{{\bf z}}{{\bf z}}
\newcommand{L_\eta}{L_\eta}
\newcommand{\Omega}{\Omega}
\newcommand{\omega}{\omega}
\newcommand{{D}_{\rm p}}{{D}_{\rm p}}
\font\tenmsb=msbm10
\font\sevenmsb=msbm7
\font\fivemsb=msbm5
\newfam\msbfam
\textfont\msbfam=\tenmsb
\scriptfont\msbfam=\sevenmsb
\scriptscriptfont\msbfam=\fivemsb
\def\Bbb#1{{\fam\msbfam\relax#1}}
\def\Bbb R{\Bbb R}
\def\Bbb R{\Bbb R}
\def\Bbb M{\Bbb M}
\newcommand{\Bbb R^n}{\Bbb R^n}
\newcommand{\Bbb R^m}{\Bbb R^m}
\newcommand{\overline\Bbb R}{\overline\Bbb R}
\newcommand{{\Bbb N}}{{\Bbb N}}
\newcommand{{\Bbb Q}}{{\Bbb Q}}
\newcommand{{\Bbb Z}}{{\Bbb Z}}
\newcommand{{\Bbb Z}^n}{{\Bbb Z}^n}
\newcommand{{\Bbb M}^{m\times n}}{{\Bbb M}^{m\times n}}
\newcommand{{\Bbb M}^{3\times 3}}{{\Bbb M}^{3\times 3}}
\newcommand{{\Bbb M}^{3\times 2}}{{\Bbb M}^{3\times 2}}
\newcommand{{\Bbb M}^{2\times 2}}{{\Bbb M}^{2\times 2}}
\newcommand{\Aa}{{\cal A}}
\newcommand{{\cal O}(\Omega)}{{\cal O}(\Omega)}
\newcommand{{\cal O}(\Bbb R^n)}{{\cal O}(\Bbb R^n)}
\newcommand{{\cal H}^2}{{\cal H}^2}
\newcommand{{\cal H}^1}{{\cal H}^1}
\newcommand{{\cal H}^{N-1}}{{\cal H}^{N-1}}
\newcommand{\hbox{\rm int}}{\hbox{\rm int}}
\def\buildrel * \over \wto{\buildrel * \over \rightharpoonup}
\def\mbox{\,a.e.}{\mbox{\,a.e.}}
\defD_{\alpha}{D_{\alpha}}
\defx_{\alpha}{x_{\alpha}}
\defy_{\alpha}{y_{\alpha}}
\defz_{\alpha}{z_{\alpha}}
\def\overline{F}{\overline{F}}
\def\bar{u}{\bar{u}}
\def\rightarrow{\rightarrow}
\def\mbox{ess sup}{\mbox{ess sup}}
\def\mbox{ess inf}{\mbox{ess inf}}
\def\spadesuit{\spadesuit}
\def\, dx_{\alpha}{\, dx_{\alpha}}
\def\bf H{\bf H}
\def{x^\e_i}{{x^\varepsilon_i}}
\def{ x_i^{\e,T}}{{ x_i^{\varepsilon,T}}}
\def{B^\e_i}{{B^\varepsilon_i}}
\def{Q^\e_i}{{Q^\varepsilon_i}}
\def\gamma{\gamma}
\begin{document}
\title{Minimising movements for oscillating energies: \\
the critical regime}
\author[1]{Nadia Ansini
\author[2]{Andrea Braides
\author[3]{Johannes Zimmer
\affil[1]{Dept. of Mathematics,
Sapienza University of Rome, P.le Aldo Moro 2, 00185 Rome, Italy}
\affil[2]{Dept. of Mathematics,
University of Rome `Tor Vergata', Via della Ricerca Scientifica, 00133 Rome, Italy}
\affil[3]{Dept. of Mathematical Sciences, University of Bath, Bath BA2 7AY, UK}
\date{}
\maketitle
\begin{abstract}
Minimising movements are investigated for an energy which is the superposition of a convex functional and fast small
oscillations. Thus a minimising movement scheme involves a temporal parameter $\tau$ and a spatial parameter $\varepsilon$,
with $\tau$ describing the time step and the frequency of the oscillations being proportional to $\frac 1 \varepsilon$. The
extreme cases of fast time scales $\tau\ll\varepsilon$ and slow time scales $\varepsilon\ll\tau$ have been investigated
in~\cite{Braides2014a}. In this article, the intermediate (critical) case of finite ratio $\varepsilon/\tau>0$ is studied. It
is shown that a pinning threshold exists, with initial data below the threshold being a fixed point of the
dynamics. A characterisation of the pinning threshold is given. For initial data above the pinning threshold, the
equation and velocity describing the homogenised motion are determined.
AMS Subject classification: 35B27 (49K40, 49J10 49J45)
Keywords: Gradient flow, wiggly energy, $\Gamma$-convergence, minimising movements
\end{abstract}
\section{Introduction}
\label{sec:Introduction}
In this paper we analyse a minimising-movement approach for gradient flows with wiggly energies,
\begin{equation}
\label{eq:gf}
x'(t) = - \mu \frac{\partial E_\varepsilon(x(t))}{\partial x}.
\end{equation}
A prototypical model of the energy is an oscillating perturbation of a quadratic energy,
\begin{equation}
\label{eq:quaden}
E_\varepsilon (x)= {1\over 2}x^2 - \varepsilon \cos \Bigl({x\over \varepsilon}\Bigr).
\end{equation}
This mathematical problem can be motivated by the analysis of interface motion in materials science. There is a range
of problems where interfaces form in a specimen and propagate so that a material particle crossing the interface
changes its stability, by transforming from an unstable or metastable state to a more stable one; see,
e.g.,~\cite{Abeyaratne1999a}. Often this evolution is load-driven, in the sense that an applied load enables a particle
to explore states of lower energy. Let us consider an interface, say between twin boundaries or phase boundaries,
macroscopically propagating with some velocity $v$. However, microscopically the interface typically does not move
homogeneously. Instead, the interface tends to propagate forward as a whole by a series of incremental steps. To
illustrate this, let us picture an interface consisting of a straight horizontal line segment, then a step up, followed
by another horizontal line segment, moving up towards a more stable state. Then it is normally advantageous for the
interface to propagate the step sideways, i.e., move upward the particle next to the step, and then move the remaining
particles consecutively. This leads to a multi-welled energy landscape, with local minima spaced periodically with high
frequency. One model for the propagation of an interface in this manner is the Frenkel-Kontorova chain with forcing,
where the motion of atom $n$ is~\cite{Abeyaratne1999a}
\begin{equation*}
m u_n''(t) = k (u_{n+1}(t) - 2 u_n(t) + u_{n-1}(t)) - W'(u_n(t)) + f(n,t),
\end{equation*}
or the continuum version
\begin{equation*}
m u''(x,t) = k u_{xx}(x,t) - W'(u(x,t)) + f(x,t).
\end{equation*}
The model considered here can be interpreted as an unforced ($f=0$) case, where the kinetics is replaced by a simpler
(gradient flow) dynamics.
We remark that the same equation appears in a related but different context in Materials Science, again originating
from transition layers. Martensitic materials can form needles of phases with pronounced tips (see, e.g., photographs
in~\cite{Abeyaratne1996a}). During creep tests, it is observed that the volume fraction of the phase fractions involved
changes rather abruptly, and it is shown that this sudden change can be attributed to a sudden split of a tip into two
tips~\cite{Abeyaratne1996a}. One can picture this as a lenticular domain of one variant trying to grow; this growth
then occurs where the tip of the lens meets a boundary between twins, and fattening of the phase happens via splitting
of the tip in two and more tips. The splitting of a needle can then be attributed to a metastable transition, moving
from one local minimum to another one. This suggests a small-scale landscape with many minima, and the energy studied
by Abeyaratne, Chu and James~\cite{Abeyaratne1996a} is a macroscopic energy augmented by small-scale oscillations
$a \varepsilon \cos(\frac x \varepsilon)$, as studied here. In addition, the kinetic law in~\cite{Abeyaratne1996a} is taken to be a
gradient flow. Specifically, there it is shown that the solution $x_\varepsilon$ to the evolution equation~\eqref{eq:gf}
converges uniformly in time to the solution of
\begin{equation*}
x(t) = - \mu \frac{\partial \bar E(x(t))}{\partial x}, \text{ with } x(0) = x_0,
\end{equation*}
with an explicitly computed driving force $\frac{\partial \bar E}{\partial x}$. This latter system is then investigated
numerically. The variational analysis carried out here can be interpreted in this light. We consider time
discretisations, as numerical algorithms would employ, but on the level of the original (not homogenised) energy
$E_\varepsilon$, rather than $\bar E$. This leads to two parameters, the time discretisation $\tau$ and oscillation scale
$\varepsilon$. The different scaling regimes that follow naturally are analysed in this paper.
We make the trivial but important remark that the limit of a sequence of gradient flows associated to an family $E_\varepsilon$
is in general not the gradient flow of the limit of the energy. For example, on one hand for the
energy~\eqref{eq:quaden}, the associated gradient flow is ($\mu=1$)
\begin{equation*}
x_\varepsilon'(t)=-x_\varepsilon(t)-\sin \Bigl({x_\varepsilon(t)\over \varepsilon}\Bigr),
\end{equation*}
for initial datum $x_0$. If $x_0\in (-1,1)$, then such solutions are trapped between stationary solutions, and they
converge to the trivial constant state $x_0$ (\emph{pinning}), while if $|x_0|\ge 1$ they can be shown to converge to a
solution $x$ of the gradient flow
\begin{equation*}
x'(t)=- {\rm sign}\,{x(t)}\sqrt{x^2(t)-1}.
\end{equation*}
On the other hand, $E_\varepsilon$ converge uniformly to the quadratic energy, whose gradient flow is trivially
\begin{equation*}
x'(t)=-x(t).
\end{equation*}
These behaviours can be obtained as limit cases of \emph{minimising movements along the sequence of energies $E_\varepsilon$} at
different \emph{time scales}. Minimising movements are defined as follows: with fixed $\varepsilon$ (the spatial scale) and
$\tau$ (the time scale), we set $x^{\varepsilon,\tau}_0= x_0$ and choose recursively $x^{\varepsilon,\tau}_k$ as a minimiser of
\begin{equation*}
x\mapsto E_\varepsilon(x)+{1\over 2\tau}|x-x^{\varepsilon,\tau}_{k-1}|^2.
\end{equation*}
This process gives the piecewise-constant trajectories
\begin{equation*}
x^{\varepsilon,\tau}(t)=x^{\varepsilon,\tau}_{\lfloor t/\tau\rfloor}.
\end{equation*}
With fixed $\tau=\tau(\varepsilon)$, a minimising movement $x(t)$ along the sequence of energies $E_\varepsilon$ at time scale $\tau$ is defined as any limit of subsequences of $x^{\varepsilon,\tau}(t)$. Simple examples show that the limit may indeed depend on the subsequence and on the choice of $\tau$. If $E_\varepsilon$ is independent of $\varepsilon$ this notion coincides with the one given by De Giorgi~\cite{De-Giorgi1993a} and at the basis of modern notions of gradient flows (see the monograph by Ambrosio, Gigli and Savar\'e~\cite{Ambrosio2008b}). Examples of problems related to varying $E_\varepsilon$ are analysed in~\cite{Braides2014a}, where in particular it is shown that for the energies above and for $\tau\ll\varepsilon$ (\emph{fast time scale}), the minimising movement $x$ coincides with the limit of the solutions $x_\varepsilon$ of the gradient flows at fixed $\varepsilon$, while for $\varepsilon\ll\tau$ (\emph{slow time scale}) it coincides with the gradient flow of the limit quadratic energy. That observation highlights the existence of a \emph{critical time scaling} when $\tau\sim\varepsilon$, for which the minimising movements are not trivially described by the limit of gradient flows or the gradient flow of the limit. The behaviour at those scales is the object of this paper. A rather different very interesting line of investigation has been taken by Menon~\cite{Menon2002a}, and independently in parallel by Smyshlyaev. In~\cite{Menon2002a}, averaging techniques are developed in the context of the time-continuous dynamical system~\eqref{eq:gf}. The homogenisation of first-order ordinary differential equations, including error estimates, is studied further in~\cite{Ibrahim2010a}. We remark that the model we consider is deterministic, where the two parameters come from spatial oscillations and a time discretisation. For stochastic models, it is also natural to consider the effective behaviour in different scaling regimes of space and noise; we refer the reader to~\cite{Dupuis2012a}.
\paragraph{Plan of the paper} A summary of the results of the paper is as follows. We consider functions
$E_\varepsilon\colon \Bbb R\rightarrow \Bbb R$ given by
\begin{equation}
\label{funzioni}
E_\varepsilon (x)= h(x) + \varepsilon\, W \Bigl({x\over \varepsilon}\Bigr),
\end{equation}
where $h$ is a strictly convex function bounded from below and $W$ is a one-periodic even Lipschitz function. We
consider a time scale $\tau=\tau(\varepsilon)$ such that $\varepsilon/\tau$ converges to $\gamma>0$. Therefore, the present analysis
complements the recent one of~\cite{Braides2014a} {(see also~\cite{Braides2008a})}, where the cases
$\gamma \in \{0, +\infty\}$ are investigated. In terms of the mechanical problem of interface propagation discussed
above, we show that pinning will occur for small initial data, while large data leads to a gradient flow evolution for
which the averaged velocity can be computed. More precisely, we prove that in that case the unique minimising movement
$x^\gamma$ with initial datum $x_0>0$ (the case $x_0<0$ is analogous by symmetry) is described as follows:
\begin{enumerate}
\item \emph{Pinning threshold:\/} There exists $T_\gamma$ such that $x^\gamma(t)=x_0$ for all $t$ if
$|x_0|\le T_\gamma$. The pinning threshold is characterised in Proposition~\ref{carac}.
\item \emph{Homogenised equation:\/} If $x_0>T_\gamma$ then $x^\gamma(t)$ is characterised as the non-increasing
function satisfying
\begin{equation*}
{\rm d\over dt}x^\gamma(t)= -\gamma\, f_\gamma(h'(x^\gamma(t)))
\end{equation*}
at almost all $t>0$. The \emph{homogenised velocity} $f_\gamma(z)$ is the \emph{average velocity} (suitably defined)
of any discrete orbit $\{y_k\}$ defined recursively by minimisation of the linearity problem
\begin{equation*}
y\mapsto z y+W(y)+{\gamma\over 2}(y-y_{k-1})^2,
\end{equation*}
which can be shown not to depend on the initial condition $y_0$.
\end{enumerate}
Mathematically, our analysis is confined to one space dimension, as it strongly relies on monotonicity properties
developed in Section~\ref{sec:Monot-behav-minim}. A central argument is a comparison of solution to the nonlinear
energy as in~\eqref{funzioni}, and one where $h$, the non-oscillating part, is suitably linearised. This argument is
developed in Section~\ref{sec:Linearized-energy}.
\section{Minimising movement along a sequence}
\label{sec:Minim-movem-along}
We recall the general definition of minimising movements for a sequence of functionals defined on a Hilbert space.
\begin{definition}
\label{DefMMseq}
Let $X$ be a separable Hilbert space, $E_\varepsilon \colon X\rightarrow [0, +\infty)$ equicoercive and lower semicontinuous and
$x_0^{\varepsilon}\rightarrow x_0$ with $E_\varepsilon (x_0^{\varepsilon}) \le C<+\infty$ and $\tau_\varepsilon >0$ converging to $0$ as $\varepsilon\rightarrow 0$. For fixed
$\varepsilon>0$, we define recursively $x_i^{\varepsilon}$ as a minimiser of the problem
\begin{equation}
\label{minpMM}
\min\Bigl\{ E_\varepsilon (x) + {1\over \tau_{\varepsilon}} \Vert x- x_{i-1}^{\varepsilon}\Vert^2\Bigr\}\,,
\end{equation}
and the piecewise-constant trajectory
\begin{equation}
\label{seqMM}
x^\varepsilon (t):= x_i^\varepsilon= x^\varepsilon_{\lfloor t/\tau_{\varepsilon}\rfloor}\,, \quad t\in [i\tau_\varepsilon, (i+1)\tau_\varepsilon)\,.
\end{equation}
A \emph{minimising movement for $E_\varepsilon$ at time scale $\tau$} from $x_0^\varepsilon$ is the limit of a subsequence $x^{\varepsilon_j}$,
\begin{equation*}
x(t)= \lim_{j\rightarrow +\infty} x^{\varepsilon_j}(t)\,,
\end{equation*}
with respect to the uniform convergence on compact sets of $[0, +\infty)$.
\end{definition}
This definition is justified by the following compactness result~\cite[Proposition 7.1]{Braides2014a}.
\begin{proposition}
For every $E_\varepsilon$ and $x_0^{\varepsilon}$ as above, there exist minimising movements for $E_\varepsilon$, from $x_0^{\varepsilon}$, with
$x(t)\in C^{1/2} ([0, +\infty); X)$.
\end{proposition}
For a comprehensive study of minimising movements for a fixed $E=E_\varepsilon$ we refer to~\cite{Ambrosio2008b}, while a
detailed analysis of some of its applications can be found in~\cite{Braides2014a}.
\section{Monotone behaviour of minimising movements}
\label{sec:Monot-behav-minim}
In the sequel we will study minimising movements for the functions $E_\varepsilon\colon \Bbb R\rightarrow \Bbb R$ given by~\eqref{funzioni},
where $h$ is a strictly convex function bounded from below. It is not restrictive to suppose that $h\ge 0$, and that
$h$ attains its global minimum in $x=0$. Furthermore, we assume that $W$ is a one-periodic even Lipschitz function with
$\Vert W^\prime\Vert_{\infty}=1$, and that the average of $W$ is $0$. The two latter assumptions serve as normalisation
only and are not restrictive.
We observe the following simple monotonicity property.
\begin{proposition}\label{monof}
Given any functions $\phi\colon\Bbb R\rightarrow \Bbb R$ and $\psi\colon\Bbb R\rightarrow \Bbb R$, and $\beta>0$, for any $x,x'\in\Bbb R$, let
$y,y'\in\Bbb R$ be minimisers of
\begin{equation*}
t\mapsto \phi(t)+ \beta (t- x)^2\,,\qquad
t\mapsto \psi(t)+ \beta (t- x')^2,
\end{equation*}
respectively. Then
\begin{equation}
\label{phi-psi}
\phi(y)-\phi(y') + \psi(y')-\psi(y)\le 2\beta (x-x')(y-y')\,.
\end{equation}
In particular, if $\psi=\phi$ and $x\le x'$ then $y\le y'$.
\end{proposition}
\noindent{\sc Proof.}\
By assumption
\begin{equation*}
\phi(y)+ \beta (y- x)^2 \le\phi(y') + \beta (y'- x)^2,
\quad
\psi(y')+ \beta (y'- x')^2 \le\psi(y)+ \beta (y- x')^2.
\end{equation*}
Summing up the two inequalities and simplifying the terms on both sides we obtain~\ref{phi-psi}. Moreover, if
$\psi=\phi$, then
\begin{equation*}
(x-x')(y-y')\ge 0,
\end{equation*}
which yields the desired inequality.
\qed
Before analysing the case of fixed ratio $\varepsilon/\tau$, we make some general remarks. According to
Definition~\ref{DefMMseq}, we define iteratively the global minimiser $x_{i+1}^\varepsilon$ to
\begin{equation}
\label{F(i)general}
F_{\varepsilon}(x, {x^\e_i})= h(x) + \varepsilon \,W \Bigl({x\over \varepsilon}\Bigr) + {1\over 2{\tau_\varepsilon}} (x- {x^\e_i})^2\,.
\end{equation}
First of all, we observe that the sequence of minimisers $(x_{i}^\varepsilon)_i$ is monotone.
\begin{proposition}[Monotone behavior of ${x^\e_i}$]
\label{Mono-xie}
Let $x_{i+1}^\varepsilon$ be a minimiser to~\eqref{F(i)general}. Then the following holds.
\begin{enumerate}
\item \label{it:monbeh1} If $x^\varepsilon_{i+1}\le {x^\e_i}$, then $ x_{i+2}^\varepsilon\le x_{i+1}^\varepsilon$\,.
\item \label{it:monbeh2} If $x^\varepsilon_{i+1}\ge {x^\e_i}$, then $ x_{i+2}^\varepsilon\ge x_{i+1}^\varepsilon$\,.
\end{enumerate}
In particular, $t\mapsto x^\varepsilon(t)$ and $t\mapsto x(t)$ are monotone functions.
\end{proposition}
\noindent{\sc Proof.}\ \ref{it:monbeh1} and~\ref{it:monbeh2} are straightforward consequences of Proposition~\ref{monof} with
$\phi(t)=\psi(t)= h(t) +\varepsilon W(t/\varepsilon)$ and $\beta = 1/(2\tau_{\varepsilon})$.
By Definition~\ref{DefMMseq}, $x^\varepsilon(t)$ is monotone in $t$, and since it converges uniformly to
$x(t)\in C^{1/2}([0, +\infty))$, on compact sets of $[0, +\infty)$, we may conclude that $x(t)$ is also a monotone
function.
\qed
\section{Linearised energy}
\label{sec:Linearized-energy}
In order to characterise the velocity $x^\prime(t_0)$ of the minimising movement scheme~\eqref{minpMM}, we study the
average velocity given by
\begin{equation}
\label{eq:vel}
{x_i^\varepsilon- x_0^\varepsilon\over i\tau_{\varepsilon}}\,.
\end{equation}
We assume, without loss of generality, that $t_0=0$ and $x(0)=x_0^\varepsilon$. We consider a (partial) linearisation of the
problem given by
\begin{equation}
\label{F(i)linear}
F^T_{\varepsilon}(x, { x_i^{\e,T}})= Tx + \varepsilon\, W \Bigl({x\over \varepsilon}\Bigr) + {1\over 2\tau_\varepsilon} (x- { x_i^{\e,T}})^2\,,
\end{equation}
where $x_0^{\varepsilon, T}:= x_0^\varepsilon$. The term $Tx$ represents the ``linear approximation'' of the potential $h$ around the
point $h(x^\varepsilon_0)$ up to translation by a constant that does not depend on $T$ and $i$. We recall that $h$ is a
strictly convex function, hence $h^\prime$ is a monotone increasing function.
\begin{proposition}[Monotone behavior of ${ x_i^{\e,T}}$ with respect to ${x^\e_i}$]
\label{monotonicity}
Let $\delta>0$ be such that $h^\prime(x_0\pm\delta)$ exists.
\begin{enumerate}
\item Let $x^{\varepsilon,T}_{i}$ be the minimiser to~\eqref{F(i)linear}. Then \label{it:mon1}
\begin{itemize}
\item if $x^{\varepsilon,T}_{i+1}\le { x_i^{\e,T}}$, then $x_{i+2}^{\varepsilon,T}\le x_{i+1}^{\varepsilon,T}$
\item if $x^{\varepsilon, T}_{i+1}\ge { x_i^{\e,T}}$, then $x_{i+2}^{\varepsilon,T}\ge x_{i+1}^{\varepsilon,T}$\,.
\end{itemize}
\item Let $T= T^{\delta^+}:= h^\prime(x_0+\delta)$, and $x^\varepsilon_i$ minimiser of~\eqref{F(i)general}. Then if
$x^{\varepsilon,T^{\delta^+}}_{i}\le x_{i}^\varepsilon$, then $ x^{\varepsilon,T^{\delta^+}}_{i+1}\le x_{i+1}^\varepsilon$. \label{it:mon2}
\item Let $T= T^{\delta^{-}}:= h^\prime(x_0-\delta)$; then if $ {x^\e_i}\le x^{\varepsilon, T^{\delta^-}}_{i}$, then
$x_{i+1}^\varepsilon\le x^{\varepsilon,T^{\delta^-}}_{i+1}$. \label{it:mon3}
\end{enumerate}
\end{proposition}
\noindent{\sc Proof.}\
The proof of~\ref{it:mon1} is a straightforward consequence of Proposition~\ref{monof} with
$\phi(t)=\psi(t)= Tt +\varepsilon W(t/\varepsilon)$ and $\beta = 1/(2\tau_{\varepsilon})$.
To prove~\ref{it:mon2}, we apply again Proposition~\ref{monof} with $\phi(t)= h(t) +\varepsilon W(t/\varepsilon)$ and
$\psi(t)= Tt +\varepsilon W(t/\varepsilon)$ with $T= T^{\delta^+}:= h^\prime(x_0+\delta)$ and $\beta =
1/(2\tau_{\varepsilon})$. By~\eqref{phi-psi}, it holds that
\begin{equation*}
\phi(x_{i+1}^\varepsilon) - \phi(x^{\varepsilon,T^{\delta^+}}_{i+1})+ \psi(x^{\varepsilon,T^{\delta^+}}_{i+1})- \psi(x_{i+1}^\varepsilon)\le 2\beta (x^{\varepsilon,T^{\delta^+}}_{i}- {x^\e_i})
(x^{\varepsilon,T^{\delta^+}}_{i+1} - x_{i+1}^\varepsilon)\,.
\end{equation*}
Therefore,
\begin{equation*}
h(x_{i+1}^\varepsilon)- h(x^{\varepsilon,T^{\delta^+}}_{i+1}) + h^\prime(x_0+\delta) (x^{\varepsilon,T^{\delta^+}}_{i+1} - x_{i+1}^\varepsilon)\le 2\beta (x^{\varepsilon,T^{\delta^+}}_{i}- {x^\e_i})
(x^{\varepsilon,T^{\delta^+}}_{i+1} - x_{i+1}^\varepsilon)\,,
\end{equation*}
which implies
\begin{align*}
& \Biggl(h^\prime(x_0+\delta) - {h(x_{i+1}^\varepsilon)- h(x^{\varepsilon,T^{\delta^+}}_{i+1})\over x_{i+1}^\varepsilon- x^{\varepsilon,T^{\delta^+}}_{i+1} }\Biggr) (x^{\varepsilon,T^{\delta^+}}_{i+1} - x_{i+1}^\varepsilon)\\
&\le 2\beta (x^{\varepsilon,T^{\delta^+}}_{i}- {x^\e_i})
(x^{\varepsilon,T^{\delta^+}}_{i+1} - x_{i+1}^\varepsilon)\,.
\end{align*}
Since $x^{\varepsilon,T^{\delta^+}}_{i}, {x^\e_i}< x_0+\delta$ and $h$ is a convex function, we get that
\begin{equation*}
h^\prime(x_0+\delta) - {h(x_{i+1}^\varepsilon)- h(x^{\varepsilon,T^{\delta^+}}_{i+1})\over x_{i+1}^\varepsilon- x^{\varepsilon,T^{\delta^+}}_{i+1} }\ge 0\,,
\end{equation*}
which gives the monotone behaviour $x^{\varepsilon,T^{\delta^+}}_{i+1}\le x_{i+1}^\varepsilon$.
\qed
\subsection{Minimising movement for fixed ratio $\varepsilon/\tau$}
\label{sec:Minim-movem-fixed}
We consider a time scale $\tau=\tau_\varepsilon$ such that $\varepsilon/\tau$ converges to $\gamma>0$. It is not restrictive to suppose
that the ratio between $\varepsilon$ and $\tau$ is fixed,
\begin{equation*}
\tau_\varepsilon= {1\over \gamma}\, \varepsilon.
\end{equation*}
We study the linearised energies in~\eqref{F(i)linear} and rescale by $1/\varepsilon$, that is,
\begin{equation*}
{F^T_{\varepsilon}(x, { x_i^{\e,T}})\over \varepsilon} = T\Bigl({x\over \varepsilon}\Bigr) + W \Bigl({x\over \varepsilon}\Bigr)
+ {\varepsilon\over 2\tau_\varepsilon} \Bigl({x- x_{i}^{\varepsilon,T}\over \varepsilon}\Bigr)^2\,.
\end{equation*}
We denote
\begin{equation}
\label{FTgamma}
F^T_{\gamma}(y, y_i^T):= T y + W (y) + {\gamma\over 2} (y-y^T_{i})^2 \, ,
\end{equation}
where $y:= x/\varepsilon$ and $y^T_{i}:= x_{i}^{\varepsilon,T}/\varepsilon$ for every $i\in{\Bbb N}$, $i\ge 1$. Note that the minimisers $y^T_{i}$
depend also on $\gamma$. However, we omit this dependence in the notation for simplicity.
\begin{proposition}
\label{monoy}
Let $y_0, z_0$ and $T, S$ be fixed with $T\le S$. Let $y_0^T=y_0$, $z_0^S=z_0$. Let $y_{i}^T$ and $z_{i}^S$ be
minimisers to $F_{\gamma}^T(y, y^T_{i-1})$ and $F_{\gamma}^S(y, z^S_{i-1})$, respectively, for every $i\in {\Bbb N}$ with
$i\ge 1$. If $z_0\le y_0$ then $z_i^{S}\le y_i^{T}$ for every $i$.
\end{proposition}
\noindent{\sc Proof.}\ By Proposition~\ref{monof}, with $\phi(t)= St + W(t)$, $\psi(t)= Tt+W(t)$, $x=z_0$, $x^\prime=y_0$, and
$\beta= \gamma/2$, it follows that
\begin{equation*}
(S-T) (z^S_1 -y^T_1) \le \gamma (z^S_1 -y^T_1) (z_0-y_0)\,.
\end{equation*}
Therefore, if $T\le S$ and $z_0\le y_0$, this yields $z^S_1 \le y^T_1$. Similarly, we can prove that the inequality
$z_i^{S}\le y_i^{T}$ is satisfied for any $i\ge 2$. \qed
\begin{theorem}
\label{tevelo}
For every $T$, the limit
\begin{equation}
\label{fgamma}
f_{\gamma}(T):=\lim_{i\rightarrow \infty} {y_0- y_i^T\over i}
\end{equation}
exists and it is independent of $y_0$. Moreover, the function $T\mapsto f_\gamma(T)$ is monotone increasing.
\end{theorem}
\noindent{\sc Proof.}\
The existence of the limit is a straightforward consequence of the subadditivity of the sequence $(y_i^T)$. More
precisely, let $h\in {\Bbb Z}$ be such that $0\le y_k^T + h\le 1$. Since $ (y_k^T + h)^T_i= y_{k+i}^T + h$, by
Proposition~\ref{monoy}, with $S\equiv T$,
\begin{equation}
\label{ineq1}
y_i^T\le y_{k+i}^T + h\le y_i^T + 1\,.
\end{equation}
Therefore, if we sum up the last inequality in~\eqref{ineq1} with
\begin{equation*}
-1\le -(y_k^T + h)\le 0\,,
\end{equation*}
we obtain
\begin{equation*}
y_i^T+ y_k^T-1\le y_{k+i}^T \le y_i^T+ y_k^T+ 1\, ,
\end{equation*}
which implies the almost subadditivity of $(y_i^T)$. We now prove that the limit
\begin{equation}
\label{fakete}
\lim_{i\rightarrow\infty} \Bigl({y_i^T\over i}\Bigr) = \inf_{i\in{\Bbb N}} \Bigl({y_i^T\over i}\Bigl)
\end{equation}
exists. Let $i= km+n$. Then
\begin{align*}
\nonumber {y_i^T\over i} & = {y_{km+n}^T\over km+n}
\le {y_{km}^T+ y_n^T+ 1\over km+n}\\
\nonumber &\le {k y_{m}^T+ y_n^T+ k\over km+n}= {k y_{m}^T\over km+n}+ { y_n^T\over km+n}+ { k\over km+n}\,. \label{km}
\end{align*}
If we fix $m$ and pass to the limit $k$ tends to $\infty$, we obtain
\begin{equation*}
\lim_{i\rightarrow\infty} {y_i^T\over i} \le {y_m^T\over m} + {1\over m}\,.
\end{equation*}
Therefore
\begin{equation*}
\inf_{i\in{\Bbb N}}{y_i^T\over i}\le \lim_{i\rightarrow\infty} {y_i^T\over i} \le \inf_{m\in{\Bbb N}}{y_m^T\over m} \,,
\end{equation*}
which proves~\eqref{fakete} and the existence of the limit in~\eqref{fgamma}.
We now prove that the function $T\mapsto f_{\gamma}(T)$ is independent of $y_0$. In fact, we can always rewrite for
$k< i$
\begin{equation*}
{y_0- y_i^T\over i}= {y_0- y_k^T+ y_k^T- y_i^T\over i} = {y_0- y_k^T\over i} + {y^T_k- y_i^T\over i-k} {i-k\over i}\,.
\end{equation*}
Hence,
\begin{equation*}
\lim_{i\rightarrow \infty} {y_0- y_i^T\over i}= \lim_{i\rightarrow \infty} {y^T_k- y_i^T\over i-k}\,.
\end{equation*}
Finally, we remark that the function $T\mapsto f_{\gamma}(T)$ is monotone increasing; {\it i.e., } if $T\le S$ then
$f_{\gamma}(T)\le f_{\gamma}(S)$. By definition~\eqref{fgamma},
\begin{equation*}
f_{\gamma}(T)=\lim_{i\rightarrow \infty} {y_0- y_i^T\over i}\,,\qquad f_{\gamma}(S)=\lim_{i\rightarrow \infty} {y_0- y_i^{S}\over i}\,;
\end{equation*}
the monotonicity follows since $- y_i^T\le - y_i^{S}$ by Proposition~\ref{monoy}.
\qed
\subsection{Characterisation of periodic orbits for the linearised problem}
The definition of $f_\gamma(T)$ reminds that of {\em Poincar\'e rotation number} in the theory of Dynamical Systems (see, e.g.,~\cite[Chapter 11]{Katok1995a}), which, in our notation, concerns the properties of the orbits of the multifunction
\begin{equation*}
A^T_\gamma(y)={\rm argmin}\Bigl\{Tz + W(z) + {1\over 2}\gamma (z-y)^2\Bigr\}.
\end{equation*}
Note that for $T>1$ this set is a singleton, but for $T\le 1$ in general it is not. We can nevertheless adapt some
arguments borrowed from Dynamical Systems to prove a characterisation of the values of $T$ for which we have periodic
orbits. By definition in this case $f_\gamma(T)$ is rational. The converse also holds true as follows.
\begin{proposition}[Periodic orbits]
Let $T> 0$, and let $\{y_i^T\}$ be defined as in Proposition~{\rm\ref{monoy}}. There exists an initial datum
$y_0=y_0^T$ and integers $p,q$ with $q\neq 0$ such that
\begin{equation}
\label{perob}
y_{kq+i}^T= y^T_i + kp
\end{equation}
if and only if $f_\gamma(T)={p\over q}$.
\end{proposition}
\noindent{\sc Proof.}\
We only have to prove the existence of $\{y_i^T\}$ satisfying (\ref{perob}) assuming that $f_\gamma(T)=p/q$.
We remark that $A=A^T_\gamma$ satisfies \\
$\bullet$ $A$ is monotonically increasing: if $y\le y'$ then $A(y)\le A(y')$; i.e., we have $z\le z'$ for all
$z\in A(y)$ and $z'\in A(y')$ (Proposition~\ref{monof}); \\
$\bullet$ $y\mapsto A(y)$ is (upper) semicontinuous: if $y_n\rightarrow y$, $z_n\in A(y_n)$ and $z_n\rightarrow z$ then $z\in A(y)$; \\
$\bullet$ $y\mapsto A(y+1)-y$ is $1$-periodic. This last property follows from the $1$-periodicity of $W$, since
\begin{align*}
A(y+1)&={\rm argmin}\Bigl\{Tz + W(z) + {1\over 2}\gamma (z-y-1)^2\Bigr\}
\\
&=1+{\rm argmin}\Bigl\{T(z-1) + W(z-1) + {1\over 2}\gamma (z-y)^2\Bigr\}
\\
&=1+{\rm argmin}\Bigl\{Tz + W(z) + {1\over 2}\gamma (z-y)^2\Bigr\}
\\
&=1+ A(y).
\end{align*}
Note that the recursive construction of $y^T_i$ translates in $y^T_i\in A(y^T_{i-1})$, and by assumption, we have
\begin{equation*}
\lim_{n} {y^T_{nq}-y_0\over n}= q f_\gamma(T)= p.
\end{equation*}
Hence, we will examine properties of $y^T_{nq}$, interpreted as the $n$-th iteration of the multifunction $A^q$ (the
$q$-fold composition of $A$) applied to $y_0$. Note that the multifunction $A^q$ is still increasing and
semicontinuous, and $y\mapsto A^q(y)-y$ is $1$-periodic.
We have to prove that there exists $y_0$ such that
\begin{equation*}
y_0+p\in A^q(y_0),
\end{equation*}
from which we obtain~\eqref{perob}. Note that we can assume that such $p$ and $q$ are the same as those defining
$f_\gamma(T)$ since this will automatically follow from~\eqref{perob}.
By the monotonicity of $A^q$ we deduce that $A^q(y)$ is a singleton except for a countable number of $y$. We may then
suppose that $A^q(0)$ is a singleton. We denote by $k_0$ the integer part of the unique element of $A^q(0)$, and
consider the multifunction
\begin{equation*}
G(y):=A^q(y)- k_0.
\end{equation*}
Note that $G$ inherits the properties of $A^q$ and that the unique element of $G(0)$ belongs to $(0,1)$.
We have to show that there exists $y$ such that $(G(y)-y)\cap{\Bbb Z}\neq \emptyset$. We reason by contradiction. Note
that the graph of $G$ can be extended to a maximal monotone graph $\cal G$ on $\Bbb R$, and that if we denote by
${\cal G}(y)$ the corresponding set such that $(y,\bar{y})\in \cal G$ if and only if $\bar{y}\in {\cal G}(y)$, then
${\cal G}(y)$ is a segment (degenerate for almost all $y$) whose endpoints belong to $G(y)$ by semicontinuity. This
implies that the graph of ${\cal G}(y)-y$ cannot intersect the horizontal lines $\bar{y}\in{\Bbb Z}$. Indeed, suppose that
otherwise there exist $y$ such $0\in {\cal G}(y)-y$ and let $\tilde{y}$ be the minimum of such points in $(0,1)$ (which
exists since the graph of ${\cal G}(y)-y$ is a continuous curve and $0\not\in{\cal G}(0)$. Then either
${\cal G}(\tilde{y})$ is a singleton, or the segment ${\cal G}(\tilde{y})-\tilde{y}$ has $0$ as the lower endpoint. In
either case, we have $0\in G(\tilde{y})-\tilde{y}$, which contradicts our hypothesis. Similarly, we may show that there
is no $y$ such that $1\in G(y)-y$. Hence, we have ${\cal G}(y)-y\subset (0,1)$ for all $y$.
By the continuity and periodicity of the graph of ${\cal G}(y)-y$, there exist $\delta>0$ such that
\begin{equation*}
\delta\le G(y)-y\le 1-\delta\quad \hbox{for all }y.
\end{equation*}
Let $y_0=0$ and $y_i\in G(y_{i-1})$. For all $n$, from
\begin{equation*}
\delta\le y_{i+1}-y_i\le 1-\delta \hbox{ for all }i\in\{0,\ldots, n-1\}
\end{equation*}
we deduce that
\begin{equation*}
n\delta\le G(y_n)=A^{nq}(0)-n k_0\le n(1-\delta);
\end{equation*}
that is,
\begin{equation*}
k_0+ \delta\le {A^{nq}(0)\over n}\le k_0+1- \delta\,.
\end{equation*}
Passing to the limit we finally get
\begin{equation*}
k_0+ \delta\le q f_\gamma(T)\le k_0+1- \delta,
\end{equation*}
which contradicts the assumption $q f_\gamma(T)\in{\Bbb Z}$.
\qed
\section{The limit equation}
\label{sec:limit-equation}
In this section, we show that the limit trajectory $x$ satisfies
\begin{equation}
\label{lique}
x'(t)=- \gamma\,f_\gamma(h^\prime(x(t)))
\end{equation}
for almost all $t>0$, with $f_\gamma$ defined in~\eqref{fgamma}. This equation fully characterises $x$ given the
initial datum $x_0$.
\begin{theorem}
Let $\gamma\in (0, + \infty)$. Let $t_0$ be such that $x^\prime(t_0)$ exists.
Then
\begin{equation*}
\gamma\,f_\gamma(h^\prime(x(t_0)^-))\le -x^\prime(t_0)\le \gamma\,f_\gamma(h^\prime(x(t_0)^+))\,.
\end{equation*}
\end{theorem}
\noindent{\sc Proof.}\
By translating the time variable if necessary we can suppose $t_0=0$ and $x_0= x(0)$. Let $\delta>0$ be such
that $h^\prime(x_0\pm\delta)$ exists. By Proposition~\ref{monotonicity}, \ref{it:mon2}--\ref{it:mon3},
\begin{equation*}
{x^{\varepsilon,T^{\delta^+}}_{i} - x_0\over i\tau_\varepsilon }\le {{x}^\varepsilon_i - x_0\over i\tau_\varepsilon }\le {x^{\varepsilon,T^{\delta^-}}_{i} - x_0\over i\tau_\varepsilon}\,.
\end{equation*}
The averaged velocity, as in~\eqref{eq:vel}, is given by
\begin{equation*}
{x_{i}^{\varepsilon,T} - x_0\over i \tau_\varepsilon }= \gamma {y_i^T- y_0\over i}\,;
\end{equation*}
with definition~\eqref{fgamma} it follows that
\begin{equation*}
-{\gamma}\, f_\gamma(h^\prime(x_0+\delta))= \lim_{i\rightarrow\infty} {x^{\varepsilon,T^{\delta^+}}_{i} - x_0\over i\tau_\varepsilon }\,,\qquad
-{\gamma}\, f_\gamma(h^\prime(x_0-\delta))= \lim_{i\rightarrow\infty} {x^{\varepsilon,T^{\delta^-}}_{i} - x_0\over i\tau_\varepsilon}\,.
\end{equation*}
Therefore, we conclude that
\begin{equation*}
{\gamma}\, f_\gamma(h^\prime(x_0^-))\le-x^\prime(0) \le {\gamma}\, f_\gamma(h^\prime(x_0^+))
\end{equation*}
as desired.
\qed
The previous result proves that equation~\eqref{lique} fully characterises $x$ when $t\mapsto x(t)$ is strictly
monotone, so that the set of $t$ such that
\begin{equation*}
\gamma\,f_\gamma(h^\prime(x(t)^-))\neq \gamma\,f_\gamma(h^\prime(x(t)^+))
\end{equation*}
is of zero (Lebesgue) measure.
By the monotonicity of $f_\gamma$, if $x$ is not strictly monotone then it is constant, so again~\eqref{lique} is
satisfied.
We now characterise the \emph{pinning set}, that is, the set of initial data for which $x(t)=x_0$ for all
$t>0$.
\begin{definition}[Pinning threshold]
For fixed $\gamma>0$, we define the \emph{pinning threshold} at scale $\gamma$ as
$T_\gamma:=\sup\{T: f_\gamma(T)=0\}$.
\end{definition}
\begin{remark}
Note that $f_\gamma$ is monotonically increasing, and thus $f_\gamma=0$ on $[0,T_\gamma]$. Hence, for all $x_0$ with
$|x_0|\le T_\gamma$ the motion is pinned; {\it i.e., } $x(t)=x_0$ for all $t$.
\end{remark}
The following proposition gives a criterion for the computation of the pinning threshold if $Ty+W(y)$ has (at most) a
unique local minimiser in the period. Then it suffices to examine the case where the iteration from that point is
trivial. Note that if $T\ge1$ the function $y\mapsto Ty+W(y)$ is strictly increasing, so that $T>T_\gamma$.
\begin{proposition}[Characterization of the pinning threshold]
\label{carac}
Assume that $W'$ has a unique local maximum in $(0,1/2)$. Let $0<T< 1$ and denote by $y_T\in(-1/2,0)$ the unique
local minimiser of $y\mapsto Ty+W(y)$ in $[-1/2,1/2]$. Then for every fixed $\gamma>0$, we have $T<T_\gamma$ if and
only if the function
\begin{equation*}
\varphi_T(y):= Ty+W(y)+{\gamma\over 2}(y-y_T)^2
\end{equation*}
has a unique global minimum in $y_T$.
\end{proposition}
\noindent{\sc Proof.}\
Suppose that $y_T$ is the unique global minimiser of $\varphi_T$. Then we can choose as initial datum $y_0=y_T$
in the computation of the velocity in Theorem~\ref{tevelo}, and obtain the trivial orbit $y_k=y_T$. Hence the velocity
is $0$ and consequently $T\le T_\gamma$. Actually, noting that local minimisers of $\varphi_T$ are a finite set
defined by the identity $T+W'(y)+\gamma(y-y_T)=0$, we have $T<T_\gamma$ by the continuous dependence of these quantities
in $T$.
Conversely, suppose that $y_T$ not be the unique global minimiser of $\varphi_T$. By definition of $T_\gamma$ in order
to show that $T_\gamma\le T$ it suffices to prove that the motion is not pinned for all $T+\delta$ for $\delta>0$.
Then, up to taking such $T+\delta$ in the place of $T$, we may directly suppose that $y_T$ is not a global minimiser of
$\varphi_T$, which is instead a value $y_1\in [y_T-n_1,y_M-n_1]$, where $y_M$ is the unique (local) maximum point of
$W'$ in $(0,1/2)$ and $n_1$ is some positive integer. Now, define the set
\begin{align*}
I&=\Bigl\{y\in \Bigl(-{1\over 2},{1\over 2}\Bigl): \hbox{ there exists a unique minimiser } \overline y<-{1\over 2}\ \hbox {of } \\
&\qquad\qquad\qquad\qquad\qquad\qquad\quad w\mapsto Tw+W(w)+{\gamma\over 2}(w-y)^2\Bigr\},
\end{align*}
which is the set of initial data for which the first iteration moves to ``another well''. By continuity, there exists
$\delta>0$ such that $[y_T-\delta,y_T+\delta]\subset I$. Then
$y_{N}\in I-k_1$ after a finite number of iterations $N$ independent of $y_1$ (for a finer estimate of $N$ in the
piecewise-quadratic case we refer to Section~\ref{sec:An-exampl-piec}), and we can proceed by induction. This gives
the positiveness of the velocity and $T\ge T_\gamma$.
\qed
\begin{remark}[Asymptotic behaviour at the pinning threshold]
\label{asybe}
If the hypotheses of Proposition~\ref{carac} are satisfied and $W$ is $C^2$ at local minimisers of $Ty+W(y)$ with
strictly positive second derivative then for all $\gamma>0$
\begin{equation*}
f_\gamma(T)\sim {1\over \log(T-T_\gamma)}
\end{equation*}
as $T\rightarrow T_\gamma^+$. This will be shown for piecewise quadratic energies $W$ in detail in the next section.
\end{remark}
\begin{proposition}[Extreme minimizing movements]
\label{prop:exmmm}
We have
\begin{equation}
\lim_{\gamma\rightarrow 0} \gamma f_\gamma(z) = z,\qquad
\lim_{\gamma\rightarrow +\infty} \gamma f_\gamma(z) = g_\infty(z),
\end{equation}
where $g_\infty$ is given by
\begin{equation}
\label{ginfty}
g_\infty(z)=
\begin{cases}\displaystyle
\Biggl(\int_0^1{1\over z+W(s)}ds\Biggl)^{-1} &\text{ if }\displaystyle{1\over z+W(s)} \text{ is integrable,}\cr\cr
0 & \text{otherwise}.
\end{cases}
\end{equation}
Moreover,
\begin{equation}
\lim_{\gamma\rightarrow+\infty} T_\gamma= \sup_\gamma T_\gamma=T_\infty,
\end{equation}
where $[-T_\infty,T_\infty]=\{T\in\Bbb R: g_\infty(T)=0\}$.
\end{proposition}
\noindent{\sc Proof.}\
Assume that $h$ and $W$ are $C^2$-functions. The convergence as $\gamma\rightarrow 0$ follows from the observation that the orbit $x^\varepsilon_k$ satisfies
\begin{equation*}
{x^\varepsilon_k-x^\varepsilon_{k-1}\over\tau}= -h^\prime(x^\varepsilon_k) +O\Bigl({\varepsilon\over\tau}\Bigr)
= -x^\varepsilon_k +o(1)
\end{equation*}
as $\gamma\rightarrow 0$. Conversely, the convergence as $\gamma\rightarrow +\infty$ follows by noting that as $\gamma\rightarrow +\infty$ the
orbits $x^\varepsilon(t)$, defined as in (\ref{seqMM}), are close to the corresponding solution of the gradient flow
\begin{equation*}
x_\varepsilon'=-h'(x_\varepsilon)-W^\prime\Bigl({x_\varepsilon\over\varepsilon}\Bigl),
\end{equation*}
whose limit satisfies $x'=-g_\infty(h'(x))$.
\qed
\begin{remark}
\label{extra}
By Theorem 8.1 in~\cite{Braides2014a}, the equations
\begin{equation*}
x'=-h'(x) \text{ and } x'=-g_\infty(h'(x))
\end{equation*}
describe the minimising movements in the cases $\varepsilon\ll\tau$ and $\tau\ll\varepsilon$, respectively. The previous proposition
shows that the same extreme minimising movements are obtained by keeping the ratio $\gamma=\varepsilon/\tau$ fixed and then
let it tend to $0$ and $+\infty$, respectively.
\end{remark}
\section{An example: the piecewise-quadratic case}
\label{sec:An-exampl-piec}
In this section, we provide an example of oscillating potential and calculate explicitly the corresponding pinning
threshold $T_\gamma$. More precisely, we consider the piecewise quadratic energy
\begin{equation*}
W(y):= \min_{k\in {\Bbb Z}} (y-k)^2\,.
\end{equation*}
Besides giving an illustrative example, we deduce the asymptotic behaviour at the pinning threshold, which depends only
on the non-degeneracy of the second derivative at local minima. In this way we deduce the asymptotic behaviour in
the general case as in Remark~\ref{asybe}.
For this choice of $W$, for $T_\infty$ as in Proposition~\ref{prop:exmmm}, it holds that
\begin{equation*}
T_\infty= 1,
\end{equation*}
and~\eqref{ginfty} becomes
\begin{equation*}
g_\infty(z)={1\over \log\bigl({z-1\over z}\bigr)} \qquad\hbox{ for } z> 1.
\end{equation*}
The function $W$ is $1$-periodic and piecewise quadratic with $\Vert W^\prime\Vert_\infty=1$. For simplicity, we fix
also $h(x)= x^2/2$. Let $T\in (0,1)$ and $y_0\in [0, 1/2)$. The minimum of the function
$F^T_{\gamma}(y, y_0)= T y + W (y) + {\gamma\over 2} (y-y_0)^2$, given in~\eqref{FTgamma}, on the interval
$[-1/2 +k, 1/2 +k]$ can be attained at the boundary or at the interior of this interval; it is given by
\begin{equation}
\label{formulayk}
y_{1,k}= {-T+2k\over 2+\gamma} + {\gamma\over 2+\gamma} y_0
\end{equation}
($y_{1,k}$ also depends on $T$ but we suppress this in the notation).
The global minimiser $y^T_1$ to $F^T_{\gamma}(y, y_0)$ can get stuck in the same well of $y_0$, that is,
$y^T_1= y_{1,0}$. Otherwise it can move into the next well, corresponding to $k=-1$, that is, $y^T_1= y_{1,-1}$.
Any single well of $W(y)$ is denoted by $W(y;k):= (y-k)^2$ for every $y\in [-1/2+k, 1/2+k]$ with $k\in {\Bbb Z}$. We define
\begin{align}
\psi (y) &:= \Bigl(Ty_{1,-1}+ W(y_{1,-1};-1) + {\gamma\over 2} (y_{1,-1}- y)^2 \Bigr) \notag\\
& {}\qquad- \Bigl( Ty_{1,0}+ W(y_{1,0};0) + {\gamma\over 2} (y_{1,0}- y)^2
\Bigr) \,. \label{defpsi}
\end{align}
To establish if $y_1^T$ gets stuck or moves, we have to study the sign of $\psi (y_0)$, since it is the difference
between the minimum value of the two wells. Therefore, if $\psi (y_0) < 0$, then the minimiser satisfies
$y^T_1= y_{1,-1}$; $\psi (y_0) \ge 0$ implies that $y^T_1= y_{1,0}$. In particular, from the sign of $\psi (y_0)$ we
expect to derive the pinning threshold $T_\gamma$.
\begin{proposition}
\label{quaqua}
Let $\gamma\in (0, +\infty)$ and let $T\in(0,1)$. Then there exist
\begin{equation*}
T_\gamma := {\gamma\over (2+\gamma)} \quad \hbox{and} \quad \delta_T:= \Bigl({2+\gamma\over 2\gamma}\Bigr)\, (T- T_\gamma),
\end{equation*}
such that the following holds.
\begin{enumerate}
\item \label{it:quaqua1}
For every $T>T_\gamma$ we have that, the following possibilities exist.
\begin{enumerate}
\item \label{it:quaqua1a} If $y_0\in [0, (-T/2)+\delta_T)$ then $y_1^T= y_{1,-1}$, where the latter is defined
in~\eqref{formulayk}. Moreover, if
\begin{equation*}
y_1^T +{T\over 2}+1= {\gamma\over 1+\gamma} \Bigl(y_0 + {T\over 2}+1\Big) < -{T\over 2} +\delta_T\,,
\end{equation*}
then the successive minimiser is given by $y^T_2= y_{2,-2}$ and so on.
\item \label{it:quaqua1b} If $y_0\in [(-T/2)+\delta_T, 1/2)$
then there exists $h\in {\Bbb N}$ given by
\begin{equation}\label{def-h}
h= \Bigg\lfloor{\log \Bigl( ({2+\gamma\over \gamma}) {T- T_\gamma \over T+ 1}\Bigr)
\over \log \Bigl({\gamma\over 2+\gamma}\Bigr)}\Bigg\rfloor + 1
\end{equation}
such that $y_1^T= y_{1,0}, \ldots,y_{h}^T= y_{h-1,0} \ge (-T/2) +\delta_T$, $y_{h,0} < (-T/2) +\delta_T$ and
$y_{h+1}^T= y_{h+1,-1}$. Similarly, if $k\in{\Bbb N}$ exists such that for some $p\in{\Bbb N}$
\begin{equation*}
y_k^T +{T\over 2}+p \ge -{T\over 2} +\delta_T \,,
\end{equation*}
then there exists $h$ as in~\eqref{def-h} such that
if we take as initial data $z_0:= y^T_k +(T/2)+p$ then we get a new sequence of minimisers such that
$z^T_1=z_{1,0}\,,\ldots ,z^T_{h-1}= z_{h-1,0}\ge -T +\delta_T$, $z_{h}^T= z_{h,0} < (-T/2) +\delta_T$ and
$z_{h+1}^T= z_{h+1,-1}$.
\end{enumerate}
\item \label{it:quaqua2} For every
\begin{equation*}
T\le T_\gamma\ \hbox{and}\ y_0\ge 0
\end{equation*} the motion is pinned.
\end{enumerate}
\end{proposition}
\noindent{\sc Proof.}\ We first derive $T_\gamma$ using the criterion given in Proposition~\ref{carac}. More precisely, since the
unique local minimiser of $y\mapsto Ty+W(y)$ in $[-1/2,1/2]$ is $y_T=-T/2$, by~\eqref{formulayk} and~\eqref{defpsi},
with $y_0=y_T$ we have that
\begin{equation*}
y_{1,-1}= -{T\over 2} -{2\over 2+\gamma} \,,\qquad
y_{1,0}= -{T\over 2} \,.
\end{equation*}
and
\begin{equation*}
\psi (y_T)= {2\over 2+\gamma} \Bigl(-T + {\gamma\over 2} + \gamma y_T\Bigr)=
-T +{\gamma\over 2 +\gamma}\,.
\end{equation*}
Therefore, we find
\begin{equation*}
T_\gamma= {\gamma\over 2 +\gamma}\,.
\end{equation*}
Moreover, for fixed $y_0= (-T/2) +\delta$, one has
\begin{equation*}
\psi\Bigl(-{T\over 2} +\delta\Bigr)\le 0 \quad \hbox{if and only if} \quad
\delta\le {2+\gamma\over 2\gamma} \Bigl(T-{\gamma\over 2+\gamma}\Bigr)\,.
\end{equation*}
Therefore, if we define
\begin{equation}
\label{deltaT}
\delta_T:= {2+\gamma\over 2\gamma} \Bigl( T- T_\gamma\Bigr)
\end{equation}
we are ready to prove the statements~\ref{it:quaqua1a} and~\ref{it:quaqua1b} of the proposition.
We now give the proof of~\ref{it:quaqua1a}. For every $T > T_\gamma$ there exists $\delta_T>0$, given
by~\eqref{deltaT}, such that for every $y_0\in [0, (-T/2)+\delta_T)$ we have that $\psi (y_0) <0$, that is,
$ y^T_1= y_{1,-1}$.
Reasoning as above, we observe that if $y_{1,-1} +(T/2)+1< (-T/2) +\delta_T$, then $y^T_2= y_{2,-2}$ and we can iterate
until this condition is satisfied. The case $y_{1,-1} +(T/2)+1\ge (-T/2) +\delta_T$ is addressed in
point~\ref{it:quaqua1b}.
Next, we give the proof of claim~\ref{it:quaqua1b}. If $y_0\in [(-T/2)+\delta_T, 1/2)$ then $\psi (y_0)\ge0$ and the
minimisers can be calculated recursively, by~\eqref{formulayk}, in the following way.
\begin{align*}
y^T_1&=y_{1,0} = {\gamma\over 2+\gamma} \, y_0 - {T\over 2+\gamma}\,, \\
y^T_2 &=y_{2,0} = {\gamma\over 2+\gamma} \, y^T_1 - {T\over 2+\gamma} \\
& \qquad = \Bigl({\gamma\over 2+\gamma}\Bigr)^2 \, y_0 - {T\over 2+\gamma} \Bigl( 1 + {\gamma\over 2+\gamma}\Bigr)\,,\\
y^T_h &= y_{h,0}
= \Bigl({\gamma\over 2+\gamma}\Bigr)^h y_0 - {T\over 2+\gamma} \sum_{n=0}^{h-1} \Bigl( {\gamma\over 2+\gamma}\Bigr)^n\,.
\end{align*}
Therefore, we may rewrite
\begin{equation}
\label{recursive-yh}
y^T_{h} = \Bigl({\gamma\over 2+\gamma}\Bigr)^h \Bigl(y_0 + {T\over 2}\Bigr) -{T\over 2}\,.
\end{equation}
Since $(y_0 + (T/2)) < {1\over 2} +(T/2)$, then we may assume that
\begin{equation*}
\Bigl({\gamma\over 2+\gamma}\Bigr)^h \Bigl(y_0 + {T\over 2}\Bigr) -{T\over 2} <
\Bigl({\gamma\over 2+\gamma}\Bigr)^h \Bigl({T+1\over 2} \Bigr) -{T\over 2} < -{T\over 2}+\delta_T\,.
\end{equation*}
Therefore
\begin{equation*}
\Bigl({\gamma\over 2+\gamma}\Bigr)^h < {2\delta_T\over T + 1};
\end{equation*}
that is,
\begin{align*}
h &> {\log \Bigl( {2\delta_T\over T+ 1}\Bigr) \over \log \Bigl({\gamma\over 2+\gamma}\Bigr)} =
{\log \Bigl( ({2+\gamma\over \gamma}) {T- T_\gamma \over T+ 1}\Bigr) \over \log \Bigl({\gamma\over 2+\gamma}\Bigr)}\,.
\end{align*}
Similarly, if $y^T_1 +(T/2)+1 \ge -(T/2) +\delta_T$ then we may reason as above, by taking as initial datum
$z_0:= y^T_1 +(T/2)+1$. Therefore, we get a new sequence of minimisers given by
\begin{align*}
z^T_1&=z_{1,0} = {\gamma\over 2+\gamma} \, z_0 - {T\over 2+\gamma} \,, \\
z^T_2 &=z_{2,0} = {\gamma\over 2+\gamma} \, z_1 - {T\over 2+\gamma} \\
& \qquad = \Bigl({\gamma\over 2+\gamma}\Bigr)^2 \, z_0 - {T\over 1+\gamma} \Bigl( 1 + {\gamma\over 2+\gamma}\Bigr) \,,\\
& \cdots \\
z^T_h &= z_{h,0}
= \Bigl({\gamma\over 2+\gamma}\Bigr)^h \Bigl(z_0 + {T\over 2}\Bigr) -{T\over 2}\,,
\end{align*}
such that $z^T_1=z_{1,0}\,,\ldots ,z^T_{h-1}= z_{h-1,0}\ge (-T/2) +\delta_T$, $z_{h}^T= z_{h,0} < (-T/2) +\delta_T$
and $z_{h+1}^T= z_{h+1,-1}$. More generally, if $k\in{\Bbb N}$ exists such that, for some $p\in{\Bbb N}$,
$y^T_k +(T/2)+p \ge (-T/2) +\delta_T$, then we may repeat the procedure above by assuming $z_0:= y^T_k +(T/2)+p$.
We now turn to case~\ref{it:quaqua2} and give the proof. If $T\le T_\gamma$, since $y_0\ge 0$, then $\psi(y_0)>0$, that
is, $y_1^T= y_{1,0}$. Moreover, by~\eqref{recursive-yh}, we have that
\begin{equation*}
y_h^T= \Bigl({\gamma\over 2+\gamma}\Bigr)^h \Bigl(y_0 + {T\over 2}\Bigr) -{T\over 2} > -{T\over 2} +\delta_T
\end{equation*}
for every $h\in{\Bbb N}$. Therefore $\lim_{h\rightarrow \infty} y_h^T= (-T/2)$, that is, for every $T\le T_\gamma$ the motion is
pinned. \qed
\begin{remark}[Behaviour at the pinning threshold]
From Proposition~\ref{quaqua}, case~\ref{it:quaqua1b} we deduce that
\begin{equation*}
f_\gamma (T)\sim {\log \bigl({\gamma\over 2+\gamma}\bigr)\over\log \Bigl(({2+\gamma\over \gamma}) {T- T_\gamma \over T+ 1}\Bigr) };
\end{equation*}
that is, for $\gamma>0$ fixed,
\begin{equation*}
f_\gamma (T)\sim {1\over|\log (T- T_\gamma) | }
\end{equation*}
as $T\rightarrow T_\gamma^+$. Note in particular that $f_\gamma$ is not Lipschitz for $T\rightarrow T_\gamma^+$.
\end{remark}
\paragraph{Acknowledgements} NA and JZ gratefully acknowledge funding by the Marie Curie Actions: Intra-European
Fellowship for Career Development (IEF2012, FP7-People) under REA grant agreement n${}^\circ$ 326044. JZ was partially
supported by the UK's Engineering and Physical Sciences Research Council Grant EP/K027743/1 and the Leverhulme Trust
(RPG-2013-261).
\def$'${$'$} \def$'${$'$} \def$'${$'$}
\def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}} \def$'${$'$}
\def$'${$'$}
|
\section{Introduction and Preliminaries}
Let $A\in \mathbb{C}^{n\times n}$. If there exists $X\in \mathbb{C}^{n \times n}$ satisfying the equations $AXA=A, ~XAX=X$ and $AX=XA$, then such an $X$ can be shown to be unique. This unique solution is called the {\it group inverse} of $A$ and is denoted by $A^{\#}$. The nomenclature group inverse arises from the fact that the positive powers of $A$ and $A^{\#}$ together with the idempotent matrix $AA^{\#}$ as the identity element, form an abelian group under matrix multiplication and was thus named by I.Erdelyi in $1967$. The group inverse of a matrix $A\in \mathbb{C}^{n \times n}$ need not always exist. It is well known that $A^{\#}$ exists if and only if $R(A) \cap N(A)=\{0\}$. In other words, $A^{\#}$ exists if and only if the index of $A$ equals $1$. Let us recall that the {\it index} of a square matrix $A$ is the smallest positive integer $k$ such that $rank(A^k)=rank(A^{k+1})$. If $A$ is invertible, then the index is defined to be zero. An easy dimensionality argument can be used to show that the
index exists for any square matrix. Thus, $A^{\#}$ exists if and only if $rank(A)=rank(A^2)$. For more details and other properties of the group inverse, we refer to the book \cite{bens}.
Let $A,B,C $ and $D \in \mathbb{C}^{n \times n}$ and $M=
\left(\begin{array}{cc}
A & B \\
C & D
\end{array} \right)$. Suppose that $A^{-1}$ exists. Then the matrix $D-CA^{-1}B$ is called the {\it Schur complement} of $A$ in $M$. Next, let $D^{-1}$ exist. Then the matrix $A-BD^{-1}C$ is called the {\it associated Schur complement} of $D$ in $M$. The usefulness of the Schur complement is well documented in numerous texts on matrices and numerical analysis (see for instance, \cite{zh}). Let us turn our attention to the principal pivot transform. Tucker introduced this notion in his studies on linear programming problems. Let $M$ be the block matrix, defined as above. Suppose that $A^{-1}$ exists. Then the principal pivot transform of $M$ relative to $A$ is defined by the matrix
\begin{center}
$\left(\begin{array}{cc}
A^{_1} & -A^{-1}B \\
CA^{-1} & D-CA^{-1} B
\end{array}\right)$.
\end{center}
Note that the bottom right block of the principal pivot transform is the Schur complement of $A$ in $M$, establishging a relationship between these notions. We refer to the recent work \cite{tsat} for some interesting properties of the principal pivot transform.
Extensions of the formulae for the Schur complement \cite{carlhaymark} and the principal pivot transform
\cite{meen}, where the usual inverse is replaced by the Moore-Penrose inverse were studied rather long ago.
We refer to these as the {\it pseudo Schur complement} and {\it pseudo principal pivot transform}, respectively. In \cite{kavi}, some of the results of \cite{tsat} and a few inheritance properties were proved, for the pseudo Schur complement and the pseudo principal pivot transform. In the present work, we consider the case of the group inverse, especially for the pseudo principal pivot transform. First, we prove a formula for the group inverse of a block matrix in the presence of some assumptions and then derive the basic properties of the principal pivot transform.
\section{The Pseudo Principal Pivot Transform in terms of the Group Inverse.}
We begin by proving a formula for the group inverse of partitioned matrices.
\begin{thm}\label{GA}
Let $A,B,C $ and $D \in \mathbb{C}^{n \times n}$. Suppose that $A^{\#}$ exists. Set $K=D-CA^\# B$ and $M=
\left(\begin{array}{cc}
A & B \\
C & D
\end{array} \right)$. Suppose that $K^{\#}$ exists. Then $ R(C^*)\subseteq R(A^*)$, $R(B)\subseteq R(A)$, $R(C)\subseteq R(K)$ and
$R(B^*)\subseteq R(K^*)$ if and only if
\begin{center}
$M^\#=\left(\begin{array}{cc}
A^\# +A^\# BK^\# CA^\# & -A^\# BK^\# \\
-K^\# CA^\# & K^\#
\end{array} \right)$ .
\end{center}
\end{thm}
\begin{proof}
First, we observe that $CA^\# A=C$ (since $R(C^*)\subseteq R(A^*)$) and $BK^\# K=B$ (since $R(B^*)\subseteq R(K^*)$). We then have
\begin{eqnarray*}
DK^\# K&=&(K+CA^\# B)K^\# K\\
&=& K+CA^\# BK^\# K\\
&=& K+CA^\# B\\
&=& D.
\end{eqnarray*}
Also,
\begin{eqnarray*}
A^{\#} B + A^{\#} BK^{\#} CA^{\#}B-A^{\#}BK^\# D&=&A^\# B + A^\# BK^\# (CA^\# B- D)\\
&=&A^\# B-A^\#BK^\# K\\
&=&0.
\end{eqnarray*}
\newpage
Set
\begin{center}
$X=\left(\begin{array}{cc}
A^\#+A^\# BK^\# CA^\# & -A^\# BK^\# \\
-K^\# CA^\# & K^\#
\end{array}\right)$.
\end{center}
From the earlier calculations one has
\begin{eqnarray*}
XM & =& \left(\begin{array}{cc}
A^\# + A^\# BK^\# CA^\# & -A^\# BK^\# \\
-K^\# CA^\# & K^\#
\end{array}\right)
\left(\begin{array}{cc}
A & B \\
C & D
\end{array}\right)\\
& =& \left(\begin{array}{cc}
A^\# A & 0 \\
0 & K^\# K
\end{array}\right)
\end{eqnarray*}
and
\begin{eqnarray*}
MX & = & \left(\begin{array}{cc}
A & B \\
C & D
\end{array}\right)
\left(\begin{array}{cc}
A^\#+A^\# BK^\# CA^\# & -A^\# BK^\# \\
-K^\# C A^\# & K^\#
\end{array}\right) \\
& = & \left(\begin{array}{cc}
AA^\# & 0 \\
0 & KK^\#
\end{array}\right),
\end{eqnarray*}
where we have used the facts that $AA^\# B=B$ (since $R(B)\subseteq R(A)$) and
$ C=KK^\# C$ (since $R(C)\subseteq R(K)$). So, $XM=MX$. Also,
\begin{eqnarray*}
MXM&=&
\left(\begin{array}{cc}
A & BK^\# K \\
C & DK^\# K
\end{array}\right)\\
&=&\left(\begin{array}{cc}
A & B\\
C & D
\end{array}\right)\\
&=&M.
\end{eqnarray*}
Further, \\
\begin{eqnarray*}
XMX
&=&
\left(\begin{array}{cc}
A^\# A & 0 \\
0 & K^\# K
\end{array}\right)
\left(\begin{array}{cc}
A^\#+A^\# BK^\# CA^\# & -A^\# BK^\# \\
-K^\# C A^\# & K^\#
\end{array}\right) \\
&=&
\left(\begin{array}{cc}
A^\#+A^\# BK^\# CA^\# & -A^\# BK^\# \\
-K^\# C A^\# & K^\#
\end{array}\right)\\
&=& X.
\end{eqnarray*}
This completes the proof of the necessity part.
\newpage
Conversely, suppose that
\begin{center}
$M^{\#}=\begin{pmatrix}
A^\# +A^\# BK^\# CA^\# & -A^\# BK^\# \\
-K^\# CA^\# & K^\#
\end{pmatrix}$.
\end{center}
Then,
\begin{center}
$MM^{\#}=\begin{pmatrix}
AA^{\#}+AA^{\#}BK^{\#}CA^{\#}-BK^{\#}CA^{\#} & -AA^{\#}BK^{\#}+BK^{\#}\\
CA^{\#}+CA^{\#}BK^{\#}CA-DK^{\#}CA^{\#} & -CA^{\#}BK^{\#}+DK^{\#}
\end{pmatrix}$.
\end{center}
Equating the top left blocks of $MM^{\#}M$ and $M$, we then have
\begin{center}
$AA^{\#}A+AA^{\#}BK^{\#}CA^{\#}A-BK^{\#}CA^{\#}A-AA^{\#}BK^{\#}C+BK^{\#}C=A$.
\end{center}
This simplifies to
\begin{center}
$A+(I-AA^{\#})BK^{\#}C(I-A^{\#}A)=A$,
\end{center}
which in turn yields
\begin{eqnarray}\label{eq1}
(I-AA^{\#})BK^{\#}C(I-A^{\#}A) &=& 0.
\end{eqnarray}
By equating the top right blocks of $MM^{\#}M$ and $M$, we get
\begin{center}
$AA^{\#}B+AA^{\#}BK^{\#}CA^{\#}B-BK^{\#}CA^{\#}B-AA^{\#}BK^{\#}D+BK^{\#}D=B$,
\end{center}
so that one has
\begin{center}
$AA^{\#}B+(I-AA^{\#})BK^{\#}K=B$.
\end{center}
This reduces to
\begin{eqnarray}\label{eq2}
(I-AA^{\#})B(I-K^{\#}K) &=& 0.
\end{eqnarray}
Again, equating the bottom left blocks of $MM^{\#}M$ and $M$,
\begin{center}
$CA^{\#}A+CA^{\#}BK^{\#}CA^{\#}A-DK^{\#}CA^{\#}A-CA^{\#}BK^{\#}C+DK^{\#}C=C,$
\end{center}
we then have
\begin{eqnarray}\label{eq3}
(I-KK^{\#})C(I-A^{\#}A) &=& 0.
\end{eqnarray}
We also have $MM^{\#}=M^{\#}M$ and so, on equating the bottom left blocks of $MM^{\#}$ and $M^{\#}M$, we have
\begin{eqnarray}\label{eq4}
K^{\#}C(I-A^{\#}A) &=&(I-KK^{\#})CA^{\#}.
\end{eqnarray}
\newpage
Premultiplying $(4)$ by $K$, we get
\begin{center}
$KK^{\#}C(I-A^{\#}A)=0$
\end{center}
and thus from equation \eqref{eq3}, one has $CA^{\#}A=C$. So, $R(C^*)\subseteq R(A^*)$. Postmultiplying $(4)$ by $A$ and using the fact that $CA^{\#}A=C$, we get
\begin{center}
$(I-KK^{\#})C=0$.
\end{center}
Thus $R(C)\subseteq R(K)$.
Now, equating top right blocks of $MM^{\#}$ and $M^{\#}M$, one has
\begin{eqnarray}\label{eq5}
(I-AA^{\#})BK^{\#}&=&A^{\#}B(I-KK^{\#}).
\end{eqnarray}
Postmultiplying equation \eqref{eq5} by $K$, we get
\begin{center}
$(I-AA^{\#})BK^{\#}K=0$
\end{center}
and from equation \eqref{eq2}, we get $AA^{\#}B=B$. Thus $R(B)\subseteq R(A)$. Finally, premultiplying \eqref{eq5} by $A$ and using the fact that $AA^{\#}B=B$,
\begin{center}
$B(I-K^{\#}K)=0$.
\end{center}
Thus $R(B^*)\subseteq R(K^*)$.
\end{proof}
Next, we state a complementary result, whose proof is similar to the proof of Theorem \ref{GA}. Note that this result uses the pseudo Schur complement $L=A-BD^\# C$, which will be called the {\it complementary Schur complement}.
\begin{thm}\label{GD}
Let $M=
\left(\begin{array}{cc}
A & B \\
C & D
\end{array}\right)$ as above. Suppose that $D^{\#}$ exists and $L=A-BD^\# C$. Suppose that $L^{\#}$ exists. Then $R(B^*)\subseteq R(D^*)$,
$R(C)\subseteq R(D)$, $R(B)\subseteq R(L)$ and $R(C^*)\subseteq R(L^*)$ if and only if
\begin{center}
$M^\#=
\left(\begin{array}{cc}
L^\# & -L^\# BD^\# \\
-D^\# C L^\# & D^\#+D^\# C^\# BD^\#
\end{array}\right)$.
\end{center}
\end{thm}
Next, we define the principal pivot transform in terms of the group inverse. As mentioned earlier, principal pivot transform involving Moore-penrose inverse of the block matrices was studied in \cite{meen}.
\newpage
\begin{defn}\label{D: D1} Let $M$ be defined as above and $A^{\#}$ exist. Then the pseudo principal pivot transform of $M$ relative to $A$ is defined by
\begin{center}
$P:=pppt(M,A)_{\#}=
\left(\begin{array}{cc}
A^\# & -A^\# B \\
CA^\# & K
\end{array}\right)$,
\end{center} where $K=D-CA^\# B$. Next, suppose that $D^{\#}$ exists. The complementary pseudo principal pivot transform of $M$ relative to $D$ is defined by
\begin{center}
$Q:=cpppt(M,D)_{\#}=
\left(\begin{array}{cc}
L & BD^\# \\
-D^\# C & D^\#
\end{array}\right)$,
\end{center}
where $L=A-BD^\# C$.
\end{defn}
Both the operations of pseudo principal transforms are involutions, in the presence of certain assumptions, as we prove next.
\begin{lem}\label{L1}
Let $M$ be defined as above. Let $A^{\#}$ and $D^{\#}$ exist.\\
$(i)$ Suppose that $R(B)\subseteq R(A)$ and $R(C^*)\subseteq R(A^*)$. Then $pppt\left(P, A^\# \right)_{\#}=M$.\\
$(ii)$ Suppose that $R(C)\subseteq R(D)$ and $R(B^*)\subseteq R(D^*)$. Then $cpppt\left(Q, D^\# \right)_{\#}=M$.
\end{lem}
\begin{proof}
$(i)$: Set $W=A^\#, X=-A^\# B$ and $Y=CA^\#$. Then $P=
\left(\begin{array}{cc}
W & X\\
Y & K
\end{array}\right)$. So,
\begin{eqnarray*}
pppt(P, A^\#)_{\#}=
pppt (P, W)_{\#}&=&
\left(\begin{array}{cc}
W^\# & -W^\# X\\
YW^\# & K-YW^\# X
\end{array}\right)\\
&=&
\left(\begin{array}{cc}
A & -A(-A^\# B) \\
CA^\# A & D-CA^\# B +CA^\# AA^\# B
\end{array}\right)\\
&=&\left(\begin{array}{cc}
A & B \\
C & D
\end{array}\right) \\
&=& M.
\end{eqnarray*}
$(ii)$: The proof is similar to part $(i)$.
\end{proof}
Finally, we derive the domain-range exchange property. Domain-range exchange property in terms of Moore-Penrose inverse of the matrix is proved in \cite{kavi}. This property is well known in the nonsingular case \cite{tsat}.
\begin{lem}\label{L2} Let $M$ be defined as earlier. Let $A^{\#}$ and $D^{\#}$ exist.\\
$(i)$ Suppose that $R(B)\subseteq R(A)$ and $R(C^*)\subseteq R(A^*)$.
Then $M$ and $P=pppt(M,A)_{\#}$ are related by the formula:
\begin{center}
$M
\left(\begin{array}{cc}
x^{1}\\
x^{2}
\end{array}\right)
=
\left(\begin{array}{cc}
AA^\# y^{1} \\
y^{2}
\end{array}\right)$
if and only if $P
\left(\begin{array}{cc}
y^{1} \\
x^{2}
\end{array}\right)
=
\left(\begin{array}{cc}
A^\# A x^{1} \\
y^{2}
\end{array}\right)$.
\end{center}
(ii) Suppose that $R(C)\subseteq R(D)$ and $R(B^*)\subseteq R(D^*)$. Then $M$ and $Q=cpppt(M,D)_{\#}$ are related by the formula:
\begin{center}
$M
\left(\begin{array}{cc}
x^{1}\\
x^{2}
\end{array}\right)
=
\left(\begin{array}{cc}
y^{1}\\
DD^\# y^{2}
\end{array}\right)$
if and only if
$Q
\left(\begin{array}{cc}
y^{1}\\
x^{2}
\end{array}\right)
=
\left(\begin{array}{cc}
x^{1}\\
D^\# Dy^{2}
\end{array}\right)$.
\end{center}
\end{lem}
\begin{proof}
We prove (i). The proof for (ii) is similar. Suppose that
\begin{center}
$M
\left(\begin{array}{cc}
x^{1}\\
x^{2}
\end{array}\right)
=
\left(\begin{array}{cc}
AA^\# y^{1}\\
y^{2}
\end{array}\right)$.
\end{center}
Then
\begin{center}
$Ax^{1}+Bx^{2}=AA^\# y^{1}$
\end{center}
and
\begin{center}
$Cx^{1}+Dx^{2}=y^{2}$.
\end{center}
Premultipling the first equation by $A^\#$ (and rearranging) we get
\begin{center}
$A^\# y^{1}-A^\# Bx^{2}=A^\# Ax^{1}$.
\end{center}
Premultiplying this equation by $C$, we then have
\begin{center}
$CA^\# y^{1}-CA^\# Bx^{2}=CA^\# Ax^{1}=Cx^{1}$.
\end{center}
So, $CA^\# y^{1}+Kx^{2} =CA^\# y^{1}+Dx^{2}-CA^\# Bx^{2} =Cx^{1}+Dx^{2} =y^{2}$. Thus,
\begin{center}
$P
\left(\begin{array}{cc}
y^{1}\\
x^{2}
\end{array}\right)
=
\left(\begin{array}{cc}
A^\# y^{1}-A^\# Bx^{2}\\
CA^\# y^{1}+Kx^{2}
\end{array}\right)
=
\left(\begin{array}{cc}
A^\# Ax^{1}\\
y^{2}
\end{array}\right)$.
\end{center}
Conversely, let
\begin{center}
$P
\left(\begin{array}{cc}
y^{1}\\
x^{2}
\end{array}\right)
=
\left(\begin{array}{cc}
A^\# Ax^{1}\\
y^{2}
\end{array}\right)$.
\end{center}
Then
\begin{center}
$A^\# y^{1}-A^\# Bx^{2}=A^\# Ax^{1}$
\end{center}
and
\begin{center}
$CA^\# y^{1}+(D-CA^\# B)x^{2}=y^{2}$.
\end{center}
Premultiplying the first equation by $A$, we have $AA^\# y^{1}-Bx^{2}=Ax^{1}$ so that $Ax^{1}+Bx^{2}=AA^\# y^{1}$. Again, premultiplying the first equation by $C$, we get $CA^\# y^{1}-CA^\# Bx^{2}=Cx^{1}$. Hence, using the second equation we have, $Cx^{1}+Dx^{2}=CA^\# y^{1}-CA^\# Bx^{2}+Dx^{2}=y^{2}$, proving that
\begin{center}
$M
\left(\begin{array}{cc}
x^{1}\\
x^{2}
\end{array}\right)
=
\left(\begin{array}{cc}
AA^\# y^{1}\\
y^{2}
\end{array}\right)$.
\end{center}
\end{proof}
|
\section{Introduction}
In the early stages of star formation, protoplanetary discs may be
subject to gravitational instability (GI) owing to their
large densities and low temperatures.
The parameter that best quantifies a disk's
susceptibility to GI is the Toomre $Q$,
defined via
\begin{equation}
Q=\dfrac{c_s\kappa}{\pi G \Sigma_0}<1
\label{mass_eq}
\end{equation}
\citep{toomre64},
where $c_s$ is
the sound speed, $\kappa$ the epicyclic frequency,
and $\Sigma_0$ the background surface density.
In a razor thin disk, the linear instability criterion for axisymmetric
disturbances is simply $Q<1$, though
non-axisymmetric nonlinear instability occurs for slightly
larger $Q$. When radiative cooling is inefficient, the system saturates
in a gravitoturbulent state that can transport significant angular
momentum \citep{Gammie2001,rice14}, while more efficient cooling causes
the system to fragment into dense clumps that may serve as the
precursors of gas giant planets \citep{cameron78,boss97}.
Note that the critical cooling time that separates the two outcomes is
vulnerable to the numerical details of its calculation and still the
subject of some debate \citep{Pdkooper2012,rice14}.
\\
A key but undeveloped area of research is the interaction between the
GI and magnetic fields, and in particular the magnetorotational instability (MRI),
an alternative mechanism of angular momentum transport
in sufficiently ionized gas \citep{balbus91,hawley95}.
In the massive early stage of a protostellar disk's life, GI
may lead to gravitoturbulence and fragmentation at large radii,
but these regions could also be
ionised to a dynamically relevant degree by cosmic rays or stellar X-rays
\citep[e.g.][]{armitage11}. Obvious questions are whether the GI and
MRI coexist and a quasi-steady state accommodating both is possible,
what the properties of this state might be, and if fragmentation is enhanced
(or mitigated) by the MRI.
Even if the MRI is quenched or
greatly impeded by non-ideal MHD effects, the gas may still couple to
large-scale magnetic fields \citep[possibly generated by Hall
currents][]{simon15},
which could then significantly modify the gravitoturbulence.
On the other hand, GI might act as a dynamo, creating
small-scale field from a low-amplitude seed. \\
The interaction between GI and magnetic fields may also be important
later in a disk's lifetime, during
FU Ori and EX-Lupi outbursts, when the accretion rate undergoes
violent jumps on timescales of 100-1000 years
\citep{evans09,aguilar12}.
It is theorised that this quasi-periodic behaviour conforms to a
`gravo-magneto' limit cycle, according to which (a)
mass accumulates in the dead
zone, via more efficient accretion at larger radii,
until (b) the high surface density initiates GI, which heats the
gas to the point that (c) collisional ionisation permits the onset of
MRI,
and (d) the excess mass is swept onto the protostar in a dramatic
accretion event \citep{armitage01,zhu10,martin11}.
One issue here is the strong heating required: can
GI adequately thermalise its turbulent motions so as to trigger the MRI?
Another issue is whether the MRI can emerge unproblematically from the
pre-existing gravitoturbulent state.\\
Global and local simulations of self-gravitating discs have been
extensively used over the last decade but very few have coupled GI
with MHD. \citet{kim2001}
studied the fragmentation criterion in magnetized galactic discs, but
their local simulations do not reproduce a fully saturated
gravitoturbulent state.
The coexistence between GI and MRI was
investigated by Fromang et al.~(2004a, 2004b) and Fromang (2005)
who showed that the turbulence
induced by MRI modes tends to reduce the strength of the gravitational
instability and prevent local clumps of gas from collapsing. That
being said,
these global simulations (however pioneering) suffered from a lack of
resolution and probably did not adequately capture the characteristic
lengthscales of either instability. On the other hand, numerical
studies of outbursts involving both MRI and GI model one or both
process as a diffusion with an effective alpha parameter (e.g., Armitage et
al.~2001, Zhu et al.~2010). Though limit
cycles can be obtained this way, there is yet no direct evidence that this
is the case when the
different turbulent flows are simulated directly.\\
One obvious response to these issues is to perform
3D vertically stratified shearing box simulations in which the
intrinsic scales of both instabilities are resolved. This is a
computationally demanding task, as the MRI inhabits lengthscales
less than the scale height $H$, while the GI saturates on scales
much greater than $H$. A preliminary (and almost unavoidable)
approach is to conduct 2D MHD simulations. Though this simpler
setup precludes the MRI, it allows us to identify
important MHD processes that should be shared by 3D simulations.
It also provides a fair description of places in the disk that are
magnetically active and yet MRI-stable, such as at certain outer
radii in massive young disks and the dead zones of older disks
at the onset of
an outburst. In this paper we present a suite of such 2D
simulations combining MHD and GI. Our computational domains are, for
the most part, threaded by a mean toroidal field of varying strengths
and the gas is endowed with a simple linear cooling law. Both ideal
and resistive MHD are tested. Though ambipolar diffusion and the Hall effect
are important (often dominant)
players in the weakly ionised plasma, they are
omitted here for simplicity. In our simulations, Ohmic (or grid) diffusion may be
interpreted as a very crude proxy for whatever
process is diffusing and destroying magnetic field. \\
The main result of our exploratory work is that the presence of an imposed magnetic field
can dramatically change the thermodynamic properties of the
gravito-turbulent state. The turbulent motions stretch, distort, and amplify the
magnetic field to strengths of order, or even exceeding, the kinetic
energy. Dissipation of this energy leads the system to
a quasi-steady state that is
markedly hotter than in hydrodynamical simulations, with a mean Toomre
$Q$ sometimes over 10. The mean $Q$ correlates with the
strength of the imposed magnetic field. Adding resistivity weakens
this phenomena but does not qualitatively change the picture.
The details of the dissipation are striking, with energy
thermalised primarily in current sheets and in the slow shocks
generated by reconnection events. Reconnection also gives rise to
`magnetic islands', or plasmoids, that persist for hundreds of orbits.
Finally, we investigate the propensity of the system to fragment as we
change the cooling time. In summary, no great qualitative change in
the critical cooling time is observed.
The structure of the paper is as follows.
In the following section we present the model, its governing
equations, and the numerical methods that we deploy in their solution.
Our main ideal MHD results appear in section 3, the principal
control parameter
being the strength of the imposed field. Effects induced by resistivity are
investigated in section 4. In section 5, we go into more detail
exploring the nature of reconnection in the simulations.
Finally, in section 6, we discuss the
astrophysical implications of this work and how it prepares for future
simulations in 3D.
\section{Model and numerical framework\label{model}}
\subsection{{Model and equations}}
The physical set-up, governing equations, and numerical approach is similar to that described
by \citet{Pdkooper2012}. We use a local Cartesian model
of an accretion disk \citep[the shearing sheet;][]{goldreich65},
whereby the axisymmetric differential rotation is approximated locally
by a linear shear flow $\mathbf{u}_0=-S x\, \mathbf{e}_y$ and a
uniform rotation rate $\boldsymbol{\Omega}=\Omega \, \mathbf{e}_z$, with
$S=(3/2)\,\Omega$ for a Keplerian equilibrium. We denote $(x,y,z)$
respectively as the shearwise, streamwise and spanwise directions,
corresponding to the radial, azimuthal and vertical directions.
We also refer to the $y$ projection of a vector field as its
toroidal component. We neglect the vertical structure of the disc
and consider it infinitely thin, so that the gas is allowed to
move only in a two-dimensional frame $(z=0)$. For simplicity, we
assume that the gas is ideal, its pressure $P$ and surface
density $\Sigma$ related by $\gamma P=\Sigma c_s^2$, where $c_s$
is the sound speed and $\gamma$ the ratio of specific heats. The
pressure is related to internal energy $U$ by $P=(\gamma-1)U$. The
evolution of surface density $\mathbf{\Sigma}$, velocity {field}
perturbations $\mathbf{u}$, magnetic field $\mathbf{B}$ and internal
energy $U$ is then governed by the 2D compressible dissipative MHD
equations:
\begin{equation}
\dfrac{\partial \Sigma}{\partial t}+\nabla\cdot \left(\Sigma \mathbf{u}\right)=0,
\label{mass_eq}
\end{equation}
\begin{multline}
\frac{\partial{\mathbf{u}}}{\partial{t}}-Sx\frac{\partial{\mathbf{u}}}{\partial{y}}+\mathbf{u}\cdot\mathbf{\nabla
u}-Su_x\mathbf{e}_y +2\boldsymbol{\Omega}\times\mathbf{u} =-\nabla\Phi\\
+\frac{1}{\Sigma}(-\mathbf{\nabla}\mathcal{P}+\mathbf{B}\cdot\mathbf{\nabla B}+\nabla\cdot \boldsymbol{\Pi}),
\label{ns_eq}
\end{multline}
\begin{equation}
\frac{\partial{\mathbf{B}}}{\partial{t}}-Sx\frac{\partial{\mathbf{B}}}{\partial{y}} = -SB_x\mathbf{e}_y+\nabla\times(\mathbf{u}\times\mathbf{B})+\eta\mathbf{\Delta B},
\label{magnetic_eq}
\end{equation}
\begin{equation}
\dfrac{\partial U}{\partial t}-Sx\dfrac{\partial U}{\partial
y}+\nabla\cdot (U\mathbf{u})
=-P\nabla\cdot\mathbf{u}-\dfrac{U}{\tau_c}+Q_D+\kappa_{th}{\Delta T} .
\label{energy_eq}
\end{equation}
To this set we must add the solenoidal condition $\nabla\cdot\mathbf{B}=0$.
In the Navier Stokes equation \eqref{ns_eq}, $\mathcal{P}$ is the sum of gas
pressure $P$ plus magnetic pressure $B^2/2$ and $\Phi$ is the
gravitational potential induced by the disc, obeying the Poisson
equation. The (molecular) viscous stress tensor is $\sf{\Pi}$ and is
defined by
\begin{equation}
\boldsymbol{\Pi} = \Sigma\nu\left[\nabla\mathbf{u} +
\left(\nabla\mathbf{u}\right)^T-\tfrac{2}{3}(\nabla\cdot\mathbf{u})\boldsymbol{I} \right].
\end{equation}
The constant kinematic viscosity and magnetic diffusivity are denoted
by $\nu$ and $\eta$. In the energy equation, we use a cooling law
that is linear in $U$ and whose typical timescale is $\tau_c$ (also
called the cooling time). The viscous and Ohmic heating is
$Q_D=\boldsymbol{\Pi} : \mathbf{\nabla v} + \eta \vert
\nabla \times \mathbf{B} \vert ^2 $. The last term on the right hand
side of the energy equation describes
thermal
conduction, which involves the temperature $T=P/ (R\Sigma)$, with $R$ the gas
constant, and the thermal conductivity
$\kappa_{th}$. We define $\Omega^{-1}$ as our
unit of time and $H_0=c_{s_0}/\Omega$ our unit of length where
$c_{s_0}$
is the uniform sound speed of the background laminar state at $t=0$.
Lastly, $\Phi$ is computed from the Poisson equation
\begin{equation}
\nabla^2\Phi = 4\pi G\rho,
\label{poisson_eq}
\end{equation}
where $\rho$ is the three-dimensional density distribution of the gas
which may be related to the surface density via
\begin{equation}
\rho(x,y,z) = \Sigma(x,y)\delta(z),
\label{poisson_eq}
\end{equation}
with $\delta$ the Dirac delta function. Note that we omit
a smoothing length, thus the self-gravitational potential
can have scales comparable to the grid size of the simulations. The
effect of a smoothing length is discussed in \cite{Pdkooper2012}.
\subsection{Diagnostics}
\label{alpha}
First let us define
$\left<. \right>=\frac{1}{L_xL_y}\int\int
\left(\int_{-\infty}^{\infty}\, .\,\, dz\right) dx dy$ as the volume
average of a quantity over a Cartesian portion of size
$L_x$ and $L_y$. A quantity that will be widely used in this paper is
the coefficient $\alpha$ which measures the
angular momentum transport. This quantity is related to
the average Reynolds stress $H_{xy}$, Maxwell stress $M_{xy}$, gravitational stress $G_{xy}$ and molecular viscous stress $\Pi_{xy}$ by:
\begin{align}
\alpha=\dfrac{2}{3\gamma \left\langle P\right\rangle}\left\langle
H_{xy}+M_{xy}+G_{xy}+\Pi_{xy} \right\rangle,
\end{align}
where
\begin{align*}
H_{xy}=\Sigma u_xu_y \quad M_{xy}=-B_xB_y \quad \text{and}
\quad G_{xy}=\dfrac{1}{4\pi G}\dfrac{\partial \Phi}{\partial
x}\dfrac{\partial \Phi}{\partial y}.
\end{align*}
It is straightforward to show that the radial flux of angular momentum
gives rise to the only source of energy in the system that can balance the
cooling. This energy, initially in the form of kinetic energy, can
be stored in magnetic fields but is irremediably converted into heat by turbulent motions.
In order to study the energy budget of the flow, we introduce
the average kinetic, magnetic,
gravitational and internal energy denoted by
\begin{align*}
E_c=\frac{1}{2}\langle\Sigma \mathbf{u}^2\rangle, \quad
E_m=\frac{1}{2}\langle\mathbf{B}^2\rangle, \quad E_G=~\langle \Sigma \Phi
+\frac{1}{8\pi G}\vert\mathbf{\nabla \Phi}\vert^2\rangle,
\end{align*}
and $U=(\gamma-1)\langle P \rangle$ respectively. Although the temperature and
thermodynamic balance can be very different from one cell to
another,
we define an average Toomre parameter in the domain
\begin{equation}
Q=\dfrac{\left\langle c_s\right\rangle \Omega}{\pi G \left\langle \Sigma\right\rangle }
\end{equation}
\subsection{Numerical methods}
We employ the 2D shearing box to simulate locally the motion of
the fluid. Because
the fluid in a gravito-turbulent disc is compressible and is mostly
heated by shocks, we use the PLUTO code \citep{mignone2007} to
perform direct numerical simulations of
Eqs. \eqref{mass_eq}-\eqref{poisson_eq}. This code uses a Godunov
scheme, a conservative finite-volume method that solves the
approximate Riemann problem at each inter-cell boundary. This scheme
is known to successfully reproduce the behaviour of conserved quantities
like mass, momentum and energy through discontinuities. The Riemann
problem is handled by the HLL solver which has the advantage of being
robust and preserving positivity. Usually HLLD solvers are more
suitable for MHD problems but we checked that our results are not
strongly modified when the HLL solver is used. In the shearing box
framework, simulations are performed in a finite domain of size
$(L_x,L_y)$, discretised on a mesh of $(N_X,N_Y)$ grid points. The
boundary conditions are periodic in $y$,
while shear-periodicity is imposed in $x$.
To compute the gravitational potential, we take advantage of the
shear-periodic boundary conditions, following \citet{Gammie2001}. At
each time step, we first shift back the density in $y$ to the time it
was last periodic $(t = t_p)$. For this, we perform a 1D forward
Fourier transform in $y$ for each $x$, multiply by a complex phase
$\exp{(-\text{i}Sk_yx(t-t_p))}$, and take the inverse 1D Fourier
transform. As the resulting surface density and gravitational
potential are periodic in $x$ and $y$, they can be expressed as a
discrete sum of Fourier modes ($\Sigma_k$, $\Phi_k$) with wavevectors
$\mathbf{k}=(k_x,k_y)$. The Fourier decomposition is done with a 2D
FFT algorithm. We then solve the Poisson equation in Fourier
space where a solution for a single mode is
\begin{equation}
\Phi_k = -\dfrac{2\pi G\Sigma_k}{|\mathbf{k}|}
\label{fourier_decompo}
\end{equation}
By multiplying $\Phi_k$ by $\text{i}k_x$ and $\text{i}k_y$, we obtain the
self-gravity force in the Fourier space. An inverse
FFT delivers the force in the real domain.
The linear stability of an infinitely thin layer
has been tested to ensure that our implementation is correct
(see Appendix A). Note that gravitational
energy and stresses are computed directly in
Fourier space in the same way as \citet{Gammie2001}.
Finally, we use the orbital advection algorithm of PLUTO, based on
splitting the equation of motion into two parts, the first containing
the linear advection operator due to the background Keplerian shear
and the second the standard MHD fluxes and source terms. This
operation allows larger time steps and eliminates numerical artifacts
at the boundaries where the Mach number associated with
the background shear flow can be very large.
\subsection{Simulation setup}
\subsubsection{Box size and resolution}
The axisymmetric linear theory for thin discs shows that the flow is
unstable for $Q\leq 1$, with the fastest growing mode possessing
a radial lengthscale of order $2\pi H\,Q$.
Although our simulations are focused on the
regime $Q\gtrsim 1$, we expect typical lengthscales
to be also $\gtrsim H$. In order to
obtain a good statistical average of the fluctuating properties, it is
then necessary that $L_x \sim L_y \gg H$. Our fixed reference
lengthscale is the initial scale height of the gas $H_0$.
As the gas heats up (or cools down) the temperature, and
hence the the scale height $H$, varies.
Previous hydrodynamic simulations show that steady turbulent flows are able to
sustain an average $Q$ around $2-3$, which translates to an average
$H\sim 1-2\, H_0$. We hence choose $L_x=L_y=40\,H_0$ to
make sure that the structures that develop in the box are much smaller
than the box size. For comparison \citet{Gammie2001} and
\citet{Pdkooper2012}
used a box of size 100 $H_0$.
An appropriate resolution is not easy to guess.
The work of \citet{Gammie2001} suggests that a resolution of 5 grid
cells per $H_0$ is the minimum required. This
ensures that the energy lost by the numerical scheme remain small
compared to the energy radiated away by the cooling law. However,
\citet{Pdkooper2012} showed that the
fragmentation criterion is still dependant on resolution when the
latter exceeds 40 points per $H_0$. In particular, increasing
resolution leads to easier fragmentation at higher values of $\tau_c$.
In fact, fragmentation appears to be
a stochastic process whose probability of occurrence decreases with
increasing $\tau_c$. The reasons for this resolution dependence remain
unclear and might depend on the algorithm or code
implementation. \citet{Pdkooper2012} argued that the numerical scheme
and resolution needs to be sufficiently accurate so as to maintain a coherent
clump of size $H$ over many dynamical timescales. In this paper, we used
a resolution of 51 points per height scale $H_0$ which translates to
$N_X=N_Y=2048$ for the entire box so that we are slightly better
resolved than the most accurate run of \citet{Pdkooper2012}.
Because of the prevalence of shocks in the compressible
gravitoturbulence, it is practically impossible to viscously
resolve the shortest scales. However, we checked that
average turbulent quantities (such as mean $Q$, the mean energies, etc)
remain relatively unchanged when using a resolution
of $N_X=N_Y=1024$, suggesting that our simulations are resolved in
this respect.
Small-scale magnetic features, on the other hand, are possible to
resolve physically if Rm is sufficiently low.
\subsubsection{Initial conditions}
\label{ic}
Initial conditions require particular attention as they determine
if the flow reaches a steady turbulent state or not. We
start our simulations with a uniform density distribution
$\Sigma_0=1$. The total mass in the box is conserved so that $\left<\Sigma\right>$ at any time is equal to $\Sigma_0$.
The initial Toomre parameter $Q$ cannot be smaller than 1 since linear
axisymetric fluctuations automatically lead to
fragmentation. To make sure that such fluctuations cannot grow, we
choose $Q_0=1.6$ at $t=0$ which corresponds to a fixed gravitational constant $G=0.2$ in all simulations.
We generate a random seed in the initial density and velocity
perturbations by injecting a small amount of energy in all $k_x$ and
$k_y$ Fourier components. We find that the noise amplitude has to be
sufficiently high to excite a turbulent flow, which confirms that
the transition to such flow is subcritical. Starting with a
sufficiently large fluctuation, the development of the turbulence is not
immediate but takes a finite time $t_\text{trans}$. For cooling times
smaller than $t_\text{trans}$, the fluid can cool down to $Q < 1$
before any heating through turbulent motions. This leads to premature
fragmentation. To avoid that, we switch on the
cooling term after a turbulent state has been reached.
For MHD simulations, initial velocity and density fields are taken
from a pre-existing gravito-turbulent state obtained by a
hydrodynamic run. A large scale uniform toroidal magnetic field
$B_{y_0}$ is then introduced into the box at $t=0$. We define the
initial beta via
\begin{equation}
\beta_0=\frac{2 \Sigma_0 c_{s_0}^2}{B_{y_0}^2},
\end{equation}
the ratio of gas to magnetic pressure of the background laminar
state. The main results of this study are restricted to the case
$\beta_0>1$. Zeldovich's theorem suggests that
no dynamo action is possible in a 2D model, and so a zero-net flux
field will decay over time.
However, the
average toroidal magnetic flux
is conserved during our simulations, which allows rms-magnetic
fluctuations to be maintained indefinitely.
\subsubsection{Cooling and diffusion parameters}
\label{diffusion_coeff}
In our model the total energy is
removed via the term $-U/\tau_c$ which mimics
radiative cooling with an adjustable timescale $\tau_c$. The validity
of this simple cooling law, and realistic values of $\tau_c$,
are important issues. The cooling time in a protostellar disk varies by
orders of magnitude between different radial locations and at different
stages of a disk's evolution. In the later T-Tauri or class-II stages,
$\tau_c$ can be $<1/\Omega$, but in younger class-0 disks the cooling time
can be considerably longer. \citet{kratter16} compute
representative values of $\tau_c$ for various sources, and deduce that (generally)
$\Omega\tau_c>10$ in massive non-fragmenting disks (their exemplar is
IRAS 16293-2422b). Following on from this work, we adopt a range
$\Omega\tau_c=1-50$.
Although
PLUTO conserves total energy, the shearing boundaries do work on the
fluid and provide a source of energy. Depending on how
these source terms are computed, numerical errors can be much larger
than roundoff errors and produce a numerical loss of energy. We checked,
however, that the numerical loss intrinsic to the code is small
compared to physical dissipation. By analysing each
term individually in the global energy budget, we were able to
quantify the ratio between the numerical loss and the total energy
content. We found that for a moderate cooling time this ratio remains
smaller than $10^{-4}$ per dynamical time, which means that on average
less than 10\% of the energy is lost after 1000 $\Omega^{-1}$. In
comparison, \citet{Gammie2001} has a relative numerical loss of the
order $10^{-3}$ per dynamical time, estimated from the difference between their numerical and predicted $\alpha$.
Internal exchange of energies are possible, at least in part, through
the action of dissipative processes such as viscous friction or Ohmic
diffusion, which convert kinetic and magnetic energy into
heat irreversibly. In addition, internal energy $U$ is redistributed
through the fluid by thermal diffusivity. In our simulations, we
introduced a uniform tiny viscosity, such that the Reynolds number
$\text{Re}=\Omega H^2/\nu=1000$, and a moderate thermal conductivity
$\kappa_{th}=0.06$. These coefficients are probably not representative
of any astrophysical disc but avoid large velocity or
temperature gradients. A test has been done with $\kappa_{th}=0$ for
which the average turbulent quantities
remain quite similar to those with $\kappa_{th}=0.06$.
Ohmic resistivity is known to play a significant role in the turbulent
dynamics of accretion discs. Its influence on self-gravitating MHD
turbulence will be first neglected in the simulations of section
\ref{mhd_runs} (so that magnetic energy is dissipated on the grid)
but taken into account in section \ref{resistive_runs}. The magnetic
Reynolds number $\text{Rm}=\Omega H^2/\eta$, defined as the typical
ratio between the advective term and the resistive term, will be
varied from 10 to 5000. For comparison, a very crude estimate of the numerical grid's magnetic
Reynolds number is $= (H_0/L)^2N_X^2 \approx 2500$, though grid diffusion will
not operate like a Laplacian nor be isotropic.
Lastly,
we recognise that ambipolar diffusion and the Hall effect play a
significant and usually dominant role in the external regions of
protoplanetary discs.
However, given that our work is exploratory, we omit
more complicated non-ideal MHD for simplicity. We hence regard
Ohmic diffusion in our model as a (very) coarse proxy for whatever diffusive
process is dominating locally (which could also include small-scale
magnetic turbulence driven by the MRI). This is discussed in more detail in
section 4.
\section{Gravito-turbulence with and without a magnetic field}
In this section, we present several pure hydrodynamical
simulations that test our code and provide a point of
comparison with later magnetized simulations. We then study the coupling
between the gravito-turbulence and a magnetic field, focussing especially
on global properties and the fragmentation
criterion in different magnetic regimes, from weakly magnetized
($\beta_0 \gg 1$) to rather strongly magnetized ($\beta_0 \sim 1$).
\subsection{Hydrodynamical simulations}
In hydrodynamic shearing box simulations with $\tau_c\gg
\Omega^{-1}$, the system settles on
a strongly turbulent state that maintains $Q$ around unity. The
heat generated by turbulent motions acts as a feedback loop that
regulates the thermodynamic state. For example, if $Q$ takes values
too
low, the instability becomes more active and enhances the gas
temperature so that the system returns to equilibrium. This idea was
first
proposed by \citet{paczynsky78} and
numerically demonstrated by \citet{Gammie2001} in the shearing box. The
latter also showed that gravito-turbulence transports a significant
amount of angular momentum and predicted that this transport is
inversely proportional to the cooling time $\tau_c$. When $\tau_c
\sim \Omega^{-1}$, the behaviour is radically different and the disc
fragments into massive clumps. For $\gamma=2$ and a numerical
resolution of $10$ points per scale height, \citet{Gammie2001} found
that the critical $\tau_{c}$ for which fragmentation occurs is
$\simeq 3 \Omega^{-1}$. In actual fact, there is no
clear transition between sustained gravitoturbulence and fragmentation,
as explained by \citet{Pdkooper2012}.
Note also that localized fragments can form stochastically
without disrupting the whole disc. \\
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{FIG_fragmentation_hydro.pdf}
\caption{The angular momentum transport coefficient $\alpha$ as a
function of $\tau_c$ in purely hydrodynamic gravitoturbulent
simulations. Green circles represent simulations with no fragments
while orange circles represent simulations where one or several transient fragments are observed. The straight lines are the theoretical predictions of \citet{Gammie2001} and the end bars correspond to the critical cooling time below which the disc is fragmenting as a whole.}
\label{fig_frag_hydro}
\end{figure}
In order to check our code and compare our results with these previous
studies, we performed several simulations without magnetic field
($B_{y_0}=0$) and a varying $\tau_c$. We used the procedure
described in section \ref{ic} to generate suitable initial
conditions. Two different adiabatic indices were considered,
$\gamma=5/3$ and $\gamma=2$. Simulations that did not fragment into one
or few massive clumps were run for at least
$600\,\Omega^{-1}$ in order to obtain well-defined saturated
states. Fig.~\ref{fig_frag_hydro} shows the average angular momentum
transport coefficient $\alpha$ of these turbulent states,
defined in section \ref{alpha}, as a function of $\tau_c$.
We show that for both $\gamma=5/3$ and $\gamma=2$, the $\alpha$
coefficient follows the theoretical prediction of \citet{Gammie2001}
(indicated by straight lines). As this prediction is derived from
an energy conservation principle, this result is just saying that
numerical energy
losses are small in our simulations. Fig.~\ref{fig_frag_hydro} also shows
that these turbulent states collapse into clumps when $\tau_c$
is decreased, though the transition is not
necessarily well defined, as in \citet{Pdkooper2012}. For $\gamma=2$, local and transient
fragments appear first for $\tau_c=6\,\Omega^{-1}$ while the entire
computational domain fragments when
$\tau_c\lesssim 5\,\Omega^{-1}$. For $\gamma=5/3$, the first fragments
appear at $\tau_c\lesssim 10\,\Omega^{-1}$ while massive
unstable clumps are formed below $\tau_c\lesssim 8-9 \,\Omega^{-1}$.
In both cases, the critical $\alpha_c$ for which the disc fragments is comparable and around $\alpha_c \simeq 0.04-0.05$.
We found that the average Toomre parameter $Q$ in steady turbulent
simulations does not depend strongly on $\gamma$. However it seems to
slightly decrease as $\tau_c$ is reduced, going from $Q=3$ when
$\tau_c=50\,\Omega^{-1}$ to $Q=2$ when $\tau_c=10\,\Omega^{-1}$. This
behaviour is not very surprising: when $\tau_c$ is decreased,
cooling is enhanced requiring a commensurate increase in turbulent
heating by GI, only possible by decreasing $Q$.
\subsection{MHD simulations: dependence of the gravito-turbulent state on $B_{y_0}$}
\label{mhd_runs}
We first study the `ideal case' in which we do not include any
explicit resistivity. We performed a series of simulations by fixing
the cooling time $\tau_c=20\,\Omega^{-1}$ and the adiabatic index
$\gamma=2$, but varying the background toroidal field $B_{y_0}$ (or
equivalently $\beta_0$). We found that for this particular cooling
time, all simulations with $\beta_0 \gg 1$ reach a steady turbulent
state without developing massive clumps. Although we started with a
pure uniform toroidal field, the geometry of the magnetic field in the
nonlinear turbulent regime becomes very intricate and tangled. In some
cases, it is amplified and the average gas to magnetic pressure ratio
measured in the saturated turbulent regime
\begin{equation}
\beta_{t}=\dfrac{2\left<\Sigma c_{s}^2\right>}{\left<B^2\right>},
\end{equation}
can differ greatly from the initial plasma parameter $\beta_0$.
In the turbulent state, three
different forces emerge in the leading order balance, namely the Lorentz,
pressure gradient, and gravitational forces. By
varying $\beta_0$, we found three different regimes
characterised by the relative importance of the Lorentz force with respect to other forces.
\subsubsection{First regime: $\beta_0,\,\beta_t\gg 1,\,\,E_G \gg E_m$}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{FIG_multi_averages1.pdf}
\caption{Simulations of gravitoturbulence in the regime $\beta_0\gg
1,\,\,E_G\gg E_m$, computed for fixed
$\tau_c=20\Omega^{-1}$ and $\gamma=2$. From left to right,
$B_{y_0}=0$, $B_{y_0}=0.01$ and $B_{y_0}=0.025$. Top panels:
snapshots of the surface density $\Sigma$ in the disc plane
($x$,$y$). Centre panels: the time evolution of the Reynolds
($H_{xy}$), gravitational ($G_{xy}$), Maxwell ($M_{xy}$) and
viscous ($\Pi_{xy}$) stresses. Bottom panels: the time evolution of
kinetic $E_c$, gravitational $-E_G$, magnetic $E_m$ and internal
energy $U$ in a logscale. The simulation in the second column with
$B_{y_0}=0.01$ was started from a hydrodynamic run. The third
simulation, with $B_{y_0}=0.025$, was initiated from a MHD run with
$B_{y_0}=0.05$, which explains why the magnetic energy and Maxwell
stress first decrease before reaching a steady state.}
\label{fig_multi_averages1}
\end{figure*}
The first regime corresponds to the case of a small Lorentz force
compared to the gravitational and pressure forces. The magnetic field
is completely slaved to the gravito-turbulence and its back-reaction
on the fluid motion is insignificant or weak. The field can be stretched or
compressed so that it grows until reconnection processes take place
and destroy it. Figure \ref{fig_multi_averages1} shows three
different simulations obtained respectively for $B_{y_0}=0$,
$B_{y_0}=0.01$ and $B_{y_0}=0.025$ ($\beta_0=\infty$, $\beta_0=20000$
and $\beta_0=3000$). In the case of a weak but non-zero magnetic field,
the flow undergoes a transient phase, before it reaches a steady turbulent state. The final
state looks much like the hydrodynamic one although turbulent structures
appear on slightly smaller scales. The
center panels of Fig.~\ref{fig_multi_averages1} indicate that the
Maxwell stress is small compared to the Reynolds and gravitational
stresses. The time evolution of the energy budget is shown in the
bottom panels. It is clear that magnetic energy remains at least an
order of magnitude smaller
than the other sources and is dynamically insignificant to a first
approximation.
Figure \ref{fig_averages_By} shows some key dimensionless quantities averaged
in space and time. The Toomre parameter $Q$ remains close to the
hydrodynamic value, although it slightly increases from $B_{y_0}=0$ to
$B_{y_0}=0.05$. This result suggests that the temperature regulation
and energy balance are only marginally affected by magnetic fields in this
regime. We note that $\beta_t$, which is directly related to $E_m/U$,
increases very rapidly as a function of $B_{y_0}$,
even if it remains much smaller than the other energy ratios.
\subsubsection{Second regime: $\beta_0,\,\beta_t> 1,\,\, E_m
> E_G$}
\label{plasmoid_regime}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{FIG_multi_averages2.pdf}
\caption{Simulations of gravitoturbulence in the regime $\beta_0\gg
1,\,\,E_m > E_G$, computed for fixed
$\tau_c=20\Omega^{-1}$ and $\gamma=2$. From left to right,
$B_{y_0}=0.05$, $B_{y_0}=0.1$ and $B_{y_0}=0.4$. Top panels:
snapshots of the surface density $\Sigma$ in the disc plane
($x$,$y$). Centre panels: the time evolution of the Reynolds
($H_{xy}$), gravitational ($G_{xy}$), Maxwell ($M_{xy}$) and
viscous ($\Pi_{xy}$) stresses. Bottom panels: the time evolution of
kinetic $E_c$, gravitational $-E_G$, magnetic $E_m$ and internal
energy $U$ in a logscale. The simulations shown in the first and
second columns were started from the hydrodynamic state while the
third was initiated from the MHD steady state
obtained for the stronger field run $B_{y_0}=0.1$.}
\label{fig_multi_averages2}
\end{figure*}
For stronger imposed fields we enter a second regime corresponding to
when the Lorentz force is
comparable to or larger than the gravitational force, but still smaller
than pressure gradients. Unlike the first regime, the
back reaction of the magnetic field on the fluid motion is no longer
negligible. Figure \ref{fig_multi_averages2} shows three
simulations obtained respectively for $B_{y_0}=0.05$, $B_{y_0}=0.1$ and $B_{y_0}=0.4$ ($\beta_0=800$, $\beta_0=200$ and $\beta_0=12.5$).
The plots in the center row indicate that the Maxwell stress $M_{xy}$
is now the largest contributor to the total stress and therefore to
the angular momentum transport. Although the total stress increases with
$B_{y_0}$ owing to the new term $M_{xy}$, $\alpha$ remains constant ($\alpha=0.022$) as it only depends on the cooling time (see \citet{Gammie2001}).
The magnetic energy is strongly enhanced in this regime and becomes
much larger than the background. Figure \ref{fig_averages_By}
shows that the ratio between magnetic energy and internal energy
increases with $B_{y_0}$ and saturates at larger $B_{y_0}$. The plasma is close to equipartition between magnetic energy and kinetic energy, as $E_c/U$ and $E_m/U$ tend to a similar value.
Figure \ref{fig_averages_By} tells us that the internal energy increases
significantly with $B_{y_0}$ because the average $Q$ and sound speed
increase by an order of magnitude between $B_{y_0}=0.05$ and
$B_{y_0}=0.4$. For $B_{y_0}=0.1$, the average temperature in the box
is multiplied by a factor ten compared to the temperature in hydrodynamic
simulations. This surprising result shows that a gravitoturbulent state can
exist at values of $Q$ much larger than in hydrodynamics.
The greater temperatures are associated with the formation
of elongated current sheets and consequent heating at those locations
through magnetic reconnection (and associated shocks). The substantial amplification of the
background field by the turbulence and its subsequent dissipation
provides a powerful and additional
source of heat. This is analysed in more detail in section
\ref{heating_source}.
A consequence of this rise in temperature is
that the gravitational instability becomes weaker, as indicated by the
gravitational stress which clearly decreases by one or two orders of
magnitude. Figure \ref{fig_averages_By} shows that most of the
gravitational energy present in the small $B_{y_0}$ regime has been
replaced by magnetic energy. It is then reasonable to ask whether
self-gravity is important at all in this state. To check that, we ran a
simulation starting from developed gravitoturbulence and then switched off
the gravitational term. We found that as soon as self gravity is
suppressed, the turbulent kinetic and magnetic energy decay to
negligible levels by $t \sim 100\, \Omega^{-1}$. This indicates that
self-gravity,
however weak it is, still plays a crucial role in sustaining the
turbulence.
The fact that graviturbulent activity persists for such
large $Q$ may have something to do with the relaxation of
angular momentum conservation by the magnetic stresses. It is likely that
transport via a tangle of magnetic
fields weakens the stabilising effect of rotation, permitting GI to
operate for larger $Q$. A similar effect is witnessed when explicit
viscosity is included in the linear theory (e.g.\ Schmit \&
Tcharnuter 1995) and probably in the
nonlinear onset of GI. It should be acknowledged that in pure hydrodynamics
it is not yet understood what sets the level of the saturated $Q$;
adding a magnetic field must further
complicate the
problem.
The top panels in Fig.~ \ref{fig_multi_averages2} show that the
plasma is characterized by dense clumps whose size are comparable to or
smaller than $H_0$. These `plasmoids' are associated with magnetic
island structures and evolve in a turbulent background that
resembles the hydrodynamic state (non-axisymmetric waves
amplified transiently and dissipated into shocks). These plasmoids
appear
to resist the shear and the shocks that propagate
through them. A detailed analysis of these structures is provided in
section \ref{plasmoid}. Note that the magnetic islands are
reminiscent of compressible turbulent 2D MHD simulations in which a
forcing term is included \citep{lee03}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{FIG_averages_By.pdf}
\caption{Top panel: the average Toomre parameter $Q$ as a function
of $B_{y_0}$. Bottom panel: ratios of kinetic, magnetic and
gravitational perturbation energy to the total internal energy for different $B_{y_0}$. The yellow dashed line is the turbulent angular momentum transport coefficient $\alpha$. All these simulations have been computed for a fixed $\tau_c=20\, \Omega^{-1}$ and a fixed adiabatic index $\gamma=2$.}
\label{fig_averages_By}
\end{figure}
\subsubsection{Third regime: $\beta_0,\,\beta_t \leq 1,\,\,E_m \gg E_c$}
As we approach $\beta_0=1$, the fluid becomes magnetically dominated
and the gravitational term is reduced. For $\beta_0=1$ and
$\beta_0=0.5$, which correspond probably to an unrealistic regime for
astrophysical discs, our simulations show that the fluid motion is
completely frozen into the magnetic field lines and unable to move due to
strong magnetic tension. The toroidal field acts like a
`straightjacket', and steady turbulent states
cannot be achieved. Flux tubes with a coherent radial length larger
than the non-axisymmetric gravitational structures develop and
sometimes form regions of high density that collapse rapidly after
a few orbits. These structures are very similar to those obtained by
\citet{lee03}. We note that this regime is extremely challenging
numerically, especially with our fine resolution, since the time step becomes very small.
\subsection{Fragmentation criterion}
\label{fragmentation_B}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{FIG_fragmentation_By.pdf}
\caption{Fragmentation criterion as a function of the cooling time
and $B_{y_0}$. Green circles indicate simulations that do
not fragment at all. Orange circles indicate simulations
that remain turbulent but for which one or several transient
fragments are observed. Red diamonds correspond to simulations where
the disk collapses into one or a few massive clumps.}
\label{fig_frag_by}
\end{figure}
In this section we vary the cooling time $\tau_c$ in order to
study the onset of fragmentation. Each simulation is
run for $100\Omega^{-1}$. For all values of the
background field $B_{y_0}$, there is a critical $\tau_c$ below
which
the disc collapses into massive clumps (within the simulation time).
Figure \ref{fig_frag_by} shows how this critical $\tau_c$ varies
with $B_{y_0}$. In the first regime,
$E_G \gg E_m$, the critical cooling time is
multiplied by a factor $\sim 2$ between $B_{y_0}=0$ and $B_{y_0}=0.05$.
Weak magnetic tension helps fragmentation by
extracting extra angular momentum from within a potentially collapsing
region, thus thwarting the Coriolis force.
In contrast, the plasmoid-dominated regime, $E_m \gtrsim E_G $, witnesses
an unexpected decrease in the critical cooling time with $B_{y_0}$.
This is probably due to
the enhanced heating generated in
the presence of a stronger imposed field. An increased
magnetic field causes the gas to be signifcantly hotter, as explored
in the previous subsection.
The associated pressure prevents fragmentation and
overwhelms the direct destabilising effect of the magnetic field in collapsing
a plasmoid, via magnetic tension and pressure.
Overall, the relative variation of the critical
cooling time with $B_{y_0}$ remains small, which suggests that
magnetic fields, despite their strong influence on the turbulent
properties, do not dramatically change the fragmentation
criterion. Note that, similarly to the hydrodynamical case, the average
$Q$ decreases when $\tau_c$ is decreased ($Q\sim 30$ for $\tau_c=20$
and $Q\sim 18$ for $\tau_c=7 \,\Omega^{-1}$ in the case of
$B_{y_0}=0.4$). No criterion for fragmentation depending on
$\alpha$ or $Q$ can be obtained simply in that case.
\section{Effects of resistivity}
\label{resistive_runs}
Up to now we have explored only `ideal' MHD --- the particulars of the grid
have been taking care of the reconnection, diffusion, and
thermalisation of magnetic field. In this section we include magnetic resistivity
explicitly to better control this process and also to push our models
to regimes relevant to the more resistive radii in protoplanetary
disks. We find that increasing diffusion, unsurprisingly, impedes the
build up of the strong fields witnessed in section \ref{mhd_runs}; the
field slips through the
turbulent gas and is no longer wound up, stretched, and amplified
as efficiently.
In protoplanetary discs, the magnetic Reynolds number Rm
is directly proportional to the gas's ionisation
fraction, which is determined by interparticle collisions,
cosmic rays and X-rays, radioisotopes,
molecular recombination, and dust grain physics
\citep{armitage11}. In a minimum mass solar nebula (MMSN) model,
estimates for the midplane Rm vary from $\sim 0.1$
at 5 AU to $10^3$ at 10 AU, to greater than $10^4$ at larger radii \citep{simon15}.
In fact, both ambipolar diffusion and the Hall effect are more important
than Ohmic diffusion at the latter two radii. The equivalent ambipolar
magnetic Reynolds number is defined as $R_A =
\Omega H^2 \nu_{in} \rho^2 x_e/ B^2$ where $\nu_{in}$ is the ion-neutral
collision rate and $x_e$ is the ionisation fraction. This number
varies roughly between 0.1 to 10 times $\beta$
between 5 and 100 AU \citep[e.g.][]{simon15}. If we are permitted to
crudely model ambipolar diffusion by Ohmic diffusion in our
simulations,
then our effective Rm
should take values between 1 and $10^4$.
The reader should be aware, however, that
the above estimates for non-ideal MHD were derived with the MMSN model,
which best describes an older type-II disk, which is insufficiently
massive to suffer GI. The relative strengths of Ohmic and ambipolar
diffusion will differ in a GI-unstable type-0 system, which will be denser
(hence the ions and neutrals better coupled) but also less
well-ionised because optically thicker. The above estimates hence only
serve as a rough guide, to fix ideas.
An additional
complicating factor is that the ionisation fraction (and hence Rm and $R_A$)
depends on height, and so the gas at
different vertical levels is not coupled to the magnetic field in the
same way.
These issues make it less than straightforward to assign a simple
Ohmic diffusivity to 2D simulations, and in fact to interpret the
role of diffusion in two dimensions generally.
One other extremely important
ingredient, neglected in our work, is the MRI. Though absent in the
heart of dead zones $\sim 5$ AU, it may appear in the more favourable
ionisation conditions at larger radii, though the details of its
prevalence are exceptionally complicated and the subject of
intensive research. The outcome depends not only on the
(poorly constrained) ionisation profile, but on the orientation,
strength, and
existence of a net magnetic field, with simulations showing that the midplane
can be completely laminar, undergo bursts of turbulence, or sustain a sluggish
form of the MRI. The
surface regions, on the other hand, may launch a magnetocentrifugal
wind or suffer vigorous turbulence
\citep{simon13b,simon15}.
We cannot hope to adequately model this
physics in our 2D simulations, but hope that its diffusive aspects can
be roughly described by a constant resistivity.
\begin{table}
\centering
$B_{y_0}=0.05$ i.e $\beta_0=800$
\begin{tabular}{c c c c c c}
\hline
Rm & $Q$ & $E_m$ & $E_m/U$ & Plasmoids \\
\hline
0 (hydro) & 3.02 & x & x & x \\
10 & 3.16 & 0.0011 & 0.002 & NO \\
100 & 3.2 & 0.005 & 0.0026 & NO\\
500 & 3.4 & 0.025 & 0.012 & NO\\
1000 & 4.9 & 0.16 & 0.031 & YES (very few) \\
5000 & 5.7 & 0.27 & 0.037 & YES\\
ideal approx. & 5.8 & 0.36 & 0.042 & YES \\
\hline
\end{tabular}
\vspace{1cm}
\centering
$B_{y_0}=0.1$ i.e $\beta_0=200$
\begin{tabular}{c c c c c c}
\hline
Rm& $Q$ & $E_m$ & $E_m/U$ & Plasmoids \\
\hline
0 (hydro) &3.02 & x & x & x \\
100 & 3.1 & 0.024 & 0.012 & NO\\
250 & 7.6 & 0.28 & 0.026 & YES (very few)\\
500 & 10.9 & 0.84 & 0.035 & YES\\
1000 & 12.5 & 1.35 & 0.045 & YES \\
5000 (hlld) & 13.6 & 1.78 & 0.05 & YES\\
ideal approx.& 13.7 & 1.88 & 0.053 & YES \\
\hline
\end{tabular}
\vspace{0.5cm}
\caption{Average $Q$, magnetic energy, and the ratio of magnetic to internal energies
for different Rm. The first table
corresponds to simulations with $B_{y_0}=0.05$, and the second to
$B_{y_0}=0.1$. The last column indicates whether plasmoids appear.
All simulations were conducted
with the hll solver except for $\text{Rm}=5000$ and
$B_{y_0}=0.1$ which was undertaken with the hlld
solver.} \label{table1}
\end{table}
\subsection{Resistive turbulent simulations}
To determine the effect
of Ohmic diffusion on MHD gravito-turbulent states, we performed a
series of simulations with explicit resistivity by taking a fixed
$B_{y_0}$ and varying the magnetic Reynolds number. We scanned a large
range of Rm, straddling midplane values typical of dead zones and larger
radii. Our simulations were initiated from a
saturated state computed in the ideal limit and run until a new steady
state was found.
Table \ref{table1}a) sums up the different results obtained for
$B_{y_0}=0.05$ (equivalently $\beta_0=800$).
As expected, when $\text{Rm}= 5000$ the turbulent
state
differs little from the `ideal case' because the numerical and physical Rm
are of the same order. For such values of Rm, Ohmic diffusion is probably
unresolved.
However, the statistical properties of the
turbulence change quite drastically when $\text{Rm}\lesssim 1000$. For
this transitional Rm, the magnetic energy is roughly halved
while $Q$ drops to 4.9 (from 5.8). Another change is
that the number of plasmoids in the box is considerably reduced while
their typical density decreases by a factor 4. Below
$\text{Rm}=500-1000$, plasmoids structures disappear and $Q$ approaches
its hydrodynamical value. At these Rm, diffusion impedes the
build up of large magnetic energies (that may be subsequently
thermalised) and the disk is hence cooler.
Results for a stronger imposed field $B_{y_0}=0.1$ ($\beta_0=200$)
are presented in Table \ref{table1}b). Now the magnetic energy
and $Q$ decrease sharply around a lower critical $\text{Rm}\sim
100-250$, and plasmoid structures disappear for
$\text{Rm}\lesssim 250$.
To conclude, the main effect of resistivity is to reduce the average
turbulent magnetic energy stored in the fluid. As a consequence, there
is less free energy to be dissipated into heat and the mean
temperature decreases. Below some critical $\text{Rm}_c$ that seems to
scale as $\beta_0$, the turbulence becomes decoupled from the
magnetic field. This result suggests that a key quantity to study the
transition between `quasi-hydrodynamical' and plasmoid-dominated
MHD turbulence might be the Elssaser number $\Lambda \sim \text{Rm}\,\beta_0^{-1}$.
\subsection{Zero net flux simulations and decay rate}
As mentioned in section \ref{ic}, a 2D turbulent flow cannot sustain
a dynamo field, which means that if we start with a zero net toroidal
flux, the magnetic field is expected to decay within a finite
time. However, the decay time can be exceptionally long, due to
compressibility, the geometry of the initial
field, and a large Rm (see also Ivers \& James 1984). It may even be
possible to sustain a magneto-turbulent state throughout a great
fraction of a disk's life. To give an
estimate of the decay timescale, we performed two different
simulations with zero net flux. In the first one, labelled `Zs', we
started from a state computed with $B_{y_0}=0.1$, removed the mean
component of the toroidal field and then let the flow evolve in time. In
the second one, labelled `Zl', we started from the same state but we
added at $t=0$ a sinusoidal $B_{y_0} \propto \sin(2\pi/L_y)$ with an
energy equivalent to the one with uniform background field. Both
simulations were performed with explicit resistivity and
$\text{Rm}=5000$. In the Zs case, the magnetic energy $E_m$ decays
to negligible values
by 150 $\Omega^{-1}$, which corresponds roughly to the decay time
expected. Indeed if the turbulent magnetic structures are of scale
$\sim H_0$, then the resistive decay time is given by
$H_0\,\text{Rm}/4\pi^2 \sim 126 \, \Omega^{-1}$. In the Zl case, we
found however that the initial magnetic field is retained over
at least $ 1000\,
\Omega^{-1}$ while magnetic energy stays virtually constant
throughout the
simulation. This is expected because the estimated decay time
for the large-scale $B_{y_0}$ is of the order $2\times10^5\, \Omega
^{-1}$, comparable or longer than the disc viscous
timescale.
\section{Current sheets and plasmoids}
\subsection{Heat sources and currents sheets}
\label{heating_source}
We showed in section \ref{mhd_runs} that for intermediate $B_{y_0}$,
the Maxwell stress produces a large contribution to the total stress and
provides an additional source of thermal energy. The build up of
magnetic energy is another source, once it is dissipated
via current sheets or related
structures \citep{parker72,cowley97}.
As the heat generated by magnetic fields
drastically alters the thermodynamic state
of the turbulence, and in particular the average $Q$, it is crucial to
better understand it.
\subsubsection{Mean features}
To identify the main source of heat in our MHD simulations,
we investigated the relative importance of each term in the averaged equation for internal energy:
\begin{equation}
\label{eq_average_heat}
\dfrac{d \langle U \rangle}{dt}= \langle - P \nabla \cdot \mathbf{u}
\rangle +\langle D_{\nu}+D_{\eta} \rangle -\langle U \rangle/\tau_c.
\end{equation}
Physically, the heat can be generated through two different processes: reversible compression or expansion of the gas
which is associated with the term
\begin{equation}
W_{PV}=- P \nabla \cdot \mathbf{u},
\end{equation}
and irreversible dissipation like viscous and Ohmic friction, whose dissipation rates are respectively:
\begin{equation}
D_\nu= \boldsymbol{\Pi} :
\mathbf{\nabla v} \quad \text{and} \quad D_\eta=\eta \left(\nabla \times \mathbf{B}\right)^2.
\end{equation}
The main difference between these sources is that pressure work
$W_{PV}$ (also called `pressure-dilatation') can have either a positive
or negative sign, meaning that the energy transfer between kinetic
and thermal modes can be in either direction. In contrast, irreversible
processes transfer energy from the kinetic to thermal channels only.
We
analysed the net heat budget for $B_{y_0}=0.1$ and Rm $=500$ by
averaging eq.~(\ref{eq_average_heat}) in time over $150 \Omega
^{-1}$. We found that almost 45\% of the budget is represented by pressure
work $W_{PV}$, 20\% by the Ohmic term $D_{\eta}$ and 6\% by the
viscous term $D_{\nu}$. The leftover is taken up by numerical
dissipation. Note that when a HLLD solver is used, the viscous term
becomes 11\% but the average $Q$ and temperature are unchanged. The
rather significant amount of numerical dissipation is not
surprising and arises because of the presence of thin shocks layers which are
difficult to resolve viscously. However, as discussed in section 2,
only a tiny fraction of energy is lost, the numerical dissipation is
mostly recycled as heat in a way approximating real
microscopic dissipation inside a
shock.
A notable result is that a large fraction of the heat comes from the
reversible expansion of the gas through the pressure dilatation term
$W_{PV}$, which oscillates between positive and negative values, with a
frequency $\approx\Omega$ but which is positive on average.
In most studies, this reversible
heating is considered irrelevant because a fluid parcel in the disc
is thought to relax adiabatically and return to its unperturbed state shortly
after the passage of a spiral wave or a shock
\citep[e.g.,][]{rafikov16}. On
average, the heat generated through an expansion is removed by a
subsequent relaxation because there is a similar degree
of compression and
expansion in the gas ($\langle \nabla\cdot \mathbf{u}
\rangle=0$). This is true when the gas
can relax adiabatically on a timescale
much shorter than the cooling time, and when a gas parcel encounters
waves or shocks on a timescale
longer than the adiabatic relaxation time.
In our simulations neither need be the case. The fluid
endures a turbulent forcing so strong that it has no
time to relax between each compression or shock crossing. A similar
net reversible compressible energy transfer, due to pressure
dilatation, has been observed in 2D and 3D hypersonic
compressible turbulence, with no radiative cooling, in a
separate non-disk context \citep{zeman91,sarkar92}. In particular,
for shear flows, this transfer can be comparable to the compressible
viscous dissipation term and contributes to a reduced growth of
turbulent kinetic energy when the
flow is integrated over a long time (the missing part being transferred to internal energy).\\
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{FIG_heat_source_noB.pdf}
\caption{A snapshot of the distribution of pressure dilatation
$W_{PV}$ in hydrodynamic gravitoturbulence ($B_{y_0}=0$).
Bright and white colors indicates heating by compression,
dark and black indicates region where the gas expands and relaxes.
The intensity has been intentionally saturated at $W_{PV}=10$
but the real maximum is $W_{{PV}_{max}}\sim 150$.}
\label{fig_heat_source}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{FIG_heat_source.pdf}
\caption{Heat sources in a quasi-steady turbulent state obtained for
$B_{y_0}=0.1$ and $\text{Rm}=500$. The top panel represents the
Ohmic dissipation $D_\eta$ whereas the bottom panel represents the
pressure dilatation $W_{PV}$.
The intensity has been
intentionally saturated at $W_{PV}=20$ but the real maximum is
$W_{{PV}_{max}}\sim 5000$, much greater than in hydrodynamical runs.}
\label{fig_heat_source2}
\end{figure}
\subsubsection{Dissipative structures}
In addition to the average budget for the internal energy, we analysed
the spatial distribution of the heat sources in the hydrodynamic case
and in a magnetized gravito-turbulent flow with $B_{y_0}=0.1$. Figure
\ref{fig_heat_source} shows that when $B_{y_0}=0$, the main sources
(here the reversible part) are located in very thin azimuthally
elongated structures. These thin layers correspond to shock waves
that propagate within the fluid and are associated with the nonlinear
evolution of large-scale gravitational wakes. The
black/dark regions correspond to expanding gas ($\nabla\cdot
\mathbf{u}>0$) where the pressure is found to be maximum.
Figure \ref{fig_heat_source2} presents a snapshot from a magnetic
simulation. The second panel shows again that the pressure
work is concentrated into very thin filaments, which reveal the
location of shocks, but their geometry is tremendously more
complicated and their number clearly increased. In comparison with the
hydrodynanical case, the surface covered by these heat sources is
multiplied by a factor $\sim 10$, for $B_{y_0}=0.1$ (this is
estimated by computing the surfaces where $W_{PV}>10$). These
intricate patterns reveal also that the heat transfer is concentrated
in smaller
scale structures, with a typical length that seems correlated
to the size of the magnetic field bundles.
Although we showed
that Ohmic dissipation is not the dominant source of heat directly,
its associated current
sheets might play a crucial indirect role by generating shocks.
Figure ~\ref{fig_heat_source2}a) shows the regions where
magnetic energy is dissipated into heat via Ohmic dissipation. These
regions clearly take the form of filamentary structures
with a small azimuthal extent (compared to $L_y$). There appears to be
a correlation between the number of such sheets and the number of shocks for which $W_{PV}$ is positive. In addition, the location of these structures seems to be found in regions of high pressure, where $\nabla\cdot
\mathbf{u}$ is actually a minimum. These regions correspond
to
the self-gravity wakes that take the form of large scale non-axisymmetric bands. \\
Magnetic reconnection in current sheets is known to accelerate the
gas, producing sometimes a pair of slow-mode shocks extending outwards
from the central sheet
\citep{petscheck64,priest86,birn07,hillier16}. These shocks are known
to be highly effective at
heating the surrounding medium. In fact, some numerical studies
indicate that the slow mode shocks are the
primary heating mechanism in the solar corona \citep{bareford15}. In some
circumstances the energy released from these shocks can be more
important than Ohmic dissipation, as seems to be the case here.
Though we do not go into a detailed analysis in this paper, it seems
plausible that the enhanced heating witnessed by magnetic
gravitoturbulence is caused by Ohmic reconnection in current sheets and in
the shocks generated by such reconnections.
\subsection{Plasmoids}
\label{plasmoid}
As pointed out in section \ref{plasmoid_regime}, when magnetic
fields have a moderate amplitude and Rm is not too small, the
turbulent flow displays coherent plasmoid structures and magnetic
islands. These patterns have been studied exhaustively in the
literature of magnetic reconnection
\citep{park84,ugai95,loureiro05,huang13,loureiro16} and 2D compressible
MHD turbulence \citep{lee03} but have not featured especially
in simulations of accretion disc dynamics. In this context they
deserve further attention as it is tempting to associate them with
planet formation, possibly as sites in which dust may accumulate.
Separately, an examination of their intrinsic balances and structure
may help unveil the role of the Lorentz force
in magnetized shear flows generally.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{FIG_plasmoid_equilibrium.pdf}
\caption{Top panel: evolution of the total mass in a plasmoid
computed from a simulation with $B_{y_0}=0.05$ and no explicit
resistivity. Bottom panel: radial forces integrated over the
interior of
the plasmoid in a frame of reference centered at the pressure
maximum.
The typical radius of the plasmoid (taken as our integral bound for averaging) is $r_p=0.4 H$.}
\label{fig_plasmo_eq}
\end{figure}
\subsubsection{Are they fragments?}
The first question is what relationship these magnetic islands have
with respect to the gravitationally bound fragments that appear in
hydro simulations of GI. Figure~\ref{fig_plasmo_eq} (top)
shows that, for $B_{y_0}=0.05$
and $\tau_c=20/\Omega$, the total mass integrated inside one of the
plasmoids does not increase with time and keeps a fixed value during
more than 30 $\Omega^{-1}$. Actually, we checked visually that they
stay stable over a much longer time. Although they form dense
structures with $\Sigma$ that can exceed 50 times the
background $\Sigma_0$, they do not seem to be regions where the gas
is collapsing, at least for sufficiently large $\tau_c$. We conclude
that they are
resolved quasi-steady objects quite different to the
fragments seen in hydrodynamical simulations of gravitational collapse.
\subsubsection{Origins}
\begin{figure*}
\centering
\includegraphics[width=0.6\textwidth]{FIG_reconnection.pdf}
\caption{Three snapshots showing the evolution of the magnetic field line topology in a small
patch centred around a region of plasmoid formation over a period of
$3\,\Omega^{-1}$.
Blue represents negative field polarity, while represents positive
$B_y$. The intensity of the colour quantifies the magnitude of the
field. The $x$-extent of the patch is $3H_0$, while the $y$-extent
is $5H_0$.}
\label{fig_reconection}
\end{figure*}
In order to understand how plasmoids form, we investigated the early stages
of a simulation in which magnetic islands appear. We found that this stage
occurs just after the onset of the turbulence. Figure
\ref{fig_reconection} shows the field line configuration near a
plasmoid forming region between $t=23\,\Omega^{-1}$ and
$t=26\,\Omega^{-1}$ for $B_{y_0}=0.1$. Initially straight azimuthal
field lines are stretched, folded and amplified by the turbulent
eddies. Strong positive (red) and negative (blue) toroidal magnetic
fields are then brought together. At $t=25\,\Omega^{-1}$, a current
sheet is forming as soon as the magnetic loop is closed. At
$t=26\,\Omega^{-1}$, the field lines become
possibly unstable to the tearing instability
\citep{biskamp86,loureiro05} and reconnect,
forming two magnetic islands. These snapshots (and many others like
them) suggest that plasmoids
are generated through a common physical mechanism and are not produced
artificially by the code.
\subsubsection{Equilibrium and structure}
\label{plasmoid_structure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{FIG_magnetic_pressure.pdf}
\caption{Top panel: total magnetic pressure in a quasi-steady
turbulent state obtained for $B_{y_0}=0.05$ and no explicit
resistivity. The bottom left and right panels are respectively the
density and magnetic pressure in the white rectangle appearing in
the top panel.}
\label{fig_plasmo_magp}
\end{figure}
Figure \ref{fig_plasmo_magp} shows a snapshot of a simulation computed
for $B_{y_0}=0.05$ with no explicit resistivity, containing a number of
plasmoids. Magnetic pressure forms strong ring-shape structures
surrounding each of these plasmoids that prevent the external gas from
penetrating within. To a first approximation
they behave as steady rigid bodies in a
sheared turbulent background. Some of them are spinning with a net
negative vorticity (in the same direction as the shear) but their
velocity profile can be quite intricate inside.
Gas pressure and
density are always maximum at their centres and decay
quasi-exponentially with distance to this axis. Plasmoids are cold
and their temperature is minimum at their centres. In order to
better understand their internal structure, we
computed their force balance. We chose one of our
simulation with $B_{y_0}=0.05$ and tracked a number of plasmoids by
calculating their position and velocity at each output time. Forces
are then computed in a particular frame of reference, centred upon a
given plasmoid (where the density is maximum). We used polar
coordinates around this origin so that $r$ denotes the distance to the
center of a plasmoid and $\theta$ the angle with the $x$ axis. Each
force is averaged inside the plasmoid by integrating in $r$ and
$\theta$. We defined the radial extent of a plasmoid as the radius
at which the density has dropped by a factor 2 from its center.
We identified one big plasmoid where the density contrast between the
background turbulent flow and the center is
$\Delta=0.13$. Fig.~\ref{fig_plasmo_eq} shows the radial force
balance in the frame of this structure. Inside the plasmoid, the fluid
is in equilibrium between the pressure gradient (which is positive and
resisting the collapse) and all other forces (that are negative
and tend to make the gas collapse). The latter are all
of the same order of magnitude although magnetic pressure is roughly
half the magnetic tension and self-gravity. Note that inertial forces
(Coriolis and nonlinear advection) affect the
equilibrium only weakly, indicating that the pressure maxima is not
maintained by vortical motions. The plasmoids hence should not be
regarded as vortices. (We did not plot the viscous force
as it is completely negligible.) Finally, we checked that this force
balance is similar for several other plasmoids.
Figure \ref{fig_multi_averages2} indicates that the plasmoids sizes
increases with $B_{y_0}$. On one hand, this behaviour might be
surprising as we explained that magnetic forces seem to push the gas
inward and force the structure to contract. But on the other hand, for
the same reason as explained in Section \ref{fragmentation_B}, the
pressure increases very rapidly with $B_{y_0}$, due to the heat
generated by magnetic fields in this regime. In addition, the
gravitational force becomes less important in comparison with pressure
forces as $Q$ is increased. Therefore, a balance is still possible and
the size of the structures can even grow with $B_{y_0}$.
\section{Discussion and Conclusion}
In summary, we performed 2D shearing sheet simulations of
gravitoturbulence in magnetised accretion disks
penetrated by a net toroidal field. For moderate plasma beta and
magnetic Reynolds number, this field was twisted, warped, and greatly
amplified by the turbulent velocity fluctuations.
Once a quasi-steady
state was achieved the final magnetic energy could, in fact, be equal to the
turbulent kinetic energy. Once thermalised,
this additional reservoir of energy leads to a dramatic
heating of the gas, and enhanced quasi-equilibrium temperatures (and thus
Toomre $Q$'s). For example, when $\beta_0\sim 100$ and Rm$>1000$, the
mean $Q$
is amplified over the hydrodynamic value by a factor 4. For the same
$\beta_0$ but a larger Ohmic resistivity, Rm$=250$, the amplification
is a factor 2. The system can thus achieve a marginal gravitoturbulent state
in which $Q\sim 20$, and the gravitational potential energy
subdominant (though absolutely necessary for the subsistence of the steady state). For lower Rm or weaker imposed
fields these striking effects subside and the system begins to resemble the
hydrodynamical regime. We tentatively attribute the persistence of GI
activity
at such high $Q$ to the breaking of angular momentum conservation by
the tangled magnetic field. The suppression of this stabilising effect
exacerbates the GI and extends the range of gravitoturbulent activity
to hot states where it would ordinarily be stable.
The thermalisation of the magnetic energy is undertaken through the
action of small-scale current sheets, and especially the slow shocks
generated by reconnection in the sheets. The resulting heating is
highly inhomogeneous and localised in an intricate network of shock
layers. The temperature fluctuations in this network may greatly
exceed the mean temperature of the disk, and may have some
consequences for chemistry and the processing of solids (see for
example, \citet{godard09} or \cite{mcnally14}). For sufficiently
large Rm, reconnection also
generates plasmoids, long-lived magnetic islands distinct
from both vortices and gravitationally collapsing blobs. If
shown to be prevalent and robust, these structures could be of
interest to planet formation theories.
Finally, we checked to see if magnetic fields had any impact on the
fragmentation criterion. By varying the cooling time, for different
imposed fields, we obtain critical $\tau_c$ below which the gas
fragments. In general, these critical values are not very different to
the hydrodynamical ones. Given the numerically dependent and
stochastic nature of fragmentation, it is difficult to set much store
on these results --- though the basic idea (that magnetic fields are
not so important) may be robust.
Our 2D ideal and resistive simulations are potentially relevant for
protostellar disk regions that are magnetically active but MRI stable.
As shown in vertically stratified simulations with the full
panoply of non-ideal MHD, such regions may span a significant range of
outer radii. Strong horizontal fields may be generated by the Hall effect,
and winds launched at the disk surfaces. Our numerical set-up does not
correctly capture these non-ideal effects, but nonetheless some of the
behaviour we witnessed might cross over. An additional uncertainty, in
any case, is the correct non-ideal regime for the outer radii of
gravitationally unstable class-0 disks. Previous work, and our
estimates, have been based
on the less massive MMSN model.
Our simulations may also be relevant for massive deadzones in older
disks, at the onset of GI-instigated outbursts. The newly GI active
region, if supplied by sufficiently strong magnetic fields
(by advection from larger radii or
locally by Hall currents) could more effectively heat the gas, as
described above and thus more easily kickstart the classical MRI, as
required by certain outburst models. That said, this enhanced heating
requires somewhat larger Rm than typically supported by dead zones, and may only
be effective at the outer edge of the zone.
A final application of these results may be to the outer part of AGN discs
which are likely to be gravitationally unstable \citep{paczynsky78},
and susceptible to fragmentation (Goodman 2003, Levin
2007). Gravitational collapse of the disk may be especially
important in star formation bursts close to the
Galactic centre. Meanwhile, the AGN gas can be relatively well ionised and
able to couple to any latent magnetic field; indeed the MRI and GI
may overlap at certain radii in especially luminous systems
(Menou \& Quataert 2001).
These exploratory 2D results point to a number of future research directions.
For a start, Ohmic diffusion could be replaced by ambipolar diffusion
to test how magnetic fields
behave in the regimes more relevant for the outer radii of
protostellar disks. However, the most interesting avenues involve 3D
vertically stratified boxes, which could include the $z$-dependent
diffusivities and the various interesting non-ideal
MHD behaviours recently discovered \citep{lesur14,bai14, simon15}. The latter would then provide magnetic fields
self-consistently. Such simulations would let us probe how the
gravitoturbulence works in the presence of MHD winds, surface
turbulence, and its action on the magnetic field. And though
numerically intensive, they would also
provide a way to simulate both the MRI and GI together and determine
if the two instabilities coexist or attempt to switch each other off.
\section*{Acknowledgements}
The authors would like to thank the anonymous reviewer for a helpful
set of comments. They are also indebted to Sijme-Jan Paardekooper and
Charles Gammie for generously reading through an earlier draft and
offering advice and criticism.
This research is partially funded by STFC
grant ST/L000636/1.
Most of the simulations were run on the
DiRAC Complexity system, operated by the University of Leicester
IT Services, which forms part of the STFC DiRAC HPC Facility
(www.dirac.ac.uk). This equipment is funded by BIS National E-
Infrastructure capital grant ST/K000373/1 and STFC DiRAC
Operations grant ST/K0003259/1. DiRAC is part of the UK National
E-Infrastructure.
\bibliographystyle{mnras}
|
\section{Introduction}\label{IntroductionSection}
A new era in radio astronomy is approaching with the upcoming continuum surveys \citep{Norris2013} planned at the SKA precursors telescopes, such as the
\textit{Westerbork Observations of the Deep APERTIF Northern-Sky} (WODAN) \citep{Rottgering2010}
at the Westerbork Synthesis Radio Telescope (WSRT), the \textit{Evolutionary Map of the Universe} (EMU) survey \citep{Norris2011} at the ASKAP array and
the \textit{MeerKAT International GigaHertz Tiered Extragalactic Exploration} (MIGHTEE) survey \citep{Van der Heyden2010} at the Meerkat observatory. A considerable improvement
is expected in sensitivity, resolution and instantaneous field of view compared to previous surveys. For instance, WODAN and EMU will jointly provide full sky
coverage at 1.3 GHz with an unprecedented sensitivity down to 10-15 $\mu$Jy/beam and resolution around 10-15 arcsec. Phased Array Feed (PAF) technology will allow
instantaneous field of view of 8 and 30 deg$^{2}$ for WODAN-APERTIF and ASKAP respectively and a corresponding increase in survey speed of a factor $\sim$20 with respect
to VLA. MIGHTEE will allow even better sensitivities (0.1-1 $\mu$Jy/beam rms) although with a reduced field of view (1 deg$^{2}$). A dramatic gain in sensitivity (a factor 100) and
field of view will be achieved with the future operations of the SKA.
New challenges are expected to be brought by these significant advances. One is related to the data product throughput (e.g. spectral-imaging data cubes) expected to be generated
by the SKA precursor telescopes, ranging from tens of gigabytes to several petabytes\footnote{ASKAP is expected to generate several petabytes per year of HI cube.}, and by the future SKA observatory, of the order of hundreds of terabytes per data
cube in SKA1 and one order of magnitude higher in SKA Phase II \citep{Kitaeff2015}. For instance, up to 3 exabytes of fully processed data are expected in one
year of full SKA1 operation \citep{Alexander2009}. Such amount of data cannot be processed nor stored and visualized on local computing resources, at least using conventional
data formats so far used in astronomy.
Furthermore, with the increase in sensitivity and surveyed sky area, a population of millions of sources will be potentially detectable making human-driven source extraction
unfeasible. For example, the EMU survey is expected to generate a catalogue of $\sim$70 millions of sources detected at the 5$\sigma$ level of 50 $\mu$Jy/beam \citep{Norris2011}.
For these reasons considerable efforts are currently focused on the development of algorithms to process imaging data and extract sources in a fast and mostly automated
way and, at the same time, on the search for new data standards and image compression formats (e.g. see \citealt{Kitaeff2014}).
While extensive studies have been performed on compact source search with several algorithms developed \citep{Hancock2012,Whiting2009,Whiting2012,Hopkins2002,Bertin1996,Hales2012,
Peracaula2015,Hopkins2015}, particularly in
the context of the ASKAP telescope, detection of extended sources in a completely unsupervised way (e.g. without requiring any a priori information or source
templates) is still a partially explored field, at least for the radio domain. This motivates investing resources on exploring completely new methods or re-adapting known
algorithms to the radio imaging case.
Different approaches have been recently proposed in such direction. Some of them make use of conventional thresholding methods in the image wavelet or curvelet domain
(e.g. see \citealt{Peracaula2011}), others employ compressive sampling techniques (e.g. \citealt{Dabbech2015}). Other studies employ the Circle Hough transform to detect circular-like objects, such as supernova remnants or bent-tail
radio galaxies \citep{Hollitt2012}. In \citet{Norris2011} several methods from the Computer Vision domain have been reviewed. Waterfalling segmentation, circular or elliptical
Hough transform and region growing were indicated as the most suited to the problem of extended source search.
In the context of the SCORPIO project \citep{Umana2015} (hereafter denoted as "Paper I", see Section~\ref{SCORPIOProjectSection}), a pathfinder of the ASKAP EMU survey, and in view of
the next-generation SKA surveys, we started to develop algorithms for automated source detection and classification. The designed method exploits some of the techniques and algorithms already
in use in other source finders, aiming to combine their best features, but also introduces new features, particularly on the background estimation, detection of
extended sources and source parameterization. We will therefore focus on these novel aspects throughout the paper. A description of the method, based on a superpixel segmentation
and hierarchical merging, is presented in Section~\ref{MethodSection}. The algorithm has been tested on SCORPIO real radio data observed at the ATCA array down to a
sensitivity of 30 $\mu$Jy/beam. Typical results achieved on sample field scenarios are presented and discussed in Section~\ref{ResultSection}, along with tests performed
on the same fields observed at different wavelengths.
\begin{figure*}
\centering%
\includegraphics[scale=0.55]{SourceFinderPipeline_last.pdf}
\caption{Schematic pipeline of the designed source finder algorithm.}
\label{SourceFinderPipelineFig}
\end{figure*}
\section{The SCORPIO project}\label{SCORPIOProjectSection}
The SCORPIO project is a blind deep radio survey of a $2\times2$ deg$^2$ sky patch toward the Galactic plane, using the ATCA array in several
configurations. The survey has been conducted at 2.1 GHz between 2011 and 2015 and achieved an average resolution around 10 arcsec. Further observations are already scheduled in 2016.
The major scientific goals of the SCORPIO project is to search for different populations of Galactic radio point sources and the study of circumstellar envelopes
(related to young or evolved massive stars, planetary nebulae and supernova remnants) which is extremely important for understanding the Galaxy evolution (e.g. ISM
chemical enrichment, star formation triggering, etc.). Besides these scientific outputs, SCORPIO will be used
as a test-bed for the EMU survey, guiding its design strategy for the Galactic plane sections. In particular, this includes exploring suitable strategies for effectively imaging and
extracting sources embedded in the diffuse emission expected at low Galactic latitudes and investigating to what extent they can be employed in the EMU survey.
The SCORPIO observations have produced a radio mosaic map of 133 single pointings with an \textrm{rms}{} down to 30 $\mu$Jy/beam. A pixel size of 1.5" is chosen for the final map.
This sensitivity and a good $uv$-plane coverage have allowed the discovery of about 1000 new faint radio point sources and to satisfactorily map tens of extended sources.
Preliminary results on a smaller pilot region of the SCORPIO field have already been published in Paper I, while the complete data reduction and analysis is
still in progress.
\section{A segmentation method for extended source detection}\label{MethodSection}
Detection of extended sources represents a hard task for source finder algorithms. The main difficulties are due to the intrinsic emission pattern, which is usually fainter compared
to compact sources (e.g. below the conventional 5$\sigma$ significance level) and spread over disjointed areas (e.g. unlike the adjacency assumption taken in compact source finders).
In addition, object borders are usually soft thus the standard edge detector algorithms are not fully sensitive to them.
Spatial filters are therefore often employed to enhance the emission at some given scale.
Another issue is related to the estimation of reliable significance levels for detection. In fact, the widely used method for local noise and background estimation is typically
biased around extended source regions, namely higher significance levels are artificially imposed for detection with respect to other image regions, free of diffuse emission. Under these
conditions the source is likely to be undetected particularly if it has a large extension.
Ideally, the source extraction task should provide a two-level hierarchical information: a segmentation of the input map into background and foreground regions associated with
a source object, and, for each of the them, a collection of nested regions representing source features (e.g. clumps, shells, blobs) also at
different scales.
To this goal, we designed a multi-stage method based on image superpixel generation and hierarchical clustering.
A schematic pipeline of the algorithm stages is shown in Fig.~\ref{SourceFinderPipelineFig} and summarized below:
\begin{enumerate}
\item \emph{Filtering}: To enhance extended structures, bright compact sources need to be filtered out from the map and a residual image generated and used as
input for the following stages. Compact source extraction, discussed with more details in Section~\ref{CompactSourceFindingSect}, requires the computation of the background
and noise maps to threshold the image at a suitable significance level.
Furthermore, a smoothing stage is introduced on the residual image to suppress texture-like features due to imaging artefacts around the brightest sources and to source residuals left
after the previous dilation stage. An edge-preserving guided filter \citep{He2013} was found to provide optimal performances among the tested filters.
\item \emph{Extended source extraction}: The smoothed residual image is used as input for the segmentation algorithm described in Section \ref{SegmentationAlgorithmSection}.
It consists of three main stages: firstly, an over-segmentation of the image into a collection of superpixels or regions is generated and a set of appearance parameters (both intensity- and
spatial-based) computed for each region; then, a saliency map is computed in the second stage from region dissimilarities and used to drive region merging at the third stage,
which is a sequence of clustering steps producing a collection of segmented regions or a binary mask as the final output.
\item \emph{Source parametrization}: A set of morphological parameters is calculated over the segmented regions and delivered to the user.
\end{enumerate}
Additional details concerning each algorithm step are given in the following sections.
\begin{figure*}
\subfloat[Field A\label{ScorpioFieldFig1}]{\includegraphics[scale=0.28]{ScorpioS17Field_bw_finalPaperVersion_light.pdf}}
\hspace{0.2cm}
\subfloat[Field B\label{ScorpioFieldFig2}]{\includegraphics[scale=0.28]{ScorpioSNRFieldZoom_bw_finalPaperVersion_light.pdf}}\\
\vspace{-0.4cm}
\subfloat[Field C\label{ScorpioFieldFig3}]{\includegraphics[scale=0.28]{ScorpioExtendedField_bw_finalPaperVersion_light.pdf}}
\hspace{0.2cm}
\subfloat[Field D\label{ScorpioFieldFig4}]{\includegraphics[scale=0.28]{ScorpioFaintSNRField_bw_finalPaperVersion_light.pdf}}
\caption{Sample SCORPIO fields (A-D) selected for algorithm testing. Flux units are reported in the z axis.}
\label{ScorpioFieldFig}
\end{figure*}
\subsection{Background and noise estimation}\label{BkgFindingSect}
As noted in Paper I, both background and noise levels are subjected to variations throughout the image, due for example to diffuse emission around
the Galactic plane or to the accuracy of the image reconstruction. Background and noise information are therefore estimated on a local basis using two alternative methods.
The first conventional method assumes a rectangular grid of sample pixels and computes the local background and noise levels over a sampling box, centered around each grid center.
Robust background/noise estimators are generally considered to reduce the bias caused by the possible presence of sources falling in the sampling box.
For instance \textit{Selavy} \citep{Whiting2009,Whiting2012} uses the median and mean absolute deviation from the median (MAD), while the inter-quartile range is adopted in
\textit{Aegean} \citep{Hancock2012}. Other methods use the previous estimators iteratively clipped up to reach a pre-specified tolerance, as
in \textit{SExtractor} \citep{Bertin1996} or in Paper I. Several estimators are available in our program: median/MAD, biweight or $\sigma$-clipped estimators.
Finally, a bicubic interpolation stage is carried out to derive local estimates on a pixel-by-pixel basis, e.g. the background and noise maps.
The second method exploits the pixel spatial information, neglected by the conventional approach, along with the pixel intensity distribution to produce less biased noise/background
estimates. Two different approaches were implemented. In the first, a superpixel partition of the image is generated (see Section~\ref{SegmentationAlgorithmSection} for more details)
with region size assumed comparable to the synthesised beam size. An outlier analysis, based on a robust estimate of the Mahalanobis distance \citep{Rousseeuw1990} on region median-MAD parameter space, is then performed
to detect significative regions (both positive or negative excesses), typically associated with sources or artefacts. Pixels belonging to that regions are marked and
excluded from the background evaluation. The background and noise maps are finally computed as above by interpolating a robust estimator computed over background-tagged
pixels in sampling boxes sliding through the entire image.
A second approach uses a flood-filling algorithm to detect and iteratively clip blobs at some predefined significance level (e.g. 5$\sigma$) with respect to the first
level estimate of the background and noise maps. Background and noise maps are re-computed at each iteration stage as described above. One or two iterations are typically sufficient.
In practice, the first method can be safely used for bright compact source filtering, in which the
background estimation is not requested to be highly accurate. The second method should be instead preferred in the search of
faint compact sources or when thresholding extended bright sources.
The size of the sampling grid is conventionally chosen to achieve sufficient interpolation accuracy at moderate computational cost. Instead, the choice of the box size is often given
in terms of the beam size (e.g. 10 or 20 larger than the synthesised beam) and may have a considerable impact in the source extraction step: estimates computed on a small
box could be severely biased by the presence of a source filling the box, while, on the other hand, a too large box could completely smooth out the local
background/noise variations. In \citet{Huynh2012} the authors compared maps
obtained by popular source finders, such as \textit{SFind} \citep{Hopkins2002}, \textit{SExtractor} \citep{Bertin1996} and \textit{Selavy} \citep{Whiting2009,Whiting2012}, and
investigated the optimal parameter settings both for real and simulated data sets. However, they note that a completely automated procedure for background estimation, possibly
independent on the distribution of sources, is still of crucial importance for future surveys.
\begin{figure*}
\subfloat[][Field A]{\includegraphics[scale=0.28]{MolongloS17Field_resampled_bw_finalPaperVersion_light.pdf}\label{MolongloScorpioFieldFig1}}
\hspace{0.1cm}
\subfloat[][Field B]{\includegraphics[scale=0.28]{MolongloSNRField_resampled_bw_finalPaperVersion_light.pdf}\label{MolongloScorpioFieldFig2}}\\
\vspace{-0.4cm}
\subfloat[][Field C]{\includegraphics[scale=0.28]{MolongloExtendedField_resampled_bw_finalPaperVersion_light.pdf}\label{MolongloScorpioFieldFig3}}
\hspace{0.1cm}
\subfloat[][Field D]{\includegraphics[scale=0.28]{MolongloFaintSNRField_resampled_bw_finalPaperVersion_light.pdf}\label{MolongloScorpioFieldFig4}}
\caption{Sample fields (A-D) selected for algorithm testing as observed in the Molonglo Galactic Plane Survey. Flux units are reported in the z axis.}
\label{MolongloScorpioFieldFig}
\end{figure*}
\subsection{Filtering compact sources}\label{CompactSourceFindingSect}
The presence of bright sources in the image significantly hardens the extended source detection task. We therefore implemented a filtering stage to remove them, based on the
following steps. Blobs of connected pixels are first extracted from the image assuming a flood-filling procedure similar to that carried out in \textit{Aegean} \citep{Hancock2012} and
\textit{Blobcat} \citep{Hales2012} source finders. A high seed threshold above the computed background is assumed, e.g. 10$\sigma$, and pixels are aggregated down to a merge
threshold, e.g. 2.6$\sigma$. Each detected blob is subjected to a further search to identify nested blobs. These are extracted by thresholding the image
curvature map $\kappa$, obtained by convolving the image with a Logarithm-of-gaussian (LoG) kernel, at some pre-specified threshold level (e.g. $\kappa >$0) or adaptively.
A 2-level hierarchy of blobs is finally obtained.
A set of morphological parameters (e.g. contour parameters, moments, shape descriptors, etc), is computed over the detected blobs and selection cuts are applied to
identify point-like candidate sources. For example, blobs with a number of pixels that is too large or with an anomalous elongated shape typically fail to pass the point-like cut.
Blobs tagged as "point-like" are removed from the input image using a morphological dilation operator with configurable kernel shape (e.g. elliptic or squared) and size, as suggested
in \citet{Peracaula2015}, and replaced with a random background realization. A kernel size larger than 5 pixels was assumed to prevent the source halo pixels to further
affect the residual image.
\begin{figure*}
\subfloat[Field A]{\includegraphics[scale=0.22]{ScorpioS17SegmResults_saliency_large_light.pdf}}
\hspace{-0.cm}
\subfloat[Field A]{\includegraphics[scale=0.22]{ScorpioS17SegmResults_large_light.pdf}}\\
\vspace{-0.4cm}
\subfloat[Field B]{\includegraphics[scale=0.22]{ScorpioSNRSegmResults_saliency_large_light.pdf}}
\hspace{-0.cm}
\subfloat[Field B]{\includegraphics[scale=0.22]{ScorpioSNRSegmResults_large_light.pdf}}\\
\vspace{-0.4cm}
\subfloat[Field C]{\includegraphics[scale=0.22]{ScorpioExtendedFieldSegmResults_saliency_large_light.pdf}}
\hspace{-0.cm}
\subfloat[Field C]{\includegraphics[scale=0.22]{ScorpioExtendedFieldSegmResults_large_light.pdf}}\\
\vspace{-0.4cm}
\subfloat[Field D]{\includegraphics[scale=0.22]{ScorpioFaintSNRSegmResults_saliency_large_light.pdf}}
\hspace{-0.cm}
\subfloat[Field D]{\includegraphics[scale=0.22]{ScorpioFaintSNRSegmResults_large_light.pdf}}
\caption{Segmentation results obtained for the test fields A-D (from top to bottom) assuming $l$=20 and $\beta$=1 (see text).
Left: Saliency maps normalized to range [0,1]; Right: Segmentation maps. Each segmented region is colored in the plot according to the mean of its pixel fluxes in mJy/beam units.
The white contour lines correspond to a manual segmentation generated by an expert astronomer.}
\label{ScorpioSegmResultsFig}
\end{figure*}
\subsection{Segmentation algorithm}\label{SegmentationAlgorithmSection}
We developed a segmentation algorithm for extraction of extended sources, based on a superpixel segmentation algorithm followed by a
hierarchical clustering stage to aggregate similar segments into final candidate source regions. The algorithm steps are described below and a summary of the relevant algorithm
parameters is reported in Table~\ref{AlgorithmParTable}:
\begin{enumerate}
\item \emph{Initialization}: Compute a set of filtered images to be used during the clustering stage, namely the image curvature $\kappa$ and an edge-sensitive map $\psi$. The latter
can be alternatively obtained by convoluting the input image with a set of Kirsch filters oriented along different directions or as the result of the Chan-Vese contour finding algorithm
\citep{ChanVese2001}.
\begin{figure*}
\subfloat[Field E]{\includegraphics[scale=0.28]{ScorpioSparseField3_bw_finalPaperVersion_light.pdf}}
\hspace{-0.cm}
\subfloat[Field E - Residual]{\includegraphics[scale=0.28]{ScorpioSparseField3ResidualMap_light.pdf}}
\hspace{-0.cm}
\caption{Left: Sample SCORPIO field E selected for algorithm testing. Flux units are reported in the z axis; Right: Residual map, normalized to range [0,1], obtained after
applying point-like source and smoothing filtering stages to the input map.}
\label{SparseFieldResultsFig}
\end{figure*}
\item \emph{Superpixel segmentation}: In this stage the image is over-segmented into $N_{\texttt{R}}$ connected regions or superpixels using flux and spatial information as input
observables. To this aim we made use of the \textit{Simple linear iterative clustering (SLIC)} algorithm developed by \citet{Achanta2012}, which uses the k-mean algorithm to
cluster pixels according to an intensity and spatial proximity measure. Segmentation is controlled by a set of input parameters, such as the desired
superpixel size $l$, typically fixed to the smallest detail to be distinguished (e.g. close to the beam size to detect compact sources or larger to search for
extended sources), the minimum number of pixels in a region ($N_{\texttt{min}}$) and a regularization parameter $\beta$ balancing spatial and intensity clustering in the
distance measure $D_{ij}$ between a pixel $i$ and a superpixel center $j$:
\begin{equation}
D_{ij} = \sqrt{ D_{ij,c}^2 + \left(\frac{\beta}{l\times l}\right)^2 D_{ij,s}^2}
\end{equation}
$D_{ij,c}$ and $D_{ij,s}$ being the intensity and spatial Euclidean distances between pixel $i$ and superpixel $j$. Higher $\beta$ enhances the spatial proximity and favors more compact
superpixels in the initial partition. In turn, lower $\beta$ favors clustering in intensity and superpixels with less regular shapes but adhering more tightly
to the object contours.
For each region $i$ an appearance parameter vector $\mathbf{x}_{i}=(\mu_{i},\sigma_{i},\mu_{i,\kappa},\sigma_{i,\kappa})$
is computed, with $\mu_{i}$ and $\mu_{i,\kappa}$ denoting
respectively the mean of flux and curvature of pixels belonging to region $i$, while $\sigma_{i}$ and $\sigma_{i,\kappa}$ are their standard deviations.
With this parameter choice, the computation and update of the region parameters after a merging can be done iteratively in a very fast way, namely without partially sorting
the region pixel vector as in the case of median and MAD estimators.
\item \emph{Saliency map estimation}: A saliency map is estimated in this step to enhance significant objects in the input image with respect to the background. Following
\citet{Zhang2013}, a saliency estimator $S_i$ is computed for each region as:
\begin{equation}
S_i= 1-\exp\left(-\frac{1}{K}\sum_{j=1}^{K}\delta_{ij}\right)\;\;\;\;\;\;\;\;\;\;\;\delta_{ij}=\frac{d_{ij,c}}{1+d_{ij,s}}
\end{equation}
where $d_{ij,c}$ is the Euclidean distance between appearance vectors $\mathbf{x}_{i}$ and $\mathbf{x}_{j}$ of region $i$ and $j$, $d_{ij,s}$ the distance between their centroids. The sum is
computed over the $K$ nearest neighbors of region $i$, typically 10\% or 20\% out of the total number of regions. Salient objects are likely to have similar pixels more confined
in space compared to similar pixels belonging to the background which
are more spatially spread in the image. To detect salient features at different scales, we combined saliency maps computed at different resolutions, e.g. corresponding to
initial partitions with different superpixel sizes. Finally, multi-resolution saliency maps are combined with the computed local noise and background maps, which
are found to be also sensitive to the diffuse emission. A saliency map with almost full pixel resolution is finally determined.
\item \emph{Superpixel tagging}: Each pixel $i$ is tagged as background/object/untagged candidate if its saliency $S_i$ is within some adaptive threshold levels:
\begin{equation}
S_i= \left\{
\begin{array}{ll}
\texttt{background} & S_i<S_{\texttt{thr}}^{\texttt{bkg}}
\vspace{0.2cm}
\\
\texttt{object} & S_i>S_{\texttt{thr}}^{\texttt{sig}}
\vspace{0.2cm}
\\
\texttt{untagged} & \text{otherwise}
\end{array}
\right.
\end{equation}
Different saliency thresholding approaches are possible. One of the most used in saliency studies \citep{Achanta2009,Perazzi2012,Kim2014,Zhang2013} assumes a global adaptive
threshold of the kind $S_{\texttt{thr}}^{\texttt{bkg,sig}}=f_{\texttt{thr}}^{\texttt{bkg,sig}}\times\langle S\rangle$, where $\langle S\rangle$ is the average (or median) saliency of
the map and $f$ is a numerical factor (e.g. $f$=1 for the background and $f$=2 for the signal; \citet{Achanta2009,Zhang2013}).
After several tests performed on different maps we obtained optimal results by combining different global threshold measures:
\begin{equation}
S_{\texttt{thr}}^{\texttt{sig}}= \texttt{max}\{f_{\texttt{thr}}^{\texttt{sig}}\times\langle S\rangle,\texttt{min}\{S_{\texttt{thr}}^{\texttt{Otsu}},S_{\texttt{thr}}^{\texttt{valley}} \}\}
\end{equation}
where $S_{\texttt{thr}}^{\texttt{Otsu}}$ is the threshold level computed through the Otsu method (e.g. see \citealt{Sezgin2004} for a review of thresholding methods)
and $S_{\texttt{thr}}^{\texttt{valley}}$ is the threshold corresponding to the first valley detected in the pixel saliency histogram.
The threshold level factor $f_{\texttt{thr}}^{\texttt{sig}}$ is chosen as a trade-off between false detection rate and object detection efficiency.
The alternative approach, more computationally expensive, is employing the local adaptive thresholding method used also for compact source extraction with
or without outlier rejection.
Superpixels are finally tagged as background, object or untagged candidates according to the majority of their pixel tags.
\item \emph{Superpixel graph}: Identify 1st- and 2nd-order neighbors to each region $i$=1,\dots,$N_{\texttt{R}}$ and build a corresponding link graph as described in \citet{Bonev2014}.
By 1st-order neighbors, we denote the regions surrounding and sharing a border with region $i$. For each region link $i-j$ in the graph, compute an edgeness $E_{ij}$ parameter
related to the amount of edge present on the shared border between region $i$ and $j$. For 1st-order neighbors, this is estimated by taking the average of $\psi$ over the pixels
located on the shared boundary, while for 2nd-order neighbors, it assumes the largest value present in the $\psi$ map.
Let us consider an asymmetric dissimilarity measure $\Delta_{ij}$ between neighbor regions $i$ and $j$ given by:
\begin{equation}
\Delta_{ij}= (1-\lambda)d(\mathbf{x}_{i},\mathbf{x}_{i\cup j}) + \lambda E_{ij}
\end{equation}
where $d(\cdot,\cdot)$ is the Euclidean distance between feature vectors, $E_{ij}$ the edgeness parameter and $\lambda$ a regularization parameters balancing distance and edgeness weights in
$\Delta_{ij}$.
The above measure expresses the change of feature vector $\mathbf{x}_{i}$ caused by a potential merging with region $j$, which is favored when the distance between feature
vectors is small and penalized when there is a border in between the two regions. Note that $\Delta_{ij}\neq\Delta_{ji}$.
Compute the adjacency matrix $\mathbf{A}$ of the graph with elements $a_{ij}$:
\begin{equation}
a_{ij}= \frac{\Delta_{ij}^{-1}}{\sum_{j}\Delta_{ij}^{-1}}
\end{equation}
properly normalized to express a transition probability from node $i$ to $j$.
\item \emph{Superpixel merging}: Following \citet{Ning2010} and \citet{Zhang2013}, merge superpixels on the basis of a maximum similarity criterion by iterating the following steps
until no more merging is possible:
\begin{enumerate}
\item Merge untagged regions to candidate background regions if their similarity is maximal among neighbor similarities.
\item Adaptively merge untagged regions if their similarity is maximal among neighbors similarities.
\end{enumerate}
Untagged regions shrink during the previous stage, while background regions grow. Signal-tagged regions are not affected in the previous stages. Superpixel parameter vector and graph
(neighbor links, dissimilarity/adjacency matrix) are updated after each iterated merging stage. When no more merging is favored, all the remaining untagged regions are labeled as
signal candidates. This stage always converges to assign all regions to either background or signal.
A suitable superpixel merging order for each of the steps described above is determined as in \citet{Bonev2014} using the Google PageRank algorithm \citep{Brin1998} on the
transition matrix $A$, that is solving the following equation:
\begin{equation}
\mathbf{p} = (1-d)\mathbf{e} + d\mathbf{A}^{T}\mathbf{p}
\end{equation}
in which $\mathbf{p}=(p_{1},p_{2},\dots,p_{N_{\texttt{R}}})$ is the desired vector with rank values (the principal eigenvector of $\mathbf{A}$), $d$ is the damping factor which
can be set to a value between 0 and 1 (e.g. $d$=0.85 as in \citet{Brin1998,Page1999}) and $\mathbf{e}$ is a column vector of all 1's. The equation is solved
by using the power iteration method \citep{Golub1983}. $\mathbf{p}$ is sorted and allows to select nodes with higher ranks for merging.
\item \emph{Source selection}: In this step sources are identified from the collection of signal candidate regions selected in the previous stage.
Following \citet{Bonev2014} the most similar signal regions are hierarchically clustered if their mutual dissimilarities ($\Delta_{ij}$, $\Delta_{ji}$)
are within a pre-specified tolerance. Only a percentage (e.g. 30\%) of top ranked merging are allowed at each clustering iteration.
A practical criterion for the merging is allowing first neighbors to always merge (e.g. a sort of flood-fill approach over superpixels) and assuming a tolerance for
2nd-order neighbors. Region parameter vectors and the dissimilarity/adjacency matrix are updated at each iteration stage and stop conditions are checked. If no
regions are merged at the current hierarchy level or the remaining number of regions is below a specified threshold the algorithm stops and the final segmentation is
returned to the user, otherwise a new iteration is started.
\item \emph{Post-processing}: Some post-processing stages can be performed on the detected sources. A first step uses the hierarchical clustering approach described above to identify
similar regions within each source and generate a list of nested sources one level down in the source hierarchy.
Further, following \citet{Yang2008}, a number of statistical and morphology-descriptor parameters are computed over the source contour and/or its pixel distribution to be
eventually employed in a source classification stage. Standard parameters include bounding box/ellipse, image/contour moments and roundness/rectangularity estimators.
More complex parameters, such as Fourier Descriptors (FDs) \citep{Zhang2003}, Hu \citep{Hu1962} and Zernike moments \citep{Singh2011}, can be computed and supplied to the user.
\end{enumerate}
\begin{table*}
\caption{Main parameters used in the source finder algorithm.}
\begin{tabular}{p{1.5cm}|p{1.5cm}|p{8cm}}
\bottomrule%
\textbf{Stage} & %
\textbf{Parameter} & %
\textbf{Description}\\%
\hline%
\multirow{2}{*}{\pbox{1.5cm}{\texttt{Background}}} &
\texttt{bkgModel} & \pbox{8cm}{\vspace{0.1cm}Model to be used for computing the background and noise maps (1=global, 2=local, 3=local robust).\vspace{0.1cm}}\\ \cline{2-3}%
& \texttt{boxSize} & \pbox{8cm}{\vspace{0.1cm}Size of the box used to compute local background/noise estimators.\vspace{0.1cm}}\\ \cline{2-3}%
& \texttt{gridSize} & \pbox{8cm}{\vspace{0.1cm}Size of the grid used when interpolating the local background/noise estimators.\vspace{0.1cm}}\\
\bottomrule
\multirow{2}{*}{\pbox{1.5cm}{\texttt{Filtering}}} & \pbox{1.5cm}{\textbf{$\sigma_{\texttt{seed}}$ $\sigma_{\texttt{merge}}$}} & \pbox{8cm}{\vspace{0.1cm}Seed and merge
threshold used to detect compact
bright blobs in the image, e.g. $\sigma_{\texttt{seed}}$=10, $\sigma_{\texttt{merge}}$=2.5.\vspace{0.1cm}}\\ \cline{2-3}%
& \textbf{$K_{\texttt{dilate}}$} & \pbox{12cm}{Kernel size to be used when dilating bright sources.}\\ \cline{2-3}%
& \textbf{$\sigma_{\texttt{smooth}}$ $K_{\texttt{smooth}}$} & \pbox{7cm}{\vspace{0.1cm}Kernel and radius parameter to be used in image residual smoothing.\vspace{0.1cm}}\\ \hline%
\multirow{2}{*}{\pbox{1.5cm}{\texttt{Superpixel}\\ \texttt{Generation}}} & \textbf{$l$} & Superpixel size used to generate the initial superpixel partition.\\ \cline{2-3}%
& \textbf{$\beta$} & \pbox{8cm}{\vspace{0.1cm}Regularization parameter controlling starting superpixel segmentation and balancing clustering spatial and color distance.
Low $\beta$ values favors spatial clustering, high $\beta$ favors color clustering\vspace{0.1cm}}\\ \hline%
\multirow{2}{*}{\pbox{1.5cm}{\texttt{Saliency}\\ \texttt{Filter}}} & \textbf{$l_{\texttt{min/max/step}}$} & Superpixel sizes to be used in multi-resolution saliency computation, e.g.
$l$=20-60, step 10.\\ \cline{2-3}%
& \textbf{$\texttt{knn}$} & \pbox{8cm}{\vspace{0.1cm}Fraction of nearest neighbors superpixel used in saliency estimation,
e.g. $\texttt{knn}$=10\%/20\%\vspace{0.1cm}}\\ \cline{2-3}%
& \textbf{$f_\texttt{sal}^\texttt{scales}$} & \pbox{8cm}{\vspace{0.1cm}Fraction of salient scales required to contribute to final saliency
estimation, e.g. $\texttt{knn}$=70\%\vspace{0.1cm}}\\ \cline{2-3}%
& \texttt{useCurvMap} & Flag to include (multi-scale) curvature maps in saliency estimation\\ \cline{2-3}%
& \texttt{useBkgMap} & Flag to include (multi-scale) background map in saliency estimation\\ \cline{2-3}%
& \texttt{useNoiseMap} & Flag to include (multi-scale) noise map in saliency estimation\\ \cline{2-3}%
& \texttt{salThrModel} & Method to be used for thresholding final saliency map (1=global, 2=local, 3=local robust)\\ \cline{2-3}%
& \textbf{$f_{\texttt{thr}}^{\texttt{bkg}}$} & \pbox{8cm}{\vspace{0.1cm}Global threshold parameter to tag background pixel candidates in
saliency map, e.g. $f_{\texttt{thr}}^{\texttt{bkg}}$=1.\vspace{0.1cm}}\\ \cline{2-3}%
& \textbf{$f_{\texttt{thr}}^{\texttt{sig}}$} & \pbox{8cm}{\vspace{0.1cm}Global threshold parameter to tag signal pixel candidates in
saliency map, e.g. $f_{\texttt{thr}}^{\texttt{bkg}}$=2.\vspace{0.1cm}}\\ \hline%
\multirow{2}{*}{\pbox{1.5cm}{
\texttt{Superpixel}\\\texttt{Merging}}} & \textbf{$\lambda$} & \pbox{8cm}{\vspace{0.1cm}Regularization parameter used in superpixel merging stage balancing appearance and edge
terms when computing superpixel dissimilarities. Low $\lambda$ values (close to zero) favors intensity similarity, high $\lambda$ values (close to 1) favors
edge penalization.\vspace{0.1cm}}\\ \cline{2-3
& \texttt{Edge Model} & \pbox{8cm}{\vspace{0.1cm}Model to be used to compute superpixel edgeness (1=Kirsch, 2=Chan-Vese).\vspace{0.1cm}}\\ \cline{2-3
& \textbf{$f_{\texttt{merge}}$} & \pbox{8cm}{\vspace{0.1cm}Fraction of top ranked superpixels selected for merging at each hierarchy level, e.g. $f_{\texttt{merge}}$=30\%.\vspace{0.1cm}}\\ \cline{2-3
& $\varepsilon_{\texttt{merge}}^{\texttt{1st,2nd}}$ & \pbox{8cm}{\vspace{0.1cm}Maximum mutual dissimilarity tolerance used for accept a selected superpixel merging for 1st or 2nd
neighbor superpixels, e.g. 5-15\%.\vspace{0.1cm}}\\ \cline{2-3
& $\Delta_{\texttt{thr}}$ & \pbox{8cm}{\vspace{0.1cm}Absolute dissimilarity threshold, when applied, to select/reject selected superpixel merging ($\Delta_{ij}\le\Delta_{thr}$).
Low $\Delta_{thr}$ values (close to zero) imply strict superpixel similarity for merging. High $\Delta_{thr}$ values relax the merging.\vspace{0.1cm}}\\ \hlin
\end{tabular}
\label{AlgorithmParTable}%
\end{table*}
\subsection{Algorithm implementation}
The described algorithms have been implemented in a C++ software library, dubbed \texttt{CAESAR}{} (\emph{Compact And Extended Source Automated Recognition}{}), allowing
image filtering, background estimation, source finding, image segmentation starting from images in FITS or ROOT format. The library is mainly based on the \textit{ROOT} \citep{Brun1997}
and \textit{R} \citep{R} frameworks for statistical objects and methods and on the \textit{OpenCV} library \citep{Bradski2000} for some of the image filtering algorithms. The source finding and
segmentation algorithms have been developed from scratch along with some of the employed filtering stages. Future developments include the algorithm fine-tuning and optimization and further design
activities for ease of deployment in a distributed computing infrastructure and integration within the pipeline frameworks of next-generation telescopes. Public distribution is planned once optimization steps are carried out.
\begin{figure*}
\subfloat[Field B]{\includegraphics[scale=0.23]{ScorpioSNRSourceDeblendResults_last_light.pdf}\label{PostProcessingResultsFig1}}
\hspace{-0.2cm}
\subfloat[Field C]{\includegraphics[scale=0.23]{ScorpioExtendedFieldDeblendResults_last_light.pdf}\label{PostProcessingResultsFig2}}
\hspace{-0.2cm}
\subfloat[Field D]{\includegraphics[scale=0.23]{ScorpioFaintSNRDeblendResults_last_light.pdf}\label{PostProcessingResultsFig3}}\\
\vspace{-0.4cm}
\subfloat{\includegraphics[scale=0.5]{ZernikeMoments_paperFinalVersion.pdf}\label{PostProcessingResultsFig4}}
\caption{Top panels: Sample segmented source images, normalized to range [0,1], in field B, C and D (solid black contours). White contours represent nested regions selected
with a multi-resolution saliency-based method (solid lines) and with a multi-scale blob detector (dashed lines); Bottom panels: Zernike moments up to order $n$=4 computed over the
segmented sources shown in the upper panels (black contoured area).}
\label{PostProcessingResultsFig}
\end{figure*}
\section{Application to SCORPIO project data}\label{ResultSection}
\subsection{Sample fields}\label{SCORPIOSampleData}
To test the designed algorithm we considered four selected fields from the SCORPIO map in which several extended structures are present
along with compact sources. The map is built as described in Paper I using data observed with the ATCA 0.75A array configuration in combination with
data observed with the ATCA EW367 configuration, in which shorter baselines are present. The effective frequency range of the radio data used is 1.4-3.1 GHz.
The sample fields, hereafter denoted as field A-D are shown in Fig.~\ref{ScorpioFieldFig}, and some details are reported below:
\begin{itemize}
\item \emph{Field A} (Fig.~\ref{ScorpioFieldFig1}): Field A (1000$\times$1000 pixels) is centered on the [DBS2003] 176 galactic stellar cluster (l=343.4830$^\circ$, b=-00.0380$^\circ$,
angular size=1.45 arcmin). Two bubble objects, S16 and S17 \citep{Churchwell2006}, are associated with the cluster but only S17 is observed in the radio domain. Two bright point-like radio
sources (SCORPIO1\_320 and SCORPIO1\_300), already known objects in radio, were identified in Paper I. SCORPIO1\_300 is located within the S17 bubble and has peak
flux around 0.04 Jy/beam. The brighter SCORPIO1\_320 (peak flux$\sim$0.14 Jy/beam) has been tentatively classified as a Massive Young Stellar Object (MYSO) candidate \citep{Urquhart2007}.
\item \emph{Field B} (Fig.~\ref{ScorpioFieldFig2}): Field B (1600$\times$1850 pixels) is centered on the Supernova Remnant (SNR) G344.7-0.1, located in the adjacency of the high
energy $\gamma$-ray source HESSJ1702-420 (see \citealt{Giacani2011}). Close to the SNR, in the north-east region of the image, another extended emission is present and most probably
associated with the MSC 345.1-0.2 supernova remnant candidate ($l$=345.062, $b$=-0.218 according to the MOST MSC survey at 843 MHz \citep{Whiteoak1996}).
\item \emph{Field C} (Fig.~\ref{ScorpioFieldFig3}): Field C (1000$\times$1000 pixels) was analyzed in detail in Paper I. Some of the extended regions of emission present were
associated with the following IRAS sources: IRAS 16566-4204, IRAS 16573-4214, IRAS 16561-4207. The first is recognized as a massive star formation region, while classification is uncertain for the others.
\item \emph{Field D} (Fig.~\ref{ScorpioFieldFig4}): Field D (1000$\times$1000 pixels) is centered on the faint SNR Candidate MSC G345.1+0.2. Below this a more intense
emission is present, associated with the G345.097+00.136 HII region.
\end{itemize}
An additional control field, free of extended sources and denoted as field E, is considered to study the algorithm response in the absence of any expected signal and tune the
detection thresholds. Field E is reported in Fig.~\ref{SparseFieldResultsFig} (left panel). This map is built using data observed with the ATCA 0.75A array configuration alone.
Due to the larger minimum baseline available extended and diffuse sources are strongly filtered out.
\begin{figure*}
\centering
\subfloat[Field A]{\includegraphics[scale=0.22]{MolongloS17FieldSegmResults_saliency_large_light.pdf}}
\hspace{0.cm}
\subfloat[Field A]{\includegraphics[scale=0.22]{MolongloS17FieldSegmResults_large_light.pdf}}\\
\vspace{-0.4cm}
\subfloat[Field B]{\includegraphics[scale=0.22]{MolongloSNRFieldSegmResults_saliency_large_light.pdf}}
\hspace{0.cm}
\subfloat[Field B]{\includegraphics[scale=0.22]{MolongloSNRFieldSegmResults_large_light.pdf}}\\
\vspace{-0.4cm}
\subfloat[Field C]{\includegraphics[scale=0.22]{MolongloExtendedFieldSegmResults_saliency_large_light.pdf}}
\hspace{0.cm}
\subfloat[Field C]{\includegraphics[scale=0.22]{MolongloExtendedFieldSegmResults_large_light.pdf}}\\
\vspace{-0.4cm}
\subfloat[Field D]{\includegraphics[scale=0.22]{MolongloFaintSNRSegmResults_saliency_large_light.pdf}}
\hspace{0.cm}
\subfloat[Field D]{\includegraphics[scale=0.22]{MolongloFaintSNRSegmResults_large_light.pdf}}
\caption{Segmentation results obtained for the Molonglo sample fields A-D (from top to bottom) assuming $l$=5 and $\beta$=1.
Left: Saliency maps normalized to range [0,1]; Right: Segmentation maps. Each segmented region is colored in the plot according to the mean of its pixel fluxes in mJy/beam units.
The contours shown with solid white lines correspond to a manual segmentation generated by an expert astronomer.}
\label{MolongloSegmResultsFig}
\end{figure*}
As discussed in Paper I the regions of extended emission present in the test fields A-D are in a few cases firmly associated with real source objects or candidates.
In most cases, however, no association with known sources has been established and an artefact nature cannot be excluded a priori without a further insight and
comparison to other surveys carried out with different telescopes or wavelength domains. As a result, no ground truth information at pixel level is available to quantify
the algorithm performances in terms of widely used measures, such as the identification efficiency and false
detection rate. The quality of the reconstruction will be therefore compared to a human-driven segmentation generated for each sample image by an expert astronomer. To enhance the
source/artefact discrimination capabilities, we considered the same sample scenarios as observed in the Molonglo Galactic Plane Survey (MGPS) at 843 MHz, reported
in Fig.~\ref{MolongloScorpioFieldFig}. The rms sensitivity over the survey is around 1-2 mJy/beam and the positional accuracy is 1-2". The lower resolution appears evident, particularly
in Field B and C in which some of the extended regions present in SCORPIO are not fully resolved and are detected as compact sources in the source finding stage. On the other hand, due
to the lower observing frequency, regions of extended emission are brighter and can be detected at higher significance levels. Furthermore, it is
unlikely that the same imaging artefacts appear in both surveys which are conducted with different telescopes. Thus, common emission features can be therefore
considered as real with a high degree of confidence.
\begin{figure*}
\subfloat[Field B - \textit{Aegean}, \textit{Blobcat}]{\includegraphics[scale=0.27]{ScorpioSNRFieldZoom_AegeanResults_finalPaperVersion_light.pdf}\label{OtherMethodResultsFig1}}
\hspace{-0.1cm}
\subfloat[Field D - \textit{Aegean}, \textit{Blobcat}]{\includegraphics[scale=0.27]{ScorpioFaintSNRField_AegeanResults_finalPaperVersion_light.pdf}\label{OtherMethodResultsFig4}}\\
\vspace{-0.4cm}
\subfloat[Field B - \textit{Chan-Vese}]{\includegraphics[scale=0.27]{ScorpioSNRFieldZoom_CVResults_finalPaperVersion_light.pdf}\label{OtherMethodResultsFig2}}
\hspace{-0.1cm}
\subfloat[Field D - \textit{Chan-Vese}]{\includegraphics[scale=0.27]{ScorpioFaintSNRField_CVResults_finalPaperVersion_light.pdf}\label{OtherMethodResultsFig5}}\\
\vspace{-0.4cm}
\subfloat[Field B - \textit{SWT}]{\includegraphics[scale=0.27]{ScorpioSNRFieldZoom_WTResults_finalPaperVersion_light.pdf}\label{OtherMethodResultsFig3}}
\hspace{-0.1cm}
\subfloat[Field D - \textit{SWT}]{\includegraphics[scale=0.27]{ScorpioFaintSNRField_WTResults_finalPaperVersion_light.pdf}\label{OtherMethodResultsFig6}}
\caption{Source finding results obtained with three different algorithms over field B (left panels) and field D (right panels) compared to the human segmentation
(solid white contours); Top: Results obtained with the Aegean (dotted green contours) and Blobcat source finders (dashed red contours); Center: Results obtained with the Chan-Vese algorithm
(dotted green contours); Right: Results obtained with the Stationary Wavelet Transform (SWT) method at scale $J$=5 (dotted green contours) and $J$=6 (dashed red contours).}
\label{OtherMethodResultsFig}
\end{figure*}
\subsection{Results}
We applied the designed segmentation algorithm to the selected test fields described in Section~\ref{SCORPIOSampleData}.
Multiple runs were performed under different choices of the algorithm parameters. The quality of the segmentation was visually inspected against
the human segmentation and a suitable choice of the algorithm parameters selected on the basis of the maximum
number of expected objects detected in all test fields at the corresponding minimum false detection rate.
A minimum region size $l$ for the initial segmentation equal to $l\sim4\times$\texttt{beam} (equivalent to $l$=20 pixels) was considered. Smaller values (e.g. $l$=10 pixels), comparable to the
beam size, were found to be too sensitive to small-scale structures (residual compact emission, artefacts) in the image and thus provide noisy segmentation results.
Larger values, e.g. $l$=30-60 pixels, were investigated as well. As $l$ increases, small-scale details of the extended sources may be smoothed out. This does not represent an issue
for field A and B in which the extended emission scale is larger by a factor of 4-5 compared to the minimum region size. Furthermore a larger value of $l$ favors the
merging of artefacts in the background region, e.g. in Field B.
The regularization parameter $\beta$, controlling initial over-segmentation, was studied. Different values were considered ($\beta$=0.01, 1, 10, 100) in correspondence to
all other scanned parameters. Results were found comparable for $\beta$=0.01-1 while for values above $\beta$=10 the superpixels start to assume very compact shapes and does
not fit well to object boundaries.
The saliency maps computed for the SCORPIO sample fields using a multi-resolution range of $l$=20-60 pixels (step 10 pixels), in combination with background and noise maps, are shown
in the left panels of Fig.~\ref{ScorpioSegmResultsFig}. It can be noted how the faint diffuse emission, previously hardly detectable without manually adjusting the map contrast, is
significantly enhanced over the background after the saliency filter. The filter mostly preserves the expected object contours and slightly smooth out small scale details.
A thresholding procedure on these saliency maps provides the initial signal and background markers for the following algorithm stages. Suitable values of the global signal
threshold factor $f_{\texttt{thr}}^{\texttt{sig}}$ were searched over all test samples. The choice of the threshold level was mainly driven by Field D and control Field E and
optimal values were found in the range 2.5-2.8. Higher values (up to 3.0) can be given to other fields at the cost of missing parts of the faint SNR source in Field D and of the
large diffuse emission in Field C. Overall, we have found that the thresholded saliency map alone already provides a reasonable source detection. It is also worth to observe
that saliency maps may constitute a valid input for different algorithms.
Different choices of the similarity regularization parameter $\lambda$ were investigated: $\lambda$= 0, 0.1, 0.5. Results obtained with $\lambda$=0.1, 0.5 are overall comparable, with
slightly better results obtained with $\lambda$=0.5, while worse results are obtained with $\lambda$=0. This analysis demonstrates that incorporating an edge information in the algorithm
improves the segmentation quality, even though edges of radio objects are considerably softer than in natural images.
The results of the segmentation stage are reported in the right panels of Fig.~\ref{ScorpioSegmResultsFig} for the four tested fields assuming $l$=20 pixels, $\beta$=1 and
$\lambda$=0.5. Each segmented region is colored according to the mean of its pixel fluxes. The human segmentation is superimposed and shown with solid white
contours. As it can be seen known objects and regions of diffuse emission are all identified and kept for later post-processing. The algorithm, at least with this choice
of parameters, is also sensitive to other faint diffuse emission which were not identified in the human segmentation.
After a deeper inspection, some of these were clearly attributed to imaging artefacts present in the input map, particularly
in the field B in which a poorly cleaned bright object outside the studied field pollutes the entire map. For the remaining objects the nature remains unclear even after a visual
inspection. This kind of artefacts represents a limitation in current SCORPIO map release. They can be removed in our analysis by increasing the threshold levels in the saliency map, at the
cost of affecting source detection especially in fields C and D.
In Fig.~\ref{SparseFieldResultsFig} (right panel) we report the results obtained over test field E using the same algorithm parameters selected for fields A-D. The left panel
shows the input map while the right panel the map given to the segmentation algorithm after the compact source filtering and smoothing stage. As desired, no signal markers are
found in the saliency map and thus no extended source detection is reported.
An example of post-processing analysis, carried out for some relevant sources present in the test fields, is reported in Fig.~\ref{PostProcessingResultsFig}. Top panels
shows the identified sources (solid black line contours) with nested components detected using two different methods. Solid white line contours are obtained by
thresholding a multi-resolution saliency map computed over source pixels. Dashed white line contours are produced by a multi-scale blob detector approach, combining Laplacian
of Gaussian (LoG) image filters at different scales. Other analysis are possible with the designed algorithm, e.g. running the hierarchical clustering over the source
region to identify the most similar areas, thus not shown here.
As discussed in Section~\ref{SegmentationAlgorithmSection} a set of parameters can be computed for each detected source, even the nested ones. As an example we report in the
bottom panel of Fig.~\ref{PostProcessingResultsFig} the set of Zernike moments computed for the three sources up to the 4-th order. Note how the moments are sensitive to the
source morphology and can be in principle considered for classification studies in combination with the other computed parameters (not in this paper purposes). A study of the
suitable set of parameters and their robustness to noise is planned to be performed using simulated data.
\begin{figure}
\centering%
\includegraphics[scale=0.38]{SourceFluxComparison.pdf}
\caption{Integrated fluxes $S$ of extended sources in the test fields A-D, reconstructed with three different algorithms (black dots: \texttt{CAESAR}, red squares: Chan-Vese, blue triangles: Wavelet
Transform J=5), as a function of the human-driven segmentation flux $S_{h}$.}
\label{SourceFluxComparisonFig}
\end{figure}
\subsection{Application to data at different wavelengths}
To evaluate the results obtained on radio data collected at different wavelengths and detector resolutions/sensitivities we considered the same test scenarios as observed in the
Molonglo Galactic Plane Survey (MGPS) at 843 MHz, shown in Fig.~\ref{MolongloScorpioFieldFig}. We applied our method to the sample Molonglo fields using the same
parameters considered in the analysis of the SCORPIO fields, with the following exceptions related to the lower resolution and size of the Molonglo maps. Smaller values of the
superpixel sizes ($l$=5-10 pixels) can be
assumed with respect to the SCORPIO maps, in which we have considered a minimum value of $l$=20 pixels. Saliency maps have been therefore computed starting from the chosen minimum superpixel
size up to a smaller maximum scale value compared to that assumed in SCORPIO maps. A less aggressive initial smoothing filter is also assumed in this case. All the other algorithm
parameters are left unchanged. The results are reported in Fig.~\ref{MolongloSegmResultsFig}. Some of the extended sources present in the field are not resolved and are
detected as compact sources in the pre-filtering stage. The white contours shown in the plots are therefore relative to the detectable extended sources.
As it can be seen, all the known sources are detected with high fidelity when comparing to the superimposed human segmentation. Additional regions of diffuse emission are detected as well.
It is unclear at the present status whether they are real or most probably reconstruction artefacts. Overall, the results demonstrate that the method is flexible to be used
also with different data under a minor tuning of parameters driven by the data itself, mainly sensitivity and resolution.
\subsection{Results with different algorithms}\label{ResultsDifferentAlgorithms}
It is valuable to consider what can be achieved on SCORPIO observed fields with other existing algorithms.
Such a test is indeed useful to be carried out as many of the available algorithms were tested with less-sensitive radio data or benchmarked against simulated data neglecting the
real background behavior and the Galactic Plane diffuse emission.
Four different methods were considered and tested. The first two, \textit{Aegean} \citep{Hancock2012} and \textit{Blobcat} \citep{Hales2012} use a flood-fill
algorithm to detect blobs in the image, starting from pixels above a seed threshold $\sigma_{\texttt{seed}}$ ($\sigma_{\texttt{seed}}$=5) with respect to the background and
aggregating adjacent pixels above a second lower threshold $\sigma_{\texttt{merge}}$ ($\sigma_{\texttt{merge}}$=2.6). Blobs are finally deblended using curvature information.
Background and noise maps were computed using the \textit{BANE} tool distributed within the \textit{Aegean} source finder. A third method, adopted by \citet{Peracaula2011}, searches
for blobs on the Stationary Wavelet Transform (SWT) of a residual image, obtained from the input map by replacing bright compact sources with a random background estimate. We implemented
this method from scratch. Finally an implementation of the Chan-Vese active contour algorithm \citep{ChanVese2001} was considered and tested over the sample data. The method
iteratively evolves an initial contour till convergence on the boundaries of the foreground region. Contour evolution is done by seeking a level set function that minimizes a fitting energy
functional depending on a set of input parameters.
In Fig.~\ref{OtherMethodResultsFig} we report the sources detected by the four methods (from top to bottom) in fields B (left panels) and D (right panels) in comparison with
the human segmentation shown with solid white contours. \textit{Aegean} and \textit{Blobcat} results are comparable. As expected, both algorithms were found to perform very well to detect
bright and faint compact sources, including blended sources, but they are biased, by design, against extended sources. A 5$\sigma$ threshold was considered for source detection
with the Wavelet method on two different scales $J$=5, 6. In these conditions, most of the extended bright sources present in the fields can be detected. Fainter features, such as
parts of the supernova remnants or diffuse regions cannot be well detected, at least at the specified significance level.
The Chan-Vese algorithm was tested over the residual image under different choices of parameters and using a simple circular level-set as initial contour. A pre-smoothing stage
is applied to the input residual image. Contours surrounding areas of negative excesses with respect to the background level were removed from the set of final detected contours. As
it can be seen, the extended source features missed by the other algorithms can be extracted with high accuracy compared to the human segmentation. Some imaging artefacts are also detected
along with real sources even with the optimal choice of the Chan-Vese parameters. Overall, the Chan-Vese method was found to outperforms the other three tested algorithms in
fully detecting extended objects.
In Fig.~\ref{SourceFluxComparisonFig} we compare the integrated flux of the extended sources present in the four fields A-D estimated with three different methods
(\texttt{CAESAR}{}: black dots, Chan-Vese: red squares, Wavelet method at scale $J$=5: blue triangles) as a function of the flux estimated using the human-driven segmentation. A total of 30 source
candidates were identified, hereafter denoted as the "reference set". Data are reported in the plot for each algorithm in case of source identification and cross-match found with
the reference set. As it can be seen, the estimated fluxes closely follow the reference, the observed spread in flux being regarded as a measure of the source reconstruction
accuracy contribution to the total flux uncertainty. Overall, better results are obtained with the \texttt{CAESAR}{} and Chan-Vese algorithm, which are able to detect fainter sources with
respect to the Wavelet method and achieve a better accuracy in flux estimation.
We are aware that we have not exhausted the list of all possible algorithms for extended source extraction and that a deeper tuning is needed for the three tested algorithms before
drawing firm conclusions on their suitability for our goals. For instance, a more refined initialization strategy is desired in the Chan-Vese method together with a finer
exploration of the parameter space. Moreover, it is known that the two-level assumption (foreground/background) at the basis of the standard Chan-Vese algorithm may not be
accurate to scenarios in which a large variation of intensity levels is present. New active contours algorithms \citep{VeseChan2002,Yang2013}, overcoming some of the standard Chan-Vese
limitations, appeared recently in the literature and could be worthy of consideration. However, we expect that none of the methods will perform accurately
over all the presented images and that a combination of different techniques is probably required at the very end. That motivated the development of a completely
different approach reported in this paper.
\section{Summary}
We described in this paper a new algorithm for the detection of extended sources in radio maps, designed for the SCORPIO project and for next-generation
radio surveys. The algorithm was tested with real radio data observed in the SCORPIO and Molonglo surveys and compared with existing algorithms. The achieved performances are
found comparable or even superior to other approaches followed in the literature. The novel points introduced are:
\begin{itemize}
\item a new procedure for computing the background in presence of extended emission;
\item an efficient filter to enhance diffuse emission, based on compact bright source removal, smoothing and saliency estimation;
\item a flexible framework providing rich information for post-processing analysis and relaxing some of the limiting requirements used for compact source
detection (e.g. pixel adjacency)
\end{itemize}
The results obtained with real data are promising and motivate further work both on the data side and on the algorithm side.
For this purpose, a new release of the SCORPIO map, with improved cleaning procedure and data flagging applied, is in progress. Preliminary results on
the studied fields show that many of the artefacts present in the first data release are now properly removed. Further, a campaign of single-dish measurement in the SCORPIO field is
already scheduled to improve the map response to extended objects beyond the limits of the ATCA telescope. Source finding will therefore largely benefit from these improved maps.
At the same time, simulation activities were started with the aim of generating extended source mock scenarios with ground truth available at pixel level to study the achieved
source detection efficiency and contamination rate with realistic noise conditions.
We are currently working on possible significant improvements also on the algorithm side, both at code and method level. Among these, improving saliency estimation and resolution has
become an active field of development in recent works, see \citet{Perazzi2012,Cheng2014,Borji2014,Shi2015}. A proper combination of different algorithms
could be a viable solution to decrease the spurious detection rate. Suitable criteria for combining nearby candidate sources is another aspect to be investigated in detail.
The current algorithm implementation is not optimized for large maps, e.g. the full SCORPIO or expected ASKAP fields, as it still requires large computation
time, e.g. from few to $\sim$15-20 minutes depending on image size, and memory requirements even on a single field, mainly related to the superpixel similarity matrixes.
A new optimized version, also designed for parallel and/or distributed processing is therefore planned to be realized, possibly compliant with ASKAP EMU software pipeline
requirements in terms of input/output products to be supported, employed technologies and processing strategies \citep{Cornwell2011,Chapman2014}.
|
\section{Introduction}
\begin{figure}
\includegraphics[trim=0cm 0cm 0cm 0cm, clip=true, width=\linewidth]{lda2vec_network_publish.pdf}
\caption{\textit{lda2vec} builds representations over both words and documents by mixing word2vec's skipgram architecture with Dirichlet-optimized sparse topic mixtures. The various components and transformations present in the diagram are described in the text.
\label{fig:lda2vec_network}
}
\end{figure}
Topic models are popular for their ability to organize document collections into a smaller set of prominent themes.
In contrast to dense distributed representations, these document and topic representations
are generally accessible to humans and more easily lend themselves to being interpreted.
This interpretability provides additional options to highlight the patterns
and structures within our systems of documents. For example,
using Latent Dirichlet Allocation (LDA)
topic models can reveal cluster of words within documents \cite{Blei:2003tn},
highlight temporal trends \cite{Charlin:2015fy},
and infer networks of complementary products \cite{McAuley:2015kd}.
See \newcite{Blei:2010fu} for an overview of topic modelling
in domains as diverse as computer vision, genetic markers, survey data,
and social network data.
Dense vector approaches to building document representations also exist:
\newcite{Le:2014vd} propose paragraph vectors that
are predictive of bags of words within paragraphs, \newcite{Kiros:2015uq} build
vectors that reconstruct the sentence sequences before and after a given sentence,
and \newcite{Ghosh:2016tia} construct contextual LSTMs that predict proceeding sentence
features. Probabilistic topic models tend to form
documents as a sparse mixed-membership of topics while neural network
models tend to model documents as dense vectors.
By virtue of both their sparsity and low-dimensionality,
representations from the
former are simpler to inspect and more immediately yield high level
intuitions about the underlying system
(although not without hazards, see \newcite{Chang:2009wd}).
This paper explores
hybrid approaches mixing sparse document representations with dense
word and topic vectors.
Unfortunately, crafting a new probabilistic topic model
requires deriving a new approximation, a procedure
which takes substantial expertise and must be customized to every model.
As a result, prototypes are time-consuming to develop
and changes to model architectures must be carefully considered.
However, with modern automatic differentiation frameworks
the practitioner can focus
development time on the model design rather than the model approximations.
This expedites the process of evaluating which model features are relevant.
This work takes advantage of the Chainer \cite{Tokui:tt} framework
to quickly develop models while also enabling us to utilize GPUs
to dramatically improve computational speed.
Finally, traditional topic models over text do not take advantage
of recent advances in distributed word representations which can capture
semantically meaningful regularities between tokens.
The examination of word co-occurrences has proven to be a fruitful research paradigm.
For example, \newcite{Mikolov:2013uz} utilize Skipgram Negative-Sampling
(SGNS) to train word embeddings
using word-context pairs formed from windows moving across a text corpus.
These vector representations ultimately encode remarkable linearities
such as $king - man + woman = queen$.
In fact, \newcite{Levy:2014vd} demonstrate that this is implicitly
factorizing a variant of the Pointwise Mutual Information (PMI) matrix
that emphasizes predicting frequent co-occurrences over rare ones.
Closely related to the PMI matrix, \newcite{Pennington:2014jd} factorize
a large global word count co-occurrence matrix to yield more efficient
and slightly more performant computed embeddings than SGNS.
Once created, these representations
are then useful for information retrieval \cite{Manning:2009kv}
and parsing tasks \cite{Levy:2014gw}. In this work, we will
take advantage of word-level representations to build document-level
abstractions.
This paper extends distributed word representations
by including interpretable document representations
and demonstrate that model inference can be performed and extended
within the framework of automatic differentiation.
\section{Model}
This section describes the model for \textit{lda2vec}. We are interested
in modifying the Skipgram Negative-Sampling (SGNS) objective in \cite{Mikolov:2013uz} to utilize
document-wide feature vectors while simultaneously learning continuous document
weights loading onto topic vectors.
The network architecture is shown in Figure \ref{fig:lda2vec_network}.
The total loss term $\mathcal{L}$ in \eqref{eq:objl} is the sum of the Skipgram Negative Sampling Loss
(SGNS) $\mathcal{L}_{ij}^{neg}$ with the addition of a Dirichlet-likelihood term over document weights, $\mathcal{L}^d$ that will be discussed later. The loss is conducted using a context vector, $\vec{c_j}$, pivot word vector $\vec{w_j}$, target word vector $\vec{w_i}$, and negatively-sampled word vector $\vec{w_l}$.
\begin{align}
\label{eq:objl}
\mathcal{L} &= \mathcal{L}^{d} + \Sigma_{ij} \mathcal{L}_{ij}^{neg} \\
\label{eq:objns} \mathcal{L}_{ij}^{neg} &= \log \sigma(\vec{c_j} \cdot \vec{w_i}) + \Sigma_{l=0}^{n}\log\sigma(-\vec{c_j}\cdot \vec{w_{l}})
\end{align}
\subsection{Word Representation}
As in \newcite{Mikolov:2013uz}, pairs of
pivot and target words $(j, i)$ are extracted when they co-occur in a moving window scanning across the corpus.
In our experiments, the window contains five tokens before and after the pivot token.
For every pivot-target pair of words the pivot word is used to predict the nearby target word.
Each word is represented with a fixed-length dense distributed-representation vector,
but unlike \newcite{Mikolov:2013uz} the same word vectors are used in both the pivot and target representations.
The SGNS loss shown in \eqref{eq:objns} attempts to discriminate context-word pairs
that appear in the corpus from those randomly sampled from a `negative' pool of
words.
This loss is minimized when the observed words are completely separated from
the marginal distribution.
The distribution from which tokens are drawn is $u^{\beta}$, where $u$
denotes the overall word frequency normalized by the total corpus size.
Unless stated otherwise, the negative sampling power $beta$ is set to $3/4$ and the number of negative samples is fixed to $n=15$ as in \newcite{Mikolov:2013uz}.
Note that a distribution of $u^{0.0}$ would draw negative tokens
from the vocabulary with no notion of popularity while a distribution
proportional with $u^{1.0}$ draws from the empirical unigram distribution.
Compared to the unigram distribution,
the choice of $u^{3/4}$ slightly emphasizes choosing infrequent words for negative samples.
In contrast to optimizing the softmax cross entropy, which requires modelling
the overall popularity of each token,
negative sampling focuses on learning word vectors conditional on a context
by drawing negative samples from each token's marginal popularity in the corpus.
\subsection{Document Representations}
\textit{lda2vec} embeds both words and document vectors into the same space and trains
both representations simultaneously. By adding the pivot and document vectors together,
both spaces are effectively joined.
\newcite{Mikolov:2013uz} provide the intuition that word vectors can be summed together
to form a semantically meaningful combination of both words. For example, the vector representation
for $Germany + airline$ is similar to the vector for $Lufthansa$. We would like to exploit the
additive property of word vectors to construct a meaningful sum of word and document vectors.
For example, if as \textit{lda2vec} is scanning a document the $j$th word is $Germany$,
then neighboring words are predicted to be similar such as $France$, $Spain$, and $Austria$.
But if the document is specifically about airlines, then we would like to construct a document vector similar to the word vector for $airline$. Then instead of predicting tokens similar to $Germany$ alone,
predictions similar to both the document and the pivot word can be made such as: $Lufthansa$,
$Condor\ Flugdienst$, and $Aero\ Lloyd$. Motivated by the meaningful sums of words vectors,
in \textit{lda2vec} the context vector is explicitly designed to be the sum of a document vector
and a word vector as in \eqref{eq:objcj}:
\begin{align}
\label{eq:objcj}\vec{c_j} &= \vec{w_j} + \vec{d_j}
\end{align}
This models document-wide relationships by preserving $\vec{d_j}$ for all word-context pairs
in a document, while still leveraging local inter-word relationships stemming from
the interaction between the pivot word vector $\vec{w_j}$ and target word $\vec{w_i}$.
The document and word vectors are summed together to form a
context vector that intuitively captures long- and short-term themes, respectively.
In order to prevent co-adaptation, we also perform dropout on both the
unnormalized document vector $\vec{d_j}$ and the pivot word vector $\vec{w_j}$
\cite{Hinton:2012tv}.
\subsubsection{Document Mixtures}
If we only included structure up to this point, the model would produce a
dense vector for every document.
However, \textit{lda2vec} strives to form interpretable representations
and to do so an additional constraint is imposed such that the document representations
are similar to those in traditional LDA models.
We aim to generate a document vector from a mixture of topic vectors
and to do so, we begin by constraining the document vector $\vec{d_j}$
to project onto a set of latent topic vectors
${\vec{t_0}, \vec{t_1}, ...,\vec{t_k}}$:
\begin{equation}
\label{eq:objdj}\vec{d_j} = p_{j0} \cdot \vec{t_0} + p_{j1} \cdot \vec{t_1}
+ ... + p_{jk} \cdot \vec{t_k}
+ ... + p_{jn} \cdot \vec{t_n}
\end{equation}
Each weight $0\leq p_{jk}\leq1$ is a fraction that denotes the membership of document $j$ in the
topic $k$.
For example, the Twenty Newsgroups model described later has 11313 documents
and $n=20$ topics so $j=0...11312$, $k=0...19$.
When the word vector dimension is set to 300, it is assumed that
the document vectors $\vec{d_j}$,
word vectors $\vec{w_i}$ and topic vectors $\vec{t_k}$ all have dimensionality 300.
Note that the topics $\vec{t_k}$ are shared and are a
common component to all documents but whose strengths are modulated by
document weights $p_{jk}$ that are unique to each document.
To aid interpretability, the document memberships are designed to be non-negative,
and to sum to unity.
To achieve this constraint, a softmax transform maps latent vectors initialized in
$\mathbb{R}^{300}$ onto the simplex defined by $p_{jk}$.
The softmax transform naturally enforces the constraint that $\Sigma_k p_{jk}=1$
and allows us interpret memberships as percentages rather than unbounded weights.
Formulating the mixture in \eqref{eq:objdj} as a sum
ensures that topic vectors $\vec{t_k}$, document vectors $\vec{d_j}$ and
word vectors $\vec{w_i}$, operate in the same space.
As a result, what words $\vec{w_i}$ are most similar to
any given topic vector $\vec{t_k}$ can be directly calculated.
While each topic is not literally a token present in the corpus,
it's similarity to other tokens is meaningful and can be measured.
Furthermore, by examining the list of most similar words
one can attempt to interpret what the topic represents.
For example, by calculating the most similar token to any
topic vector (e.g. $argmax_i(\vec{t_0} \cdot \vec{w_i})$) one may discover that
the first topic vector $\vec{t_{0}}$ is similar to the tokens
\textit{pitching}, \textit{catcher}, and \textit{Braves} while the second topic vector $\vec{t_{1}}$
may be similar to \textit{Jesus}, \textit{God}, and \textit{faith}. This provides us the
option to interpret the first topic as \textit{baseball} topic, and as a result
the first component in every document proportion $p_{j0}$ indicates how much
document $j$ is in the \textit{baseball} topic.
Similarly, the second topic may be interpreted
as \textit{Christianity} and the second component of any document proportion $p_{j1}$
indicates the membership of that document in the \textit{Christianity} topic.
\subsubsection{Sparse Memberships}
Finally, the document weights $p_{ij}$ are sparsified by optimizing the document
weights with respect to a Dirichlet likelihood with a low concentration parameter
$\alpha$:
\begin{align}
\label{eq:objdl} \mathcal{L}^{d} &=\lambda \Sigma_{jk}\ (\alpha - 1) \log p_{jk}
\end{align}
The overall objective in \eqref{eq:objdl} measures the likelihood of document $j$
in topic $k$ summed over all available documents.
The strength of this term is modulated by the tuning parameter $\lambda$.
This simple likelihood encourages the document proportions coupling in each
topic to be sparse when $\alpha<1$ and homogeneous when $\alpha>1$.
To drive interpretability, we are interested in finding sparse memberships and so set
$\alpha={n}^{-1}$ where $n$ is the number of topics.
We also find that setting the overall strength of the Dirichlet
optimization to $\lambda=200$ works well.
Document proportions are initialized to be relatively homogeneous,
but as time progresses, the $\mathcal{L}^d$
encourages document proportions vectors to become
more concentrated (e.g. sparser) over time.
In experiments without this
sparsity-inducing term (or equivalently when $\alpha=1$)
the document weights $p_{ij}$ tend to have probability
mass spread out among all elements. Without any sparsity
inducing terms the existence of so many non-zero weights makes
interpreting the document vectors difficult. Furthermore, we find that
the topic basis are also strongly affected, and the topics become incoherent.
\subsection{Preprocessing and Training}
The objective in \eqref{eq:objl} is trained in individual minibatches at a time while using the Adam optimizer \cite{Kingma:2014us} for two hundred epochs across the dataset. The Dirichlet likelihood term $\mathcal{L}^d$ is typically computed over all documents, so in modifying the objective to minibatches we adjust the loss of the term to be proportional to the minibatch size divided by the size of the total corpus.
Our software is open source, available online, documented and unit tested\footnote{The code for \textit{lda2vec} is available online at \url{https://github.com/cemoody/lda2vec}}.
Finally, the top ten most likely words in a given topic are submitted to the online \textit{Palmetto}\footnote{The online evaluation tool can be accessed at \url{http://palmetto.aksw.org/palmetto-webapp/}} topic quality measuring tool and the coherence measure $C_v$ is recorded. After evaluating multiple alternatives,
$C_v$ is the recommended coherence metric in \newcite{Roder:2015ev}.
This measure averages the Normalized Pointwise Mutual Information (NPMI) for every pair of words within a sliding window of size 110 on an external corpus and returns mean of the NPMI for the submitted set of words. Token-to-word similarity is evaluated using the \texttt{3COSMUL} measure \cite{Levy:2014wb}.
\begin{figure}
\begin{tabular}{|r|r|r|}
\hline
\# of topics & $\beta$ & Topic Coherences \\
\hline
20 & 0.75 & \textbf{0.567} \\
30 & 0.75 & 0.555 \\
40 & 0.75 & 0.553 \\
50 & 0.75 & 0.547 \\
20 & 1.00 & 0.563 \\
30 & 1.00 & 0.564 \\
40 & 1.00 & 0.552 \\
50 & 1.00 & 0.558 \\
\hline
\end{tabular}
\caption{Average topic coherences found by \textit{lda2vec} in the Twenty Newsgroups dataset are given. The topic coherence has been demonstrated to correlate with human evaluations of topic models \cite{Roder:2015ev}. The number of topics chosen is given, as well as the negative sampling exponent parameter $\beta$. Compared to $\beta=1.00$, $\beta=0.75$ draws more rare words as negative samples. The best topic coherences are found in models $n=20$ topics and a $\beta=0.75$. \label{fig:20ngcoherence} }
\end{figure}
\begin{figure*}
\centering
\begin{tabular}{|l|l|l|l|l|}
\hline
Topic Label & ``Space'' & ``Encryption'' & ``X Windows'' & ``Middle East'' \\
\hline
Top tokens & astronomical & encryption & mydisplay & Armenian \\
& Astronomy & wiretap & xlib & Lebanese \\
& satellite & encrypt & window & Muslim \\
& planetary & escrow & cursor & Turk \\
& telescope & Clipper & pixmap & sy \\
\hline
Topic Coherence & 0.712 & 0.675 & 0.472 & 0.615 \\
\hline
\end{tabular}
\caption{Topics discovered by \textit{lda2vec} in the Twenty Newsgroups dataset. The inferred
topic label is shown in the first row. The tokens with highest similarity to the topic are shown
immediately below. Note that the twenty newsgroups corpus contains corresponding
newsgroups such as \textit{sci.space}, \textit{sci.crypt},
\textit{comp.windows.x} and \textit{talk.politics.mideast}. \label{fig:20ngtopics}}
\end{figure*}
\section{Experiments}
\subsection{Twenty Newsgroups}
This section details experiments in discovering the salient topics in the Twenty Newsgroups dataset, a popular corpus for machine learning on text.
Each document in the corpus was posted to one of twenty possible newsgroups.
While the text of each post is available to \textit{lda2vec},
each of the newsgroup partitions is not revealed to the algorithm
but is nevertheless useful for post-hoc qualitative evaluations of the discovered topics.
The corpus is preprocessed using the data loader
available in Scikit-learn \cite{Pedregosa:2012tv} and tokens are identified using the SpaCy parser \cite{Honnibal:2015jm}. Words are lemmatized to group multiple inflections into single tokens. Tokens that occur fewer than ten times in the corpus are removed, as are tokens that appear to be URLs, numbers or contain special symbols within their orthographic forms. After preprocessing, the dataset contains 1.8 million observations of 8,946 unique tokens in 11,313 documents. Word vectors are initialized to the pretrained values found in \newcite{Mikolov:2013uz} but otherwise updates are allowed to these vectors at training time.
A range of \textit{lda2vec} parameters are evaluated by varying the number of topics $n\in{20, 30, 40, 50}$ and the negative sampling exponent $\beta\in{0.75, 1.0}$. The best topic coherences were achieved with $n=20$ topics and with negative sampling power $\beta=0.75$ as summarized in Figure \ref{fig:20ngcoherence}. We briefly experimented with variations on dropout ratios but we did not observe any substantial differences.
Figure \ref{fig:20ngtopics} lists four example topics discovered in the Twenty Newsgroups dataset. Each topic is associated with a topic vector that lives in the same space as the trained word vectors and listed are the most similar words to each topic vector.
The first topic shown has high similarity to the tokens \textit{astronomical}, \textit{Astronomy}, \textit{satellite}, \textit{planetary}, and \textit{telescope} and is thus likely a `Space'-related topic similar to the `sci.space' newsgroup.
The second example topic is similar to words semantically related to `Encryption', such as \textit{Clipper} and \textit{encrypt}, and is likely related to the `sci.crypt' newsgroup. The third and four example topics are `X Windows' and `Middle East' which likely belong to the `comp.windows.x' and `talk.politics.mideast' newsgroups.
\begin{figure*}
\centering
\begin{tabular}{|l|l|l|l|l|l|}
\hline
``Housing Issues'' & ``Internet Portals'' & ``Bitcoin'' & ``Compensation'' & ``Gadget Hardware'' \\
\hline
more housing & DDG. & btc & current salary & the Surface Pro \\
basic income & Bing & bitcoins & more equity & HDMI \\
new housing & Google+ & Mt. Gox & vesting & glossy screens \\
house prices & DDG & MtGox & equity & Mac Pro \\
short-term rentals & iGoogle & Gox & vesting schedule & Thunderbolt \\
\hline
\end{tabular}
\caption{Topics discovered by \textit{lda2vec} in the Hacker News comments dataset. The inferred
topic label is shown in the first row. We form tokens from noun phrases to capture the unique vocabulary of this specialized corpus. \label{fig:hntopics} }
\end{figure*}
\begin{figure*}
\begin{tabular}{|l|l|l|l|l|l}
\hline
Artificial sweeteners & Black holes & Comic Sans & Functional Programming & San Francisco \\
\hline
glucose & particles & typeface & FP & New York \\
fructose & consciousness & Arial & Haskell & Palo Alto \\
HFCS & galaxies & Helvetica & OOP & NYC \\
sugars & quantum mechanics & Times New Roman & functional languages & New York City \\
sugar & universe & font & monads & SF \\
Soylent & dark matter & new logo & Lisp & Mountain View \\
paleo diet & Big Bang & Anonymous Pro & Clojure & Seattle \\
diet & planets & Baskerville & category theory & Los Angeles \\
carbohydrates & entanglement & serif font & OO & Boston \\
\hline
\end{tabular}
\caption{Given an example token in the top row, the most similar words available in the Hacker News comments corpus are reported. \label{fig:hnsim} }
\end{figure*}
\subsection{Hacker News Comments corpus}
This section evaluates \textit{lda2vec} on a very large corpus of Hacker News \footnote{See \url{https://news.ycombinator.com/}} comments. Hacker News is social content-voting website and community whose focus is largely on technology and entrepreneurship. In this corpus, a single document is composed of all of the words in all comments posted to a single article.
Only stories with more than 10 comments are included, and only comments from users with more than 10 comments are included.
We ignore other metadata such as votes, timestamps, and author identities.
The raw dataset \footnote{The raw dataset is freely available at \url{https://zenodo.org/record/45901}} is available for download online. The corpus is nearly fifty times the size of the Twenty Newsgroups corpus which is sufficient for learning a specialized vocabulary.
To take advantage of this rich corpus, we use the SpaCy to tokenize whole noun phrases and entities at once \cite{Honnibal:2015jm}.
The specific tokenization procedure\footnote{The tokenization procedure is available online at \url{https://github.com/cemoody/lda2vec/blob/master/lda2vec/preprocess.py}} is also available online, as are the preprocessed datasets \footnote{A tokenized dataset is freely available at \url{https://zenodo.org/record/49899}} results.
This allows us to capture phrases such as \textit{community policing measure} and prominent figures such as \textit{Steve Jobs} as single tokens.
However, this tokenization procedure generates a vocabulary substantially different from the one available in the Palmetto topic coherence tool and so we do not report topic coherences on this corpus. After preprocessing, the corpus contains 75 million tokens in 66 thousand documents with 110 thousand unique tokens. Unlike the Twenty Newsgroups analysis, word vectors are initialized randomly instead of using a library of pretrained vectors.
We train an \textit{lda2vec} model using 40 topics and 256 hidden units and report the learned topics that demonstrate the themes present in the corpus. Furthermore, we demonstrate that word vectors and semantic relationships specific to this corpus are learned.
In Figure \ref{fig:hntopics} five example topics discovered by \textit{lda2vec} in the Hacker News corpus are listed. These topics demonstrate that the major themes of the corpus are reproduced and represented in learned topic vectors in a similar fashion as in LDA \cite{Blei:2003tn}.
The first, which we hand-label \textit{Housing Issues} has prominent tokens relating to housing policy issues such as housing supply (e.g. \textit{more housing}), and costs (e.g. \textit{basic income} and \textit{house prices}). Another topic lists major internet portals, such as the privacy-conscious search engine `Duck Duck Go' (in the corpus abbreviated as \textit{DDG}), as well as other major search engines (e.g. \textit{Bing}), and home pages (e.g. \textit{Google+}, and \textit{iGoogle}). A third topic is that of the popular online curency and payment system \textit{Bitcoin}, the abbreviated form of the currency \textit{btc}, and the now-defunct Bitcoin trading platform \textit{Mt. Gox}. A fourth topic considers salaries and compensation with tokens such as \textit{current salary}, \textit{more equity} and \textit{vesting}, the process by which employees secure stock from their employers. A fifth example topic is that of technological hardware like \textit{HDMI} and \textit{glossy screens} and includes devices such as \textit{the Surface Pro} and \textit{Mac Pro}.
Figure \ref{fig:hnsim} demonstrates that token similarities are learned in a similar fashion as in SGNS \cite{Mikolov:2013uz} but specialized to the Hacker News corpus. Tokens similar to the token \textit{Artificial sweeteners} include other sugar-related tokens like \textit{fructose} and food-related tokens such as \textit{paleo diet}. Tokens similar to \textit{Black holes} include physics-related concepts such as \textit{galaxies} and \textit{dark matter}.
The Hacker News corpus devotes a substantial quantity of text to fonts and design, and the words most similar to \textit{Comic Sans} are other popular fonts (e.g. \textit{Times New Roman} and \textit{Helvetica}) as well as font-related concepts such as \textit{typeface} and \textit{serif font}. Tokens similar to \textit{Functional Programming} demonstrate similarity to other computer science-related tokens while tokens similar to \textit{San Francisco} include other large American cities as well smaller cities located in the San Francisco Bay Area.
\begin{figure}
\begin{tabular}{|p{4cm}|l|}
\hline
Query & Result \\
\hline
California + technology & Silicon Valley \\
\hline
digital + currency & Bitcoin \\
\hline
Javascript - browser + server & Node.js \\
\hline
Mark Zuckerberg - \newline Facebook + Amazon & Jeff Bezos \\
\hline
NLP - text + image & computer vision \\
\hline
Snowden - United States + Sweden & Assange \\
\hline
Surface Pro - Microsoft + Amazon & Kindle \\
\hline
\end{tabular}
\caption{Example linear relationships discovered by \textit{lda2vec} in the Hacker News comments dataset. The first column indicates the example input query, and the second column indicates the token most similar to the input. \label{fig:hnrel}}
\end{figure}
Figure \ref{fig:hnrel} demonstrates that in addition to learning topics over documents and similarities to word tokens,
linear regularities between tokens are also learned. The `Query' column lists a selection of tokens that when combined yield a token vector closest to the token shown in the `Result' column. The subtractions and additions of vectors are evaluated literally, but instead take advantage of the \texttt{3COSMUL} objective \cite{Levy:2014wb}. The results show that relationships between tokens important to the Hacker News community exists between the token vectors. For example, the vector for \textit{Silicon Valley} is similar to both \textit{California} and \textit{technology},
\textit{Bitcoin} is indeed a \textit{digital currency},
\textit{Node.js} is a technology that enables running \textit{Javascript} on \textit{servers} instead of on client-side \textit{browsers},
\textit{Jeff Bezos} and \textit{Mark Zuckerberg} are CEOs of \textit{Amazon} and \textit{Facebook} respectively,
\textit{NLP} and \textit{computer vision} are fields of machine learning research primarily dealing with \textit{text} and \textit{images} respectively,
Edward \textit{Snowden} and Julian \textit{Assange} are both whistleblowers who were primarily located in the \textit{United States} and \textit{Sweden}
and finally the \textit{Kindle} and the \textit{Surface Pro} are both tablets manufactured by \textit{Amazon} and \textit{Microsoft} respectively. In the above examples semantic relationships between tokens encode for attributes and features including: location, currencies, server v.s. client, leadership figures, machine learning fields, political figures, nationalities, companies and hardware.
\subsection{Conclusion}
This work demonstrates a simple model, \textit{lda2vec}, that extends SGNS \cite{Mikolov:2013uz} to build unsupervised document representations that yield coherent topics.
Word, topic, and document vectors are jointly trained and embedded in a common representation space that preserves semantic regularities between the learned word vectors while still yielding sparse and interpretable document-to-topic proportions in the style of LDA \cite{Blei:2003tn}. Topics formed in the Twenty Newsgroups corpus yield high mean topic coherences which have been shown to correlate with human evaluations of topics \cite{Roder:2015ev}. When applied to a Hacker News comments corpus, \textit{lda2vec} discovers the salient topics within this community and learns linear relationships between words that allow it solve word analogies in the specialized vocabulary of this corpus.
Finally, we note that our method is simple to implement in automatic differentiation frameworks and can lead to more readily interpretable unsupervised representations.
\bibliographystyle{acl}
|
\section{Introduction}
Optimisation problems of both practical and theoretical importance deal with the search of an optimal configuration for a set of variables to achieve some specified goals.
Potential solutions may be encoded with real-valued, discrete, binary or permutation decision variables depending on the problem to be solved.
Optimisation methods for real-valued functions can be roughly classified into two groups: direct and gradient-based methods
(\citealp{Chong:Zak:2013}; \citealp[Chap. 2]{Givens:Hoeting:2013}).
In direct search methods only the objective function is used to guide the search strategy, whereas gradient-based methods consider the first and/or second-order derivatives of the objective function during the search process.
Constraints may be present and are usually taken into account in the definition of the objective function or in the decision variables representation.
Direct search methods can be applied without modifications to many optimisation tasks, but they are usually slow requiring many function evaluations for convergence.
On the contrary, gradient-based methods quickly converge to an optimal solution, but are not efficient in non-differentiable or discontinuous problems.
Both direct and gradient-based techniques depend on the chosen initial starting values, so they can get stuck in suboptimal solutions.
Furthermore, they are not efficient in handling problems with discrete decision variables, and cannot be efficiently implemented on parallel machines.
Problems where the decision variables are expressed using discrete or binary values are usually referred to as combinatorial optimisation problems, and consist in searching for the best solution from a set of discrete items (\citealp{Papadimitriou:Steiglitz:1998}; \citealp[Chap. 3]{Givens:Hoeting:2013}).
Typical examples are the knapsack problem, the minimum spanning tree, the traveling salesman problem, and the vehicle routing problem. Although in principle these type of problems can be solved with exact algorithms, the time required to solve them increases exponentially as the size of the problem grows.
A large number of heuristics and metaheuristics algorithms have been proposed for solving complex optimisation tasks .
Specific (ad-hoc) heuristic techniques are able to identify solutions in a reasonably short amount of time, but the solutions obtained are generally not guaranteed to be optimal or accurate.
On the contrary, metaheuristics offer a tradeoff between exact and heuristics methods, in the sense that they are generic techniques that offer good solutions, often the global optimal value sought, in a moderate execution time by efficiently and effectively exploring the search space \citep{Luke:2013}.
This class of algorithms typically implements some form of stochastic optimisation and includes:
Evolutionary Algorithm \citep[EA;][]{Back:Fogel:Michalewicz:2000:vol1, Back:Fogel:Michalewicz:2000:vol2},
Iterated Local Search \citep[ILS;][]{Lourenco:Stutzle:2003},
Simulated Annealing \citep[SA;][]{Kirkpatrick:Gelatt:Vecchi:1983},
Tabu Search \citep[TS;][]{Glover:Laguna:2013},
and Ant Colony Optimisation \citep[ACO;][]{Dorigo:Stutzle:2004}.
EAs are stochastic iterative algorithms in which a population of individuals evolve by emulating the biological processes observed in natural evolution and genetics \citep{Eiben:Smith:2003, DeJong:2006, Simon:2013}.
Each individual of the population represents a tentative solution to the problem. The quality of the proposed solution is expressed by the value of a fitness function assigned to each individual. This value is then used by EAs to guide the search and improve the fitness of the population.
Compared to other metaheuristics algorithms, EAs are able to balance between exploration of new areas of the search space and exploitation of good solutions.
The trade-off between exploration and exploitation is controlled by some tuning parameters, such as the population size, the genetics operators (i.e. selection, crossover, and mutation), and the probability of applying them.
Genetic Algorithms (GAs) are search and optimisation procedures that are motivated by the principles of natural genetics and natural selection.
GAs are the ``earliest, most well-known, and most widely-used EAs'' \citep[p. 35]{Simon:2013}.
R\ offers several tools for solving optimisation problems. A comprehensive listing of available packages is contained in the CRAN task view on ``Optimization and Mathematical Programming'' \citep{CRANTaskView:Optimization}.
An extensive treatment of optimisation techniques applied to problems that arise in statistics and how to solve them using R\ is provided by \citet{Nash:2014}. A gentle introduction to metaheuristics optimisation methods in R\ is contained in \citet{Cortez:2014}.
The R\ package \CRANpkg{GA} is a flexible general-purpose set of tools for optimisation using genetic algorithms and it is fully described in \citet{Scrucca:2013}. Real-valued, integer, binary and permutation GAs are implemented, whether constrained or not. Users can easily define their own objective function depending on the problem at hand. Several genetic operators for selection, crossover, and mutation are available, and more can be defined by experienced R\ users.
This paper describes some recent additions to the \pkg{GA} package. The first improvement involves the option to use hybrid GAs. Although GAs are able to identify the region of the search space where the global optimum is located, they are not especially fast at finding the optimum when in a locally quadratic region.
Hybrid GAs combine the power of GAs with the speed of a local optimiser, allowing researchers to find a global solution more efficiently than with the conventional evolutionary algorithms.
Because GAs can be easily and conveniently executed in parallel machines, the second area of improvement is that associated with parallel computing. Two approaches, the master-slave and islands models, have been implemented and are fully described.
Several examples, using both real-world data examples and benchmark functions, are presented and discussed.
\section{GA package}
In the following we assume that the reader has already installed the latest version ($\ge 3.0$) of the package from CRAN with
\begin{example}
> install.packages("GA")
\end{example}
and the package is loaded into an R\ session using the usual command
\begin{example}
> library(GA)
\end{example}
\section{Hybrid genetic algorithms}
\label{sec:hga}
EAs are very good at identifying near-optimal regions of the search space (\textit{exploration}), but they can take a relatively long time to locate the exact local optimum in the region of interest (\textit{exploitation}).
More effective algorithms might try to incorporate efficient local search algorithms into EAs.
There are different ways in which local searches or problem-specific information can be integrated in EAs \citep[see ][Chap. 10]{Eiben:Smith:2003}.
For instance, a local search may be started from the best solution found by a GA after a certain number of iterations, so that, once a promising region is identified, the convergence to the global optimum can be speed up.
These evolutionary methods have been named in various ways, such as \emph{hybrid GAs}, \emph{memetic GAs}, and \emph{genetic local search algorithms}.
Some have argued that the inclusion of a local search in GAs implies the use of a form of Lamarckian evolution. This fact has been criticised from a biological point of view, but ``despite the theoretical objections, hybrid genetic algorithms typically do well at optimization tasks'' \citep[p. 82]{Whitley:1994}.
In case of real-valued optimisation problems, the \pkg{GA} package provides a simple to use implementation of hybrid GAs by setting the argument \code{optim = TRUE} in a \code{ga()} function call.
This allows to perform local searches using the base R\ function \code{optim()}, which makes available general-purpose optimisation methods, such as Nelder–Mead, quasi-Newton with and without box constraints, and conjugate-gradient algorithms.
Having set \code{optim = TRUE}, the local search method to be used and other parameters can be controlled with the optional argument \code{optimArgs}. This must be a list with the following structure and defaults:
\begin{example}
optimArgs = list(method = "L-BFGS-B",
poptim = 0.05,
pressel = 0.5,
control = list(fnscale = -1, maxit = 100))
\end{example}
where
\begin{longtable}[l]{lp{0.75\textwidth}}
\code{method} & The method to be used among those available in \code{optim} function (see Details section in \code{help(optim)}). By default, the BFGS with box constraints is used, where the bounds are those provided in the \code{ga()} function call).\\
\code{poptim} &
A value in the range $(0,1)$ which gives the the probability of applying the local search at each iteration.\\
\code{pressel} &
A value in the range $(0,1)$ which specifies the pressure selection.\\
\code{control} &
A list of parameters for fine tuning the \code{optim} algorithm (see \code{help(optim)} for details).
\end{longtable}
In the implementation available in \pkg{GA}, the local search is applied stochastically during the GA iterations with probability \code{poptim} $\in [0,1]$; by default, once every $1/0.05 = 20$ iterations on average.
The local search algorithm is started from a random selected solution drawn with probability proportional to fitness and with the selection process controlled by the parameter \code{pressel} $\in [0,1]$.
The latter value is used in the function \code{optimProbsel()} for computing the probability of selection for each individual of the genetic population. Smaller values of \code{pressel} tend to assign equal probabilities to all the solutions, and larger values tend to assign larger values to those solutions having better fitness.
As an example, consider the following output which presents a vector of fitness values \code{f} assgined to different solutions, and the corresponding probabilities of selection obtained by varying the selection pressure parameter:
\begin{example}
> f <- c(1, 2, 5, 10, 100)
> data.frame(f = f,
"0" = optimProbsel(f, 0),
"0.2" = optimProbsel(f, 0.2),
"0.5" = optimProbsel(f, 0.5),
"0.9" = optimProbsel(f, 0.9),
"1" = optimProbsel(f, 1),
check.names = FALSE)
f 0 0.2 0.5 0.9 1
1 1 0.2 0.1218 0.03226 0.00009 4.930e-32
2 2 0.2 0.1523 0.06452 0.00090 3.309e-24
3 5 0.2 0.1904 0.12903 0.00900 2.220e-16
4 10 0.2 0.2380 0.25806 0.09000 1.490e-08
5 100 0.2 0.2975 0.51613 0.90001 1.000e+00
\end{example}
When no pressure selection is set, i.e. at 0, the same probability is assigned to all. Larger probabilities are assigned to larger $f$ values as the pressure value increases. In the extreme case of pressure selection equal to 1, only the largest $f$ has assigned a probability of selection equal to 1, whereas the others have no chance of being selected.
When a \code{ga()} function call is issued with \code{optim = TRUE}, a local search is always applied at the end of GA evolution (even in case of \code{poptim = 0}), but now starting from the solution with the highest fitness value. The rationale for this is to allow for local optimisation as a final improvement step.
\subsection{Portfolio selection}
In portfolio selection the goal is to find the optimal portfolio, i.e. the portfolio that provides the highest return and lowest risk. This is achieved by choosing the optimal set of proportions of various financial assets \citep[Chap. 16]{Ruppert:Matteson:2015}.
In this section an example of mean–variance efficient portfolio selection \citep[Chap. 13]{Gilli:etal:2011} is illustrated.
Suppose we have selected 10 stocks from which to build a portfolio. We want to determine how much of each stock to include in our portfolio.
The \emph{expected return rate} of our portfolio is
$$
E(R) = \sum_{i=1}^{10} w_i E(R_i),
$$
where $E(R_i)$ is the expected return rate on asset $i$, and $w_i$ is the fraction of the portfolio value due to asset $i$. Note that the portfolio weights $w_i$ must satisfy the constraints $w_i \ge 0$, and $\sum_{i=1}^{10} w_i = 1$.
At the same time, we want to minimise the \emph{variance of portfolio returns} given by
$$
\sigma^2_p = w' \Sigma w,
$$
where $\Sigma$ is the covariance matrix of stocks returns, and $w' = (w_1, \ldots, w_{10})$, under the constraint that the portfolio must have a minimum expected return of 1\%, i.e $E(R) \ge 0.01$.
Consider the following stocks with monthly return rates obtained by Yahoo finance using the \CRANpkg{quantmod} package:
\begin{example}
> library(quantmod)
> myStocks <- c("AAPL", "XOM", "GOOGL", "MSFT", "GE", "JNJ", "WMT", "CVX", "PG", "WFC")
> getSymbols(myStocks, src = "yahoo")
> returns <- lapply(myStocks, function(s)
monthlyReturn(eval(parse(text = s)),
subset = "2013::2014"))
> returns <- do.call(cbind,returns)
> colnames(returns) <- myStocks
\end{example}
The monthly return rates for the portfolio stocks are shown in Figure~\ref{fig1:portfolio} and obtained with the code:
\begin{example}
> library(timeSeries)
> plot(as.timeSeries(returns), at = "chic", minor.ticks="month",
mar.multi = c(0.2, 5.1, 0.2, 1.1), oma.multi = c(4, 0, 4, 0),
col = .colorwheelPalette(10), cex.lab = 0.8, cex.axis = 0.8)
> title("Portfolio Returns")
\end{example}
\begin{figure}[htbp]
\centering
\includegraphics[width=\textwidth]{portfolio1}
\caption{Monthly return rates for a portfolio of selected stocks.}
\label{fig1:portfolio}
\end{figure}
Summary statistics for the portfolio stocks are computed as:
\begin{example}
> nStocks <- ncol(returns) # number of portfolio assets
> R <- colMeans(returns) # average monthly returns
> S <- cov(returns) # covariance matrix of monthly returns
> s <- sqrt(diag(S)) # volatility of monthly returns
> plot(s, R, type = "n", panel.first = grid(),
xlab = "Std. dev. monthly returns", ylab = "Average monthly returns")
> text(s, R, names(R), col = .colorwheelPalette(10), font = 2)
\end{example}
The last two commands draw a graph of the average vs standard deviation for the monthly returns (see Figure~\ref{fig2-4:portfolio}a). From this graph we can see that there exists a high degree of heterogenity among stocks, with AAPL having the largest standard deviation and negative average return, whereas some stocks have small volatility and high returns, such as WFC and MSFT. Clearly, the latter are good candidate for inclusion in the portfolio. The exact amount of each stock also depends on the correlation among stocks through the variance of portfolio returns $\sigma^2_p$, and so we need to formalise our objective function under the given constraints.
\begin{figure}[htbp]
\centering\footnotesize
\begin{minipage}{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{portfolio2}
(a)
\end{minipage}
\begin{minipage}{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{portfolio3}
(b)
\end{minipage}
\begin{minipage}{\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{portfolio4}\\
(c)
\end{minipage}
\caption{(a) Plot of average monthly returns vs the standard deviation for the selected stocks. (b) Portfolio stocks composition estimated by HGA. (c) Trace of HGA iterations.}
\label{fig2-4:portfolio}
\end{figure}
In order to compute the GA fitness function, we define the following functions:
\begin{example}
> weights <- function(w) # normalised weights
{ drop(w/sum(w)) }
> ExpReturn <- function(w) # expected return
{ sum(weights(w)*R) }
> VarPortfolio <- function(w) # objective function
{
w <- weights(w)
drop(w
}
\end{example}
We may define the fitness function to be maximised as the (negative) variance of the portfolio penalised by an amount which is function of the distance between the expected return of the portfolio and the target value:
\begin{example}
> fitness <- function(w) # fitness function
{
ER <- ExpReturn(w)-0.01
penalty <- if(ER < 0) 100*ER^2 else 0
-(VarPortfolio(w) + penalty)
}
\end{example}
A hybrid GA with local search can be obtained with the following call:
\begin{example}
> GA <- ga(type = "real-valued", fitness = fitness,
min = rep(0, nStocks), max = rep(1, nStocks), names = myStocks,
maxiter = 1000, run = 200, optim = TRUE)
> summary(GA)
+-----------------------------------+
| Genetic Algorithm |
+-----------------------------------+
GA settings:
Type = real-valued
Population size = 50
Number of generations = 1000
Elitism = 2
Crossover probability = 0.8
Mutation probability = 0.1
Search domain =
AAPL XOM GOOGL MSFT GE JNJ WMT CVX PG WFC
Min 0 0 0 0 0 0 0 0 0 0
Max 1 1 1 1 1 1 1 1 1 1
GA results:
Iterations = 216
Fitness function value = -0.00049345
Solution =
AAPL XOM GOOGL MSFT GE JNJ WMT CVX PG WFC
[1,] 0.031021 0.11981 0.035005 0.5248 0 0 0.17327 0.26192 0.18141 0.9932
> plot(GA)
\end{example}
The last command produces the graph on Figure~\ref{fig2-4:portfolio}c, which shows the trace of best, mean, and median values during the HGA iterations. The vertical dashes at the top of the graph indicate where the local search occurred.
It is interesting to note that the inclusion of a local search greatly speedup the termination of the GA search, which converges after 216 iterations. Without including the local optimisation step, a fitness function value within a 1\% from the maximum value found above is attained after $1,633$ iterations, whereas the same maximum fitness value cannot be achieved even after $100,000$ iterations.
The estimated portfolio weights and the corresponding expected return and variance are computed as:
\begin{example}
> (w <- weights(GA@solution))
AAPL XOM GOOGL MSFT GE JNJ WMT CVX
0.013369 0.051632 0.015085 0.226166 0.000000 0.000000 0.074671 0.112875
PG WFC
0.078178 0.428025
> ExpReturn(w)
[1] 0.016168
> VarPortfolio(w)
[1] 0.00049345
> barplot(w, xlab = "Stocks", ylab = "Portfolio weights",
cex.names = 0.7, col = .colorwheelPalette(10))
\end{example}
The last command draws a barchart of the optimal portfolio selected, and it is shown in Figure~\ref{fig2-4:portfolio}b.
\subsection{Poisson change-point model}
In the study of stochastic processes a common problem is to determine whether or not the functioning of a process has been modified over time. Change-point models assume that such a change is occurring at some point in time in a relatively abrupt manner \citep{Lindsey:2004}.
In a single change-point model the distribution of a response variable $Y_t$ at time $t$ is altered at the unknown point in time $\tau$, so we can write
\begin{equation}
Y_t \sim
\begin{cases}
f(y_t; \theta_1) & t < \tau \\
f(y_t; \theta_2) & t \ge \tau
\end{cases}
\label{eq:change-point}
\end{equation}
where $f(\cdot)$ is some given parametric distribution depending on $\theta_k$ for $k = \{1,2\}$, and $\tau$ is an unknown parameter giving the change-point time. Some or all of the elements of the vector of parameters $\theta_k$ in model \eqref{eq:change-point} may change over time. In more complex settings, the distribution function itself may be different before and after the change point.
Given a sample $\{y_t; t=1,\ldots,T\}$ of observations over time, the log-likelihood function of the change-point problem is
\begin{equation}
\ell(\theta_1, \theta_2, \tau; y_1, \ldots, y_T) =
\sum_{t < \tau} \log f(y_t; \theta_1) +
\sum_{t \ge \tau} \log f(y_t; \theta_2)
\label{eq:change-point-loglik}
\end{equation}
Further, for a Poisson change-point model we assume that $f(y_t; \theta_k)$ is the Poisson distribution with mean parameter $\theta_k$.
Consider the British coal-mining disasters dataset which provides the annual counts of disasters (having at least 10 deaths) from 1851 to 1962 \citep{Jarrett:1979, Raftery:Akman:1986}.
The data from Table~1 of \citet{Carlin:Gelfand:Smith:1992} are the following:
\begin{example}
> data <- data.frame(
y = c(4, 5, 4, 1, 0, 4, 3, 4, 0, 6, 3, 3, 4, 0, 2, 6, 3, 3, 5, 4, 5, 3, 1,
4, 4, 1, 5, 5, 3, 4, 2, 5, 2, 2, 3, 4, 2, 1, 3, 2, 2, 1, 1, 1, 1, 3,
0, 0, 1, 0, 1, 1, 0, 0, 3, 1, 0, 3, 2, 2, 0, 1, 1, 1, 0, 1, 0, 1, 0,
0, 0, 2, 1, 0, 0, 0, 1, 1, 0, 2, 3, 3, 1, 1, 2, 1, 1, 1, 1, 2, 4, 2,
0, 0, 0, 1, 4, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 1),
year = 1851:1962,
t = 1:112)
\end{example}
Graphs of annual counts and cumulative sums over time are shown in Figure~\ref{fig1:coalmine}. These can be obtained using the following code:
\begin{example}
> plot(y ~ year, data = data, ylab = "Number of mine accidents/yr")
> plot(cumsum(y) ~ year, data = data, type = "s",
ylab = "Cumsum number of mine accidents/yr")
\end{example}
Both graphs seems to suggest a two-regime behaviour for the number of coal-mining disasters.
\begin{figure}[htbp]
\centering\footnotesize
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{coalmine1a}
(a)
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{coalmine1b}
(b)
\end{minipage}
\caption{Plots of the number of yearly coal-mining accidents (a) and cumulative sum of mine accidents (b) from 1851 to 1962 in Great Britain.}
\label{fig1:coalmine}
\end{figure}
We start the analysis by fitting a no change-point model, i.e. assuming a homogeneous Poisson process with constant mean. Clearly, in this simple case the MLE of the Poisson parameter is the sample mean of counts. However, for illustrative purposes we write down the log-likelihood and we maximise it with a hybrid GA.
\begin{example}
> loglik1 <- function(th, data)
{
mu <- exp(th) # Poisson mean
sum(dpois(data$y, mu, log = TRUE))
}
> GA1 <- ga(type = "real-valued",
fitness = loglik1, data = data,
min = log(1e-5), max = log(6), names = "th",
maxiter = 200, run = 50,
optim = TRUE)
> exp(GA1@solution[1,])
1.7054
> mean(data$y)
[1] 1.7054
\end{example}
For the change-point model in \eqref{eq:change-point}, the mean function can be expressed as
$$
\mu_t = \exp\left\{ \theta_1 + (\theta_2 - \theta_1) I(t \ge \tau) \right\},
$$
where $\tau$ is the time of change-point, $\theta_1$ is the mean of the first regime, i.e. when $t < \tau$, $\theta_2$ is the mean of the second regime, i.e. when $t \ge \tau$, and $I(\cdot)$ denotes the indicator function (which is equal to 1 if its argument is true and 0 otherwise). In R\ the above mean function and the log-likelihood from \eqref{eq:change-point-loglik} can be written as
\begin{example}
> meanFun <- function(th, t)
{
tau <- th[3] # change-point parameter
th <- th[1:2] # mean-related parameters
X <- cbind(1, t >= tau) # design matrix
exp(drop(X
}
> loglik2 <- function(th, data)
{
mu <- meanFun(th, data$t) # vector of Poisson means
sum(dpois(data$y, mu, log = TRUE))
}
\end{example}
The vector \code{th} contains the three parameters that have to be estimated from the sample dataset \code{data}.
Note that, for convenience, it is defined as $(\theta_1, \theta^*_2, \tau)'$, where $\theta^*_2 = (\theta_2 - \theta_1)$ is the differential mean effect of second regime.
Maximising the log-likelihood in \code{loglik2()} by iterative derivative-based methods is not viable due to lack of differentiability with respect to $\tau$. However, hybrid GAs can be efficiently used in this case as follows:
\begin{example}
> GA2 <- ga(type = "real-valued",
fitness = loglik2, data = data,
min = c(log(1e-5), log(1e-5), min(data$t)),
max = c(log(6), log(6), max(data$t)+1),
names = c("th1", "th2", "tau"),
maxiter = 1000, run = 200,
optim = TRUE)
> summary(GA2)
+-----------------------------------+
| Genetic Algorithm |
+-----------------------------------+
GA settings:
Type = real-valued
Population size = 50
Number of generations = 1000
Elitism = 2
Crossover probability = 0.8
Mutation probability = 0.1
Search domain =
th1 th2 tau
Min -11.5129 -11.5129 1
Max 1.7918 1.7918 113
GA results:
Iterations = 364
Fitness function value = -168.86
Solution =
th1 th2 tau
[1,] 1.1306 -1.2344 41.804
> (mean <- exp(cumsum(GA2@solution[1,1:2]))) # mean function parameters
th1 th2
3.09756 0.90141
> (tau <- GA2@solution[1,3]) # change-point
tau
41.804
\end{example}
Note that both the estimated change-point and the means are quite close to those reported by \citet{Raftery:Akman:1986}, and \citet{Carlin:Gelfand:Smith:1992}, using Bayesian methodology.
The two estimated models can be compared using a model selection criterion, such as the Bayesian information criterion \citep[BIC;][]{Schwartz:1978} defined as
$$
\mathrm{BIC} = 2 \ell(\hat{\theta};y) - \nu\log(n)
$$
where $\ell(\hat{\theta}; y)$ is the log-likelihood evaluated at the MLE $\hat{\theta}$, $n$ is the number of observations, and $\nu$ is the number of estimated parameters. Using this definition, larger values of BIC are preferable.
\begin{example}
> (tab <- data.frame(
loglik = c(GA1@fitnessValue, GA2@fitnessValue),
df = c(ncol(GA1@solution), ncol(GA2@solution)),
BIC = c(2*GA1@fitnessValue - log(nrow(data))*ncol(GA1@solution),
2*GA2@fitnessValue - log(nrow(data))*ncol(GA2@solution))))
loglik df BIC
1 -203.86 1 -412.43
2 -168.86 3 -351.88
\end{example}
A comparison of BIC values clearly indicates a preference for the change-point model.
We may summarise the estimated model by drawing a graph of observed counts over time with the estimated means before and after the change-point:
\begin{example}
> mu <- meanFun(GA2@solution, data$t)
> col <- c("red3", "dodgerblue2")
> with(data,
{ plot(t, y)
abline(v = tau, lty = 2)
lines(t[t < tau], mu[t < tau], col = col[1], lwd = 2)
lines(t[t >= tau], mu[t >= tau], col = col[2], lwd = 2)
par(new=TRUE)
plot(year, cumsum(y), type = "n", axes = FALSE, xlab = NA, ylab = NA)
axis(side = 3); mtext("Year", side = 3, line = 2.5)
})
\end{example}
and a graph of observed cumulative counts and the estimated cumulative mean counts:
\begin{example}
> with(data,
{ plot(t, cumsum(y), type = "s", ylab = "Cumsum number of mine accidents/yr")
abline(v = tau, lty = 2)
lines(t[t < tau], cumsum(mu)[t < tau], col = col[1], lwd = 2)
lines(t[t >= tau], cumsum(mu)[t >= tau], col = col[2], lwd = 2)
par(new=TRUE)
plot(year, cumsum(y), type = "n", axes = FALSE, xlab = NA, ylab = NA)
axis(side = 3); mtext("Year", side = 3, line = 2.5)
})
\end{example}
Both graphs are reported in Figure~\ref{fig2:coalmine}. The latter plot is particularly illuminating of the good fit achieved by the selected model.
\begin{figure}[htbp]
\centering\footnotesize
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{coalmine2a}
(a)
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{coalmine2b}
(b)
\end{minipage}
\caption{Summary plots for the change-point model fitted to the British coal-mining accidents dataset: (a) plot of observed counts over time with the estimated means before and after the estimated change-point (vertical dashed line); (b) plot of observed cumulative counts (step function) and the cumulative estimated mean counts.}
\label{fig2:coalmine}
\end{figure}
\subsection{S-I-R model for influenza epidemic}
The S-I-R model is a simple epidemiology compartmental model proposed by \citet{Kermack:McKendrick:1927}, which assumes a fixed population with only three compartments or states:
\begin{itemize}
\item $S(t)$ = number of susceptible, i.e. the number of individuals susceptible to the disease not yet infected at time $t$;
\item $I(t)$ = number of infected, i.e. the number of individuals who have been infected at time $t$ with the disease and are capable of spreading the disease to those in the susceptible category;
\item $R(t)$ = number of recovered, i.e. those individuals who have been infected and then removed from the disease, either due to immunisation or due to death. Members of this compartment are not able to be infected again or to transmit the infection to others.
\end{itemize}
Using a fixed population, i.e. with constant size $N = S(t) + I(t) + R(t)$, \citet{Kermack:McKendrick:1927} derived the following system of quadratic ODEs:
\begin{align*}
\frac{dS}{dt} & = - \beta S I \\
\frac{dI}{dt} & = \beta S I - \gamma I \\
\frac{dR}{dt} & = \gamma I
\end{align*}
where $\beta > 0$ is the rate (constant for all individuals) at which an infected person infects a susceptible person, and $\gamma > 0$ is the rate at which infected people recover from the disease.
The flow of the S-I-R model can be represented in the following scheme:
$$
\fbox{$S(t)$} \quad\xrightarrow{\beta S I}\quad
\fbox{$I(t)$} \quad\xrightarrow{\gamma I}\quad
\fbox{$R(t)$}
$$
where boxes represent the compartments and arrows indicate flows between compartments. Note that
$\frac{dS}{dt} + \frac{dI}{dt} + \frac{dR}{dt} = 0$, then $S(t) + I(t) + R(t) = N$, and the initial condition $S(0) > 0, I(0) > 0, R(0) = 0$.
Thus, the system can be reduced to a system of two ODEs.
For our data analysis example, we consider the influenza epidemic in an English boarding school from 22nd January to 4th February 1978 as described in \citet[p. 325--326]{Murray:2002}.
There were 763 resident boys in the school, and one (the initial infective) returned from winter break with illness. Over the course of 13 days, 512 boys were infected by the flu.
\begin{example}
> day <- 0:14
> Infected <- c(1,3,6,25,73,222,294,258,237,191,125,69,27,11,4)
> N <- 763
> init <- c(S = N-1, I = 1, R = 0)
> plot(day, Infected)
\end{example}
We aim at estimating the values of $\beta$ and $\gamma$ based on the observed data by minimising the following loss function:
\begin{equation}
RSS(\beta, \gamma) = \sum_t e(t)^2 = \sum_t \left( I(t) - \hat{I}(t) \right)^2,
\label{eq:RSS_sir}
\end{equation}
where $I(t)$ is the number of infected observed at time $t$, and $\hat{I}(t)$ is the corresponding number of infected predicted by the model, which depends on the unknown parameters $\beta$ and $\gamma$. Nonlinear least squares can be used to fit this model to data, but it strongly depends on the initial values as shown below. A more robust approach can be pursued by using GAs.
First of all, we need to define a function which computes the values of the derivatives in the ODE system at time $t$. This function is then used, together with the initial values of the system and the time sequence, by function \code{ode()} in the R\ package \CRANpkg{deSolve} to solve the ODE system:
\begin{example}
> library(deSolve)
> SIR <- function(time, state, parameters)
{
par <- as.list(c(state, parameters))
with(par, { dS <- -beta * S * I
dI <- beta * S * I - gamma * I
dR <- gamma * I
list(c(dS, dI, dR))
})
}
> RSS.SIR <- function(parameters)
{
names(parameters) <- c("beta", "gamma")
out <- ode(y = init, times = day, func = SIR, parms = parameters)
fit <- out[,3]
RSS <- sum((Infected - fit)^2)
return(RSS)
}
\end{example}
The function \code{RSS.SIR()} computes the predicted number of infected $\hat{I}(t)$ from the solution of ODE system for the input \code{parameters} values, and returns the objective function in \eqref{eq:RSS_sir} to be minimised.
Then, a \code{ga()} function call can be used with local search to find the optimal values of parameters $(\beta, \gamma)$ in S-I-R model. Note that the fitness function is specified as a local function which simply returns the negative of the objective function.
In this case, fine tuning of local search is specified through the optional argument \code{optimArgs}: the selection pressure is set with \code{pressel} at a higher value, so better solutions have higher probability of being used as starting point for the local search, and \code{maxit} gets a two-values vector specifying the maximum number of iterations to be used, respectively, during the GA evolution and after the final iteration.
\begin{example}
> GA <- ga(type = "real-valued",
fitness = function(par) -RSS.SIR(par),
min = c(0,0), max = c(0.1,0.5),
names = c("beta", "gamma"),
popSize = 25, maxiter = 1000, run = 200,
optim = TRUE,
optimArgs = list(pressel = 0.8,
control = list(maxit = c(10,100))))
> summary(GA)
+-----------------------------------+
| Genetic Algorithm |
+-----------------------------------+
GA settings:
Type = real-valued
Population size = 25
Number of generations = 1000
Elitism = 1
Crossover probability = 0.8
Mutation probability = 0.1
Search domain =
beta gamma
Min 0.0 0.0
Max 0.1 0.5
GA results:
Iterations = 503
Fitness function value = -4507.1
Solution =
beta gamma
[1,] 0.0021806 0.44516
\end{example}
Based on the estimated parameters other quantities of interest can be computed. For instance, $1/\gamma = 1/0.44516 \approx 2.25$ is the average recovery time which expresses the duration of infection (in days), and $\beta/\gamma \times 100 = 0.0021806/0.44516 \times 100 \approx 0.49\%$ is the infection's contact rate.
\begin{figure}[htbp]
\centering\footnotesize
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{sir_fig1a}
(a)
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{sir_fig1b}
(b)
\end{minipage}
\caption{Influenza epidemic in an English boarding school in winter 1978: (a) plot of the number of infected students; (b) model predictions from the S-I-R model with parameters estimated by hybrid GAs.}
\label{fig1:sir}
\end{figure}
The graph in Figure~\ref{fig1:sir}b provides a graphical summary of quantities involved in S-I-R model and the dynamic evolution of epidemia:
\begin{example}
> t <- seq(0, 15, length = 100)
> fit <- data.frame(ode(y = init, times = t, func = SIR,
parms = GA@solution[1,]))
> col <- brewer.pal(4, "GnBu")[-1]
> matplot(fit$time, fit[,2:4], type = "l",
xlab = "Day", ylab = "Number of subjects",
lwd = 2, lty = 1, col = col)
> points(day, Infected)
> legend("right", c("Susceptibles", "Infecteds", "Recovereds"),
lty = 1, lwd = 2, col = col, inset = 0.05)
\end{example}
We note that \citet{Murray:2002} reported solution $(\beta = 0.00218, \gamma = 0.441)$ gives a RSS equal to $4535.9$, larger than the optimal solution found by HGAs which is equal to $4507.1$. Furthermore, direct optimisation depends on starting values and often converges to sub-optimal solutions as, for instance, the following:
\begin{example}
> optim(c(0.001,0.4), RSS.SIR, method = "L-BFGS-B", lower = GA@min, upper = GA@max)
$par
[1] 0.0021434 0.3954033
$value
[1] 8764.9
$counts
function gradient
96 96
$convergence
[1] 52
$message
[1] "ERROR: ABNORMAL_TERMINATION_IN_LNSRCH"
\end{example}
\section{Parallel genetic algorithms}
\label{sec:PGA}
Parallel computing in its essence involves the simultaneous use of multiple computing resources to solve a computational problem.
This is viable when a task can be divided into several parts that can be solved simultaneously and independently, either on a single multi-core processors machine or on a cluster of multiple computers.
Support for parallel computing in R\ is available since 2011 (version 2.14.0) through the base package \pkg{parallel}.
This provides parallel facilities previously contained in packages \pkg{multicore} and \pkg{snow}. Several approaches to parallel computing are available in R\ \citep{McCallum:Weston:2011}, and an extensive and updated list of R\ packages is reported in the CRAN task view on \textit{High-Performance and Parallel Computing with R} \citep{CRAN:HighPerfParComp}.
GAs are regarded as ``embarrassingly parallel'' problems, meaning that they require a large number of independent calculations with negligible synchronisation and communication costs.
Thus, GAs are particularly suitable for parallel computing, and it is not surprising that such idea has been often exploited to speed up computations (see for instance \citet{Whitley:1994} in the statistical literature).
\citet{Luque:2011} identify several types of parallel GAs. In the master-slaves approach there is a single population, as in sequential GAs, but the evaluation of fitness is distributed among several processors (\textit{slaves}). The \textit{master} process is responsible of the distribution of the fitness function evaluation tasks performed by the slaves, and for applying genetic operators such as selection, crossover, and mutation (see Figure~\ref{fig:GPGA}). Since the latter operations involve the entire population, it is also known as global parallel GAs (GPGA). This approach is generally efficient when the computational time involving the evaluation of the fitness function is more expensive than the communication overhead between processors.
Another approach is the case of distributed multiple-population GAs, where the population is partitioned into several subpopulations and assigned to separated islands. Independent GAs are executed in each island, and only occasionally sparse exchanges of individuals are performed among these islands (see Figure~\ref{fig:ISLPGA}). This process, called migration, introduces some diversity into the subpopulations, thus preventing the search from getting stuck in local optima. In principle islands can evolve sequentially, but increased computational efficiency is obtained by running GAs in each island in parallel. This approach is known as coarse-grained GAs or island parallel GAs (ISLPGA).
\begin{figure}[htb]
\centering
\begin{minipage}[b]{.45\linewidth}
\centering
\includegraphics[width=\textwidth]{master_slaves}
\vspace{2ex}
\end{minipage}%
\hspace{1cm}%
\begin{minipage}[b]{.45\linewidth}
\centering
\includegraphics[width=\textwidth]{islands_evol}
\end{minipage}\\[-10pt]
\begin{minipage}[t]{.45\linewidth}
\caption{Master-slaves or global parallel GA scheme (GPGA). The master process stores the population, executes genetic operations, and distributes individuals to the slaves, which only evaluate the fitness of individuals.}
\label{fig:GPGA}
\end{minipage}%
\hspace{1cm
\begin{minipage}[t]{.45\linewidth}
\caption{Islands parallel GA scheme (ISLPGA). In a multiple-population parallel GA each process is a simple GA which evolves independently. Individuals occasionally migrate between one island and its neighbours.}
\label{fig:ISLPGA}
\end{minipage}%
\end{figure}
By default, searches performed with the \pkg{GA} package occur sequentially. In some cases, particularly when the evaluation of the fitness function is time consuming, parallelisation of the search algorithm may be able to speed up computing time.
Starting with version 2.0, the \pkg{GA} package provides facilities for using parallel computing in genetic algorithms following the GPGA approach.
Recently, with version 3.0, the ISLPGA model has also been implemented in the \pkg{GA} package. The following subsections describes usage of both approaches.
Parallel computing in the \pkg{GA} package requires the following packages to be installed: \pkg{parallel} (available in base R), \CRANpkg{doParallel}, \CRANpkg{foreach}, and \CRANpkg{iterators}. Moreover, \CRANpkg{doRNG} is needed for reproducibility in the ISLPGA case.
\subsection{Global parallel implementation}
The GPGA approach to parallel computing in \pkg{GA} can be easily obtained by manipulating the optional argument \code{parallel} in the \code{ga()} function call. This argument accepts several different values. A logical value may be used to specify if parallel computing should be used (\code{TRUE}) or not (\code{FALSE}, default) for evaluating the fitness function. A numeric value can also be supplied, in which case it gives the number of cores/processors to employ; by default, all the available cores, as provided by \code{detectCores()}, are used.
Two types of parallel functionalities are available depending on system OS: on Windows only \textit{snow} type functionality is present, whereas on POSIX operating systems, such as Unix, GNU/Linux, and Mac OSX, both \textit{snow} and \textit{multicore} (default) functionalities are available. In the latter case, a string can be used as the argument to \code{parallel} to set out which parallelisation tool should be used.
A final option is available if a researcher plans to use a cluster of multiple machines. In this case, \code{ga()} can be executed in parallel using all, or a subset of, the cores available to each machine assigned to the cluster. However, this option requires more work from the user, who needs to set up and register a parallel back end. The resulting cluster object can be passed as input value to the \code{parallel} argument.
\subsection{Islands parallel implementation}
The ISLPGA approach to parallel computing in \pkg{GA} has been implemented in the \code{gaisl()} function. This function accepts the same input arguments as the \code{ga()} function \citep[see][Section~3]{Scrucca:2013}, with the following additional arguments:
\vspace{\dimexpr-2\parsep-2\parskip\relax}%
\begin{center}
\begin{tabular}[t]{lp{0.75\textwidth}}
\code{numIslands} &
An integer value which specifies the number of islands to use in the genetic evolution (by default is set to $4$). \\
\code{migrationRate} &
A value in the range $(0,1)$ which gives the proportion of individuals that undergo migration between islands in every exchange (by default equal to $0.10$). \\
\code{migrationInterval} &
An integer value specifying the number of iterations at which exchange of individuals takes place. This interval between migrations is called an \emph{epoch}, and it is set at 10 by default.
\end{tabular}
\end{center}
The implemented ISLPGA uses a simple \textit{ring topology}, in which each island is connected unidirectionally with another island, hence forming a single continuous pathway (see Figure~\ref{fig:ISLPGA}). Thus, at each exchange step the top individuals, selected according to the specified \code{migrationRate}, substitute random individuals (with the exception of the elitist ones) in the connected island.
By default, the function \code{gaisl()} uses \code{parallel = TRUE}, i.e. the islands algorithm is run in parallel, but other values can also be provided as described in the previous subsection. Note that it is possible to specify a number of islands larger than the number of available cores. In such a case, the parallel algorithm will be run using blocks of islands, with the block size depending on the maximal number of cores available or the number of processors as specified by the user.
It has been noted that using parallel islands GAs often leads to, not only faster algorithms, but also superior numerical performance even when the algorithms run on a single processor. This because each island can search in very different regions of the whole search space, thus enhancing the exploratory attitude of evolutionary algorithms.
\subsection{Simulation study}
In this Section results from a simulation study are presented and discussed. The main goal is to compare the performance of sequential GAs with the two forms of parallel algorithms implemented in the \pkg{GA} package, namely GPGA and ISLPGA, for varying number of cores and different fitness computing times.
A fictitious fitness function is used to allow for controlling the computing time required at each evaluation. This is achieved by including the argument \code{pause} which suspend the execution for a specified time interval (in seconds):
\begin{example}
> fitness <- function(x, pause = 0.1)
{
Sys.sleep(pause)
x*runif(1)
}
\end{example}
The simulation design parameters used are the following:
\begin{example}
> ncores <- c(1, 2, 4, 8, 16) # number of cores/processors
> pause <- c(0.01, 0.1, 1, 2) # pause during fitness evaluation
> nrep <- 10 # number of simulation replications
\end{example}
Thus, \code{ncores} specifies that up to 16 cores or CPU processors are used in the parallel GAs solutions for increasing time spent on fitness evaluation as specified by \code{pause} (in seconds). Each combination of design parameters is replicated \code{nrep = 10} times and results are then averaged.
GAs are run under the GPGA approach using \code{popSize = 50} and {maxiter = 100}. For ISLPGA runs the \code{numIslands} argument is set at the specified number of cores, with \code{popSize = 160} and \code{maxiter = 100}. The increased population size allows to work with at least 10 individuals on each island when \code{numIslands} is set at the maximum number of cores. In both cases, the remaining arguments in \code{ga()} or \code{gaisl()} function are set at their defaults.
The study was performed on a 16 cores Intel\textsuperscript{\textregistered} Xeon\textsuperscript{\textregistered} CPU E5-2630 running at 2.40GHz and with 128GB of RAM.
The R\ code used in the simulation study is provided in the accompanying supplemental material.
\begin{figure}[!htbp]
\centering
\includegraphics[width=\textwidth]{performance_GPGA_fig1}
\caption{Empirical GPGA performance for varying number of cores/processors and different fitness computing times. Graph on the left panel shows the average running times, whereas graph on the right panel shows the speedup factor compared to the sequential run (i.e. when only 1 core is used). In the latter plot, the dashed line represents the ``ideal'' linear speedup.}
\label{fig1:PGA}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=\textwidth]{performance_ISLPGA_fig1}
\caption{Empirical ISLPGA performance for varying number of cores/processors and different fitness computing times. Graph on the left panel shows the average running times, whereas graph on the right panel shows the speedup factor compared to the sequential run (i.e. when only 1 core is used). In the latter plot, the dashed line represents the ``ideal'' linear speedup.}
\label{fig1:ISLPGA}
\end{figure}
Graphs in the left panel of Figures \ref{fig1:PGA} and \ref{fig1:ISLPGA} show the average execution times needed for varying number of cores and different fitness computing times. As expected, increasing the number of cores allows to run GAs faster, but the improvement is not linear, in particular for the GPGA approach.
By using a machine with $P$ cores/processors, we would like to obtain an increase in calculation speed of $P$ times. However, this is typically not the case because in the implementation of a parallel algorithm there are some inherent non-parallelisable parts and communication costs between tasks \citep{Nakano:2012}.
The speedup achieved using $P$ processors is computed as $s_P = t_1/t_P$, where $t_i$ is the execution time spent using $i$ cores.
Graphs in the right panel of Figures \ref{fig1:PGA} and \ref{fig1:ISLPGA} show the speedup obtained in our simulation study. For the GPGA approach the speedup is quite good but it is always sub-linear, in particular for the less demanding fitness evaluation time and when the number of cores increases. On the other hand, the ISLPGA implementation shows a very good speedup (nearly linear).
Amdahl's law \citep{Amdahl:1967} is often used in parallel computing to predict the theoretical maximum speedup when using multiple processors. According to this, if $f$ is the fraction of non-parallelisable task, i.e. the part of the algorithm that is strictly serial, and $P$ is the number of processors in use, then the speedup obtained on a parallel computing platform follows the equation
\begin{equation}
S_P = \frac{1}{f + (1-f)/P} .
\label{eq:AmdahlLaw}
\end{equation}
In the limit, the above ratio converges to $S_{\max} = 1/f$, which represents the maximum speedup attainable in theory, i.e. by a machine with an infinite number of processors.
Figures \ref{fig2:PGA} and \ref{fig2:ISLPGA} show the observed speedup factors $S_P$ and the estimated Amdahl's law curves fitted by nonlinear least squares. In all the cases, Amdahl's law appears to well approximate the observed behaviour. The horizontal dashed lines are drawn at the maximum speedup $S_{\max}$, which is computed based on the estimated fraction of non-parallelisable task $f$ (see also Table~\ref{tab1:ISLPGA}). As the time required for evaluating the fitness function increases, the maximum speedup attainable also increases. As noted earlier, the ISLPGA approach shows an improved efficiency compared to the simple GPGA.
\begin{table}[!htb]
\centering
\caption{Fraction of non-parallelisable task $(f)$ estimated by nonlinear least squares using the Amdahl's law, and corresponding theoretical speedup $(S_{\max})$ for the GPGA and ISLPGA approaches.}
\label{tab1:ISLPGA}
\begin{tabular}{lrrrrcrrrr}
\toprule
& \multicolumn{4}{c}{GPGA} && \multicolumn{4}{c}{ISLPGA} \\
\cline{2-5} \cline{7-10} \\[-2ex]
& 0.01 & 0.1 & 1 & 2 && 0.01 & 0.1 & 1 & 2\\
\cline{2-5} \cline{7-10} \\[-2ex]
$f$ & 0.0695 & 0.0209 & 0.0122 & 0.0114 &
& 0.0069 & 0.0036 & 0.0031 & 0.0025 \\
$S_{\max}$ & 14.38 & 47.76 & 81.88 & 87.88 &
& 145.29 & 278.57 & 327.12 & 408.58 \\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}[!htb]
\centering
\includegraphics[width=\textwidth]{performance_GPGA_fig2}
\caption{Amdahl's law curves for the GPGA approach. Points refer to the speedup factors observed using different number of cores/processors, whereas the curves are estimated using nonlinear least squares. Horizontal dashed lines refer to the maximum speedup theoretically attainable. Each panel corresponds to a different fitness computing time (in seconds), and vertical axes are on log scale.}
\label{fig2:PGA}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=\textwidth]{performance_ISLPGA_fig2}
\caption{Amdahl's law curves for the ISLPGA approach. Points refer to the speedup factors observed using different number of cores/processors, whereas the curves are estimated using nonlinear least squares. Horizontal dashed lines refer to the maximum speedup theoretically attainable. Each panel corresponds to a different fitness computing time (in seconds), and vertical axes are on log scale.}
\label{fig2:ISLPGA}
\end{figure}
\subsection{ARIMA order selection}
Autoregressive moving average (ARMA) models are a broad class of parametric models for stationary time series popularised by \citet{Box:Jenkins:1976}. They provide a parsimonious description of a stationary stochastic process in terms of two polynomials, one for the auto-regression and the second for the moving average.
Nonstationay time series can be modelled by including an initial differencing step (``integrated'' part of the model). This leads to autoregressive integrated moving average (ARIMA) models, a popular modelling approach in real-world processes.
ARIMA models can be fitted by MLE after identifying the order $(p,d,q)$ for the autoregressive, integrated, and moving average components, respectively. This is typically achieved by preliminary inspection of the autocovariance function (ACF) and partial autocovariance function (PACF).
Model selection criteria, such as the Akaike information criterion (AIC), the corrected AIC (AICc), and the Bayesian information criterion (BIC), are also used for order selection.
The function \code{auto.arima()} in package \CRANpkg{forecast} provides an automatic algorithm which combines unit root tests, minimisation of the AICc in a stepwise greedy search, and MLE, to select the order of an ARIMA model.
Here, an island parallel GAs approach is used for order selection.
Consider the quarterly U.S. GNP from 1947(1) to 2002(3) expressed in billions of chained 1996 dollars and seasonally adjusted. The data are available on package \CRANpkg{astsa} and described in \citet{Shumway:Stoffer:2013}.
\begin{example}
> data(gnp, package="astsa")
> plot(gnp)
\end{example}
The plot of the time series obtained with the last command is shown in Figure~\ref{fig1-2:usgnp}a.
\begin{figure}[htbp]
\centering\footnotesize
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{usgnp1}
(a)
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{usgnp2}
(b)
\end{minipage}
\caption{(a) Plot of quarterly U.S. GNP from 1947(1) to 2002(3). (b) Trace of island parallel GAs search for ARIMA order selection.}
\label{fig1-2:usgnp}
\end{figure}
The selection of the ``optimal'' ARIMA$(p,d,q)$ model can be pursued by using binary GAs to maximise the BIC. The decision variables to be optimised are expressed in binary digits using the following function:
\begin{example}
> decode <- function(string, bitOrders)
{
string <- split(string, rep.int(seq.int(bitOrders), times = bitOrders))
orders <- sapply(string, function(x) { binary2decimal(gray2binary(x)) })
return(unname(orders))
}
\end{example}
For example, using 3 bits for encoding $p$ and $q$, and 2 bits for $d$, an ARIMA(3,1,1) model can be expressed with the binary string $(0,1,0,\;0,1,\;0,0,1)$:
\begin{example}
> decode(c(0,1,0, 0,1, 0,0,1), bitOrders = c(3,2,3))
[1] 3 1 1
\end{example}
Note that the \code{decode()} function assumes that the input binary string is expressed using Gray encoding, which ensures that consecutive values have the same Hamming distance \citep{Hamming:1950}.
The fitness function to be used in the GA search is defined as follows:
\begin{example}
> fitness <- function(string, data, bitOrders)
{
orders <- decode(string, bitOrders)
mod <- try(Arima(data, order = orders, include.constant = TRUE, method = "ML"),
silent = TRUE)
if(inherits(mod, "try-error")) NA else -mod$bic
}
\end{example}
Note that the objective function is defined as (minus) the BIC for the specified ARIMA model, with the latter fitted using the \code{Arima()} function available in the R\ package \pkg{forecast}.
An island binary parallel GA is then used to search for the best ARIMA model, using a migration interval of 20 generations, and the default migration rate of 0.1:
\begin{example}
> GA <- gaisl(type = "binary", nBits = 8,
fitness = fitness, data = gnp, bitOrders = c(3,2,3),
maxiter = 1000, run = 100, popSize = 50,
numIslands = 4, migrationInterval = 20)
> plot(GA)
> summary(GA)
+-----------------------------------+
| Genetic Algorithm |
| Islands Model |
+-----------------------------------+
GA settings:
Type = binary
Number of islands = 4
Islands pop. size = 12
Migration rate = 0.1
Migration interval = 20
Elitism = 1
Crossover probability = 0.8
Mutation probability = 0.1
GA results:
Iterations = 280
Epochs = 14
Fitness function values = -2259.615 -2259.615 -2259.615 -2259.615
Solutions =
x1 x2 x3 x4 x5 x6 x7 x8
[1,] 0 1 1 1 1 0 0 1
[2,] 0 1 1 1 1 0 0 1
[3,] 0 1 1 1 1 0 0 1
[4,] 0 1 1 1 1 0 0 1
\end{example}
Figure~\ref{fig1-2:usgnp}b shows the trace of the ISLPGA search for each of the four islands used. All the islands converge to the same final solution, as also shown by the summary output above.
The selected model is an ARIMA(2,2,1), which can be fitted using:
\begin{example}
> (orders <- decode(GA@solution[1,], c(3,2,3)))
[1] 2 2 1
> mod <- Arima(gnp, order = orders, include.constant = TRUE, method = "ML")
> mod
Series: gnp
ARIMA(2,2,1)
Coefficients:
ar1 ar2 ma1
0.2799 0.1592 -0.9735
s.e. 0.0682 0.0682 0.0143
sigma^2 estimated as 1451: log likelihood=-1119.01
AIC=2246.02 AICc=2246.21 BIC=2259.62
\end{example}
It is interesting to compare the above solution with that obtained with the automatic procedure implemented in \code{auto.arima()} using the same criterion:
\begin{example}
> mod1 <- auto.arima(gnp, ic = "bic")
> print(mod1)
Series: gnp
ARIMA(1,2,1)
Coefficients:
ar1 ma1
0.3243 -0.9671
s.e. 0.0665 0.0162
sigma^2 estimated as 1486: log likelihood=-1121.71
AIC=2249.43 AICc=2249.54 BIC=2259.62
> mod1$bic
[1] 2259.622
> mod$bic
[1] 2259.615
\end{example}
The model returned by \code{auto.arima()} is an ARIMA(1,2,1), so a simpler model where an AR(1) component is chosen instead of an AR(2). The BIC values are almost equivalent, with a slightly smaller value for the ARIMA(2,2,1) model identified by ISLPGA. However, by looking at some diagnostic plots it seems that a second-order AR component is really needed to account for autocorrelation at several lags as indicated by the Ljung-Box test of autocorrelation (see Figure~\ref{fig3:usgnp}; the code used to produce the plots is available in the supplementary material).
\begin{figure}[htbp]
\centering
\includegraphics[width=0.9\textwidth]{usgnp3}
\caption{ACF of residuals and p-values for the Ljung-Box test of autocorrelation for the ARIMA(1,2,1) model (top graphs) and the ARIMA(2,2,1) model (bottom graphs) fitted to the quarterly U.S. GNP data from 1947(1) to 2002(3).}
\label{fig3:usgnp}
\end{figure}
\subsection{Empirical Bayes beta-binomial model for rates estimation}
Consider the problem of estimating the probability $p_i$ of an event based on the observed number of successes $x_i$ out of $n_i$ trials, for $i=1,\ldots,N$ independent observations.
A reasonable model assumes a binomial distribution for the number of successes, i.e.
$$
x_i|p_i \sim Bin(p_i, n_i),
$$
with known trials $n_i > 0$ and unknown parameters $p_i$.
Suppose that the $p_i$s are generated from a common distribution, which we may take to be the Beta distribution, i.e.
$$
p_i \sim Be(\alpha, \beta).
$$
This a conjugate prior for the binomial likelihood, so the posterior distribution turns out to be
$$
p_i|x_i \sim Be(\alpha + x_i, \beta + n_i - x_i).
$$
The unknown rate $p_i$ can then be estimated by the posterior mean
$E(p_i|x_i) = \dfrac{\alpha + x_i}{\alpha + \beta + n_i}$,
or by the maximum a posteriori estimate,
$MAP(p_i|x_i) = \dfrac{\alpha + x_i - 1}{\alpha + \beta + n_i - 2}$.
In the Empirical Bayes approach the parameters $\alpha$ and $\beta$ of the prior distribution are estimated using the observed data. This is done by maximising the marginal likelihood of $x$ obtained by integrating the distribution of $x_i|p_i$ with respect to the parameter $p_i$. Thus, omitting the subscript $i$, we may write
\begin{align*}
f(x|\alpha,\beta,n) & = \int_{0}^{1} Bin(x|p,n) Be(p|\alpha,\beta) dp \\
& = \int_{0}^{1} \left\{ \binom{n}{x} p^x (1-p)^{n-x}
\frac{p^{\alpha-1}(1-p)^{\beta-1}}{B(\alpha,\beta)}
\right\} dp \\
& = \binom{n}{x} \frac{B(\alpha + x, \beta +n -x)}{B(\alpha,\beta)},
\end{align*}
where $B(\alpha,\beta)=\Gamma(\alpha)\,\Gamma(\beta)/\Gamma(\alpha+\beta)$ is the beta function, with $\Gamma(t) = \int_0^\infty x^{t-1} e^{-x}\,dx$. This is the density of a Beta-Binomial distribution, for which we can write
$$
x_i \sim BeBin(\alpha, \beta, n_i)
$$
with
$$
E(x_i) = n_i \frac{\alpha}{\alpha+\beta},
$$
and
$$
Var(x_i) = n_i \frac{\alpha\beta}{(\alpha+\beta)^2}
\frac{\alpha+\beta+n_i}{\alpha+\beta+1}.
$$
Under the independence assumption, the marginal log-likelihood is then
\begin{equation}
\ell(\alpha,\beta) =
\sum_{i=1}^{n} \left\{
\log\binom{n_i}{x_i} + \log B(\alpha + x_i, \beta + n_i - x_i)
- \log B(\alpha, \beta)
\right\}.
\label{eq:mloglik_betabinom}
\end{equation}
In the Empirical Bayes approach the general idea is to estimate the parameters of the prior distribution from the data, rather than fixing them based on prior knowledge. Thus, the MMLE of parameters $(\alpha,\beta)$ are obtained by maximising the marginal log-likelihood in \eqref{eq:mloglik_betabinom}, which are then used to obtain the posterior distribution.
Consider the data on mortality rates in 12 hospitals performing cardiac surgery on babies \citep[p. 15]{Spiegelhalter:etal:1996} and available in the R\ package \CRANpkg{SMPracticals}.
\begin{example}
> data("cardiac", package = "SMPracticals")
> x <- cardiac$r
> n <- cardiac$m
> Hospitals <- rownames(cardiac)
> plot(n, x/n, type = "n",
xlab = "Number of operations (n)",
ylab = "Mortality rates (x/n)")
> text(n, x/n, Hospitals)
\end{example}
As shown in Figure~\ref{fig12:surgical}a there exists a large variability on the number of operations $n_i$, ranging from Hospital A with 47 cardiac operations to Hospital D with more than 800. The ratios $x_i/n_i$ are the MLE for the mortality rates, but they strongly depend on the number of surgeries performed. For example, the mortality rate of 0 for the Hospital A is likely the result of the limited number of operations carried out.
\begin{figure}[htbp]
\centering\footnotesize
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{surgery1}
(a)
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{surgery2}
(b)
\end{minipage}
\caption{(a) Plot of mortality rates for cardiac surgery on babies at 12 Hospitals. (b) Contour plot of the marginal log-likelihood surface with axes for the parameters on the log scale.}
\label{fig12:surgical}
\end{figure}
The marginal log-likelihood in \eqref{eq:mloglik_betabinom} can be written as
\begin{example}
> mloglik <- function(par, x, size)
{
a <- par[1]
b <- par[2]
sum(lchoose(size, x) + lbeta(a+x, b+size-x) - lbeta(a, b))
}
\end{example}
A plot of the log-likelihood surface is shown in Figure~\ref{fig12:surgical}b and can be obtained using the following code:
\begin{example}
> ngrid <- 200
> a <- exp(seq(-5, 4, length.out = ngrid))
> b <- exp(seq(-5, 8, length.out = ngrid))
> grid <- expand.grid(a, b)
> mll <- function(par) mloglik(par, x, n)
> l <- matrix(apply(grid, 1, mll), ngrid, ngrid)
> image(a, b, l, col = spectral.colors(51), log = "xy",
xlab = expression(alpha), ylab = expression(beta), axes = FALSE)
> axis(1); axis(2); box()
> contour(a, b, l, nlevels = 21, add = TRUE)
> contour(a, b, l, levels = quantile(l,c(0.99,0.999)), drawlabels = FALSE, add = TRUE)
\end{example}
We opted to use parallel GAs evolving in four islands with the default immigration policies, using also a local optimisation search to speed up convergence to the optimal solution.
\begin{example}
> GA <- gaisl("real-valued",
fitness = mloglik, x = x, size = n,
min = exp(c(-5,-5)), max = exp(c(4,8)), names = c("a", "b"),
numIslands = 4, optim = TRUE,
maxiter = 1000, run = 200)
> plot(GA, log = "x")
> summary(GA)
+-----------------------------------+
| Genetic Algorithm |
| Islands Model |
+-----------------------------------+
GA settings:
Type = real-valued
Number of islands = 4
Islands pop. size = 25
Migration rate = 0.1
Migration interval = 10
Elitism = 1
Crossover probability = 0.8
Mutation probability = 0.1
Search domain =
a b
Min 0.0067379 0.0067379
Max 54.5981500 2980.9579870
GA results:
Iterations = 220
Epochs = 22
Fitness function values = -38.753 -38.753 -38.753 -38.753
Solutions =
a b
[1,] 8.2535 99.637
[2,] 8.2535 99.637
[3,] 8.2535 99.637
[4,] 8.2535 99.637
\end{example}
Looking at the trace of GA evolution in each island as shown in Figure~\ref{fig34:surgical}a, we can see that the algorithm soon achieves the optimal value for all the islands and then remain constants until a stopping rule is meet. The solution found is $(\hat{\alpha} = 8.2535, \hat{\beta} = 99.637$), which can be used to compute the posterior mean and the MAP estimate. For completeness we also compute the MLE and pooled MLE values:
\begin{example}
> (MLE <- x/n)
[1] 0.000000 0.121622 0.067227 0.056790 0.037915 0.066327 0.060811 0.144186 0.067633
[10] 0.082474 0.113281 0.066667
> (pooledMLE <- sum(x)/sum(n))
[1] 0.073916
> par <- GA@solution[1,]
> (posteriorMean <- (par[1] + x)/(par[1] + par[2] + n))
[1] 0.053286 0.102597 0.071636 0.059107 0.050969 0.069938 0.067425 0.121569 0.070671
[10] 0.079328 0.102376 0.068934
> (MAP <- (par[1] + x - 1)/(par[1] + par[2] + n - 2))
[1] 0.047442 0.099466 0.067826 0.058144 0.048135 0.067089 0.064018 0.119210 0.067926
[10] 0.075181 0.100178 0.067083
\end{example}
The estimates are shown graphically with the code
\begin{example}
> plot(n, MLE, log = "x",
xlab = "Number of operations",
ylab = "Estimated mortality rates")
> axis(1, at = seq(50,800,by=50), tck=-0.01, labels = FALSE)
> axis(2, at = seq(0.01,0.15,by=0.01), tck=-0.01, labels = FALSE)
> grid(equilogs = FALSE)
> points(n, posteriorMean, col = spectral.colors(2)[1], pch = 19)
> points(n, MAP, col = spectral.colors(2)[2], pch = 15)
> abline(h = pooledMLE, lty = 3)
> legend("bottomright", legend = c("MLE", "Pooled MLE", "Posterior mean", "MAP"),
col = c(1,1,spectral.colors(2)), pch = c(1,NA,19,15), lty = c(NA,3,NA,NA),
ncol = 2, inset = 0.03, cex = 0.8)
\end{example}
From Figure~\ref{fig34:surgical}b we can see that EB estimates for the mortality rates are shrunk toward the overall average (corresponding to the pooled MLE), with the posterior mean uniformly larger than the MAP due to the fact that the distribution is skewed to the right. EB prior estimation has a small effect on the Hospitals with larger number of surgical operations, whereas it has a large impact on those hospitals with small number of operations (e.g. Hospital A) or on those with more extreme rates (e.g. Hospitals H, B, K, and E).
\begin{figure}[htbp]
\begin{minipage}{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{surgery3}
(a)
\end{minipage}
\begin{minipage}{0.5\textwidth}
\centering
\includegraphics[width=\textwidth]{surgery4}
(b)
\end{minipage}
\caption{(a) Trace of GA evolution in each island, with the x-axis on the log scale to enhance the first few iterations. (b) Plot of estimated mortality rates vs the number of surgical operations, with the x-axis on the log scale.}
\label{fig34:surgical}
\end{figure}
Although better approaches are available \citep[sec. 3.5]{Carlin:Louis:2000}, and in particular that proposed in \citet{Laird:Louis:1987}, equi-tail \emph{naive} Empirical Bayes confidence intervals can be easily obtained from the quantiles of the Beta distribution:
\begin{example}
> level <- 0.95
> EBconfint <- data.frame(lower = numeric(length(x)),
upper = numeric(length(x)))
> for(i in 1:nrow(EBconfint))
{
EBconfint[i,] <- qbeta(c((1-level)/2, (1+level)/2),
shape1 = (par[1] + x[i]),
shape2 = (par[2] + n[i]))
}
> (tab <- data.frame(x, n, MLE, pooledMLE, MAP, posteriorMean, EBconfint))
x n MLE pooledMLE MAP posteriorMean lower upper
1 0 47 0.000000 0.073916 0.047442 0.053286 0.023805 0.093613
2 18 148 0.121622 0.073916 0.099466 0.102597 0.063953 0.133335
3 8 119 0.067227 0.073916 0.067826 0.071636 0.040438 0.104879
4 46 810 0.056790 0.073916 0.058144 0.059107 0.042638 0.071677
5 8 211 0.037915 0.073916 0.048135 0.050969 0.028915 0.075727
6 13 196 0.066327 0.073916 0.067089 0.069938 0.042261 0.097039
7 9 148 0.060811 0.073916 0.064018 0.067425 0.038717 0.097751
8 31 215 0.144186 0.073916 0.119210 0.121569 0.080397 0.145603
9 14 207 0.067633 0.073916 0.067926 0.070671 0.043145 0.097148
10 8 97 0.082474 0.073916 0.075181 0.079328 0.044698 0.115510
11 29 256 0.113281 0.073916 0.100178 0.102376 0.067903 0.125637
12 24 360 0.066667 0.073916 0.067083 0.068934 0.045445 0.089036
\end{example}
and shown graphically in Figure~\ref{fig5:surgical} using the following code
\begin{example}
> ord <- order(tab$posteriorMean)
> par(mar = c(4,6,2,1))
> with(tab[ord,],
{ plot(0, 0, ylim = range(ord), xlim = c(0,0.15), xaxt = "n", yaxt = "n",
xlab = "Estimated mortality rates", ylab = "")
axis(side = 1, at = seq(0,0.15,by=0.01))
axis(side = 2, at = seq(ord), las = 2,
labels = paste0(rownames(cardiac)[ord], " (", x, "/", n, ")"))
grid()
abline(v = pooledMLE, lty = 2)
mclust:::errorBars(seq(ord), lower, upper, col = spectral.colors(2)[1], horizontal = TRUE)
points(posteriorMean, seq(ord), pch = 19, col = spectral.colors(2)[1])
points(MLE, seq(ord), pch = 1)
})
\end{example}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\textwidth]{surgery5}
\caption{MLEs and posterior means with 95\% confidence intervals for the surgical mortality rates in each hospital. Numbers in brackets show the observed number of deaths and the total number of operations. The dashed vertical line indicates the population mean failure rate (pooled MLE).}
\label{fig5:surgical}
\end{figure}
\subsection{Benchmark function optimisation}
\label{sec:benchopt}
\citet{Mullen:2014} compared several optimisation algorithms using 48 benchmark functions available in the \CRANpkg{globalOptTests} package. \pkg{GA} was one of the several R\ packages investigated in such a comparison. However, with the settings used in this study, its overall performance was not particularly brilliant, ranking 14th out of 18 methods, thus leaving plenty of room for improvements.
One of the problematic case is the Griewank function, which is defined as
$$
f(x_1, \ldots, x_d) = 1 + \frac{1}{4000} \sum_{i=1}^d x^2_i - \prod_{i=1}^d \cos(x_i/\sqrt{i}).
$$
This a multimodal, non-separable function, with several local optima within the search region.
For any dimensionality $d$, it has one global minimum of zero located at the point $(0, \ldots, 0)$. Figure~\ref{fig1:Griewank} shows some perspective plots for $d=2$ at different zooming levels.
\begin{figure}[htbp]
\centering\footnotesize
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{Griewank_fig1}
(a)
\end{minipage}
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{Griewank_fig2}
(b)
\end{minipage}
\caption{Perspective plots of two-dimensional Griewank function. At larger scale the function appears convex (a), but zooming reveals a complex structure with numerous local minima (b).}
\label{fig1:Griewank}
\end{figure}
We replicated the simulation study in \citet{Mullen:2014} using the standard sequential GA (\code{GA}), the parallel island GA with 4 islands (\code{GAISL}), the hybrid GA with local search (\code{HGA}), and the island GA with local search (\code{HGAISL}).
Results for the Griewank function based on 100 replications are shown in Figure~\ref{fig1:GriewankBenchmark}.
The use of hybrid GAs, particularly in combination with the islands evolution, clearly yields more accurate solutions and with less dispersion.
The same behavior has been observed in many other benchmark functions available in the \pkg{globalOptTests} package.
\begin{figure}[htbp]
\centering\footnotesize
\includegraphics[width=0.7\textwidth]{GriewankBenchmark_fig1}
\caption{Results from 100 replications of Griewank function optimisation using standard GAs (\code{GA}), island GAs (\code{GAISL}), hybrid GAs with local search (\code{HGA}), and island GAs with local search (\code{HGAISL}).}
\label{fig1:GriewankBenchmark}
\end{figure}
\citet[Section 5]{Mullen:2014} also provided a measure of accuracy for each optimisation method considered by counting the number of successful runs, with the latter defined as a solution less than $0.005$ from the minimum of the objective function. The empirical accuracy scores obtained in our simulations are shown in Table~\ref{tab1:benchopt}, and these can be compared with those reported in Mullen's paper and its supplemental material. Hybrid GAs including local optimisation search (HGA) yield a large improvement on accuracy (ranking 2nd with a score of 3717), and when combined with island evolution (HGAISL) achieve the highest overall score (ranking 1st with a score equal to 3954).
\begin{table}[!htb]
\centering
\caption{Benchmark functions accuracy scores for GAs and some hybrid and islands evolution variants (larger values are better).}
\label{tab1:benchopt}
\begin{tabular}{ccccc}
\toprule
GA from Mullen's paper & GA & GAISL & HGA & HGAISL \\
\midrule
2259 & 2372 & 2587 & 3717 & 3954 \\
\bottomrule
\end{tabular}
\end{table}
\section{Summary}
\pkg{GA} is a flexible R\ package for solving optimisation problems with genetic algorithms. This paper discusses some improvements recently added to the package.
We have discussed the implementation of hybrid GAs, which employ local searches during the evolution of a GA to improve accuracy and efficiency.
Further speedup can also be achieved by parallel computing. This has been implemented following two different approaches. In the first one, the so-called master-slave approach, the fitness function is evaluated in parallel, either on a single multi-cores machine or on a cluster of multiple computers. In the second approach, called islands model, the evolution takes place independently on several sub-populations assigned to different islands, with occasional migration of solutions between islands.
Both enhancements often lead to high-quality solutions more efficiently.
Future plans include the possibility to improve overall performance by rewriting some key functions in C++ using the \CRANpkg{Rcpp} package. In particular, coding of genetic operators in C++ should provide sensible benefits in terms of computational speedup.
Finally, the package \CRANpkg{memoise} enables to store the results of an expensive fitness function call and returns the cached result when the same input arguments occur again. This strategy could be conveniently employed in the case of binary GAs.
\section*{Acknowledgements}
The author acknowledge the CINECA award under the ISCRA initiative (\url{http://www.hpc.cineca.it/services/iscra}) for the availability of high performance computing resources and support.
\baselineskip=14pt
\bibliographystyle{chicago}
|
\section{Hamiltonian Constraint}
The anisotropic scaling of space and time in HL theory breaks Lorentz invariance. In this way, the HL action is no longer a topological invariance and the theory is dynamical. Since GR is recovered in the low energy limit, the matter content will be introduced in the model as in General Relativity, i.e., via Schutz formalism of perfect fluids coupled to gravity \cite{schutz1,schutz2}. Then, the total action will be given by $S=S_{HL}+S_f$, in which $S_{HL}$ denotes the action of the FRW minisuperspace in HL theory and $S_f$ denotes the action of the matter content. Given the total action, the Hamiltonian constraint can be obtained by canonical methods. Following Dirac's algorithm of quantization, we associate to this Hamiltonian constraint an operator which annihilates the wave function of the universe. Due to the matter content, there is a parameter of time in the model and the resulting equation will be a Schr\"odinger-like equation.
In $(1+1)$ dimensions, the HL action is given by \cite{wang}
\begin{equation}
S_{HL}=\int{dtdxN(t)\sqrt{h_{11}}\left(\mathcal{L}_K-\mathcal{L}_V\right)}.
\end{equation}
In the above equation, $N(t)$ is the lapse function, $h_{11}$ is the spatial part of the metric, $\mathcal{L}_K$ is the kinetic term and $\mathcal{L}_V$ is the potential Lagrangian. The kinetic term depends on the extrinsic curvature $K_{11}$ of the constant leaves $t=\text{constant}$ as
\begin{equation}
\mathcal{L}_K=K^{11}K_{11}-\lambda K^2,
\end{equation}
where $\lambda$ is a dimensionless constant. Note that, although $\lambda$ breaks Lorentz invariance, the action is still invariant under the group of diffeomorphism of the spatial slices. The potential Lagrangian $\mathcal{L}_V$ is given by
\begin{equation}
\mathcal{L}_V=-2\Lambda-\alpha a^{i}a_{i},
\label{potential lagrangian}
\end{equation}
where $\alpha$ is a coupling constant and $a_{i}=\partial_i\ln{N}$.
For the matter content, the action of a perfect fluid coupled to gravity in Schutz formalism \cite{schutz1,schutz2} is given by
\begin{equation}
S_{f}=\int{d^2x\sqrt{-g}p},
\end{equation}
where $p$ is the fluid's pressure, which in turn is related to the fluid's density by the EoS $p=w\rho$. The two-velocity of the fluid depends on its specific enthalpy $\mu$, the specific entropy $S$ and two other potentials $\phi$ and $\theta$ with no clear physical meaning \cite{ahmed}. It is given by
\begin{equation}
U_\nu=\frac{1}{\mu}\left(\phi_{,\nu}+\theta S_{,\nu}\right).
\label{two-velocity}
\end{equation}
By thermodynamical arguments, Lapchinski and Rubakov \cite{Lapchinskii} found that the pressure depends on the Schutz potentials as
\begin{equation}
p=\frac{w\mu^{1+1/w}}{(1+w)^{1+1/w}}e^{-S/w}.
\end{equation}
Let us consider the FRW metric
\begin{equation}
ds^2=-N(t)^2dt^2+a(t)^2dx^2,
\end{equation}
where $a(t)$ is the scale factor. In this case, the extrinsic curvature tensor is given by
\begin{equation}
K_{11}=\frac{1}{N}(-\dot{h}_{11}+2\nabla_1N_1)=-\frac{a\dot{a}}{N}.
\label{extrinsic}
\end{equation}
The two-velocity field of the fluid is $U_{\nu}=N\delta^{0}_{\nu}$. By Eq. (\ref{two-velocity}) we have
\begin{equation}
\mu=\left(\frac{\dot{\phi}+\theta \dot{S}}{N}\right).
\label{enthalpy}
\end{equation}
Plugging Eqs. (\ref{extrinsic}) and (\ref{enthalpy}) into the total action $S=S_{HL}+S_{f}$ leads to
\begin{equation}\begin{aligned}
S=&\int{dtdx\left[(1-\lambda)\frac{\dot{a}^2}{Na}+2\Lambda aN\right.}\\&\left.+\frac{w}{(1+w)^{1+1/w}}\left(\frac{\dot{\phi}+\theta\dot{S}}{N}\right)^{1+1/w}Na e^{-S/w}\right].
\end{aligned}
\label{total action}
\end{equation}
Given the total action, we can define the conjugated momenta
\begin{equation}
\left\{\begin{aligned}
&p_a=2(1-\lambda)\frac{\dot{a}}{Na},\\
&p_\phi=\frac{a \mu^{1/w}}{(1+w)^{1/w}}e^{-S/w},\\
&p_{S}=\theta p_{\phi}.
\label{momenta}
\end{aligned}\right.
\end{equation}
The Hamiltonian is then given by
\begin{equation}
\begin{aligned}
H&=p_a\dot{a}+p_\phi\left(\dot{\phi}+\theta\dot{S}\right)-L,\\
&=N\left(\frac{ap_a^2}{4(1-\lambda)}+\frac{p_\phi^{1+w}e^{S}}{a^{w}}-2\Lambda a\right).
\end{aligned}
\label{hamiltonian1}
\end{equation}
Note that the above expression does not contain any linear conjugated momentum. However, if we perform the following canonical transformation
\begin{equation}
\left\{\begin{aligned}
&T=-p_Se^{-S}\phi^{-(1+w)},\\
&p_T=p_{\phi}^{1+w}e^{S},\\
&\bar{\phi}=\phi+(1+w)\frac{p_S}{p_\phi},\\
&\bar{p}_\phi=p_\phi,
\label{canonical transformation}
\end{aligned}\right.
\end{equation}
we arrive at
\begin{equation}
H=N\left(\frac{ap_a^2}{4(1-\lambda)}+\frac{p_T}{a^{w}}-2\Lambda a\right).
\label{hamiltonian2}
\end{equation}
The conjugated momentum $p_T$ appears linearly in the above expression. In this way, $T$ will be a parameter of time after quantization. Hamilton's equations applied to Eq. (\ref{hamiltonian2}) give us
\begin{equation}
\left\{\begin{aligned}
&\dot{p}_T=0\Rightarrow p_T=\text{constant},\\
&\dot{T}=\frac{N}{a^{w}},\\
&\dot{p}_{a}=N\left[-\frac{p_a^2}{4(1-\lambda)}+2\Lambda +w \frac{p_T}{a^{1+w}}\right],\\
&\dot{a}=N\frac{ap_a}{2(1-\lambda)},
\label{equations of motion}
\end{aligned}\right.
\end{equation}
so that $T$ is time in the gauge $N=a^w$.
Finaly, the super-Hamiltonian constraint comes from varying the action with respect to $N$. It is given by
\begin{equation}
\mathcal{H}=\frac{ap_a^2}{4(1-\lambda)}+\frac{p_T}{a^{w}}-2\Lambda a\approx 0.
\end{equation}
\section{Quantization}
To proceed with Dirac's algorithm of quantization of constrained systems \cite{dirac}, we perform the substitutions $p_a\to-i\partial/\partial a$ and $p_T\to -i \partial/\partial T$ and demand that the super-Hamiltonian operator annihilates the wave function, i.e., $\mathcal{H}\Psi=0$. We still have an operator ordering ambiguity, since $\hat{a}$ and $\hat{p}_a$ do not commute. With the operator ordering $ap_a^2\to \hat{p}_a\hat{a}\hat{p}_a$, the resulting Schr\"odinger-like equation
\begin{equation}
-\frac{1}{4(1-\lambda)}\frac{\partial}{\partial a}\left(a \frac{\partial \psi(a,T)}{\partial a}\right)-2\Lambda a \psi(a,T)=\frac{i}{a^{w}}\frac{\partial \psi(a,T)}{\partial T}
\label{Sturm-Liouville}
\end{equation}
is already in the Sturm-Liouville form. Note that, in order for the above equation to be formally self-adjoint, the inner product must be given by
\begin{equation}
\left<f(a),g(a)\right>=\int_{0}^{\infty}{f^{\ast}(a)g(a)\frac{da}{a^{w}}}.
\end{equation}
Given Eq. (\ref{Sturm-Liouville}), we can perform a Liouville transformation to obtain an equivalent Schr\"odinger equation
\begin{equation}
-\frac{\partial^2 y(x,t)}{\partial x^2}+V(x)y(x,t)=i\frac{\partial y(x,t)}{\partial t},
\label{Sch}
\end{equation}
with corresponding inner product
\begin{equation}
\left<f(x),g(x)\right>=\int{f^{\ast}(x)g(x)dx}.
\end{equation}
This can be done in the following way \cite{everitt}. For a Sturm-Liouville differential equation
\begin{equation}
-\left(p(a)\phi(a)\right)'+q(a)\phi(a)=\lambda \omega(a) \phi(a),
\end{equation}
where $\omega(a)$ is called the weight function, we define
\begin{equation}\left\{\begin{aligned}
&x=\int{\left(\frac{\omega(a)}{p(a)}\right)^{1/2}da},\\
&Y(x)=\left(p(a)\omega(a)\right)^{1/4}\phi(a),\\
&V(x)=\omega(a)^{-1}q(a)\\&\phantom{V(x)}-\left(\omega(a)^{-3}p(a)\right)^{1/4}\left\{p(a)\left[\left(p(a)\omega(a)\right)^{-1/4}\right]'\right\}'.
\end{aligned}\right.\end{equation}
As a result, the new function $Y(x)$ satisfies the differential equation
\begin{equation}
-Y''(x)+V(x)Y(x)=\lambda Y(x).
\end{equation}
To perform the Liouville transformation in Eq. (\ref{Sturm-Liouville}), two distinct cases must be analyzed: $w=1$ (radiation fluid with EoS $p=\rho$) and $w\neq 1$. In each case we must be careful if $\lambda<1$ or $\lambda>1$. For $\lambda<1$, the sign of the first term in Eq. (\ref{Sturm-Liouville}) will be negative and $t=T$ will be the natural time. For $\lambda>1$, the first term in Eq. (\ref{Sturm-Liouville}) changes sign, so the time will be $t=-T$. This caution is necessary since we want to obtain a Schr\"odinger equation of the form (\ref{Sch}).
\subsection{Case $w=1$:}
We define the new variable
\begin{equation}
x=\int{\left(\frac{\omega(a)}{p(a)}\right)^{1/2}da}=\sqrt{2m}\int{da/a}=\sqrt{2m}\ln{(a)},
\end{equation}
where $m=m_{<}\equiv 2(1-\lambda)$ if $\lambda<1$ and $m=m_{>}=2(\lambda-1)$ if $\lambda>1$. We also define the new wave function
\begin{equation}
y(x,t)=\left(p(a)\omega(a)\right)^{1/4}\psi(a,t)=\frac{1}{\left(2m\right)^{1/4}}\psi(a,t).
\end{equation}
A simple calculation shows that, in the new variable $x$, the wave function $y(x,t)$ satisfies
\begin{equation}
-\frac{\partial^2y(x,t)}{\partial x^2}+2\tilde{\Lambda} \exp{\left(\frac{2x}{\sqrt{2m}}\right)}y(x,t)=i\frac{\partial y(x,t)}{\partial t},
\label{Liouville1}
\end{equation}
where $\tilde{\Lambda}=-\Lambda$ if $\lambda<1$ and $\tilde{\Lambda}=\Lambda$ if $\lambda>1$. We see an exponential (Liouville) potential which grows with $x$ when $\tilde{\Lambda}>0$ and decreases without bound when $\tilde{\Lambda}<0 $. For $\tilde{\Lambda}=0$ we have a free particle Schr\"odinger equation.
\subsection{Case $w\neq 1$:}
In this case we define
\begin{equation}\left\{\begin{aligned}
&x=\int{\left(\frac{\omega(a)}{p(a)}\right)^{1/2}da}=\sqrt{2m}\int{da/a^{\frac{1+w}{2}}}\\&\phantom{x}=\sqrt{2m}\left(\frac{2}{1-w}\right)a^{\frac{1-w}{2}},\\
&y(x,t)=\left(p(a)\omega(a)\right)^{1/4}\psi(a,T)\\&\phantom{y(x,t)}=\frac{1}{\left(2m\right)^{1/4}}a^{\frac{1-w}{4}}\psi(a,t).
\label{transf. variables}
\end{aligned}\right.\end{equation}
A direct (tedious) calculation shows that
\begin{equation}
-\frac{\partial y(x,t)}{\partial x^2}+\left[\tilde{\Lambda}x^{2\left(\frac{1+w}{1-w}\right)}-\frac{1}{4x^2}\right]y(x,t)=i\frac{\partial y(x,t)}{\partial t},
\end{equation}
where we define
\begin{equation}\begin{aligned}
&\tilde{\Lambda}\equiv - 2\Lambda \left(2m_{<}\right)^{-\frac{1+w}{1-w}}\left(\frac{1-w}{2}\right)^{2\left(\frac{1+w}{1-w}\right)}\,\,\, \text{if} \,\,\,\lambda<1,\\
&\tilde{\Lambda}\equiv 2\Lambda \left(2m_{>}\right)^{-\frac{1+w}{1-w}}\left(\frac{1-w}{2}\right)^{2\left(\frac{1+w}{1-w}\right)}\,\,\, \text{if} \,\,\,\lambda>1.
\end{aligned}\end{equation}
Note that the resulting potential has a regular polynomial factor $x^{2\left(\frac{1+w}{1-w}\right)}$, plus an inverse square potential $-\frac{1}{4x^2}$, which does not depend on the EoS.
\section{Some solvable cases}
In this section, a few cases of interest will be analyzed. In particular, the cases $w=1$ (radiation fluid), $w=0$ (dust) and $w=-1$ (dark energy) will be studied, since explicit solutions can be found in these cases. The expectation values of the scale factors in each case will be found and compared with the classical predictions of Ref. \cite{pitelli1}.
\subsection{Case $w=1$ (radiation):}
\subsubsection{$\Lambda=0$:}
In this case, the equivalent Schr\"odinger equation becomes
\begin{equation}
-\frac{\partial^2y(x,t)}{\partial x^2}=i\frac{\partial y(x,t)}{\partial t}.
\end{equation}
This is just the free particle Schr\"odinger equation. No boundary conditions are necessary at $x=\pm \infty$, since square integrability is sufficient to uniquely determine the solution.
By a separation of variables $y(x,t)=X(x)e^{-i E t}$ we have $X_{k}(x)=e^{ikx}$, so that
\begin{equation}
y_k(x,t)=e^{ikx}e^{-i k^2 t},
\end{equation}
where $k^2=E$. A wave packet
\begin{equation}
y(x,t)=\int_{-\infty}^{\infty}{y_k(x,t)A(k)dk}
\end{equation}
can then be constructed by choosing $A(k)=e^{-\gamma k^2}$, with $\gamma>0$. This leads to
\begin{equation}\begin{aligned}
y(x,t)&=\int_{-\infty}^{\infty}{e^{i k x}e^{-i k^2 t}e^{-\gamma k^2}dk}\\&=\sqrt{\frac{\pi}{it+\gamma}}e^{-\frac{x^2}{4(it+\gamma)}}.
\end{aligned}\end{equation}
We thus have, in terms of the original variable $a$,
\begin{equation}
\Psi(a,t)=\left(2m\right)^{1/4}\sqrt{\frac{\pi}{it+\gamma}}e^{-\frac{m\left(\ln{a}\right)^2}{2(it+\gamma)}}.
\end{equation}
The expectation value of the scale factor is given by
\begin{equation}
\left<a\right>\!(t)=\frac{\int_{0}^{\infty}{a\left|\Psi(a,T)\right|^2\frac{da}{a}}}{\int_{0}^{\infty}{\left|\Psi(a,t)\right|^2\frac{da}{a}}}=e^{\frac{\gamma^2+t^2}{4m\gamma}}.
\label{scale factor 1}
\end{equation}
This shows an universe bouncing from a contraction epoch to an expansion era. However, $t$ is not the cosmic time. In fact, $t$ is related to the cosmic time $\tau$ by $d\tau=N(t)dt=\left<a\right>\!(t)dt$. We see that the cosmic time cannot be expressed in terms of elementary functions. Therefore, $\tau$ is found numerically and the behavior of $\left<a\right>(\tau)$ is showed graphically in Fig. \ref{fig1}. Note that, as $\tau\to\infty$, $\left<a\right>\!(\tau)$ approaches a straight line. In Ref. \cite{pitelli1} we see that the classical solution, i.e., the solution of Eq. (\ref{equations of motion}) is given by
\begin{equation}
a(\tau)=A+B\tau.
\end{equation}
Therefore, for asymptotically large universes we recover the classical behavior of the universe.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=8.5cm]{fig1}
\end{center}
\caption{Behavior of the expectation value of the scale factor as a function of the cosmic time $\tau$. We considered $\lambda=1/2$ and $\gamma=1$ in Eq. (\ref{scale factor 1}). The continuous line represents $\left<a\right>\!(\tau)$, while the dashed line represents the asymptotic behavior of $\left<a\right>\!(\tau)$.}
\label{fig1}
\end{figure}
\subsubsection{$\Lambda\neq 0$:}
By Eq. (\ref{Liouville1}), we see that the potential of the Schr\"odinger equation in the $(x,t)$ variables has the form $V(x)=2\tilde{\Lambda}\exp{\left(\frac{2x}{\sqrt{m}}\right)}$. Since analytical solutions can only be found when the exponential potential grows with $x$, I will, therefore, confine myself to two possibilities: $\Lambda>0$ for $\lambda<1$ and $\Lambda<0$ for $\lambda>1$. In this way, the normalized eigenstates are given by (see Ref. \cite{hoker})
\begin{equation}
\begin{aligned}
\Psi_E(a,t)=
&\left(2m\right)^{1/4}\left[\frac{2\sqrt{m}}{\pi}\sinh{\pi\left|\nu\right|}\right]^{1/2}\\&\times K_\nu\left(2\sqrt{\left|\Lambda\right| m}a\right)e^{-i E t},
\end{aligned}
\end{equation}
where $\nu=i\sqrt{2Em}$ and $K_{\nu}$ is the modified Bessel function of order $\nu$. It can be shown that the spectrum of energy is continuous and $E=0$ is a greatest lower bound. However, I was not able to find an explicit solution in this case.
\subsection{Case $w=0$ (dust):}
\subsubsection{$\Lambda=0$:}
The interesting feature with this case is that the same Schr\"odinger equation in Liouville form works for every fluid with EoS $p=w\rho$, namely,
\begin{equation}
-\frac{\partial^2y(x,t)}{\partial x^2}-\frac{1}{4x^2}y(x,t)=i\frac{\partial y(x,t)}{\partial t},
\label{-1/4}
\end{equation}
where $x=\sqrt{2m}\left(\frac{2}{1-w}\right)a^{\frac{1-w}{2}}$ and $y(x,t)=\frac{1}{\left(2m\right)^{1/4}}a^{\frac{1-w}{4}}\psi(a,t)$. With this potential, a boundary condition is necessary in order to solve Eq. (\ref{-1/4}). In Ref. \cite{gitman}, this problem has been solved and every possible boundary condition has been found. In order to obtain an explicit solution to the problem, I will choose a particular boundary condition, namely, Friedrichs boundary condition. For this boundary condition, the energy spectrum is continuous and positive. The propagator respecting Friedrichs boundary condition can be found in Ref. \cite{efthimiou} and is given by
\begin{equation}
G(x,x';t)=\frac{1}{2t}\sqrt{x x'}i^{-1}e^{i\frac{x^2+x'^2}{4t}}J_{0}\left(\frac{x x'}{2t}\right),
\end{equation}
where $J_0$ is the $0-th$ order Bessel function.
Given an initial wave packet $y(x,0)$, we can find the solution $y(x,t)$ by means of the integral
\begin{equation}
y(x,t)=\int_{0}^{\infty}{G(x,x';t)y(x',0)dx'}.
\end{equation}
Choosing $y(x,0)=x^{1/2}e^{-\gamma x^2}$, with $\gamma>0$, we have
\begin{equation}
y(x,t)=\frac{x^{1/2}i^{-1}}{4\gamma t-i}e^{\frac{i x^2}{4t}}e^{-\frac{x^2}{16\gamma t^2-4it}}.
\end{equation}
For $w=0$ we have $x=2\sqrt{2m}a^{1/2}$ and $\psi(a,t)=(2m)^{1/4}a^{-1/4}y(x,t)$ by Eq. (\ref{transf. variables}). The wave function $\psi(a,t)$ is then given by
\begin{equation}
\psi(a,t)=\frac{\sqrt{2}\sqrt{2m}i^{-1}}{4\gamma t-i}e^{i \frac{2m a}{t}}e^{-\frac{2ma}{4\gamma t^2-i t}}.
\end{equation}
The expectation value of the scale factor can be found to be
\begin{equation}
\left<a\right>\!(t)=\frac{\int_0^{\infty}{a\left|\psi(a,t)\right|^2\frac{da}{a^{0}}}}{\int_0^{\infty}{\left|\psi(a,t)\right|^2\frac{da}{a^{0}}}}=\frac{1+16\gamma^2t^2}{16\gamma m}.
\end{equation}
Since the scale factor is given by $d\tau=Ndt$ and $N=a^{0}$, $t=\tau$ in this case. Therefore, the expectation value of the scale factor as a function of the cosmic time $\tau$ is given by
\begin{equation}
\left<a\right>\!(\tau)=\frac{1+16\gamma^2\tau^2}{16\gamma m}.
\end{equation}
The classical solution in this case (see Ref. \cite{pitelli1} and Eq. (\ref{equations of motion})) is given by
\begin{equation}
a(\tau)=A+B \tau-\frac{p_T}{2}\tau^2.
\end{equation}
Note that the quantum theory smooth out the classical singularity, giving a bouncing model with an even expectation value of the scale factor.
\subsubsection{$\Lambda\neq 0$:}
In this case, the potential of the Schr\"odinger equation of the universe has the form $V(x)=\left[\tilde{\Lambda}x^{2}+\frac{1}{4x^2}\right]$. Analytical solutions can be found for potentials of the form $V(x)=x^2-\frac{1}{4x^2}$. In this way, I consider the cases $\Lambda>0$ if $\lambda<1$ and $\Lambda<0$ if $\lambda>1$. We arrive at the following Schr\"odinger equation
\begin{equation}
-\frac{\partial^2y(x,t)}{\partial x^2}+\left[|\tilde{\Lambda}|x^2-\frac{1}{4x^2}\right]y(x,t)=i\frac{\partial y(x,t)}{\partial t}.
\end{equation}
The propagator for the above equation can be found in \cite{efthimiou} and is given by
\begin{equation}\begin{aligned}
G(x,x';t)&=\frac{\sqrt{|\tilde{\Lambda}}|\sqrt{xx'}i^{-1}}{\sin{\left(2\sqrt{|\tilde{\Lambda}|t}\right)}}\\&\times\exp{\left[i\frac{\sqrt{|\tilde{\Lambda}|}}{2}\cot{\left(2\sqrt{|\tilde{\Lambda}|t}\right)\left(x^2+x'^2\right)}\right]}\\&\times J_{0}\left(\frac{\sqrt{|\tilde{\Lambda}|}xx'}{\sin{\left(2\sqrt{|\tilde{\Lambda}}|t\right)}}\right).
\end{aligned}\end{equation}
By choosing an initial wave packet $y(x,0)=x^{1/2}e^{-\sigma x^2}$, with $\sigma>0$, we have
\begin{widetext}
\begin{equation}\begin{aligned}
y(x,t)&=\int_0^{\infty}{G(x,x';t)y(x',0)dx'}=\frac{ i^{-1}\sqrt{|\tilde{\Lambda}| x} }{2 \sigma \sin{\left(2\sqrt{|\tilde{\Lambda}|} t\right)} -i \sqrt{|\tilde{\Lambda}| } \cos \left(2 \sqrt{|\tilde{\Lambda}| } t\right)}\times \\ \times&\exp \left[\frac{\sqrt{|\tilde{\Lambda}| } x^2 \left(-i \sqrt{|\tilde{\Lambda }|}+\sigma \tan \left(\sqrt{|\tilde{\Lambda }|} t\right)-\sigma \cot \left(\sqrt{|\tilde{\Lambda}| } t\right)\right)}{2 \sqrt{|\tilde{\Lambda}| } \cot \left(2 \sqrt{|\tilde{\Lambda} } t\right)+4 i \sigma }\right].
\end{aligned}\label{general}\end{equation}
\end{widetext}
Since $x=2\sqrt{2m}a^{1/2}$ and $\psi(a,t)=(2m)^{1/4}a^{-1/4}y(x,t)$ we have the following expression for the wave function $\psi(a,t)$ of the universe:
\begin{widetext}
\begin{equation}\begin{aligned}
\psi(a,t)&=\frac{ 2i^{-1}\sqrt{|\tilde{\Lambda}|m} }{2 \sigma \sin{\left(2\sqrt{|\tilde{\Lambda}|}t\right)} -i \sqrt{|\tilde{\Lambda}| } \cos \left(2 \sqrt{|\tilde{\Lambda}| } t\right)}\times\\ &\times\exp \left[\frac{8m\sqrt{|\tilde{\Lambda}| } a \left(-i \sqrt{|\tilde{\Lambda }|}+\sigma \tan \left(\sqrt{|\tilde{\Lambda }|} t\right)-\sigma \cot \left(\sqrt{|\tilde{\Lambda}| } t\right)\right)}{2 \sqrt{|\tilde{\Lambda}| } \cot \left(2 \sqrt{|\tilde{\Lambda} } t\right)+4 i \sigma }\right].
\end{aligned}\end{equation}
\end{widetext}
The expectation value of the scale factor as a function of the cosmic time $\tau$ is then given by
\begin{equation}
\left<a\right>\!(\tau)=\frac{4\sigma^2\sin^2{\left(2\sqrt{|\tilde{\Lambda}|}\tau\right)}+|\tilde{\Lambda}|\cos^2{\left(2\sqrt{|\tilde{\Lambda}|}\tau\right)}}{16m|\tilde{\Lambda}|\sigma}.
\end{equation}
This shows an oscillatory behavior as predicted by the classical theory \cite{pitelli1}. Once again the big-bang singularity is smoothed and the expectation value of the scale factor is an even function of the cosmic time.
\subsection{Case $w=-1$ (dark energy):}
\subsubsection{$\Lambda=0$:}
In this case $x=\sqrt{2m}a$ and $\psi(a,t)=(2m)^{1/4}a^{-1/2}y(x,t)$ by Eq. (\ref{transf. variables}). By Eq. (\ref{general}), the wave function $\psi(a,t)$ becomes
\begin{equation}
\psi(a,t)=\frac{\sqrt{2m}i^{-1}}{4\gamma t -i}e^{i\frac{m a^2}{2t}}e^{-\frac{2m a^2}{16 \gamma t^2-4i t}}.
\end{equation}
The expectation value of the scale factor is given by
\begin{equation}
\left<a\right>\!(t)=\frac{\int_0^{\infty}{a\left|\psi(a,t)\right|^2\frac{da}{a^{-1}}}}{\int_0^{\infty}{\left|\psi(a,t)\right|^2\frac{da}{a^{-1}}}}=\frac{\sqrt{\pi}}{4\sqrt{m\gamma}}\sqrt{1+16\gamma^2t^2}.
\end{equation}
The proper time $\tau$ can be found from $d\tau=a^{-1}dt$. It is given by
\begin{equation}
\tau=\sqrt{\frac{m}{\pi \gamma}}\text{arcsinh}\left(4\gamma t\right).
\end{equation}
Therefore, the expectation value of the scale factor as a function of the cosmic time has the form
\begin{equation}
\left<a\right>\!(\tau)=\frac{\sqrt{\pi}}{4\sqrt{m\gamma}}\cosh{\left(\sqrt{\frac{\pi \gamma}{m}}\tau\right)}.
\label{w=-1}
\end{equation}
This inflationary universe is what we expect from a fluid with equation of state $p=-\rho$.
\subsubsection{$\Lambda\neq 0$:}
In this case we have
\begin{equation}
-\frac{\partial y(x,t)}{\partial x^2}+\left[\tilde{\Lambda}-\frac{1}{4x^2}\right]y(x,t)=i\frac{\partial y(x,t)}{\partial t},
\end{equation}
The only effect of $\tilde{\Lambda}$ is to shift the energy, i.e., if $y(x,t)$ is a solution of the above equation with $\tilde{\Lambda}=0$, then $y(x,t)e^{-i \tilde{\Lambda}t}$ is a solution of the above equation with $\tilde{\Lambda}\neq0$. Therefore, the scale factor does not change and it is still given by Eq. (\ref{w=-1}).
\acknowledgments
This work was partially supported by FAPESP Grant. No. 2013/09357-9.
\section{Conclusions}
In this paper, the $(1+1)$-dimensional FRW universe filled with a perfect fluid in HL theory has been quantized. The resulting Schr\"odinger equation of the universe has some interesting properties. For a radiation fluid, it corresponds to a free particle equation when the cosmological constant is null and has a Liouville potential when $\Lambda\neq 0$. For all other kinds of fluids, the Schr\"odinger equation have a potential composed by a characteristic inverse square potential, not depending on the EoS, plus a regular polynomial (when $\Lambda\neq 0$) which depends on the equation of state.
A few solvable cases were studied and the evolution of the scale factor was found. Consistency with the classical predictions found in Ref. \cite{pitelli1} was proved. In all cases, the universe bounces around the big-bang singularity $a=0$ and tends to the classical universe when the cosmological time $\tau$ goes to infinity. This also happens in usual quantum cosmology \cite{alvarenga}. In this way, in GR and in HL theory, quantization seems to smooth out the big-bang singularity at $a=0$, while still retaining the classical behavior as the universe becomes asymptotically large.
|
\section{Introduction}
Let $\Gamma$ be a discrete non-elementary subgroup of $\mathbf{PU}(1,n)$, and let $\Lambda_\Gamma$ be its limit set; this is a subset of the boundary of the complex hyperbolic $n$-space, $\partial \mathbf{H}_{\mathbf{C}}^n$. According to a theorem of Bishop and Jones (\cite{BishopJones}, see also \cite{Stratmann}), the Hausdorff dimension of the conical limit set $\Lambda_\Gamma^c$, with respect to a Gromov metric on the boundary, is equal to the Growth exponent $\delta_\Gamma$. According to Bishop and Jones, the proof of this theorem ``uses nothing but the definitions and a few simple properties of M\"{o}bius transformations''. To be precise, Bishop and Jones deal only with the real hyperbolic space, but their theorem was generalized further to the case of pinched curvature by Paulin \cite{PaulinBJ}.
The reader should bear in mind that the operation of $\mathbf{PU}(1,n)$ on the boundary is \emph{conformal} with respect to a given Gromov metric. Conformality usually makes questions related to dimension of sets or measures more tractable than their non-conformal counterparts. In the presence of non-conformality, various phenomena may arise, we refer the reader to \cite{ChenPesin}.
In this paper, we wish to demonstrate the usefulness of the celebrated Ledrappier-Young formula in studying Hausdorff dimension of limit sets in non-conformal setting. More precisely, we are interested in estimating the Hausdorff dimension of $\Lambda_\Gamma^c$ with respect to a \emph{spherical} metric on the boundary.
In fact looking at the boundary $\partial \mathbf{H}_{\mathbf{C}}^n$ (minus some point) endowed with the Gromov metric (resp. the spherical metric) amounts to looking at the classical Heisenberg space of height $n-1$, $\mathcal{H}_{n-1}$ (resp. the Euclidean space $\mathbf{R}^{2n-1}$). The Heisenberg space $\mathcal{H}_{n-1}$ has Hausdorff dimension $2n$ whereas Euclidean space $\mathbf{R}^{2n-1}$ has Hausdorff dimension $2n-1$. The main question we are interested in is thus a special case of the more general ``Gromov problem'' of relating the (Hausdorff) dimension of some given subset $A$ of the Heisenberg space $\mathcal{H}$ endowed with the Heisenberg metric, $\mathrm{dim}_{\mathcal{H}}(A)$, to the dimension $\mathrm{dim}_E(A)$ of $A$ with respect to the Euclidean metric.
This general problem has been worked out by Balogh \emph{et al.} in \cite{BaloghTyson}. Let us state their result now (see also theorem \ref{th.balogh} below). If we denote by $\delta$ the dimension of $A$ with respect to the Heisenberg metric, then the following sharp inequalities hold:
\[ \sup \left\{ \delta-1,\frac{\delta}{2} \right\} \leq \mathrm{dim}_E(A) \leq \inf \left\{ \delta,n-1+\frac{\delta}{2} \right\} \text. \]
Here, sharpness means that these inequalities cannot be improved without further assumptions. In this paper we are going to improve on the lower bound -- this is usually the difficult part in dealing with Hausdorff dimension -- when $A$ is some special kind of fractal set, namely the limit set of a discrete subgroup $\Gamma$ of $G$ satisfying some mild hypotheses.
We will assume that $\Gamma$ is Zariski-dense and has finite Bowen-Margulis-Sullivan measure. I should emphasize that I am not able to prove an exact formula in this general setting. Indeed, we obtain the following
\begin{theoremA}[theorem \ref{th.lower.bound}]
Let $\Gamma$ be a non-elementary discrete subgroup of $\mathbf{PU}(1,n)$, Zariski-dense, with finite BMS measure. If $\mu$ is some Patterson-Sullivan measure of exponent $\delta_\Gamma$, then for $\mu$-almost every $\xi$,
\[ \delta_\Gamma - \frac{1}{2} \mathrm{dim}(\lambda,Z) \leq \underline{\dim}(\mu,\xi) \]
where the lower pointwise dimension is with respect to the \emph{spherical} metric on the boundary.
The same inequality holds if we replace $\underline{\dim}(\mu,\xi)$ with $\mathrm{dim}_E(\Lambda_\Gamma^c)$.
\end{theoremA}
Here, $\underline{\dim}(\mu,\xi)$ is the lower pointwise dimension of $\mu$ at $\xi$, \emph{i.e.}
\[ \underline{\dim}(\mu,\xi) = \liminf_{\rho \to 0} \frac{\log \mu(B(\xi,\rho))}{\log \rho} \]
where $B(\xi,\rho)$ is the spherical ball of radius $\rho$ and centre $\xi$. The number $\mathrm{dim}(\lambda,Z)$ can be interpreted as the (almost sure) dimension of Patterson-Sullivan along chains (\emph{i.e.} boundaries of complex geodesics), with respect to a Gromov metric. See theorem C below in this introduction, and lemma \ref{lemma.def.dim} \emph{infra} for the precise definition of $\mathrm{dim}(\lambda,Z)$.
So the general lower bound
\[ \sup \{ \delta_\Gamma - 1,\frac{\delta_\Gamma}{2} \} \leq \underline{\dim}(\mu,\xi) \]
(which holds by virtue of Balogh \emph{et al.} inequality) is made more precise; indeed the number $\dim(\lambda,Z)$ lies somewhere between $0$ and $\inf\{\delta_\Gamma,2\}$.
We actually improve slightly on the general result, indeed we prove
\begin{theoremB}[corollary \ref{cor.referee}]
Let $\Gamma$ be a non-elementary discrete group of $G$, Zariski-dense and geometrically finite; assume furthermore that $\Gamma$ is \emph{not} a lattice. Then $\dim(\lambda,Z)<2$. In particular, we get the strict inequality
\[ \delta_\Gamma - 1 < \mathrm{dim}_E(\Lambda_\Gamma) \]
where $\mathrm{dim}_E(\Lambda_\Gamma)$ is the Hausdorff dimension of the limit set with respect to the Euclidean metric on the boundary.
\end{theoremB}
Of course, this is better than the general inequality only if $\delta_\Gamma \geq 2$.
As an obvious corollary of theorem A we get the fact that if $\dim(\lambda,Z)=0$, then the Hausdorff dimension of the limit set $\Lambda_\Gamma$ must be \emph{equal} to $\delta_\Gamma$. In general, computing $\dim(\lambda,Z)$ seems to be a difficult problem. We will define a class of Schottky subgroups (which we call ``Schottky subgroups in good position'') and prove the following
\begin{theoremC}[corollary \ref{cor.schottky}]
Let $\Gamma$ be a Schottky subgroup in good position in $\mathbf{PU}(1,n)$. Then $\dim(\lambda,Z)=0$; in particular, the Hausdorff dimension of $\Lambda_\Gamma$, with respect to the spherical metric, is equal to $\delta_\Gamma$.
\end{theoremC}
As an intriguing consequence, we get the fact that if $\Gamma$ is a Schottky subgroup in good position, then $\delta_\Gamma \leq 2(n-1)$ (because by virtue of Balogh et al inequality, one must then have $\delta_\Gamma \leq \frac{\delta_\Gamma}{2}+n-1$). Wether one should expect general (Zariski dense) Schottky subgroups to satisfy this inequality seems to be an interesting question.
In proving theorems A, B and C, we use the following version of Ledrappier-Young's well-known formula.
\begin{theoremD}[theorem \ref{th.additivity}]
Let $G$ be a metric locally compact second countable group which acts in a Borel way on a standard Borel space $X$ with uniformly discrete stabilizers. Let $H$ be a closed normal subgroup of $G$ and assume that the metric group $G/H$ has the Besicovitch covering property. Let $\lambda$ be a Borel probability measure on $X$. Assume that the hypotheses stated at the beginning of section \ref{section.dimension} are satisfied, so that we may define the dimension of $\lambda$ along $G$ and $H$, $\mathrm{dim}(\lambda,G)$ and $\mathrm{dim}(\lambda,H)$ respectively, as well as the transverse dimension $\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,G/H)$. Then the following holds:
\[ \mathrm{dim}(\lambda,G)=\mathrm{dim}(\lambda,H)+\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,G/H) \text. \]
\end{theoremD}
We do not get into details as they are rather technical, see sections 2 to 4. We will apply this theorem to $X=\Gamma \backslash \mathbf{PU}(1,n)$, $G=N$ in some Iwasawa decomposition $\mathbf{PU}(1,n)=KAN$, $H$ is the centre of $G$, and $\lambda$ is the Bowen-Margulis-Sullivan measure on $X$. In particular, $\mathrm{dim}(\lambda,N)$ is equal to the growth exponent $\delta_\Gamma$.
Let us describe our strategy in proving theorem A. Let $\mu$ be the Patterson-Sullivan measure (of exponent $\delta_\Gamma$) associated with $\Gamma$ (recall that $\Gamma$ has finite BMS measure, so $\mu$ is essentially unique). We use the fact that the boundary $\partial \mathbf{H}_{\mathbf{C}}^n$ is the one-point compactification of the Heisenberg group and we look at the way Patterson-Sullivan measure decomposes in Heisenberg space, with respect to the natural fibration of this space along the central direction.
This is where the Ledrappier-Young formula enters the scene. According to this formula, the dimension of Patterson-Sullivan measure, with respect to the Gromov metric on the boundary (or, equivalently, with respect to the Heisenberg metric on the Heisenberg group) is equal to the sum $\mathrm{dim}(\lambda,Z)+\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,N/Z)$ where $\mathrm{dim}(\lambda,Z)$ can be interpreted as the dimension of Patterson-Sullivan along the central direction, and $\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,N/Z)$ is the dimension of Patterson-Sullivan ``transverse'' to this central direction.
We must then use this information to estimate the dimension of Patterson-Sullivan, now with respect the to spherical metric. Easy computations show that the ``transverse dimension'' is left unmodified, whereas the dimension along the central direction is divided by 2. The Ledrappier-Young formula holds no longer, but super-additivity of dimension does (proposition \ref{prop.additivity.1}). That's why in the end all we get is a lower bound.
The plan of this paper is as follows. In section 2 we sketch the basic theory of conditional measures along a group operation. In section 3 we define the dimension of conditional measures along a group operation, as well as transverse dimension (with respect to some normal subgroup) and we study the elementary relations between these numbers. In section 4 we state and prove a version of Ledrappier-Young formula in the setting of our theory. The most important result of sections 2 to 4 is theorem \ref{th.additivity}. Then in sections 5 and 6 we recall some facts of complex hyperbolic geometry and Patterson-Sullivan theory. In section 7 we apply our Ledrappier-Young theorem to study Hausdorff dimension of limit sets with respect to the spherical metric, as explained above. In section 8 we construct Schottky subgroups in good position of $\mathbf{PU}(1,n)$. In section 9 we discuss our results.
Most of our notations are standard. If $\mu$ is some Borel measure on some metric space $X$, we let
\[ \underline{\dim}(\mu,x) = \liminf_{\rho \to 0} \frac{\log \mu(B(x,\rho))}{\log \rho}, \quad \overline{\dim}(\mu,x) = \limsup_{\rho \to 0} \frac{\log \mu(B(x,\rho))}{\log \rho} \text. \]
If $X,Y$ are measurable spaces (recall that a measurable space is a set endowed with a $\sigma$-algebra of subsets) and if $\phi : X \to Y$ is a measurable map, the push-forward of a measure $\lambda$ on $X$ is the measure $\phi_* \lambda$ defined by $\phi_* \lambda (B)=\lambda(\phi^{-1}(B))$ for all measurable subset $B$ of $Y$. If $\phi$ is a measurable isomorphism and $\mu$ a measure on $Y$, we denote the push-forward $(\phi^{-1})_* \mu$ by $\phi^* \mu$.
If $G$ is a group, the right translation $h \mapsto hg^{-1}$ by some element $g \in G$ is denoted by $R_g$.
I would like to thank the referee for their useful remarks and for pointing out (and correcting) a mistake in proposition 38. I also would like to thank my thesis advisor Jean-Fran\c{c}ois Quint for suggesting this problem and for his constant support.
\section{Disintegration of a measure along a group operation}
In this section we set up the ``language'' of conditional measures along a group operation. This is a very useful, albeit slightly technical, way of looking at things. The knowledgeable reader might want to skip it and jump straight into section 4, where the Ledrappier-Young formula we need is stated and proved.
On the other hand the lay man (or woman) interested in knowing all the details could read the definitions below bearing in mind the following setting: $G$ is the real line acting in the usual way (i.e. as a linear flow) on the torus $X$. The stabilizers may be non-trivial but they are certainly discrete. We are going to disintegrate some finite Borel measure $\lambda$ carried by $X$ along the operation of $G$ (it is very important to note that we do not expect $\lambda$ to be $G$-invariant in any way). As is well known, if this operation has dense orbits (i.e. it is an irrational flow) the elementary theory of disintegration does not suit our needs. This is why we need the more sophisticated concept of conditional measures along a group operation.
In defining them, one is basically looking at some suspension of the dynamical systems ``$G$ acting on $X$'' in order to get rid of the troubles caused by ``wild orbits'' (to be more specific, the real concern is standard Borelness of the quotient space). This is why we introduce the so-called ``lacunary sections'' $\Sigma$. They allow us to lift our measure $\lambda$ to some quotient space $G \times \Sigma$ where $G$ acts \emph{smoothly} (\emph{i.e.} the quotient space is standard Borel). We lose finiteness of $\lambda$, while retaining $\sigma$-finiteness. We may then disintegrate the lifted measure appropriately, and check that the conditional measures thus obtained are indeed ``canonical'' in some way (i.e. they do not depend on the choice of $\Sigma$).
The reader looking for a more detailed account may want to look at our doctoral thesis \cite{thesis}.
\subsection{Definition}
We begin with some definitions and basic facts. First, we state a useful theorem of Kechris.
\begin{theorem}
Let $G$ be a locally compact second countable topological group, and $X$ a standard Borel space on which $G$ acts in a Borel way. Fix a compact symmetric neighbourhood $U$ of the identity in $G$. There exists a Borel subset $\Sigma \subset X$ such that
\begin{enumerate}
\item $\Sigma$ is a $U$-lacunary section: if $x' \in \Sigma$ and $g \in U$ are such that $gx' \in \Sigma$, then $gx'=x'$;
\item $\Sigma$ is $U^2$-complete: $U^2 \Sigma=X$.
\end{enumerate}
\end{theorem}
\begin{proof}
A close inspection of the proof of the main theorem in \cite{Kechris2} yields this improved version.
\end{proof}
If there is no need to make the set $U$ explicit, we will use the phrase ``let $\Sigma$ be a complete lacunary section''.
We will make some assumptions on the operation of $G$ on $X$. If the stabilizer of any point $x \in X$ is a discrete subset of $G$, we say that the operation has \emph{discrete stabilizers}; in other words, for every $x \in X$ there is is a neighbourhood $V$ of identity in $G$ such that the only $g \in V$ which fixes $x$ is the identity. If the neighbourhood $V$ can be chosen independently of $x$, we say that the operation has \emph{uniformly discrete stabilizers}.
Now let $X$ be a standard Borel space, $\lambda$ a $\sigma$-finite Borel measure on $X$, and $G$ a locally compact second countable topological group acting on $X$ in a Borel way with discrete stabilizers. We recall, following \cite{BenoistQuint}, how one can disintegrate the measure $\lambda$ along this group operation.
Let $\Sigma$ be a complete lacunary section. The map
\[ a_\Sigma : G \times \Sigma \to X\quad (g,x') \mapsto gx' \]
has countable fibers. Hence, we can define a Borel measure $a_\Sigma^* \lambda$ on $G \times \Sigma$ in the following way:
\[ a_\Sigma^* \lambda = \int d\lambda(x) \mathrm{Card}_{a_\Sigma^{-1}(x)} \]
where $\mathrm{Card}_{a_\Sigma^{-1}(x)}$ is the counting measure on the fiber $\{(g,x')\ ;\ gx'=x\}$. It is not difficult to check that $a_\Sigma^*\lambda$ is a $\sigma$-finite measure. We may thus choose some finite measure on $G \times \Sigma$, equivalent to $a_\Sigma^* \lambda$, and push it down $\Sigma$ (through the projection $G \times \Sigma \to \Sigma$) to get a finite Borel measure $\lambda_\Sigma$ (which we call a pseudo-image of $\lambda$). We may then disintegrate $a_\Sigma^* \lambda$ over $\lambda_\Sigma$:
\[ a_\Sigma^* \lambda = \int \mathrm{d} \lambda_\Sigma(x')\ \sigma_\Sigma(x') \otimes \delta_{x'} \]
(where $\delta_{x'}$ is the Dirac measure supported on $\{x'\}$).
The conditional measures $\sigma_\Sigma(x')$ are (almost everywhere) uniquely defined up to a scalar constant. Replacing $\lambda_\Sigma$ with another pseudo-image amounts to replacing $\sigma_\Sigma(x')$ (for almost every $x'$) with some multiple $c(x') \sigma_\Sigma(x')$.
The following result is basic to the theory. We keep the standing notations and assumptions. Also, we denote by $\mathcal M_\sigma^1(G)$ (resp. ${\mathcal{M}}_{\mathrm{r}}^{\mathrm{1}}(G)$) the space of projective classes of $\sigma$-finite (resp. Radon) non-zero measures on $G$. If $\mu$ is a non-zero measure, the projective class $[\mu]$ of $\mu$ is the equivalence class
\[ [\mu] = \{ t \mu\ ;\ t>0\}\text. \]
The space $\mathcal M_{\mathrm{r}}(G)$ of all non-zero Radon measures on $G$ has a natural Borel structure (generated by narrow topology), and we endow the quotient space ${\mathcal{M}}_{\mathrm{r}}^{\mathrm{1}}(G)$ with the quotient Borel structure.
\begin{proposition}[\cite{BenoistQuint}, proposition 4.2] \label{prop.benoistquint}
There is a mapping $\sigma_{G,\lambda} : X \to \mathcal M^1_\sigma(G)$ with the following property: for any complete lacunary section $\Sigma$, there is a conegligible subset $X' \subset X$ such that for almost every $x' \in \Sigma$ and any $g \in G$, if $gx'$ belongs to $X'$ then
\[ R_g^*\sigma_{G,\lambda}(gx')=[\sigma_\Sigma(x')] \text. \]
The mapping $\sigma_{G,\lambda}$ is unique up to a negligible set. Furthermore, $\sigma_{G,\lambda}$ is essentially $G$-equivariant, \emph{i.e.} there is a conegligible set $X' \subset X$ such that if $x$ and $gx$ belong to $X'$, then
\[ \sigma_{G,\lambda}(gx)=(R_g)_* \sigma_{G,\lambda}(x) \text. \]
If $\lambda$ is finite, then $\sigma_{G,\lambda}$ maps $X$ into ${\mathcal{M}}_{\mathrm{r}}^{\mathrm{1}}(G)$, and is Borel.
\end{proposition}
Note that for $x \in X$, $\sigma_{G,\lambda}(x)$ is not, strictly speaking, a measure on $G$. Nonetheless, in many situation there is no problem in dealing with $\sigma_{G,\lambda}(x)$ as if it were a genuine measure. Hopefully, this abuse of language will not cause too much harm to the reader.
The mapping $\sigma_{G,\lambda}$ will be called the \emph{disintegration of $\lambda$ along $G$}.
We will also need the following fact.
\begin{lemma} \label{lemma.finiteness}
Assume the operation of $G$ on $X$ has uniformly discrete stabilizers. Then for any lacunary section $\Sigma$ and any compact subset $K$ of $G$, the measure $a^*_\Sigma \lambda (K \times \Sigma)$ is finite.
\end{lemma}
\begin{proof}
Let $U$ be some relatively compact symmetric neighbourhood of the identity such that $\Sigma$ is $U^2$-lacunary. Since the operation has uniformly discrete stabilizers, we may shrink $U$ so that for any $x \in X$, the only $g$ in $U^2$ such that $gx=x$ is the identity. We deduce that for any $x \in X$, there is at most one pair $(g,x') \in U \times \Sigma$ such that $gx'=x$. Consequently, $a_\Sigma^* \lambda(U \times \Sigma) \leq 1$, and also $a_\Sigma^* \lambda (gU \times \Sigma) \leq 1$ for any $g \in G$. Now it becomes clear that $a_\Sigma^* \lambda (K \times \Sigma)$ must be finite for any compact subset $K$ of $G$.
\end{proof}
We state another lemma which we will use freely. We skip the easy proof.
\begin{lemma}
For $\lambda$-almost every $x \in X$, the identity element of $G$ belongs to the support of $\sigma_{G,\lambda}(x)$.
\end{lemma}
If $\Gamma$ is some discrete subgroup of finite Bowen-Margulis-Sullivan measure of $\mathbf{PU}(1,n)$ (say), let $X=\Gamma \backslash \mathbf{PU}(1,n)$, and let $G$ be the $N$ group in the Iwasawa decomposition $\mathbf{PU}(1,n)=KAN$. If we disintegrate the Bowen-Margulis-Sullivan measure $\lambda$ on the frame bundle $X$ along $N$, what we get is exactly the so-called ``horospheric measures'' (as seen on $N$). See \cite{Roblin}.
\subsection{Transitivity of disintegration}
We keep the notations and assumptions of the previous subsection. Let also $H$ be a closed subgroup of $G$. Recall that the operation of $G$ on $X$ has discrete stabilizers; obviously, the operation of $H$ on $X$ has discrete stabilizers as well. The disintegration of $\lambda$ along $G$ and $H$, respectively, gives two mappings
\[ \sigma_{G,\lambda} : X \to \mathcal M_\sigma^1(G),\quad \sigma_{H,\lambda} : X \to \mathcal M_\sigma^1(H) \text. \]
For any $x \in X$, we may also disintegrate $\sigma_{G,\lambda}(x)$ along the operation of $H$ on $G$ by left translation, thus obtaining a mapping
\[ \sigma_{H,\sigma_{G,\lambda}(x)} : G \to \mathcal M_\sigma^1(H) \]
(\emph{stricto sensu} we are disintegrating some measure in the projective measure class $\sigma_{G,\lambda}(x)$ and the disintegration does not depend on the choice of this measure).
The meaning of the following proposition should be intuitive. The proof, however, is rather lengthy and technical. We skip it and refer the reader to our thesis, \cite{thesis}, section 2.1.3.
\begin{proposition} \label{prop.transitivity}
For $\lambda$-almost every $x \in X$ and $\sigma_{G,\lambda}(x)$-every $g \in G$, the following holds:
\[ \sigma_{H,\sigma_{G,\lambda}(x)}(g)=\sigma_{H,\lambda}(gx) \text. \]
\end{proposition}
\section{Dimension of conditional measures} \label{section.dimension}
We do not pursue the most general theory of conditional measures along group operations any further. From now on, we will deal with ``hyperbolic transformations'' of the space $X$ that are equivariant with respect to $G$. This added structure allows us to define the dimension of conditional measures and prove some useful results, the most important of which is theorem 23 in section 4 below.
Let us state the standing hypotheses in this section: $G$ is a locally compact second countable topological group, acting in a Borel way on a standard Borel space $X$, with discrete stabilizers. The space $X$ carries a Borel probability measure $\lambda$. We disintegrate $\lambda$ along $G$, thus getting a mapping
\[ \sigma : X \to \mathcal M_{\mathrm{r}}^1(G) \text. \]
Now we make some further assumptions. Namely, we assume that we are given a Borel automorphism $\alpha_X : X \to X$ as well as a group automorphism $\alpha_G : G \to G$ such that the following conditions hold:
\begin{enumerate}
\item For every $x \in X$ and $g \in G$, $\alpha_X(gx)=\alpha_G(g) \alpha_X(x)$;
\item The automorphism $\alpha_X : X \to X$ leaves invariant the measure $\lambda$ (\emph{i.e.} the push-forward $(\alpha_X)_* \lambda$ is equal to $\lambda$);
\item The group $G$ is endowed with a compatible metric $d$ which is right $G$-invariant and such that $\alpha_G$ acts on $G$ as a contracting similitude, \emph{i.e.} there is some real constant $\alpha<1$, such that
\[ d(\alpha_G g,\alpha_G h)=\alpha d(g,h) \]
for any $g,h \in G$.
\item The measure $\lambda$ is $\alpha_X$-ergodic, \emph{i.e.} if $A$ is some Borel subset of $X$ such that $\alpha_X A=A$, then either $\lambda(A)=0$ or $\lambda(A)=1$.
\end{enumerate}
The reader should have in mind the following picture: $\alpha_X$ is some hyperbolic automorphism of $X$, and $G$ parametrizes a sub-foliation of the stable foliation associated with this hyperbolic automorphism. Starting from section 5, we will apply the ongoing theory to the following objects :
\begin{itemize}
\item $G$ is some connected closed normal subgroup of $N$ in the Iwasawa decomposition $\mathbf{PU}(1,n)=KAN$, acting on the right on the quotient space $X=\Gamma \backslash \mathbf{PU}(1,n)$ (where $\Gamma$ is some discrete subgroup of $\mathbf{PU}(1,n)$ with finite Bowen-Margulis-Sullivan measure); in fact $G$ will be either $N$ or its centre $Z$, but the theory would apply just as well to any connected closed subgroup containing $Z$ (recall that $Z$ is also the derived subgroup of $N$);
\item $\lambda$ is the Bowen-Margulis-Sullivan measure;
\item $\alpha_X$ is some non-trivial element of the real line acting on $\Gamma \backslash \mathbf{PU}(1,n)$ as the ``frame flow'';
\item $\alpha_G$ is the corresponding Heisenberg homothety
\item the metric $d$ carried by $G$ is the usual (restriction of) Heisenberg metric on the Heisenberg group $N$.
\end{itemize}
\subsection{Basic facts}
First note the following
\begin{lemma} \label{lemma.alphaequiv}
For $\lambda$-almost every $x$,
\[ \sigma(\alpha_X x)=(\alpha_G)_* \sigma(x) \text. \]
\end{lemma}
\begin{proof}
This is a consequence of the ``uniqueness'' part in proposition \ref{prop.benoistquint}.
\end{proof}
We will need a technical definition and a few related facts.
\begin{definition}
A metric space $Y$ is called \emph{doubling} if there is a constant $N \geq 2$ such that any closed ball $B(y,r)$ can be covered by $N$ balls of radius $r/2$. The smallest of such numbers $N$ is the \emph{doubling constant} of $Y$.
\end{definition}
The relevance of this notion to our work is because of the following
\begin{lemma}
Let $G$ be a locally compact topological group endowed with a right-invariant metric, and assume that $G$ admits a group automorphism $\alpha_G$ which is a similarity transformation with similarity ratio $<1$. Then $G$ is a doubling space.
\end{lemma}
\begin{proof}
To begin with, note that any closed ball in $G$ is compact. Indeed, since $G$ is locally compact we can find a radius $\rho>0$ small enough so that the closed ball $B(e,\rho)$ is a compact set. If $R>0$ is arbitrary, we have $\alpha_G^k B(e,R) \subset B(e,\rho)$ for any $k$ large enough; since $\alpha_G$ is continuous, this implies that $B(e,R)$ is itself a compact set. So we see that any bounded closed subset of $G$ must be compact.
We now prove that $G$ is a doubling space. Since $B(e,1)$ is a compact set and $B(e,\alpha/2)$ has non-void interior, we may find a finite sequence $g_1,\ldots,g_N \in G$ such that the right translates $B(e,\alpha/2)g_i$ cover $B(e,1)$ ($i=1,\ldots,n$). Now fix some real number $R>0$ and let $k$ be the integer such that $\alpha^{k+1} \leq R < \alpha^k$. We have $B(e,R) \subset B(e,\alpha^k)$, so $\alpha_G^{-k} B(e,R) \subset B(e,1)$, and we deduce that the right translates $B(e,\alpha^{k+1}/2) \alpha^k_G (g_i)$ ($i=1,\ldots,n$) cover $B(e,R)$. Since $\alpha^{k+1} \leq R$, we see that $B(e,R)$ can be covered by $N$ balls of radius $R/2$ (recall that the metric on $G$ is right invariant). Thus $G$ is a doubling space.
\end{proof}
A metric group satisfying the hypotheses of this lemma will be called \emph{a doubling group}.
\begin{proposition}[\cite{KaRaSu}, lemma 2.2]
Let $Y$ be a doubling metric space and $\mu$ a Borel probability measure on $Y$. There is a constant $D$, depending only on the doubling constant of $Y$, such that for $\mu$-almost every $y \in Y$,
\[ \liminf_{\rho \to 0} \frac{\log \mu(B(y,\rho))}{\log \rho} \leq D \text. \]
\end{proposition}
The next lemma is both straightforward and basic to our work.
\begin{lemma} \label{lemma.def.dim}
Under the hypotheses stated at the beginning of section 3, for $\lambda$-almost every $x$, the conditional measure $\sigma(x)$ is \emph{exact-dimensional}, and its dimension does not depend on $x$. In other words, for $\lambda$-almost every $x \in X$ and $\sigma(x)$-almost every $g \in G$
\[ \lim_{\rho \to 0} \frac{\log \sigma(x)(B(g,\rho))}{\log \rho} \]
exists, is finite, and does not depend on $x$ nor on $g$.
\end{lemma}
\begin{proof}
First we prove that the limit (maybe infinite) exists for almost every $x$ when $g$ is the identity of $G$. It is enough to show that
\[ \lim_{n \to \infty} \frac{\log \sigma(x)(B(e,\alpha^n))}{n \log \alpha} \]
exists for $\lambda$-almost every $x$. Let
\[ I(x)= \log \frac{\sigma(x)(B(e,\alpha))}{\sigma(x)(B(e,1))} \text. \]
Note that $I(x) \leq 0$. By lemma \ref{lemma.alphaequiv} and an obvious induction we get
\[ \sum_{k=0}^{n-1} I(\alpha_X^{-k} x)= \log \frac{\sigma(x)(B(e,\alpha^n))}{\sigma(x)(B(e,1))} \]
($n \geq 1$). Upon dividing both sides by $n \log \alpha$ and letting $n \to \infty$, we see that
\[ \lim_{n \to \infty} \frac{\log \sigma(x)(B(e,\alpha^n))}{n \log \alpha} = \lim_{n \to \infty} \frac{1}{n \log \alpha}\sum_{k=0}^{n-1} I(\alpha_X^{-k}x) \]
where the right side exists and is equal, for $\lambda$-almost every $x \in X$, to
\[ \frac{1}{\log \alpha} \int_X I(x) \mathrm{d} \lambda(x) \in [0,+\infty] \]
by virtue of the ergodic theorem, since $\lambda$ is $\alpha_X^{-1}$-ergodic. Now the fact that this limit is almost surely finite is an obvious consequence of the previous proposition. Let $\delta\in [0,\infty[$ be this finite number, thus
\[ \lim_{\rho \to 0} \frac{\log \sigma(x)(B(e,\rho))}{\log \rho} = \delta \]
for every $x$ in some conull subset $X'$ of $X$.
It is easy to check that, since $X'$ is conull, $gx$ must belong to $X'$ for $\lambda$-almost every $x \in X$ and $\sigma(x)$-almost every $g \in G$. Also, for $\lambda$-almost every $x$ and $\sigma(x)$-almost every $g \in G$, we have
\[ \sigma(gx) = (R_g)_* \sigma(x) \]
whence
\[ \delta = \lim_{\rho \to 0} \frac{\log \sigma(gx)(B(e,\rho))}{\log \rho} = \lim_{\rho \to 0} \frac{\log \sigma(x)(B(g,\rho))}{\log \rho} \]
because the metric carried by $G$ is right-invariant. The proposition is proven.
\end{proof}
\begin{definition}
The limit in the previous lemma is called \emph{dimension of $\lambda$ along $G$}, and is denoted by $\dim(\lambda,G)$.
\end{definition}
Note the following formula:
\begin{equation} \label{eq.dimension}
\mathrm{dim}(\lambda,G) = \frac{1}{\log \alpha} \int_X \log \left( \frac{\sigma(x)(\alpha_H B)}{\sigma(x)(B)} \right) \mathrm{d}\lambda(x)
\end{equation}
for any relatively compact neighbourhood $B$ of the identity.
We now state some useful facts in the setting of doubling metric spaces.
\begin{lemma}[\cite{KaRaSu}, proposition A.2]
Let $Y$ be a doubling metric space, and $\mu$ be a finite Borel measure on $Y$. Let $A$ be some Borel subset of $Y$ such that $\mu(A)>0$. For almost every $y \in A$, we have
\[ \limsup_{\rho \to 0} \frac{\mu(A \cap B(y,\rho))}{\mu(B(y,\rho))} = 1 \text. \]
\end{lemma}
Note that contrary to the well-known density theorem of Lebesgue (which holds in euclidian space and more generally in metric spaces satisfying the Besicovitch covering property) the left side is an upper limit, not a genuine limit.
\begin{proposition} \label{prop.additivity.1}
Let $Y$ be a complete doubling metric space, $Z$ a complete separable metric space, and $\pi : Y \to Z$ a Lipschitz mapping. Let $\mu$ be a Borel probability measure on $Y$, $\nu$ the push-forward of $\mu$ through $\pi$, and
\[ \mu = \int_Z d\nu(z) \mu_z \]
be the disintegration of $\mu$ along $\pi$.
Assume that there exists a constant $\gamma \geq 0$ and a Borel mapping $\delta : Z \to [0,\infty[$ such that for $\mu$-almost every $y$ in $Y$ the following hold:
\[ \underline{\dim}(\mu_{\pi(y)},y) \geq \gamma,\quad \underline{\dim}(\nu,\pi(y)) \geq \delta(\pi(y)) \text. \]
Then for $\mu$-almost every $y \in Y$, we have
\[ \overline{\dim}(\mu,y) \geq \gamma + \delta(\pi(y)) \text. \]
If, instead of assuming that $Y$ is doubling, we assume that this space is complete separable and satisfies the Besicovitch covering property, we get the stronger conclusion
\[ \underline{\dim}(\mu,y) \geq \gamma+\delta(\pi(y)) \]
for $\mu$-almost every $y$.
\end{proposition}
\begin{proof}
See \cite{LY} lemma 11.3.1 when $Y$ satisfies the Besicovitch covering property. If $Y$ is a doubling space just copy the proof and use the previous lemma instead of Lebesgue density theorem to obtain the weaker conclusion.
\end{proof}
\begin{remark} This is a key proposition for the main result, so it may be worth taking some time to comment on this inequality. Choose some continuous mapping $f: [0,1] \to [0,1]$ whose graph has Hausdorff dimension 2, see \emph{e.g.} \cite{Wingren}. Let $\nu$ be the Lebesgue measure on $Z=[0,1]$, and $\mu$ be the push-forward of $\nu$ through the mapping $x \mapsto (x,f(x))$, so that $\nu$ is itself the push-forward of $\mu$ through the projection onto the first factor. We may disintegrate $\mu$ above $\nu$, and obviously we get
\[ \mu=\int_0^1 d\nu(x) \delta_{(x,f(x))} \]
where $\delta_{(x,y)}$ is the Dirac mass at $(x,y)$. In particular, the previous proposition amounts to the obvious inequality
\[ 2 \geq 1\text. \]
This illustrates the fact that dimension is only ``super-additive'', and we should not expect equality to hold in general, unless some significant geometric assumption is made on the measures we are looking at.
\end{remark}
In the same way one proves the following
\begin{proposition} \label{prop.additivity.2}
Let $Y$ be a complete doubling metric space, $Z$ a standard Borel space, and $\pi : Y \to Z$ a Borel mapping. Let $\mu$ be a Borel probability measure on $Y$, $\nu$ the pushforward of $\mu$ through $\pi$, and
\[ \mu = \int_Z d\nu(z) \mu_z \]
be the disintegration of $\mu$ along $\pi$.
Assume that there exists a constant $\gamma \geq 0$ such that for $\mu$-almost every $y$ in $Y$ the following holds:
\[ \underline{\dim}(\mu_{\pi(y)},y) \geq \gamma \text. \]
Then for $\mu$-almost every $y \in Y$, we have
\[ \overline{\dim}(\mu,y) \geq \gamma \text. \]
If, instead of assuming that $Y$ is doubling, we assume that this space is complete separable and satisfies the Besicovitch covering property, we get the stronger conclusion
\[ \underline{\dim}(\mu,y) \geq \gamma \]
for $\mu$-almost every $y$.
\end{proposition}
Note that here we do not assume that $Z$ is a Polish metric space, and even if this is so, $\pi$ might not be Lipschitz, so we are not just applying the previous proposition with $\delta=0$. Nonetheless the proof is similar and straightforward and we skip it.
\subsection{Monotonicity of dimension; transverse dimension}\label{subsection.monotonicity}
We keep the previous assumptions and we consider a closed subgroup $H$ of $G$ such that $\alpha_G(H)=H$, so that $\alpha_G$ induces an automorphism $\alpha_H$ of $H$. Obviously, the dimension of $\lambda$ along $H$ is well-defined; the aim of this subsection is to compare the two dimensions.
The following proposition is intuitively clear.
\begin{proposition} \label{prop.monotonicity}
We have
\[ \mathrm{dim}(\lambda,G) \geq \mathrm{dim}(\lambda,H) \text. \]
\end{proposition}
\begin{proof}
Let us disintegrate $\lambda$ along $G$ and $H$, thus obtaining mappings
\[ \sigma_G : X \to \mathcal {\mathcal{M}}_{\mathrm{r}}^{\mathrm{1}}(G),\quad \sigma_H : X \to \mathcal {\mathcal{M}}_{\mathrm{r}}^{\mathrm{1}}(H) \text. \]
Let $\pi : G \to G/H$ be the quotient mapping. Recall that $G/H$ is a locally compact second countable topological space, so that both $G$ and $G/H$ are standard Borel space and $\pi$ is a Borel mapping.
Fix some relatively compact neighbourhood $B$ of the identity element of $G$. For any $x \in X$, we let $\nu^B(x)$ be the Borel probability measure on $G$ obtained by conditioning $\sigma(x)$ on $B$, \emph{i.e.}
\[ \nu^B(x) = \frac{\sigma_G(x)|B}{\sigma_G(x)(B)} \text. \]
For $\lambda$-almost every $x$ in $X$, if we disintegrate the measure $\nu^B(x)$ along $\pi$, the conditional measure we get above $\pi(g)$ is, almost surely, proportional to the restriction of $\sigma_H(gx)$ on some neighbourhood of the identity element in $H$ (proposition \ref{prop.transitivity}). In particular, these conditional measures are almost surely exact dimensional, with dimension $\dim(\lambda,H)$.
By virtue of proposition \ref{prop.additivity.2}, we get that
\[ \overline{\dim}(\nu^B(x),g) \geq \dim(\lambda,H) \]
for $\nu^B(x)$-almost every $g$ in $G$. Now $B$ is a neighbourhood of the identity element in $G$, so that $\nu^B(x)$ must be exact dimensional of dimension $\dim(\lambda,G)$, and we have thus proved that
\[ \dim(\lambda,G) \geq \dim(\lambda,H) \text. \]
\end{proof}
We will now improve on this result, and introduce, following Ledrappier and Young \cite{LY}, a \emph{transverse dimension}, under the supplementary assumption that $H$ is a normal subgroup. We can endow the topological quotient group $G/H$, which is locally compact and second countable, with a natural metric. More precisely, we let
\[ d(gH,g'H) = \inf\{ d(gh,g'h)\ ;\ h \in H \} \]
and we may check that this defines a right invariant metric, such that the quotient mapping
\[ \pi: G \to G/H \]
is $1$-Lipschitz, and that the group automorphism $\alpha_{G/H} : G/H \to G/H$ induced by $\alpha_G$ is a similitude with same ratio as $\alpha_G$ itself.
We may now define the transverse to $H$ dimension of $\lambda$ along $G$. If $B$ is some relatively compact neighbourhood of identity in $G$, we let, as in the proof of the previous proposition,
\[ \nu^B(x) = \frac{\sigma(x)|B}{\sigma(x)(B)},\quad \theta^B(x)=\pi_* \nu^B(x) \]
\emph{i.e.} $\nu^B(x)$ is the probability measure obtained by conditioning $\sigma(x)$ on $B$ and $\theta^B(x)$ is the push-forward of $\nu^B(x)$ through $\pi$.
\begin{proposition} \label{prop.transverse.dimension}
There is a finite number $\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,G/H)$ such that the following holds. Let $B$ be a compact neighbourhood of identity in $G$. For $\lambda$-almost every $x \in X$, and $\theta^B(x)$-almost every $gH \in G/H$,
\[ \liminf_{\rho \to 0} \frac{\log \theta^B(x) (B(gH,\rho))}{\log \rho} = \underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,G/H) \text. \]
In other words, $\theta^B(x)$ has lower pointwise dimension equal to $\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,G/H)$ almost everywhere.
\end{proposition}
\begin{proof}
First, we prove that the lower dimension of $\theta^B(x)$ at the identity element of $G/H$ is constant $\lambda$-almost everywhere. In order to shorten notations we will write $B^T(\rho)$ for the set $\pi^{-1}(B(\pi(e),\rho))$ (where $\pi(e)=H$ is the identity element of $G/H$). Since the metric carried by $G$ is right-invariant, we have $B^T(\rho)=B(e,\rho)H$. Now let $\delta_B(x)$ be the lower dimension of $\theta^B(x)$ at the identity element, \emph{i.e.}
\[ \delta_B(x)=\liminf_{\rho \to 0}\frac{\log \theta^B(x)(B(\pi(e),\rho))}{\log \rho} = \liminf_{\rho \to 0} \frac{\log \sigma(x)(B \cap B^T(\rho))}{\log \rho} \text.\]
For $\lambda$-almost every $x$ we have (lemma \ref{lemma.alphaequiv})
\[ \delta_B(\alpha_X x)=\liminf_{\rho \to 0} \frac{\log \sigma(x)(\alpha_G^{-1} B \cap B^T(\alpha^{-1} \rho))}{\log \rho} \]
and the right side is obviously equal to $\delta_{\alpha_G^{-1} B}(x)$. Now assume for a moment that $B$ is an open ball $B(e,R)$; we then have $B \subset \alpha_G^{-1} B$, so that the relation $\delta_B(\alpha_X x)=\delta_{\alpha_G^{-1} B}x$ implies the relation $\delta_B(\alpha_X x) \leq \delta_B(x)$. By a straightforward application of Birkhov's ergodic theorem (bearing in mind the ergodicity of $\lambda$) we see that $\delta_B$ is almost surely constant, and the relation $\delta_B \circ \alpha_X = \delta_{\alpha_G^{-1} B}$ implies that the almost certain value of $\delta_B$ is equal to the almost certain value of $\delta_{\alpha_G^k B}$ for any integer $k$. Thus the almost certain value of $\delta_B$ does not depend on the radius $R$ if $B$ is the open ball $B(e,R)$; now if $B$ is any relatively compact neighbourhood of the identity, we can find radii $R'$ and $R''$ such that
\[ B(e,R') \subset B \subset B(e,R'') \]
and we get that $\delta_B$ is almost surely constant and that its almost certain value does not depend on $B$.
Denote by $\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,G/H)$ the almost certain value of $\delta_B$ for any relatively compact neighbourhood $B$ of the identity. Now let $X'$ be a conull subset of $X$ such that $\delta_B(x)=\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,G/H)$ for any relatively compact neighbourhood $B$ of the identity, and, furthermore, such that $\sigma(gx)=(R_g)_* \sigma(x)$ if $x$ and $gx$ belong to $X'$.
For any $x \in X'$ and any $g \in G$ such that $gx \in X'$, we have
\[ \underline{\dim} (\theta^B(x),\pi(g))=\liminf_{\rho \to 0} \frac{\log \sigma(x)(B \cap \pi^{-1}B(\pi(g),\rho))}{\log \rho} \text. \]
Also, $\pi^{-1}(B(\pi(g),\rho))=B^T(\rho)g$ because $H$ is normal in $G$. The previous quantity is thus equal to
\[ \liminf_{\rho \to 0} \frac{\log \sigma(gx)(Bg^{-1} \cap B^T(\rho))}{\log \rho} \text. \]
Now if $g$ belongs to $B$, the set $Bg^{-1}$ is a relatively compact neighbourhood of the identity, so that we get
\[ \underline{\dim}(\theta^B(x),\pi(g))=\delta_{Bg^{-1}}(x) \]
and the right side is equal to $\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,G/H)$. The proposition is proven.
\end{proof}
\begin{definition}
The quantity $\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,G/H)$, whose existence was proven in the previous proposition, is called ``transverse to $H$ dimension of $\lambda$ along $G$''.
\end{definition}
\begin{proposition} \label{prop.subadd}
Under the previous hypotheses, the following holds:
\[ \mathrm{dim}(\lambda,G) \geq \mathrm{dim}(\lambda,H)+\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,G/H) \text. \]
\end{proposition}
\begin{proof}
We just need to argue as in the proof of proposition \ref{prop.monotonicity}, applying proposition \ref{prop.additivity.1} in lieu of proposition \ref{prop.additivity.2}.
\end{proof}
\section{Ledrappier-Young formula}
This section is a translation of theorem C' in \cite{LY} into the language and setting of conditional measures along a group action.
\subsection{Two technical lemmata}
\subsubsection{Classical statements}
We first state two more-or-less classical lemmas, which we are going to generalize to the setting of conditional measures along a group operation. In proving the generalized results, we will use their ``classical'' counterparts.
I should mention that lemmas \ref{lemma.classical.1} and \ref{lemma.classical.2} are basic to the proof of theorem C' in (\emph{ibid.}). We are basically going to copy the arguments of Ledrappier and Young, in our language, simply replacing the ``classical'' lemmas with the ``generalized'' lemmas to follow .
Let $Y$ be some complete separable metric space satisfying Besicovitch covering property, and consider two Radon measures $\mu,\nu$. Assume, for simplicity, that $\mu$ is finite. For any $y \in Y$ and any radius $\rho > 0$, let
\[ \phi_\rho(y) = \frac{\nu(B(y,\rho))}{\mu(B(y,\rho))} \in [0,+\infty] \]
and
\[ \phi_*(y) = \inf_{\rho > 0} \phi_\rho(y) \text. \]
\begin{lemma} \label{lemma.classical.1}
The positive Borel function $- \log \phi_*$ is $\nu$-integrable, and
\[ \int - \log \phi_* \mathrm{d} \nu \leq C(Y) \mu(Y) \]
where $C(Y)$ is the Besicovitch constant of $Y$.
\end{lemma}
\begin{proof}
Let $E_t$ be the set of all those $y$ such that $\phi_*(y) < e^{-t}$ ($t > 0$). If $y$ belongs to $E_t$, there is some radius $r(y)\in ]0,1[$ such that $\nu(B(y,r(y))) < e^{-t} \mu(B(y,r(y)))$. Let, by virtue of Besicovitch covering property, $A$ be a subset of $E_t$ such that the closed balls $B(y,r(y))$ ($y \in A$) cover $E_t$ with multiplicity less than $C(Y)$. We have
\[ \nu(E_t) \leq \sum_{y \in A} \nu(B(y,r(y))) \leq e^{-t} C(Y) \mu(Y) \text. \]
Now integrate
\[ \int_0^\infty \nu(E_t) \leq C(Y) \mu(Y) \int_0^\infty e^{-t} \mathrm{d}t \]
and the left side is equal to
\[ \int - \log \phi_* \mathrm{d} \nu \]
by a classical application of Fubini theorem.
\end{proof}
\begin{lemma}[\cite{Mattila}, theorem 2.12] \label{lemma.classical.2}
If $\nu$ is absolutely continuous with respect to $\mu$, we have
\[ \lim_{\rho \to 0} \phi_\rho(y) = \frac{\mathrm{d} \nu}{\mathrm{d} \mu}(y) \]
for $\mu$-almost every $y$.
\end{lemma}
\subsubsection{Generalized statements}
We begin with the notations. Let $G$ be a locally compact second countable topological group acting in a Borel way on a standard Borel space $X$ with \emph{uniformly discrete stabilizers}. Let $H$ be a closed normal subgroup of $G$. We assume that the (metrizable) quotient group $G/H$ is endowed with a compatible metric $d$, right invariant and \emph{proper}, which means that any closed ball is a compact set. Note that we do not need to endow $G$ (nor $H$) with any metric. We denote by $B(gH,\rho)$ the open ball of radius $\rho$ in $G/H$. Let $\pi$ be the quotient map $G \to G/H$.
We make the assumption that the metric space $G/H$ satisfies the \emph{Besicovitch covering property} (see, \emph{e.g.}, \cite{Mattila}).
Now let $\lambda$ be some Borel probability measure on $X$. We disintegrate $\lambda$ along $G$ and $H$, thus obtaining maps
\[ \sigma_G : X \to \mathcal M_{\mathrm{r}}^1(G),\quad \sigma_H : X \to \mathcal M_{\mathrm{r}}^1(H) \text. \]
Fix some compact neighbourhood $A \subset B$ of the identity in $G$. For any $x \in X$ and $\rho > 0$ we let
\[ \phi_\rho(x) = \frac{\pi_* (\sigma(x)|A)(B(H,\rho))}{\pi_*(\sigma(x)|B)(B(H,\rho))} \text. \]
To shorten notations, we denote by $B^T_\rho(g)$ ($g \in G$) the inverse image $\pi^{-1}(B(\pi(g),\rho))$, so that
\[ \phi_\rho(x) = \frac{\sigma(x)(B^T_\rho(e) \cap A)}{\sigma(x)(B^T_\rho(e) \cap B)} \text. \]
Let also
\[ \phi_*(x) = \inf_{\rho > 0} \phi_\rho(x) \text. \]
We know state:
\begin{lemma} \label{lemma.technical.1}
The integral
\[ \int - \log \phi_*(x) \mathrm{d} \lambda(x) \]
is finite.
\end{lemma}
\begin{lemma} \label{lemma.technical.2}
For $\lambda$-almost every $x$, we have
\[ \lim_{\rho \to 0} \phi_\rho(x) = \frac{\sigma_H(x)(A)}{\sigma_H(x)(B)} \text. \]
\end{lemma}
\subsubsection{Proof of lemma \ref{lemma.technical.1}}
Fix some compact symmetric neighbourhood of identity $\Delta$ in $G$ such that $\Delta^4 \subset A$. Let $\Sigma$ be a $\Delta$-lacunary, $\Delta^2$-complete section to the operation of $G$ on $X$. Denote by $a$ the mapping $G \times \Sigma$, $a(g,x')=gx'$. Choose a pseudo-image $\lambda_\Sigma$ of $a^* \lambda$, and remember the notation $\sigma_\Sigma$ (\emph{cf supra}, subsection 2.1). We have
\[ \begin{split} \int -\log \phi_* \mathrm{d} \lambda & \leq \int d(a^* \lambda)(g,x') \mathbf{1}_{\Delta^2}(g) (- \log \phi_* (gx')) \\ & =\int d\lambda_\Sigma(x') \int_{\Delta^2} \mathrm{d}(\sigma_\Sigma(x'))(g) (-\log \phi_*(gx')) \text.
\end{split} \]
For $\lambda_\Sigma$-almost every $x'$ and $\sigma_\Sigma(x')$-almost every $g$, we have $\sigma(gx')=[(R_g)_* \sigma_\Sigma(x')]$. Fix such an $x'$. By Fubini theorem, using the fact that $\phi_*$ is less than $1$, we have
\[ \int_{\Delta^2} - \log \phi_*(gx') \mathrm{d}(\sigma_\Sigma(x'))(g) = \int_0^\infty \sigma_\Sigma(x') \left\{ g \in \Delta^2\ ;\ \phi_*(gx') < e^{-t} \right\} \mathrm{d}t \text. \]
Now let
\[ \tilde \phi_\rho(x',g) = \frac{\sigma_\Sigma(x')(\Delta^2 \cap B_\rho^T(g))}{\sigma_\Sigma(x')(B \Delta^2 \cap B_\rho^T(g))} \]
and
\[ \tilde \phi_* (x',g) = \inf_{\rho >0} \tilde \phi_\rho(x',g)\text{.} \]
I claim that
\[ \tilde \phi_*(x',g) \leq \phi_*(gx') \]
for $\lambda_\Sigma$-almost every $x'$ and $\sigma_\Sigma(x')$-almost every $g \in \Delta^2$. Indeed, if $g \in \Delta^2$ is such that $\sigma(gx')=[(R_g)_* \sigma_\Sigma(x')]$, we have
\[ \begin{split} \phi_\rho(gx') &= \frac{\sigma(gx')(A \cap B_\rho^T(e))}{\sigma(gx')(B \cap B_\rho^T(e))} \\
& = \frac{\sigma_\Sigma(x')(Ag \cap B_\rho^T(g))}{\sigma_\Sigma(x')(Bg \cap B^T_\rho(g))} \\ & \geq \frac{\sigma_\Sigma(x')(\Delta^2 \cap B_\rho^T(g))}{\sigma_\Sigma(x')(B \Delta^2 \cap B_\rho^T(g))=\tilde \phi_\rho(g,x')}
\end{split}
\]
since $\Delta^2 \subset Ag$ and $Bg \subset B \Delta^2$ for any $g \in \Delta^2$. The claim follows.
Now apply lemma \ref{lemma.classical.1} to get, for $\lambda_\Sigma$-almost every $x'$,
\[ \int_{\Delta^2} \mathrm{d} \sigma_\Sigma(x')(g) (-\log \tilde{\phi}_*(x',g)) \leq C(G/H) \sigma_\Sigma(x')(B \Delta^2) \]
where $C(G/H)$ is the Besicovitch constant of $G/H$. We need only integrate (with respect to $\mathrm{d} \lambda_\Sigma(x')$) to obtain
\[ \int - \log \phi_* \mathrm{d} \lambda \leq C(G/H) \int \sigma_\Sigma(x')(B \Delta^2) \mathrm{d} \lambda_\Sigma(x') \]
and the right side is finite by virtue of lemma \ref{lemma.finiteness}.
\subsubsection{Proof of lemma \ref{lemma.technical.2}}
\begin{enumerate}
\item Let $A' \subset B'$ be two relatively compact neighbourhood of the identity in $G$. Let $\Sigma$ be a lacunary section to the operation of $G$ on $X$. Recall the notation $\sigma_\Sigma$ (\emph{cf supra}). For $\lambda$-almost every $x'$ and $\sigma_\Sigma(x')$-almost every $g \in A'$, we have
\[ \lim_{\rho \to 0} \frac{\sigma_\Sigma(x')(A' \cap B_\rho^T(g))}{\sigma_\Sigma(x')(B' \cap B_\rho^T(g))} = \frac{\sigma_H(gx')(A' \cap H)}{\sigma_H(gx')(B' \cap H)} \text. \]
Indeed, this is a straightforward consequence of lemma \ref{lemma.classical.2} and proposition \ref{prop.transitivity}.
\item Now let, for $x$ in $X$,
\[ \theta(x) = \limsup_{\rho \to 0} \left| \frac{\sigma(x)(A\cap B_\rho^T)}{\sigma(x)(B \cap B_\rho^T)} - \frac{\sigma_H(x)(A \cap H)}{\sigma_H(x)(B \cap H)} \right| \text. \]
We are going to show that for any $\varepsilon > 0$, the set of all $x$ such that $\theta(x) \geq \varepsilon$ is a null set (with respect to $\lambda$).
\item Fix some radius $r>0$, small enough in a way we will make precise soon. In order to shorten notations we denote by $U$ the open ball $B(e,r)$; $U$ is a relatively compact symmetric open neighbourhood of the identity. Let $\Sigma$ be a $U$-lacunary, $U^2$-complete section to the operation of $G$ on $X$. For almost every $x' \in \Sigma$ and $\sigma_\Sigma(x')$-almost every $g$, we have $\sigma(gx')=[(R_g)_* \sigma_\Sigma(x')]$; hence, for any $\rho>0$,
\[ \frac{\sigma(gx')(A \cap B_\rho^T)}{\sigma(gx')(B \cap B_\rho^T)} = \frac{\sigma_\Sigma(x')(Ag \cap B_\rho^T(g))}{\sigma_\Sigma(x')(Bg \cap B_\rho^T(g))} \text. \]
We assume $r$ is small enough so that the sets
\[ A_{- r} = \bigcap_{g \in U^2} Ag,\quad B_{-r} = \bigcap_{g \in U^2} Bg \]
are neighbourhood of the identity. Let also
\[ A_{+r}=AU^2,\quad B_{+r}=BU^2 \]
and note that these sets are relatively compact.
\item For almost any $x' \in \Sigma$ and $\sigma_\Sigma(x')$-almost any $g \in U^2$, we have
\[ \frac{\sigma_{\Sigma}(x')(A_{-r} \cap B_\rho^T(g))}{\sigma_\Sigma(x')(B_{+r} \cap B_\rho^T(g))}
\leq \frac{\sigma(gx')(A \cap B_\rho^T)}{\sigma(gx')(B \cap B_\rho^T)} \leq \frac{\sigma_{\Sigma}(x')(A_{+r} \cap B_\rho^T(g))}{\sigma_\Sigma(x')(B_{-r} \cap B_\rho^T(g))} \text. \]
By virtue of 1. \emph{supra}, we get
\begin{align*} & \frac{\sigma_H(gx')(A_{-r} \cap H)}{\sigma_H(gx')(B_{+r} \cap H)} \leq \liminf_{\rho \to 0} \frac{\sigma(gx')(A \cap B_\rho^T)}{\sigma(gx')(B \cap B_\rho^T)} \\
\leq{} & \limsup_{\rho \to 0} \frac{\sigma(gx')(A \cap B_\rho^T)}{\sigma(gx')(B \cap B_\rho^T)} \leq \frac{\sigma_H(gx')(A_{+r} \cap H)}{\sigma_H(gx')(B_{-r} \cap H)} \text.
\end{align*}
\item Let
\[ \theta_r'(x) = \left| \frac{\sigma_H(x)(A \cap H)}{\sigma_H(x)(B \cap H)} - \frac{\sigma_H(x)(A_{-r} \cap H)}{\sigma_H(x)(B_{+r} \cap H)} \right| + \left| \frac{\sigma_H(x)(A \cap H)}{\sigma_H(x)(B \cap H)}-\frac{\sigma_H(x)(A_{+r} \cap H)}{\sigma_H(x)(B_{-r} \cap H)} \right| \text. \]
According to 4. we know that for almost every $x' \in \Sigma$ and $\sigma_\Sigma(x')$-almost every $g \in U^2$, we have $\theta(gx') \leq \theta_r'(gx')$. Since $\Sigma$ is $U^2$-complete, we obtain
\begin{align*} \lambda \{ x \in X \ ;\ \theta(x) \geq \varepsilon \} & \leq a^* \lambda \{ (g,x') \in U^2 \times \Sigma \ ;\ \theta(gx') \geq \varepsilon \} \\
& \leq a^* \lambda \{ (g,x') \in U^2 \times \Sigma\ ;\ \theta_r'(gx') \geq \varepsilon \} \\
& = \int_{\theta_r' \geq \varepsilon} d\lambda(x) \kappa(x)
\end{align*}
where $\kappa(x)$ is the number of all $(g,x') \in U^2 \times \Sigma$ such that $gx'=x$. This number is bounded uniformly in $x$ by some constant $K$ independent of $r$ small enough; we skip the proof of this easy fact, which is a consequence of the facts that the stabilizers are uniformly discrete and that $G$ is a doubling group.
All in all, we have
\[ \lambda\{ x\ ;\ \theta(x) \geq \varepsilon \} \leq K\lambda\{ x\ ;\ \theta_r'(x) \geq \varepsilon \} \]
for any $r>0$ small enough. Now $\sigma_H(x)$ is (almost surely) a Radon measure, and $A,B$ are relatively compact open sets; whence
\[ \lim_{r \to 0} \theta'_r(x)=0 \]
almost surely. The lemma is thus proven.
\end{enumerate}
\subsection{Additivity of dimension} \label{sect.LY}
We keep the notations and hypotheses set at the beginning of section \ref{section.dimension}. We also consider a closed normal subgroup $H$ of $G$ that is $\alpha_G$-invariant, \emph{i.e.} $\alpha_G H=H$. Recall that the quotient group $G/H$ is endowed with a natural metric, see section \ref{subsection.monotonicity}. We denote by $\pi$ the quotient map $G \to G/H$. Now we make two supplementary hypotheses:
\begin{enumerate}
\item The operation of $G$ on $X$ has uniformly discrete stabilizers.
\item The metric space $G/H$ satisfies the Besicovitch covering property.
\end{enumerate}
Under these hypotheses, we now state the main result of this section.
\begin{theorem} \label{th.additivity}
The following equality holds:
\[ \mathrm{dim}(\lambda,G)=\mathrm{dim}(\lambda,H)+\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,G/H) \text. \]
\end{theorem}
\begin{proof}
Due to proposition \ref{prop.subadd}, we just need to establish
\[ \mathrm{dim}(\lambda,G) - \mathrm{dim}(\lambda,H) \leq \underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,G/H) \text. \]
Fix some open relatively compact neighbourhood $B$ of the identity in $G$. We must prove that the push-forward measure
\[ \pi_* \left( \frac{\sigma_G(x)|B}{\sigma_G(x)(B)} \right) \]
has lower dimension $\geq \mathrm{dim}(\lambda,G)-\mathrm{dim}(\lambda,H)$ at the identity of $G/H$, for $\lambda$-almost every $x$ (proposition \ref{prop.transverse.dimension}).
Introduce the ``transverse ball''
\[ B_\rho^T(g)=\pi^{-1}(B(\pi(g),\rho)),\quad \rho>0,\quad g\in G \text. \]
Note that $B_\rho^T(gg')=B_\rho^T(g)g'$ because the metric on $G$ is right-invariant. To shorten notations we denote the transverse ball at identity $B^T_\rho(e)$ by $B_\rho^T$.
We only need to prove that for $\lambda$-almost every $x$,
\[ \liminf_{n \to \infty} \frac{1}{n \log \alpha} \log \left( \frac{\sigma(x)(B \cap B_{\alpha^n}^T)}{\sigma(x)(B)} \right) \geq \mathrm{dim}(\lambda,G)-\mathrm{dim}(\lambda,H)\text. \]
The key of the argument is the relation
\[ \frac{\sigma(x)(B \cap B_{\alpha^n}^T)}{\sigma(x)(B)}
= \frac{\sigma(x)(B \cap B_{\alpha^n}^T)}{\sigma(x)(\alpha_G B \cap B_{\alpha^n}^T)} \times \frac{\sigma(x)(\alpha_G B)}{\sigma(x)(B)} \times \frac{\sigma(x)(\alpha_G B \cap B_{\alpha^n}^T)}{\sigma(x)(\alpha_G B)}\text{.} \]
This relation implies, by an obvious induction (recall that $\sigma(\alpha x)=\alpha_* \sigma(x)$ almost everywhere and $\alpha_{G/H} B_\rho^T=B_{\alpha \rho}^T$) that for any $p, n \geq 1$,
\[ \begin{split} \frac{\sigma(x)(B \cap B_{\alpha^n}^T)}{\sigma(x)(B)} & = \left( \prod_{i=0}^{p-1} \frac{\sigma(\alpha_X^{-i}x)(B \cap B_{\alpha^{n-i}}^T)}{\sigma(\alpha_X^{-i}x)(\alpha B \cap B_{\alpha^{n-i}}^T)} \times \frac{\sigma(\alpha_X^{-i}x)(\alpha B)}{\sigma(\alpha_X^{-i}x)(B)} \right)\times\\& \frac{\sigma(\alpha_X^{-p}x)(B \cap B_{\alpha^{n-p}}^T)}{\sigma(\alpha_X^{-p}x)(B)} \text.
\end{split}
\]
If $p \leq n$, the last factor on the right-hand side is obviously less than $1$. Now let
\[ \phi_\rho(x) = \log \frac{\sigma(x)(\alpha_G B \cap B_\rho^T)}{\sigma(x)(B \cap B_\rho^T)}, \tau(x) = \log \frac{\sigma(x)(\alpha_G B)}{\sigma(x)(B)}, \phi_*(x) = \inf_{\rho > 0} \phi_\rho(x) \]
so that for any $p \leq n$ we have
\begin{equation} \label{eq.dem.1} \frac{1}{n \log \alpha} \log \frac{\sigma(x)(B \cap B_{\alpha^n}^T)}{\sigma(x)(B)} \geq \frac{1}{n \log \alpha} \sum_{i=0}^{p-1} \tau(\alpha_X^{-i}x) - \frac{1}{n\log \alpha} \sum_{i=0}^{p-1} \phi_{\alpha^{n-i}}(\alpha_X^{-i}x)\text.
\end{equation}
Fix some small $\varepsilon > 0$ and let $p_n$ be the integral part of $(1-\varepsilon)n$, so that the numbers $\alpha^{n-i}$ converge to $0$ as $n$ tends to infinity, and uniformly so with respect to $i \leq p_n$.
Let us take care of the first member of the right-hand side. We have
\begin{equation} \liminf_{n \to \infty} \frac{1}{n \log \alpha} \sum_{i=0}^{p_n-1} \tau(\alpha_X^{-i} x) = (1-\varepsilon) \int_X \frac{1}{\log \alpha} \log \left( \frac{\sigma(x)(\alpha_G B)}{\sigma(x)(B)} \right) = (1- \varepsilon) \mathrm{dim}(\lambda,G)
\end{equation}
by virtue of Birkhov's ergodic theorem and formula \eqref{eq.dimension}.
We now handle the remaining member of the right-hande side. For $\lambda$-almost every $x$ there is a real number $\rho(x)>0$ such that for any $\rho < \rho(x)$ there holds
\[ \phi_\rho(x) \geq \log\left( \frac{\sigma_H(x)(\alpha_G B)}{\sigma_H(x)(B)} \right) - \varepsilon \]
(lemma \ref{lemma.technical.2}). Choose $\rho_0>0$ small enough that, letting
\[ M = \{x\in X\ ;\ \rho(x)>\rho_0\} \]
we have
\[ \int_{X \setminus M} \phi_* \mathrm{d} \lambda \geq -\varepsilon \]
(lemma \ref{lemma.technical.1}). We split the sum on the left-hand side of \eqref{eq.dem.1} over indices $i$ such that $\alpha_X^{-i} x \in M$ and indices $i$ such that $\alpha_X^{-i} x \notin M$. First,
\[\begin{split} \liminf_{n \to \infty}& \frac{-1}{n \log \alpha} \sum_{0 \leq i \leq p_n} \mathbf{1}_M(\alpha_X^{-i}x) \phi_{\alpha^{n-i}}(\alpha_X^{-i}x) \geq \\ & \liminf_{n \to \infty} \frac{-1}{n \log \alpha} \sum_{0 \leq i \leq p_n} \mathbf{1}_M(\alpha_X^{-i}x) \left( \log \left( \frac{\sigma_H(\alpha_X^{-i}x)(\alpha B)}{\sigma_H(\alpha_X^{-i}x)(B)} \right) - \varepsilon \right) \\
& = -(1-\varepsilon)\frac{1}{\log \alpha} \int_M \left( \log \left(\frac{\sigma_H(x)(\alpha_G B)}{\sigma_H(x)(B)}\right) - \varepsilon\right) d\lambda(x)\\ & \geq - \mathrm{dim}(\lambda,H) + \frac{\varepsilon}{\log \alpha} \text.
\end{split}
\]
Second,
\[ \begin{split}
\liminf_{n \to \infty}& \frac{1}{n} \sum_{0 \leq i \leq p_n} \mathbf{1}_{\complement M}(\alpha_X^{-i} x) \phi_{\alpha^{n-i}}(\alpha_X^{-i}x)\\ & \geq \liminf_{n \to \infty} \frac{1}{n} \sum_{0 \leq i \leq p_n} \mathbf{1}_{\complement M}(\alpha_X^{-i}x) \phi_*(\alpha_X^{-i}x) \\ & = (1-\varepsilon) \int_{\complement M} \phi_* d\lambda \geq -\varepsilon(1-\varepsilon) \text.
\end{split}
\]
All in all, we get, for all $\varepsilon>0$, the inequality
\[ \liminf_{n \to \infty} \frac{1}{n \log \alpha} \log \left( \frac{\sigma(x)(B \cap B_{e^{-n}}^T)}{\sigma(x)(B)} \right) \geq (1-\varepsilon) \mathrm{dim}(\lambda,G) - \mathrm{dim}(\lambda,H) + \frac{\varepsilon + \varepsilon(1-\varepsilon)}{\log \alpha} \]
The theorem is thus proved.
\end{proof}
\section{Complex hyperbolic spaces}
\subsection{Basic facts}
Fix an integer $n \geq 1$. We denote by $G$ the group $\mathbf{PU}(1,n)$, \emph{i.e.} the projective unitary group associated with an hermitian form of signature $(1,n)$. We fix once and for all an Iwasawa decomposition $G=KAN$. Here, $K$ is a maximal compact subgroup of $G$, $A$ is a Cartan subgroup, and $N$ a maximal unipotent subgroup. We identify the quotient $G/K$ with complex hyperbolic space of (complex) dimension $n$, $\mathbf{H}_{\mathbf{C}}^n$. The hyperbolic metric of $\mathbf{H}_{\mathbf{C}}^n$ is denoted by $d$. The group of isometric transformations of $\mathbf{H}_{\mathbf{C}}^n$ is equal to $G$. The complex hyperbolic space has sectional curvature lying between $-4$ and $-1$.
We denote by $\partial \mathbf{H}_{\mathbf{C}}^n$ the boundary at infinity of $\mathbf{H}_{\mathbf{C}}^n$. Recall that $\mathbf{H}_{\mathbf{C}}^n$ is homeomorphic to the $2n$-dimensional ball $\mathbf{B}^{2n}$ and thus $\partial \mathbf{H}_{\mathbf{C}}^n$ is homeomorphic to the $(2n-1)$-dimensional sphere $\mathbf{S}^{2n-1}$.
A holomorphic totally geodesic submanifold of (complex) dimension one is called a \emph{complex geodesic}. These objects play a major role in the questions we will be interested in. The boundary of a complex geodesic is called a \emph{chain}. Recall the following easy, although important, fact:
\begin{proposition}[\cite{Goldman}, theorem 3.1.11]
\begin{enumerate}
\item Any pair of distinct points in $\mathbf{H}_{\mathbf{C}}^n \cup \partial \mathbf{H}_{\mathbf{C}}^n$ lies on a unique complex geodesic.
\item Given a nonzero tangent vector $v$ in the tangent space $T_x \mathbf{H}_{\mathbf{C}}^n$ there is a unique complex geodesic containing $x$ and tangent to $v$.
\item Any pair of distinct points of the boundary $\partial \mathbf{H}_{\mathbf{C}}^n$ lies on a unique chain.
\end{enumerate}
\end{proposition}
Recall also that the boundary $\partial \mathbf{H}_{\mathbf{C}}^n$ carries a natural CR-structure. We will not explicitly make use of this CR-structure but we note the following fact, in relation with the previous theorem.
\begin{proposition}[\cite{Goldman}, theorem 4.3.5]
Let $\xi \mapsto P_\xi$ be the natural CR-structure on the boundary $\partial \mathbf{H}_{\mathbf{C}}^n$. If $\xi$ belongs to $\partial \mathbf{H}_{\mathbf{C}}^n$ and $v \in T_\xi \partial \mathbf{H}_{\mathbf{C}}^n$ is a non-zero tangent vector at $\xi$, there is a chain passing through $x$ and tangent to $v$ if and only if $v$ does not belong to $P_\xi$.
\end{proposition}
\subsection{Heisenberg group and the boundary at infinity}
Recall that $G=KAN$ is a fixed Iwasawa decomposition. The unipotent group $N$ is isomorphic to Heisenberg group which we now introduce.
Let $V$ be a real vector space of dimension $2(n-1)$, and let $\omega$ be a non-degenerate skew-symmetric bilinear form on $V$. It is well known that two such forms are always conjugate, so we may agree (for example) that $V=\mathbf{R}^{2(n-1)}$ and
\[ \omega(x,y) = \sum_{i=1}^{n-1} x_{2i-1} y_{2i}-x_{2i} y_{2i-1} \text. \]
The product set $\mathcal H_n = V \times \mathbf{R}$ endowed with the group law
\[ (v,s) \cdot (w,t) = (v+w, s+t+\omega(v,w)) \]
and the product topology is a locally compact second countable topological group called the \emph{Heisenberg group}. The centre $Z$ of $\mathcal H_n$ is $\{0\} \times \mathbf{R}$, and is also equal to the derived subgroup.
We now define the \emph{Heisenberg metric} on $\mathcal H_n$ in the following way. First, let $\| \cdot \|_H$ be the ``Heisenberg pseudo-norm''
\[ \| (v,t) \|_H = (\|v\|^4+t^2)^{1/4} \]
(where $\|v\|$ is the euclidean norm on $\mathbf{R}^{2(n-1)}$). Then let
\[ d_H((v,t),(w,s)) = \| (v,t) \cdot (w,s)^{-1} \|_H \]
($v,w \in \mathcal H_n$). Obviously, $d_H$ is a right-invariant metric. If we denote by $h_\lambda$ ($\lambda \in \mathbf{C}^*$) the transformation
\[ h_\lambda(v,t) = (\lambda v,|\lambda|^2 t) \]
we define a similitude transformation of $\mathcal H_n$ with similitude ratio $|\lambda|$, and $h_\lambda$ is a group automorphism as we readily check. Such a transformation is called a \emph{Heisenberg similitude}.
Let us return to the Iwasawa decomposition $G=KAN$. The group $AN$ fixes a unique point $\xi_+ \in \partial \mathbf{H}_{\mathbf{C}}^n$, and $A$ itself fixes another point $\xi_- \in \partial \mathbf{H}_{\mathbf{C}}^n$. We look at the operation of $N$ on $\partial \mathbf{H}_{\mathbf{C}}^n$. This operation is transitive on the complement of $\{ \xi_+\}$. The mapping $n \mapsto n \xi_-$ ($n \in N$) is a homeomorphism $\phi$ from $N$ onto $\partial \mathbf{H}_{\mathbf{C}}^n \setminus \{\xi_+ \}$, and this homeomorphism is the restriction of a homeomorphism from the Alexandrov compactification $N \cup \{\infty\}$ onto $\partial \mathbf{H}_{\mathbf{C}}^n$.
The centre $Z$ of $N$ is mapped by $\phi$ onto the chain passing through $\xi_-$ and $\xi_+$. More generally, $\phi$ maps the translates of $Z$ in $N$ (\emph{i.e.} the fibers of the quotient mapping $N \to N/Z$) onto chains passing through $\xi_+$.
In other words, $\phi$ defines a quotient bijection from $N/Z$ onto the sets of all chains passing through $\xi_+$.
\subsection{Metrics on the boundary at infinity}
Recall that the boundary $\partial \mathbf{H}_{\mathbf{C}}^n$ is diffeomorphic to the $(2n-1)$-sphère $\mathbf{S}^{2n-1}$. We can endow $\partial \mathbf{H}_{\mathbf{C}}^n$ with the usual ``spherical metric''. Of course, the metric itself depends on the choice of the diffeomorphism, but the bilipschitz equivalence class is uniquely defined, because the sphere is compact. In other words, if $d$ and $d'$ are two spherical metrics on $\partial \mathbf{H}_{\mathbf{C}}^n$, there exist a uniform constant $C>0$ such that
\[ \frac{1}{C} d(\xi,\eta) \leq d'(\xi,\eta) \leq C d(\xi,\eta) \]
for any $\xi,\eta \in \partial \mathbf{H}_{\mathbf{C}}^n$. In particular, any notion invariant under scale changes (for example local lower or upper dimension of some measure at some point) does not depend on the choice of a spherical metric.
Now there is another bilipschitz equivalence class of metrics on $\partial \mathbf{H}_{\mathbf{C}}^n$, less elementary albeit more ``natural'' in some sense. Recall that $\mathbf{H}_{\mathbf{C}}^n$ has pinched negative curvature, so that we may consider Busemann functions
\[ b_\xi(x,y) = \lim_{t \to \infty} d(x,\xi_t)-d(y,\xi_t), \quad x,y \in \mathbf{H}_{\mathbf{C}}^n,\quad \xi \in \partial \mathbf{H}_{\mathbf{C}}^n \]
where $\xi_t$ parametrizes some unit-speed geodesic such that $\lim_{t \to \infty} \xi_t = \xi$. For any $x \in \mathbf{H}_{\mathbf{C}}^n$, the \emph{Gromov metric} from $x$ between $\xi$ and $\eta$ in $\partial \mathbf{H}_{\mathbf{C}}^n$ is then
\[ d_x(\xi,\eta) = e^{-\frac{1}{2} \left( b_\xi(x,p)+b_\eta(x,p) \right)} \]
where $p$ is some point on the (real) geodesic from $\xi$ to $\eta$. This metric is compatible with the topology of $\partial \mathbf{H}_{\mathbf{C}}^n$. The point is that the family $(d_x)_{x \in \mathbf{H}_{\mathbf{C}}^n}$ satisfies the following properties: for any $x,y \in \mathbf{H}_{\mathbf{C}}^n$, $\xi,\eta \in \partial \mathbf{H}_{\mathbf{C}}^n$ and $g \in G$,
\begin{enumerate}
\item Conformality: $d_y(\xi,\eta)=e^{\frac{1}{2}(b_\xi(x,y)+b_\eta(x,y))} d_x(\xi,\eta)$
\item Equivariance: $d_{gx}(g\xi,g\eta)=d_x(\xi,\eta)$
\end{enumerate}
Conformality property above makes it clear that all metrics $(d_x)_{x \in \mathbf{H}_{\mathbf{C}}^n}$ are pairwise bilipschitz equivalent.
The following lemma results from easy computations.
\begin{lemma} \label{lemma.gromov.spherical}
Gromov metrics are not bilipschitz equivalent to spherical metrics. Nonetheless, if $d_G, d_E$ are a Gromov and a spherical, respectively, metric on the boundary, there is a constant $C>0$ such that
\[ \frac{1}{C} d_G(\xi,\eta)^2 \leq d_E(\xi,\eta) \leq C d_G(\xi,\eta) \]
for any $\xi,\eta \in \partial \mathbf{H}_{\mathbf{C}}^n$.
\end{lemma}
In fact, the boundary $\partial \mathbf{H}_{\mathbf{C}}^n$ has dimension $2n-1$ with respect to spherical metrics, and $2n$ with respect to Gromov metrics. More generally, we recall the following
\begin{theorem}\label{th.balogh}
Let $S$ be some subset of $\partial \mathbf{H}_{\mathbf{C}}^n$ and let $\alpha,\beta$ be the Hausdorff dimensions of $S$ with respect to spherical metrics, and Gromov metrics, respectively. Then
\[ \max\{\alpha,2\alpha-2n\} \leq \beta \leq \min\{2 \alpha,\alpha + 1 \} \]
and these inequalities are sharp.
\end{theorem}
\begin{proof}
See \cite{BaloghTyson} theorem 2.4, and lemma \ref{lemma.loclip} \emph{infra}.
\end{proof}
Before ending this subsection, let us state for future reference the following
\begin{lemma} \label{lemma12}
Let $d_H$ and $d_E$ be the Heisenberg metric and Euclidean metric respectively on Heisenberg space $\mathcal H_n$.
\begin{enumerate}
\item For any compact subset $K \subset \mathcal H_n$, there is a constant $C>1$ such that
\[ \frac{1}{C} d_H(h,h')^2 \leq d_E(h,h') \leq C d_H(h,h') \]
for any $h,h' \in K$.
\item If $h,h' \in \mathcal H_n$ are such that $h-h'$ belongs to the centre, then
\[ d_E(h,h')=d_H(h,h')^2 \text. \]
\item The quotient, on $\mathcal H_n/Z$, of the Heisenberg metric, is equal to the quotient of the euclidean metric.
\end{enumerate}
\end{lemma}
\begin{proof}
Straightforward computations.
\end{proof}
\subsection{The unit tangent bundle}
Let $T^1 \mathbf{H}_{\mathbf{C}}^n$ be the unit tangent bundle of $\mathbf{H}_{\mathbf{C}}^n$, that is, the set of all pairs $(x,u)$, where $u \in \mathbf{H}_{\mathbf{C}}^n$ and $u \in T_x \mathbf{H}_{\mathbf{C}}^n$ is a unitary tangent vector at $x$. The operation of $G$ on $\mathbf{H}_{\mathbf{C}}^n$ extends to an operation on $T^1 \mathbf{H}_{\mathbf{C}}^n$. One may identify $T^1 \mathbf{H}_{\mathbf{C}}^n$ with the quotient space $G/M$. Here, $M$ is the centralizer, in $K$, of $A$ (remember that we fixed once and for all an Iwasawa decomposition $G=KAN$).
If $u$ belongs to $T^1 \mathbf{H}_{\mathbf{C}}^n$, we denote by $u^+$ and $u^-$ the forward and backward, respectively, endpoints in $\partial \mathbf{H}_{\mathbf{C}}^n$ of the geodesic defined by $u$. Likewise, if $gM$ belongs to $G/M$, we denote by $g^+$ and $g^-$ the points $u^-$ and $u^-$, respectively, where $u$ is the element of $T^1 \mathbf{H}_{\mathbf{C}}^n$ corresponding to $gM$.
Since $A$ is included in the normalizer of $M$, the operation of $A$ on $G$ by translation on the right induces an operation of $A$ on $G/M$, \emph{i.e.} $(gM,a) \mapsto gaM$. There is a (unique) isomorphism $\mathbf{R} \to A$, $t \mapsto a_t$, such that this operation of $A$ on $G/M$ is identified with the geodesic flow on $T^1 \mathbf{H}_{\mathbf{C}}^n$.
On the contrary, $N$ does not normalize $M$, so the operation of $N$ on $G$ by translation on the right does not give rise to an operation on $G/M$. Note, though, that $M$ does normalize $N$, so that $N$-orbits are well-defined in $G/M$. Actually, for any $v \in G/M$, the $N$-orbit $vN$ id the unstable manifold passing through $v$.
We need to introduce the so-called Hopf coordinates on $T^1 \mathbf{H}_{\mathbf{C}}^n$, \emph{i.e} the mapping
\[ T^1 \mathbf{H}_{\mathbf{C}}^n \to \partial^2 \mathbf{H}_{\mathbf{C}}^n \times \mathbf{R},\quad u \mapsto (u^-, u^+, b_{u^-}(u,o)) \]
where we denote, as is customary, by $\partial^2 \mathbf{H}_{\mathbf{C}}^n$ the set of all pairs $(\xi,\eta) \in \partial \mathbf{H}_{\mathbf{C}}^n \times \partial \mathbf{H}_{\mathbf{C}}^n$ with $\xi \neq \eta$, and $o$ is some fixed ``base point'' in $\mathbf{H}_{\mathbf{C}}^n$. This is a diffeomorphism. To shorten notations, we will write $u=(\xi,\eta,s)$ if $\xi=u^-$, $\eta=u^+$ and $s=b_\xi(u,o)$.
We will need the following facts.
\begin{lemma}[\cite{PaulinHersonsky}, appendix] \label{lemma.loclip}
For any $g \in G$, let $\phi_g : N \to \partial \mathbf{H}_{\mathbf{C}}^n \setminus \{ g^-\}$ be the mapping $n \mapsto (gn)^+$.
\begin{enumerate}
\item If $N$ is endowed with Heisenberg metric and $\partial \mathbf{H}_{\mathbf{C}}^n$ with a Gromov metric, $\phi_g$ is locally bilipschitz.
\item If $N$ is endowed with euclidean metric and $\partial \mathbf{H}_{\mathbf{C}}^n$ with a spherical metric, $\phi_g$ is locally bilipschitz.
\end{enumerate}
\end{lemma}
\section{Patterson-Sullivan theory}
\subsection{Limit set and growth exponent}
A good reference for this section is \cite{Roblin}. We keep the notations and conventions of the previous section. Let $\Gamma$ be a discrete subgroup of $G$. If $x$ is some point of $\mathbf{H}_{\mathbf{C}}^n$, the set of accumulation points of the orbit $\Gamma \cdot x$ on $\mathbf{H}_{\mathbf{C}}^n \cup \partial \mathbf{H}_{\mathbf{C}}^n$ is a subset $\Lambda_\Gamma$ of the boundary, namely $\Lambda_\Gamma = \overline{\Gamma \cdot x} \cap \partial \mathbf{H}_{\mathbf{C}}^n$. This set does not depend on $x$. It is called the limit set of $\Gamma$. If $\Lambda_\Gamma$ a finite set, $\Gamma$ is called elementary, otherwise $\Gamma$ is called non-elementary.
The growth exponent of $\Gamma$,
\[ \delta_\Gamma = \limsup_{R \to \infty} \frac{1}{R} \mathrm{Card} \{\gamma \in \Gamma\ ;\ d(x,\gamma x) \leq R \} \]
does not depend on $x$. It is a finite number, $0 < \delta_\Gamma \leq 2n-1$.
The study of $\Lambda_\Gamma$ and $\delta_\Gamma$ goes back a long way. We state the following important result, though we will not make use of it.
\begin{theorem}[\cite{BishopJones}, \cite{Stratmann}, \cite{PaulinBJ}]
Assume $\Gamma$ is non-elementary. Then the subset $\Lambda_\Gamma^c \subset \Gamma$ of conical limit points of $\Gamma$ has Hausdorff dimension $\delta_\Gamma$ with respect to Gromov metrics.
\end{theorem}
Recall that a point $\xi \in \partial \mathbf{H}_{\mathbf{C}}^n$ is a conical limit point if there is an infinite sequence $(\gamma_n)$ of (pairwise distincts) elements of $\Gamma$ such that the distance from $\gamma_n x$ to the geodesic $]x,\xi[$ is bounded (uniformly in $n$), for some $x$ (and, thus, for any $x$).
This theorem raises the following
\begin{question}
Let $\Gamma$ be a discrete non-elementary subgroup of $G$. What is the Hausdorff dimension of $\Lambda_\Gamma^c$ with respect to the spherical metric on the boundary?
\end{question}
\subsection{Conformal densities}
\begin{definition}
Let $\Gamma$ be a non-elementary discrete subgroup of $G$. Let $\beta$ be some real number $\geq 0$. A $\Gamma$-conformal density of exponent $\beta$ is a family $(\mu_x)_{x \in \mathbf{H}_{\mathbf{C}}^n}$ of finite measures on $\partial \mathbf{H}_{\mathbf{C}}^n$ which satisfies
\begin{enumerate}
\item $\Gamma$-equivariance:
\[ \gamma_* \mu_x = \mu_{\gamma x} \]
for any $x \in \mathbf{H}_{\mathbf{C}}^n$ and any $\gamma \in \Gamma$.
\item Conformality: for any $x,y \in \mathbf{H}_{\mathbf{C}}^n$, $\mu_x$ and $\mu_y$ are equivalent measures and the Radon-Nikodym derivative is given by
\[ \frac{\mathrm{d} \mu_y}{\mathrm{d} \mu_x}(\xi) = e^{-\beta b_\xi(y,x)} \]
almost everywhere.
\end{enumerate}
\end{definition}
The following well-known theorem is basic.
\begin{theorem}[\cite{Roblin}]
Let $\Gamma$ be a non-elementary discrete subgroup of $G$, with growth exponent $\delta_\Gamma$. There exist a $\Gamma$-conformal density of exponent $\delta_\Gamma$.
\end{theorem}
\subsection{Bowen-Margulis-Sullivan measure}
Now, let $\Gamma$ be a discrete non-elementary subgroup of $G$. Let $\mu$ be a $\Gamma$-conformal density of exponent $\delta_\Gamma$. Fix some arbitrary point $x \in \mathbf{H}_{\mathbf{C}}^n$. We define the Bowen-Margulis-Sullivan measure $m_{\mathrm{BMS}}$ on $T^1 \mathbf{H}_{\mathbf{C}}^n$:
\begin{equation} \label{eq.BMS}
\mathrm{d} m_{\mathrm{BMS}}(u) = e^{\delta_\Gamma(b_\xi(x,u)+b_\eta(x,u))} \mathrm{d} \mu_x(\xi) \mathrm{d} \mu_x(\eta) \mathrm{d}s
\end{equation}
where $u=(\xi,\eta,s)$.
This Radon measure does not depend on the choice of $x$. It is invariant under the geodesic flow as well as under $\Gamma$. Consequently, the measure on $\Gamma \backslash T^1 \mathbf{H}_{\mathbf{C}}^n$ defined by passing to the quotient is a Radon measure, invariant under the geodesic flow. Remember that $T^1 \mathbf{H}_{\mathbf{C}}^n$ is identified with $G/M$, so that we get a Radon measure on $\Gamma \backslash G/M$, invariant under the operation of $A$ on the right.
\begin{definition}
We say that a discrete non-elementary subgroup $\Gamma$ of $G$ has finite BMS measure if the associated Bowen-Margulis-Measure on $\Gamma \backslash T^1 \mathbf{H}_{\mathbf{C}}^n$ (or $\Gamma \backslash G/M$) is finite.
\end{definition}
\begin{theorem}[\cite{Roblin}]
Let $\Gamma$ be a discrete non-elementary subgroup of $G$. If $\Gamma$ has finite BMS measure, the $\Gamma$-conformal density of exponent $\delta_\Gamma$ is unique, atomless, its support is the limit set $\Lambda_\Gamma$, and the conical limit set $\Lambda_\Gamma^c$ has full measure. Furthermore, the BMS measure is (strongly) mixing with respect to the geodesic flow.
\end{theorem}
The Bowen-Margulis-Measure on $\Gamma \backslash G/M$ does not exactly suit our needs, because, as we said, $N$ does not act on the right on this space. Hence we are lead to consider the unique $M$-invariant lifting of this measure to $\Gamma \backslash G$. We still call this measure on $\Gamma \backslash G$ the Bowen-Margulis-Measure. Note that the right action of $A$ on $\Gamma \backslash G/M$ extends to a right action on $\Gamma \backslash G$. The space $\Gamma \backslash G$ is sometimes called \emph{the frame bundle of $\Gamma \backslash \mathbf{H}_{\mathbf{C}}^n$}, and the operation of $\mathbf{R}$ on $\Gamma \backslash G$, $(t, \Gamma g) \mapsto \Gamma g a_t$ is called the \emph{frame flow}. The following theorem is crucial.
\begin{theorem}[\cite{Winter}]
Let $\Gamma$ be a discrete non-elementary subgroup of $G$. Assume that $\Gamma$ is Zariski-dense and has finite BMS measure. Then the BMS measure on $\Gamma \backslash G$ is (strongly) mixing under the right operation of $A$.
\end{theorem}
We now state some easy facts and do routine checks.
The space $X=\Gamma \backslash G$ is a standard Borel space on which $N$ operates (on the right) in a Borel way with discrete stabilizers. We may disintegrate the Bowen-Margulis-Sullivan measure on $\Gamma \backslash G$ along $N$. Assume now that $\Gamma$ is Zariski-dense and has finite BMS measure. We obtain a measurable mapping $\sigma : X \to {\mathcal{M}}_{\mathrm{r}}^{\mathrm{1}}(N)$. Also, let $a$ be some non-identity element of $A$. There is a real number $t \neq 0$ such that for any $x \in X$ and $n \in N$,
\[ x n a = xa (a^{-1} n a) = xa h_{e^t}(n) \]
where $h_{e^t}$ is the Heisenberg similitude with ratio $e^t$. We may assume that $t<0$, \emph{i.e.} $h_{e^t}$ is a contracting similitude of $N$ (endowed with the Heisenberg metric).
Last, we know that the BMS measure $\lambda$ on $X$ is $a$-ergodic because of the previous theorem. The conditions stated at the beginning of section \ref{section.dimension} are thus satisfied, and we may consider the dimension of $\lambda$ along $N$ and $Z$, $\mathrm{dim}(\lambda,N)$ and $\mathrm{dim}(\lambda,Z)$ respectively, as well as the transverse dimension $\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,N/Z)$. One easily checks that
\[ \mathrm{dim}(\lambda,N) \in [0,2n],\quad \mathrm{dim}(\lambda,Z) \in [0,2],\quad \underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,N/Z) \in [0,2(n-1)] \text. \]
Indeed, recall that by definition $\mathrm{dim}(\lambda,N)$ (for example) is the dimension of the conditional measures of $\lambda$ along $N$. Now $N$ has Hausdorff dimension $2n$ (with respect to the Heisenberg metric) so we see that $\mathrm{dim}(\lambda,N) \in [0,2n]$. In the same way, $Z$ has Hausdorff dimension $2$ (with respect to the restricted Heisenberg metric), and $N/Z$ has Hausdorff dimension $2(n-1)$, indeed it is isometric to $\mathbf{R}^{2(n-1)}$.
We now apply the Ledrappier-Young formula. This yields
\begin{equation} \label{eq.LY.BMS}
\mathrm{dim}(\lambda,N) = \mathrm{dim}(\lambda,Z)+\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,N/Z)
\end{equation}
Let us check the hypotheses 1 and 2 (see beginning of section \ref{sect.LY}). First, the right operation of $N$ on $\Gamma \backslash G$ may not have uniformly discrete stabilizers, but the workaround is easy, as this operation is in fact essentially free, \emph{i.e.} the stabilizer of $\lambda$-almost every point is trivial. Indeed, if $g \in G$ is such that $gn=\gamma g$ with $n \in N$ different from the identity and $\gamma \in \Gamma$, then $g^+$ must be a parabolic limit point. There are only countably many parabolic limit points, and the Patterson-Sullivan measure is atomless (because $\Gamma$ has finite BMS measure), so the essential freeness of the operation of $N$ on $\Gamma \backslash G$ is a consequence of the very definition of BMS measure, equation \ref{eq.BMS}. Second, the quotient metric space $N/Z$ (endowed with the quotient of Heisenberg metric) satisfies Besicovitch covering property because it is isometric to euclidean space $\mathbf{R}^{2(n-1)}$.
We state for future reference the following
\begin{lemma} \label{lemma14}
Keep the previous assumptions and notations. For any $g \in G$, denote by $\phi_g$ the mapping $N \to \partial \mathbf{H}_{\mathbf{C}}^n \setminus \{ g^-\}$,
\[ \phi_g(n) = (gn)^+ \text. \]
Then for $\lambda$-almost every $x=\Gamma g \in \Gamma \backslash G$, the push-forward $(\phi_g)_* \sigma(x)$ is equivalent to the Patterson-Sullivan measure (restricted to the complement of $g^+$), and the Radon-Nikodym derivative is a continuous mapping
\[ \partial \mathbf{H}_{\mathbf{C}}^n \setminus \{ g^+ \} \to ]0,+\infty[ \text. \]
\end{lemma}
We skip the straightforward proof.
\section{Hausdorff dimension of limit sets}
We keep the notations and definitions of the previous section.
\subsection{A lower bound for Hausdorff dimension}
Let $\Gamma$ be a non-elementary discrete subgroup of $G=\mathbf{PU}(1,n)$ ($n \geq 2$). If $\Lambda_\Gamma^c$ is the set of conical limit points of $\Gamma$, the Hausdorff dimension of $\Lambda_\Gamma^c$, with respect to the Gromov metric on the boundary at infinity, is equal to $\delta_\Gamma$. Therefore, if we denote by $\mathrm{dim}_E(\Lambda_\Gamma^c)$ the Hausdorff dimension of $\Lambda_\Gamma^c$, with respect to the spherical metric, the following inequalities hold:
\[ \frac{\delta_\Gamma}{2} \leq \mathrm{dim}_E(\Lambda_\Gamma^c) \leq \delta_\Gamma \]
(lemma \ref{lemma.gromov.spherical}). In fact, by virtue of theorem \ref{th.balogh}, we know that
\[ \max\left\{\frac{\delta_\Gamma}{2},\delta_\Gamma-1\right\} \leq \mathrm{dim}_E(\Lambda_\Gamma^c) \text.\]
Our aim is to make this result more precise under some mild assumptions. Let us sketch our argument again. There is a natural fibration on $\mathcal H_n$, namely the centre $Z$ gives rise to a mapping
\[ \mathcal H_n \to \mathcal H_n /Z \text. \]
Now the main difference between Heisenberg metric and euclidean metric on $\mathcal H_n$ is ``on the fibers'', because the quotient metrics on $\mathcal H_n/Z$ are actually the same. Fiberwise, on the other hand, we have
\[ d_E(h,h')=d_H(h,h')^2 \quad(h,h' \text{ belong to the same fiber}) \text. \]
What is left is to understand how the dimension along fibers and the transversal dimension account for the dimension of Patterson-Sullivan measure itself. In general there is no exact relation but thanks to our ``Ledrappier-Young'' formula here we know that dimension is indeed additive as far as Heisenberg metric is concerned. On the other hand when Heisenberg space is endowed with euclidean metric, dimension is only super-additive -- which is why we only get an inequality in the end.
We now state and prove our theorem.
\begin{theorem} \label{th.lower.bound}
Let $\Gamma$ be a non-elementary discrete subgroup of $G=\mathbf{PU}(1,n)$, Zariski-dense, with finite BMS measure.
If $\mu$ is some Patterson-Sullivan measure of exponent $\delta_\Gamma$, then for $\mu$-almost every $\xi$,
\[ \delta_\Gamma - \frac{1}{2} \mathrm{dim}(\lambda,Z) \leq \underline{\dim}(\mu,\xi) \leq \delta_\Gamma \]
where the dimension is with respect to the spherical metric on the boundary.
The same inequality holds if we replace $\underline{\dim}(\mu,\xi)$ with $\mathrm{dim}_E(\Lambda_\Gamma^c)$.
\end{theorem}
\begin{proof}[Proof of the theorem]
The theorem is a direct consequence of the following two facts:
\begin{itemize}
\item[A.] $\displaystyle{\delta_\Gamma = \mathrm{dim}(\lambda,Z)+\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,N/Z)}$
\item[B.] $\displaystyle{\underline{\dim}(\mu,\xi) \geq \frac{1}{2} \mathrm{dim}(\lambda,Z)+\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,N/Z)}$ for $\mu$-almost every $\xi \in \partial \mathbf{H}_{\mathbf{C}}^n$
\end{itemize}
which we now prove in order.
To prove fact A, we just need to recall equation \ref{eq.LY.BMS} and prove the relation $\mathrm{dim}(\lambda,N)=\delta_\Gamma$. This holds because the Patterson-Sullivan measure has pointwise lower dimension $\delta_\Gamma$ almost everywhere (with respect to the Gromov metric), see \cite{Ledrappier} theorem 4.3, and because of lemmas \ref{lemma14} and \ref{lemma.loclip}.1.
We now prove fact B. Fix some compact neighbourhood $B$ of the identity in $N$. For any $x \in X$, let
\[ \nu^B(x)=\frac{\sigma(x)|B}{\sigma(x)(B)} \]
and let $\theta^B(x)$ be the pushforward of $\nu^B(x)$ through the quotient mapping $\pi: N \to N/Z$. Let us disintegrate $\nu^B(x)$ above $\theta^B(x)$:
\[ \nu^B(x) = \int \mathrm{d}(\theta^B(x))(v)\ \phi_v^B(x) \]
where $\phi_v^B(x)$ is (almost surely) a Radon measure concentrated on the fiber $\pi^{-1}(v)$.
We know that for $\lambda$-almost every $x \in X$, $\theta^B(x)$ has lower pointwise dimension equal to $\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,N/Z)$ almost everywhere, and $\phi_v^B(x)$ has exact dimension equal to $\mathrm{dim}(\lambda,Z)$, for $\theta^B(x)$-almost every $v \in N/Z$. Recall that $N/Z$ is endowed with the quotient metric derived from Heisenberg metric, and this metric coincides with the quotient metric derived from euclidean metric. Also, fibers of the form $\pi^{-1}(v)$ are endowed with the restriction of Heisenberg metric.
Now we endow $N$ with euclidean metric, and each fiber of the form $\pi^{-1}(v)$ is endowed with the restriction of euclidean metric. According to lemma \ref{lemma12}.2, the conditional measure $\pi_v^B(x)$ is of exact dimension $\dfrac{\mathrm{dim}(\lambda,Z)}{2}$ for $\lambda$-almost every $x$ and $\theta^B(x)$-almost every $v$. According to proposition \ref{prop.additivity.1} we deduce that the lower pointwise dimension of $\nu^B(x)$, with respect to the Euclidean metric, is (almost everywhere) at least equal to
\begin{equation} \underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,N/Z)+\frac{1}{2} \mathrm{dim}(\lambda,Z)\text. \label{eq.formula.prve}
\end{equation}
By considering an increasing sequence of compact neighbourhood of identity $B_1 \subset B_2 \subset \ldots$ and using lemma \ref{lemma.loclip}.2, we see that the pointwise lower dimension of Patterson-Sullivan measure is almost everywhere greater than the previous number.
\end{proof}
In fact, we can show that the lower inequality must be strict unless $\Gamma$ is a lattice. I am grateful to the referee for pointing out and correcting a mistake in the previous version of this proposition.
\begin{proposition}
Let $\Gamma$ be a non-elementary discrete subgroup of $G=\mathbf{PU}(1,n)$, Zariski-dense, with finite BMS measure, such that $\Lambda_\Gamma \neq \partial \mathbf{H}_{\mathbf{C}}^n$. Then $\dim(\lambda,Z)<2$. In particular, we get the strict inequality
\[ \delta_\Gamma - 1 < \mathrm{dim}_E(\Lambda_\Gamma^c) \text. \]
\end{proposition}
\begin{proof}
Let us describe briefly the rationale behind this proposition. If $\dim(\lambda,Z)=2$, then it is a classical fact that $m_{\mathrm{BMS}}$ must in fact be $Z$-invariant. This implies that the limit set mut be in some way saturated with respect to chains (recall that the boundary of a complex geodesic is called a chain). Since $\Gamma$ is Zariski-dense, its limit set $\Lambda_\Gamma$ may not be included in a chain, so it must really be a ``big'' set, and we show in fact that it must be the whole boundary.
Now we assume that $\dim(\lambda,Z)=2$. Then it is well-known that for $m_{\mathrm{BMS}}$-almost every $x \in X$ (recall that $X=\Gamma \backslash G$), $\sigma_Z(x)$ must be the Haar measure on $Z$, and $\sigma(x)$ must then be $Z$-invariant. This key fact is kind of folklore; it is proven (in slightly different languages) in \cite{MargulisTomanov}, as well as in \cite{Hoch} and \cite{EinsiedlerLindenstrauss}.
This means that for almost every $\xi$ on the boundary (with respect to the Patterson-Sullivan measure), if we send $\xi$ to infinity and identify $\partial \mathbf{H}_{\mathbf{C}}^n \setminus \{\xi\}$ with the Heisenberg space $\mathcal{H}_n$, then $\Lambda_\Gamma \setminus \{ \xi \}$ is $Z$-invariant when seen as a subset of $\mathcal{H}_n$. Here we are using the fact that $\Lambda_\Gamma$ is the support of the Patterson-Sullivan measure as well as lemma \ref{lemma14}.
Let us reformulate this: for almost every $\xi$ with respect to the Patterson-Sullivan measure, and for \emph{every} $\eta \in \Lambda_\Gamma \setminus \{\xi\}$, the unique chain passing through $\xi$ and $\eta$ is contained in $\Lambda_\Gamma$.
Clearly this implies that there is some chain $C$ that is contained in the limit set $\Lambda_\Gamma$. Since $\Gamma$ is Zariski-dense, $\Lambda_\Gamma$ cannot be equal to $C$; so the Patterson-Sullivan measure of $\Lambda_\Gamma \setminus C$ is $>0$ (as $\Lambda_\Gamma$ is the support of the Patterson-Sullivan measure and $C$ is a closed subset of $\Lambda_\Gamma$).
Now we use the above reformulation again in order to pick some point $\xi$ not in $C$ such that for \emph{any} $\eta \in C$, the chain passing through $\xi$ and $\eta$ is contained in $\Lambda_\Gamma$.
Let us send $\xi$ at infinity and look at what is going on in the Heisenberg space $\mathcal{H}_n$. The subset $L=\Lambda_\Gamma \setminus \{\xi\} \subset \mathcal{H}_n$ satisfies the following properties:
\begin{itemize}
\item It is vertically saturated; \emph{i.e.} for any $h\in L$, the translate $hZ$ is contained in $L$.
\item There is some fixed vertical chain $C=h_0 L$ and some point $h_1$ not belonging to $C$ such that every chain passing through $h_1$ and $C$ is included in $L$.
\end{itemize}
We finish the proof assuming that $n=2$; in the general setting, one would have to argue by induction on the dimension of the smallest $k$-chain contained in the limit set $\Lambda_\Gamma$ (a $k$-chain is the boundary of a $k$-complex geodesic).
Let us check that $L=\mathcal{H}_2$; we identify $\mathcal{H}_2$ with $\mathbf{C} \times \mathbf{R}$. In order to simplify notations, we assume that the vertical chain $Z$ is included in $L$, and also that the point $(1,0) \in \mathbf{C} \times \mathbf{R}$ belongs to $L$.
\begin{description}
\item[Claim.] Fix some real number $y$. The circle with centre $1/2+iy$ and radius $\sqrt{1/4+y^2}$ is contained in the vertical projection of $L$ onto $\mathbf{C}$.
\end{description}
Define
\[ v_0= \frac{1}{2}+i \qquad s_0=-2y\qquad r_0=\sqrt{\frac{1}{4}+y^2} = \| v_0 \| \text. \]
For any $\theta \in [0,2\pi]$, let
\[ v(\theta)=v_0+r_0 e^{i \theta}\qquad s(\theta)=s_0 - 2 \mathrm{Im}( \overline{v_0} v(\theta)) \text. \]
Then the mapping $\theta \mapsto (v(\theta),s(\theta))$ parametrizes the unique chain passing through $(0,s_0)$ and $(1,0)$. Of course the non-trivial assertion is the fact that this mapping does indeed parametrize a chain, see \cite{Goldman} equation (4.12) page 129. This proves the claim, and we deduce that $L=\partial \mathbf{H}_{\mathbf{C}}^2 \setminus \{\xi\}$.
We have shown that $\dim(\lambda,Z)=2$ implies $\Lambda_\Gamma = \partial \mathbf{H}_{\mathbf{C}}^n$; hence the conclusion.
\end{proof}
\begin{corollary} \label{cor.referee}
Let $\Gamma$ be a non-elementary discrete group of $G$, Zariski-dense and geometrically finite; assume furthermore that $\Gamma$ is \emph{not} a lattice. Then $\dim(\lambda,Z)<2$. In particular, we get the strict inequality
\[ \delta_\Gamma - 1 < \mathrm{dim}_E(\Lambda_\Gamma) \text. \]
\end{corollary}
\begin{proof}
Recall that geometrical finiteness of $\Gamma$ implies that the limit set $\Lambda_\Gamma$ is the union of the conical limit set $\Lambda_\Gamma^c$ and the parabolic limit set $\Lambda_\Gamma^p$, the latter being countable (hence $\Lambda_\Gamma$ and $\Lambda_\Gamma^c$ have the same Hausdorff dimension); also, the Hausdorff dimension of the limit set $\Lambda_\Gamma$ (with respect to the Gromov metric) is equal to the critical exponent $\delta_\Gamma$.
Now, by the previous proposition, we know that if $\dim(\lambda,Z)=2$, the limit set must be equal to the boundary $\partial \mathbf{H}_{\mathbf{C}}^n$; thus $\delta_\Gamma$ is equal to $2n-1$ and this implies that $\Gamma$ (which is geometrically finite) is a lattice.
\end{proof}
Before ending this paragraph, let us remark that in view of the main result in \cite{LedXie}, it seems likely that $\dim(\lambda,Z)$ should be strictly less than $\delta_\Gamma$ when $\Gamma$ is Zariski-dense (and has finite BMS measure) but I have not been able to adapt the methods of Ledrappier and Xie to prove this result.
\subsection{Schottky subgroups in good position}
We will now describe a class of Schottky subgroups of $G=\mathbf{PU}(1,n)$ for which, with the notations of theorem \ref{th.lower.bound}, $\dim(\lambda,Z)=0$. As an immediate corollary, the dimension of the limit set associated with such a Schottky subgroup, with respect to the spherical metric, is equal to $\delta_\Gamma$. Recall that a Schottky subgroup is convex cocompact, so that any limit point is a conical limit point.
Let $W \subset G$ be a finite set of hyperbolic transformations, at least two, and, for each $w \in W\cup W^{-1}$, let $B(w)$ be an open subset of $\partial \mathbf{H}_{\mathbf{C}}^n$. We make the following assumptions:
\begin{enumerate}
\item If $w \in W$, then $w^{-1}$ does not belong to $W$.
\item The closures $\overline{B(w)}$ are pairwise disjoint.
\item For any $w \in W \cup W^{-1}$,
\[ w \left( \partial \mathbf{H}_{\mathbf{C}}^n \setminus B(w^{-1})\right) \subset B(w) \text. \]
\item No chain passes through three of these open subsets $B(w)$.
\end{enumerate}
Recall that a chain is the boundary of a complex geodesic.
It is easy to construct a set $W$ satisfying these hypotheses. First, choose some hyperbolic isometries $w_1',\ldots,w_k'$ such that if $i \neq j$, the chain passing through the fixed points of $w_i'$ does not pass through a fixed point of $w_j'$. Then let $w_i = (w_i')^n$ ($1 \leq i \leq k$), and the set $W=\{w_1,\ldots,w_k\}$ suits our needs, if only $n$ is great enough.
Now let $\Gamma$ be the subgroup of $G$ generated by $W$. It is easy to check that $\Gamma$ is the free group of basis $W$, $F(W)$. Therefore for any element $\gamma \in \Gamma$, we may speak of the reduced decomposition and the length of $\gamma$ (with respect to $W$). We say that $\Gamma$ is a Schottky subgroup \emph{in good position}.
If $f \in \Gamma$ has reduced decomposition $f=f_1 \cdots f_p$ (\emph{i.e.} $f_i f_{i+1} \neq e$ for $i < p$ and $f_i \in W \cup W^{-1}$ for $i \leq p$), we denote by $B(f)$ the set
\[ f_1 \cdots f_{n-1} B(f_n) \text. \]
If $f,g \in \Gamma$ are distinct elements (not equal to identity), the sets $B(f)$ and $B(g)$ have empty intersection.
If $f,g \in \Gamma$ have reduced decomposition $f_1 \cdots f_p$, $g_1 \cdots g_q$ respectively, we denote by $f \wedge g$ the longest word $h$ with reduced decomposition $h_1 \cdots h_r $ such that $f_i = g_i = h_i$ for $1 \leq i \leq r$. If $f \wedge g$ is the empty word, we say that $f$ and $g$ are disjoint words; on the other hand, if $f \wedge g = f$, we say that $f$ is a prefix of $g$. If $f$ is a prefix of $g$, then $B(g) \subset B(f)$.
\begin{lemma}
Let $f,g,h \in \Gamma$ be such that none of them is a prefix of an other. Then the sets $B(f),B(g),B(h)$ are pairwise disjoints, and no chain passes through these three sets.
\end{lemma}
\begin{proof}
We argue by contradiction. Assume $C$ is a chain passing through $B(f),B(g)$ and $B(h)$. We may assume that $f \wedge g \wedge h$ is the empty word: if not, we replace $C$ with $(f \wedge g \wedge h)^{-1} C$, $f$ with $(f \wedge g \wedge h)^{-1} f$, \emph{etc}.
If $f,g,h$ are pairwise disjoints, the conclusion is a direct consequence of hypothesis 4 above.
Now assume for example that $k = f \wedge g$ is not the empty word. Note that $k$ is not equal to $f,g$ or $h$ because none of these words is a prefix of an other. Then $k^{-1} C$ is a chain passing through $B(k^{-1} f)$, $B(k^{-1}g)$ and $B(k^{-1} f)$. If $l(w)$ is the length of the element $w \in \Gamma$, we have
\[ l(k^{-1}f)+l(k^{-1}g)+l(k^{-1}h) < l(f)+l(g)+l(h) \]
and we may argue by induction on $l(f)+l(g)+l(h)$ to prove the lemma.
\end{proof}
\begin{lemma}
Any chain contains at most two points of $\Lambda_\Gamma$.
\end{lemma}
\begin{proof}
Since $\Gamma$ is free, for any $\xi \in \Lambda_\Gamma$ and any integer $n \geq 1$, there is a unique word $\xi(n) \in \Gamma$ of length $n$ such that $\xi$ belongs to $B(\xi(n))$. We let $\xi(n)=\xi_1 \cdots \xi_n$, where $\xi_i$ belongs to $W \cup W^{-1}$ ($1 \leq i \leq n$) and $\xi_i \xi_{i+1} \neq e$.
Now assume on the contrary that $\xi,\eta,\zeta$ are three distinct points of $\Lambda_\Gamma$ that belong to some chain $C$. Since $\xi,\eta,\zeta$ are different, we may find an index $n$ such that the three words
\[ f=\xi_1 \cdots \xi_n,\quad g=\eta_1 \cdots \eta_n,\quad h=\zeta_1 \cdots \zeta_n \]
are pairwise different. By virtue of the previous lemma, no chain passes through $B(f),B(g)$ and $B(h)$. This contradicts the fact that $C$ does pass through these sets.
\end{proof}
We are now able to prove the following fact.
\begin{proposition}
Let $\Gamma$ be a Schottky subgroup in good position of $\mathbf{PU}(1,n)$ ($n \geq 2$). Let $\lambda$ be the BMS measure on $\Gamma \backslash G$ and recall that $Z$ is the centre of $N$. Then
\[ \mathrm{dim}(\lambda,Z)=0 \text. \]
\end{proposition}
\begin{proof}
Let $\sigma : X \to {\mathcal{M}}_{\mathrm{r}}^{\mathrm{1}}(N)$ be the mapping obtained by disintegrating $\lambda$ along $N$ on the right. For $\lambda$-almost every $x=\Gamma g$, the support of $\sigma(x)$ is the inverse image, by the mapping $\phi : N \to \partial \mathbf{H}_{\mathbf{C}}^n \setminus \{g^-\}$, $n \mapsto (gn)^+$, of the set $\Lambda_\Gamma \setminus \{g^-\}$. For any $n \in N$, the translate $nZ$ is the inverse image, by $\phi$, of the chain passing through $(gn)^+$ and $g^-$.
Now, for $\lambda$-almost every $x = \Gamma g$, and $\sigma(x)$-almost every $n \in N$, both points $g^-$ and $(gn)^+$ belong to the limit set. Consequently, the chain passing through these points does not contain any other point of $\Lambda_\Gamma$, so that $nZ$ meets the support of $\sigma(x)$ only at $n$.
In other words, $\sigma(x)$ is concentrated on a Borel section of the quotient mapping $N \to N/Z$. Therefore, the conditional of $\lambda$ along $Z$ are Dirac measures, and in particular $\mathrm{dim}(\lambda,Z)=0$.
\end{proof}
\begin{corollary} \label{cor.schottky}
The Hausdorff dimension of $\Lambda_\Gamma$, with respect to the spherical metric, is equal to $\delta_\Gamma$.
\end{corollary}
Before ending this paragraph, let us remark that being in good position is not a \emph{generic} condition. Indeed it is clear by definition that a Schottky subgroup in bad position will remain so after a small perturbation, because in the definition the sets $B(w)$ are taken to be open. On the other hand, being in good position is an \emph{open} condition since we may replace the sets $B(w)$ with their closure (maybe after shrinking them a little bit).
\subsection{Discussion}
First, let us discuss our hypotheses.
Zariski density is absolutely crucial. Indeed, let us say, for example, that $n=2$. Pick some one-dimensional complex geodesic $H \subset \mathbf{H}_{\mathbf{C}}^2$. There is a bijective mapping $\phi : H \to \mathbf{H}_{\mathbf{R}}^1$ onto the real hyperbolic line, such that
\[ d_{\mathbf{H}_{\mathbf{R}}^1}(\phi(\xi),\phi(\eta))=\frac{1}{2} d_{\mathbf{H}_{\mathbf{C}}^2}(\xi,\eta) \]
(where the letter $d$ denotes hyperbolic metric on both spaces). As a consequence, if $\Gamma$ is some uniform lattice of $\mathbf{PSL}_2(\mathbf{C})$, we may identify $\Gamma$ with some discrete subgroup of $ \mathbf{PU}(1,2)$ with finite Bowen-Margulis-Sullivan measure, such that $\delta_\Gamma=2$ and $\Lambda_\Gamma=H$. The Hausdorff dimension of $\Lambda_\Gamma$ with respect to the \emph{spherical} metric is then equal to $1$. Of course, such a $\Gamma$ is not Zariski-dense (it is contained in some conjugate of $\mathbf{PU}(1,1)$).
On the other hand, it is not clear to the author wether finiteness of the Bowen-Margulis-Sullivan measure is a mere technical hypothesis. It would be interesting to find out what can happen if we assume only divergence of the Poincaré series.
Now we discuss our results and ask a few questions.
As was already mentioned in the introduction to this paper, corollary \ref{cor.schottky} has an interesting consequence: if $\Gamma$ is a well-positionned Schottky subgroup of $\mathbf{PU}(1,n)$, one must have
\[ \delta_\Gamma \leq 2(n-1) \text. \]
This raises the following
\begin{questions}
Let $\Gamma$ be some Zariski-dense Schottky subgroup of $\mathbf{PU}(1,n)$.
\begin{enumerate}
\item Does the previous inequality hold?
\item Is the Hausdorff dimension of $\Lambda_\Gamma$ with respect to the \emph{spherical} metric equal to $\delta_\Gamma$?
\item Does the dimension of the Bowen-Margulis-Sullivan measure along the centre $Z$ have to be zero?
\end{enumerate}
\end{questions}
Recall that $Z$ is the centre of the Heisenberg group. Of course, 3 implies 2 and 2 implies 1. I would like to underline the fact that these questions are genuine, and I have absolutely no heuristics whatsoever that would imply that any of these assertions should be considered plausible.
It is worth noting that another family of subsets of the Heisenberg space that exhibit coincidence of dimension with respect to both Euclidean and Heisenberg metrics was constructed by Balogh \emph{et al} in \cite{BaloghTyson} and \cite{BaloghLifts}.
Now let us pass to general, non necessarily Schottky subgroups. Let $\Gamma$ be a Zariski dense discrete subgroup of $\mathbf{PU}(1,n)$ with finite Bowen-Margulis-Sullivan measure. Theorem \ref{th.lower.bound} makes it plain that in order to understand the Hausdorff dimension (with respect to the spherical metric) of $\Lambda_\Gamma$, one would like to be able to compute the dimension of the Bowen-Margulis-Sullivan measure along the centre $Z$, $\mathrm{dim}(\lambda,Z)$. Because of the Ledrappier-Young formula, this is equivalent to computing the dimension of the Bowen-Margulis-Sullivan measure transverse to the centre $Z$, $\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,N/Z)$. Indeed, recall that the Ledrappier-Young formula yields
\[ \delta_\Gamma = \mathrm{dim}(\lambda,Z)+\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,N/Z) \text. \]
\begin{questions}
Let $\Gamma$ be some Zariski dense discrete subgroup of $\mathbf{PU}(1,n)$ with finite Bowen-Margulis-Sullivan measure. Let $\delta_\Gamma$ be the growth exponent of $\Gamma$.
\begin{enumerate}
\item Is it true that if $\delta_\Gamma \leq 2(n-1)$, then $\mathrm{dim}(\lambda,Z)=0$?
\item Is it true that if $\delta_\Gamma > 2(n-1)$, then $\underline{\mathrm{dim}}^{\mathrm{T}}(\lambda,N/Z)=2(n-1)$?
\end{enumerate}
\end{questions}
If the answer to both questions was affirmative, this would, by virtue of theorem \ref{th.lower.bound}, and thanks to the formula of Balogh \emph{et al}, imply the following formula, where we denote by $\dim_E(\Lambda_\Gamma^c)$ the Hausdorff dimension of the conical limit set with respect to the spherical metric:
\begin{equation} \tag{*} \dim_E(\Lambda_\Gamma^c) = \left \{ \begin{array}{ccc} \delta_\Gamma & \text{if} & \delta_\Gamma \leq 2(n-1) \\ 2(n-1)+\frac{1}{2} (\delta_\Gamma-2(n-1)) & \text{if} & \delta_\Gamma > 2(n-1)
\end{array} \right.
\end{equation}
What this (hypothetical) formula means is that the Hausdorff dimension of the (conical) limit set with respect to the spherical metric is really as high as it can be. Geometrically, this would imply in some way that the limit set is as nearly transverse to the centre $Z$ as possible.
To put differently, we are asking wether there is a ``transverse measure saturation'' phenomenon going on; to be more specific, does it hold that when the growth exponent is ``small'' (\emph{i.e.} $\delta_\Gamma \leq 2(n-1)$) the transverse measure accounts for all of the dimension (\emph{i.e.} the dimension along $Z$ is zero), whereas as the growth exponent gets larger, the transverse measure is in some way ``saturated'' (perhaps absolutely continuous with respect to the Lebesgue measure on the quotient space $N/Z$) and the measure along $Z$ is forced to take its share of the dimension?
In pondering these questions, one should bear in mind related examples:
\begin{itemize}
\item In the Euclidean $\mathbf{R}^n$ space, if $\mu$ is some exact-dimensional (say) probability measure of dimension $\delta$, and if we randomly pick some $k$-plane $P \subset \mathbf{R}^n$ (with respect to the standard Lebesgue measure on the space of $k$-planes in $\mathbf{R}^n$), then the orthogonal projection of $\mu$ onto $P$ has dimension
\[ \inf\{ \delta,k \} \]
and furthermore if we disintegrate $\mu$ above $P$ (along the orthogonal projection), the conditional measures have dimension $0$ if $\delta \leq k$ and $\delta - k$ otherwise. This is basically Marstrand's theorem, see \cite{Mattila} or \cite{thesis}.
\item Falconer's formula giving the almost sure Hausdorff dimension of self-affine subsets of the Euclidean space \cite{Falconer2}. If we pretend that when the boundary at infinity is endowed with the spherical metric, the elements of $\Gamma$ are affine transformations of the euclidean space $\mathbf{R}^{2n-2}$, and try and apply Falconer's formula, in order to compute the Hausdorff dimension of the (conical) limit set with respect to the spherical metric, we do end up with the above formula $(*)$.
\item In \cite{LedLin}, Ledrappier and Lindenstrauss exhibit a related ``transverse measure saturation'' phenomenon (though for some reason in that setting the phenomenon holds only in small dimension).
\end{itemize}
I reckon that I have no idea wether there are actually non-trivial (\emph{i.e.} non lattices) $\Gamma$ that satisfy $\delta_\Gamma > 2(n-1)$. In the quaternionic hyperbolic setting, it is known that there is in fact a gap, see \cite{Corlette}.
\begin{question}
Can we find a Zariski dense discrete subgroup $\Gamma$ of $\mathbf{PU}(1,n)$ with finite Bowen-Margulis-Sullivan measure, that is not a lattice, and such that
\[ \dim(\lambda,Z)>0 \quad \text{?} \]
\end{question}
Let us note as well the following question.
\begin{question}
Let $\Gamma$ be some Zariski dense discrete subgroup of $\mathbf{PU}(1,n)$ with finite Bowen-Margulis-Sullivan measure. Is the Hausdorff dimension of the conical limit set with respect to the spherical measure on the boundary equal to
\[ \delta_\Gamma - \frac{1}{2} \dim(\lambda,Z) \quad \text{?}\]
\end{question}
This is weaker than question 3 above.
Finally, I would like to mention that in the real hyperbolic setting, I am able to compute the dimension of the Bowen-Margulis-Sullivan measure along (connected) subgroups of the unipotent group $N$ (where $\mathbf{PO}(1,n)=KAN$ is an Iwasawa decomposition). This stands in stark contrast to the complex hyperbolic setting we have been looking at in this paper, where computing $\dim(\lambda,Z)$ seems more difficult. The ``transverse measure saturation'' phenomenon we mentioned earlier does take place in the real hyperbolic setting. In fact this is essentially a consequence of Marstrand's theorem. This problem is related to a paper of Oh and Mohammadi, \cite{OhMohammadi}. The reader is referred to \cite{preprintOM}.
|
\section{Introduction}
\label{sec:introduction}
Core-collapse supernovae (CCSNe) mark the violent death of massive
stars. It is believed that the initial collapse of a star's iron core
results in the formation of a proto-neutron star and the launch of a
hydrodynamic shock wave. The latter, however, fails to immediately
explode the star, but stalls and must be \emph{revived} by a
yet-uncertain supernova ``mechanism'' on a $\sim$$0.5-1\,\mathrm{s}$
timescale to explode the star (e.g.,
\cite{bethe:90,janka:12a,burrows:13a}). If the shock is not revived, a
black hole is formed and no or only a very weak explosion results
(e.g., \cite{oconnor:11,lovegrove:13,piro:13}). If the shock is
revived, it reaches the stellar surface and produces the spectacular
electromagnetic display of a Type II or Type Ib/c supernova. The Type
classification is based on the explosion light curve and spectrum, which depend
largely on the nature of the progenitor star (e.g.,
\cite{filippenko:97}). The time from core collapse to breakout of the
shock through the stellar surface and first supernova light is
minutes to days, depending on the radius of the progenitor and energy
of the explosion (e.g., \cite{kistler:13,matzner:99,morozova:15}).
Any core collapse event generates a burst of neutrinos that
releases most of the proto-neutron star's gravitational binding energy
($\sim 3 \times 10^{53}\,\mathrm{erg} \approx 0.15\,M_\odot c^2$) on a
timescale of order $10$~seconds. This neutrino burst was detected from
SN 1987A and confirmed the basic theory of CCSNe
\cite{baade:34a,bethe:90,hirata:87,bionta:87}.
Gravitational waves (GWs) are emitted by aspherical mass-energy
dynamics that includes quadrupole or higher-order contributions. Such
asymmetric dynamics are expected to be present in the pre-explosion
stalled-shock phase of CCSNe and may be crucial to the CCSN explosion
mechanism (see, e.g., \cite{bhf:95,herant:95,couch:15,lentz:15}). GWs
can serve as probes of the magnitude and character of these
asymmetries and thus may help in constraining the CCSN mechanism
\cite{ott:09b,logue:12,abdikamalov:14}.
Stellar collapse and CCSNe were considered as potential sources of
detectable GWs already for resonant bar detectors in the 1960s
\cite{weber:66}. Early analytic and semi-analytic estimates of the GW
signature of stellar collapse and CCSNe (e.g.,
\cite{ruffini:71,saenzshapiro:79,thuan:74,ipsermanagan:84,thorne:87})
gave optimistic signal strengths, suggesting that first-generation
laser interferometer detectors could detect GWs from CCSNe in the
Virgo cluster (at distances $D\gtrsim 10\,\mathrm{Mpc}$). Modern
detailed multi-dimensional CCSN simulations (see, e.g.,
\cite{dimmelmeier:08,yakunin:10,ott:11a,mueller:e12,mueller:13gw,ott:12a,ott:13a,abdikamalov:14,kuroda:14,yakunin:15}
and the reviews in \cite{ott:09,kotake:13review,fryernew:11}) find GW
signals of short duration ($\lesssim 1\,\mathrm{s}$) and emission
frequencies in the most sensitive $\sim$$10-2000\,\mathrm{Hz}$ band of
ground based laser interferometer detectors. Predicted total emitted
GW energies are in the range $10^{-12}-10^{-8}\, M_\odot c^2$ for
emission mechanisms and progenitor parameters that are presently
deemed realistic. These numbers suggest that the early predictions were optimistic and that
even second-generation laser interferometers (operating from 2015$+$) such as Advanced LIGO \cite{aLIGO}, Advanced
Virgo \cite{aVirgo}, and KAGRA \cite{kagra} will only be able to detect GWs from very nearby CCSNe at $D \lesssim 1-100\,\mathrm{kpc}$.
Only our own Milky Way and the Magellanic Clouds are within that
range. The expected event rate is very low and estimated to $\lesssim
2-3\,\mathrm{CCSNe}/100\,\mathrm{yr}$
\cite{vandenbergh:91,cappellaro:93,tammann:94,li:11,diehl:06,maoz:10}.
However, there are also a number of analytic and semi-analytic GW emission
models of more extreme scenarios, involving non-axisymmetric
rotational instabilities, centrifugal fragmentation, and accretion
disk instabilities. The emitted GW signals may be sufficiently strong
to be detectable to much greater distances of $D \gtrsim
10-15\,\mathrm{Mpc}$, perhaps even with first-generation laser
interferometers (e.g.,
\cite{ott:06prl,fryer:02,piro:07,vanputten:04}). These emission
scenarios require special and rare progenitor characteristics, but
they cannot presently be strictly ruled out on theoretical grounds. In a sphere of radius $\sim 15\,\mathrm{Mpc}$ centered on
Earth, the CCSN rate is $\gtrsim 1/\mathrm{yr}$
\cite{kistler:13,ando:05}. This makes Virgo cluster CCSNe interesting
targets for constraining extreme GW emission scenarios.
Previous observational constraints on GW burst sources applicable to
CCSNe come from all-sky searches for short-duration GW burst
signals
\cite{ando:05tama,FirstBurst,abbott:07burst,ligo_hfburst:09,ligo_burst_s5y1:09,S5y2Burst,S6Burst}.
These searches did not target individual astrophysical events.
Targeted searches have the advantage over all-sky searches that
potential signal candidates
in the data streams have to
arrive in a well-defined temporal \emph{on-source window} and have to
be consistent with coming from the sky location of the source. Both
constraints can significantly reduce the noise background and
improve the sensitivity of the search (e.g., \cite{sutton:10}).
Previous targeted GW searches have been carried out for gamma-ray
bursts
\cite{abbott:08m31,abbott:08grb,abadie_s5sgrb:10,ligo_051103:12,abadie:12grb,aasi:13stampgrb,aasi:14grb,aasi:14ipngrb},
soft-gamma repeater flares \cite{abbott:08sgr,ligo_a5_sgr:11}, and
pulsar glitches \cite{velaglitch:11}.
A recent study \cite{gossan:16} confirmed that targeted searches with
Advanced LIGO and Virgo at design sensitivity should be able to detect
neutrino-driven CCSNe out to several kiloparsecs
and rapidly rotating CCSNe out to tens of kiloparsecs,
while more extreme GW emission scenarios will be detectable to several megaparsecs.
In this paper, we present the first targeted search for GWs from CCSNe
using the first-generation Initial LIGO (iLIGO)~\cite{LIGO}, GEO\,600~\cite{grote:10},
and Virgo~\cite{VIRGO} laser interferometer detectors. The data
searched were collected over 2005--2011 in the S5, A5, and S6 runs of
the iLIGO and GEO\,600 detectors, and in the VSR1--VSR4 runs of the
Virgo detector. From the set of CCSNe observed in this period
\cite{snlistweb}, we make a preliminary selection of four targets for
our search: SNe~2007gr, 2008ax, 2008bk, and 2011dh. These CCSNe
exploded in nearby galaxies ($D\lesssim10\,\mathrm{Mpc}$), have well
constrained explosion dates, and at least partial coverage by
coincident observation of more than one interferometer. SNe 2008ax and
2008bk occurred in the \emph{astrowatch} (A5) period between the S5
and S6 iLIGO science runs. In A5, the principal goal was detector
commissioning, not data collection. Data quality and sensitivity were
not of primary concern. Preliminary analyses of the
gravitational-wave data associated with SNe 2008ax and 2008bk showed
that the sensitivity was much poorer than the data for SNe~2007gr and
2011dh. Because of this, we exclude SNe 2008ax and 2008bk and focus
our search and analysis on SNe~2007gr and 2011dh.
We find no evidence for GW signals from SNe~2007gr or 2011dh in the
data. Using gravitational waveforms from CCSN
simulations, waveforms generated with phenomenological astrophysical
models, and \textit{ad-hoc} waveforms, we measure the
sensitivity of our search. We show that none of the considered
astrophysical waveforms would likely be detectable at the distances of
SNe~2007gr and 2011dh for the first-generation detector networks.
Furthermore, even a very strong gravitational wave could potentially be
missed due to incomplete coverage of the CCSN on-source window by
the detector network.
Motivated by this, we provide a statistical approach for model
exclusion by combining observational results for multiple CCSNe. Using
this approach, we quantitatively estimate how increased detector
sensitivity and a larger sample of targeted CCSNe will improve our
ability to rule out the most extreme emission models. This suggests
that observations with second-generation ``Advanced'' interferometers
\cite{aLIGO,aVirgo,kagra} will be able to put interesting constraints
on GW emission of extragalactic CCSN at $D\lesssim 10\,\mathrm{Mpc}$.
The remainder of this paper is structured as follows. In
Section~\ref{sec:ccsne}, we discuss the targeted CCSNe and the
determination of their on-source windows. In Section~\ref{sec:networks},
we describe the detector networks, the coverage of the on-source
windows with coincident observation, and the data searched. In
Section~\ref{sec:overview}, we present our search methodology and the
waveform models studied. We present the search results in
Section~\ref{sec:results} and conclusions in
Section~\ref{sec:summaries}.
\section{Targeted Core-Collapse Supernovae}
\label{sec:ccsne}
\begin{table*}[t]
\caption{Core-collapse supernovae selected as triggers for the gravitational-wave search described in this paper. Distance gives the best current estimate
for the distance to the host galaxy. $t_1$ and $t_2$ are the UTC
dates delimiting the on-source window. $\Delta t$ is the temporal
extent of the on-source window. iLIGO/Virgo run indicates the data taking campaign during which the supernova explosion was
observed. Detectors lists the interferometers taking data during at
least part of the on-source window. The last column provides the
relative coverage of the on-source window with science-quality or
Astrowatch-quality data of at least two detctors. For SN 2007gr,
the relative coverage of the on-source window with the most
sensitive network of four active interferometers is 67\%. See the
text in Section~\ref{sec:ccsne} for details and references on the
supernovae and Section~\ref{sec:networks} for details on the
detector networks, coverage, and data quality. }
\begin{center}
\begin{tabular}{ll@{\!}rrllrrrr}
\hline
\hline
Identifier & Type &\multicolumn{1}{c}{Host}&\multicolumn{1}{c}{Distance}
& \multicolumn{1}{c}{$t_{1}$} & \multicolumn{1}{c}{$t_{2}$}
& \multicolumn{1}{c}{$\Delta t$} & \multicolumn{1}{c}{iLIGO/Virgo} & \multicolumn{1}{c}{Active}&\multicolumn{1}{c}{Coincident}\\
& &\multicolumn{1}{c}{Galaxy} &\multicolumn{1}{c}{[Mpc]} &\multicolumn{1}{c}{[UTC]} &\multicolumn{1}{c}{[UTC]} & \multicolumn{1}{c}{[days]} &\multicolumn{1}{c}{Run} & \multicolumn{1}{c}{Detectors}&\multicolumn{1}{c}{Coverage}\\
\hline
SN 2007gr & Ic & NGC 1058&$10.55$$\pm$$1.95$
& 2007 Aug $10.39$ & 2007 Aug $15.51$ & 5.12 & S5/VSR1 & H1,H2,L1,V1&93\%\\
SN 2008ax & IIb & NGC 4490 &$9.64$$+1.38$$-1.21$
& 2008 Mar $2.19$ & 2008 Mar 3.45 & 1.26 & A5 & G1,H2 &8\% \\
SN 2008bk & IIP & NGC 7793 &$3.53$$+0.21$$-0.29$
& 2008 Mar $13.50$ & 2008 Mar 25.14 & 11.64 & A5 & G1,H2 &38\%\\
SN 2011dh & IIb & M51 &$8.40$$\pm$$0.70$
& 2011 May $30.37$ & 2011 May 31.89 & 1.52 & S6E/VSR4 & G1,V1 &37\%\\
\hline
\hline
\end{tabular} \label{table:triggers}
\end{center}
\end{table*}
For the present search it is important to have an estimate of the time
of core collapse for each supernova. This time
coincides (within one to a few seconds; e.g.,
\cite{ott:09}) with the time of strongest GW emission. The better the
estimate of the core collapse time, the smaller the \emph{on-source
window} of detector data that must be searched and the smaller the
confusion background due to non-Gaussian non-stationary detector
noise.
For a Galactic or Magellanic Cloud CCSN, the time of core collapse
would be extremely well determined by the time of arrival of the neutrino
burst that is emitted coincident with the GW signal
\cite{pagliaroli:09}. A very small on-source window of seconds to
minutes could be used for such a special event.
For CCSNe at distances $D\gtrsim 1\,\mathrm{Mpc}$, an observed coincident
neutrino signal is highly unlikely \cite{ikeda:07,leonor:10}. In this
case, the time of core collapse must be inferred based on estimates of
the explosion time, explosion energy, and the radius of the
progenitor. The explosion time is defined as the time at which the
supernova shock breaks out of the stellar surface and the
electromagnetic emission of the supernova begins. Basic information
about the progenitor can be obtained from the lightcurve and spectrum
of the supernova (e.g., \cite{filippenko:97}). Much more information
can be obtained if pre-explosion imaging of the progenitor is
available (e.g., \cite{smartt:09b}). A red supergiant progenitor with
a typical radius of $\sim$$500 - 1500\,R_\odot$ produces a Type IIP
supernova and has an explosion time of $\sim$$1-2\,\mathrm{days}$
after core collapse and a typical explosion energy of
$10^{51}\,\mathrm{erg}$; sub-energetic explosions lead to longer
explosion times (e.g., \cite{kistler:13,matzner:99,morozova:15}). A
yellow supergiant that has been partially stripped of its
hydrogen-rich envelope, giving rise to a IIb supernova (e.g.,
\cite{bersten:12}), is expected to have a radius of
$\sim$$200-500\,R_\odot$ and an explosion time of $\lesssim
0.5\,\mathrm{days}$ after core collapse
\citep{bersten:12,morozova:15}. A blue supergiant, giving rise to a
peculiar type IIP supernova (such as SN 1987A), has a radius of
$\lesssim$$100\,R_\odot$ and an explosion time of
$\lesssim$$2-3\,\mathrm{hours}$ after core collapse. A Wolf-Rayet star
progenitor, giving rise to a Type Ib/c supernova, has been stripped of
its hydrogen (and helium) envelope by stellar winds or binary
interactions and has a radius of only a $\mathrm{few}$ to
$\sim$$10\,R_\odot$ and shock breakout occurs within
$\sim$$10-100\,\mathrm{s}$ of core collapse
\cite{kistler:13,matzner:99}.
The breakout of the supernova shock through the surface of the
progenitor star leads to a short-duration high-luminosity burst of
electromagnetic radiation with a spectral peak dependent on the radius
of the progenitor. The burst from shock-breakout preceeds the rise of
the optical lightcurve which occurs on a timescale of days after shock
breakout (depending, in detail, on the nature of the progenitor star;
\cite{filippenko:97,kasen:09,bersten:12,morozova:15}).
With the exception of very few serendipitous
discoveries of shock breakout bursts (e.g.,
\cite{soderberg:08,gezari:08}), core-collapse supernovae in the
2007--2011 time frame of the present GW search were usually discovered
days after explosion and their explosion time is constrained by one or
multiple of (\emph{i}) the most recent non-detection, i.e., by the
last date of observation of the host galaxy without the supernova present;
(\emph{ii}) by comparison of observed lightcurve and spectra with
those of other supernovae for which the explosion time is well known;
(\emph{iii}) by lightcurve extrapolation \cite{cowen:10}; or,
(\emph{iv}), for type IIP supernovae, via lightcurve modeling using
the expanding photosphere method (EPM; e.g.,
\cite{kirshner:74,dessart:05}).
More than 100 core-collapse supernovae were discovered in the optical
by amateur astronomers and professional astronomers (e.g.,
\cite{snlistweb}) during the S5/S6 iLIGO and the VSR2, VSR3, VSR4 Virgo data
taking periods. In order to select optically discovered core-collapse
supernovae as triggers for this search, we impose the following
criteria: (\emph{i}) distance from Earth not greater than
$\sim$$10-15\,\mathrm{Mpc}$. Since GWs from core-collapse supernovae
are most likely very weak and because the observable GW amplitude
scales with one-over-distance, nearer events are greatly
favored. (\emph{ii}) A well constrained time of explosion leading to an
uncertainty in the time of core collapse of less than $\sim$2 weeks.
(\emph{iii}) At least partial availability of science-quality data of
coincident observations of more than one interferometer in the
on-source window.
The core-collapse supernovae making these cuts are SN~2007gr,
SN~2008ax, SN~2008bk, and SN~2011dh. Table~\ref{table:triggers}
summarizes key properties of these supernovae and we discuss each in
more detail in the following.
{\bf SN 2007gr}, a Type Ic supernova, was discovered on 2007 August
15.51 UTC\,\cite{madison:07}. A pre-discovery empty image taken by
KAIT~\cite{filippenko:01} on August 10.44 UTC provides a baseline
constraint on the explosion time. The progenitor of this supernova was
a compact stripped-envelope star
\cite{crockett:08b,mazzali:10,eldridge:13,chen:14} through which the supernova
shock propagated within tens to hundreds of seconds. In order to be
conservative, we add an additional hour to the interval between
discovery and last non-detection and arrive at a GW on-source window
of 2007 August 10.39 UTC to 2007 August 15.51 UTC. The sky location of
SN 2007gr is
$\mathrm{R.A.}\!=02^{\mathrm{h}}43^{\mathrm{m}}27^{\mathrm{s}}.98$, $
\mathrm{Decl.}\!= +37^\circ20{'}44{''}.7$ \cite{madison:07}. The host
galaxy is NGC 1058. Schmidt~\emph{et al.} \cite{schmidt:94} used
EPM to determine the distance to SN 1969L, which exploded in the same
galaxy. They found $D = (10.6+1.9-1.1)\,\mathrm{Mpc}$. This is broadly
consistent with the more recent Cepheid-based distance estimate of $D
= (9.29\pm0.69)\,\mathrm{Mpc}$ to NGC 925 by
\cite{silbermann:96}. This galaxy is in the same galaxy group as NGC
1058 and thus presumed to be in close proximity. For the purpose of
the present study, we use the conservative combined distance estimate
of $D = (10.55\pm 1.95\,\mathrm{Mpc})$.
\begin{table*}[t]
\caption{Overview of GW interferometer science runs from which we draw
data for our search. H1 and H2 stand for the LIGO Hanford
$4$-$\mathrm{km}$ and $2$-$\mathrm{km}$ detectors, respectively. L1
stands for the LIGO Livingston detector. V1 stands for the Virgo
detector and G1 stands for the GEO\,600 detector. The duty factor
column indicates the approximate fraction of science-quality data
during the observation runs. The coincident duty factor column
indicates the fraction of time during which at least two detectors
were taking science-quality data simultaneously. The A5 run was
classified as \emph{astrowatch} and was not a formal science
run. The H2 and V1 detectors operated for only part of A5. The Virgo
VSR1 run was joint with the iLIGO S5 run, the Virgo VSR2 and VSR3 runs
were joint with the iLIGO S6 run, and the GEO\,600 detector (G1)
operated in iLIGO run S6E during Virgo run VSR4. When iLIGO and Virgo
science runs overlap, the coincident duty factor takes into
account iLIGO, GEO\,600, and Virgo detectors.} \centering
\begin{tabular}{l@{\hspace{0.3em}}c@{\hspace{1em}}c@{\hspace{1em}}c@{\hspace{1em}}c}
\hline
\hline
Run&Detectors&Run Period&Duty Factors& Coin.\ Duty Factor\\
\hline
S5&H1,H2,L1,G1&2005/11/04--2007/10/01&$\sim$75\% (H1), $\sim$76\% (H2), $\sim$65\% (L1), $\sim$77\% (G1)&$\sim$87\%\\
A5&G1,H2,V1&2007/10/01--2009/05/31& $\sim$81\%(G1), $\sim$18\% (H2), $\sim$5\% (V1)& $\sim$18\%\\
S6&L1,H1,G1&2009/07/07--2010/10/21&$\sim$51\% (H1), $\sim$47\% (L1), $\sim$56\% (G1)&$\sim$67\%\\
S6E&G1&2011/06/03--2011/09/05&$\sim$77\%&$\sim$66\%\\
VSR1/S5&V1&2007/05/18--2007/10/01&$\sim$80\%&$\sim$97\%\\
VSR2/S6&V1&2009/07/07--2010/01/08&$\sim$81\%&$\sim$74\%\\
VSR3/S6&V1&2010/08/11--2010/10/19&$\sim$73\%&$\sim$94\%\\
VSR4/S6E&V1&2011/05/20--2011/09/05&$\sim$78\%&$\sim$62\%\\
\hline
\hline
\end{tabular}
\label{tab:runs}
\end{table*}
{\bf SN 2008ax}, a Type IIb supernova \cite{chornock:08}, was
discovered by KAIT on 2008 March 3.45 UTC\,\cite{mostardi:08}. The
fortuitous non-detection observation made by Arbour on 2008 March 3.19
UTC \cite{arbour:08}, a mere 6.24 h before the SN discovery, provides
an excellent baseline estimate of the explosion time. Spectral
observations indicate that the progentior of SN 2008ax was almost
completely stripped of its hydrogen envelope, suggesting that is
exploded either as a yellow supergiant or as a Wolf-Rayet
star \cite{chornock:11,crockett:08}. Most recent observations and
phenomenological modeling by \cite{folatelli:15} suggest that the
progenitor was in a binary system and may have had a blue-supergiant
appearance and an extended ($30-40\,R_\odot$) low-density (thus,
low-mass) hydrogen-rich envelope at the time of explosion. To be
conservative, we add an additional day to account for the uncertainty
in shock propagation time and define the GW on-source window as 2008
March 2.19 UTC to 2008 March 3.45 UTC. The coordinates of SN 2008ax
are $\mathrm{R.A.}\!
=12^{\mathrm{h}}30^{\mathrm{m}}40^{\mathrm{s}}.80$, $
\mathrm{Decl.}\!= +41^\circ38{'}14{''}.5$ \cite{mostardi:08}. Its
host galaxy is NGC 4490, which together with NGC 4485 forms a pair of
interacting galaxies with a high star formation rate. We adopt the
distance $D = (9.64 +1.38 -1.21)\, \mathrm{Mpc}$ given by
Pastorello~\emph{et al.}~\cite{pastorello:08}
{\bf SN 2008bk}, a Type IIP supernova, was discovered on 2008 March
25.14 UTC \cite{monard:08}. Its explosion time is poorly constrained
by a pre-explosion image taken on 2008 January 2.74
UTC\,\cite{monard:08}. Morrell \& Stritzinger \cite{morrell:08}
compared a spectrum taken of SN 2008bk on 2008 April 12.4 UTC to a
library of SN spectra \cite{howell:05} and found a best fit to the
spectrum of SN 1999em taken at 36~days after explosion
\cite{morrell:08}. However, the next other spectra available for SN
1999em are from 20 and 75 days after explosion, so the uncertainty of
this result is rather large. EPM modeling by
Dessart~\cite{dessart:11pc} suggests an explosion time of March
$19.5\pm5$ UTC, which is broadly consistent with the lightcurve data
and hydrodynamical modeling presented in \cite{hamuy:12}. The
progenitor of SN 2008bk was most likely a red supergiant with a radius
of $\sim$$500\,R_\odot$ \cite{maund:14,vandyk:12,mattila:08}, which
suggests an explosion time of $\sim$$1\,\mathrm{day}$ after core
collapse \cite{kistler:13,matzner:99,morozova:15}. Hence, we assume a
conservative on-source window of 2008 March 13.5 UTC to 2008 March
25.14 UTC. The coordinates of SN 2008bk are $\mathrm{R.A.}\!=
23^{\mathrm{h}}57^{\mathrm{m}}50^{\mathrm{s}}.42$, $ \mathrm{Decl.}\!=
-32^\circ 33{'} 21{''}.5$ \cite{li:08_2008bk}. Its host galaxy is NGC
7793, which is located at a Cepheid-distance $D = (3.44 +0.21
-0.2)\,\mathrm{Mpc}$ \cite{pietrzynski:10}. This distance estimate is
consistent with $D = (3.61 +0.13 -0.14)\,\mathrm{Mpc}$ obtained by
\cite{jacobs:09} based on the tip of the red giant branch method
(e.g., \cite{lee:93}). For the purpose of this study, we use a
conservative averaged estimate of $D = (3.53 +0.21
-0.29)\,\mathrm{Mpc}$.
{\bf SN 2011dh}, a type IIb supernova, has an earliest discovery date in the literature
of 2011 May 31.893, which was by amateur astronomers
\cite{griga:11,marion:14,ergon:14,horesh:13}. An earlier discovery
date of 2011 May 31.840 is given by Alekseev \cite{rochester2011dh}
and a most recent non-detection by Dwyer on 2011 May 31.365
\cite{rochester2011dh}. The progenitor of SN 2011dh was with high
probability a yellow supergiant star \cite{vandyk:13} with a radius of
a $\mathrm{few}\,100\,R_\odot$
\cite{bersten:12,vinko:12,vandyk:11}. We conservatively estimate an
earliest time of core collapse of a day before the most recent
non-detection by Dwyer and use an on-source window of 2011 May 30.365
to 2011 May 31.893. SN 2011dh's location is $\mathrm{R.A.}\!=
13^{\mathrm{h}}30^{\mathrm{m}}05^{\mathrm{s}}.12, $ $\mathrm{Decl.}\!=
+47^\circ 10{'} 11{''}.30$ \cite{atel:3406} in the nearby spiral
galaxy M51. The best estimates for the distance to M51 come
from Vink\'o~\emph{et al.}~\cite{vinko:12}, who give $D =
8.4\pm0.7\,\mathrm{Mpc}$ on the basis of EPM modeling of SN 2005cs and
SN 2011dh. This is in agreement with Feldmeier~\emph{et al.}
\cite{feldmeier:97}, who give $D=8.4\pm0.6\,\mathrm{Mpc}$ on the basis
of planetary nebula luminosity functions. Estimates using surface
brightness variations \cite{tonry:01} or the Tully-Fisher relation
\cite{tully:88} are less reliable, but give a somewhat lower distance
estimates of $D=7.7\pm0.9$ and $D = 7.7\pm1.3$, respectively. We adopt
the conservative distance $D = 8.4\pm0.7\,\mathrm{Mpc}$ for the purpose of
this study.
\section{Detector Networks and Coverage}
\label{sec:networks}
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\textwidth]{figures/coverage/plotDuty_search_paper.pdf}
\caption{On-source windows as defined for the four core-collapse
supernovae considered in Section~\ref{sec:ccsne}. The date given
for each core-collapse supernova is the published date of discovery.
Overplotted in color are the stretches of time covered with
science-quality and Astrowatch-quality data of the various GW
interferometers. The percentages given for each core-collapse
supernova and interferometer is the fractional coverage of the
on-source window with science or astrowatch data by that
interferometer. See
Table~\ref{table:triggers} and Sections~\ref{sec:ccsne} and
\ref{sec:networks} for details.}
\label{fig:coverage}
\end{figure*}
This search employs data from the $4\,\mathrm{km}$ LIGO Hanford, WA
and LIGO Livingston, LA interferometers (denoted \textbf{H1} and
\textbf{L1}, respectively), from the $2\,\mathrm{km}$ LIGO Hanford, WA
interferometer (denoted as \textbf{H2}), from the $0.6\,\mathrm{km}$
GEO 600 detector near Hannover, Germany (denoted as \textbf{G1}), and
from the $3\,\mathrm{km}$ Virgo interferometer near Cascina, Italy
(denoted as \textbf{V1}).
Table~\ref{tab:runs} lists the various GW interferometer data taking
periods (``runs'') in the 2005--2011 time frame from which we draw
data for our search. The table also provides the duty factor and
\emph{coincident} duty factor of the GW interferometers. The duty
factor is the fraction of the run time a given detector was
taking science-quality data. The coincident duty factor is the
fraction of the run time at least two detectors were taking science
quality data. The coincident duty factor is most relevant for GW
searches like ours that require data from at least two detectors to
reject candidate events that are due to non-Gaussian instrumental or
environmental noise artifacts (``glitches'') but can mimic real
signals in shape and time-frequency content (see, e.g.,
\cite{LIGO,ligo_burst_s5y1:09}).
One notes from Table~\ref{tab:runs} that the duty factor for the
first-generation interferometers was typically $\lesssim50-80\%$.
The relatively low duty factors are due to a combination of
environmental causes (such as distant
earthquakes causing loss of interferometer lock) and interruptions for
detector commissioning or maintenance.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\columnwidth]{figures/spectra/spectra.pdf}
\caption{Noise amplitude spectral densities of the GW interferometers
whose data are analyzed for SNe 2007gr and 2011dh (see
Section~\ref{sec:networks}). The curves are the results of averaging
$1/S(f)$ over the on-source windows of the SNe (see
Table~\ref{table:triggers}). We plot the G1 noise spectrum also for
SN~2008ax to demonstrate the improvement in high-frequency
sensitivity due to GEO-HF \cite{grote:13squeeze} for SN~2011dh.}
\label{fig:noise}
\end{figure}
The CCSNe targeted by this search and described in
Section~\ref{sec:ccsne} are the only 2007--2011 CCSNe located within
$D\lesssim 10-15 \,\mathrm{Mpc}$ for which well-defined on-source
windows exist and which are also covered by extended stretches of coincident
observations of at least two interferometers. In
Figure~\ref{fig:coverage}, we depict the on-source windows for
SNe 2007gr, 2008ax, 2008bk, and 2011dh. We indicate with regions of
different color times during which the various interferometers were
collecting data.
SN 2007gr exploded during the S5/VSR1 joint run between the iLIGO, GEO
600, and Virgo detectors. It has the best coverage of all considered
CCSNe: 93\% of its on-source window are covered by science-quality
data from at least two of H1, H2, L1, and V1. We search for GWs from
SN 2007gr at times when data from the following detector networks are
available: H1H2L1V1, H1H2L1, H1H2V1, H1H2, L1V1. The G1 detector was
also taking data during SN 2007gr's on-source window, but since its
sensitivity was much lower than that of the other detectors, we do not
analyze G1 data for SN 2007gr.
SNe 2008ax and 2008bk exploded in the A5 \emph{astrowatch} run between
the S5 and S6 iLIGO science runs (cf.~Table~\ref{tab:runs}). Only
the G1 and H2 detectors were operating at sensitivities much lower
than those of the $4$-km L1 and H1 and the $3$-km V1 detectors. The
coincident duty factor for SN~2008ax is only 8\% while that for
SN~2008bk is 38\%. Preliminary analysis of the available coincident GW
data showed that due to a combination of low duty factors and low
detector sensitivity, the overall sensitivity to GWs from these CCSNe
was much lower than for SNe~2007gr and 2011dh. Because of this, we
exclude SNe~2008ax and 2008bk from the analysis presented in the rest
of this paper.
SN 2011dh exploded a few days before the start of the S6E/VSR4 run
during which the V1 and G1 interferometers were operating
(cf.~Table~\ref{tab:runs}). G1 was operating in GEO-HF mode
\cite{grote:13squeeze} that improved its high-frequency ($f \gtrsim
1\,\mathrm{kHz}$) sensitivity to within a factor of two of V1's
sensitivity. While not officially in a science run during the
SN~2011dh on-source window, both G1 and V1 were operating and
collecting data that passed the data quality standards necessary for
being classified as science-quality data (e.g.,
\cite{VSRdetchar:12,S6detchar:15,mciver:12}). The coincident G1V1 duty
factor is 37\% for SN 2011dh.
In Figure~\ref{fig:noise}, we plot the one-side noise amplitude spectral
densities of each detector averaged over the on-source windows of
SNe~2007gr and 2011dh. In order to demonstrate the high-frequency
improvement in the 2011 G1 detector, we also plot the G1 noise
spectral density for SN~2008ax for comparison.
\section{Search Methodology}\label{sec:overview}
\begin{figure*}[t]
\begin{minipage}[l][][t]{0.495\textwidth}
\vspace*{\fill}
\flushleft
\includegraphics[width=0.97\linewidth]{figures/plotFAP/cwb_FAP_SN2007gr_HHLV.pdf}
\end{minipage}
\begin{minipage}[r][][t]{0.495\textwidth}
\vspace*{\fill}
\flushright
\includegraphics[width=0.97\linewidth]{figures/plotFAP/X_FAP_SN2007gr_HHLV.pdf}
\end{minipage}
\caption{False Alarm Probability [FAP,
Equation~(\ref{eqn:fap})] distributions of the background events
for SN~2007gr and the H1H2L1V1 detector network
(cf.\ Section~\ref{sec:networks}). The FAP indicates the
probability that an event of a given ``loudness''
(significance) is consistent with background noise. The
left panel shows the FAP distribution determined by the
{\textsc{cWB}} pipeline as a function of its loudness measure, $\rho$,
(see \cite{klimenko:08} for details). The right panel depicts
the same for {\textsc{X-Pipeline}} as a function of its loudness measure,
$\Lambda_c$, (see \cite{sutton:10,Was:2012zq} for
details). The shaded regions indicate $1-\sigma$ error
estimates for the FAP. }
\label{fig:bckgrddist}
\end{figure*}
Two search algorithms are employed in this study: {\textsc{X-Pipeline}}
\cite{sutton:10,Was:2012zq} and Coherent WaveBurst ({\textsc{cWB}})
\cite{klimenko:08}. Neither algorithm requires detailed assumptions
about the GW morphology and both look for subsecond GW transients in
the frequency band 60\,Hz to 2000\,Hz. This is the most sensitive band
of the detector network, where the amplitude of the noise spectrum of
the most sensitive detector is within about an order of magnitude of
its minimum. This band also encompasses most models for GW emission from
CCSNe (cf.~\cite{ott:09,kotake:13review,fryer:11}). The benefit of
having two independent algorithms is that they can act as a cross
check for outstanding events. Furthermore, sensitivity studies using
simulated GWs show some complementarity in the signals detected by
each pipeline; this is discussed further in Section~\ref{sec:results}.
The two algorithms process the data independently to identify potential GW
events for each supernova and network combination. Each algorithm assigns a
``loudness'' measure to each event; these are described in more detail below.
The two algorithms also evaluate measures of signal consistency across
different interferometers and apply thresholds on these measures (called
coherence tests) to reject background noise events.
We also reject events that occur at times of environmental noise
disturbances that are known to be correlated with transients in the GW data
via well-established physical mechanisms; these so-called ``category 2'' data
quality cuts are described in~\cite{S5y2Burst}.
The most important measure of an event's significance is its false
alarm rate (FAR): the rate at which the background noise produces
events of equal or higher loudness than events that pass all coherent
tests and data quality cuts. Each pipeline estimates the FAR using
background events generated by repeating the analysis on time-shifted
data --- the data from the different detectors are offset in time, in
typical increments of $\sim 1$\,s. The shifts remove the chance of
drawing a sub-second GW transient into the background sample since the
largest time of flight between the LIGO and Virgo sites is 27
milliseconds (between H1 and V1). To accumulate a sufficient sampling
of rare background events, this shifting procedure is performed
thousands of times without repeating the same relative time shifts
among detectors.
Given a total duration $T_\mathrm{off}$ of off-source (time-shifted) data,
the smallest false alarm rate that can be measured is $1/T_\mathrm{off}$.
On-source events from each combination of CCSN, detector
network, and pipeline are assigned a FAR using the time-slide
background from that combination only.
The event lists from the different CCSNe, detector networks, and pipelines
are then combined and the events ranked by their FAR.
The event with lowest FAR is termed the {\it loudest event}.
In order for the loudest event to be considered as a GW detection
it must have a False Alarm Probability (FAP) low enough that it is implausible to have been caused
by background noise. Given a FAR value $R$,
the probability $p(R)$ of noise producing one or more events of FAR less than
or equal to $R$ during one or more CCSN on-source windows of total duration $T_\mathrm{on}$ is
\begin{equation}
\label{eqn:fap}
p = 1-\exp{(-R T_\mathrm{on})} \, .
\end{equation}
The smallest such false alarm probability (FAP) that can be measured
given an off-source (time-shifted) data duration $T_\mathrm{off}$ is
approximately $T_\mathrm{on}/T_\mathrm{off}$. Several thousand time
shifts are therefore sufficient to measure FAP values of $O(10^{-3})$.
We require a FAP below 0.001, which exceeds 3-$\sigma$ confidence,
in order to consider an event to be a possible GW detection candidate.
Figure~\ref{fig:bckgrddist} shows examples of the FAP as a function
of event loudness for {\textsc{cWB}} and {\textsc{X-Pipeline}} for the H1H2L1V1 network
during the SN~2007gr on-source window.
If no on-source events have a FAP low enough to be considered
GW candidates, then we can set upper limits on the strength of any
GW emission by the CCSNe.
This is done by adding to the data simulated GW signals of various
amplitudes (or equivalently sources at various distances) and repeating
the analysis. For each amplitude or distance we measure the fraction
of simulations that produce an event in at least one pipeline with FAP
lower than the loudest on-source event, and which survive our coherence
tests and data quality cuts; this fraction is the \textit{detection efficiency}
of the search.
\subsection{Coherent WaveBurst}
\label{sec:cwb}
The {\textsc{cWB}} \cite{klimenko:08} analysis is performed as described in
\cite{ligo_burst_s5y1:09}, and it is based on computing a constrained
likelihood function. In brief: each detector data stream is decomposed
into 6 different wavelet decompositions (each one with different time and frequency resolutions).
The data are whitened, and the largest 0.1 percent of wavelet
magnitudes in each frequency bin and decomposition for each
interferometer are retained (we call these ``black pixels''). We also retain ``halo'' pixels, which are those that surround each
black pixel.
In order to choose pixels that are more likely related to a
GW transient ({\it candidate event}) we identify clusters of them. Once all of the wavelet decompositions
are projected into the same time frequency plane, clusters are defined as sets of contiguous retained pixels (black or halo).
Only the pixels involved in a cluster are used in the subsequent calculation of the
likelihood. These clusters also need to be consistent between
interferometers for the tested direction of arrival. For each cluster
of wavelets, a Gaussian likelihood function is computed, where the
unknown GW is reconstructed with a maximum-likelihood
estimator.
The likelihood analysis is repeated over a grid of sky
positions covering the range of possible directions to the GW source.
Since the sky location of each of the analyzed CCSNe is well known, we
could choose to apply this procedure only for the known CCSN sky
location. However, the detector noise occasionally forces the {\textsc{cWB}}
likelihood to peak in a sky location away from the true sky
location. As a consequence, some real GW events could be assigned a smaller
likelihood value, lowering the capability to detect them.
Because of this, we consider triggers that fall within an error region
of $0.4$~degrees of the known CCSN sky location and that pass the
significance threshold, even if they are not at the peak of the {\textsc{cWB}}
reconstructed sky position likelihood. The $0.4$~degree region is
determined empirically by trade-off studies between detection
efficiency and FAR.
For SN~2011dh, the noise spectra were very different for the G1 and V1
detectors, with the consequence that the network effectively had only
one detector at frequencies up to several hundred Hz, and therefore
location reconstruction was very poor. As a consequence we decided to
scan the entire sky for candidate events for this CCSN.
The events reported for a given network configuration are internally
ranked for detection purposes by {\textsc{cWB}} using the coherent network amplitude statistic $\rho$
defined in \cite{drago:11}.
Other constraints related to the degree of similarity of the reconstructed
signal across different interferometers (the ``network correlation coefficient'' $cc$)
and the ability of the network to reconstruct both polarizations of the GW signal
(called {\it{regulators}}) are applied to reject background events;
these are also described in \cite{drago:11}.
\subsection{{\textsc{X-Pipeline}}}
In the {\textsc{X-Pipeline}} \cite{sutton:10,Was:2012zq,chatterji:06} analysis, the
detector data are first whitened, then Fourier transformed. A total
energy map is made by summing the spectrogram for each detector, and
``hot'' pixels are identified as the 1\% in each detector with the
largest total energy. Hot pixels that share an edge or vertex
(nearest neighbors and next-nearest neighbors) are clustered. For each
cluster, the raw time-frequency maps are recombined in a number of
linear combinations designed to give maximum-likelihood estimates of
various GW polarizations given the known sky position of the CCSN. The
energy in each combination is recorded for each cluster, along with
various time-frequency properties of the cluster. The procedure is
repeated using a series of Fourier transform lengths from $1/4\,$s,
$1/8\,$s, \ldots $1/128\,$s. Clusters are ranked internally using a
Bayesian-inspired estimate $\Lambda_c$ of the likelihood
ratio for a circularly polarized GW, marginalized over the unknown GW
amplitude $\sigma_h$ with a Jeffreys (logarithmic) prior
$\sigma_h^{-1}$; see \cite{Searle:2007uv,searle:09,Was:2012zq} for
details.
When clusters from different Fourier transform lengths overlap in
time-frequency, the cluster with the largest likelihood $\Lambda_c$ is
retained and the rest are discarded. Finally, a post-processing
algorithm tunes and applies a series of pass/fail tests to reject
events due to background noise; these tests are based on measures of
correlation between the detectors for each cluster. The tuning of
these tests is described in detail in \cite{sutton:10}. For more
details see also \cite{gossan:16}.
\begin{table*}
\caption{
Injection waveforms from detailed multi-dimensional CCSN
simulations described in the text.
For each waveform, we give the
emission type, journal reference, waveform identifier,
angle-averaged root-sum-squared strain $h_\mathrm{rss}$,
the frequency $f_\mathrm{peak}$ at which the GW energy spectrum peaks, the
emitted GW energy $E_\mathrm{GW}$, and available polarizations.
See \cite{gossan:16,szczepanczykdcc:15} for details.
}
\begin{tabular}{lclcrcc}
\hline
\hline
\multicolumn{1}{c}{Emission Type}
&\multicolumn{1}{c}{Ref.}
&\multicolumn{1}{c}{Waveform Identifier}
&\multicolumn{1}{c}{$h_\mathrm{rss}$}
&\multicolumn{1}{c}{$f_\mathrm{peak}$}
&\multicolumn{1}{c}{$E_\mathrm{GW}$}
&\multicolumn{1}{c}{Polarizations} \\
&&
&\multicolumn{1}{c}{$[10^{-22} @ 10\,\mathrm{kpc}]$}
&\multicolumn{1}{c}{$[\mathrm{Hz}]$}
&\multicolumn{1}{c}{$[10^{-9} M_{\odot} c^{2}]$}\\
\hline
Rotating Core Collapse
&\cite{dimmelmeier:08}
&Dim1-s15A2O05ls
&1.052
&774
&\phantom{0}\phantom{0}7.685
&$+$\\
Rotating Core Collapse
&\cite{dimmelmeier:08}
&Dim2-s15A2O09ls
&1.803
&753
&\phantom{0}27.873
&$+$\\
Rotating Core Collapse
&\cite{dimmelmeier:08}
&Dim3-s15A3O15ls
&2.690
&237
&\phantom{0}\phantom{0}1.380
&$+$\\
\hline
2D Convection
&\cite{yakunin:10}
&Yakunin-s15
&1.889
&888
&\phantom{0}\phantom{0}9.079
&$+$\\
\hline
3D Convection
&\cite{mueller:e12}
&M\"uller1-L15-3
&1.655
&150
&\phantom{0}$3.741\times10^{-2}$
&$+$, $\times$\\
3D Convection
&\cite{mueller:e12}
&M\"uller2-N20-2
&3.852
&176
&\phantom{0}$4.370\times10^{-2}$
&$+$, $\times$\\
3D Convection
&\cite{mueller:e12}
&M\"uller3-W15-4
&1.093
&204
&\phantom{0}$3.247\times10^{-2}$
&$+$, $\times$\\
\hline
Protoneutron Star Pulsations
&\cite{ott:09}
&Ott-s15
&5.465
&971
&429.946
&$+$\\
\hline
\hline
\end{tabular}
\label{tab:numwaveforms}
\end{table*}
\subsection{Simulated Signals and Search Sensitivity}
\label{sec:simulations}
An important aspect of the GW search presented in this study is to
understand how sensitive the GW detector networks are to GWs emitted
by the considered CCSNe. We establish sensitivity via Monte Carlo
simulation in the following way:
\begin{enumerate}
\item We determine the loudest event in the on-source window that is
consistent with the CCSN location (and the angular uncertainty
of the search algorithms).
\item We ``inject'' (add) theoretical waveforms scaled to a specific distance
(or emitted GW energy) every $100\,\mathrm{s}$ plus a randomly
selected time in $[-10,10]\,\mathrm{s}$ into the time-shifted
background data. We compare the loudness of the recovered injections
with the loudest on-source event and record the fraction of the
injections that passed the coherent tests and data quality cuts and
were louder than the loudest on-source event. This
fraction is the \emph{detection efficiency}.
\item We repeat step (2) for a range of distances (or emitted GW energies)
to determine the detection efficiency as a function of distance (or
emitted GW energy).
\end{enumerate}
We refer the reader to~\cite{gossan:16} for more
details on the injection procedure.
In this paper, we employ three classes of GW signals for our Monte
Carlo studies: (1) representative waveforms from detailed
multi-dimensional (2D axisymmetric or 3D) CCSN simulations; (2)
semi-analytic phenomenological waveforms of plausible but extreme
emission scenarios; and (3) \textit{ad-hoc} waveform models with tuneable frequency
content and amplitude to establish upper limits on the energy emitted
in GWs at a fixed CCSN distance.
We briefly summarize the nature of these waveforms below.
We list all employed waveforms in
Tables~\ref{tab:numwaveforms} and \ref{tab:phenowaveforms} and
summarize their key emission metrics. In particular, we provide the
angle-averaged root-sum-squared GW strain,
\begin{equation}
\label{eq:hrss}
h_\mathrm{rss}= \sqrt{\int \left\langle h^2_+(t) + h^2_\times(t) \right\rangle_\Omega \mathrm{d}t}\,\, ,
\end{equation}
and the energy $E_\mathrm{GW}$ emitted in GWs, using the expressions given
in \cite{gossan:16}.
\subsubsection{Waveforms from Multi-Dimensional CCSN Simulations}
Rotation leads to a natural axisymmetric quadrupole (oblate)
deformation of the collapsing core. The tremendous acceleration at
core bounce and proto-neutron star formation results in a strong
linearly-polarized burst of GWs followed by a ring-down
signal. Rotating core collapse is the most extensively studied GW
emission process in the CCSN context (see, e.g.,
\cite{mueller:82,zwerger:97,dimmelmeier:02,kotake:03,ott:04,dimmelmeier:08,ott:07prl,abdikamalov:14}
and \cite{ott:09,kotake:13review,fryer:11} for reviews). For the
purpose of this study, we select three representative rotating core
collapse waveforms from the 2D general-relativistic study of
Dimmelmeier~\emph{et al.}~\cite{dimmelmeier:08}. The simulations
producing these waveforms used the core of a $15$-$M_\odot$ progenitor
star and the Lattimer-Swesty nuclear equation of state
\cite{lseos:91}. The waveforms are enumerated by Dim1--Dim3 prefixes
and are listed in Table~\ref{tab:numwaveforms}. They span the range
from moderate rotation (Dim1-s15A2O05ls) to extremely rapid rotation
(Dim3-s15A3O15ls). See \cite{dimmelmeier:08} for details on the
collapse dynamics and GW emission.
\begin{table*}[t]
\caption{
Injection waveforms from phenomenological and \textit{ad-hoc} emission
models described in the text.
For each waveform, we give the
emission type, journal reference, waveform identifier,
angle-averaged root-sum-squared strain $h_\mathrm{rss}$,
the frequency $f_\mathrm{peak}$ at which the GW energy spectrum peaks, the
emitted GW energy $E_\mathrm{GW}$, and available polarizations.
See \cite{gossan:16,szczepanczykdcc:15} for details.
As sine-Gaussian waveforms are \textit{ad-hoc}, they can be
rescaled arbitrarily and do not have a defined
physical distance or $E_\mathrm{GW}$ value.
}
\begin{tabular}{lclcrcc}
\hline
\hline
\multicolumn{1}{c}{Emission Type}
&\multicolumn{1}{c}{Ref.}
&\multicolumn{1}{c}{Waveform Identifier}
&\multicolumn{1}{c}{$h_\mathrm{rss}$}
&\multicolumn{1}{c}{$f_\mathrm{peak}$}
&\multicolumn{1}{c}{$E_\mathrm{GW}$}
&\multicolumn{1}{c}{Polarizations}\\
&&
&\multicolumn{1}{c}{$[10^{-20} @ 10\,\mathrm{kpc}]$}
&\multicolumn{1}{c}{$[\mathrm{Hz}]$}
&\multicolumn{1}{c}{$[M_{\odot} c^{2}]$}\\
\hline
Long-lasting Bar Mode
&\cite{ott:10dcc}
&LB1-M0.2L60R10f400t100
&\phantom{0}1.480
&800
&\phantom{0}$2.984\times10^{-4}$
&$+,\times$\\
Long-lasting Bar Mode
&\cite{ott:10dcc}
&LB2-M0.2L60R10f400t1000
&\phantom{0}4.682
&800
&\phantom{0}$2.979\times10^{-3}$
&$+,\times$\\
Long-lasting Bar Mode
&\cite{ott:10dcc}
&LB3-M0.2L60R10f800t100
&\phantom{0}5.920
&1600
&\phantom{0}$1.902\times10^{-2}$
&$+,\times$\\
Long-lasting Bar Mode
&\cite{ott:10dcc}
&LB4-M1.0L60R10f400t100
&\phantom{0}7.398
&800
&\phantom{0}$7.459\times10^{-3}$
&$+,\times$\\
Long-lasting Bar Mode
&\cite{ott:10dcc}
&LB5-M1.0L60R10f400t1000
&23.411
&800
&\phantom{0}$7.448\times10^{-2}$
&$+,\times$\\
Long-lasting Bar Mode
&\cite{ott:10dcc}
&LB6-M1.0L60R10f800t25
&14.777
&1601
&\phantom{0}$1.184\times10^{-1}$
&$+,\times$\\
\hline
Torus Fragmentation Instability
& \cite{piro:07}
&Piro1-M5.0$\eta$0.3
&\phantom{0}2.550
&2035
&\phantom{0}$6.773\times10^{-4}$
&$+,\times$\\
Torus Fragmentation Instability
& \cite{piro:07}
&Piro2-M5.0$\eta$0.6
&\phantom{0}9.936
&1987
&\phantom{0}$1.027\times10^{-2}$
&$+,\times$\\
Torus Fragmentation Instability
& \cite{piro:07}
&Piro3-M10.0$\eta$0.3
&\phantom{0}7.208
&2033
&\phantom{0}$4.988\times10^{-3}$
&$+,\times$\\
Torus Fragmentation Instability
& \cite{piro:07}
&Piro4-M10.0$\eta$0.6
&28.084
&2041
&\phantom{0}$7.450\times10^{-2}$
&$+,\times$\\
\hline
sine-Gaussian
&\cite{abadie:12s6burst}
&SG1-235HzQ8d9linear
&---
&235
&---
&$+$\\
sine-Gaussian
&\cite{abadie:12s6burst}
&SG2-1304HzQ8d9linear
&---
&1304
&---
&$+$\\
sine-Gaussian
&\cite{abadie:12s6burst}
&SG3-235HzQ8d9elliptical
&---
&235
&---
&$+,\times$\\
sine-Gaussian
&\cite{abadie:12s6burst}
&SG4-1304HzQ8d9elliptical
&---
&1304
&---
&$+,\times$\\
\hline
\hline
\end{tabular}
\label{tab:phenowaveforms}
\end{table*}
In non-rotating or slowly rotating CCSNe, neutrino-driven convection
and the standing accretion shock instability (SASI) are expected to
dominate the GW emission. GWs from convection/SASI have also been
extensively studied in 2D (e.g.,
\cite{jm:97,mueller:04,kotake:07a,murphy:09,kotake:09,marek:09b,mueller:13gw,yakunin:10,yakunin:15})
and more recently also in 3D \cite{mueller:e12,ott:13a}. For the
present study, we select a waveform from a 2D Newtonian (+
relativistic corrections) radiation-hydrodynamics simulation of a CCSN
in a $15$-$M_\odot$ progenitor by Yakunin~\emph{et
al.}~\cite{yakunin:10}. This waveform and its key emission metrics
are listed as Yakunin-s15 in Table~\ref{tab:numwaveforms}. Note that
since the simulation producing this waveform was axisymmetric, only
the $+$ polarization is available.
CCSNe in Nature are 3D and produce both GW polarizations ($h_+$ and
$h_\times$). Only a few GW signals from 3D simulations are presently
available. We draw three waveforms from the work of M\"uller~\emph{et
al}.~\cite{mueller:e12}. These and their key GW emission
characteristics are listed with M\"uller1--M\"uller3 prefixes in
Table~\ref{tab:numwaveforms}. Waveforms M\"uller1-L15-3 and
M\"uller2-W15-4 are from simulations using two different progenitor
models for a $15$-$M_\odot$ star. Waveform M\"uller2-N20-2 is from a
simulation of a CCSN in a $20$-$M_\odot$ star. Note that the
simulations of M\"uller~\emph{et al.}~\cite{mueller:e12} employed an
\textit{ad-hoc} inner boundary at multiple tens of kilometers. This prevented
decelerating convective plumes from reaching small radii and high
velocities. As a consequence, the overall GW emission in these
simulations peaks at lower frequencies than in simulations that do not
employ an inner boundary
(cf.~\cite{yakunin:10,yakunin:15,mueller:13gw,ott:13a}).
In some 2D CCSN simulations \cite{burrows:06,burrows:07a}, strong
excitations of an $\ell = 1$ \emph{g}-mode (an oscillation mode with
gravity as its restoring force) were observed. These oscillations were
found to be highly non-linear and to couple to GW-emitting $\ell = 2$
modes. The result is a strong burst of GWs that lasts for the duration
of the large-amplitude mode excitation, possibly for hundreds of
milliseconds \cite{ott:09,ott:06prl}. More recent simulations do not
find such strong \emph{g}-mode excitations (e.g.,
\cite{marek:09,mueller:13gw}). We nevertheless include here one
waveform from the simulations of \cite{burrows:07a} that was
reported by Ott~\cite{ott:09}. This waveform is from a simulation with
a $15$-$M_\odot$ progenitor and is denoted as Ott-s15 in
Table~\ref{tab:numwaveforms}.
\subsubsection{Phenomenological Waveform Models}
In the context of rapidly rotating core collapse, various
non-axisymmetric instabilities can deform the proto-neutron star into
a tri-axial (``bar'') shape (e.g.,
\cite{lai:95,brown:01,shibata:05,rmr:98,rotinst:05,ott:07prl,scheidegger:10b}),
potentially leading to extended ($\sim$$10\,\mathrm{ms} -
\,\mathrm{few}\,\mathrm{s}$) and energetic GW emission. This emission
occurs at twice the proto-neutron star spin frequency and with
amplitude dependent on the magnitude of the bar deformation
\cite{fryer:02,scheidegger:10b,ott:07prl}. Since few long-term 3D
simulations are available, we resort to the simple phenomenological
bar model described in \cite{ott:10dcc}. Its parameters are the length
of the bar deformation, $L$, in km, its radius, $R$, in km, the mass,
$M$, in $M_\odot$, involved in the deformation, the spin frequency,
$f$, and the duration, $t$, of the deformation. We select six
waveforms as representative examples. We sample the potential
parameter space by chosing $M = \{0.2,1.0\}\,M_\odot$, $f =
\{400,800\}\,\mathrm{Hz}$, and $t = \{25, 100,
1000\}\,\mathrm{ms}$. We list these waveforms as ``Long-lasting Bar
Mode'' in Table~\ref{tab:phenowaveforms} and enumerate them as
LB1--LB6. The employed model parameters are encoded in the full
waveform name. One notes from Table~\ref{tab:phenowaveforms} that the
strength of the bar-mode GW emission is orders of magnitude greater
than that of any of the waveforms computed from detailed
multi-dimensional simulations listed in Table~\ref{tab:numwaveforms}.
We emphasize that the phenomenological bar-mode waveforms should be
considered as being at the extreme end of plausible GW emission
scenarios. Theoretical considerations (e.g., \cite{ott:09}) suggest
that such strong emission is unlikely to obtain in CCSNe.
Observationally, however, having this emission in one
or all of the CCSNe has not been ruled out.
We also consider the phenomenological waveform model proposed by Piro
\& Pfahl \cite{piro:07}. They considered the formation of a dense
self-gravitating $M_\odot$-scale fragment in a thick accretion torus
around a black hole in the context of collapsar-type gamma-ray
bursts. The fragment is driven toward the black hole by a combination
of viscous torques and energetic GW emission. This is an extreme but
plausible scenario. We generate injection waveforms from this model
using the implementation described in \cite{santamaria:11dcc}. The
model has the following parameters: mass $M_\mathrm{BH}$ of the black
hole in $M_\odot$, a spatially constant geometrical parameter
controlling the torus thickness, $\eta = H/r$, where $H$ is the disk
scale height and $r$ is the local radius, a scale factor for the
fragment mass (fixed at $0.2$), the value of the phenomenological
$\alpha$-viscosity (fixed at $\alpha = 0.1$), and a starting radius
that we fix to be $100 r_g = 100 GM_\mathrm{BH}/c^2$. We employ four
waveforms, probing black hole masses $M_\mathrm{BH} =
\{5,10\}\,M_\odot$ and geometry factors $\eta = \{0.3, 0.6 \}$. The
resulting waveforms and their key emission metrics are listed as
``Torus Fragmentation Instability'' and enumerated by Piro1--Piro4 in
Table~\ref{tab:phenowaveforms}. The full waveform names encode the
particular parameter values used. As in the case of the bar-mode
emission model, we emphasize that also the Torus Fragmentation
Instability represents an extreme GW emission scenario for CCSNe. It
may be unlikely based on theoretical considerations (e.g.,
\cite{ott:09,santamaria:11dcc}), but has not been ruled out
observationally.
\subsubsection{Ad-Hoc Waveforms: sine-Gaussians}
Following previous GW searches, we also employ \textit{ad-hoc} sine-Gaussian
waveforms to establish frequency-dependent upper limits on the emitted
energy in GWs. This also allows us to compare the sensitivity of our
targeted search with results from previous all-sky searches for GW
bursts (e.g., \cite{ligo_burst_s5y1:09,S5y2Burst,S6Burst})
Sine-Gaussian waveforms are, as the name implies, sinusoids in a
Gaussian envelope. They are analytic and given by
\begin{flalign}
\label{eqn:sgp}
h_+(t)&= A \frac{1+\alpha^2}{2} \exp {(-t^2/\tau^2)} \sin(2\pi f_0 t) \, , \\
\label{eqn:sgc}
h_\times(t)&= A \alpha \exp {(-t^2/\tau^2)} \cos(2\pi f_0 t) \, .
\end{flalign}
Here, $A$ is an amplitude scale factor, $\alpha=\cos\iota$ is the
ellipticity of the waveform with $\iota$ being the inclination angle,
$f_0$ is the central frequency, and $\tau=Q/(\sqrt{2}\pi f_0)$, where
$Q$ is the quality factor controlling the width of the Gaussian and
thus the duration of the signal. Since the focus of our study is more
on realistic and phenomenological waveforms, we limit the set of
sine-Gaussian waveforms to four, enumerated SG1--SG4 in
Table~\ref{tab:phenowaveforms}. We fix $Q=8.9$ and study linearly
polarized ($\cos\iota = 0$) and elliptically polarized ($\cos\iota$
sampled uniformly on $[-1,1]$) waveforms at $f = \{235,
1304\}\,\mathrm{Hz}$. We choose this quality factor and these
particular frequencies for comparison with
\cite{ligo_burst_s5y1:09,S5y2Burst,S6Burst}.
\subsection{Systematic Uncertainties}
Our efficiency estimates are subject to a number of uncertainties.
The most important of these are calibration uncertainties in the
strain data recorded at each detector, and Poisson uncertainties
due to the use of a finite number of injections (Monte Carlo uncertainties).
We account for each of these uncertainties in the sensitivities reported
in this paper.
We account for Poisson uncertainties from the finite number of injections
using the Bayesian technique described in \cite{Paterno:2004cb}.
Specifically, given the total number of injections performed at some
amplitude and the number detected, we compute the 90\% credible
lower bound on the efficiency assuming a uniform prior on $[0,1]$
for the efficiency. All efficiency curves reported in this paper are
therefore actually 90\% confidence level lower bounds on the efficiency.
Calibration uncertainties are handled by rescaling quoted $h_\mathrm{rss}$ and
distance values following the method in \cite{ligo_hfburst:09}.
The dominant effect is from the uncertainties in the amplitude calibration;
these are estimated at approximately 10\% for G1, H1, and H2,
14\% for L1, and 6\%-8\% for V1 at
the times of the two CCSNe studied~\cite{Abadie2010223,marion:08}.
The individual detector amplitude uncertainties are combined into a single uncertainty
by calculating a combined root-sum-square signal-to-noise ratio and
propagating the individual uncertainties assuming each error is independent
(the signal-to-noise ratio is used as a proxy for the
loudness measures the two pipelines use for ranking events).
This combination depends upon the relative sensitivity of each detector,
which is a function of frequency, so we compute the total uncertainty at
a range of frequencies across our analysis band for each CCSN and select the largest
result, 7.6\%, as a conservative estimate of the total 1-$\sigma$ uncertainty.
This 1-$\sigma$ uncertainty is then scaled by a factor of 1.28
(to 9.7\%) to obtain the factor by which our amplitude and distance
limits must be rescaled in order to obtain values consistent with a 90\%
confidence level upper limit.
\section{Search Results}\label{sec:results}
As discussed in Section~\ref{sec:overview}, on-source events from
each combination of CCSN, detector network, and pipeline are assigned
a false alarm rate by comparing to time-slide background events.
Table~\ref{fapcwb3} lists the FAR values of the loudest event found
by each pipeline for each network and CCSN. The lowest FAR,
1.7\,$\times10^{-6}$\,Hz, was reported by {\textsc{cWB}} for the analysis of
SN~2007gr with the H1H2L1V1 network. This rate can be converted to a
false alarm probability (FAP) using equation (\ref{eqn:fap}).
The total duration of data processed by {\textsc{cWB}} or {\textsc{X-Pipeline}} for the two
CCSNe was $T_\mathrm{on} = 873461$\,s. Equation (\ref{eqn:fap}) then
yields a false-alarm probability of 0.77 for the loudest event; this is
consistent with the event being due to background noise.
We conclude
that none of the events has a FAP low enough to be considered as a
candidate GW detection.
\begin{table}[htb]
\label{fapcwb3}
\centering
\caption{False alarm rate (FAR) of the loudest event found by each pipeline
for each detector network. No on-source events survived the coherent tests
and data quality cuts for the {\textsc{cWB}} analysis of the H1H2L1 and H1H2 networks
for SN~2007gr. The lowest FAR, 1.7\,$\times10^{-6}$\,Hz, corresponds to a FAP of 0.77.}
\begin{tabular}{l@{\hspace*{1em}}l@{\hspace*{1em}}l}
\hline
\hline
\multicolumn{1}{c}{Network} & \multicolumn{1}{c}{{\textsc{cWB}}} & \multicolumn{1}{c}{{\textsc{X-Pipeline}}} \\
\hline
H1H2L1V1 & 1.7\,$\times10^{-6}$\,Hz & 2.5\,$\times10^{-6}$\,Hz \\
H1H2L1 & no events & 1.1\,$\times10^{-5}$\,Hz \\
H1H2V1 & 1.2\,$\times10^{-5}$\,Hz & 5.3\,$\times10^{-6}$\,Hz \\
H1H2 & no events & 7.1\,$\times10^{-5}$\,Hz \\
L1V1 & 4.8\,$\times10^{-5}$\,Hz & 4.1\,$\times10^{-3}$\,Hz \\
G1V1 & 1.2\,$\times10^{-5}$\,Hz & 2.7\,$\times10^{-5}$\,Hz \\ \hline
\hline
\end{tabular}
\end{table}
We note that the loudest events reported by {\textsc{cWB}} and {\textsc{X-Pipeline}} are both from
the analysis of SN~2007gr with the H1H2L1V1 network; this is consistent with
chance as this network combination accounted for more than 60\% of the data
processed. In addition, the times of the loudest X-pipeline and cWB events differ by more than a day,
so they are not due to a common physical trigger.
\subsection{Detection efficiency vs.~distance}
Given the loudest event, we can compute detection efficiencies for the
search following the procedure detailed in Section~\ref{sec:simulations}.
In brief, we measure the fraction of simulated signals that produce
events surviving the coherent tests and data quality cuts and which have
a FAR (or equivalently FAP) lower than the loudest event.
Figures~\ref{fig:sn2007gr} and \ref{fig:sn2011dh} show the efficiency
as a function of distance for the CCSN waveforms from
multi-dimensional simulations and the phenomenological waveforms
discussed in Section \ref{sec:simulations} and summarized in
Tables~\ref{tab:numwaveforms} and \ref{tab:phenowaveforms}. For
SN~2007gr, the maximum distance reach is of order 1\,kpc for waveforms
from detailed multi-dimensional CCSN simulations, and from $\sim$100\,kpc to
$\sim$1\,Mpc for GWs from the phenomenological models (torus fragmentation
instability and long-lived rotating bar mode). The variation in
distance reach is due to the different peak emission frequencies of
the models and the variation in detector sensitivities with frequency
as shown in Figure~\ref{fig:noise}. The distance reaches for
SN~2011dh are lower by a factor of several than those for SN~2007gr;
this is due to the difference in sensitivity of the operating
detectors, as also evident in Figure~\ref{fig:noise}. Finally, we
note that at small distances the efficiencies asymptote to the
fraction of the on-source window that is covered by coincident data,
approximately 93\% for SN~2007gr and 37\% for SN~2011dh (up to a few
percent of simulated signals are lost due to random coincidence with
data quality cuts).
We do not show the efficiencies for the multi-dimensional simulation CCSN waveforms for SN~2011dh,
as the detection efficiency was negligible in this case. This is due to the fact that
the relative orientation of the G1 and V1 detectors -- rotated approximately $45^\circ$
with respect to each other -- means that the two detectors are sensitive to orthogonal
GW polarizations. In order for the coherent cuts to reject background noise
{\textsc{X-Pipeline}} needs to assume some relationship between these two polarizations.
We require that the $h_+$ and $h_\times$ polarizations are out of phase by $90^\circ$,
as would be expected for emission from a rotating body with a non-axisymmetric
quadrupole deformation. We choose this because the strongest GW emission models
are for rotating non-axisymmetric systems (the fragmentation instability and long-lived bar mode).
Unfortunately, the waveforms from multi-dimensional CCSN simulations are either
linearly polarized (i.e.~have only one polarization) or exhibit
randomly changing phase. Hence, they cannot be detected by the
search performed for SN~2011dh with {\textsc{X-Pipeline}}.
The tuning of {\textsc{cWB}} did not use these constraints, however the G1 noise floor was
about a factor of 2 higher than V1 around 1000\,Hz and the difference was even greater at lower frequencies.
This issue weakened the internal {\textsc{cWB}} measures of correlation of the reconstructed
signal between the two interferometers and severely reduced the detection efficiencies
at distances beyond a few parsecs.
The distances shown in Figures~\ref{fig:sn2007gr} and \ref{fig:sn2011dh}
show the probability of a GW signal producing an event with FAP lower than
that of the loudest event.
The physical interpretation of the efficiency $\epsilon$ at a distance $d$ for a
given model is related to the prospect of excluding the model with observations.
Explicitly, the non-observation of any events with FAP lower than the
loudest event gives a frequentist exclusion of that GW emission model for a
source at distance $d$ with confidence $\epsilon$.
However, in this search the loudest event had a large FAP (0.77).
In order for an event to be considered as a possible detection it would need to
have a FAP of order $10^{-3}$ or less; we find that imposing this more stringent
requirement lowers the maximum distance reach by approximately 5\%-25\%
depending on the waveform model.
Unfortunately, none of the models have distance reaches out to the $\sim$10\,Mpc distance
of SN~2007gr or SN~2011dh; we conclude that our search is not able to constrain
the GW emission model for either of these CCSNe.
\begin{figure*}[t]
\begin{minipage}[c][][t]{0.495\textwidth}
\vspace*{\fill}
\flushleft
\includegraphics[width=0.96\linewidth]{figures/EffVsDist/OR_SN2007gr_mueller5_eff-vs-distance_replot.pdf}
\includegraphics[width=0.96\linewidth]{figures/EffVsDist/OR_SN2007gr_numerical3_eff-vs-distance_replot.pdf}
\end{minipage}
\begin{minipage}[c][][t]{0.495\textwidth}
\vspace*{\fill}
\flushright
\includegraphics[width=0.96\linewidth]{figures/EffVsDist/OR_SN2007gr_piro5_eff-vs-distance_replot.pdf}
\includegraphics[width=0.96\linewidth]{figures/EffVsDist/OR_SN2007gr_rotbar5_eff-vs-distance_replot.pdf}
\end{minipage}
\caption{\label{fig:sn2007gr}
SN~2007gr detection efficiency versus distance for the
waveforms from multi-dimensional CCSN simulations (left) and the phenomenological waveforms (right)
described in Tables~\ref{tab:numwaveforms} and \ref{tab:phenowaveforms}.
Simulated GW signals are added into detector data with a range of
amplitudes corresponding to different source distances. A simulated signal
is considered detected if {\textsc{cWB}} or {\textsc{X-Pipeline}} reports an event that survives
the coherent tests and data quality cuts with a FAR value lower than
that of the loudest event from the SN~2007gr and SN~2011dh on-source windows.
These efficiencies are averaged over all detector network combinations for SN~2007gr.
The efficiencies are limited to $\le93\%$ at small distances due to the fact that
this was the duty cycle for coincident observation over the SN~2007gr on-source window.
The numbers in brackets for each model are the distances at which the efficiency equals $50\%$
of the asymptotic value at small distances. }
\end{figure*}
\begin{figure*}[t]
\begin{minipage}[c][][t]{0.495\textwidth}
\vspace*{\fill}
\flushleft
\includegraphics[width=0.96\linewidth]{figures/EffVsDist/OR_SN2011dh_piro_long_eff-vs-distance_replot.pdf}
\end{minipage}
\begin{minipage}[c][][t]{0.495\textwidth}
\vspace*{\fill}
\flushright
\includegraphics[width=0.96\linewidth]{figures/EffVsDist/OR_SN2011dh_rotbar_long_eff-vs-distance_replot.pdf}
\end{minipage}
\caption{\label{fig:sn2011dh}
SN~2011dh detection efficiency versus distance for the phenomenological waveforms
described in Table~\ref{tab:phenowaveforms}.
Simulated GW signals are added into detector data with a range of
amplitudes corresponding to different source distances. A simulated signal
is considered detected if either {\textsc{cWB}} or {\textsc{X-Pipeline}} reports an event that survives
the coherent tests and data quality cuts with a FAR value lower than
that of the loudest event from the SN~2007gr and SN~2011dh on-source windows.
The efficiencies are limited to $\le37\%$ at small distances due to the fact
that this was the duty cycle for coincident observation over the SN~2011dh
on-source window; some simulations are also vetoed by data quality cuts.
The numbers in the brackets are the distances at which the efficiency
equals $50\%$ of its maximum value for each model. }
\end{figure*}
\subsection{Constraints on Energy Emission}
In addition to the astrophysically motivated phenomenological and
multi-dimensional CCSN simulation waveforms, we employ the \textit{ad-hoc}
sine-Gaussian waveforms specified by equations (\ref{eqn:sgp}) and (\ref{eqn:sgc})
to establish frequency-dependent upper limits on the emitted energy in GWs.
This also allows us to compare the sensitivity of our
targeted search with results from previous all-sky searches for GW
bursts (e.g., \cite{ligo_burst_s5y1:09,S5y2Burst,S6Burst})
The detection efficiency is computed using the same procedure as for
the other waveforms. However, since these \textit{ad hoc} waveforms
have no intrinsic distance scale, we measure the efficiency as a
function of the root-sum-square amplitude $h_\mathrm{rss}$, defined by
equation~(\ref{eq:hrss}). For this study, we use the two
sine-Gaussian waveforms described in Section~\ref{sec:simulations},
which have central frequencies of 235\,Hz and 1304\,Hz. These are
standard choices for all-sky burst searches~\cite{S6Burst}.
Table~\ref{tab:energy} lists the $h_\mathrm{rss}$ values at which the
efficiency reaches half of its maximum value. Note that we use the
half-maximum efficiency rather than 50\% efficiency here, since the
maximum efficiency is limited by the fraction of the on-source window
that is covered by coincident data. The half-maximum gives a measure
of the distance reach of the instruments independent of their duty cycle.
These $h_\mathrm{rss}$ values can be converted to limits on energy emission by
assuming a specific angular emission pattern of the source
\cite{sutton:13}. For simplicity, we assume isotropic emission, for
which
\begin{equation}
\label{eqn:energy}
E_\mathrm{GW} = \frac{\pi^2 c^3}{G}D^2 f^2_0h_\mathrm{rss}^2 \, .
\end{equation}
Here $f_0$ is the peak frequency of the GW and $D$ is the distance of
the source. We use distances of 10.55\,Mpc for SN 2007gr and 8.40\,Mpc
for SN 2011dh. Table~\ref{tab:energy} also lists the energy emission
values at which the efficiency reaches half of its maximum value. If
the total amount of energy emitted in GWs was larger than the numbers
quoted in the Table, we would have had a greater than $50\%$ chance of
seeing a signal from the CCSN at the estimated distance, provided
coincident observation with the most sensitive detector network. Note,
however, that the on-source window did not have $100\%$ coverage (see
Section~\ref{sec:networks}).
The most stringent constraints are a few percent of a mass-energy
equivalent of a solar mass emitted in GWs at 235\,Hz, where the noise
floor is low. The 1304\,Hz results indicate that with this data
set, we should not expect to be able to detect extra-Galactic GWs at
kHz frequencies, since the limits are less stringent, $O(10)$~$M_\odot
c^2$ or more.
The above results can be compared with the energy available in CCSNe,
which are powered by the gravitational energy released in core
collapse. The total available energy is set by the binding energy of a
typical $1.4\,M_\odot$ neutron star and is roughly
$3\times10^{53}\,\mathrm{erg}$, corresponding to $\sim$$0.15\,M_\odot
c^2$ (e.g., \cite{lattimer:01}). The observation of neutrinos from
SN~1987A confirmed that $\sim$$99\%$ of that energy is emitted in the
form of neutrinos in proto-neutron star cooling (e.g.,
\cite{vissani:15}). The typical CCSN explosion kinetic energy is
$\sim$$10^{51}\,\mathrm{erg}$ ($\sim$$10^{-3}\,M_\odot
c^2$). Considering these observational constraints, the energy emitted
in GWs is unlikely to exceed $O(10^{-3})\,M_\odot c^2$. Hence, the energy
constraints obtained by this search for SNe 2007gr and 2011dh are not
astrophysically interesting.
\begin{table*}
\caption{Gravitational-wave energy emission constraints at
half-maximum detection efficiency for SN~2007gr and SN~2011dh.
These assume distances of 10.55\,Mpc for SN 2007gr and 8.40\,Mpc for
SN 2011dh.
\label{tab:energy}}
\begin{center}
\begin{tabular}{r|c|c|c|c|c|c}
\hline
\hline
\multirow{2}{*}{Waveform~~}& \multicolumn{3}{|c|}{SN~2007gr} & \multicolumn{3}{|c}{SN~2011dh} \\ \cline{2-7}
& $h_\mathrm{rss}$\,[Hz${}^{-1/2}$] & $E_\mathrm{GW}\,[\mathrm{erg}]$ & $E_\mathrm{GW}\,[M_{\odot}c^2]$
& $h_\mathrm{rss}$\,[Hz${}^{-1/2}$] & $E_\mathrm{GW}\,[\mathrm{erg}]$ & $E_\mathrm{GW}\,[M_{\odot}c^2]$ \\ \hline
SGel2 SG235Q9 & $5.4\times10^{-22}$ & 6.7$\times10^{52}$ & 0.038 & $9.1\times10^{-21}$ & 1.2$\times10^{55}$ & 6.8 \\ \hline
SGlin2 SG235Q9 & $6.6\times10^{-22}$ & 1.0$\times10^{53}$ & 0.058 & $4.8\times10^{-20}$ & 3.4$\times10^{56}$ & 1.9$\times10^2$ \\ \hline
SGel2 SG1304Q9 & $2.1\times10^{-21}$ & 3.1$\times10^{55}$ & 17 & $2.2\times10^{-21}$ & 2.3$\times10^{55}$ & 13 \\ \hline
SGlin2 SG1304Q9 & $2.5\times10^{-21}$ & 4.6$\times10^{55}$ & 26 & n/a & n/a & n/a \\ \hline
\hline
\end{tabular} \label{table:res}
\end{center}
\end{table*}
\subsection{Model Exclusion Confidence}
As we have seen, it is unlikely that we will have coincident science-quality
data covering an entire multi-day on-source window for any given CCSN.
In the present analysis, the coverage of the on-source windows is
approximately 93\% for SN~2007gr and 37\% for SN~2011dh. Considering that
data-quality cuts typically remove another few percent of livetime, we
cannot expect to exclude even fairly strong GW emission at the 90\% confidence level
for a single CCSN. However, by combining observations of multiple CCSNe, it
is straightforward to exclude the simple model in which all CCSNe produce
identical GW signals; i.e., assuming \textit{standard-candle} emission.
Consider a CCSN model $M_{\mathrm{SN}}$ which predicts a particular GW
emission pattern during the CCSN event (e.g., one of the waveforms
considered in Section~\ref{sec:simulations}).
In the case that no GW candidates are observed, we can
constrain that model using observations from multiple CCSN events at
known distances $d_i$ using the measured detection efficiencies
$\epsilon_i(d_i)$ for each supernova (e.g., as in Figure~\ref{fig:sn2007gr}).
These $\epsilon_i(d_i)$ can be combined into an overall model exclusion
probability~\cite{kalmus:13a}, $P_\mathrm{excl}$:
\begin{equation}
P_\mathrm{excl} = 1 - \prod_{i =1}^N (1-\epsilon_i(d_i)) \label{eq:reach}
\end{equation}
It is also straightforward to marginalize over uncertainties in the $d_i$
(as in Table~\ref{table:triggers}) by the replacement
\begin{equation}
\label{eqn:marginalise}
\epsilon_i(d_i) \to \epsilon_i \equiv \int_0^\infty \!\!\!d\bar{d} \,\pi_i(\bar{d}) \epsilon_i(\bar{d})
\end{equation}
where $\pi_i$ is our prior on the distance to CCSN $i$ (e.g., a Gaussian).
In the light of the measured sensitivity ranges in
Figures~\ref{fig:sn2007gr} and \ref{fig:sn2011dh}, it is clear that we
cannot exclude any of the considered models of GW emission for
SN~2007gr and SN~2011dh with the current data. However, iLIGO and
Virgo are being upgraded to advanced configurations, with a final
design sensitivity approximately a factor of ten better than for the
period 2005-2011 considered in this paper. It is therefore instructive
to consider what model exclusion statements the advanced detectors
will be able to make using future CCSNe similar to SN~2007gr and
SN~2011dh. We will focus on the phenomenological waveform models of plausible but more
extreme GW emission, where we expect to reach sooner large standard candle
model exclusion probabilities.
Specifically, we analyze the rotating bar and torus fragmentation scenarios
(see also the discussion in \cite{gossan:16}).
Figure~\ref{fig:exclusion} presents model exclusion confidence plots
for four of the phenomenological waveform models. These plots are
based on the measured efficiencies for SN~2007gr and SN~2011dh, but
assume the detector noise spectra have been lowered by a factor of
$A$, so the search would be expected to have the same efficiency for a
particular source at $A$-times the distance, and the number of CCSNe
in the sample has been increased by a factor of $p$. For example,
$A=10$ represents having a sensitivity 10 times better than the
2005--2011 data, which is realistic for Advanced LIGO and Advanced
Virgo, while $p=2$ corresponds to having two CCSNe similar to
SN~2007gr and two similar to SN~2011dh. The curves correspond to the
experimentally derived values based on the 2005-2011 data set. It is
worth stressing that the power of excluding models from this data set
almost exclusively originates from SN~2007gr, given the more sensitive
interferometers available at the time of that supernova. For example,
in the bottom left panel of Fig.~\ref{fig:exclusion}, the curves, when
A is smaller than 9, depend almost exclusively on the contribution of
SN~2007gr. In this regard, the presented model exclusion probabilities
will be reached with less than $2p$ CCSNe if the sample contains more
data sets comparable in coverage and sensitivity to the rescaled
SN~2007gr data set than a rescaled SN~2011dh data set. In summary,
Fig.~\ref{fig:exclusion} shows that it is a reasonable expectation
that extended coincident observations with advanced-generation
detectors will rule out extreme CCSN emission models.
\begin{figure*}[t]
\begin{minipage}[c][][t]{0.495\textwidth}
\vspace*{\fill}
\flushleft
\includegraphics[width=0.96\linewidth]{figures/ModelExclusion/rotbar_M1p2L60R10f400t1000.pdf}
\includegraphics[width=0.96\linewidth]{figures/ModelExclusion/rotbar_M1p2L60R10f800t125.pdf}
\end{minipage}
\begin{minipage}[c][][t]{0.495\textwidth}
\vspace*{\fill}
\flushright
\includegraphics[width=0.96\linewidth]{figures/ModelExclusion/piro_M5d0eta0d6fac0d2.pdf}
\includegraphics[width=0.96\linewidth]{figures/ModelExclusion/piro_M10d0eta0d6fac0d2.pdf}
\end{minipage}
\caption{\label{fig:exclusion}
Expected model exclusion probabilities for example waveforms as a function
of amplitude sensitivity rescaling, $A$, and supernova sample
size rescaling, $p$, based on the SN~2007gr and SN~2011dh sample
(e.g. $p=5$ corresponds to 10 supernovae).
The naming convention is described in Table \ref{tab:phenowaveforms}.
Currently none of the emission models can be exluded,
but for the advanced detectors with better sensitivity and
more nearby CCSNe it is realistic to expect to
rule out some of the extreme emission models.
}
\end{figure*}
\subsection{Sensitivity Advantage of the Triggered Search}
As noted in Section~\ref{sec:introduction}, targeted searches have the
advantage over all-time all-sky searches that potential signal candidates in the
data streams have to arrive in a well-defined temporal on-source window
and have to be consistent with coming from the sky location of the source.
Both constraints can significantly reduce the noise background.
Here we assess the improved sensitivity of a triggered search
by comparing our $h_\mathrm{rss}$ sensitivities to linearly polarized sine-Gaussian
waveforms for SN~2007gr to those of an all-sky search of the same data.
The most straightforward way to compare two searches is to fix the FAR
threshold and compare the $h_\mathrm{rss}$ values at 50\% efficiency.
The S5/VSR1 all-sky all-time search \cite{S5y2Burst} using {\textsc{cWB}} was run
on 68.2 days of coincident H1H2L1V1 data with thresholds to give
a FAP of 0.1 or less in the frequency band up to 2000\,Hz.
The livetime for the {\textsc{cWB}} SN~2007gr analysis of the H1H2L1V1 network
was 3.25 days,
so a FAP of 0.1 corresponds to a FAR of $3.56\times10^{-7}$\,Hz.
Including calibration and Monte Carlo uncertainties,
the ${h_\mathrm{rss}}$ values at 50\% efficiency for this FAR are
$5.0\times10^{-22}\,\mathrm{Hz}^{-1/2}$ at 235\,Hz and
$2.2\times10^{-21}\,\mathrm{Hz}^{-1/2}$ at 1304\,Hz.
After adjusting for systematic differences in the antenna responses
and noise spectra\footnote{In particular, during the on-source window of SN~2007gr the noise spectral
density for L1 was about 50 percent worse at low frequencies than the average value
during the whole of S5.} between the S5/VSR1 all-sky search and the SN~2007gr
search, the effective all-sky ${h_\mathrm{rss}}$ values are
$7.0\times10^{-22}\,\mathrm{Hz}^{-1/2}$ at 235\,Hz and
$2.9\times10^{-21}\,\mathrm{Hz}^{-1/2}$ at 1304\,Hz,
approximately 30\% to 40\% higher than the targeted search.
Equivalently, the distance reach of our targeted search is larger
than that of the all-time all-sky search by 30\% to 40\% at this FAP.
Alternatively, we can compare the two searches without adjusting to
a common FAR. After allowing for systematic differences in the antenna
responses and noise spectra between the S5/VSR1 all-sky search and the
SN~2007gr search, we find that the ${h_\mathrm{rss}}$ values at 50\% efficiency
are identical (to within a few percent).
However, the FAR of the SN~2007gr search is lower by an order of magnitude:
$1.8\times10^{-9}\,$Hz compared to $1.7\times10^{-8}\,$Hz for the all-sky search.
This is consistent with expectations for restricting from an all-sky
search to a single sky-position search.
Furthermore, the FAP for a trigger produced by the SN~2007gr search will
be smaller than that of a trigger from the all-sky search at the same FAR
because the SN~2007gr on-source window (3.5 days for {\textsc{cWB}} and {\textsc{X-Pipeline}} combined) is a factor of 20 shorter
than the all-sky window (68.2 days).
So if we consider a surviving trigger
that is just above threshold in the two searches, the SN~2007gr trigger
will have an FAP a factor of approximately 200 lower than an all-sky trigger
with the same $h_\mathrm{rss}$.
\section{Summary and Discussion}
\label{sec:summaries}
We presented the results of the first iLIGO-GEO-Virgo search for
gravitational-wave (GW) transients in coincidence with optically
detected core-collapse supernovae (CCSNe) observed between 2007 and
2011. Two CCSNe, SN~2007gr and SN~2011dh, satisfied our criteria of
proximity, well-constrained time of core collapse, and occurrence
during times of coincident high-sensitivity operation of at least two
GW detectors. No statistically significant GW events were observed
associated with either CCSN.
We quantified the sensitivity of the search as a function of distance
to the CCSNe using both representative waveforms from detailed
multi-dimensional CCSN simulations and from semi-analytic
phenomenological models of plausible but extreme emission
scenarios. The distances out to which we find signals detectable for
SNe 2007gr and 2011dh range from $O(\lesssim 1)$\,kpc for waveforms from
detailed simulations to $O(1)$\,Mpc for the more extreme phenomenological models. From
the known distances of our two target supernovae, we estimated the
minimum energy in gravitational waves corresponding to our sensitivity
limits using \textit{ad-hoc} sine-Gaussian waveforms. These range from
$O(0.1)\,M_\odot c^2$ at low frequencies to $\gtrsim O(10)\,M_\odot
c^2$ above 1\,kHz.
This first search for GWs from extragalactic CCSNe places the
most stringent observational constraints to-date on GW emission in
core-collapse supernovae. A comparison of our search's sensitivity
with the standard all-sky, all-time search for generic GW bursts in
the same GW detector data shows a 35\%-40\% improvement in distance
reach at fixed false alarm probability. This improvement comes from
knowledge of the sky positions of the CCSNe and approximate
knowledge of the collapse times. It is, hence, clearly beneficial to carry
out targeted searches for GWs from CCSNe.
The results of our search do not allow us to constrain astrophysically
meaningful GW emission scenarios.
We have extrapolated our results to the
sensitivity level expected for Advanced LIGO and Virgo.
Considering the improved detector sensitivity and assuming the analysis of
multiple CCSNe, we find that at design sensitivity (c.~2019, \cite{Aasi:2013wya})
this network will be able to constrain
the extreme phenomenological emission models for extragalactic CCSNe
observed out to distances of $\sim$10\,Mpc. Detection of the most realistic GW
signals predicted by multi-dimensional CCSN simulations will require a
Galactic event even at the design sensitivity of the Advanced
detectors. These are consistent with the results of the study in \cite{gossan:16},
which used data from iLIGO and Virgo recoloured to match the
advanced detector design sensitivities. We conclude that
third-generation detectors with a sensitivity improvement
of a factor of $10-20$ over the Advanced detectors may be needed to
observe GWs from extragalactic CCSNe occurring at a rate of $1-2$ per
year within $\sim$10\,Mpc.
\begin{acknowledgments}\label{sec:acknowledgments}
We thank L.~Dessart for applying the expanding photosphere method to
SN 2008bk to derive an approximate explosion date and A.~Howell for
access to his supernova spectra fit software \textsc{superfit}. The authors
gratefully acknowledge the support of the United States National
Science Foundation for the construction and operation of the LIGO
Laboratory, the Science and Technology Facilities Council of the
United Kingdom, the Max-Planck-Society and the State of
Niedersachsen/Germany for support of the construction and operation of
the GEO\,600 detector, and the Italian Istituto Nazionale di Fisica
Nucleare and the French Centre National de la Recherche Scientifique
for the construction and operation of the Virgo detector. The authors
also gratefully acknowledge the support of the research by these
agencies and by the Australian Research Council, the Council of
Scientific and Industrial Research of India, the Istituto Nazionale di
Fisica Nucleare of Italy, the Spanish Ministerio de Educaci\'on y
Ciencia, the Conselleria d'Economia Hisenda i Innovaci\'o of the
Govern de les Illes Balears, the Foundation for Fundamental Research
on Matter supported by the Netherlands Organisation for Scientific
Research, the Polish Ministry of Science and Higher Education, the
FOCUS Programme of Foundation for Polish Science, the Royal Society,
the Scottish Funding Council, the Scottish Universities Physics
Alliance, the National Aeronautics and Space Administration, the
Carnegie Trust, the Leverhulme Trust, the David and Lucile Packard
Foundation, the Research Corporation, and the Alfred P. Sloan
Foundation.
This document has been assigned LIGO Laboratory document number \ligodoc.
\end{acknowledgments}
\bibliographystyle{apsrev}
|
\section{Introduction}
\label{sec:introduction}
The observed population of hot-Jupiters is thought to have migrated inward after their formation. The angle between the spin axis of the star and the orbit normal of the hot-Jupiter is a useful probe for the migration history of the planet. In the most simple interpretation, planets found in well-aligned orbits are thought to have migrated in the protoplanetary disk via planet-gas interactions \citep[e.g.][]{1996Natur.380..606L}, while those found in high obliquity orbits underwent dynamical interactions, such as planet-planet scattering \citep[e.g.][]{1996Sci...274..954R}, Kozai-Lidov induced eccentricity migration \citep[e.g.][]{2003ApJ...589..605W,2007ApJ...669.1298F}, or were born in primordially tilted disks \citep[e.g.][]{2010MNRAS.401.1505B,2012Natur.491..418B}. Of the 74 planets with spin-orbit measurements\footnote{Measured by the Rossiter-McLaughlin effect, a technique that does not impose strong selection biases on the spin-orbit orientation of the systems measured. Sample selected from R\'{e}ne Heller's Holt-Rossiter-McLaughlin Encyclopedia (\url{http://www2.mps.mpg.de/homes/heller/}). Where multiple spin-orbit angles are quoted, the authors examined the discovery paper and chose the most robust observation. Only hot-Jupiters are included.}, 23\% are found in misaligned orbits.
However, interpreting the misalignment statistic is made harder by potential post-migration evolution of the orbit geometry. It is thought that star-planet interactions, such as tidal and magnetic drag, can realign the spin direction of the convective envelope of a host star \citep{2010ApJ...718L.145W,2012MNRAS.423..486L,2013ApJ...769L..10R,2014ApJ...784...66X,2014ApJ...790L..31D}. This is supported by the observed trend that massive planets in close-in orbits around cooler stars tend to be aligned, while smaller planets around hotter stars (which lack convective envelopes), or at longer periods (where tidal forces are weak), exhibit a wide range of obliquity angles \citep[e.g.][]{2010ApJ...718L.145W,2010ApJ...719..602S,2012ApJ...757...18A,2015ApJ...801....3M}. It should be noted that this framework has some observational shortcomings. \citet{2015ApJ...801....3M} and \citet{2015arXiv151105570L} compared the photometric variability (a potential proxy for line-of-sight spin axis inclination) of planet-hosting stars to other stars of similar properties that do not host transiting hot-Jupiters. They found the stellar type -- spin-orbit angle trend persists at long periods, beyond the bounds of tidal interactions. In addition, short period planets around cool stars have been found in severe misalignment \citep{2010MNRAS.402L...1P,2015ApJ...814L..16Z}, these planets should have realigned the convective envelope of their stars under the tidal theory.
Within this tidal realignment framework, we can postulate that the primordial spin-orbit angles of planets around early-type stars are more likely recoverable, especially those that exhibit rapid rotation and have not yet been spun-down by planet-star interactions. Unlike late-type stars that undergo magnetic braking, stars hotter than $T_\text{eff} \sim 6250\,\text{K}$ do not spin-down significantly, and are generally more rapidly rotating. One key problem with characterising planets around rapidly rotating, early type stars is that precise radial velocity measurements needed for the Rossiter-McLaughlin effect \citep{1924ApJ....60...15R,1924ApJ....60...22M} are difficult to obtain.
In this study, we present Doppler tomographic analyses to measure the spin-orbit angles of two systems, orbiting rapidly rotating F-type stars. We present a new spectroscopic transit observation for the hot-Jupiter HAT-P-56b \citep{2015AJ....150...85H}, and a re-analysis of archival observations for the hot-Jupiter KELT-7b \citep{2015AJ....150...12B}. The key properties of these systems are presented in Table~\ref{tab:literature_properties}. They all orbit rapidly rotating F-dwarfs, with projected rotational velocities of $70$, and $38\,\text{km\,s}^{-1}$, respectively (see Section~\ref{sec:lsd}). Of all transiting hot-Jupiter hosts known, only the A-stars WASP-33 \citep{2010MNRAS.407..507C}, KOI-13 \citep{2014ApJ...790...30J}, and HAT-P-57 \citep{2015AJ....150..197H} have higher rotation rates.
\begin{table*}
\caption{\label{tab:literature_properties}Key properties of the KELT-7 and HAT-P-56 systems from the literature}
\begin{tabular}{lrr}
\hline\hline
& \textbf{KELT-7} & \textbf{HAT-P-56}\\
Source & \citet{2015AJ....150...12B} & \citet{2015AJ....150...85H}\\
\hline
RA & 05:13:11.0 & 06:45:24.0\\
DEC & +33:19:05 & +27:15:08\\
$V_\text{mag}$ & 8.54 & 10.91\\
$M_\star \, (M_\odot)$ & $1.535_{-0.054}^{+0.066}$ & $1.296\pm0.036$ \\
$R_\star \, (R_\odot)$ & $1.732_{-0.045}^{+0.043}$ & $1.428\pm0.030$\\
$T_\text{eff}\,(\text{K})$ &$6789_{-49}^{+50}$ & $6566\pm50$\\
$v\sin i\,(\text{km\,s}^{-1})$ & $65.0_{-5.9}^{+6.0}$ & $40.06\pm0.50$\\
$M_p\,(M_\text{Jup})$ & $1.28\pm0.18$ & $2.18\pm0.25$\\
$R_p\,(R_\text{Jup})$ & $1.533_{-0.047}^{+0.046}$ & $1.466\pm0.040$\\
Period (days) & $2.7347749\pm0.0000039$ & $2.7908327\pm0.0000047$\\
$|\lambda|\,(^\circ)$ & $9.7\pm5.2$ & -\\
\hline
\end{tabular}
\end{table*}
The vast majority of known spin-orbit angles have been measured via the Rossiter-McLaughlin effect. When the planet transits the host star, it successively blocks out part of the rotating stellar surface, and thereby induces a net shift in the centroid of the stellar spectral lines, measured as an apparent in-transit radial velocity variation that is dependent on the transit geometry. The Rossiter-McLaughlin effect was observed for KELT-7b by \citet{2015AJ....150...12B}. The system was reported to be in prograde, spin-orbit aligned geometries. The spectral lines of rapidly rotating host stars are severely broadened by rotation, resulting in blending of individual lines. In such cases it is often possible to directly measure the deviation in the rotational broadening kernel of the spectral lines induced by the transiting planet -- a technique known as Doppler tomography. The technique has already been employed to measure the spin-orbit angles of a number of planetary systems \citep{2010MNRAS.403..151C,2010MNRAS.407..507C,2010A&A...523A..52M,2012ApJ...760..139B,2012A&A...543L...5G,2013ApJ...771...11A,2014ApJ...790...30J,2015ApJ...810L..23J,2015AJ....150..197H}. This technique allows us to directly detect the Doppler shadow of the planet, providing a more accurate measurement of the spin-orbit angle, as well improved characterisation of other transit parameters. We present spin-orbit angles for KELT-7b and HAT-P-56b measured via the Doppler tomographic technique.
\section{Spectroscopic observations and reductions}
\label{sec:obs_and_analyses}
\subsection{TRES transit spectroscopy observations}
\label{sec:tres}
Spectroscopic observations of the transits were obtained with the Tillinghast Reflector Echelle Spectrograph (TRES) on the 1.5\,m telescope at the Fred Lawrence Whipple Observatory, Mount Hopkins, Arizona, USA. The spectrograph has a resolving power of $\lambda / \Delta \lambda \equiv R = 44000$, sampling the spectral range of 3850-9100\,\AA\, over 51 echelle orders. We use the archival spectroscopic transit data of KELT-7b, originally presented by \citet{2015AJ....150...12B}. The observation was performed on 2013-10-19 UT, a total of 29 spectra were observed, each with 900\,s exposure time made up of three exposures in order to optimize the removal of cosmic rays. The set of spectra achieved an average signal-to-noise per resolution element of $\text{S/N} \sim 160$ over the Mg b line region. Wavelength calibration is achieved by a sequence of Th-Ar hollow cathode lamp exposures that bracket each 900s exposure. The transit of HAT-P-56b was observed on 2016-01-03 UT, with a total of 30 spectra obtained. Similar to the KELT-7b observations, each spectrum was combined from an average of three short exposures, with a total exposure time of 540\,s, achieving an average S/N of $\sim35$. Details of the spectral reduction and extraction is similar to that described in \citet{2010ApJ...720.1118B} for FIES observations. The pipeline was modified to work with the TRES spectrograph. All the parameters pertaining to the detector and spectrograph were modified, like the CCD format including overscan regions, the gain and readout noise, an initial guess for the position of the ThAr lines. Furthermore, tweaks were made for the handling of 3D cosmic ray removal, 3D profiling to remove pixel-to-pixel variations. However, these modifications are mostly minor changes and the bulk of the pipeline did not need major modifications to work with TRES.
\subsection{Retrieving the stellar broadening profile}
\label{sec:lsd}
Extracting the line profile from the spectra of rapidly rotating stars is complicated by the lack of unblended lines. We follow the technique set out in \citet{1997MNRAS.291..658D} and \citet{2010MNRAS.407..507C}, and perform a Least-Squares Deconvolution (LSD) to recover the line broadening kernel of each spectrum. This usually involves deconvolving the observed spectrum against a weighted delta-function line list to derive the broadening kernel of the star. Following \citet{2015AJ....150..197H}, we use unbroadened synthetic spectral templates, rather than weighted delta functions, as the template of the deconvolution. Synthetic spectral templates are generated using the spectral synthesis program SPECTRUM\footnote{\texttt{http://www1.appstate.edu/dept/physics/spectrum/spectrum.html}} \citep{1994AJ....107..742G}, with the ATLAS9 model atmospheres \citep{2004astro.ph..5087C}. We assume no rotation, microturbulence, macroturbulence, and instrumental broadening for the spectral template.
Each echelle order of the TRES spectrum is first blaze corrected and continuum normalised. We then derive broadening profiles for consecutive sections of the spectrum, each spanning three echelle orders $(\sim 200\,\text{\AA})$. A 20\% trapezium apodisation is applied to the observed spectrum and template to reduce the artifacts that are induced by the deconvolution. The broadening profiles from each section are average combined to form the final rotational profile for each exposure. A total of 34 echelle orders were used, spanning the spectral range 3900--6250\,\AA. This region was chosen to best avoid the telluric absorption lines. We found that deconvolutions of spectra stitched from three consecutive echelle orders yielded lower noise in the final rotational profile than either deconvolution of individual echelle orders, or deconvolution of the entire stitched spectrum. The $\sim 200\,$\AA\, long spectral regions contain enough information to allow an effective deconvolution, and are small enough that the sections can be weighted to arrive at the highest signal-to-noise averaged profile. The radial velocity shift of the star through the transit sequence, determined from the published orbit, is then subtracted, such that the centroid of each rotation kernel is shifted to $0\,\text{km\,s}^{-1}$.
\subsection{Measuring $v\sin i$ from broadening profile}
\label{sec:vsini}
Spin-orbit angles derived from Doppler tomography and Rossiter-McLaughlin analyses are often degenerate with the rotational velocity of the star. However, $v \sin i$ is difficult to measure from the spectrum due to degenerate effects with other broadening parameters, such as macro turbulence, and the assumed limb darkening parameters. \citet{2012ApJ...757..161T} found the $v\sin i$ estimates from the Stellar Parameter Classification (SPC) pipeline \citep[used in the discovery papers,] []{2012Natur.486..375B} were systematic offset from those of the Spectroscopy Made Easy \citep[SME][]{1996A&AS..118..595V} analyses. Additional systematic offsets in \teff and $\log g$\xspace between the spectral retrieval procedures have been noted \cite[e.g.][]{2012ApJ...757..161T,2013A&A...558A.106M}, but are small enough that the derived line profiles and Doppler tomographic signals are not affected. However, we note that the host star properties, which are part of the global modelling in Section~\ref{sec:modelling}, will affect the final derived stellar properties (e.g. $M_\star$, $R_\star$, and resulting planet properties).
To check if the $v\sin i$ can be accurately recovered from the combination of broadening factors, we generated a series of spectra with $v\sin i=50\,\text{km\,s}^{-1}$, macroturbulence of 0 to $10\,\text{km\,s}^{-1}$, and instrumental broadening of $6\,\text{km\,s}^{-1}$ (as per TRES resolution). These are deconvolved against an unbroadened template to derive broadening profiles for each test synthetic spectrum (Figure~\ref{fig:vmacro}). We fit the broadening kernel with the convolution of a rotation term \citep[modelled analytically from][]{2005oasp.book.....G} and a Gaussian term to account for macroturbulence \citep[expected for F-stars at $6500\,\text{K}$ to be $\sim 6\,\text{km\,s}^{-1}$][]{2014MNRAS.444.3592D} and instrumental broadening. This is different to the SPC approach, which cross correlates a series of spectral templates to the observed spectra, and maximises the cross correlation function peak. Figure~\ref{fig:vmacro} shows the $v\sin i$ can be recovered to within $0.2\,\text{km\,s}^{-1}$. From these tests, we also note that the $v\sin i$ measurement can be overestimated when we use a template that does not account for macroturbulence and instrumental broadening.
\begin{figure*}
\includegraphics[width=12cm]{plots/vmacro.eps}
\caption{\label{fig:vmacro} We test the influence of macroturbulence (vmacro) on the $v\sin i$ derived from LSD broadening kernels. A series of synthetic spectra are generated, with $v\sin i=50\,\text{km\,s}^{-1}$, and macroturbulence of 0 to $10\,\text{km\,s}^{-1}$. Their derived broadening kernels are plotted on the left panel. The middle panel shows that $v\sin i$ can be accurately recovered if we account for both rotational and macroturbulent broadening in the LSD profile fitting. The recovered macroturbulence values are plotted on the left panel.}
\end{figure*}
A $v\sin i$ measurement is made for all available TRES out-of-transit spectra of each object. The median and standard deviation of the $v\sin i$ measurements are $69.2\pm0.2\,\text{km\,s}^{-1}$ and $35.7\pm0.7\,\text{km\,s}^{-1}$ for KELT-7 and HAT-P-56 respectively We also derive macroturbulence velocities of $4.3\pm0.3\,\text{km\,s}^{-1}$ and $7.1\pm1.1\,\text{km\,s}^{-1}$ for each star respectively. The $v\sin i$ we derive for KELT-7 is consistent to that from the discovery paper to within errors ($65_{-5.9}^{+6.0}\,\text{km\,s}^{-1}$), while we derive a slower velocity for HAT-P-56 ($40.1\pm0.5\,\text{km\,s}^{-1}$ from discovery page), a difference likely attributed to the incorporation of macroturbulence in our analysis.
\section{Global modelling of the transit geometry}
\label{sec:modelling}
To derive the spin-orbit angle of the planets, we performed a global modelling of the Doppler tomographic and photometric transit datasets. Since the parameters, such as transit depth, shape, and duration, are shared among the photometric and Doppler tomographic observations, a global fit is required to properly constrain the transit parameters and propagate associated uncertainties.
For KELT-7b, we included all the available photometric follow-up observations detailed in \citet{2015AJ....150...12B}. These included the University of Louisville Moore Observatory 2012-10-04 $g'$ band transit, FLWO KeplerCam 2012-10-23 $z'$, 2012-11-03 $z'$, 2013-11-22 $g'$, and 2013-10-19 $i'$ band transits, Bryne Observatory at Sedgwick 2012-11-14 $g'$ and 2014-01-13 $i'$ band transits, Canela's Robotic Observatory 2012-12-08 $V$, 2013-01-29 $i'$ band transits, and the Whitin Observatory at Wellesley College 2013-01-27 $i'$ band transit.
For HAT-P-56b, we included the K2 long cadence light curve available for the target. The K2 light curve reduction and detrending process are described in \citet{2015AJ....150...85H} and \citet{2015MNRAS.454.4159H}.
The photometric transits were modelled as per \citet{2002ApJ...580L.171M}. Free parameters include the planet-star radius ratio $R_p/R_\star$, normalised orbital distance $a/R_\star$, orbital inclination inc, the transit centre time $T_0$, and period $P$. The quadratic limb darkening parameters for each band are taken from \citet{2011A&A...529A..75C}, interpolated using the tools described in \citet{2013PASP..125...83E} to the atmospheric parameters of each star, and fixed during the global fitting. To account for the long-cadence nature of the HAT-P-56 K2 light curves, we integrated the model over 30-minutes about each time stamp at 10 evenly spaced points. To ensure the per-point photometric uncertainties are accurate, we also inflated the uncertainties such that the reduced $\chi^2$ is at unity when compared against the best fitting model.
To model the Doppler tomographic transit observations, we first created an averaged out-of-transit rotational profile. The `shadow' of the planet is modelled as a Gaussian intrusion to the average rotational profile at each time step. The Gaussian has width of $R_p/R_\star \times v\sin i$, area of $1-f(t)$ (where $f(t)$ is the flux, blocked by the planet, that makes up the transit light curve), centred about $v_p(t)$ (where $v_p(t)$ is the projected rotational velocity for the region of the star occulted by the planet). Given the low S/N of the signal, a Gaussian function is a quick and effective model for the Doppler shadow of the planet \citep[e.g.][]{2016arXiv160200322C}. The parameter $v_p(t)$ is dependent on the spin-orbit angle $|\lambda|$, and the projected rotational velocity of the star $v \sin i$. As instrumental systematics can induce variations to the rotational profile at each time step, at each iteration we also fit for a dilation in the height and width of the rotational profile. An example of the rotational profile fitting, and the Doppler shadow of the planet, are shown in Figure~\ref{fig:vprof_KELT7} (Top). The flux blocked by the planet directly correlates with the area of the Doppler shadow, and we also constructed a transit light curve directly from the Doppler tomographic signal (Figure~\ref{fig:vprof_KELT7} Bottom).
\begin{figure}
\begin{tabular}{c}
\includegraphics[width=8cm]{plots/KELT-7/vprof.eps}\\
\includegraphics[width=8cm]{plots/KELT-7/lc_ccfprofile.pkl.eps}
\end{tabular}
\caption{\label{fig:vprof_KELT7}Top: Broadening kernels derived from a sample of five consecutive exposures during the transit of KELT-7b. The broadening kernel from each exposure is marked by the grey points. The average out-of-transit broadening kernel is marked by the blue lines. Models generated from a set of 100 randomly selected steps in the converged MCMC chain are plotted in red. The residual for each exposure from the average out-of-transit profile is shown below. The Doppler shadow of the planet can be seen in the residuals. Bottom: Measuring the area of the Doppler shadow is equivalent to measuring a transit light curve. The transit light curve of KELT-7b, directly measured from the Doppler tomographic signal, is plotted. }
\end{figure}
The per point uncertainties in the rotational profile residuals were estimated by taking the standard deviation of the baseline regions of the rotational profile. We accounted for correlated noise in the broadening kernel via a Gaussian Processes approach \citep[e.g.][]{2015AJ....150..197H}. Applications of Gaussian process regression to astronomical signals have been extensively covered in the literature and shown to reliably retrieve model parameters from data-sets influenced by stochastic noise sources \citep[e.g.][]{2012MNRAS.419.2683G,2014MNRAS.445.3401G}. The Gaussian process regression was modelled with the \emph{George} module \citep{2014arXiv1403.6015A}. We employed a radial exponential kernel to model the co-variance $\Sigma_{ij}$ between points $i$, $j$, with velocities $v_{i}$ and $v_{j}$:
\begin{equation}
\Sigma_{ij} = \sigma_{i}^2 \delta_{ij} + A \exp \left(- \frac{|v_{i}-v_{j}|}{\tau} \right)\,,
\end{equation}
where $\sigma_{i}$ is the per-point uncertainty for point $i$, and $\delta_{ij}$ is the Kronecker delta function. The Gaussian process hyper-parameters $A$ and $\tau$ specify the amplitude and the scale length of the covariance between the velocity points, respectively.
We explored the parameter space with a Markov chain Monte Carlo (MCMC) analysis, using the \emph{emcee} implementation \citep{2013PASP..125..306F} of an affine-invariant ensemble sampler. The transit parameters are also constrained by the spectroscopic stellar parameters for each star. At each iteration, we calculate an expected $a/R_\star$ using the orbital period, stellar mass and radii expected for the spectroscopic $T_\text{eff}$, $\log g$, [Fe/H] of the spectrum. The stellar mass and radius are interpolated from the spectroscopic parameters via the \citet{2010A&ARv..18...67T} relationships. The expected $a/R_\star$ is then compared to the tested $a/R_\star$, thereby constraining the fit. The spectroscopic $v \sin i$ measurement and uncertainty from Section~\ref{sec:lsd} is applied to the fit as Gaussian prior. Gaussian priors were also imposed on the transit centre $T_0$ and period $P$, since these were derived in the discovery papers from the discovery and follow-up light curves, and are therefore much better constrained than from follow-up light curves alone. Uniform priors were assumed for all other parameters, including the hyper-parameters of the Gaussian process regression.
The derived values and uncertainties are shown in Table~\ref{tab:parameters}. The Doppler tomographic signals, from the rotational profile fit to each TRES exposure, are shown in Figure~\ref{fig:dopplergram}.
\begin{table*}
\caption{\label{tab:parameters}Derived values for MCMC walker parameters}
\begin{tabular}{lrr}
\hline\hline
& \textbf{KELT-7b} & \textbf{HAT-P-56b}\\
\hline
Period (days) $^a$ & $2.734780_{-0.000003}^{+0.000003}$& $2.790833_{-0.000004}^{+0.000004}$ \\
$T_0$ (BJD) $^a$ & $2456355.2293_{-0.0001}^{+0.0001}$ & $2456553.6164_{-0.0003}^{+0.0003}$ \\
$R_p/R_\star$ & $0.0922_{-0.0004}^{+0.0004}$ & $0.099_{-0.002}^{+0.002}$ \\
$a/R_\star$ & $5.50_{-0.06}^{+0.06}$ & $6.7_{-0.4}^{+0.5}$ \\
$\text{inc}\,(^\circ)$ & $83.7_{-0.2}^{+0.2}$ & $82.6_{-0.6}^{+0.7}$\\
$|\lambda|\,(^\circ)$ & $2.7_{-0.6}^{+0.6}$ & $7_{-2}^{+2}$\\
$v\sin i\,(\text{km\,s}^{-1})$ $^b$ & $69.3_{-0.2}^{+0.2}$ & $36.4_{-0.7}^{+0.7}$\\
$\ln(A)$ $^c$ & $-9.17_{-0.04}^{+0.05}$ & $-7.65_{-0.05}^{+0.05}$ \\
$\ln(\tau)$ $^c$ & $3.6_{-0.1}^{+0.1}$ & $2.24_{-0.1}^{+0.1}$\\
$T_\mathrm{eff}$ (K) $^a$ & $6513_{-53}^{+49}$& $6568_{-53}^{+51}$\\
$\log g$ & $4.14_{-0.03}^{+0.03}$ $^a$ & $4.26_{-0.05}^{+0.06}$\\
\hline
\end{tabular}
\begin{flushleft}
$^a$ Gaussian priors according to literature values were imposed.\\
$^b$ Gaussian priors according to $v\sin i$ estimates from Section~\ref{sec:vsini} were imposed. \\
$^c$ Gaussian process hyper-parameters $A$ and $\tau$ describe the amplitude and the scale length of the covariance between velocity points in the broadening profile modelling.
\end{flushleft}
\end{table*}
\begin{figure*}
\begin{tabular}{cc}
\includegraphics[width=7cm]{plots/KELT-7/ccfprofile.pkl.eps} &
\includegraphics[width=7cm]{plots/HATP56/dopplergram.eps} \\
\end{tabular}
\caption{\label{fig:dopplergram}The Doppler tomographic signal for KELT-7b (left) and HAT-P-56b (right) from the TRES observations. The top panels show the signal induced by the planet in the residual between each broadening kernel and the averaged out-of-transit kernel. The middle panels show the model of the best fit geometry. The bottom panels show the residual after the model is subtracted. }
\end{figure*}
To test the dependence of our results on choice of Gaussian process co-variance kernel, we also tested an exponential squared kernel:
\begin{equation}
\Sigma_{ij} = \sigma_{i}^2 \delta_{ij} + A \exp \left(- \frac{(v_{i}-v_{j})^2}{2\tau} \right)\,,
\end{equation}
yielding $|\lambda| = 2.6_{-0.6}^{+0.6}\,^\circ$ for KELT-7b and $|\lambda| = 7_{-2}^{+2}\,^\circ$ for HAT-P-56b.
Using the Matern 3/2 kernel:
\begin{equation}
\Sigma_{ij} = \sigma_{i}^2 \delta_{ij} + A \left( 1+\sqrt{\frac{3(v_{i}-v_{j})^2}{\tau}} \right) \exp \left(-\sqrt{\frac{3 (v_{i}-v_{j})^2}{\tau}}\right)\,,
\end{equation}
yields $|\lambda| = 2.7_{-0.6}^{+0.6}\,^\circ$ for KELT-7b and $|\lambda| = 8_{-2}^{+2}\,^\circ$ for HAT-P-56b. In each case, the Gaussian process hyper-parameters converged to a solution without the need to apply priors. To check for the effect of accounting for stochastic noise via Gaussian process, we also modelled the Doppler tomographic observations without allowing for co-variance between points, and derived similar results, but with smaller uncertainties, of $|\lambda|$ of $2.4_{-0.4}^{+0.4}\,^\circ$ for KELT-7b and $6_{+1}^{+2}\,^\circ$ for HAT-P-56b.
To test the effect of a systematic offset in the assumed stellar parameters on our final result, we deconvolved the KELT-7 spectra against a 7000\,K template, and modelled this new set of broadening profiles in our global MCMC analysis, whilst imposing a stellar parameter Gaussian prior of $T_\mathrm{eff}=7000$\,K, $\log g=4.2$ for the MCMC jump parameters. We arrive at the same set of planet parameters, with no significant change in the best fit values or uncertainties.
\section{Discussion}
\label{sec:discussion}
We find KELT-7b and HAT-P-56b to in spin-orbit aligned geometries, with $|\lambda|$ of $2.8 \pm 0.6\,^\circ$ and $8 \pm 2\,^\circ$ for the two systems respectively.
The transit parameters we derived for KELT-7b and HAT-P-56b are in general agreement with those in the discovery papers. KELT-7b is confirmed to be in a spin-orbit aligned geometry, in agreement with the Rossiter-McLaughlin analysis in \citet{2015AJ....150...12B}. In fact, \citet{2015AJ....150...12B} were able to detect the shadow of the planet in the cross correlation function of the spectra. Because of the grazing nature of the transit of HAT-P-56b, the uncertainties in its transit parameters are larger than the other systems. The uncertainties we derive for HAT-P-56b are are larger than those in the discovery paper, we find $R_p/R_\star = 0.099_{-0.002}^{+0.002}$ and $a/R_\star = 6.7_{-0.4}^{+0.5}$, compared to $R_p/R_\star = 0.1054\pm0.0009$ and $a/R_\star = 6.37\pm0.11$ from the discovery paper. The uncertainty in the planet radius is unchanged, since it is dominated by the uncertainty in the stellar radius, rather than in $R_p/R_\star$. We also note a weak stellar pulsation signal is seen in the Doppler tomographic analysis of KELT-7, manifested as diagonal stripes in Figure~\ref{fig:dopplergram}). Pulsations were not detected in the discovery KELT light curves of the star, but their presence not surprising given the star lies close to the instability strip. Similar pulsations are also seen in the Doppler tomographic observations of WASP-33b \citep{2010MNRAS.407..507C,2015ApJ...810L..23J} and HAT-P-57b \citep{2015AJ....150..197H}.
\subsection{The spin-orbit angle distribution of hot-Jupiters around F-type stars}
\label{sec:stats}
A total of 6 hot-Jupiter systems around stars with $T_\text{eff} > 6250\,\text{K}$ have now been found in spin-orbit alignment $(|\lambda| < 10^\circ)$, while 15 have higher obliquities.
Have these systems undergone tidal synchronisation and realignment? The characteristic timescale for stellar spin synchronisation is given by \citet{2010ApJ...723..285H} and \citet{2012ApJ...757....6H}, who reviewed the tidal theory in the context of hot-Jupiter systems. Adopting Equation 3 in \citet{2012ApJ...757....6H}, the characteristic timescale to modify the spin of the host star, $T_\text{spin}$, is:
\begin{align}
T_\text{spin} =& \frac{3.1\times10^9\,\text{years}}{(1-e^2)^{1/2}} \left(\frac{a}{0.02 \,\text{AU}} \right)^{7.5} \left( \frac{R_\star}{R_\odot}\right)^{-8} \left( \frac{k_0^2}{0.1} \right) \notag\\
&\times \left( \frac{30 \,\text{days}}{P_\text{rot}} \right) \left(\frac{M_\star + M_p}{M_\odot} \right)^{1/2} \left(\frac{M_p}{M_\text{Jup}} \right)^{-2} \notag\\
&\times \left( \frac{\sigma_\star}{7.8\times10^{-8}} \right)^{-1}\,.
\end{align}
We assume a stellar gyration radius of $k_0^2=0.1\,R_\star ^2$ \citep{2000MNRAS.315..543H}, stellar dissipation coefficients $\sigma_\star$ of $10^{-12}$ for KELT-7, and $10^{-8}$ for HAT-P-56 \citep[from Figure 3 of][]{2012ApJ...757....6H}. The expected $T_\text{spin}$ is $\sim 10^{16}$ years for KELT-7 and $10^{12}$ years for HAT-P-56b. That is, tidal dissipation in the stellar envelope is expected to be weak for KELT-7 and HAT-P-56, at least based on standard equilibrium tide models. In fact, tidal interactions are expected to be weak for all the spin-orbit aligned hot-Jupiters around F-dwarfs. Only the transiting brown dwarf KELT-1b \citep{2012ApJ...761..123S}, found around a rapidly rotating F-star in a 1.2\,d period orbit, has a short tidal realignment timescale ($10^8$ years).
KELT-7b and HAT-P-56b were found to be in projected spin-orbit alignment. If we assume alignment in the line-of-sight of the stellar spin axis as well, we can examine the spin-orbit coupling of these systems. The spins of KELT-7 and HAT-P-56 are both super-synchronous with respect to the orbital period of the planet. The maximum rotation period for KELT-7 is $1.08\pm0.03$ days, and $1.8\pm0.2$ days for HAT-P-56, based on the spectroscopic $v \sin i$ measurements. From the K2 light curves, \citet{2015AJ....150...85H} found that HAT-P-56 is a possible $\gamma$ Dor-pulsator with a primary pulsation period of $1.64\pm0.03$ days, and a secondary peak in the periodogram of $1.74\pm0.02$ days that is consistent with the $v\sin i$ derived rotation period. Neither the KELT discovery light curves, nor the archival SuperWASP light curves \citep{2010A&A...520L..10B} yielded a photometric modulation period for KELT-7 in our analysis. For comparison, \citet{2013MNRAS.436.1883W} found that \emph{Kepler} systems with planets of $R_p > 6 \,R_\text{E}$ and $\text{Period}<10$\,days are preferentially found in the stellar-spin -- planet-orbit synchronised states. The four that were systems found to be in super-synchronous states had orbital periods greater than 5 days, the largest of which had a radius of $0.7\,R_\mathrm{J}$. For a consistency check on the assumption that these two systems are also in line-of-sight alignment, we can compare the $v\sin i$ of these stars to that expected from the rotation periods of Kepler stars of similar stellar parameters \citep{2013A&A...557L..10N}. For F-dwarfs like KELT-7 ($6600 < T_\mathrm{eff} < 6800$), 68\% of stars have rotation periods that lie within 1.2 -- 5.9 days. The $v\sin i$-derived rotation period of the KELT-7 is 1.08 days, consistent with the population, and with an aligned geometry. For stars like HAT-P-56 ($6400 < T_\mathrm{eff} < 6600$), 68\% of stars lie within rotation periods of 1.4 -- 8.6 days. The rotation period of HAT-P-56 from $v\sin i$ is 1.8 days, again consistent with the distribution, and with alignment.
Along with the CoRoT-11 system \citep{2010A&A...524A..55G,2012A&A...543L...5G}, KELT-7 and HAT-P-56 are the only spin-orbit aligned super-synchronous systems with planetary orbital periods $<5$\,days. Figure~\ref{fig:porb_prot} shows the orbital period $P_\text{orb}$ of spin-orbit aligned systems against their stellar rotation period $P_\text{rot}$. With the exception of the Kepler candidates from \citet{2013MNRAS.436.1883W}, the $P_\text{rot}$ values are derived from the spectroscopic $v\sin i$ measurements, which should be representative of the stellar spin period if we assume these systems are truly aligned. The only close-in systems in super-synchronous states are found around F-dwarfs, but this may be a selection bias due to the lack of rapidly rotating, cooler stars. For these aligned super-synchronous systems, the angular momentum exchange between the star and the planet is expected to slow down the rotation of the star, and extend the orbital period of the planet. However, the timescale for tides to modify the orbital period is similar to that of the stellar spin synchronisation timescales \citep{2012ApJ...757....6H} for KELT-7 and HAT-P-56, and should not have affected the orbital periods of the planets. We note that a number of other spin-orbit misaligned systems are also found in super-synchronous states (CoRoT-3b, KOI-13b, WASP-7b, WASP-8b, WASP-33b, WASP-38b).
\begin{figure*}
\includegraphics[width=14cm]{plots/Porb_Prot.eps}
\caption{\label{fig:porb_prot}The orbital periods $P_\text{orb}$ and stellar rotation periods $P_\text{rot}$ of close-in hot-Jupiter systems in spin-orbit alignment $(|\lambda| < 10^\circ)$. The solid line marks 1:1 spin-orbit synchronisation. KELT-7 and HAT-P-56 are super-synchronous with respect to the orbital period of their planets. The rotation of Kepler candidates were derived by \citet{2013MNRAS.436.1883W} from their light curves. Otherwise, the spin rotation periods are inferred from the spectroscopic $v\sin i$, assuming $i=90^\circ$, and as such represent the upper limit of the rotational periods. Super-synchronous rotation for short period systems ($P<5$\,days) are only found in systems with host stars of $T_\text{eff}>6250\,\text{K}$. We note that one of the $|\lambda|$ solutions for HAT-P-57b is of low obliquity and super-synchronous, but given the ambiguity that multiple $|\lambda|$ values are allowed \citep{2015AJ....150..197H}, it is left off the plot.}
\end{figure*}
To further check for tidal evolution of the stellar spin, we can also compare the derived rotation periods of F-stars hosting large transiting planets against a similar sample without transiting hot-Jupiters. Figure~\ref{fig:prot_hist} shows the distribution of rotation periods for host stars with $T_\text{eff}>6250\,\text{K}$, binned into aligned $(|\lambda| < 10^\circ)$ and misaligned groups. We also show the rotation period distribution for equivalent stars from the Kepler sample \citep{2013A&A...557L..10N}. To check for the distinction between the populations, we run a two-sample Kolmogorov-Smirnov over 100000 iterations. At each iteration, we draw samples from the rotation period distribution via a bootstrap process, and then draw the rotation period of each star from a Gaussian distribution with the standard deviation as their respective uncertainties. The rotation period of the Kepler F-dwarf sample and that of F-dwarfs hosting spin-orbit aligned systems cannot be distinguished $(p=0.6\pm0.2)$. However, we can tentatively reject the null hypothesis that the rotation periods of the misaligned sample originated from the same population as the Kepler F-dwarfs $(p=0.017\pm0.019)$, although it is marred by small-number statistics. Similarly, a K-sample Anderson-Darling test cannot distinguish between the non-transit planet hosting F-dwarf sample and the spin-orbit aligned sample ($p=0.7\pm0.2$), but can distinguish against the misaligned sample ($p=0.009\pm0.008$). This is expected given these systems are already known to have misaligned $|\lambda|$ angles, and are likely to have line-of-sight $(i)$ misalignments too. The uncertainties are derived by a Monte Carlo exercise, drawing the rotation period of each star from a Gaussian distribution with the standard deviation as their respective uncertainties. This suggests that 1) there is no evidence that the rotation periods of hot stars hosting spin-orbit aligned planets have being modified by tidal interactions, 2) the systems with low projected obliquities are likely to have low true obliquities too. We note that this analysis suffers from an observational selection bias against rapidly rotating stars. Planets found against rapid rotators are difficult to confirm, and therefore lacking in the literature. The same analysis on planets orbiting cool stars $(T_\text{eff}<6250\,\text{K})$ could not distinguish between any of the populations, since these cools stars are spun-down with age, and do not exhibit a sharp rotation period distribution.
\begin{figure}
\includegraphics[width=8cm]{plots/prot_hist.eps}
\caption{\label{fig:prot_hist}The distribution of stellar rotation periods for systems orbiting hot stars $(T_\text{eff}>6250\,\text{K})$. The rotation periods are inferred from their spectroscopic $v\sin i$, assuming $i=90^\circ$. The top panel shows the distribution for well-aligned systems, middle panel shows systems with misalignment in $|\lambda|$. The bottom panel shows the distribution for the rotation periods for stars of the equivalent spectral type measured from Kepler photometry by \citet{2013A&A...557L..10N}}
\end{figure}
Given the lack of evidence for tidal evolution in the rotation periods of most hot host stars, we can examine the set of spin-orbit angles for these systems around hot stars in the context of migration mechanisms. 23\% of the systems around hot stars are found in spin-orbit aligned arrangements. While the fraction of aligned systems is significantly lower than that of the overall distribution, it is still different from the relatively even $\lambda$ distribution expected from dynamical interactions such as eccentric migration via stellar binary Kozai-Lidov cycles \citep[e.g.][]{2012ApJ...754L..36N,2015ApJ...799...27P} or planet-planet scattering \citep[e.g.][]{2011ApJ...742...72N}. Nevertheless, dynamical interactions, compared to in-disk co-planar migration, are likely responsible for a significant fraction of hot-Jupiters around hot stars. It should be noted that inhomogeneity of the star-forming cloud, or binary-induced disk tilting, will also cause primordial spin-orbit misalignment \citep{2010MNRAS.401.1505B,2012Natur.491..418B}. We also note \citet{2012ApJ...758L...6R} suggest that internal gravity waves at the convective core -- radiative envelope boundary can induce arbitrary surface spins for hot stars, independent of star-planet interactions.
\section*{Acknowledgements}
\label{sec:acknowledgements}
GZ thanks Chelsea Huang for HAT-P-56b K2 light curves. The modelling in this paper was performed on the Smithsonian Institution High Performance Cluster (SI/HPC). We also thank Jessica Mink for running the TRES pipeline and maintaining the TRES archive. We acknowledge Andrew H. Szentgyorgyi, Gabor F\H{u}r\'{e}sz, and John Geary, who played major roles in the development of the TRES instrument.
\bibliographystyle{mn2e}
|
\section{Introduction}
Extracting single sentences from corpora with the use of Natural Language Processing (NLP) tools can be useful for a number of purposes including the detection of candidate sentences for automatic exercise generation. Such sentences are also known as \textit{seed sentences} \cite{sumita2005measuring} or \textit{carrier sentences} \cite{smith2010gap} in the Intelligent Computer-Assisted Language Learning (ICALL) literature. Interest for the use of corpora in language learning has arisen already in the 1980s, since the increasing amount of digital text available enables learning through authentic language use \cite{o2007corpus}. However, since sentences in a text form a coherent discourse, it might be the case that for the interpretation of the meaning of certain expressions in a sentence, previously mentioned information, i.e. a \emph{context}, is required \cite{poesio2011computational}. Corpus sentences whose meaning is hard to interpret are less optimal to be used as exercise items \cite{kilgarriff2008gdex}, however, having access to a larger linguistic context is not possible due to copy-right issues sometimes \cite{volodina2012semi}.
In the followings, we explore how we can automatically assess whether a sentence previously belonging to a text can also be used as a stand-alone sentence based on the linguistic information it contains. We consider a sentence \emph{context-dependent} if it is not meaningful in isolation due to: (i) the presence of expressions referring to textual content that is external to the sentence, or (ii) the absence of one or more elements which could only be inferred from the surrounding sentences.
Understanding the main factors giving rise to context dependence can improve the trade-off between discarding (or penalizing) sub-optimal candidates and maximizing the variety of examples and thus, their authenticity. Such a system may not only facilitate teaching professionals' work, but it can also aid the NLP community in a number of ways, e.g. evaluating automatic single-sentence summaries, detecting ill-formed sentences in machine translation output or identifying dictionary examples.
Although context dependence has been taken into consideration to some extent in previous work, we offer an in-depth investigation of this research problem. The theoretical contribution of our work is a set of criteria relevant for assessing context dependence of single sentences based on a qualitative analysis of human evaluators' comments. This is complemented with a practical contribution in the form of a rule-based system implemented using the proposed criteria which can reliably categorize corpus examples based on context dependence both when evaluated using relevant datasets and according to human raters' judgments. The current implementation of the system has been tested on Swedish data, but the criteria can be easily applied to other languages as well.
\section{Background}
\label{sec:background}
\subsection{Corpus Examples Combined with NLP for Language Learning}
\label{bg-gdex}
In a language learning scenario, corpus example sentences can be useful both as exercise items and as vocabulary examples. Previous work on exercise item generation has adopted different strategies for carrier sentence selection. In some cases, sentences are mainly required to contain a lexical item or a linguistic pattern that constitutes the target of the exercise, but context dependence is not explicitly addressed \cite{sumita2005measuring,arregik2011automatic}. Another alternative has been using dictionary examples as carrier sentences, e.g. from WordNet \cite{pino2009semi}. Such sentences are inherently context-independent, however, they pose some limitations on the linguistic aspects to target in the exercises. In \newcite{pilan2014rule} we presented and compared two algorithms for carrier sentence selection for Swedish, using both rule-based and machine learning methods. Context dependence, which had not been specifically targeted in that phase, emerged as a key issue for sub-optimal candidate sentences during an empirical evaluation.
Identifying corpus examples for illustrating lexical items is the main purpose of the GDEX (Good Dictionary Examples) algorithm \cite{husak2010automatic,kilgarriff2008gdex} which has also inspired a Swedish algorithm for sentence selection \cite{volodina2012semi}. GDEX incorporates a number of linguistic criteria (e.g. sentence length, vocabulary frequency) based on which example candidates are ranked. Some of these are related to context dependence (e.g. incompleteness of sentences, presence of personal pronouns), but they are somewhat coarser-grained criteria not focusing on syntactic aspects. A system using GDEX for carrier sentence selection is described in \newcite{smith2010gap} who underline the importance of the well-formedness of a sentence and who determine a sufficient amount of context in terms of sentence length. \newcite{segler2007investigating} focuses on vocabulary example identification for language learners. Teachers' sentence selection criteria has been modeled with logistic regression, the main dimensions examined being syntactic complexity and similarity between the original context of a word and an example sentence.
\subsection{Linguistic Aspects Influencing Context Dependence}
\label{bg-crit}
The relationship between sentences in a text can be expressed either explicitly or implicitly, i.e. with or without specific linguistic elements requiring extra-sentential information \cite{mitkov2014anaphora}. The explicit forms include words and phrases that imply structural discourse relations or are anaphoric \cite{webber2003anaphora}. In a text, the way sentences are interconnected can convey an additional relational meaning besides the one which we can infer from the content of each sentence separately. Examples of such elements include \emph{structural connectives}: conjunctions, subjunctions and “paired” conjunctions \cite{webber2003anaphora}.
Another form of reference to previously mentioned information is \emph{anaphora}. The phenomenon of anaphora consists of a word or phrase (\emph{anaphor}) referring back to a previously mentioned entity (\emph{antecedent}). \newcite{mitkov2014anaphora} outlines a number of different anaphora categories based on their form and location, the most common being pronominal anaphora which has also been the focus of recent research within NLP \cite{poesio2011computational,ng2010supervised,nilsson2010hybrid}. A number of resources available today have noun phrase coreference annotation, such as the dataset from the SemEval-2010 Task \cite{recasens2010semeval} and SUC-CORE for Swedish \cite{nilsson2013suc}.
Besides the anaphora categories described in \newcite{mitkov2014anaphora}, \newcite{webber2003anaphora} argue that adverbial connectives (\textit{discourse connectives}), e.g. \textit{istället} `instead', also behave anaphorically, among others because they function more similarly to anaphoric pronouns than to structural connectives. A valuable resource for developing automatic methods for handling discourse relations is the Penn Discourse Treebank \cite{prasad2008penn} containing annotations for both implicit and explicit discourse connectives. Using this resource \newcite{pitler2009using} present an approach based on syntactic features for distinguishing between discourse and non-discourse usage of explicit discourse connectives (e.g. \textit{once} as a temporal connective corresponding to "as soon as" vs. the adverb meaning "formerly"). Another phenomenon connected to context dependence is \emph{gapping} where the second mention of a linguistic element is omitted from a sentence \cite{poesio2011computational}.
\section{Datasets}
\label{sec:datasets}
Instead of creating a corpus specifically tailored for this task with gold standard labels assigned by human annotators, which can be a rather time- and resource-intensive endeavor, we explored how different types of existing data sources which contained inherently context-(in)dependent sentences could be used for our purposes.
Language learning coursebooks contain not only texts, but also single sentences in the form of exercise items, lists and language examples illustrating a lexical or a grammatical pattern. We collected sentences belonging to these two latter categories from COCTAILL \cite{volodina22you}, a corpus of coursebooks for learners of Swedish as a second language. Most exercises contained gaps which might have misled the automatic linguistic annotation, therefore they have not been included in our dataset.
Dictionaries contain example sentences illustrating the meaning and the usage of an entry. One of the characteristics of such sentences is the absence of referring expressions which would require a larger context to be understood \cite{kilgarriff2008gdex}, therefore they can be considered suitable representatives of context-independent sentences. We collected instances of good dictionary example sentences from two Swedish lexical resources: SALDO \cite{borin2013saldo} and the Swedish FrameNet (SweFN) \cite{heppin2012rocky}. These sentences were manually selected by lexicographers from a variety of corpora.
Sentences explicitly considered dependent on a larger context are less available due to their lack of usefulness in most application scenarios. Two previous evaluations of corpus example selection for Swedish are described in \newcite{volodina2012semi} and \newcite{Pilan-Ildiko2013-9}, we will refer to these as \textsc{Eval1} and \textsc{Eval2} respectively. In the former case, evaluators including both lexicographers and language teachers had to provide a score for the appropriateness of about 1800 corpus examples on a three-point scale. In \textsc{Eval2}, about 200 corpus examples selected with two different approaches were rated by a similar group of experts based on their understandability (readability) for language learners, as well as their appropriateness as exercise items and as good dictionary examples. The data from both evaluations contained human raters' comments explicitly mentioning that certain sentences were context-dependent. We gathered these instances to create a negative sample. Since comments were optional, and context dependence was not the focus of these evaluations, the amount of sentences collected remained rather small, 92 in total. It is worth noting that this data contains spontaneously occurring mentions based on raters' intuition, rather than being labeled following a description of the phenomenon of context dependence as it would be customary in an annotation task.
The sentences from all data sources mentioned above constituted our development set. The amount of sentences per data source is presented in Table \ref{table:data}, where \textsc{CInd} indicates positive, i.e. context-independent samples, and \textsc{CDep} the negative, context-dependent ones. The suffix \textsc{-LL} stands for sentences collected from language learning materials while \textsc{-D} represents dictionary examples.
\begin{table}[ht]
\centering
\small
\begin{tabular}{|c|c|c|c|}
\hline
\bf Source & \bf Code & \bf Nr. sent & \bf Total\\
\hline
COCTAILL & \textsc{CInd-LL} & 1739 & \\
SALDO & \textsc{CInd-D} & 4305 & \bf 8729 \\
SweFN & \textsc{CInd-D} & 2685 & \\
\hline
\textsc{Eval1} & \textsc{CDep} & 22 & \\
\textsc{Eval2} & \textsc{CDep} & 70 & \bf 92 \\
\hline
\end{tabular}
\caption{\label{table:data} Number of sentences per source.}
\end{table}
\setlength{\textfloatsep}{0.7cm}
\section{Methodology}
As the first step in developing the algorithm, we aimed at understanding the presence or absence of which linguistic elements make sentences dependent on a larger context by analyzing our negative sample. Although the number of instances in the context-independent category was considerably higher, certain linguistic characteristics of such sentences could have been connected to aspects not relevant to our task. Negative sentences on the other hand, although modest in number, were explicit examples of the target phenomenon. Information about the cultural context may also be relevant for this task, however, we only concentrated on linguistic factors which can be effectively captured with NLP tools.
We aimed at covering a wide range of potential application scenarios, therefore we developed a method that was independent of: (i) information from surrounding sentences and (ii) the exact intended use for the selected sentences. The first choice was motivated by the fact that, even though most previous related methods (see section \ref{bg-crit}) rely on information from neighboring sentences as well, sometimes a larger context might not be available either due to the nature of the task (e.g. output of single-sentence summarization systems) or copy-right issues. Secondly, for a more generalizable approach, we aimed at assessing sentences based on whether their information content can be treated as an autonomous unit rather than according to whether they provide the appropriate amount and type of context to, for example, be solved as exercise items of a certain type. This way the method could serve as a generic basis to be tailored to specific applications which may pose additional requirements on the sentences.
Being that the amount of negative samples was rather restricted, we opted for the qualitative method of \emph{thematic analysis} \cite{boyatzis1998transforming,braun2006using} aiming at discovering \emph{themes}, i.e. categories, in our negative sample. Once we collected a set of context-dependent sentences, we started coding our data, in other words, manually labeling the instances with \textit{codes}, a word or a phrase shortly describing the type of element that inhibited the interpretation of the sentence in isolation (for some examples see Table \ref{table:TA_results} on the next page). In the subsequent phases, we grouped together codes into themes, i.e. broader categories, according to their thematic similarity in a mixed deductive-inductive fashion. We started out with an initial pool of themes inspired by phenomena proposed in previous literature relevant to context dependence. Some of the codes, however, could not be placed in any of these themes. For part of these we have found a theme candidate in the literature after the pattern emerged during the code grouping phase. In other cases, in absence of an existing category matching some instances of the \textsc{CDep} data, we created our own theme labels.
\begin{table*}[ht]
\centering
\small
\begin{tabular}{|l|c|c|c|c|}
\hline
\bf Theme & \bf ID & \bf Nr & \bf Example code & \bf Example \textsc{CDep} sentence \\
\hline
Incomplete sentence&\textsc{IncompSent}&12&incorrect sent.&\it \textbf{'' p}iper hon och alla skrattar .\\
&&&tokenization&`\textbf{'' s}he whines and everyone laughs.'\\
\hline
Implicit anaphora&\textsc{ImpAnaphora}&11&omitted verb&\it Till jul skulle hon \textbf{[X]}.\\
&&&&`For Christmas she should have \textbf{[X]}.'\\
\hline
Pronominal anaphora&\textsc{PNAnaphora}&23&pronoun as&\it Eller också sitter \textbf{den} i taket.\\
&&&subject&`Or \textbf{it} sits on the roof.'\\
\hline
Adverbial anaphora 1&\textsc{AdvAnaphora1}&12&locative adverb&\it \textbf{Då} ska folk kunna lämna området .\\
(Temporal and locative) & & & & `\textbf{Then} people can leave the area.'\\
\hline
Adverbial anaphora 2&\textsc{AdvAnaphora2}&22& adv. anaphora&\it Vissa gånger sover hon inte \textbf{heller}.\\
(Discourse connectives) & & & &`Sometimes she does not sleep \textbf{either}.'\\
\hline
Structural connectives&\textsc{StructConn}&17&coordinating&\it \textbf{Men} de pratade inte på samma ställe.\\
&&& conjunction&`\textbf{But} they did not talk at the same place.'\\
\hline
Answers to closed &\textsc{CEQAnswer}&11& yes/no answer & \it \textbf{Ja,} men det är ju jul.\\
ended questions & & & & `\textbf{Yes,} but it is of course Christmas.'\\
\hline
Context-depend & \textsc{CDPC} & 8 & unusual noun- & \it Du lämnar \textbf{planen}, \textbf{tolvan}! \\
properties of concepts & & & noun comb. &`You leave the \textbf{field}, \textbf{twelve}!'\\
\hline
\end{tabular}
\caption{\label{table:TA_results} Thematic analysis results.}
\end{table*}
Besides thematic analysis, we carried out also a quantitative analysis based on the distribution of part of speech tags in both our positive and negative sample in order to identify potential differences that could support and complement the information emerged in the themes.
In the following step, we implemented a rule-based algorithm for handling context dependence using the findings from the qualitative and quantitative analyses. Since most emerged aspects could be translated into rather easily detectable linguistic clues, and a sufficiently large dataset annotated with the different context-dependent phenomena was not available for Swedish, we opted for a heuristic-based system. We applied the algorithm and observed its performance on our development data. Our primary focus was on evaluating how precisely are context-dependent elements identified in \textsc{CDep}, but we complemented this also with observing the percentage of false positives for context dependence in our positive sample.
Finally, in order to test candidate selection empirically, a new set of sentences has been retrieved from different corpora. These sentences were then first given to our system for assessment, then the subset of candidates not containing context dependent elements were given to evaluators for an external validation.
\section{Data Analysis Results}
\subsection{Qualitative Results Based on Thematic Analysis}
The list of themes collected during our qualitative analysis is presented in Table \ref{table:TA_results}. For each theme, we provide an identifier (\emph{ID}), the number of occurrence in the \textsc{CDep} dataset (\emph{Nr\footnote{Occasionally sentences included more than one theme.}}) together with an example code and an example sentence\footnote{Tokens relevant to each theme are in bold and [X] indicates the position of an omitted element.}.
The total number of codes emerged from the data was 22, which we mapped to 8 themes. Some of the themes were related to the categories mentioned in previous literature which we described in section \ref{sec:background}. These included pronominal anaphora \cite{mitkov2014anaphora}, adverbial anaphora \cite{webber2003anaphora}, connectives \cite{miltsakaki2004annotating}. Incomplete sentences \cite{didakowski2012automatic} contained incorrectly tokenized sentences, titles and headings. Moreover, we distinguished three themes among different anaphoric expressions: pronominal anaphora, adverbial anaphora (with temporal and locative adverbs) and discourse connectives, i.e. adverbials expressing logical relations. Under the implicit anaphora theme we grouped different forms of gapping.
Two themes that emerged from the data during the thematic analysis were answers to closed ended questions and context-dependent properties of concepts. In the case of the former category, answers were mostly of the yes/no type. As for the latter theme, our data showed that the unexpectedness of the context of a word (especially if this is short, such as a sentence) can also play a role in whether a sentence is interpretable in isolation. Previous literature \cite{barsalou1982context} defines this phenomenon as “context-dependent properties of concepts”. While the “core meanings” of words are activated “independent of contextual relevance”, context-dependent properties are “only activated by relevant contexts in which the word appears” \cite[p. 82]{barsalou1982context}. In (\ref{cdpc}) we provide an example of both context-independent and context-dependent properties of the noun \textit{tak} `roof', from the \textsc{Eval2} data.
\begin{example}
\label{cdpc}
\begin{itemize}
\item[(a)] \textit{Troligen berodde olyckan på all snö som låg på taket.}
\newline
`The accident probably depended on all the snow that covered the roof.'
\item[(b)] \textit{Fler än hundra levande kunde dras fram under taket .}
\newline
`More than a hundred [people] were pulled out from under the roof alive.'
\end{itemize}
\end{example}
Sentence (\ref{cdpc}b) was considered context-dependent by human raters, while (\ref{cdpc}a) was not. Being covered in snow (\ref{cdpc}a) appears a more easily interpretable property of roof without a larger context than having something being pulled out from under it. The context that activates the context-dependent property of roof in (\ref{cdpc}b) is that the roof had collapsed, which, however, is missing from the sentence.
Finally, for 7 sentences in our \textsc{CDep} data, no clear elements causing context dependence could have been clearly identified, these are omitted from Table \ref{table:TA_results}, but they have been preserved in the experiments.
\subsection{Quantitative Comparison of Positive and Negative Samples}
Besides carrying out a thematic analysis, we compared our positive and negative samples also based on quantitative linguistic information in search of additional evidence for the emerged themes and to detect further aspects that could be potentially worth targeting. Overall part of speech (POS) frequency counts showed some major differences between the \textsc{CDep} and \textsc{CIndep} sentences. There was a tendency towards a nominal content in context-independent sentences, where 21.6\% of all POS tags were nouns. However, this value was 9\% lower for context-dependent sentences, which would suggest a preference for a higher density of concepts in context-independent sentences. Pronouns, on the other hand, were more frequent in context-dependent sentences (12.6\% in total) than in context-independent ones (7\% less frequent).
The qualitative analysis revealed that elements responsible for context dependence commonly occurred at the beginning of the sentence. Therefore, we compared the percentage of POS categories for this position in the two groups of sentences. Context-independent sentences showed a strong tendency towards having a noun in sentence-initial position, almost one fourth of the sentences fit into this category. On the other hand, only 3\% of the positive examples started with a conjunction, but 16\% of context-depend items belonged to this group.
\section{An Algorithm for the Assessment of Context Dependence}
\label{sec:alg}
Inspired by the results of the thematic analysis and the quantitative comparison described above, we implemented a heuristics-based system for the automatic detection of context dependence in single sentences.
For retrieving example sentences the system uses the concordancing API of Korp \cite{borin2012korp}, a corpus-query system giving access to a large amount of Swedish corpora. All corpora were annotated for different linguistic aspects including POS tags and dependency relation tags which served as a basis for the implementation. The system scores each sentence based on the amount of phenomena detected that match an implemented context dependence theme.
Users can decide whether to \emph{filter}, i.e. discard sentences that contain any element indicating context dependence. Alternatively, sentences can be \emph{ranked} according to the amount of context-dependent issues detected: sentences without any such elements are ranked highest, followed by instances minimizing these aspects.
All themes have an equal weight of 1 when computing the final ranking score, except for pronominal anaphora in which case, if pronouns have antecedent candidates, the weight is reduced to 0.5. In the followings, we provide a detailed description of the implementation of the themes listed in Table \ref{table:TA_results}.
\newline
\noindent
\textbf{Incomplete sentence.}
To detect incomplete sentences the algorithm scans instances for the presence of an identified dependency root, the absence of which is considered to cause context dependence. Moreover, orthographic clues denoting sentence beginning and end are inspected. Sentence beginnings are checked for the presence of a capital letter optionally preceded by a parenthesis, quotation mark or a dash, frequent in dialogues. Sentences beginning with a digit are also permitted. Sentence end is checked for the presence of major sentence delimiters (e.g. period, exclamation mark).
\newline
\noindent
\textbf{Implicit anaphora.}
Candidate sentences are checked for gapping, in other words, omitted elements. Our system categorizes as gapped (elliptic) a sentence which either lacks a finite verb or a subject. Finite verbs are all verbs that are not infinite, supine or participle. Modal verbs are considered finite in case they form a verb group with another verb. Subjects include also logical subjects, and in the case of a verb in imperative mode, no subject is required.
\newline
\newline
\noindent
\textbf{Explicit pronominal anaphora.}
We considered in this category the third person singular pronouns \textit{den} `it' (common gender) and \textit{det} `it' (neuter gender) as well as demonstrative pronouns (e.g. \textit{denna} `this', \textit{sådan} `such' etc.). We did not include here the animate third person pronouns \textit{han} `he' and \textit{hon} `she' since corpus-based evidence suggests that these are often used in isolated sentences in coursebooks \cite{scherrer2007rivstart} as well as in conversation \cite{mitkov2014anaphora}.
Similarly to the English pronoun \textit{it}, the Swedish equivalent \textit{det} can also be used non-anaphorically in expositions, clefts and expressions describing a local situation, such as time and weather \cite{holmes2003swedish,li2009identification,gundel2005pronouns} as the examples in (\ref{det}) show.
\begin{example}
\label{det}
\begin{itemize}
\item[(a)] \textit{det} with weather-related verbs
\newline
\textit{Det regnar.}
\newline
`It is raining.'
\item[(b)] Cleft
\newline
\textit{Det är sommaren (som) jag älskar.}
\newline
`It is the summer (that) I like.'
\item[(c)] Exposition
\newline
\textit{Det är viktigt att du kommer.}
\newline
`It is important that you come.'
\end{itemize}
\end{example}
Our system treats as non-anaphoric the pronoun \textit{det} if it is expletive (pleonastic) syntactically according to the output of the dependency parser which covers expositions and clefts. To handle cases like (2a), weather-related verbs have been collected from lexical resources. The list currently comprises 14 items. First, verbs related to the class \emph{Weather} in the Simple+ lexicon \cite{kokkinakis2000annotating} have been collected. Then for each of these, the child nodes from the SALDO lexicon have been added. Finally, the list has been complemented with a few manual additions.
For potentially anaphoric pronouns, the system tries to identify antecedent candidates in a similar way to the robust pronoun resolution algorithm proposed in \newcite{mitkov1998robust}. We count proper names and nouns occurring with the same gender and number to the left of the anaphora. This is complemented with an infinitive marker headed by a verb as potential candidate for \textit{det}. Since certain types of information useful for antecedent disambiguation were not available through our annotation pipeline or lexical resources for Swedish (e.g. gender for named entities, animacy), the final step for scoring and choosing candidates is not applied in this initial version of the algorithm. Lastly, pronouns followed by a relative clause introduced by \textit{som} `which' were considered non-anaphoric.
\newline
\noindent
\textbf{Explicit adverbial anaphora.}
Adverbs emerged as an undesirable category during both \textsc{Eval1} and \textsc{Eval2}. However, a deeper analysis of our development data revealed that not all adverbs have equal weight when determining the suitability of a sentence. Some are more anaphoric then others. We collected a list of anaphoric adverbs based on \newcite{teleman1999svenska}. Certain time and place adverbials, also referred to as demonstrative pronominal adverbs \cite{webber2003anaphora} are used anaphorically (e.g. \textit{där} `there', \textit{då} `then'). Sentences containing these adverbs are considered context-independent only when: (i) they are the head of an adverbial of the same type that further specifies them, e.g. \textit{där på landet} `there on the countryside'; (ii) they appear with a determiner, which in Swedish builds up a demonstrative pronoun, e.g. \textit{det där huset} `that house'.
\newline
\noindent
\textbf{Discourse connectives.}
Discourse connectives, i.e. adverbs expressing logical relations, fall usually into the syntactic category of conjunctional adverbials in the dependency parser output. Several conjunctional adverbials appear in the context-dependent sentences from \textsc{Eval1} and \textsc{Eval2}. Our system categorizes a sentence containing a conjuctional adverb context-independent when a sentence contains:
(i) at least 2 coordinate clauses; (ii) coordination or subordination at the same dependency depth or a level higher, that is, a sibling node that is either a conjunction or a subjunction.
\newline
\noindent
\textbf{Structural connectives.}
Sentences with conjunctions as dependency roots are considered context-dependent unless they are paired conjunctions with both elements included (e.g. \textit{antingen ... eller} `either ... or'). Conjunctions in sentence initial position are also treated as an indication of context dependence except when there are at least two clauses or conjuncts in the sentence.
\noindent
\textbf{Answers to closed ended questions.}
To identify sentences that are answers to closed ended questions, the algorithm tries to match POS-tag patterns of sentence-initial interjections (e.g. \textit{ja} `yes', \textit{nej} `no') and adverbs surrounded by minor delimiters (e.g. dash), the initial delimiter being optional in the case of interjections.
\newline
\noindent
\textbf{Context-dependent properties of concepts.}
Apart from the theme implementations described above, we are currently investigating the usefulness of word co-occurrence information for this theme. The corpus query tool Korp for instance offers an API providing mutual information scores. The intuition behind this idea is that the frequency of words appearing together is positively correlated with the unexpectedness of the association between them.
\newline
\section{Performance on the Datasets}
We evaluated our system both on the hand-coded negative example sentences collected from \textsc{Eval1} and \textsc{Eval2} (\textsc{CDep}) and the positive samples comprised of the good dictionary examples (\textsc{CIndep-D}) and the coursebook sentences (\textsc{CIndep-LL}). The performance when predicting different aspects of context dependence is presented in Table \ref{table:precision}.
\begin{table}[ht]
\centering
\small
\begin{tabular}{|c|ccc|}
\hline
\bf Theme & \bf Precision & \bf Recall & \bf F1 \\
\hline
\textsc{IncompSent} &0.75 &0.5 &0.6 \\
\textsc{ImpAnaphora}& 0.33 &0.36 &0.35\\
\textsc{PNAnaphora} & 0.75 &0.78 &0.77\\
\textsc{AdvAnaphora1}& 0.91 &0.83&0.87\\
\textsc{AdvAnaphora2}&0.87 &0.59&0.70\\
\textsc{StructConn} &0.7 &0.82 &0.76\\
\textsc{CEQAnswer} & 1.0 &0.55&0.71\\
\hline
Average & \bf 0.76& 0.63 &0.60\\
\hline
\end{tabular}
\caption{\label{table:precision} Theme prediction performance in \textsc{CDep} sentences.}
\end{table}
We focused on maximizing precision, i.e. on correctly identifying as many themes as possible in the hand-coded \textsc{CDep} sentences, recall values were of lower importance since we aimed at avoiding every context-dependent sentence rather than retrieving them all. Most themes were correctly identified, all themes except one was predicted with a precision of at least 0.7 and above. The only theme that yielded a lower result was that of implicit anaphoras. The error analysis revealed that these cases were mostly connected to an incorrect dependency parse of the sentences, mainly subjects tagged as objects in sentences with an inverted (predicate-subject) word order.
As mentioned previously, we strived for minimizing sub-optimal sentences in terms of context dependence, while trying to avoid being excessively selective to maintain a varied set of examples. To assess performance with respect to this latter aspect, we inspected also the percentage of sentences identified as context-dependent in dictionary examples (\textsc{CInd-D}) and coursebook sentences (\textsc{CInd-LL}). The percentage of predicted themes per dataset is shown in Table \ref{table:cl_results} where \emph{Total} stands for the percentage of sentences with at least one predicted theme.
\begin{table}[ht]
\centering
\small
\begin{tabular}{|c|c|c|}
\hline
\bf Theme & \bf \textsc{CInd-D} & \bf \textsc{CInd-LL} \\
\hline
IncompSent & 2.37 & 3.39\\
ImpAnaphora & 4.61 & 5.80\\
PNAnaphora & 9.39 & 11.0\\
AdvAnaphora1 & 3.59 & 2.93\\
AdvAnaphora2 & 9.95 & 3.74\\
StructConn & 3.70 & 0.92\\
CEQAnswer & 0.37 & 2.59\\
\hline
\bf Total & \bf 33.35 & \bf 26.74\\
\hline
\end{tabular}
\caption{\label{table:cl_results} Percentage of sentences with a predicted theme in the \textsc{CInd} datasets.}
\end{table}
We can observe that even though all sentences are expected to be context-independent, our system labeled as context-dependent about three out of ten good dictionary examples and coursebook sentences. The error analysis revealed that some of these instances did indeed contain context-dependent elements, e.g. the conjunction \textit{men} `but' in sentence-initial position. In \textsc{CInd-LL} in the case of some sentences containing anaphoric pronouns an image provided the missing context in the coursebook, thus not all predicted cases were actual false positives, but rather, they indicated some noise in the data. As for dictionary examples, the presence of such sentences may also suggest that the criterion of context dependence can vary somewhat depending on the type of lexicon or lexicographers' individual decisions.
Some sentences exhibited more than one phenomenon connected to context dependence. Multiple themes were predicted in 30.43\% of the \textsc{CDep} sentences, but only 6.54\% and 7.25 of the \textsc{CInd-D} and \textsc{CInd-LL} sentences respectively.
\section{User-based Evaluation Results}
The algorithm was tested also empirically during an evaluation of automatic candidate sentence selection for the purposes of learning Swedish as a second language. The evaluation data consisted of 338\footnote{We excluded 8 sentences with incomplete evaluator scores during the calculation of the results.} sentences retrieved from a variety of modern Swedish corpora and classified as not containing context dependence themes according to our algorithm (with the exception of 4 control sentences that were context-dependent). These were all unseen sentences not present in the datasets described in section \ref{sec:datasets}. In the evaluation setup, all implemented themes were used as filters, i.e. sentences containing any recognized element connected to context dependence, described in section \ref{sec:alg}, were discarded. Besides context dependence, the evaluated system incorporated also other selection criteria (e.g. readability), but for reasons of relevance and space these aspects and the associated results are not discussed here.
The selected sentences were given for evaluation to 5 language teachers who assessed the suitability of these sentences based on 3 criteria: (i) their degree of being independent of context, (ii) their CEFR\footnote{The Common European Framework of Reference for Languages (CEFR) is a scale describing proficiency levels for second language learning \cite{councilofeurope2001}.} level and (iii) their overall suitability for language learners. Teachers were required to assess this latter aspect without a specific exercise type in mind, but considering a learner reading the sentence instead. Sentences were divided into two subsets, each being rated by at least 2 evaluators. Teachers had to assign a score between 1 to 4 to each sentence according to the scale definition in Table \ref{table:eval_scale}.
\begin{table}[ht]
\centering
\begin{tabular}{|c|l|}
\hline
\multicolumn{2}{|l|}{\it The sentence...}\\
\hline
1 & \it ... doesn't satisfy the criterion. \\
2 & \it ... satisfies the criterion to a smaller extent.\\
3 & \it ... satisfies the criterion to a larger extent.\\
4 & \it ... satisfies the criterion entirely.\\
\hline
\end{tabular}
\caption{\label{table:eval_scale} Evaluation scale. }
\end{table}
The results were promising, the average score over all evaluators and sentences for context independence was 3.05, and for overall suitability 3.23. For context-independence, 61\% of the sentences received score 3 or 4 (completely satisfying the criterion) from at least half of the evaluators, and 80\% of the sentences received an average score higher than 2.5. This latter improves significantly on the percentage of context-dependent sentences that we reported previously in \newcite{Pilan-Ildiko2013-9}, where about 36\% of all selected sentences were explicitly considered context-dependent by evaluators.
Furthermore, we computed the Spearman correlation coefficient for teachers' scores of overall suitability and context dependence to gain insight into how strongly associated these two aspects were according to our evaluation data. The correlation over all sentences was $\rho$=0.53, which indicates that not being context-dependent is positively associated with overall suitability. Therefore, context dependence is worth targeting when selecting carrier sentences.
\section{Conclusion and Future Work}
We described a number of criteria that influence context dependence in corpus examples when presented in isolation. Based on the thematic analysis of a set of context-dependent sentences, we implemented a rule-based algorithm for the automatic assessment of this aspect which has been evaluated not only on our datasets but also with the help of language teachers with very positive results.
About 76\% of themes were correctly identified in context-dependent sentences, while the amount of false positives in the context-independent data was maintained rather low.
Approximately 80\% of candidate sentences selected with a system incorporating the presented algorithm were deemed context-independent in our user-based evaluation. The results also showed a positive correlation between sentences being context-independent and overall suitable for language learners.
In the future, we are planning to explore the extension of the algorithm to other languages as well as to experiment with machine learning approaches for this task using, among others, the resources mentioned in this paper.
|
\section{Introduction}
\label{sec:intro}
In the introduction of his 1916 paper, Eugene Cook Bingham states: ``In the study of plastic flow it has already been shown that most homogeneous solids will flow somewhat after the manner of liquids, if subjected to sufficient pressure. Copper, steel, lead, ice, menthol, glass, and asphalt fall in this class insofar as they may be regarded as homogeneous solids. But ordinarily plastic substances are not homogeneous solids but suspensions of finely divided solids in fluids, such as paint in oil, lime in water, and especially clay in water. Numerous papers have been devoted to the explanation of this latter type of plasticity.'' \cite{Bingham:1916}. A hundred years later, in spite of an overwhelming number of works on the subject, the details of plastic flow in such amorphous materials as the clay suspensions studied by Bingham, nowadays referred to as ``soft glassy materials,'' still resist complete understanding \cite{Lin:2016}.
Creep, the slow deformation that results from the application of a constant stress on a solidlike material, is a common manifestation of plasticity. Six years before Bingham's paper, at the age of 23, Edward Neville da Costa Andrade showed that when stretched under a constant load, some metallic wires deform such that the strain grows with time as a power law of exponent 1/3 \cite{Andrade:1910}. This behaviour, known as ``power-law creep,'' has been interpreted in terms of peculiar dislocation motion within the crystalline matrix due to thermally activated processes \cite{Cottrell:1952,Mott:1953,Cottrell:1997} or due to dislocation jamming \cite{Miguel:2002}. Although also observed in a variety of disordered materials including emulsions, microgels and colloidal gels and glasses, the origin of power-law creep remains mostly elusive in such soft amorphous systems \cite{Sentjabrskaja:2015}.
When the applied stress is increased above some critical value, creep gives way to fracture or to viscous flow depending on the material microstructure. In his 1916 paper, Bingham focused on ``English china clay.'' Using a capillary viscometer, he showed that this colloidal dispersion was well described by the following affine constitutive relation between the stress $\sigma$ and the shear rate $\dot\gamma$: $\sigma=\sigma_\text{y}+\eta\dot\gamma$, where $\sigma_\text{y}$ is the yield stress of the material and $\eta$ is the viscosity of the material when made to flow far above $\sigma_\text{y}$. Ten years later, this empirical relation was generalized by Winslow Herschel and Ronald Bulkley to $\sigma=\sigma_\text{y}+k\dot\gamma^n$, where $k$ and $n$ are known as the consistency index and flow index respectively \cite{Herschel:1926}. This relation accounts for shear-thinning above the yield stress and the exponent $n$ typically lies between 0.2 and 0.8 \cite{Roberts:2001,Bonn:2015}. The Herschel-Bulkley (HB) constitutive relation was found to account correctly for the behaviour of a wide range of disordered materials such as foams, emulsions and microgels, which were recently coined ``simple yield stress fluids'' by contrast with other soft glassy materials that display more complex rheological behaviour including thixotropy or shear localization \cite{Moller:2009b,Balmforth:2014,Bonn:2015}.
An additional well-known feature of glassy systems is their history-dependence. For instance, when a silica glass is thermally quenched below the glass temperature, it stores stresses that remain trapped within the material. These ``residual'' or ``internal'' stresses are routinely used to strengthen so-called prestressed glasses in industrial applications. Similarly, soft glassy materials store residual stresses when mechanically quenched from a presheared fluidised state to rest at zero strain or stress \cite{Ballauff:2013}. The configuration reached by the quenched microstructure is generally metastable and evolves more and more slowly as the system explores more stable configurations. Such a physical aging may subsequently compete with rejuvenation induced by shear \cite{Cloitre:2000,Viasnoff:2002}.
The main goal of the present contribution is to explore creep in a carbopol dispersion, a system made of acidic cross-linked polymer particles that swell and jam upon neutralization \cite{Ketz:1988,Piau:2007,Putz:2009}. Such carbopol ``microgels'' generally nicely follow the HB behaviour and are considered as non-aging, non-thixotropic yield stress materials \cite{Curran:2002,Piau:2007,Lee:2011}. However, a number of recent results have reported unexpected results for such a simple yield stress fluid, including rheological hysteresis \cite{Putz:2009,Divoux:2013}, transient shear localization \cite{Divoux:2012}, influence of rest time on stress overshoots upon shear start-up \cite{Divoux:2011b} and influence of confinement on flow properties \cite{Geraud:2013a}. Creep below the yield stress has only been scarcely explored in carbopol and previous works focused on the debate about the existence of a true yield stress \textit{vs} a Newtonian plateau at low shear rate \cite{Moller:2009a,Dimitriou:2013}.
Here, we focus on a carbopol ETD~2050 dispersion whose preparation, microstructure and basic rheological features are discussed in Section~\ref{sec:material}. Section~\ref{sec:results} presents the results of creep experiments performed after preshear followed by a rest time under a zero shear rate in cone-and-plate geometry. We first show that our microgels display robust power-law creep with an exponent $\alpha=0.39\pm 0.04$ for stresses ranging from the yield stress $\sigma_\text{y}$ down to a typical value $\sigma_\text{c}\simeq 0.2 \sigma_\text{y}$. For applied shear stresses lower than $\sigma_\text{c}$, the microgels experience a more complex creep behaviour, characterized by an initial decrease of the strain. We attribute this anomalous creep to residual stresses of the order of $\sigma_\text{c}$ that slowly relax after quenching the system from a presheared fluidised state and that compete with the applied stress. This behaviour appears to be independent of the carbopol concentration (Appendix~B) but is no longer observed if one prevents residual stresses by imposing a zero shear stress during the rest time (Appendix~C). In Section~\ref{sec:discussion}, these results are discussed in light of the current literature and compared to previous works on other systems in terms of glassy dynamics and aging phenomena. Beside findings specific to carbopol microgels, we emphasize the relevance of creep studies for a better general understanding of soft glassy materials and for deeper investigations of the subtle interplay between microstructure and deformation below the yield stress.
\section{Materials and methods}
\label{sec:material}
\subsection{Sample preparation, microstructure and rheology}
\label{sec:preparation}
Our samples are made of carbopol ETD~2050 following the preparation protocol described in \cite{Geraud:2013a,Geraud:2013b}. The carbopol powder, a mixture of homopolymers and copolymers of acrylic acid cross-linked with a poly\-alkyl poly\-ether, is dissolved in deionized water at 50$^\circ$C under magnetic stirring for 30~min. The carbopol mass fraction is $C=1$~\%~wt throughout this paper, except in Appendix~B where mass fractions of 0.6~\%~wt and 2~\%~wt are briefly explored. After equilibration at room temperature for 30~min, the acid solution is neutralized by addition of sodium hydroxyde (NaOH at 0.1~M) until the pH reaches a value of $7.0\pm 0.5$. This leads to a dilution of the carbopol solution by about 10~\% so that the final carbopol mass fraction is close to 0.9~\%~wt. Neutralization induces the swelling and jamming of polymer particles \cite{Roberts:2001,Baudonnet:2004}. The sample is subsequently homogenized for 24~h by stirring with a mixer (RW20, Ika, tip: R1303) at 2000~rpm. Finally, the sample is centrifuged for 10~min at 3200~rpm in order to remove air bubbles.
The microstructure of carbopol ETD~2050 dispersions has been characterized by dynamic light scattering and confocal imaging \cite{Lee:2011,Gutowski:2012,Geraud:2013b}. It was shown that these systems can be considered as a dense, amorphous assembly of soft jammed particles made of swollen polymers, namely as a ``microgel'' following a widespread --yet rather loose-- definition \cite{Baker:1949,Saunders:1999,Roberts:2001}. The typical size of the particles is about $2~\mu$m as estimated from confocal microscopy on the same system as that investigated here \cite{Geraud:2013b}. Such a length scale is consistent with light scattering measurements performed on samples made of the same carbopol and prepared in similar conditions \cite{Lee:2011}. More specifically, \cite{Lee:2011} report the existence of two different length scales, which points to a heterogeneous structure where regions of higher- and lower- than-average cross-link density coexist. This suggests that carbopol ETD~2050 consists of randomly cross-linked polymers rather than polymer particles with heavily cross-linked cores and dangling free ends.
As a result of the percolation of the network of swollen polymer particles, carbopol dispersions typically show a yield stress for weight concentrations as low as 0.06~\% \cite{Roberts:2001,Oppong:2006,Oppong:2011}. For concentrations larger than typically 0.2~\%~wt, rheological measurements suggest that the system can be described as a ``space-filling paste of swollen microgel particles'' \cite{Gutowski:2012}. As shown in Appendix~A (see Fig.~\ref{fig:flowcurve}), the flow curve of our carbopol ETD~2050 microgel for $C=1$~\%~wt nicely follows HB rheology, $\sigma=\sigma_\text{y}+k\dot{\gamma}^n$, with parameters $\sigma_\text{y} =10.2$~Pa, $n=0.60$ and $k=3.3$~Pa.s$^n$ that are fully consistent with those reported on the same system with similar preparation protocols \cite{Divoux:2010,Divoux:2012,Gutowski:2012,Geraud:2013a}. The large extent of the linear regime up to about one strain unit, the low-frequency values of the viscoelastic moduli $G'\simeq 37$~Pa $\gg G''\simeq 3$~Pa and their weak increase with frequency (see Fig.~\ref{fig:oscill} in Appendix~A) are also in line with previous measurements available in the carbopol literature \cite{Ketz:1988,Benmouffok:2010,Divoux:2011b,Gutowski:2012,Geraud:2013b,Jofore:2015}.
These various rheological features are characteristic of ``soft glassy materials'' \cite{Liu:1996b,Fielding:2000,Derec:2003}. Although one should keep in mind that there is no universality among these systems due to the huge variety of microstructures, this ``simple'' yield stress rheology makes carbopol microgels good candidates for generic studies of creep phenomena.
\subsection{Creep test and flow cessation protocols}
\label{sec:creepprotocol}
All the experiments reported here were performed with a stress-imposed rheometer (Anton Paar MCR 301) equipped with a cone-and-plate geometry of diameter 40~mm and angle $2^\circ$. To minimize wall slip, we used a sand-blasted cone (CP40-2/S) with a roughness of 5.5~$\mu$m. The bottom plate includes a Peltier element that keeps the temperature fixed to 25$^\circ$C. A specific cover is used to limit evaporation and allows us to work on the same sample loading for at most $\sim 12$~h.
In Sect.~\ref{sec:creep} below, we perform creep experiments by monitoring the strain response $\gamma(t)$ to a constant shear stress $\sigma$. As in most soft glassy materials, a good control of the initial state of the microgel sample is crucial to warrant reproducibility and to ensure proper analysis of the experimental data. Therefore, prior to any experiment (including the previous viscoelastic measurements), the freshly loaded sample is first thoroughly fluidised thanks to a strong preshear under a controlled shear rate $\dot{\gamma}_\text{p}$ for a duration $t_\text{p}$. The sample is then allowed to rest under a zero shear rate during a waiting time $t_w$. Such a protocol allows us to start subsequent tests in a reproducible initial state. We note that it slightly differs from usual creep protocols performed from a rest period under a zero shear stress \cite{Cloitre:2000,Purnomo:2007}. Experiments performed by imposing a zero shear stress rather than a zero shear rate during the rest time are briefly presented in Appendix~C. We check for reproducibility (and for lack of evaporation or other long-term degradation of the sample) by systematically measuring the viscoelastic moduli during the rest time under small-amplitude oscillatory shear (strain amplitude 1~\% and frequency 1~Hz). The variation in $G'$ from one experiment to the other over the same loading is found to be less than 1~\%.
Unless otherwise specified, we use $\dot{\gamma}_\text{p}=\SI{100}{\per\second}$, $t_\text{p}=\SI{60}{\second}$ and $t_\text{w}=\SI{300}{\second}$ as default parameters for the preshear and rest protocol. The creep test is started immediately after the rest time and the strain $\gamma(t)$ is monitored for at least 300~s and up to one hour. Thus, the origin of time $t$ is taken here after an ``aging'' duration $t_w$ following preshear. The influence of $\dot{\gamma}_\text{p}$ and $t_\text{p}$ will be addressed in Sect.~\ref{sec:interplay}. Aging issues, i.e. a possible dependence with $t_w$, will be discussed in Sect.~\ref{sec:aging}.
We checked that a constant torque is reached within about 10~ms and that it is kept perfectly constant by the rheometer throughout the creep experiment. However, inertio-elastic oscillations occur upon application of stress and the material is truly submitted to a constant stress only once these oscillations are damped, typically after a couple of seconds (see Fig.~\ref{fig:creep_fit} and discussion below). After a creep test under a given shear stress $\sigma$, the preshear and rest protocol is repeated over again and another creep test is performed for a different value of $\sigma$. Successive values of $\sigma$ are chosen in a random order to avoid any artefact arising from a systematic drift in the sample or in the setup. This procedure is iterated until the elastic modulus at rest differs by more than 10~\% from its initial value, indicating evaporation and/or sample alteration, thus requiring to load a new sample. To improve the statistics, creep experiments are reproduced three to six times for the same shear stress on different loadings of the same microgel preparation batch.
Finally, in Sect.~\ref{sec:interplay}, we shall investigate stress relaxation after preshear through flow cessation experiments. The reason for these additional experiments is that our rheometer does not provide the full temporal stress response under small-amplitude oscillatory strain but only the amplitude of the stress oscillations, so that stress relaxation cannot be monitored simultaneously to the viscoelastic moduli. The preshear protocol for flow cessation experiments is the same as for creep experiments, namely preshear under a controlled shear rate $\dot{\gamma}_\text{p}$ for a duration $t_\text{p}$. After preshear, a zero shear rate is applied and the stress $\sigma(t)$ is monitored over time $t$ whose origin is taken immediately after preshear.
\section{Results}
\label{sec:results}
\subsection{Creep tests}
\label{sec:creep}
\subsubsection{Analysis of a typical creep test}
\label{sec:singlecreep}
\begin{figure*}
\includegraphics[width=0.9\textwidth]{exemple_fit_tot.pdf}
\caption{(a)~Strain response $\gamma(t)$ for a creep experiment at $\sigma=\SI{5}{\pascal}$ (preshear at $\dot{\gamma}_\text{p}=\SI{100}{\per\second}$ for $t_\text{p}=\SI{60}{\second}$, rest time $t_\text{w}=\SI{300}{\second}$). The red solid line is the best Andrade fit, $\gamma(t)=\gamma_0 + (t/\tau)^\alpha$, with $\gamma_0 = 11.2~\%$, $\alpha=0.36$ and $\tau=\SI{1.3e5}{\second}$, following the procedure described in the text. The inset shows $\gamma - \gamma_0$ with $\gamma_0 = 11.2~\%$ as a function of time in logarithmic scales. (b)~Shear rate response $\dot{\gamma}(t)$ corresponding to the experiment in (a). The red solid line shows the best power-law fit with exponent $\alpha-1=-0.64$.}
\label{fig:creep_fit}
\end{figure*}
Figure~\ref{fig:creep_fit}(a) displays a typical strain response $\gamma(t)$ recorded during a creep test under a shear stress $\sigma$ well below the yield stress $\sigma_\text{y}\simeq 10$~Pa but larger than 2~Pa. The short-time response involves damped oscillations [see inset of Fig.~\ref{fig:creep_fit}(a) for $t<1$~s] that are characteristic of the coupling between the instrument inertia and the material elasticity in stress-controlled rheometry \cite{Ferry:1980}. Although analyzing such oscillations may provide valuable information \cite{Baravian:1998,Baravian:2007,Benmouffok:2010,Ewoldt:2007}, we shall disregard them in the present work as we focus on the long-term creep behaviour of our microgels. In order to quantify the response to creep, we first consider the shear rate $\dot\gamma(t)$ obtained by differentiating $\gamma(t)$ with respect to time and plotted in Fig.~\ref{fig:creep_fit}(b). As seen by using logarithmic scales, $\dot\gamma(t)$ decays as a power law of time, $\dot\gamma(t)\sim t^{\alpha-1}$ with $\alpha-1\simeq -0.64$ over almost three decades in time.
Based on the value $\alpha$ of the exponent extracted from $\dot\gamma(t)$, we then go back to the strain data and fit $\gamma(t)$ by a power law:
\begin{equation}
\gamma(t)=\gamma_0 + (t/\tau)^\alpha\,,
\label{eq:strain}
\end{equation}
with only two adjustable parameters, namely the initial strain $\gamma_0$ and the prefactor $\tau$. Such a fit is restricted to times larger than a few seconds in order to exclude short-time oscillations. The result is shown as a red line in Fig.~\ref{fig:creep_fit}(a) and the inset confirms that $\gamma(t)-\gamma_0$ grows as a power law of time. Such a power law is strikingly similar to the power-law creep with exponent 1/3 first reported by Andrade in metallic wires stretched under a constant load \cite{Andrade:1910}. This behaviour, referred to as ``Andrade-like'' scaling, will be discussed in more details in Sect.~\ref{sec:andrade} below.
The initial strain $\gamma_0$ corresponds to the instantaneous, elastic deformation that would be observed in the absence of inertio-elastic oscillations i.e. in an ``ideal'' creep experiment. The time $\tau$ in Eq.~(\ref{eq:strain}) corresponds to the point where the strain has increased by 100~\% starting from $\gamma_0$. However, $\tau$ cannot be interpreted as some ``characteristic time'' of the material since it appears in a power law. As a matter of fact, $\tau$ is generally much larger than the creep duration itself ($\tau=\SI{1.3e5}{\second}$ in the present case) and has no specific physical significance. In the following, we shall thus discuss only the exponent $\alpha$ and the strain $\gamma_0$.
\subsubsection{Creep dependence on the applied stress}
\label{sec:multiplecreep}
\begin{figure}
\includegraphics[width=0.45\textwidth]{regimes_creep.pdf}
\caption{Shear rate responses $\dot{\gamma}(t)$ for creep experiments at $\sigma=\SI{15}{}$, \SI{14}{}, \SI{12}{}, \SI{7.4}{}, \SI{5.0}{}, \SI{3.4}{} and \SI{1.0}{\pascal} from top to bottom ($\dot{\gamma}_\text{p}=\SI{100}{\per\second}$, $t_\text{p}=\SI{60}{\second}$, $t_\text{w}=\SI{300}{\second}$). The yellow solid lines show the best power-law fits of $\dot{\gamma}(t)$ over the corresponding fitting intervals (see text). The negative parts of the short-time oscillations do not show due to the use of logarithmic scales. Blue dashes separate creep experiments at $\sigma>\sigma_\text{y}$ from those at $\sigma<\sigma_\text{y}$.}
\label{fig:creep_regimes}
\end{figure}
Shear rate responses to several creep tests for imposed stresses ranging from 1~Pa to 15~Pa are gathered in Fig.~\ref{fig:creep_regimes}. While experiments performed above the yield stress $\sigma_\text{y}\simeq 10$~Pa show a transition from a creep regime to a flowing regime with a non-zero steady-state shear rate (see dark brown curve for $\sigma=15$~Pa) or at least an upturn in the shear rate (see curves for $\sigma=12$ and 14~Pa), creep tests for $\sigma<\sigma_\text{y}$ all decrease to vanishingly small shear rates. This behaviour is typical of the yielding transition. Below the yield stress, the system slowly creeps and the shear rate tends towards zero. Above the yield stress, the material eventually flows although such a solid-to-fluid transition takes increasingly long times as the applied stress approaches the yield stress.
The creep behaviour of microgels made of carbopol ETD~2050 --yet with a different preparation protocol-- has been extensively studied {\it above} the yield stress by Divoux {\it et al.} \cite{Divoux:2011a,Divoux:2012}. It was shown that the fluidisation time diverges as a power law of the viscous stress and that the transition towards a flowing state involves a solid--fluid coexistence through the presence of long-lived transient shear bands. Although the goal of the present study is to focus on creep {\it below} the yield stress, we emphasize that an Andrade-like scaling was already reported in \cite{Divoux:2011a} with an exponent of about 1/3 and that the creep regime was shown to involve a locally homogeneous deformation field at scales larger than a few tens of micrometers thanks to ultrasonic velocimetry.
\begin{figure*}
\includegraphics[width=0.9\textwidth]{andrade_tot.pdf}
\caption{Andrade fit parameters, (a)~exponent $\alpha$ and (b)~initial elastic deformation $\gamma_0$, as a function of the applied stress $\sigma$. The data are averaged over three to six independent creep experiments on the same carbopol batch but for different loadings. Error bars correspond to the standard deviation of the Andrade parameter estimations. Black dots correspond to estimations based on power-law fits of the shear rate response (yielding $\alpha$) then on fits of the strain response with two free parameters (yielding $\gamma_0$ and $\tau$) while red squares correspond to full fits of the strain response given by Eq.~(\ref{eq:strain}) with three free parameters. The gray line in (a) shows the mean value of $\langle\alpha\rangle=0.39$ (average over $\sigma=1.8$--8.6~Pa). The gray line in (b) is a linear fit $\sigma=G'_0\gamma_0+\sigma_0$ with $G'_0=\SI{32}{\pascal}$ and $\sigma_0=\SI{0.9}{\pascal}$ (fit over $\sigma=2$--10~Pa). The gray area shows the range of yield stresses $\sigma_\text{y}$ found by systematically fitting the flow curve (see Fig.~\ref{fig:flowcurve}) for the various loadings prior to creep experiments. The vertical dashed line shows the typical stress of 2~Pa below which residual stresses significantly affect the creep response (see Sect.~\ref{sec:interplay}).}
\label{fig:andrade}
\end{figure*}
The parameters $\alpha$ and $\gamma_0$, extracted from power-law fits of $\dot{\gamma}(t)$ and $\gamma(t)$ as described above, are plotted as a function of the imposed stress $\sigma$ with black symbols in Fig.~\ref{fig:andrade}. The time intervals over which the fits are performed depend on $\sigma$ and are indicated by yellow lines in Fig.~\ref{fig:creep_regimes}. The exponent $\alpha$ is seen to remain roughly constant for $0.2 \sigma_\text{y}\lesssim\sigma\lesssim \sigma_\text{y}$ and its mean is $\langle\alpha\rangle=0.39$. For stresses above the yield stress, the exponent seems to increase significantly. This could be due to the fact that for $\sigma\gtrsim \sigma_\text{y}$, the amplitude of the initial inertio-elastic oscillations may reach into the nonlinear regime and significantly affect the material, e.g. through microstructural rearrangements. Yet, the fitting interval gets smaller as the creep regime becomes shorter under larger applied stresses, which makes the error bars on $\alpha$ increase dramatically. The value of the Andrade exponent $\alpha$ for $\sigma\lesssim \sigma_\text{y}$ will be further discussed in Sect.~\ref{sec:andrade} in light of the current literature.
Not unexpectedly, the initial elastic deformation $\gamma_0$ increases linearly with $\sigma$ over the whole range of applied shear stresses and the proportionality factor $G'_0=32$~Pa is fully consistent with the elastic modulus $G_0=34$~Pa measured independently through small oscillatory shear. However, it is important to note that the linear fit shown in Fig.~\ref{fig:andrade}(b) involves a non-negligible value for $\sigma=0$. In other words, one has $\sigma=G'_0\gamma_0+\sigma_0$ where $\sigma_0=0.9$~Pa is significant and points to a specific behaviour at low applied stresses. Correspondingly, the exponent $\alpha$ sharply deviates from its mean value when the stress is decreased below a characteristic value $\sigma_\text{c}\simeq 2$~Pa.
\begin{figure}
\includegraphics[width=0.45\textwidth]{creep_bas_stress_sigma.pdf}
\caption{Strain responses $\gamma(t)$ for creep experiments at low stresses, $\sigma=\SI{0.6}{},\SI{0.8}{},\SI{1.0}{},\SI{1.2}{},\SI{1.4}{},\SI{1.6}{\pascal}$, from bottom to top ($\dot{\gamma}_\text{p}=\SI{100}{\per\second}$, $t_\text{p}=\SI{60}{\second}$, $t_\text{w}=\SI{300}{\second}$). The yellow lines are the power-law fits obtained with the procedure described in Sect.~\ref{sec:singlecreep} and leading to exponents $\alpha=0.29$, 0.23, 0.38 and 0.40 from bottom to top. The gray lines correspond to the best fits of $\gamma(t)$ by Eq.~\ref{eq:strain} with three adjustable parameters and yield $\alpha=0.84$, 0.54, 0.47 and 0.44 from bottom to top.}
\label{fig:low_stress1}
\end{figure}
To test the robustness of the above estimates, Fig.~\ref{fig:andrade} also displays the parameters $\alpha$ and $\gamma_0$ obtained by fitting directly the strain response to Eq.~(\ref{eq:strain}) with three free parameters. Over the range $\sigma=2$--10~Pa, the results are almost undistinguishable from those obtained with the previous fitting procedure. For $\sigma\lesssim \sigma_\text{c}$ however, the two estimates of $\alpha$ strongly differ. This may be due to the fact that for low stresses, differentiating $\gamma(t)$ leads to noisy shear rate responses involving negative values of $\dot\gamma(t)$ which are not taken into account when fitting the logarithm of the data to a straight line (see Fig.~\ref{fig:creep_regimes} for $\sigma=1$~Pa). This also suggests that Eq.~(\ref{eq:strain}) may fail in describing the strain response for $\sigma\lesssim\sigma_\text{c}$.
In order to get better insight into the creep behaviour at low applied stresses, Fig.~\ref{fig:low_stress1} displays strain responses recorded for $\sigma<2$~Pa in semilogarithmic scales. For $1\le\sigma<2$~Pa, $\gamma(t)$ can still be well fitted to power laws, leading to the values of $\alpha$ reported in Fig.~\ref{fig:andrade}(a). It could be tempting to interpret the low values of $\alpha$ obtained by first fitting $\dot\gamma(t)$ (yellow solid lines in Fig.~\ref{fig:low_stress1}) as logarithmic creep. However, fitting $\gamma(t)$ with three free parameters over the same time range leads to much larger values of $\alpha$ (gray solid lines in Fig.~\ref{fig:low_stress1}). This is indicative of a high sensitivity to the fitting procedure and questions the validity of the power-law model. Even more strikingly, at even lower stresses ($\sigma<1$~Pa), strain responses no longer show a monotonic increase once the initial oscillations have died out and $\gamma(t)$ even goes to negative values for the smallest applied stress of 0.6~Pa. As shall be shown in the next section, this anomalous behaviour is the signature of a subtle interplay between creep and residual stresses following preshear.
\subsection{Interplay between creep and residual stresses}
\label{sec:interplay}
\subsubsection{Evidence for residual stresses}
\label{sec:residual}
\begin{figure*}
\includegraphics[width=0.9\textwidth]{stress_res_tot.pdf}
\caption{(a)~Stress relaxation $\sigma(t)$ after preshear at $\dot{\gamma}_\text{p}=$10, 100 and \SI{1000}{\per\second} (from top to bottom) for $t_\text{p}=\SI{60}{\second}$. A zero shear rate is imposed immediately after preshear. (b)~Residual stress $\sigma_\text{r}$, defined as the stress after a relaxation over $t_\text{w}=\SI{600}{\second}$, as a function of the preshear duration $t_\text{p}$ for a fixed $\dot{\gamma}_\text{p}=\SI{100}{\per\second}$. The red line is $\langle \sigma_\text{r} \rangle = \SI{0.63}{\pascal}$. (c)~$\sigma_\text{r}$ as a function of the preshear rate $\dot{\gamma}_\text{p}$ for a fixed $t_\text{p}=\SI{60}{\second}$. The red line is the best power-law fit $\sigma_\text{r} = 1.6 \dot{\gamma}_\text{p}^{-0.2}$.}
\label{fig:residual}
\end{figure*}
In order to interpret the unusual creep behaviour under low applied stresses, we now turn to the flow cessation experiments shown in Fig.~\ref{fig:residual}(a). The stress $\sigma(t)$ is seen to slowly relax and does not seem to tend to zero at long times. This indicates that internal stresses remain within the material over long time scales. For obvious practical reasons, we have to stop the relaxation after a given waiting time $t_\text{w}$, here $\SI{600}{\second}$. We then define the ``residual stress'' simply as $\sigma_\text{r}=\sigma(t_\text{w})$. This residual stress is of the order of 1~Pa and decreases with the preshear rate $\dot{\gamma}_\text{p}$. Figures~\ref{fig:residual}(b) and (c) show that for a given $\dot{\gamma}_\text{p}$, $\sigma_\text{r}$ is independent of the preshear duration $t_\text{p}$ but that it decreases as $\dot{\gamma}_\text{p}^{-0.2}$. This means that the material only keeps a memory of the preshear intensity (as long as it has been previously thoroughly fluidised by preshear).
As will be discussed in more details below in Sect.~\ref{sec:discuss_residual}, residual stresses result from the quench from a high shear rate to zero shear rate, which traps the microstructure into a configuration that slowly relaxes over time. Here, we note that the values taken by $\sigma_\text{r}$ roughly correspond to the stress $\sigma_\text{c}$ below which the Andrade exponent significantly differs from $\langle\alpha\rangle=0.39$ in Figs.~\ref{fig:andrade}(a) and \ref{fig:low_stress1}. It is also of the same order of magnitude as the value $\sigma_0=0.9$~Pa found in Fig.~\ref{fig:andrade}(b). Therefore, the anomalous creep behaviour observed at low stress most likely arises from an effect of this residual stress.
\subsubsection{Interpretation of anomalous creep at low applied stresses}
\label{sec:creep_low}
\begin{figure*}
\includegraphics[width=0.9\textwidth]{creep_bas_stress_tot.pdf}
\caption{(a)~Strain responses $\gamma(t)$ for creep experiments at $\sigma=\SI{1}{\pascal}$ after preshear at $\dot{\gamma}_\text{p}=$5, 20, $\SI{1000}{\per\second}$ from bottom to top ($t_\text{p}=\SI{60}{\second}$, $t_\text{w}=\SI{300}{\second}$). (b)~Strain response $\gamma(t)$ over one hour for a creep experiment at $\sigma=\SI{0.5}{\pascal}$ ($\dot{\gamma}_\text{p}=\SI{100}{\per\second}$, $t_\text{p}=\SI{60}{\second}$, $t_\text{w}=\SI{300}{\second}$).
(c)~Stress relaxation $\sigma(t)$ immediately after preshear ($\dot{\gamma}_\text{p}=\SI{100}{\per\second}$, $t_\text{p}=\SI{60}{\second}$). Same loading as in (b). Note that the origin of time corresponds to 300~s after preshear in (a,b) while preshear ends at $t=0$ in (c).}
\label{fig:low_stress2}
\end{figure*}
To confirm the link between anomalous creep and residual stresses, we go back to creep experiments at low imposed stress but we now vary the preshear rate $\dot{\gamma}_\text{p}$ for a fixed $\sigma=1$~Pa in Fig.~\ref{fig:low_stress2}(a). The strain response is observed to decrease initially only for the lowest preshear rate i.e. for the largest residual stress after the rest period of 300~s following preshear. When the preshear rate is increased such that the corresponding residual stress lies significantly below the applied stress, Andrade-like creep is recovered.
Figures~\ref{fig:low_stress2}(b) and (c) focus on two one-hour long experiments performed on the same loading, respectively a creep test at a very low stress of 0.5~Pa and a stress relaxation for the same preshear protocol. Interestingly, the strain $\gamma(t)$ in Fig.~\ref{fig:low_stress2}(b) decreases as long as the stress $\sigma(t)$ is larger than 0.5~Pa in Fig.~\ref{fig:low_stress2}(c) and the time at which $\sigma(t)$ reaches 0.5~Pa is of the same order of magnitude as the time at which $\gamma(t)$ goes through a minimum value of about $-0.4\%$ and starts increasing (see red dashes).
Our interpretation of anomalous creep at low stress is thus as follows. When brought to rest at a zero shear rate, the presheared material stores some internal stresses that do not fully relax even on very long time scales. After a waiting time of 300~s, the resulting macroscopic stress $\sigma_\text{r}$ is typically about 1~Pa. When a stress $\sigma$ is subsequently applied, the material actually ``feels'' an effective stress $\sigma-\sigma_\text{r}$. When $\sigma<\sigma_\text{r}$, this results in a decreasing strain response after the initial elastic deformation. During such a ``backward creep,'' the strain may then reach negative values. To put it differently, relaxing residual stresses tend to deform the material in the direction opposite to that of preshear. This is most probably because microstructural deformations induced by preshear in the forward direction are quenched during the rest time at zero shear rate so that releasing the stress (or applying $\sigma<\sigma_\text{r}$) tends to relax the structure in the reverse direction. As the microstructure slowly relaxes, the applied stress eventually overcomes residual stresses and the strain $\gamma(t)$ starts increasing, although this may take very long times. On the other hand, when $\sigma\gtrsim\sigma_\text{r}$, the strain increases right from the start of the creep experiment but residual stresses slow down the strain response, leading to the flatter $\gamma(t)$ observed for $1<\sigma<2$~Pa. This competition between structural relaxation and the applied stress leads to the anomalous creep response observed for $\sigma<\sigma_\text{c}\simeq 2$~Pa. Long-time behaviours similar to those reported here in Figs.~\ref{fig:low_stress1} and \ref{fig:low_stress2} have been observed previously in other microgels under low stresses \cite{Cloitre:2000,Purnomo:2007} and raise the question of aging. This will be further discussed in Sect.~\ref{sec:aging}.
\section{Discussion}
\label{sec:discussion}
Our main findings concern (i)~the evidence for robust power laws that characterize the creep behaviour of ETD~2050 carbopol microgels below the yield stress $\sigma_\text{y}$ and down to $\sigma_\text{c}\simeq 0.2 \sigma_\text{y}$ and (ii)~the presence of residual stresses after a rest time under a zero shear rate following preshear. These residual stresses are of the same order of magnitude as $\sigma_\text{c}$ and compete with the applied stress in creep experiments for $\sigma\lesssim\sigma_\text{c}$ and that raise the question of long-term relaxation and glassy dynamics. As shown in Appendix~B, these findings are robust to a change of carbopol concentration. Moreover, in Appendix~C, we test the more classical protocol where a zero shear stress (rather than a zero shear rate) is imposed during the rest time. These additional measurements show that power-law creep is still present and characterized by the same exponent but that anomalous creep is no longer observed at low applied stresses. Since this zero-stress protocol cancels out any residual stress, this confirms the interplay between creep and residual stresses when a quench to a zero shear rate is applied following preshear. In this section, we further discuss our results in light of previous works on soft materials.
\subsection{Andrade-like creep in soft materials}
\label{sec:andrade}
Power-law creep has been reported in a number of soft systems including cellulose solutions \cite{Plazek:1960}, aging polymers \cite{Cheriere:1997}, hexagonal columnar phases \cite{Bauer:2006}, collagen solutions \cite{Gobeaux:2010}, carbopol microgels similar to those studied here \cite{Moller:2009a,Divoux:2011a,Dimitriou:2013}, emulsions \cite{Paredes:2013,Dinkgreve:2015}, colloidal glasses \cite{Siebenburger:2012,Chan:2014}, colloidal gels made of attractive carbon black particles \cite{Grenard:2014}, natural gum \cite{Jaishankar:2012}, protein gels \cite{Brenner:2009,Brenner:2013,Leocmach:2014} and living cells \cite{Desprat:2005,Balland:2006,Kollmannsberger:2011,Hecht:2015}. The exponent $\alpha$ ranges from 0.2 to 0.7 depending on the system. Therefore, Andrade-like creep is ubiquitous in soft matter and the main open question is how it can be interpreted in terms of microstructure.
It appears that the various materials listed above can be broadly divided into two categories. A first category encompasses biological and/or fibrous-like materials such as living cells, protein, collagen or cellulose gels, whose creep eventually leads to irreversible failure. These materials are generally characterized not only by power-law creep but also by a ``power-law rheology'' in the sense that their linear viscoelastic moduli depend upon frequency as power laws, $G'\sim G''\sim f^\alpha$, with the same exponent as the strain in creep experiments. Power-law rheology is generally understood as deriving from the fractal nature of the underlying microstructural network \cite{Muthukumar:1989,Patricio:2015,Hung:2015}. Since the viscoelastic spectrum and the creep compliance are related through a Laplace transform, the link between power-law rheology and power-law creep is straightforward at least in the small deformation regime \cite{Desprat:2005,Tschoegl:2012}. Interestingly, based on Scott-Blair's fractional approach and material ``quasiproperties'' \cite{Blair:1947}, such a link can be extended to more complex linear rheologies and even to nonlinear viscoelasticity \cite{Jaishankar:2014}. Moreover, power-law creep, as well as the critical behaviour of some protein gels approaching failure \cite{Leocmach:2014}, is well captured by a family of models known as ``fiber-bundle models'' (FBMs) \cite{Kun:2003b,Kovacs:2008,Jagla:2011} that were originally devised for solid fibrous materials such as fiber-matrix composites \cite{Nechad:2005a,Nechad:2005b}. However, so far, some important assumptions of FBMs, such as local \textit{vs} global load sharing \cite{Kun:2003a} or damage accumulation \cite{Halasz:2012} by the ``fibers'' that constitute the microstructure, have not been evidenced experimentally in real soft systems.
The second category, which is more relevant to the present work, gathers a number of ``soft glassy materials'' that are characterized by an amorphous microstructure made of a dense assembly of particles. These materials include hexagonal columnar phases, colloidal glasses, emulsions and carbopol microgels. They generically show an almost frequency-independent elastic modulus so that there is not such an obvious link between their linear viscoelasticity and power-law creep as for the previous class of materials. Rather, the origin of power-law creep in soft glassy materials and its relation to their structural properties remain to be fully unveiled. In particular, a recent single-mode reduction of mode-coupling theory (MCT) generalized to transient regimes suggests that under certain conditions and in the low-frequency range, the elastic modulus is essentially frequency-independent and the viscous modulus scales as $G''\sim f^\alpha$ where $\alpha$ corresponds to the Andrade exponent \cite{Frahsa:2015}. Here, we note from Fig.~\ref{fig:oscill}(a) that $G''\sim f^{0.44}$ so that this exponent is indeed compatible with the Andrade exponent. Clearly, more systematic work is required to decide whether this is mere coincidence or whether there is a fundamental connection between the frequency-dependence of the viscous modulus and power-law creep in soft glassy materials. Let us also point out that a recent model of an elasto-plastic material with kinematic hardening predicts power-law creep where the Andrade exponent $\alpha$ is linked to the Herschel-Bulkley exponent $n$ through $\alpha=n/(n-1)$ for $\sigma<\sigma_\text{y}$ and $\alpha=0$ for $\sigma=\sigma_\text{y}$ \cite{Dimitriou:2013}. Unfortunately, with $0<n<1$, this corresponds to negative values of $\alpha$ i.e. to decreasing strains inconsistent with the Andrade-like creep observed experimentally.
By analogy with the motion of dislocations and vacancies that gives rise to Andrade creep in crystalline solids \cite{Cottrell:1952,Mott:1953,Cottrell:1997,Miguel:2002,Miguel:2008,Nabarro:2004}, it has been speculated that power-law creep arises from collective particle dynamics within the dense, disordered microstructure. A phenomenological model based on this idea and introducing a time-dependent fraction of mobile and arrested particles correctly predicts the Andrade exponent $\alpha\simeq 0.4$ observed in emulsions from the scaling of the steady-state flow curve with the volume fraction \cite{Paredes:2013}. Direct evidence for such heterogeneous dynamics under creep came only very recently from both simulations and experiments on hard-sphere colloidal glasses \cite{Chaudhuri:2013,Sentjabrskaja:2015}. Using simultaneous rheology and confocal microscopy, Sentjabrskaja {\it et al.} could make a quantitative link between the strain response $\gamma(t)$ and dynamical heterogeneities that remain spatially localized and sub-diffusive during creep but grow in size and become transiently super-diffusive at the onset of steady flow. In particular, they showed that the mean squared displacement $\Delta^2(t)$ follows the same power law as the strain, $\Delta^2(t)\sim\gamma(t)\sim t^\alpha$, with $\alpha\simeq 0.4$ for $\sigma=0.9\sigma_\text{y}$ in simulations and $\alpha\simeq 0.5$ for $\sigma\simeq\sigma_\text{y}$ in experiments. Although applied stresses smaller than the yield stress were not investigated experimentally, probably due to the very small value of $\sigma_\text{y}\simeq 10$~mPa, this study provides promising key results on microscopic dynamics during the creep of colloidal glasses. Interestingly, simulations show that dynamical heterogeneities may arrange into macroscopic shear bands during power-law creep \cite{Chaudhuri:2013} while local measurements --yet on rather large spatial scales-- have suggested that the deformation remains homogeneous \cite{Divoux:2011a,Grenard:2014}. This obviously deserves more work and confocal microscopy appears as an excellent tool to address this issue.
Of course the present measurements, which are restricted to global rheological data, do not provide any insight into local dynamics. However, we note that the Andrade exponent takes very similar values in carbopol, emulsions and colloidal glasses. Our results also show that power-law creep is observed below the yield stress down to very small applied stresses. We believe that microscopy experiments coupled to rheology are in line to fully understand creep in carbopol microgels where the above picture of mobile \textit{vs} arrested particles is likely to fail since the microstructure is constituted of swollen polymer particles that are compressed and deformed against each other.
\subsection{Residual stresses in soft materials}
\label{sec:discuss_residual}
Over the last few years, interest into residual stresses in soft materials has grown spectacularly. It is acknowledged that residual stresses (also sometimes referred to as ``internal'' stresses) result from history-dependent microstructural features that are frozen-in upon cessation of shear and generally decrease with the preshear rate or stress. Recent studies include polyelectrolyte microgels \cite{Mohan:2013,Mohan:2015}, colloidal gels \cite{Osuji:2008b,Negi:2009b,Negi:2010c} as well as colloidal glasses \cite{Ballauff:2013}. Numerical simulations of random jammed assemblies of non-Brownian elastic spheres have shown that residual stresses can be attributed to quenched angular distortions of the microstructure that slowly relax over time \cite{Mohan:2013}.
\begin{figure*}
\includegraphics[width=0.9\textwidth]{ageing_tot.pdf}
\caption{(a)~Time evolution of the elastic modulus $G'$ (red upper curve) and of the viscous modulus $G''$ (blue bottom curve) of a 1~\%~wt carbopol microgel after preshear at $\dot{\gamma}_\text{p}=\SI{100}{\per\second}$ for $t_\text{p}=\SI{60}{\second}$. $G'$ and $G''$ are recorded under a small-amplitude oscillatory shear of strain amplitude $\gamma=1~\%$ and frequency $f=\SI{1}{\hertz}$. The black solid line is the best logarithmic fit of the elastic modulus, $G'(t)=G_1+G_2\log(t)$, with $G_1=\SI{32.7}{\pascal}$ and $G_2=\SI{3.91}{\pascal}$. Inset: $G'(t)$ in semilogarithmic scales. (b)~Strain relaxation $\gamma(t)$ following flow cessation in a 1~\%~wt carbopol microgel. A preshear stress $\sigma=\SI{20}{\pascal}$, much larger than the yield stress and corresponding to a steady-state shear rate of about $\SI{5}{\per\second}$, is applied for $\SI{300}{\second}$. The system is then quenched to $\sigma=0$ at $t=0$. The red solid line is the best fit by a sum of three exponential functions with relaxation times 0.7, 5.3 and $\SI{56}{\second}$. The initial value $\gamma_0$ is inferred from this fit.}
\label{fig:aging}
\end{figure*}
To the best of our knowledge, in carbopol microgels, residual stresses have only been mentioned by a few studies and in the specific contexts of bubble rise \cite{Piau:2007,Mougin:2012}, penetrometry \cite{Boujlel:2012a}, capillary rise \cite{Geraud:2014} and surface tension measurements \cite{Jorgensen:2015}. While Coussot {\it et al.} have raised the possibility that residual stresses interfere with the creep response at low applied stresses in bentonite, mustard and a hair gel mostly constituted of carbopol \cite{Coussot:2006}, no quantitative analysis of this phenomenon had been performed as in the present work. As seen in Fig.~\ref{fig:residual}(c) for a fixed $t_\text{p}$, the residual stress decreases as a weak power law of $\dot{\gamma}_\text{p}$ with an exponent $-0.2$ which seems rather insensitive to the concentration [see Fig.~\ref{fig:concentration}(c) in Appendix~B]. Interestingly, residual stresses measured from experiments on colloidal glasses and from molecular dynamics simulations are consistent with such a scaling, although power-law fits were not performed by the authors (see Fig.~2 in \cite{Ballauff:2013}). Moreover, it could be argued that the logarithmic scaling evidenced in polyelectrolyte microgels, $\sigma_\text{r}\sim\log(\sigma_\text{p}-\sigma_\text{y})$ with $\sigma_\text{p}$ the preshear stress \cite{Mohan:2013,Mohan:2015}, is actually compatible with the present weak power law of $\dot{\gamma}_\text{p}$.
\subsection{Aging or not aging?}
\label{sec:aging}
Andrade-like creep and/or residual stresses are commonly predicted from various non-ergodic models for glassy behaviour, including the ``Soft Glassy Rheology'' (SGR) model derived from local probabilistic descriptions \cite{Fielding:2000}, the ``Shear Transformation Zone'' (STZ) theory \cite{Falk:2011,Bouchbinder:2011}, MCT \cite{Ballauff:2013,Frahsa:2015} and more phenomenological approaches \cite{Dinkgreve:2015,Joshi:2015}. There, the progressive slowing down of the shear rate observed under creep is attributed to aging, i.e. to the fact that the system explores deeper and deeper potential wells in the complex energy landscape that results from the wealth of ever-evolving configurations accessible to the glassy system. In the case of step-strain response, aging can be defined as ``the property that a significant part of the stress relaxation takes place on time scales that grow with the age $t_\text{w}$ of the system'' \cite{Fielding:2000}.
These observations obviously question the possible implication of aging in the interplay between creep and residual stresses observed in our experiments. As a matter of fact, decreasing strain responses under small applied stresses have been reported in polyelectrolyte microgels \cite{Cloitre:2000}. Together with logarithmic strain recovery curves after flow cessation, these peculiar creep responses have been interpreted in terms of aging phenomena. More precisely, the $\gamma(t)$ curves measured under a stress $\sigma\simeq 0.05\sigma_\text{y}$ first increase and then keep decreasing logarithmically over the course of several hours without any sign of levelling off. The time at which $\gamma(t)$ starts to decrease was shown to correspond to the ``age'' of the system, defined as the waiting time $t_\text{w}$ at rest between preshear and the creep measurement, so that all strain responses under low stresses can be collapsed by plotting $\gamma(t)-\gamma(t_\text{w})$ as function of $(t-t_\text{w})/t_\text{w}$. Similar results were later reported in thermosensitive pNIPAM microgels for $\sigma\simeq0.1\sigma_\text{y}$ \cite{Purnomo:2007} and successfully modeled with an STZ theory \cite{Bouchbinder:2011}. Such a striking behaviour, characterized by a response in the direction opposite to the preshear flow, was ascribed to structural relaxation and strong aging effects: after a rest time $t_\text{w}$, configurations with relaxation times faster than $t_\text{w}$ have relaxed so that, when a small stress is applied, the strain increases at short times but decreases for $t\gtrsim t_\text{w}$ as configurations with relaxation times longer than $t_\text{w}$ relax. Above a typical stress $\sigma_\text{c}\simeq 0.2\sigma_\text{y}$, ever-increasing strain responses are observed that are still dependent on $t_\text{w}$ due to competition between aging and partial rejuvenation. This transition between normal and anomalous creep occurs at a relative stress $\sigma_\text{c}/\sigma_\text{y}\simeq 0.2$ which is quantitatively very close to that found in our experiments.
\begin{figure}
\includegraphics[width=0.45\textwidth]{alpha_vs_tw.pdf}
\caption{Andrade exponent as a function of the waiting time $t_w$ after preshear for creep experiments at $\sigma=5$~Pa. The red line is $\langle \alpha\rangle = 0.37$. Experiments with $C=1$~\%~wt, $\dot{\gamma}_\text{p}=\SI{100}{\per\second}$ and $t_\text{p}=\SI{60}{\second}$.}
\label{fig:aging_andrade}
\end{figure}
In order to decide whether aging could be at play in the long-time creep behaviour of our carbopol microgels, we first test for the time evolution of the elastic modulus $G'(t)$ over one hour after preshear in Fig.~\ref{fig:aging}(a). As already reported on similar ETD~2050 carbopol samples \cite{Divoux:2011b}, we find that the elastic modulus recovers a value of about 40~Pa within a few seconds but that it subsequently increases logarithmically with time [see inset of Fig.~\ref{fig:aging}(a)] while the viscous modulus concomitantly decreases. Therefore, we observe a slow consolidation, i.e. enhanced elasticity and smaller dissipation over time, that could be interpreted as the signature of some aging process \cite{Fielding:2000,Derec:2003,Coussot:2006}. However, focusing on strain relaxation after preshear provides a somewhat different picture. Figure~\ref{fig:aging}(b) indeed shows that following a fast quench from a stress value $\sigma=20$~Pa well above the yield stress down to zero stress, the strain $\gamma(t)$ does not decrease logarithmically as observed in polyelectrolyte microgels by \cite{Cloitre:2000}. Rather it is well fitted by a sum of three exponentials, which hints at a viscoelastic-like relaxation mechanism. Thus, it appears that our system reaches an equilibrium state within a few minutes in contrast with strongly aging materials (compare with Fig.~1 in \cite{Cloitre:2000}). We also note that in the case of thermosensitive pNIPAM microgels, strain relaxations were not strictly logarithmic and seemed to level off after a few minutes (see Fig.~4 in \cite{Purnomo:2007}). This could be indicative of aging effects intermediate between the present study and those in \cite{Cloitre:2000}.
\begin{figure*}
\includegraphics[width=0.9\textwidth]{bas_stress_tw_tot.pdf}
\caption{(a)~Strain responses at $\sigma/\sigma_\text{y}=0.05$ for various wating times after preshear: $t_\text{w}=6$, 9, 42, 120, 300 and 600~s from bottom to top. Here, for the sake of consistency with previous works on aging systems \cite{Cloitre:2000,Purnomo:2007}, we explicitly denote by $t-t_\text{w}$ the time elapsed after the stress is applied at $t=t_\text{w}$. (b)~Same data plotted as a function of $(t-t_\text{w})/t_\text{w}$. Experiments with $C=1$~\%~wt, $\dot{\gamma}_\text{p}=\SI{100}{\per\second}$ and $t_\text{p}=\SI{60}{\second}$.}
\label{fig:aging_lowstress}
\end{figure*}
To further test the influence of the ``age'' of the system on its creep behaviour, we performed series of creep experiments with different rest times $t_\text{w}$ during which the viscoelastic moduli are monitored through small-amplitude oscillatory shear (strain amplitude 1~\% and frequency 1~Hz). When the applied stress is well above $\sigma_\text{c}$, we observe power-law creep whatever the age of the system. Moreover, as seen in Fig.~\ref{fig:aging_andrade} for $\sigma\simeq 0.5\sigma_\text{y}$, the Andrade exponent $\alpha$ does not display any systematic variation with $t_w$ and the average of 0.37 is consistent with $\alpha=0.39\pm 0.04$ found previously. This indicates that, if present, aging has little effect on Andrade-like creep.
In Fig.~\ref{fig:aging_lowstress}(a), a stress $\sigma=0.05\sigma_\text{y}$, well below $\sigma_\text{c}$, is applied for different values of $t_\text{w}$. There, a clear influence of $t_w$ on anomalous creep is observed: for short waiting times, the strain decrease is faster and much sharper. Such strain responses significantly differ from those of polyelectrolyte microgels (compare with Fig.~2 in \cite{Cloitre:2000}). Here, neither a strain maximum nor a long-time logarithmic decrease is observed. Moreover, as shown in Fig.~\ref{fig:aging_lowstress}(b), the strain response does not scale as $(t-t_\text{w})/t_\text{w}$. The influence of $t_\text{w}$ on $\gamma(t)$ at low stresses may thus be simply interpreted as the signature of the relaxation of residual stresses over (viscoelastic) time scales of a few tens of seconds: when $t_\text{w}\lesssim 100$~s, Fig.~\ref{fig:residual}(a) indeed shows that residual stresses sharply decrease so that for shorter $t_\text{w}$, a higher level of residual stress induces a stronger strain decrease but the strain eventually levels off and starts to increase as the applied stress becomes larger than the residual stress.
To conclude this discussion, the above comparison with previous results on microgels suggests that aging does not have a paramount influence on the creep behaviour of our carbopol samples. Even if we observe long-lasting stress and strain relaxations, these are most likely to be of viscoelastic origin than due to true glassy dynamics.
\section{Conclusion}
\label{sec:conclu}
We have shown that microgels constituted of carbopol ETD~2050 display power-law creep with an exponent of about 0.4 whatever the carbopol concentration. Additionally, when quenched from a fluidised state to a constant strain, these samples store residual stresses that relax in a viscoelastic-like fashion (rather than logarithmically as observed in other systems with strong aging effects). These residual stresses result in an anomalous creep behaviour under low shear stresses: for $\sigma<\sigma_\text{c}\simeq 0.2 \sigma_\text{y}$, the strain initially decreases but levels off and eventually increases after a time that depends on the preshear rate, on the rest time $t_\text{w}$ and on the applied stress. Here, we do not observe strain responses with long-time logarithmic decreasing trends that scale as $t_\text{w}$, which points to the absence of significant aging effects.
Future work will explore other types of carbopol microgels, such as carbopol Ultrez 10 and carbopol 940 or 941, in order to check for the generality of Andrade-like creep and residual stresses. Investigations based on confocal microscopy should also provide deeper insight into the origin of power-law creep and of stress relaxation in microgels.
\begin{acknowledgements}
The authors wish to thank P.~Coussot, T.~Divoux and G.~Ovar\-lez for enlightening discussions.
\end{acknowledgements}
\section*{Appendix A: Rheological characterization of carbopol ETD~2050 samples}
\subsubsection*{Flow curve measurement}
\begin{figure}
\includegraphics[width=0.45\textwidth]{flow_curve_tot.pdf}
\caption{Flow curve, shear stress $\sigma$ \textit{vs} shear rate $\dot{\gamma}$, of a 1~\% wt carbopol microgel measured in a sand-blasted cone-and-plate geometry by decreasing the shear rate in logrithmically-spaced steps of 5~s each. The red solid line is the best Herschel-Bulkley fit, $\sigma=\sigma_\text{y}+k\dot{\gamma}^n$, with $\sigma_\text{y} =10.2$~Pa, $n=0.60$ and $k=3.3$~Pa.s$^n$, inferred from the minimization procedure described in the text. The upper left inset shows $\sigma-\sigma_\text{y}$ with $\sigma_\text{y} = 10.2$~Pa as a function of $\dot{\gamma}$ in logarithmic scales. The upper right inset shows the residuals $\chi^2$ of the best power-law fit of $\sigma-\sigma_\text{y}$ \textit{vs} $\dot{\gamma}$ when varying the value of the yield stress $\sigma_\text{y}$ (see text).}
\label{fig:flowcurve}
\end{figure}
The flow curve of our 1~\% wt carbopol microgel recorded through a downward shear-rate sweep is shown in Fig.~\ref{fig:flowcurve} together with the best HB fit, $\sigma=\sigma_\text{y}+k\dot{\gamma}^n$, yielding $\sigma_\text{y} =10.2$~Pa, $n=0.60$ and $k=3.3$~Pa.s$^n$. Let us emphasize here our method for estimating the yield stress $\sigma_\text{y}$. Instead of fitting the flow curve with the full HB model that involves three free parameters, which can raise some convergence issues due to nonlinearity depending on the fitting algorithm and on the initial guess for the parameters \cite{Mullineux:2008}, we first set $\sigma_\text{y}$ to some arbitrary value and compute the best linear fit of $\ln(\sigma-\sigma_\text{y})$ \textit{vs} $\ln\dot{\gamma}$. Fitting a straight line in logarithmic coordinates rather than fitting $\sigma-\sigma_\text{y}$ \textit{vs} $\dot{\gamma}$ as a power law allows us to give the same weight to small shear rates and to larger ones. We compute the residuals $\chi^2$ of this fit, defined as the sum of the squared distances from the experimental data to the fit. We then vary $\sigma_\text{y}$ systematically over a realistic range and look for a minimum of $\chi^2$ \textit{vs} $\sigma_\text{y}$. As shown in the upper right inset of Fig.~\ref{fig:flowcurve}, $\chi^2$ goes through a well-defined minimum: the stress $\sigma_\text{y} =10.2$~Pa corresponding to this minimum is thus taken as the yield stress of our microgel. It can be checked in the upper left inset of Fig.~\ref{fig:flowcurve} that with this value of $\sigma_\text{y}$, the difference $\sigma-\sigma_\text{y}$ follows a strict power law of $\dot{\gamma}$ with no significant deviation over the whole range of shear rates. This method allows us to confidently estimate the yield stress of our microgels to within 1~\% for a given flow curve measurement. In the case of Fig.~\ref{fig:flowcurve}, this procedure yields $\sigma_\text{y} =10.2$~Pa, $n=0.60$ and $k=3.3$~Pa.s$^n$. We note that a direct HB fit of the same data with three free parameters leads to significantly different estimates ($\sigma_\text{y} =9.0$~Pa, $n=0.55$ and $k=4.5$~Pa.s$^n$), which illustrates the importance of using a careful fitting procedure as described here.
Still, it should be noted that we observe a reproducibility of the best HB fit parameters of only about 10~\%. For instance, for the same preparation batch, we found values of $\sigma_\text{y}$ ranging from 9.5 to 10.5~Pa from one loading of the cone-and-plate geometry to the other. Since the flow curves are measured through decreasing shear rate sweeps, the loading history is efficiently erased at high shear rates and such variations can only be explained by small differences in the sample volume from one loading to the other. We checked that the various flow curves are simply shifted along the stress axis from one measurement to the other and that the scatter is typically 1~Pa, which confirms that variations in HB parameters essentially stem from variations in the loaded sample volume.
\subsubsection*{Linear and nonlinear viscoelastic measurements}
\begin{figure*}
\includegraphics[width=0.65\textwidth]{oscillatory_caracterisation_tot.pdf}
\caption{(a)~Elastic modulus $G'$ (red) and viscous modulus $G''$ (blue) \textit{vs} frequency $f$ after a rest time $t_\text{w}=300$~s following preshear under a constant shear rate $\dot{\gamma}_\text{p}=\SI{100}{\per\second}$ for a duration $t_\text{p}=\SI{60}{\second}$. The strain amplitude is fixed to $\gamma=1~\%$ and the frequency is logarithmically swept down. The orange solid line shows that $G''\sim f^{0.44}$ for $f\gtrsim\SI{0.2}{\hertz}$. (b)~$G'$ (red, left axis), $G''$ (blue, left axis) and stress amplitude $\sigma$ (black, right axis) \textit{vs} strain amplitude $\gamma$ after the same preparation protocol as in (a). The frequency is fixed to $f=\SI{1}{\hertz}$ and the strain amplitude $\gamma$ is logarithmically swept up with a duration of 16~s per point. One has $G'(\gamma^*)=G''(\gamma^*)$ for $\gamma^*\simeq 200~\%$ which corresponds to $\sigma^*\simeq\SI{36}{\pascal}$. The red solid line is the best linear fit, $\sigma=G_0\gamma$ for $\gamma<80~\%$, leading to $G_0 = \SI{34}{\pascal}$. The dotted line indicates the end of the linear regime at $\gamma\simeq 80~\%$.}
\label{fig:oscill}
\end{figure*}
The linear viscoelastic moduli of our microgels are shown as a function of oscillation frequency $f$ in Fig.~\ref{fig:oscill}(a). The storage modulus $G'(f\rightarrow 0)=37$~Pa increases weakly over the whole frequency range. Here again, we observe variations of $G'$ by about 10~\% from one loading to the other. The loss modulus $G''$ remains always smaller than $G'$ and can be well fitted by a power law $G''\sim f^{0.44}$ at high frequencies.
The oscillatory strain sweep of Fig.~\ref{fig:oscill}(b) shows that the linear regime, characterized by an elastic modulus $G_0=34$~Pa, extends up to strain amplitudes of about 100\%. The nonlinear regime involves a sharp drop of the storage modulus and a local maximum in the loss modulus. This corresponds to a case of ``weak strain overshoot'' as classified by Huyn {\it et al.} in their review on large-amplitude oscillatory shear \cite{Hyun:2011} and appears as a distinctive feature of soft glassy materials, including microgels \cite{deSouzaMendes:2014}. In systems like emulsions and microgels, the maximum in $G''$ is generally attributed to enhanced dissipation due to local irreversible particle rearrangements that progressively invade the whole sample before yielding and complete fluidisation \cite{Mason:1995c,Knowlton:2014}. We note that the point at which $G'$ and $G''$ cross corresponds to a stress of about 36~Pa, significantly above the yield stress measured from the flow curve in Fig.~\ref{fig:flowcurve}. Since this article is mostly devoted to creep experiments performed below the yield stress, where the strain shall not increase beyond 100~\%, we do not expand more on the nonlinear behaviour of our carbopol ETD~2050 samples and refer the reader to the cited literature for more details.
\section*{Appendix B: Influence of the carbopol concentration}
The robustness of our findings has been tested by considering two other concentrations of carbopol ETD~2050, namely $C=0.6$~\%~wt and 2~\%~wt. As shown in Fig.~\ref{fig:concentration}(a) for similar normalized stresses, $\sigma/\sigma_\text{y}\simeq 0.4$, power-law creep is observed in all three samples. Table~\ref{tab:concentration} gathers the rheological parameters of the various samples as well as the results of the analysis of strain responses in the Andrade-like regime as described in Sect.~\ref{sec:multiplecreep}. We find a remarkably robust mean Andrade exponent of $\alpha\simeq 0.4$ for all concentrations. Here again, the prefactor $G'_0$ deduced from the initial elastic deformation $\gamma_0$ is in good agreement with the elastic modulus $G'$. The parameter $\sigma_0$ is non-zero and increases with $C$ in the same fashion as the elastic modulus and the yield stress. This suggests that residual stresses also come into play for $C=0.6$~\%~wt and 2~\%~wt.
\begin{figure*}
\includegraphics[width=0.9\textwidth]{concentration_tot.pdf}
\caption{Creep and residual stresses in carbopol microgels with concentrations $C=0.6$~\% wt (blue), 1~\% wt (black) and 2~\% wt (red). (a)~Strain responses $\gamma(t) - \gamma_0$ for $\sigma\simeq 0.4\sigma_\text{y}$ plotted in logarithmic scales together with Andrade fits (yellow solid lines): [$C$, $\sigma/\sigma_\text{y}$, $\gamma_0$, $\alpha$] = [0.6~\% wt, 0.38, 9.9~\%, 0.40], [1~\% wt, 0.36, 8.9~\%, 0.37], [2~\% wt, 0.35, 16.2~\%, 0.40]. (b)~Strain responses $\gamma(t)$ for $\sigma/\sigma_\text{y}\le 0.1$ plotted in semilogarithmic scales for [$C$, $\sigma/\sigma_\text{y}$] = [0.6~\% wt, 0.04], [1~\% wt, 0.04], [2~\% wt, 0.1]. (c)~Residual stress $\sigma_\text{r}$ (after a relaxation over $t_\text{w}=\SI{600}{\second}$) as a function of the preshear rate $\dot{\gamma}_\text{p}$ for $t_\text{p}=\SI{60}{\second}$. Solid lines are power laws with exponent $-0.2$.}
\label{fig:concentration}
\end{figure*}
Indeed, for applied stresses such that $\sigma/\sigma_\text{y}< 0.1$, the same decreasing trend is observed in the strain response for all three concentrations [see Fig.~\ref{fig:concentration}(b)]. The presence of residual stresses for $C=0.6$~\%~wt and 2~\%~wt is further confirmed through flow cessation experiments performed as in Sect.~\ref{sec:residual}. Figure~\ref{fig:concentration}(c) shows that $\sigma_\text{r}$ decreases roughly as $\dot{\gamma}_\text{p}^{-0.2}$ for all concentrations with an amplitude that scales with $C$ like $G'$ and $\sigma_\text{y}$ (see Table~\ref{tab:concentration}). We conclude that a similar interplay between creep and residual stress occurs at low stresses whatever the carbopol concentration.
\begin{table}
\caption{Elastic modulus $G'$ and yield stress $\sigma_\text{y}$ (see Appendix~A), average Andrade exponent $\langle\alpha\rangle$ and fitting parameters $G'_0$ and $\sigma_0$ (see Sect.~\ref{sec:singlecreep}), and residual stress $\sigma_\text{r}$ measured 600~s after preshear at $\dot{\gamma}_\text{p}=\SI{100}{\per\second}$ (see Sect.~\ref{sec:residual}) for carbopol microgels with different concentrations $C$.}
\label{tab:concentration}
\begin{tabular}{clclclclclclc}
\hline\noalign{\smallskip}
$C$ (\%~wt) & $G'$ (Pa) & $\sigma_\text{y}$ (Pa) & $\langle\alpha\rangle$ & $G'_0$ (Pa) & $\sigma_0$ (Pa) & $\sigma_\text{r}$ (Pa)\\
\noalign{\smallskip}\hline\noalign{\smallskip}
0.6 & $13 \pm 1$ & $3.2 \pm 0.2$ & $0.40\pm 0.07$ & $9.8$ & $0.15$ & 0.20 \\
1 & $37 \pm 4$ & $10.0 \pm 0.5$ & $0.39\pm 0.04$ & $32$ & $0.9$ & 0.65 \\
2 & $105 \pm 10$ & $34.5 \pm 0.7$ & $0.38\pm 0.03$ & $120$ & $4.7$ & 4.9 \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
\section*{Appendix C: Influence of the rest protocol}
In order to test the influence of the rest protocol, Fig.~\ref{fig:rest1} compares creep experiments following a rest period performed either under a zero shear stress or under a zero shear rate (or more precisely under small strain oscillations with amplitude 1\% and frequency 1~Hz, which were checked to effectively correspond to a zero shear rate). As shown in Fig.~\ref{fig:rest1}(a) for $\sigma=\SI{5}{Pa}$, the Andrade exponent is robust to a change of rest protocol but the initial deformation $\gamma_0$ is slightly smaller when rest is imposed under a zero shear rate. Moreover, when the imposed stress is decreased to $\sigma=\SI{0.4}{Pa}$, anomalous creep characterized by a decreasing strain is recovered only in the case of rest under a zero shear rate [see Fig.~\ref{fig:rest1}(b)]. In the case of rest performed under a zero shear stress, Andrade-like response persists down to the lowest imposed stresses as shown by the fit in Fig.~\ref{fig:rest1}(b).
Andrade fit parameters are displayed in Fig.~\ref{fig:rest2} which confirms the good collapse of the exponents $\alpha$ independent of the rest protocol, provided the imposed stress is large enough and anomalous creep is avoided. The linear fits of $\gamma_0$ vs $\sigma$ show that the difference noted above in the initial deformation does not stem from the slope $G'_0$ which remains close to the elastic modulus of the microgel but from the intercept $\sigma_0$ which is significantly larger in the case of rest under zero shear rate (0.7~Pa) than in the case of rest under zero shear stress (0.2 ~Pa). These results indicate that imposing a zero shear stress during the rest time essentially cancels out residual stresses so that power-law creep is observed even at very low applied stresses.
\begin{figure*}
\includegraphics[width=0.9\textwidth]{appC_fig1.pdf}
\caption{Influence of the rest protocol on creep in a 1~\% wt carbopol microgel. Following preshear at $\dot{\gamma}_\text{p}=\SI{100}{\per\second}$ for $t_\text{p}=\SI{10}{\second}$, a rest time of $t_\text{w}=\SI{300}{\second}$ is imposed either under a zero shear rate ($\bullet$) or under a zero shear stress ($\square$) on the same loading of the cone-and-plate geometry. (a)~Strain responses $\gamma(t) - \gamma_0$ for $\sigma=\SI{5}{Pa}$ plotted in logarithmic scales together with Andrade fits (red solid lines) yielding respectively $\gamma_0=11.4\%$ and $\alpha=0.38$ for rest under a zero shear rate and $\gamma_0=13.4\%$ and $\alpha=0.35$ for rest under a zero shear stress. (b)~Strain responses $\gamma(t)$ for $\sigma=\SI{0.4}{Pa}$ plotted in semilogarithmic scales. The red solid line is the Andrade fit in the case of rest under a zero shear stress with $\gamma_0=1.2\%$ and $\alpha=0.33$.}
\label{fig:rest1}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.9\textwidth]{appC_fig2.pdf}
\caption{Andrade fit parameters for the two rest protocols used in Fig.~\ref{fig:rest1} on the same loading ($\bullet$: zero shear rate, $\square$: zero shear stress): (a)~exponent $\alpha$ and (b)~initial elastic deformation $\gamma_0$, as a function of the applied stress $\sigma$. The vertical dashed line shows the typical stress of 2~Pa below which anomalous creep is observed when rest is imposed at a zero shear rate. Solid lines in (b) are linear fits $\sigma=G'_0\gamma_0+\sigma_0$ with $G'_0=\SI{26.3}{\pascal}$ and $\sigma_0=\SI{0.7}{\pascal}$ for rest under a zero shear rate (red line) and $G'_0=\SI{27.0}{\pascal}$ and $\sigma_0=\SI{0.2}{\pascal}$ for rest under a zero shear stress (blue line).}
\label{fig:rest2}
\end{figure*}
|
\section{Introduction}
\label{sec1}
Solar filaments/prominences are characterized as cool and dense
structures that lie above the solar surface in the hot corona
\citep{Labrosse10,Mackay10}. Filaments exist in the magnetic dips
within magnetic configurations known as filament channels or flux
ropes \citep{Aulanier02,Liu12}. Filaments and/or associated
channels sometimes interact and show interesting dynamics in the
chromosphere and low corona
\citep{Uralov02,Schmieder04,Su07,Bone09,Kumar10,Chandra11,Li12,Liu10,Filippov11,Jiang13,Jiang14,Joshi14b}.
During interaction under some specific conditions, these magnetic
structures can reconnect and change their foot point connectivity.
They can also merge to form one common filament. The first type is
known as 'slingshot' reconnection and the second type is 'merging'
(see paper by \cite{Linton01,Linton05}). Besides, large-scale
flux-rope interaction/merging in the outer corona has also been
observed in the form of CME-CME interaction and their merging
\citep{Gopalswmay01,Joshi13b}.
Some observational studies suggest the observational evidence of slingshot
magnetic reconnection between filaments
\citep{Kumar10,Chandra11,Filippov11,Jiang13}. \cite{Kumar10} and
\cite{Chandra11} first reported the interaction, reconnection, and
footpoint connectivity change between two nearby filaments using
H$\alpha$ observations on 2003 November 20. Later on,
\cite{Filippov11} reported a few observational cases showing pairs
of large filaments joining and exchanging their halves. More
recently, \cite{Jiang13} reported another observational evidence
of partial slingshot reconnection during interaction of two
filaments on 2011 December 3.
There are relatively few numerical simulations, which have been
performed for the slingshot reconnection between flux ropes
\citep{Linton01,Linton05,Torok11a}. \cite{Linton01} and
\cite{Linton05} presented different types of flux rope interaction
using three dimensional magnetohydrodynamic (MHD) simulations for
convection zone conditions. Later on, \cite{Torok11a} simulated
the 2003 November 20 filament interaction using a three
dimensional zero $\beta$ MHD model for coronal conditions and
interpreted it in terms of 'slingshot' reconnection between two
magnetic flux ropes.
Merging of two filament channels after dynamic interactions have
also been observed by \cite{Schmieder04,Bone09,Jiang14,Joshi14b}.
\cite{Schmieder04} found evidence of merging of two segments with
dextral chiralities to form a long dextral filament. Recently,
\cite{Jiang14} reported the interaction and merging of two
sinistral filaments on 2001 December 6 and found that they form a
new long magnetic channel. More recently, \cite{Joshi14b} reported
an interesting dynamic event of merging of two filament channels
and formation of a long compound flux rope on 2014 January 1. On
the basis of numerical simulations, some authors discussed various
conditions necessary for interaction/merging of filaments.
\cite{DeVore05} and \cite{Aulanier06} modeled filaments as differentially
sheared arcades and found that two filaments, occupying a single
polarity inversion line (PIL) in a bipolar large-scale magnetic
configuration, easily merged if their chiralities were identical
and axial magnetic fields were aligned. This is in accordance with
empirical rules for filament interaction found by
\cite{Martin94} and \cite{Schmieder04}. However, in a quadrupolar
configuration the situation is more complex and ambiguous
\citep{DeVore05,Linton06,Romano11}.
\cite{Linton01} and \cite{Linton05} analyzed numerically the reconnection of two twisted flux tubes contacting at different angles. The result of interaction depends on the twist handedness of the tubes and the angle between their axial magnetic fields. A pair of oppositely twisted flux tubes shows a bounce interaction, if their axial magnetic fields are parallel, and slingshot reconnection, if their axial magnetic fields are anti-parallel or perpendicular. \cite{Linton01} and \cite{Linton05} considered isolated flux tubes without a surrounding magnetic field. But flux ropes containing filaments are not isolated flux tubes. They are imbedded into coronal magnetic fields created mostly by photospheric sources and follow basically photospheric PILs. For example, observations of filaments crossing each other at different heights are very rare. Oppositely twisted flux ropes with antiparallel axial magnetic fields need the presence of an additional PIL between them to be in equilibrium in the coronal magnetic field. Therefore, the slingshot interaction of filaments of different chiralities may be different.
In this paper, we present observations
of interaction of two adjacent filament channels of different chirality
associated with two adjacent PILs within a fan-spine configuration.
This kind of dynamic interaction has not been discussed in detail before.
We discuss the filament-channel interaction dynamics,
the probable magnetic reconnection at a null point above
them, different helical motions, and the apparent exchange of
heated plasma between different filament channels. The structure
of the paper is as follows: Section~\ref{sec2} deals with the
description of the observational data set used in the paper.
Morphology and magnetic structure of the filaments are discussed
in Section~\ref{sec3}. Different events of interactions and
associated plasma dynamics are described in Section~\ref{sec4}. In
Section~\ref{sec5}, we present an interpretation of the observed
phenomenon in the light of filament flux-rope models and
discussion. Main results and conclusions are listed in
Section~\ref{sec6}.
\section{Observations}
\label{sec2}
We used Big Bear Solar Observatory (BBSO) and {\it National Solar Observatory} (NSO)/
{\it Global Oscillation Network Group} (GONG) H$\alpha$
observations for the present study. The BBSO high resolution
images are collected from archive \url{http://www.bbso.njit.edu/}
are used to investigate the chiralities of the filaments.
The GONG data are available in the
data archive at \url{http://halpha.nso.edu/archive.html} with
full-disk images in 6563 \AA~line. The images have spatial
resolution of 1$"$ and a cadence of around 1 min \citep{Har11}.
The GONG H$\alpha$ observations are used to get the information about
the filament activation and partial eruption dynamics. We also
used data of the {\it Atmospheric Imaging Assembly} (AIA;
\cite{Lem12}) instrument on board the {\it Solar Dynamics
Observatory} (SDO). It observes the full disk of the Sun in
ultra-violet (UV) and extreme ultra-violet (EUV) wavelengths with
a minimum cadence of 12 s and a pixel size of 0.6$"$. We used AIA
images in 304 and 171, 193, 131 and 94 \AA\ wavelength channels. The line-of-sight
(LOS) photospheric magnetic field data are obtained by the {\it
Helioseismic and Magnetic Imager} (HMI; \cite{Schou12}), with a
spatial resolution of 1$"$ and a minimum cadence of 45 s. It is
also an instrument on board {\it SDO}.
\section{Morphological and Magnetic Structure of the Filaments}
\label{sec3}
Figures~\ref{fig1}(a) and \ref{fig1}(c) represent the BBSO H$\alpha$
images at $\sim$18:46 UT on 2014 April 15 and $\sim$17:01 UT on
2014 April 16, respectively. Figure~\ref{fig1}(b) shows the
\textit{SDO}/AIA 304 \AA\ images at $\sim$18:46 UT
on 2014 April 15. These images show the filaments and filament
channels about 3-4 days before the first interaction that starts on
2014 April 18. Two dark filaments, named as the northern filament
(NF) and southern filament (SF), are clearly seen in the
H$\alpha$ image (Figures~\ref{fig1}(a) and (c)). The extended filament
channels of both filaments can be seen in the \textit{SDO}/AIA 304
\AA\ EUV image (Figure~\ref{fig1}(b)). Figure~\ref{fig1}(d) shows
the \textit{SDO}/HMI LOS magnetogram at 18:46:19 UT on 2014 April
15. To find out the filament positions and approximate endpoint
locations, we tracked the filament spines from the H$\alpha$ image
(see Figure~\ref{fig1}(a)) and overplotted them in the LOS
magnetogram (Figure~\ref{fig1}(d)). NF/SF axis are shown by the
red/orange colors, respectively in Figure~\ref{fig1}(d). After comparing
H$\alpha$, EUV 304 \AA\ images, and the LOS magnetogram we
determine that the eastern/western ends of both filaments are
anchored in negative/positive polarity, respectively. Both
filaments are stretched approximately from the south-east to the
north-west along two different PILs.
The handedness or chirality of filaments can be determined using
high-resolution H$\alpha$ images and the position of filament ends
relative to the photospheric LOS magnetic fields.
Figure~\ref{fig1}(d) hints that the eastern ends of both filaments
are anchored in negative polarities, while the western ends are
rooted in positive polarities. In this case, the axial
magnetic field in both filaments is directed from west to east.
Accordingly, NF is sinistral, while SF is dextral because they are
separated by positive polarity.
The H$\alpha$ images in Figures~\ref{fig1}(a) and~\ref{fig1}(c) also show
that fine threads within NF and SF bodies deviated counterclockwise and
clockwise from their axes, respectively. The filament barbs are
left-bearing/ right-bearing for the NF/SF, which corresponds to
sinistral chirality of NF and dextral chirality of SF
\citep{Martin94,Martin98a}. Some of the visible barbs are marked
with the red arrows in Figures~\ref{fig1}(a) and~\ref{fig1}(c).
The magnetic configuration surrounding the filaments can be
deduced from the analysis of \textit{SDO}/AIA 171 \AA\ images
(Figure~\ref{fig2}) and the potential-field source-surface (PFSS)
magnetic-field extrapolation \citep{Schrijver03}
(Figure~\ref{fig3}). Figures~\ref{fig2}(a)--(d) represent
\textit{SDO}/AIA 171 \AA\ images at 10:48:59 UT on 2014 April 18
and at 00:36:11 UT, 10:28:59 UT, and 17:57:59 UT on 2014 April
19, respectively. To compare the coronal loop structure with a LOS
magnetogram, we overplot positive (green) and negative (blue) LOS
magnetic field contours on the AIA 171 \AA\ image
(Figure~\ref{fig2}(a)). We clearly see arcades, connecting the
central positive polarity with negative polarities on both sides,
above the filaments. We also see long loops that connect the
negative polarities to the remote region of positive polarity,
near to the western endpoint of SF. All these images show the
fan-spine configuration over the filaments. A null point is
expected to be above the central positive polarity between the
filaments. The outer spine field line emanates from the null point
but is directed not radially into the outer corona, as it is
usually assumed in fan-spine configurations, but deviates to the
west and touches the photosphere near the western endpoint of SF
within an area of positive polarity. This configuration is
confirmed by the PFSS magnetic field extrapolation
(Figure~\ref{fig3}). To perform the PFSS extrapolation we used the
PFSS software package available in IDL SolarSoftWare.
Figure~\ref{fig3}(a) shows the full disk magnetogram with
extrapolated magnetic field lines, while the zoomed region
corresponding to the black box is shown in Figure~\ref{fig3}(b).
The calculated field lines match quit well to the structure of
coronal loops in \textit{SDO}/AIA 171 \AA\ images.
The filament channels were approaching each other from April 15 to
April 18 as seen in \textit{SDO}/AIA 304 \AA\ images in
Figure~\ref{fig4}. The filament channels manifest themselves as
long dark structures. They are quite separated on April 15
(Figure~\ref{fig4}(a)), come slowly closer to each other during
April 16-17 (Figure~\ref{fig4}(b)-(c)), and become very close on
April 18 (Figure~\ref{fig4}(d)). The closing of the filament
channels is marked by the white arrows in all the panels in
Figure~\ref{fig4}.
\section{Dynamic Interactions of the Filament Channels}
\label{sec4}
We observed three events of filament channel interactions during
2014 April 18--20. In this section we describe the detail observation
of these interaction dynamics in multiwavelength channels.
\subsection{First Event of Interaction and Associated Dynamics}
\label{sec4.1}
The sequence of the images showing the first event
of interaction are represented in Figure~\ref{fig5}. The left
panel shows the \textit{SDO}/AIA 304 \AA\ images ((a)--(d)), while
the right panel show the NSO/GONG H$\alpha$ images ((e)--(h)). In
Figure~\ref{fig5}(a), the \textit{SDO}/AIA 304 \AA\ image is
overplotted with \textit{SDO}/HMI LOS magnetogram contours.
Green/blue contours show the positive/negative polarity regions,
respectively. In Figure~\ref{fig5}(a), we see two nearby filament
channels. However, at the same time in the H$\alpha$ image only SF
is visible (Figure~\ref{fig5}(e)). The first event of interaction
starts around 20:14 UT on 2014 April 18 with the activation of
the middle part of SF. The initial activation area is shown by the
small white circle in Figure~\ref{fig5}(a). After the activation,
the middle part of the filament partially erupts towards the
north. The activated filament is seen in Figure~\ref{fig5}(b).
Along with the partial eruption we also see the counterclockwise
rotation of filament threads around the long filament axis, if we
observe it from the east end (see the AIA 304 \AA\ animation
associated with Figure~\ref{fig5}). This counterclockwise rotation
of the threads can be the manifestation of the redistribution of the
twist along the flux rope due to its expansion and swelling during
the activation \citep{Parker74}.
Negative helicity of a flux rope corresponds to the dextral chirality of a
filament in flux-rope models.
The partial failed eruption of SF seems to trigger reconnection at
the magnetic null that lie above the filament channels. It is
exhibited by EUV brightenings at several places on either side of
both filament channels simultaneously with the partial eruption
and helical motion. These bright regions are marked by the white
circles in Figure~\ref{fig5}(c). Such brightenings are believed to
appear due to hits of the chromosphere by fast electrons and
heated plasma from the region of reconnection. After the partial
failed eruption the heated as well as cool material of SF moves
along the axis towards the eastern and western ends of the
filament (directions are shown by the arrows in
Figure~\ref{fig5}(d)). The two separate filament channels are
still observed very close to each other at 20:46:07 UT. Apart from
the chromospheric brightenings, no influence of SF activation on
NF was observed.
The distance--time plot of the hot plasma movement is presented in
Figure~\ref{fig6}(a). The rough trajectory along which the distance
measurements was made is shown by dashed black line in
Figure~\ref{fig5}(c). We tracked a bright plasma blob that moves
towards the western end of SF. The east-most point was used as a
reference point for the distance measurements. For more accurate
results, we repeated the measurements three times and the standard
deviations was used as errors. The linear fit to these data points
is used to estimate an average speed. It is evident that the hot
plasma moved with the average speed of $\sim$40 $\rm km~s^{-1}$
between 20:25 and 20:50 UT.
\subsection{Second Event of Interaction and Associated Dynamics}
\label{sec4.2}
Figure~\ref{fig7} represents the selected \textit{SDO}/AIA 304
\AA\ and NSO/GONG H$\alpha$ images showing the second event of
interaction and associated dynamics. Second event of interaction
started at $\sim$15:33 UT on 2014 April 19 with a compact
brightening near the place of the closest approaching of the
filament channels, just between them. Another small brightening
appeared on the southern side of SF. The locations of these
compact brightenings are marked by the white circles in
Figure~\ref{fig7}(a). The activation of NF started at $\sim$15:37
UT near its eastern end (Figure~\ref{fig7}(b)). At the same time,
we also see a remote brightening (RB) on the west. The location of
RB region is shown in Figure~\ref{fig7}(b) with the white circle.
Thereafter, the bright features propagate from the eastern end of
NF to its middle part. The pattern of filament bright and dark
threads looks like the upper part of a right-handed helix, which
is consistent with the sinistral chirality of NF. What is most
surprising, after reaching the place of the closest approaching of
the filament channels the heated plasma propagates not to the
north-west along the axis of the NF channel but to the west and
south-west nearly along the axis of SF. At first glance, one might
fancy that the eastern part of NF and the western part of SF form
a joint magnetic structure allowing plasma to move easily from the
eastern end of NF to the western end of SF. However, it is very
doubtful if they can form such a structure because their chirality
and helicity are opposite. We will discuss the problem in more
detail in Section~\ref{sec5}.
Signature of brightening around the magnetic null can also be seen in other AIA
channels. Figure~\ref{fig8} shows the \textit{SDO}/AIA 171, 193, 131 and 94
\AA\ image at $\sim$15:41 UT just after the partial filament eruption.
We can see the brightening near the null point, which can be understood
due to the magnetic reconnection there. The remote brightening signature
is also visible in these channels.
The heated plasma first moves towards northwest direction with a
speed of $\sim$90 $\rm km~s^{-1}$ and then towards southwest
direction towards the western end of SF with a speed of $\sim$140
$\rm km~s^{-1}$. The kinematics of the plasma flows is shown in
Figure~\ref{fig6}(b). It represents the distance--time profiles of
the plasma flows towards the north-west (red curve) and the
south-west (green curve). The trajectories along which the
northwest directed (white dashed line) and southwest directed
(black dashed line) displacement measurements have been performed
are shown in Figure~\ref{fig7}(d). Two compact bright areas
appeared on both sides of NF near its eastern end closer to the
end of the second event (Figures~\ref{fig7}(d), (g) and (h)). They
can be considered as the compact ribbons formed by the partial
eruption of the NF and the associated flare. The overplotted \textit{SDO}/AIA 171 \AA\ image
at around $\sim$16:21 UT shows the loop like structures joining
the two bright ribbons and can be considered as the post flare loops.
\subsection{Third Event of Interaction and Associated Dynamics}
\label{sec4.3}
The third event in many features is similar to the second event of
interaction. Figure~\ref{fig9} show the interaction dynamics in
\textit{SDO}/AIA 304 \AA\ and NSO/GONO H$\alpha$ observations.
It starts at $\sim$00:14 UT on 2014 April 20 with
brightenings near the place of the closest approaching of the
filament channels and the southern side of SF. Immediately after
that at $\sim$00:15 UT we observe an activation and failed
eruption of the eastern part of NF with formation of two bright
ribbons on both sides of it. These ribbons can be formed as a result of reconnection between the legs of surrounding arcades during the partial eruption of the NF inside the northern arcades of the fan-spine structure. Soon after the activation, a remote
brightening appears near the western end of SF at nearly the same
place as in the second event (Figure~\ref{fig9}(b), (e) and (f)).
We also observed the brightening at the magnetic null point just
after the partial eruption of NF in All the EUV channels
(Figures~\ref{fig9}(b) and~\ref{fig10}). Figure~\ref{fig10} represents
\textit{SDO}/AIA 171, 193, 131 and 94 \AA\ images at $\sim$00:18 UT
on 2014 April 20 also showing the brightening at the null. This brightening
is due to some magnetic reconnection at the null point. The outer
spine lines and the remote brightening can be seen in the hotter
AIA channels (Figures~\ref{fig10}(c) and (d)).
Heated plasma of the eastern part of NF forms a wide bright
helical structure with intensive internal motions. Some part of
the hot plasma moves from the middle of NF to the north-west along
its axis to the western end, while a fraction of bright material
moves to the western end of SF along the curved path nearly the
same as in the second event. Different directions of plasma
flows are shown by the arrows in Figure~\ref{fig9}(c).
Several long threads as a whole shift from the northern side of
the NF channel to the southern side. This movement corresponds to
clockwise rotation of a right-hand helix around its axis, as seen
from the east, if the threads belong to its upper part and reveals
untwisting of the helix. The eastern part of the helix looks more
twisted, with threads more transversal to the axis. At the ending
phase of the event, there are many blobs moving along the threads
to the eastern end of NF. Their rotation (counterclockwise) is
opposite to the rotation of the whole threads in the middle part
of the helix because they presumably move along the upper part of
the right-hand helix to its eastern end.
The distance--time profiles of these plasma motions are
represented in Figure~\ref{fig11}. Figure~\ref{fig11}(a) shows the
results for plasma moving to the north-west along the NF channel.
We measured the profiles along two different trajectories shown by
white dashed lines in Figure~\ref{fig9}(d). Heated plasma moves
with average speeds of $\sim$110 and $\sim$160 $\rm km~s^{-1}$
along trajectories 1 and 2, respectively. Figure~\ref{fig11}(b)
represents the profiles of plasma motion along the SF channel
towards its western end. Since plasma moves along the curved path,
we measured two profiles along straight lines, one for the
northwestward motion and another for the southwestward motion. The
speeds are $\sim$90 and $\sim$45 $\rm km~s^{-1}$, respectively.
\section{Interpretation and Discussion}
\label{sec5}
Two filaments gradually approach each other in their middle parts
during their passage through the solar disk on 2014 April. We
specify chiralities of the filament to be opposite. The NF fine
structure definitely reveals the sinistral chirality, which is in
accordance with the general hemispheric rule for the southern
hemisphere. The chirality of SF is evident to be dextral for
many reasons despite the violation of the
hemispheric rule (see Section~\ref{sec3} and Figure~\ref{fig1}
for more details). During several episodes from April 18 to April
20, the filaments show an activation and formation of a temporal
structure that joins them into a united system. It looks puzzling
because usually filaments with parallel axial magnetic fields and opposite chirality do not merge or
reconnect with the formation of new stable or two different filaments
from their halves.
In our interpretation of these observations we follow the
flux-rope model \citep[e.g.,][]{Canou10,Guo10,Joshi14,Filippov15} of filaments considering filament plasma
accumulated in lower parts of helical flux tubes. Dextral
filaments are contained within left-handed helices, while
sinistral filaments fill right-handed ones. In our case, the axial
magnetic fields of the two flux ropes are parallel but the
azimuthal fields have different sense of rotation. Therefore, when
these two flux ropes come close together side-by-side, both their
axial and azimuthal field components have the same directions and
cannot reconnect.
Two flux ropes of similar helicity either show
merging or slingshot reconnection during their interaction
\citep{Linton01,Torok11a}. However, two flux ropes of different
helicity with parallel axial magnetic fields can only undergo a bounce interaction \citep{Linton01}
when they are brought together. They repulse from each other and
cannot reconnect and form the joint structure. In our case, it is
clear that SF/NF have dextral/sinistral chiralities, respectively with parallel axial magnetic fields. Therefore,
the associated flux ropes should have different signs of twist,
which is the condition for the bounce interaction.
We believe that although the events look like interaction of two
filament channels, the most important interaction occurs between a
flux rope and the surrounding coronal magnetic field of special
structure. We clearly observe a fan-spine configuration of coronal
loops over the filament channels, with a presumed null point above
them (see Figure~\ref{fig2}). The PFSS magnetic field extrapolation
confirms the existence of the fan-spine magnetic configuration (see Figure~\ref{fig3}).
The initial magnetic field-line distribution and subsequent plasma
dynamics are shown in the schematic representation in
Figure~\ref{fig12}.
The coronal structure is similar to a
"pseudostreamer" \citep{Wang07,Rachmeler14} with two flux ropes at
the base. The left column represents the 3D disk view (Figure~\ref{fig12}(a)), while the right column shows the projected view of selected 3D field lines (Figure~\ref{fig12}(g)).
However, in contrast to the "pseudostreamer" the outer
spine field line emanating from the null point is not directed
radially into the outer corona, but deviates to the west and
touches the photosphere near to the endpoint of SF within an area
of positive polarity (see Figures~\ref{fig2} and~\ref{fig3}).
In projection on the disk, the outer spine
field line runs nearly parallel to the SF axis, so the plasma motion
along the spine can easily be mixed up with the motion along the
SF axis. We believe it is most probable in the observations of the
filament interaction in our case.
In the first event the case is looking simple, i.e., reconnection between the inner green and the outer blue line (Figures~\ref{fig12}(b),~\ref{fig12}(h) and~\ref{fig5}). However in the second and third case the senario is little complex.
There is no anti-parallel field lines belonging to flux ropes that
contain SF and NF. But if each of the flux ropes approaches the
null point, its azimuthal field can reconnect with outer field
lines of the opposite lobe, i.e., circular
field lines of the red flux rope can reconnect with the outer
green line. This case seems to take place in the second (Figures~\ref{fig12}(c),~\ref{fig12}(i), and~\ref{fig7}) and third (Figures~\ref{fig12}(e),~\ref{fig12}(k), and~\ref{fig9}) events of the filament interaction. The
locations of reconnections are shown by pink stars in
Figure~\ref{fig12}. Due to the projected view of 3D field lines in a 2D plan, the reconnected field lines are appear as a single line in panels (j) and (l) of Figure~\ref{fig12}. But actually it represents the two different sets to field lines as shown in 3D view (Figures~\ref{fig12}(d) and~\ref{fig12}(f)).
After the reconnection some amount of heated plasma confined
previously within the flux rope is able to propagate along the
field lines of the surrounding magnetic configuration. In
particular, it can move along the spine and this motion mimics
the movement along the SF axis. Penetration of the flux-rope
plasma into the outer structure can be illustrated by a simple 2-D
model.
Let us consider the coronal magnetic field with fan-spine
structure as a sum of a vertical homogeneous magnetic field $B_0$
and a vertical 2-D dipole located at $x = 0$, $z = z_d$ with the
dipole moment $M$. If $y$ is the axis of translational symmetry,
$x$ is the horizontal axis and $z$ is the vertical axis with the
origin at the photospheric level, the external field is described
by $y$-component of vector potential $\bf A$
\begin{equation}
A_y^e = B_0 x + \frac{M x}{x^2 + (z - z_d)^2} .
\end{equation}
We put into this field a flux rope in the simplest form of a
straight linear current along the $y$-axis. According to the
boundary condition for the coronal current $I$ on the photosphere,
its vector potential can be written as
\citep{VanTend78,Molodenskii87,Filippov01}
\begin{equation}
A_y^I = \frac{I}{c}[\ln \left((x - x_0)^2 + (z + z_0)^2 \right) -
\ln \left((x - x_0)^2 + (z - z_0)^2 \right)] ,
\end{equation}
where $x_0$ and $z_0$ are the coordinates of the coronal current.
Neglecting the weight of the flux rope its equilibrium position
$(x_0, z_0)$ is defined by the equations:
\begin{equation}
B_0 - M \frac{ x_0^2 - (z_0 - z_d)^2}{\left(x_0^2 +
(z_0 - z_d)^2\right)^2} = 0,
\end{equation}
\begin{equation}
\frac{I}{c z_0} - M \frac{2 x_0 (z_0 - z_d)}{\left(x_0^2 +
(z_0 - z_d)^2\right)^2} = 0.
\end{equation}
Figure~\ref{fig13}(left) shows field lines described by Equations
(1) and (2) for dimensionless parameters $B_0 = 1, M = -4, z_d =
-1$, and the value of $I/c$ being close to the critical value
of the current $I_c$ over which the stable equilibrium is
impossible \citep{Molodenskii87,Filippov01}. Spaces between
several magnetic surfaces $A_y$ = const are shadowed by different
tints. Figure~\ref{fig13}(right) shows the same magnetic surfaces
for the slightly increased value of $I_c$. The magnetic flux
$\Phi$ conservation between the current and the photosphere is
taken into account in the form
\begin{equation}
\Phi = M x_0 \left(\frac{1}{ x_0^2 +
(z_0 - z_d)^2 } -\frac{1}{ x_0^2 +
z_d^2 } \right) + \frac{I}{c} \ln \frac{2 z_0}{r_0} = const,
\end{equation}
where $r_0 = 0.01$ is the radius of a flux tube with nearly
homogeneous current density, which should be taken into account to
avoid divergency.
When the equilibrium position becomes higher, some of previously
closed field lines reconnect with open field lines at the null
point. Plasma (possibly previously heated) confined between some
closed magnetic surfaces is able to propagate along the open
field lines into the upper corona and to the photosphere. Such
scenario we expect to happen in the present case of filament
interactions during 2014 April 18 -- 20.
The first event and reconnection on April 18 was initiated by the activation and
partial eruption of SF that bring the inner green field lines
toward the null point and trigger the reconnection (Figures~\ref{fig12}(b) and~\ref{fig12}(h)). The
observed brightenings (shown by the chartreuse color) near the
foot points of the fan lines on both sides of the filaments is
strong evidence of the null point reconnection (Figure~\ref{fig5}(c)). The accelerated
electrons move after the reconnection towards
the foot points of the fan lines and produce brightenings there.
In the second event on April 19, the activation and partial eruption of NF
bring its field lines to the magnetic null and trigger
reconnection (Figures~\ref{fig12}(c) and~\ref{fig12}(i)). The brightening at the
foot points of the fan lines was also observed in this event. Some
part of the heated NF plasma travels along the outer spine over
the SF towards the western foot point
(Figures~\ref{fig12}(d) and~\ref{fig12}(j)).
The third event on April 20 is quite
similar to the second one with a partial eruption of NF again, but there is additional flow of heated
NF plasma along the axis of NF towards the western end
(along the white arrows in Figure~\ref{fig9}(c)).
Apparent brightening near the null
point has also been observed just after the partial eruption in both
the second and third cases, which provides a signature of magnetic reconnection
(see Figures~\ref{fig8} and~\ref{fig10}).
In \textit{SDO}/AIA images the filaments and filament channels,
arcades and loops look the same after second event.
However, we believe that there should be some changes in the magnetic configuration
as shown in panel(j) of Figure~\ref{fig12}. The flux rope with helical field lines (black color)
and the outer green spine lines reconnect (Figures~\ref{fig12}(c) and~\ref{fig12}(d)) and create a new domain between a few
new reconnected lines joining the two systems (shown by dotted black lines in
Figures~\ref{fig12}(d) and ~\ref{fig12}(j). We believe that a similar reconnection is occurring
during the third event (Figure~\ref{fig12}(e) and~\ref{fig12}(k)).
The activation of the filaments can be understood by some photospheric
magnetic changes. Looking at the magnetic field evolution in this region,
we note flux cancellation close to the eastern end of SF, emerging flux close
to the middle and the western end of SF along with the expending motions
(see the \textit{SDO}/HMI animation attached with Figure~\ref{fig1}).
These motions could create shear and cancelling flux and be the trigger of the
activation of the filaments. The SF remains stable after the first event. From
this we could guess that the twist of SF flux rope became less and the overlying
arcades are more potential. On the other hand the NF being overlaid by less
and more sheared arcades can rise and reach the null point.
\section{Conclusions}
\label{sec6}
In this paper, we discuss the observations of the interactions
between two near-by filament channels of different chirality with parallel axial magnetic fields. We
found a key role of the interaction of partially erupting
filaments and associated flux ropes with the overlying fan-spine
magnetic structure. On the basis of the analysis of coronal EUV
images and magnetic field calculations, we come the following main results.
\begin{enumerate}
\item \textit{SDO}/AIA EUV observations show close connection between two
near-by filament channels during three episodes of activation and
interaction of the filaments with different chiralities on 18, 19, and 20 April 2014, respectively.
\item The observations as well as the PFSS calculations clearly show
the existence of fan-spine magnetic configuration over the two
filaments.
\item Although the events look like interaction of two
filament channels, the most important interaction occurs between
the flux ropes containing the filaments and the surrounding
coronal magnetic field of special structure. The activations and
partial eruptions of the filaments are believed to be responsible
for the reconnection between their magnetic field lines with the coronal
field at the magnetic null that lies above them.
\end{enumerate}
The most interesting aspect in these events is the geometry of
the spine line of the coronal fan-spine structure. In projection
on the solar disk, the spine line runs nearly parallel to the SF
axis, close to it. The remote point, where the outer spine line is
anchored in the photosphere, is located near the western end of
SF. It leads to the wrong impression that plasma from one filament
channel easily penetrates to the other filament channel with the
opposite chirality. In fact, the filament plasma penetrates into
the coronal structure and propagates along coronal field lines,
which are located above the SF axis. We strongly believe that the
activations and partial eruptions of both filaments were
responsible for the reconnection at the magnetic null. Plasma of
the filaments then moves along the spine line to the remote
footpoint, which lies near the western end of SF.
Interaction of two nearby filament channels with different filament chiralities within large-scale
coronal fan-spine magnetic structure have not been observed
before, although fan-spine configurations have been discussed in
case of solar eruption in "pseudostreamers" \citep{Wang07,Torok11b,Rachmeler14,Yang15},
jets \citep{Pariat09,Filippov09,Filippov15} and flares \citep{Wang14,Joshi15}.
Observations of filament interactions and merging are crucial for
better understanding of reconnection between large-scale coronal
flux ropes. It also provide the information of interaction the flux
ropes with the overlying magnetic configurations.
These magnetic structures are responsible for
different types of eruptions. High-resolution observations provide
important inputs for the MHD modeling of flux-rope interactions
and reconnections, which are needed to understand the physics of
flux-rope dynamics more clearly.
\acknowledgments The authors thank the referee for his/her valuable comments/suggestions.
We thank SDO/AIA, SDO/HMI, BBSO and GONG/NSO
teams for providing their data for the
present study. This work is supported by the BK21 plus program
through the National Research Foundation (NRF) funded by the
Ministry of Education of Korea. NCJ thank School of Space
Research, Kyung Hee University for providing Postdoctoral grant. We are thankful to Dr. Pascal D{\'e}moulin for his valuable suggestions.
|
\section{Introduction}
Let $X\subset\CC\PP^{\nu}$ be a $n$-dimensional projective variety. The embedding $X \hookrightarrow \mathbb{CP}^{\nu}$ is called balanced if $$\int_{X} \big(\frac{z_{i}\bar{z}_{j}}{|z|^{2}}-\frac{\delta_{ij}}{\nu+1}\big)\omega_{FS}^{n}=0, $$ where $z_{0}, \cdots, z_{\nu}$ are the homogenous coordinates of $\mathbb{CP}^{\nu}$, $|z|^{2}=\sum|z_{i}|^{2}$, and $\omega_{FS}$ is the Fubini-Study metric. In \cite{ZS}, S. Zhang proved that $X$, as an algebraic cycle in $\CC\PP^\nu$, is Chow polystable if and only if the embedding $X\subset \CC\PP^{\nu}$ can be translated to a balanced one via an element $u\in \mathrm{SL}(\nu+1)$.
A theorem due to Donaldson \cite{Dos} shows the connection of the balanced embedding and the existence of K\"{a}hler metric with constant scalar curvature. More precisely, let $(X,L)$ be a polarized manifold of dimension $n$ such that the automorphism group ${\rm Aut}(X,L)$ is finite. If there is a K\"{a}hler metric $\omega$ with constant scalar curvature representing $c_{1}(L)$, then Donaldson's theorem asserts that for $k\gg 1$, $L^{k}$ induces a balanced embedding $\Phi_{k}: X\hookrightarrow \mathbb{CP}^{\nu_{k}}$ with $\Phi_{k}^{*}\mathcal{O}_{\mathbb{CP}^{\nu_{k}}}(1)=L^{k}$. Furthermore, $$\|\omega-k^{-1}\Phi_{k}^{*}\omega_{FS}\|_{C^{r}(X)}\rightarrow 0,$$ when $k \rightarrow\infty$, in the $C^{r}$-sense for any $r>0$. The K\"{a}hler metric $k^{-1}\Phi_{k}^{*}\omega_{FS}$ is called a {\em balanced matric.}
If $(X,L)$ is a polarized Calabi-Yau manifold, Yau's theorem on the Calabi conjecture says that there exists a unique Ricci-flat K\"{a}hler-Einstein metric $\omega$ with $\omega\in c_{1}(L)$, i.e. the Ricci curvature ${\rm Ric}(\omega)\equiv 0$ (cf. \cite{Yau1}). By Donaldson's theorem, $L^{k}$ induces a balanced embedding $X\hookrightarrow \mathbb{CP}^{\nu_{k}}$ for $k\gg 1$, and the Ricci-flat K\"{a}hler-Einstein metric $\omega$ can be approximated by the balanced metrics. If $(X,L)$ is a principally polarized Abelian variety, it is proven first in \cite{No} (and also independently \cite{WY}) that the standard embedding induced by the classical theta functions of level $k$ is balanced. Moreover, in this case the convergence of balanced metrics to the flat metric can be verified via a complete elementary way without quoting \cite{Dos}.
In \cite{SYZ}, Strominger, Yau and Zaslow propose a geometric way of constructing mirror Calabi-Yau manifolds via dual special lagrangian fibration, which is the celebrated SYZ conjecture. Later, a new version of the SYZ conjecture is proposed by Gross, Wilson, Kontsevich, and Soibelman,
(cf. \cite{GW,KS,KS2}) by using the collapsing of Ricci-flat K\"{a}hler-Einstein metrics.
Let $(\mathcal{X} \rightarrow \Delta, \mathcal{L}) $ be a maximal unipotent degeneration of polarized Calabi-Yau $n$-manifolds, i.e. the relative canonical bundle $\mathcal{K}_{\mathcal{X} / \Delta} $ is trivial, such that $0\in \Delta$ is a large complex limit point, and $\omega_{t}$ be the Ricci-flat K\"{a}hler-Einstein metric satisfying $\omega_{t}\in c_{1}(\mathcal{L}|_{X_{t}}) $ for $t\in \Delta^{\circ}$.
The collapsing version of SYZ conjecture asserts that
$$(X_{t}, {\rm diam}_{\omega_{t}}^{-2}(X_{t}) \omega_{t})\rightarrow (B,d_{B})$$ in the Gromov-Hausdorff sense, when $t\rightarrow 0$, where $(B,d_{B})$ is a compact metric space.
Furthermore, there is an open dense subset $B_{0}$ of $ B$, which is smooth, and
is of real dimension $n$, and admits a real affine structure. The metric $d_{B}$ is induced by a Monge-Amp\`ere metric $g_{B}$ on $B_{0}$, i.e. under affine coordinates $y_{1}, \cdots, y_{n}$, there is a potential function $\varphi$ such that $$g_{B}= \sum_{ij} \frac{ \partial^{2} \varphi}{ \partial y_{i} \partial y_{j}} dy_{i} dy_{j}, \ \ {\rm and} \ \ \det \Big(\frac{\partial^{2}\varphi}{ \partial y_{i} \partial y_{j} }\Big ) ={\rm const.}.$$ Clearly it is true for Abelian varieties.
This conjecture is verified by Gross and Wilson for fibred K3 surfaces with only type $I_{1}$ singular fibers in \cite{GW}, and is studied for higher dimensional HyperK\"ahler manifolds in \cite{GTZ, GTZ2}.
Bernd Siebert raises a question to relate the balanced embeddings of $X_{t}$ to the metric limit of rescaled Ricci-flat K\"{a}hler-Einstein metrics $\varepsilon_{t}\omega_{t}$ for a certain family of constants $\varepsilon_{t}$. In the Gross-Siebert program (cf. \cite{Gro,GS1}), theta functions are constructed on certain degenerations of polarized Calabi-Yau manifolds $(\mathcal{X}\rightarrow {\rm Spec}\mathbb{C}[[t]],\mathcal{L}) $ as the canonical basis of the space of sections for $\mathcal{L}$ (cf. \cite{GHKS,GS,GHKS2}), which is predicted by the Homological Mirror Symmetry conjecture. In particular, these theta functions recover the classical theta functions in the case of principally polarized Abelian varieties. If $(\mathcal{X},\mathcal{L})$ is an analytic family (cf. \cite{RS}), Siebert asks whether the theta functions give the balanced embeddings of polarized Calabi-Yau manifolds, and furthermore whether there is a family version of the Donaldson's theorem for the degeneration of Calabi-Yau manifolds near large complex limits.
In this note, we study this question in the case of principally polarized Abelian varieties, and establish the connection between the limit metric $g_{B}$ and the balanced embeddings.
In Section 2, let us recall the basic setup for a maximal unipotent family of principally polarized Abelian varieties over the punctured disc, and then state our main result. Theorem \ref{thm01} says that after a certain base change, one can find a filling-in to complete the family of Abelian varieties to a degeneration, such that the whole degeneration can be simultaneously balanced embedded in a projective space over a disc by the {\em canonical} theta functions constructed via Gross-Siebert program. Theorem \ref{thm02} studies the relationship between the balanced filling-in and the Gromov-Hausdorff limit of flat K\"{a}hler metrics on the nearby fibers. In Section 3, we review the construction of theta functions on degenerations of Abelian varieties in the Gross-Siebert program. Finally, Theorem \ref{thm01} and Theorem \ref{thm02} are proved in Section 4.
\noindent {\bf Acknowledgements:} The authors would like to thank Helge Ruddat, Mark Gross and Bernd Siebert for many helpful discussions. The work was done when both authors are in residence at the Mathematical Sciences Research Institute in Berkeley, California, during the 2016 Spring semester.
\section{Set up and Main Theorems}
In this paper, we always denote
$M\cong \mathbb{Z}^{n}$, $M_{\mathbb{R}}=M\otimes_{\mathbb{Z}}\mathbb{R}$, $N=\mathrm{Hom}_{ \mathbb{Z}}(M, \mathbb{Z})$, $N_{\mathbb{R}}=N\otimes_{\mathbb{Z}}\mathbb{R}$, $T_{\mathbb{C}}=N\otimes_{\mathbb{Z}}\mathbb{C}^{*}$, and $\langle\cdot,\cdot\rangle$ the pairing between $N_{\mathbb{R}}$ and $M_{\mathbb{R}}$.
\subsection{A family of Abelian varieties}\label{ppav}
This subsection gives the basic setup of this paper, which is a family of Abelian varieties over the punctured disc approaching to a large complex limit.
Let $Z(\cdot, \cdot): M_{\mathbb{R}} \times M_{\mathbb{R}} \rightarrow \mathbb{R}$ be a
positive definite bilinear form satisfying $Z(M,M)\subset \mathbb{Z}$. If we define the quadratic function
\begin{equation}\label{z-phi}
\overline{\varphi}(y)=\frac{1}{2}Z(y,y)
\end{equation}
on $M_{\mathbb{R}}$, and the affine linear function
\begin{equation}\label{e0.0}\alpha_{\gamma}(\cdot)= Z(\gamma, \cdot)+\frac{1}{2} Z(\gamma, \gamma),
\end{equation}
for any $\gamma \in M$, then \begin{equation}\label{e0.1} \overline{\varphi}(y+\gamma)= \overline{\varphi} (y) + \alpha_{\gamma} (y).\end{equation}
The couple $\{M, Z\}$ determines a family of principally polarized Abelian varieties over the punctured disc $(\pi: \mathcal{X}_{\eta}\rightarrow \Delta^{\circ}, \mathcal{L}_{\eta})$ as the following.
We define an $M$-action on $M\times \mathbb{Z}$ via $(m,r) \mapsto (m,r+ Z(\gamma,m))$ for any $\gamma\in M$, which induces an $M$-action on $ T_{\mathbb{C}} $ by
\begin{equation}\label{e0.100}\mathcal{Z}^{m} \mapsto \mathcal{Z}^{m} s^{Z(\gamma, m)}, \ \ \ \gamma\in M, \end{equation} for any $s\in \Delta^{\circ}$. More explicitly,
let
$$e_{1}, \cdots, e_{n}\in M$$
be a basis of $M$, we have coordinates $z_{1}= \mathcal{Z}^{e_{1}}, \cdots, z_{n}=\mathcal{Z}^{e_{n}}$ on $T_{\mathbb{C}}\cong (\mathbb{C}^\times)^{n}$, and the $M$-action is that $ (z_{1}, \cdots, z_{n}) \mapsto (z_{1} s^{Z(\gamma, e_{1})}, \cdots, z_{n} s^{Z(\gamma, e_{n})}) $ for any $\gamma\in M$. We claim that the quotient $X_{s}=T_{\mathbb{C}}/M$ is a principally polarized Abelian variety with period matrix $$\left[ I, \frac{\log s}{2\pi \sqrt{-1}}Z_{ij} \right]$$ where $Z_{ij}=Z(e_{i},e_{j}) \in \mathbb{Z} $.
Denote $e_{1}^{*}, \cdots, e_{n}^{*}$ the dual basis of $N$, and $N_{\mathbb{C}}=N\times_{\mathbb{Z}}\mathbb{C} \cong \mathbb{C}^{n}$. We have a natural embedding $N\hookrightarrow N_{\mathbb{C}} $ as the real part, and by abusing notions, we regard $e_{1}^{*}, \cdots, e_{n}^{*}$ as a $\mathbb{C}$-basis of $N_{\mathbb{C}}$.
The universal covering $N_{\mathbb{C}} \rightarrow T_{\mathbb{C}} $ is given by $ w_{i} \mapsto z_{i}= \exp 2\pi \sqrt{-1} w_{i} $, $i=1, \cdots, n$, where $w_{1}, \cdots, w_{n}$ are coordinates of $N_{\mathbb{C}}$ respecting to $e_{1}^{*}, \cdots, e_{n}^{*}$. Then
$T_{\mathbb{C}}\cong N_{\mathbb{C}}/N$ where $N$ acts on $N_{\mathbb{C}}$ given by $ w_{i} \mapsto w_{i} + \langle \mu, e_{i}\rangle= w_{i}+ \mu_{i}$ for any $\mu=\sum\limits_{i}\mu_{i}e_{i}^{*}\in N$. We have an $N \times M$-action on $\mathbb{C}^{n}$ by $$w_{i} \mapsto w_{i} +\langle \mu, e_{i}\rangle + \frac{\log s}{2\pi \sqrt{-1}}Z(\gamma, e_{i}), $$ for any $(\mu, \gamma)\in N\times M$, and we obtain $X_{s}=T_{\mathbb{C}}/M=\mathbb{C}^{n}/\Lambda_{s}$, where the lattice $\Lambda_{s}={\rm span}_{\mathbb{Z}} \{e_{1}^{*}, \cdots, e_{n}^{*}, \frac{\log s}{2\pi \sqrt{-1}}Z( e_{1}, \cdot) , \cdots, \frac{\log s}{2\pi \sqrt{-1}}Z( e_{n}, \cdot)\}$. Furthermore, we construct a family of Abelian varieties $\mathcal{X}_{\eta}=(T_{\mathbb{C}} \times \Delta^{\circ} )/M\rightarrow \Delta^{\circ}$ over the punctured disc $\Delta^{\circ}$ with fiber $X_{s}$.
We extend the $N \times M$-action on $\mathbb{C}^{n}$ to $\mathbb{C}^{n}\times \mathbb{C}$ by $$(w, \lambda) \mapsto (w + \mu + \frac{\log s}{2\pi \sqrt{-1}}Z(\gamma, \cdot), \lambda \exp \pi\sqrt{-1}(- \frac{\log s}{2\pi \sqrt{-1}}Z(\gamma,\gamma)-2 \langle w, \gamma \rangle)), $$ for any $(\mu, \gamma)\in N\times M$, where $w=(w_{1}, \cdots, w_{n})\in \mathbb{C}^{n}$. The quotient $(\mathbb{C}^{n}\times \mathbb{C})/(N \times M)$ is the relative ample bundle $\mathcal{L}_{\eta}$. The classical Riemann theta function (cf. \cite{BL})
$$
\vartheta=\sum_{\gamma\in M}\exp \pi \sqrt{-1}\left( \frac{\log s}{2\pi \sqrt{-1}}Z(\gamma, \gamma)+ 2 \langle w,\gamma \rangle\right)$$
is the distinguished section of $\mathcal{L}_{\eta}$.
On any $X_{s}$, the first Chern class $c_{1}(\mathcal{L}_{\eta}|_{X_{s}})$ is represented by the flat K\"{a}hler metric \begin{equation}\label{e0.10}\omega_{s} =\frac{-\pi \sqrt{-1}}{ \log |s|}\sum_{ij} Z ^{ij} dw_{i}\wedge d\bar w_{j}, \end{equation} where $ Z ^{ij}$ denotes the inverse of the matrix $Z_{ij}=Z(e_{i},e_{j}) $.
Let $x_{1}, \cdots, x_{n}, y_{1}, \cdots, y_{n}$ denote the coordinates on $N_{\mathbb{C}}$ with respect to the basis $e_{i}^{*}$, $1\leqslant i\leqslant n$, $ ( \frac{\log |s|}{2\pi \sqrt{-1}}+ \arg (s))Z( e_{j}, \cdot) $, $1\leqslant j\leqslant n$, where $0\leqslant \arg(s) <2\pi$. We also regard $ y_{1}, \cdots, y_{n}$ (resp. $x_{1}, \cdots, x_{n}$) as coordinates on $M_{\mathbb{R}} $ (resp. $N_{\mathbb{R}} $) respecting to $e_{1}, \cdots, e_{n}$.
Then
$$w_{i}=x_{i}+(\arg (s)-\sqrt{-1}\frac{\log |s|}{2\pi } )\sum_{j=1}^nZ_{ij}y_{j},\ i=1, \cdots, n,$$ and
as a symplectic form, $$\omega_{s}=\sum\limits_{ij}dx_{i} \wedge dy_{j}. $$ The corresponding Riemannian metric is $$g_{s}=\frac{-2\pi}{\log |s|}\sum_{i,j} Z^{ij}dx_{i}dx_{j}-\frac{\log |s|}{2\pi} \sum_{i,j} Z_{ij}dy_{i}dy_{j}, $$ and, when $s \rightarrow 0$, $$(X_{s}, \frac{-2\pi}{\log |s|}g_{s}) \longrightarrow (B, g_{B}= \sum_{ij} Z_{ij}dy_{i}dy_{j})$$ in the Gromov-Hausdorff topology, where $B=M_{\mathbb{R}}/M$.
We can regard $\overline{\varphi}$ as a multivalue function on $B=M_{\mathbb{R}}/M $ by (\ref{e0.1}), and on any small open subset on $B$, the difference of any two sheets of $\overline{\varphi}$ is a linear function defined in (\ref{e0.0}). Note that the Hessian matrix of $\overline{\varphi}$ is well-defined, and $\frac{\partial^{2} \overline{\varphi}}{\partial y_{i} \partial y_{j}}=Z_{ij}$.
The Riemannian metric $g_{B}$ is the Monge-Amp\`{e}re metric with potential $\overline{\varphi}$ respecting to the affine coordinates $y_{1}, \cdots, y_{n}$, i.e. \begin{equation}\label{e0.11}g_{B}= \sum_{ij} \frac{\partial^{2} \overline{\varphi}}{\partial y_{i} \partial y_{j}}dy_{i}dy_{j}, \ \ {\rm and} \ \ \det ( \frac{\partial^{2} \overline{\varphi}}{\partial y_{i} \partial y_{j}}) \equiv \det(Z_{ij}). \end{equation}
\subsection{Main results}
For any $k\gg 1$, let $M_{k}=kM$, and $\varphi$ be an $M_{k}$-periodic convex piecewise linear function such that the slopes of $\varphi$ are in $N$, and \begin{equation}\label{e0.2}\varphi(y+\gamma)= \varphi (y) + \alpha_{\gamma} (y) \end{equation} for any $\gamma\in kM$, which induces an $M_{k}$-invariant rational polyhedral decomposition $\mathcal{P} $ of $M_{\mathbb{R}}$, i.e.
$\sigma $ is a cell of $\mathcal{P} $ if and only if $\varphi$ is linear on $\sigma$.
The Mumford's construction (cf. \cite{Mum}) gives a degeneration of principally polarized Abelian varieties $(\pi: \mathcal{X}\rightarrow \Delta, \mathcal{L})$ from the data $(M_{k},\mathcal{P}, \varphi) $ such that $ \pi: \mathcal{X}_{\Delta^{\circ}}\rightarrow \Delta^{\circ}$ is the base change of $ \mathcal{X}_{\eta}$ via $t\mapsto t^{k}=s$, where $ \mathcal{X}_{\Delta^{\circ}}=\pi^{-1}(\Delta^{\circ})$, and $\mathcal{L}|_{\mathcal{X}_{\Delta^{\circ}}}$ is the pull-back of $\mathcal{L}_{\eta}^{k}$. The central fiber $X_{0}=\pi^{-1}(0)$ is reduced and reducible with only toric singularities. The intersection complex of $X_{0}$ is $(B_{k}, \tilde{\mathcal{P}})$ where $B_{k}=M_{\mathbb{R}}/M_{k}$ and $\tilde{\mathcal{P}}$ is the quotient rational polyhedron decomposition of $\mathcal{P}$. The irreducible components are one to one corresponding to $n$-cells in $\tilde{\mathcal{P}}$, and for any $n$-cells $\sigma\in \tilde{\mathcal{P}}$, the respective irreducible component is the toric variety $X_{\sigma}$ defined by $\sigma$, where we regard $\sigma$ as a polytope in $M_{\mathbb{R}}$. Furthermore, the restriction of $\mathcal{L}$ on $X_{\sigma}$ is the toric ample line bundle defined by $\sigma$.
For any $m\in B_{k}(\mathbb{Z})=M/M_{k}$, a section $\vartheta_{m}$ of $\mathcal{L}$ is constructed in Section 6 of \cite{GHKS} such that the restriction of $ t^{\varphi(m)-\bar{\varphi} (m)}\vartheta_{m}$ on any $X_{t}=\pi^{-1}(t)$, $t\neq0$, is a classical Riemann theta function (See also \cite{GS}). And the restriction of $\vartheta_{m}$ on any component $X_{\sigma}$ of $X_{0}$ is a monomial section of $\mathcal{L}|_{X_{\sigma}} $.
Note that the choice of $(\mathcal{P}, \varphi)$ is not unique, and different choices give different filling-ins $X_{0}$. However there is a canonical one studied in \cite{AN}, which satisfies $\varphi (m)=\bar{\varphi} (m)$ for any $m\in M$. More precisely, let $\varphi: M_{\mathbb{R}} \rightarrow \mathbb{R}$ be the convex piecewise linear function such that the graph of $\varphi$ is the lower bound of the convex hull of $ \{(m, \overline{\varphi}(m))| m\in M \}\subset M_{\mathbb{R}}\times \mathbb{R}$, and let $\mathcal{P}$ be the rational polyhedral decomposition of $ M_{\mathbb{R}}$ induced by $\varphi$, i.e. a cell $\sigma\in\mathcal{P} $ if and only if $\varphi$ is linear on $\sigma$. It is clear that (\ref{e0.2}) is satisfied, and \begin{equation}\label{enew01} \varphi(m)= \overline{\varphi}(m), \ \ \ \ {\rm for \ \ any} \ m\in M. \end{equation}
Let $\overline{\mathcal{P}}$ be the boundary of the convex hull of the lattice points on the graph of $\overline{\varphi}$, then the polyhedral decomposition $\mathcal{P}$ is obtained in such a way that each cell of $\mathcal{P}$ is precisely the projection of a face of $\overline{\mathcal{P}}$ onto $M_\RR$.
The decomposition $\mathcal{P}$ is the {\em mostly divided} polyhedral decomposition that one can have. Any cell $\sigma$ of $\mathcal{P}$ intersects with the lattice $M$ only at its vertices, i.e. there is no integral point in the interior of $\sigma$.
We further assume that for any $\sigma\in\mathcal{P}$, the slop of $\varphi|_{\sigma} $ is integral, i.e.
\begin{equation}\label{e0.3}
d\varphi|_{\sigma} \in N \end{equation}
for simplicity. It is not a further restriction, and the reason is as the following. Note that $d\varphi|_{\sigma} \in N\otimes_{\mathbb{Z}}\mathbb{Q}$. For any $m\in M$, there is a $\nu \in \mathbb{N}$ such that $\nu d\varphi|_{\sigma} \in N$ for any cell $\sigma\in\mathcal{P}$ with $m\in \sigma$, i.e. $\sigma$ belongs to the star of $m$. By the $M$-action and (\ref{e0.2}), $\nu d\varphi|_{\sigma}+ \nu d\alpha_{\gamma} \in N$ for any $\gamma\in M$, and thus $\nu \varphi$ satisfies (\ref{e0.3}).
If we replace $Z $ by $\nu Z $, then
the new family of Abelian varieties constructed from $\nu\overline{\varphi}$ is the base change of the original family by $s \mapsto s^{\nu}$.
Hence we assume (\ref{e0.3}).
Before we present the main theorems, we look at some lower dimensional cases.
If $\dim_{\mathbb{R}}M_{\mathbb{R}}=1$, $B_{k}$ is a cycle, $1$-cells in $\mathcal{P} $ are intervals, and $X_{0}$ is the Kodaira type I$_{k}$ fiber, i.e. a cycle of $k$ rational curves. It is well known to experts that
$X_{0}$ can be balanced embedded.
When $\dim_{\mathbb{R}}M_{\mathbb{R}}=2$, there are only two possible choices of $\mathcal{P} $ by \cite{AN}. One is that any $2$-cell in $\mathcal{P} $ is a standard simplex, and the other one is that $\mathcal{P} $ consists of cubes. Hence $X_{0}$ consists either finite many $\mathbb{CP}^{2}$ or finite many $\mathbb{CP}^{1}\times \mathbb{CP}^{1}$.
\begin{center}
\setlength{\unitlength}{0.5cm}
\begin{picture}(3.3,2.6)(-2.5,-0.25)
\linethickness{.075mm}
\put(-2,0.5){\line(0,1){1}}
\put(-2,1.5){\line(1,1){1}}
\put(-2,0.5){\line(1,0){1}}
\put(-1,0.5){\line(1,1){1}}
\put(-1,2.5){\line(1,0){1}}
\put(0,1.5){\line(0,1){1}}
\end{picture}
\setlength{\unitlength}{0.5cm}
\begin{picture}(4.3,3.6)(-2.5,-0.25)
\linethickness{.075mm}
\multiput(-2,0)(1,0){4}
{\line(0,1){3}}
\multiput(-2,0)(0,1){4}
{\line(1,0){3}}
\end{picture}
\setlength{\unitlength}{0.5cm}
\begin{picture}(4.3,3.6)(-2.5,-0.25)
\linethickness{.075mm}
\multiput(-2,0)(1,0){4}
{\line(0,1){3}}
\multiput(-2,0)(0,1){4}
{\line(1,0){3}}
\put(0,0){\line(1,1){1}}
\put(-1,0){\line(1,1){2}}
\put(-2,0){\line(1,1){3}}
\put(-2,1){\line(1,1){2}}
\put(-2,2){\line(1,1){1}}
\end{picture}\end{center}
The first result of this paper shows that the embedding of the canonical degeneration $\mathcal{X}$ of \cite{AN} by theta functions constructed in \cite{GHKS} is balanced.
\begin{theorem}\label{thm01} For any $k\gg 1$, $B_{k}=M_{\mathbb{R}}/M_{k}$, $B_{k}(\mathbb{Z})=M/M_{k}$, and let $(\pi: \mathcal{X}\rightarrow \Delta, \mathcal{L})$ be the degeneration of principally polarized Abelian varieties from the triple $(M_{k},\mathcal{P}, \varphi) $. The theta functions $\{\vartheta_{m}| m\in B_{k}(\mathbb{Z})\}$ define a relative balanced embedding $$ \Phi_{ k} =[\vartheta_{m_{1}}, \cdots, \vartheta_{m_{k^{n}}}]: \mathcal{X}\rightarrow \mathbb{CP}^{k^{n}-1}\times \Delta,$$ i.e. for any $t\in \Delta$, $\Phi_{ k}|_{X_{t}}:X_{t}\hookrightarrow \mathbb{CP}^{k^{n}-1}$ is a balanced embedding.
\end{theorem}
Again it was proven in \cite{WY} and \cite{No} that any individual Abelian variety $X_{t}$, $t\neq0$, can be balanced embedded in certain $ \mathbb{CP}^{N}$ via theta functions. If one set $t\rightarrow 0$, the limit variety $X_{0}$ in the projective space is also balanced. However Theorem \ref{thm01} uses a group of different theta functions that guarantee the balanced embedding varying holomorphicly when $t$ approaches to $0$. Furthermore, Theorem \ref{thm01} identifies the balanced limit $X_{0}$ to be the canonical filling-in of $ ( \mathcal{X}_{\Delta^{\circ}}, \mathcal{L}|_{\mathcal{X}_{\Delta^{\circ}}})$ constructed in \cite{AN}.
Now we study the connection between the metric limit $(B, g_{B})$ and the balanced filling-in when $k\rightarrow \infty$. For a fixed $k\gg 1$, $B_{k}$ has a natural affine structure induced by $M_{\mathbb{R}}$. If $f$ is a convex function on an open subset $U$ of $B_{k}$ with respect to the affine structure, where we also regard $U$ as a subset of $M_{\mathbb{R}}$ by the quotient, then for any $y_{0}\in U$, we let $$ \partial f (y_{0}) =\{\upsilon \in N_{\mathbb{R}}| f(y)\geq \langle \upsilon, y-y_{0} \rangle + f(y_{0}) \ \ {\rm for \ \ all} \ \ y\in M_{\mathbb{R}} \}, $$ and we define the Monge-Amp\`{e}re measure $${\rm MA}(f)(E)={\rm Vol}(\bigcup_{y\in E} \partial f (y)),$$ for any Borel subset $E\subset U$, where ${\rm Vol}$ denotes the standard Euclidean measure on $N_{\mathbb{R}}$ (cf. \cite{BPS} and \cite{TW}). It is well known that ${\rm MA}(f+\alpha)={\rm MA}(f)$ for any linear function $\alpha$ on $M_{\mathbb{R}}$, and if $f$ is smooth, $${\rm MA}(f)=\det \big(\frac{\partial^{2}f}{\partial y_{i} \partial y_{j}}\big)dy_{1}\wedge \cdots \wedge dy_{n}. $$
Note that we can regard the function $\varphi$ as a multivalue function on $B_{k}$, and by (\ref{e0.2}), the difference between any two sheets of $\varphi$ is a linear function. Thus we have a well-defined Monge-Amp\`{e}re measure ${\rm MA}(\varphi)$. The next theorem shows that after some rescaling, this Monge-Amp\`{e}re measure converges to the Monge-Amp\`{e}re measure of the potential function $\bar{\varphi}$ of $g_{B}$ when $k\rightarrow \infty$, and furthermore, the rescaled potential function $\varphi$ also converges to $\bar{\varphi}$ in the $C^{0}$-sense, which shows the link between the balanced embeddings and the Gromov-Hausdorff limit.
\begin{theorem}\label{thm02} For any $k \gg 1$, the Monge-Amp\`{e}re measure of $\varphi$ is $${\rm MA}(\varphi)=\det (Z_{ij}) \sum_{m\in B_{k}(\mathbb{Z})}\delta_{m}, $$ where $\delta_{m}$ is the Dirac measure at $m\in B_{k}(\mathbb{Z})$.
If $\chi_{k}: B\rightarrow B_{k}$ is induced by the dilation $y_{i} \mapsto k y_{i}$, $i=1, \cdots, n$, of $M_{\mathbb{R}}$, then $\frac{1}{k^{2}} \chi_{k}^{*}\varphi- \overline{\varphi}$ is a well-defined function on $B$, and
$$\sup_{B} \left |\frac{1}{k^{2}} \chi_{k}^{*}\varphi- \overline{\varphi}\right | \rightarrow 0, $$
when $k\rightarrow\infty$. Furthermore, $$ \frac{1}{k^{n}} \chi_{k}^{*}{\rm MA}(\varphi) \rightharpoonup {\rm MA}(\bar{\varphi})=\det(Z_{ij})dy_{1}\wedge \cdots \wedge dy_{n}$$ in the weak sense.
\end{theorem}
In \cite{Liu}, the non-archimedean Monge-Amp\`{e}re equation is solved for degenerations of Abelian varieties, where the approximation of continuous potential functions by piecewise linear functions are also used. See \cite{BFJ} for the non-archimedean Monge-Amp\`{e}re equations for more general cases. Here in Theorem \ref{thm02}, we are working on the intersection complexes instead of the dual intersection complexes as in \cite{BFJ,Liu}, and our piecewise linear functions are from the balanced embedded degenerations of Abelian varieties.
We end this section by giving a remark of Calabi-Yau manifolds balanced embedded by theta functions.
In Section 3 of \cite{DKLR}, it is shown that the Calabi-Yau hypersurfaces $$X_{t}=\{ [z_{0}, \cdots, z_{4}]\in \mathbb{CP}^{n} |t( z_{0}^{n+1}+ \cdots + z_{n}^{n+1})+ z_{0} \cdots z_{n}=0\} $$ are balanced embedded for any $t\in \mathbb{C}$. The proof involves two finite group actions on $(X_{t}, \mathcal{O}_{\mathbb{CP}^{n}}(1)|_{X_{t}})$. The first one is $Ab_{n+1}=\{(a_{0}, \cdots, a_{n})| a_{i} \in \mathbb{Z}_{n+1}, a_{0}+ \cdots+ a_{n}=0\}/ \mathbb{Z}_{n+1}$, which acts on $X_{t}$ by $$(z_{0}, \cdots, z_{n}) \mapsto (\zeta^{a_{0}}z_{0}, \cdots, \zeta^{a_{n}}z_{n})$$ where $\zeta=\exp \frac{2\pi \sqrt{-1}} {n+1}$. The second one is the symmetric group $ \mathbb{S}_{n+1}$ on $n+1$ elements, which acts on $X_{t}$ by translating $z_{0}, \cdots, z_{n}$. The same argument as in the proof of Theorem \ref{thm01} also shows the result of $X_{t}$ being balanced.
On the other hand, $z_{0}, \cdots, z_{n}$ as sections of $ \mathcal{O}_{\mathbb{CP}^{n}}(1)|_{X_{t}}$ are theta functions constructed in the Gross-Siebert program for $|t|\ll 1$ (at least for $n=3$) by Example 6.3 in \cite{GHKS}, and thus $X_{t}$ are balanced embedded by theta functions.
\section{Construction of theta functions}
We recall the construction of theta functions on degenerations of Abelian varieties, and we follow the arguments in Example 6.1 of \cite{GHKS} and Section 2 of \cite{GS} closely. See also \cite{ABC} for the elliptic curve case.
Let $(\mathcal{P}, \varphi)$ be the same as in the above section, i.e.
the graph of $\varphi$ is the lower boundary of the upper convex hull
$$\mathrm{conv} \{ (m, \overline{\varphi}(m))| m\in M \}\subset M_{\mathbb{R}}\times \mathbb{R},$$
and $\mathcal{P}$ be the rational polyhedral decomposition of $ M_{\mathbb{R}}$ induced by $\varphi$. If we define
$$\Delta_{\varphi}=\{(m,r)\in M_{\mathbb{R}}\times \mathbb{R}| r \geqslant \varphi (m)\}, $$ the standard construction for toric degenerations gives a toric variety $X_{\Delta_{\varphi}}$ with a line bundle $L_{\Delta_{\varphi}}$.
\setlength{\unitlength}{0.8cm}
\begin{center}
\begin{picture}(4.3,3.6)(-2.5,-0.25)
\put(-2,0){\line(1,0){4.4}}
\put(2.45,-.05){$M_{\mathbb{R}}$}
\put(0,0){\vector(0,1){2.2}}
\put(0,2.45){\makebox(0,0){$\mathbb{R}$}}
\qbezier(0.0,0.0)(1.0,0.0)
(2.0,2.0)
\qbezier(0.0,0.0)(-1.0,0.0)
(-2.0,2.0)
\put(0,0){\circle*{0.15}}\put(1,0){\circle*{0.15}}\put(-1,0){\circle*{0.15}}\put(2,0){\circle*{0.15}}\put(-2,0){\circle*{0.15}}
\put(0,0){\line(2,1){1}}\put(1,0.5){\line(2,3){1}}\put(0,0){\line(-2,1){1}}\put(-1,0.5){\line(-2,3){1}}
\put(2,1){$\bar{\varphi}(y)=y^{2}$} \put(1,1){$\varphi$} \put(-1,1.5){$\Delta_{\varphi}$}
\end{picture}
\end{center}
For any $l\in\mathbb{N}$,
$H^{0}(X_{\Delta_{\varphi}}, L_{\Delta_{\varphi}}^{l}) $
is generated by {\em monomial} sections
\begin{equation}\label{Z}
\{ \mathcal{Z}^{(m,r,l)}\mid \ \ \ \forall (m,r,l)\in C(\Delta_{\varphi}) \cap (M\times\mathbb{Z}\times \{l\})\}
\end{equation}
where
$$C(\Delta_{\varphi})=\overline{\{(lm',lr',l)|(m',r')\in \Delta_{\varphi}, l\in \mathbb{R}_{\geq 0}\}}\subset M_{\mathbb{R}}\times\mathbb{R}\times \mathbb{R}. $$
Note that we have a {\em canonical} regular function
$$\bar{\pi}=\mathcal{Z}^{(0,1)}=\mathcal{Z}^{(0,1,0)}: X_{\Delta_{\varphi}} \rightarrow \mathbb{C}. $$ The toric boundary is $\bar{X}_{0}=\{\mathcal{Z}^{(0,1)}=0\}$, a toric variety with infinite many irreducible components, and $\bar{X}_{t}=\{\mathcal{Z}^{(0,1)}=t\}\cong T_{\mathbb{C}}$,
for any $t
\neq 0$. We have a family of toric varieties $\bar{X}_{t}$ degenerating to a singular toric varieties $\bar{X}_{0}$.
The degeneration of principally polarized Abelian varieties $(\pi: \mathcal{X}\rightarrow \Delta, \mathcal{L})$ is constructed as the quotient of an $M_{k}$-action on $(X_{\Delta_{\varphi}}, L_{\Delta_{\varphi}})$ as the following.
\begin{lemma}\label{le00} There is an $M$-action on $(X_{\Delta_{\varphi}}, L_{\Delta_{\varphi}})$ such that the projection $\bar{\pi}$ is $M$-invariant, i.e. $\bar\pi(m\cdot)=\bar\pi(\cdot)$ for any $m\in M$; the induced $M$-action on monomial rational functions is given by
$$ \mathcal{Z}^{ (m,r)} \mapsto \mathcal{Z}^{(m, r+ Z(\gamma, m))}, \ \ \ \gamma\in M, \ \ \ (m,r)\in M\times\mathbb{Z};$$
and the induced $M$-action on $H^{0}(X_{\Delta_{\varphi}}, L_{\Delta_{\varphi}})$ is given by $$ \mathcal{Z}^{ (m,r,1)} \mapsto \mathcal{Z}^{(m+ \gamma, r+ \alpha_{\gamma}(m),1)}, \ \ \ \gamma\in M, \ \ \ (m,r)\in \Delta_{\varphi}$$ for monomial sections.
\end{lemma}
\begin{proof} If $\Sigma \subset N_{\mathbb{R}}\times \mathbb{R}$ denotes the normal fan of $\Delta_{\varphi}$, then one-dimensional rays of $\Sigma$ are one-to-one correspondence to the maximal dimensional cells in $\mathcal{P}$, and the primitive generator of a ray has the form $(-d\varphi |_{\sigma}, 1)$ for an $n$-dimensional cell $\sigma$ of $\mathcal{P}$. Then $M$ acts on $N\times \mathbb{Z} $ by $\check{T}^{0}_{\gamma}: (\mu,l)\mapsto (\mu- ld\alpha_{\gamma}, l) $ for any $\gamma \in M $, which preserves $\Sigma$, and thus induces an $M $-action on $X_{\Delta_{\varphi}}$.
The dual $M $-action on $M\times\mathbb{Z} $ is the transpose, i.e. $T^{0}_{\gamma}: (m,r) \mapsto (m,
r+ d\alpha_{\gamma}(m))$ for any $\gamma \in M $. Thus the $M $-action preserves the regular function $\bar{\pi}=\mathcal{Z}^{(0,1)}$, and the induced $M$-action on monomial rational functions is given by
$$ \mathcal{Z}^{ (m,r)} \mapsto \mathcal{Z}^{T^{0}_{\gamma} (m,r)}= \mathcal{Z}^{(m, r+ Z(\gamma, m))},$$ for any $ \gamma\in M$ and $ (m,r)\in M\times\mathbb{Z}$.
For constructing the $M$-action on $L_{\Delta_{\varphi}}$, we consider
the $M $ action on $C(\Delta_{\varphi}) $ defined by
\begin{equation}\label{e1.1} T_{\gamma}: (m,r,l) \mapsto (m+ l\gamma, r+ d\alpha_{\gamma}(m)+lc_{\gamma},l) \ \ \ \ \ \gamma \in M, \end{equation} where
$ d\alpha_{\gamma}(m)= Z(\gamma, m)$ by (\ref{e0.0}) and $c_{\gamma}=
\frac{1}{2} Z(\gamma, \gamma)$.
We have $ T_{\gamma} (m,\varphi(m),1) = (m+\gamma, \varphi(m+\gamma),1)$ by (\ref{e0.2}). This action lifts the $M$-action on $X_{\Delta_{\varphi}}$ to an $M$-action on $L_{\Delta_{\varphi}}^{l}$.
More precisely the $M$-action on $L_{\Delta_{\varphi}}^{l}$ is given by \begin{equation}\label{e1.1+} \mathcal{Z}^{ (m,r,l)} \mapsto \mathcal{Z}^{T_{\gamma} (m,r,l)}= \mathcal{Z}^{ (m+ l\gamma, r+ d\alpha_{\gamma}(m)+lc_{\gamma},l)} \end{equation} for monomial sections.
\end{proof}
For any $k\in \mathbb{N}$, $M_{k}=kM$ is a subgroup, and acts on $(X_{\Delta_{\varphi}}, L_{\Delta_{\varphi}})$ induced by the $M$-action in the above lemma.
Note that the map $\bar \pi$ is $M_{k}$-invariant, and $M_k$ acts properly and discontinuously on $\bar{\pi}^{-1}(\Delta)$ for the unit disc $\Delta \subset \mathbb{C}$.
The quotient is the degeneration of principally polarized Abelian varieties $\pi: \mathcal{X}=\bar{\pi}^{-1}(\Delta)/M_{k}\rightarrow \Delta$, and $\pi^{-1}(t)=X_{t}=\bar{X}_{t}/M_{k}$. The central fiber $X_{0}$ of $\mathcal{X}$ is a union of finite irreducible toric varieties, and the corresponding intersection complex is $B_{k}=M_{\mathbb{R}}/M_{k} $ with rational polyhedron decomposition $\tilde{\mathcal{P}}$ induced by $\mathcal{P}$. There is a one to one corresponding between the $n$-dimensional cells of $\tilde{\mathcal{P}}$ and the irreducible components of $X_{0}$. More precisely,
for any $n$-dimensional cell $\sigma$ of $\tilde{\mathcal{P}}$, we regard it as a rational polytope in $M_{\mathbb{R}} $, and it defines a polarized toric variety $(X_{\sigma}, L_{\sigma})$. The irreducible component of $X_{0}$ corresponding to $\sigma$ is isomorphic to $X_{\sigma}$.
The quotient of $L_{\Delta_{\varphi}}$ by the $M_{k}$-action is the relative ample line bundle $\mathcal{L}$ on $\mathcal{X}$.
The restriction of $\mathcal{L}$ on any irreducible component $X_{\sigma}$ of $X_{0}$ is $L_{\sigma}$.
The $M_{k}$-invariant sections descent to sections of $\mathcal{L}$.
We claim that and $\mathcal{X}_{\Delta^{\circ}}=\pi^{-1}(\Delta^{\circ})$ is a base change of $\mathcal{X}_{\eta}$ (c.f. Section \ref{ppav}) via $t \mapsto t^{k}=s$.
Note that for any $t\neq 0$, $M_{k}$ acts on $\bar{X}_{t}=T_{\mathbb{C}}$ by
\begin{equation} \label{TZ}
\mathcal{Z}^{(m,r)} :=\mathcal{Z} ^{m}t^{r}\mapsto \mathcal{Z}^{T^{0}_{\gamma} (m,r)}:=\mathcal{Z}^{(m,r)}t^{ Z(\gamma,m)}.
\end{equation}
Since $t^{ k Z(k^{-1}\gamma,m)} $, $ k^{-1}\gamma \in M$, we obtain that $\bar{X}_{t}/M_{k} =X_{s}$ by (\ref{e0.100}) where $ t^{k}=s$.
For any $m \in B_{k}(\mathbb{Z})= M/M_{k}$, we define the theta function \begin{equation}\label{e1.2} \vartheta_{m}= \sum_{\gamma \in M_{k}} \mathcal{Z}^{T_{\gamma} (m,\varphi(m),1)}= \sum_{\gamma \in M_{k}} \mathcal{Z}^{(m+ \gamma, \varphi(m+\gamma),1)}, \end{equation} which is a section of $\mathcal{L}$ (cf. Example 6.1 of \cite{GHKS} and Section 2 of \cite{GS}). By abusing of notation, we will use $m$ to denote {\em both} a point in $M$ and its image under
the quotient map
$$M \longrightarrow B_{k}(\mathbb{Z})=M/kM\subset B_k=M_\RR/kM $$ without any confusing.
We obtain a basis $\{\vartheta_{m}|m \in B_{k}(\mathbb{Z}) \}$ of $H^{0}(\mathcal{X}, \mathcal{L})$. For any irreducible component $X_{\sigma}\subset X_{0}$, $\vartheta_{m}|_{X_{\sigma}}$ is a monomial section of $L_{\sigma}$, and it is not a zero section if and only if $m\in \sigma$.
For any $t\neq0$,
\begin{equation}\begin{split} \vartheta_{m}(w )& = \sum_{\gamma \in M_{k}} \mathcal{Z}^{m+ \gamma}t^{ \varphi(m+\gamma)} \\ & = \sum_{\gamma \in M_{k}} \exp \pi\sqrt{-1}(2\langle w, m+ \gamma\rangle+ \frac{\log t}{2 \pi \sqrt{-1}}Z(m+\gamma, m+\gamma)), \end{split} \end{equation} by $\varphi(m+\gamma)=\bar{\varphi}(m+\gamma)$, where $w=w_{1}e^{*}_{1}+\cdots +w_{n}e^{*}_{n}$. Thus it is the classical theta function
$$\vartheta_{m}(w)=\vartheta \begin{bmatrix}
m \\[0.2em]
0
\end{bmatrix} (w, \frac{\log t}{2 \pi \sqrt{-1}}Z_{ij} )
$$ on $X_{t}$.
If we regard $\vartheta_{m}|_{X_{t}}$ as on $X_{s}$, $s=t^{k}$, then
$$ \vartheta_{m}(w) = \sum_{\gamma' \in M} \exp \pi\sqrt{-1}(2k \langle w, \frac{m}{k}+ \gamma'\rangle+ k\frac{\log s}{2 \pi \sqrt{-1}}Z(\frac{m}{k}+\gamma',\frac{m}{k}+ \gamma')),$$
and a direct calculation shows
$$ \vartheta_{ m}(w+\mu + \frac{\log s}{2 \pi \sqrt{-1}}Z(p,\cdot))= \vartheta_{ m}(w)\exp k\pi \sqrt{-1}(-2\langle w,p\rangle- \frac{\log s}{2 \pi \sqrt{-1}}Z(p,p)),$$
for any $\mu\in N$ and $p\in M$. Thus $\mathscr{L}|_{X_{t}} \cong \mathscr{L}_{\eta}^{k}|_{X_{s}}$, i.e.
$\mathscr{L}|_{\mathcal{X}_{\Delta^{*}}}$ is the pull-back of $\mathscr{L}_{\eta}^{k}$.
\begin{remark}
Notice that, in particular, the monodromy action $\log t\to \log t+2\pi \sqrt{-1}$ acts trivially on $\vartheta_m$ via
\begin{eqnarray*}
&&\vartheta \begin{bmatrix}
m \\[0.2em]
0
\end{bmatrix} (w, \left(\frac{\log t}{2 \pi \sqrt{-1}}+1\right )Z_{ij} ) \\
&=& \sum_{\gamma \in M_{k}} \exp \pi\sqrt{-1}(2 \langle w, m+ \gamma \rangle+ (\frac{\log t}{2 \pi \sqrt{-1}}+1)Z(m+\gamma,m+ \gamma))\\
&=&\vartheta \begin{bmatrix}
m \\[0.2em]
0
\end{bmatrix} (w, \frac{\log t}{2 \pi \sqrt{-1}}Z_{ij} )
\end{eqnarray*} since $ Z(m+\gamma,m+ \gamma)=2\bar{\varphi}(m+\gamma) \in 2\mathbb{Z} $ for $m\in B_k(\ZZ)$.
\end{remark}
\begin{example}\label{extend}
We illustrate the explicit formula of the theta function $\vartheta_{m}$ in (\ref{e1.2}) in local coordinates for a special 1-dimensional family. Let $M\cong \mathbb{Z}$, $k=3$, and $\bar{\varphi}(y)=y^2$ on $M_{\mathbb{R}}$.
Note that $(0, 0)$ is a vertex of $\Delta_{\varphi}$, and the $\mathbb{C}$-algebra $\mathbb{C} [T_{(0,0)}\Delta_{\varphi}\cap M]$ is generated by $z_{1}=\mathcal{Z}^{(1,1)}$, $z_{2}=\mathcal{Z}^{(-1,1)}$ and $t=\mathcal{Z}^{(0,1)}$, where $T_{(0,0)}\Delta_{\varphi}$ is the tangent cone of $\Delta_{\varphi}$ at $(0, 0)$. The toric variety $Y_{0}={\rm Spec}(\mathbb{C} [T_{(0,0)}\Delta_{\varphi}\cap M])$ is defined in $\CC^3$ by equation $z_{1}z_{2}=t^{2}$, and an open subset of $Y_{0}$ is biholomorphic to a neighborhood $U_{0}$ of the zero strata of $X_{0}$ in $ \mathcal{X}$ corresponding to the vertex $(0, 0)$. Let us fix a trivialization of $\mathscr{L}|_{U_0}\cong \mathscr{O}_{Y_0}|_{U_0} $ via the indentification $\mathcal{Z}^{(m,r,1)}\mapsto \mathcal{Z}^{(m,r)}$ for any $(m,r)\in \Delta_{\varphi}\cap M$. Then we have
\begin{eqnarray*} & \vartheta_{0}& = 1+ \sum_{\nu \in \mathbb{Z}, \nu>0} (z_{1}^{3\nu}+z_{2}^{3\nu})t^{9\nu^{2}-3\nu},\\ &\vartheta_{1}& = z_{1}+ \sum_{\nu \in \mathbb{Z}, \nu>0}(z_{1}^{1+3\nu}t^{9\nu^{2}+3\nu}+z_{2}^{3\nu-1}t^{9\nu^{2}-9\nu+2}), \\ & \vartheta_{2}& = z_{2}+ \sum_{\nu \in \mathbb{Z}, \nu>0}(z_{1}^{3\nu -1}t^{9\nu^{2}-9\nu+2}+z_{2}^{3\nu+1}t^{9\nu^{2}+3\nu}), \end{eqnarray*} by (\ref{e1.2}). In particular, $\vartheta_i$ extends to the central fiber.
\end{example}
Notice that the central extention of the product group $B_k(\ZZ)\times T_k$ with $T_k:=N/N_k$ is precisely the {\em finite Heisenberg group }
$\HH_k=\mu_k\times B_k(\ZZ)\times T_k$ (cf. \cite[Section 3]{Mum-th}) with the multiplication rule:
\begin{equation}
(\mu,a,b)\cdot (\mu',a',b')=(\mu\mu'\displaystyle\exp \frac{2\pi\sqrt{-1}\langle b,a\rangle}{k}, a+a',b+b')
\end{equation}
for any $(\mu,a,b),(\mu',a',b')\in \mu_k\times B_k(\ZZ)\times T_k$, where $\mu_k$ is the cyclotomic group of order $k$.
\begin{lemma}\label{le01}
The group
$B_k(\ZZ)\times T_k$ acts on $(\mathcal{X},\mathscr{L})$ and induces a representation of $\HH_k$ on the $H^{0}(\mathcal{X}, \mathscr{L})=\mathrm{Span}_{\mathcal{H}}\{\vartheta_m\}_{m\in B_k(\ZZ)}$ via
\begin{itemize}
\item[i)] For any $a\in B_k(\ZZ)$
$$\mathfrak{T}_{a}\vartheta_{m} = \vartheta_{a+m} $$
\item[ii)] For any $b\in T_k$
$$ \mathfrak{S}_{b}\vartheta_{m}=\vartheta_{m} \exp \frac{2\pi\sqrt{-1}\langle b, m \rangle}{k},$$
\end{itemize}
for all $m\in B_k(\ZZ)$, where $\mathcal{H}$ denotes the ring of holomorphic functions on $\Delta$. In particular, the representation of $\HH_k$ on $H^0(\mathcal{X},\mathscr{L})$ is irreducible.
\end{lemma}
\begin{proof} The $M$-action on $(X_{\Delta_{\varphi}}, L_{\Delta_{\varphi}})$ in Lemma \ref{le00} induces the $B_{k}(\mathbb{Z})$-action on $(\mathcal{X}, \mathcal{L})$, which acts on $H^{0}(\mathcal{X}, \mathcal{L})$ given by \begin{eqnarray*}\mathfrak{T}_{a}\vartheta_{m}& = & \sum_{\gamma \in M_{k}} \mathcal{Z}^{T_{a}(m+ \gamma, \varphi(m+\gamma),1)} \\ & = & \sum_{\gamma \in M_{k}} \mathcal{Z}^{(m+ \gamma+a, \varphi(m+\gamma)+\alpha_{a}(m+\gamma) ,1)} \\ & = & \sum_{\gamma \in M_{k}} \mathcal{Z}^{(m+ \gamma+a, \varphi(m+\gamma+a),1)}\\ & = &\vartheta_{a+m}\end{eqnarray*} for any $m\in B_{k}(\mathbb{Z})$ and $a\in B_{k}(\mathbb{Z})$ by (\ref{e1.2}). We obtain i), and next we prove ii).
Note that there is a natural injective homeomorphism $\iota_{k}:T_{k} \hookrightarrow T_{\mathbb{C}} $ such that $$\mathcal{Z}^{m}(\iota_{k}(b))=\exp \frac{2\pi\sqrt{-1}\langle b, m\rangle}{k} $$ for any $b\in T_{k}$ and $m\in M$. The standard $T_{\mathbb{C}} $-action on $(X_{\Delta_{\varphi}}, L_{\Delta_{\varphi}}^{l})$ induces a $T_{k}$-action, which preserves the regular function $\bar{\pi}=\mathcal{Z}^{(0,1)} $, acts on monomial rational functions by $$\mathcal{Z}^{(m, r)}\mapsto \mathcal{Z}^{(m, r)} \exp \frac{2\pi\sqrt{-1}\langle b, m\rangle}{k} $$ for any $(m,r)\in M\times\mathbb{Z}$, and acts on monomial sections of $L_{\Delta_{\varphi}}^{l}$ via $$\mathcal{Z}^{(m, r,l)}\mapsto \mathcal{Z}^{(m, r,l)} \exp \frac{2\pi\sqrt{-1}\langle b, m\rangle}{k} $$ for any $(m,r,l)\in C( \Delta_{\varphi})$. The induced $T_{k}$-action on $\bigoplus_{l=0}^{\infty}H^{0}(X_{\Delta_{\varphi}}, L_{\Delta_{\varphi}}^{l})$ commutes with the $M_{k}$-action by $$ \mathcal{Z}^{T_{\gamma}(m, r,l)} \exp \frac{2\pi\sqrt{-1}\langle b, m\rangle}{k}= \mathcal{Z}^{(m+ l\gamma, r+ d\alpha_{\gamma}(m)+lc_{\gamma},l)} \exp \frac{2\pi\sqrt{-1}\langle b, m+ l\gamma\rangle}{k}, $$ for any $\gamma\in M_{k}$ by (\ref{e1.1+}). Thus
the $T_{k}$-action commutes with the $M_{k}$-action on $(X_{\Delta_{\varphi}}, L_{\Delta_{\varphi}}^{l})$, which induces a $T_{k}$-action on $(\pi:\mathcal{X}\rightarrow \Delta, \mathcal{L})$. We denote $ \mathfrak{S}$ the induced $T_{k}$-action on $H^{0}(\mathcal{X}, \mathcal{L})$, which satisfies $$
\mathfrak{S}_{b}\vartheta_{m}= \sum_{\gamma \in M_{k}} \mathcal{Z}^{(m+ \gamma, \varphi(m+\gamma),1)}\exp \frac{2\pi\sqrt{-1}\langle b, m +\gamma \rangle}{k}=\vartheta_{m}\exp \frac{2\pi\sqrt{-1}\langle b, m \rangle}{k}, $$
for any $b\in T_{k}$ and any $m\in B_{k}(\mathbb{Z})$,
by (\ref{e1.2}).
\end{proof}
\begin{remark} The direct calculations show that on any $X_{t}$, $t\neq 0$, the action of $B_k(\ZZ)\times T_k$ on $(X_t,\mathscr{L}|_{X_t})$ is given by $ w\mapsto w+\frac{b}{k}+ \frac{\log t}{2\pi\sqrt{-1}}Z(\cdot,a)$,
\begin{eqnarray*}(\mathfrak{T}_{a}\vartheta_{m})(w)& = &\exp\pi\sqrt{-1}(2\langle w, a\rangle+\frac{\log t}{2\pi\sqrt{-1}}Z(a,a))\vartheta_{m}(w+ \frac{\log t}{2\pi\sqrt{-1}}Z(\cdot,a))\\ & = & \vartheta_{a+m} (w),\\
( \mathfrak{S}_{b}\vartheta_{m})(w)&= & \vartheta_m(w+\frac{b}{k})=\exp \frac{2\pi\sqrt{-1}\langle b, m \rangle}{k}\vartheta_{m}(w),\end{eqnarray*} for any $m\in B_k(\ZZ)$ and $(a,b)\in B_k(\ZZ)\times T_{k}$. In particular, there the finite torus $B_k(\ZZ)\times T_{k}$ acts on the image of the projective embedding $\Phi_k(X_t)$.
\end{remark}
\section{Proofs of Main Theorems}
Before we start the proof, let us recall the Hermitian metric on $ \mathscr{L}|_{X_t}\to X_t$ with $s=t^k$ is given by
\begin{equation} h(w):=\exp\left ( \frac{2\pi}{\log|t|}Z^{ij}y_iy_j\right )= \exp k\left ( \frac{2\pi}{\log|s|}Z^{ij}y_iy_j\right )
\end{equation}
with
$$w_{i}=x_{i}+(\arg (s)-\sqrt{-1}\frac{\log |s|}{2\pi } )\sum_{j=1}^nZ_{ij}y_{j},\ i=1, \cdots, n\ . $$ On $X_{t}\cong X_{s}$, we have $$\omega_{t}=-\sqrt{-1}\partial\overline{\partial}\log h=k \omega_{s}, $$ by (\ref{e0.10}).
\begin{lemma} $\forall a\in M, b\in N$ and function $f: N_{\mathbb{C}} \rightarrow \mathbb{C}$ we have
\begin{equation}
h(w)|\mathfrak{S}_b f|^2 =(h|f|^2)(w+\frac{b}{ k});
\text{ } h(w)|\mathfrak{T}_a f|^2 =(h|f|^2)\left (w+\frac{\log s}{2\pi\sqrt{-1}}Z(\cdot, \frac{a}{ k})\right)\ .
\end{equation}
\end{lemma}
\begin{proof} See \cite[Proposition 3]{WY}.
\end{proof}
As a consequence, for $a\in B_k(\ZZ),\ b\in T_k$ and $ f\in H^0(X_t,\mathscr{L}|_{X_t})$ we have
\begin{eqnarray*}
\|(\mathfrak{S}_b f)(w)\|^2_\mathrm{FS}
&=&\int\frac{|(\mathfrak{S}_bf)(w)|^2}{\sum_{m}|\vartheta_m(w)|^2}\Phi_k^\ast\omega^n_\mathrm{FS}\\
&=&\int\frac{h|(\mathfrak{S}_bf)(w)|^2}{\sum_{m}h|\mathfrak{S}_b\vartheta_m(w)|^2}\Phi_k^\ast\omega^n_\mathrm{FS}\\
&=&\int\frac{(h|f|^2)\left(w+\frac{b}{k}\right)}{(\sum_{m}h|\vartheta_m|^2)\left(w+\frac{b}{k}\right)}(\mathfrak{S}_b\circ\Phi_k)^\ast\omega^n_\mathrm{FS}\\
&=&\int\left(\frac{|f|^2}{\sum_{m}|\vartheta_m|^2}\Phi_k^\ast\omega^n_\mathrm{FS}\right)\left(w+\frac{b}{k}\right)\\
&=&\|f(w)\|^2_\mathrm{FS}
\end{eqnarray*}
and
\begin{eqnarray*}
\|(\mathfrak{T}_a f)(w)\|^2_\mathrm{FS}
&=&\int\frac{|(\mathfrak{T}_af)(w)|^2}{\sum_{m}|\vartheta_m(w)|^2}\Phi_k^\ast\omega^n_\mathrm{FS}\\
&=&\int\frac{h|(\mathfrak{T}_af)(w+ \frac{\log s}{2\pi\sqrt{-1}}Z(\cdot ,\frac{a}{ k}))|^2}{\sum_{m}h|\mathfrak{T}_a\vartheta_m(w+ \frac{\log s}{2\pi\sqrt{-1}}Z(\cdot ,\frac{a}{ k}))|^2}(\mathfrak{T}_a\circ\Phi_k)^\ast\omega^n_\mathrm{FS}\\
&=&\int\left(\frac{|f|^2}{\sum_{m}|\vartheta_m|^2}\Phi_k^\ast\omega^n_\mathrm{FS}\right)(w+ \frac{\log s}{2\pi\sqrt{-1}}Z(\cdot ,\frac{a}{ k}))\\
&=&\|f(w)\|^2_\mathrm{FS}\ .
\end{eqnarray*}
Hence the finite group generated by image of
$$ \{\mathfrak{T}_a,\mathfrak{S}_b \mid (a,b)\in B_k(\ZZ)\times T_k\} \subset \mathrm{GL}(H^0(X_{ s},\mathscr{L}|_{X_{ s}}))$$
actually lies in $\mathrm{U}(k^n)$ with respect to the Fubini-Study metric induced via the embedding
$$
\Phi_k=[\vartheta_{m_1},\cdots,\vartheta_{m_{k^n}}]:\mathcal{X}\longrightarrow \CC\PP^{k^n-1}\ .
$$
\begin{proof}[Proof of Theorem \ref{thm01}]
It follows from above that the action generated by $B_k(\ZZ)\times T_k$ via $\mathfrak{T}$ and $\mathfrak{S}$ lies in $\mathrm{U}(k^n)$. And Lemma \ref{le01} implies that sections $\{\vartheta_m\}_{m\in B_k(\ZZ)}$ forms an orthonormal basis with respect to the pull back of Fubini-Study metric via the map $\Phi_k$, that is, the embedding $\Phi_k$ is balanced for each $t\in \Delta^\circ$.
On the other hand, $\Phi_k(X_t)$ being balanced for each $t\ne 0$ implies that the Chow point for $\Phi_k(X_t)$ lies on the $0$-level set the moment map
\begin{equation*}
\begin{array}{cccc}
\mu_{{\rm SU}}:{\rm Chow}_{\mathbb{CP}^{k^n-1}}(d,n) & \overset{}{ \xrightarrow{\hspace*{1.7cm}}} & \mathfrak{su}(k^n) \\
X_t & \longmapsto & \sqrt{-1}\cdot \displaystyle \int_{X_t}\left(\frac{\vartheta_m \bar\vartheta_{m'}}{\sum_m|\vartheta_m|^2}-\frac{\delta_{mm'}}{k^n}\right )\frac{ \Phi_k^\ast\omega^n_\mathrm{FS}}{n!} &
\end{array}%
\end{equation*}%
of the $\mathrm{SU}(k^n)$-action on the Chow variety of $n$-dimensional degree $d$ cycles in $\CC\PP^{k^n-1}$ (c.f. \cite[Proposition 17]{W} and \cite[Theorem 1.4]{ZS}),
which is proper via standard Kirwan-Kempf-Ness theory in \cite{Kir} (c.f. also in \cite{Th}). Notice that $\{\vartheta_m(\cdot, \frac{\log t}{2\pi\sqrt{-1}}Z)\}_m$ vary holomorphically with respect to $t$ and can be extended to $X_0$ by the construction of theta functions in Section 3, these imply that $\Phi_k$ has bounded image in $\CC\PP^{k^n-1}$, by Riemann mapping Theorem, the {\em unique} continuous extension $\Phi_k(X_0)$ must lies in $ \mu_{\rm SU}^{-1}(0)$, that is, the embedding $\Phi_k(X_0)$ is balanced as well.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm02}] For any $m\in M$, we denote $\check{m}\subset N_{\mathbb{R}}$ the dual polytope of $m$ with respect to $\varphi$. More precisely, if $ \check{m}^{o}\subset N_{\mathbb{R}}\times \mathbb{R}$ denotes the dual cone of the tangent cone $T_{(m,\varphi(m))}\Delta_{\varphi} \subset M_{\mathbb{R}}\times \mathbb{R}$ of $\Delta_{\varphi}$ at the vertex $(m, \varphi(m))$, then $\check{m}= \check{m}^{o} \cap (N_{\mathbb{R}}\times \{1\})$. The Monge-Amp\`{e}re measure of $\varphi$ is $${\rm MA}(\varphi)=\sum_{m\in B_{k}(\mathbb{Z})}{\rm Vol}(\check{m})\delta_{m}, $$ where ${\rm Vol}(\check{m})$ is the Euclidean volume of $\check{m}$ (cf. Proposition 2.7.4 in \cite{BPS}).
Since the $M$-action $\check{T}^{0}$ on $N_{\mathbb{R}}\times \mathbb{R} $ preserves the fan $\Sigma$ of $\Delta_{\varphi}$, we obtain $\check{T}^{0}_{m-m'}(\check{m}')=\check{m}$ for any two $m$ and $m'\in M$, and thus ${\rm Vol}(\check{m})=V$ is independent of $m$.
Let $D$ be the fundamental domain of the $M_{k}$-action $\check{T}^{0}$ on $N_{\mathbb{R}}\times \{1\} $. Since $\check{T}^{0}_{\gamma} (0,l)= (- d\alpha_{\gamma}, l) =(-Z(\gamma,\cdot),1)$ for any $\gamma \in M $, we let $D$ be the convex hull of $(0,1), (-Z(k e_{1},\cdot),1), \cdots, (-Z(k e_{n},\cdot),1)$ in $N_{\mathbb{R}}\times \{1\} $, where $e_{1}, \cdots, e_{n}$ is a basis of $M$. The Euclidean volume of $D$ is ${\rm Vol}(D)=k^{n}\det (Z_{ij})$, where $Z_{ij}=Z(e_{i},e_{j}) $, and we have $$ {\rm Vol}(D)=\sum_{m\in B_{k}(\mathbb{Z})}{\rm Vol}(\check{m})=k^{n} V.$$ We obtain $V=\det (Z_{ij})$, and the conclusion $${\rm MA}(\varphi)=\det (Z_{ij}) \sum_{m\in B_{k}(\mathbb{Z})}\delta_{m}. $$
Let $\chi_{k}: B\rightarrow B_{k}$ be the diffeomorphism induced by the dilation $y_{i} \mapsto k y_{i}$, $i=1, \cdots, n$, of $M_{\mathbb{R}}$. For any smooth function $f$ on $B$, $$ \frac{1}{k^{n}}\int_{B} f \chi_{k}^{*}( \sum_{m\in B_{k}(\mathbb{Z})}\delta_{m})=\frac{1}{k^{n}} \sum_{m'\in (\frac{1}{k}M)/M}f(m') \rightarrow \int_{B}f dy_{1}\wedge \cdots \wedge dy_{n}$$ when $k\rightarrow \infty$, which implies that $$ \frac{1}{k^{n}} \chi_{k}^{*}{\rm MA}(\varphi) \rightharpoonup \det(Z_{ij})dy_{1}\wedge \cdots \wedge dy_{n}={\rm MA}(\bar{\varphi})$$ in the weak sense.
By (\ref{e0.1}) and (\ref{e0.2}), $(\varphi-\bar{\varphi})(y+\gamma)=(\varphi-\bar{\varphi})(y)$ for any $y\in M_{\mathbb{R}}$ and any $\gamma \in M$, and $\varphi-\bar{\varphi}$ is a well-defined function on $B_{k}$. Since $$ \sup_{ M_{\mathbb{R}}}|\varphi-\bar{\varphi}|=\sup_{y \in M_{\mathbb{R}}/ M}|\varphi-\bar{\varphi}|(y),$$ we obtain
$$
\sup_{ B}\left | \frac{1}{k^{2}} \chi_{k}^{*} \varphi- \frac{1}{k^{2}} \chi_{k}^{*} \bar{\varphi}\right| =\frac{1}{k^{2}}\sup_{ M_{\mathbb{R}}}|\varphi-\bar{\varphi}| \rightarrow 0,
$$
when $k\rightarrow\infty$, and the conclusion by $\frac{1}{k^{2}}\bar{\varphi}(ky)=\bar{\varphi}(y)$.
\end{proof}
\section{Appendix}
In this section, we state the following Theorem which unified the proof of balanced embedding for projective space and princpally polarized Abelian varieties.
\newcommand{\mathrm{SU}}{\mathrm{SU}}
\begin{theorem}Let $X\subset \PP^N$ be a subvariety and $G<\mathrm{SU}(N+1)$ be a compact subgroup which leaves the embedding $X\hookrightarrow\PP^N$ invariant. Suppose the centralizer of $c_G< \mathrm{SU}(N+1)$ of $G$ inside $\mathrm{SU}(N+1)$ is trivial. Then the embedding $X\subset \PP^N$ is balanced, i.e.
\begin{equation}\label{mu}
\int_X\mu_{\PP^N}\frac{\omega_\mathrm{FS}^n}{n!}=0.
\end{equation}
\end{theorem}
\begin{proof}
Notice that the moment map of $\mathrm{SU}(N+1)$-action on the Chow variety $\mathrm{Chow}_{\PP^N}(n,d)$ of dimension $n$ and degree $d$ cycle in $\PP^N$ is precisely given by \eqref{mu}. In particular, it is $\mathrm{SU}(N+1)$, and hence $G$-equivariant. This implies that for any $g\in G$
\begin{eqnarray*}\label{mu}
&&\int_{ X}\mu_{\PP^N}\frac{\omega_\mathrm{FS}^n}{n!}
=\int_{g\cdot X}\mu_{\PP^N}\frac{\omega_\mathrm{FS}^n}{n!}=\int_{X}\mu_{\PP^N}\circ g\frac{\omega_\mathrm{FS}^n}{n!}\\
&=&\mathrm{Ad}_g\left(\int_{X}\mu_{\PP^N}\frac{\omega_\mathrm{FS}^n}{n!}\right).
\end{eqnarray*}
By our assumption, we have $\displaystyle \int_{ X}\mu_{\PP^N}\frac{\omega_\mathrm{FS}^n}{n!}\in \mathfrak{c}_G=0$, where $\mathfrak{c}_G=\mathrm{Lie}(c_G)$ is the Lie algebra. And our proof is thus completed.
\end{proof}
|
\section{Introduction}
An early attempt to unify theories of gravitation and electromagnetism was realized by Kaluza~\cite{KaluzaZUP966.1921} in five-dimensional theory based on classical general relativity. That was later extended by Klein~\cite{KleinZP895.1926} with quantum interpretations. The idea of extra dimensions played a key role in development of an existing new fundamental theory of physics, superstring theory (M-theory) which requires the spacetime to have ten (eleven) dimensions.
Braneworld models were proposed in order to tackle the hierarchy problem. The question why there is such a large gap between the electroweak scale at $\sim 1$ TeV and the Planck scale at $\sim 10^{16}$ TeV has been addressed in~\cite{Arkani-HamedPLB263.1998}. The first string realization of low-scale gravity and
braneworld models were given,
pointing out the motivation of TeV
strings from the stabilization of mass hierarchy and
the graviton emission in the bulk~\cite{AntoniadisPLB257.1998}. Randall and Sundrum proposed their braneworld models~\cite{RandallPRL3370.1999,RandallPRL4690.1999} where the hierarchy problem can be addressed, as the large size of the extra dimension plays crucial role to fill the gap between the electroweak and Planck effective scales. According to the Randall-Sundrum model, our Universe is a three-brane (domain wall) embedded in five-dimensional bulk spacetime; one extra dimension is large, and the bulk is a slice of the anti-de Sitter (AdS) spacetime, i.e., spacetime with a negative cosmological constant. Later, this model was further extended by Shiromizu et al.~\cite{ShiromizuPRD024012.2000}. Reissner-Nordstr\"{o}m-like static, a spherically symmetric black hole solution with a tidal charge parameter (instead of electric charge), localized on a three-brane in five-dimensional gravity in the Randall-Sundrum model was obtained, without finding the bulk metric, by Dadhich et al.~\cite{DadhichPLB1.2000}. So far, several black hole solutions~\cite{EmparanJHEP007.2000,BronnikovPRD024025.2003,ShankaranarayananIJMPD1095.2004}, wormholes~\cite{BronnikovPRD064027.2003}, and nonuniform stars and gravastars~\cite{OvalleIJMPD837.2009} in the Randall-Sundrum model have been obtained (see reviews~\cite{MaartensLRR13.2010,GermaniPRD124010.2001} and references therein.).
The discovery of the braneworld model opens up a new window to test modified general relativity. Thus, up to now, many physical effects related to test the particle motion and geodesic structure of braneworld black holes~\cite{ScheeGRG1795.2009,AAAPRD044022.2010} and neutron and compact stars~\cite{KotrlovaCQG225016.2008,StuchlikCQG175002.2011} have been studied.
One of the most important properties of black holes is their characteristic oscillations, which are called \textit{quasinormal modes} that carry information about them. The quasinormal modes determine, e.g., ringdown of gravitational waves created while a black hole is born. They are characterized by black hole parameters being dependent on initial perturbations. They have complex frequencies -- real and imaginary parts of the quasinormal frequencies represent frequencies of the real oscillations and their dissipation rate, respectively.
Recently, the interferometric LIGO detectors have measured the first ever gravitational wave signals from the merging of two black holes~\cite{AbbottPRL061102.2016}. Later, it has been shown that current precision of the experiment leaves some possibilities for alternative theories of gravity~\cite{KonoplyaPLB350.2016,Abramowicz.2016}. Cardoso et al.~\cite{Cardoso.2016} stated that ringdown waveforms indicate the existence of the stable light rings regardless of existence of the horizons. According to~\cite{Chirenti.2016}, the ringdown part of the GW150914 signal has excluded formation of gravastar by the merger of two rotating compact objects. So far, characteristic ringdown signals (quasinormal modes) of the various black holes have been studied in great detail by a number of authors within perturbation theory\cite{StarinetsPRD124013.2002,VazquezJHEP008.2002,MaedaPRD086012.2005, AbdallaNPB40.2006,BertiPRD024013.2006,ChenPLB282.2007,ZhidenkoPRD024007.2008, MorganJHEP117.2009,HodCQG105016.2011}.
Since there is still room open for alternative theories of gravity, we aim to study in this paper perturbations of the black holes localized in the Randall-Sundrum braneworld. We concentrate on their stabilities, scattering effects and quasinormal modes. After this work was almost completed, electromagnetic perturbations of current braneworld solution by Molina et al.~\cite{Molina.2016} appeared in arXiv. Despite the fact that some our results of the quasinormal frequencies of the electromagnetic perturbations are repeating the results of~\cite{Molina.2016}, we keep them in order to compare them with profiles of scalar and gravitational perturbations. Moreover, we study scattering and absorption problem in the electromagnetic case as well. The paper is organized as follows: In section~\ref{sec-spacetime} we briefly describe the spacetime geometry and its main properties. In section~\ref{sec-pert-eqs} the equations for scalar, electromagnetic, and axial and polar gravitational perturbations are introduced. We give some numerical results, such as quasinormal frequencies obtained by the sixth order WKB method in low and large multipole number limits, and stability analysis in section~\ref{results}. In section~\ref{scattering} classical scattering problem is solved by using the standard $S$-matrix. In section~\ref{absorption} we study absorption cross section of massless scalar waves by the braneworld black hole in comparison with the Schwarzschild and Reissner-Nordstr\"{o}m black holes. Finally, we present some concluding remarks in section~\ref{summary}. Throughout the paper we use the geometric system of units $c=G=\hbar=1$ and a spacelike signature $(-,+,+,+)$.
\section{Black hole in the braneworld}\label{sec-spacetime}
We focus on the static, spherically symmetric black hole geometry localized on a braneworld, described by the line element~\cite{DadhichPLB1.2000}
\bear \label{metric}
ds^2=-f(r)dt^2+\frac{dr^2}{f(r)}+r^2d\theta^2+r^2\sin^2\theta d\phi^2\ ,
\ear
where
\bear \label{lapse}
f(r)=1-\frac{2M}{r}+\frac{\beta}{r^2}\ ,
\ear
and $\beta$ is a constant parameter. One can see from~(\ref{lapse}) that if $\beta=0$ the spacetime metric (\ref{metric}) reduces to the Schwarzschild one. Moreover, for $\beta\geq0$ the spacetime metric (\ref{metric}) is identical to the Reissner-Nordstr\"{o}m black hole one with two horizons which both are smaller than the one of the Schwarzschild black hole ($0\leq r_-\leq r_+\leq 2M$). However, in the braneworld $\beta$ can have negative value ($\beta<0$) too. In this case, black hole has only one horizon which is always bigger than the one of the Schwarzschild black hole
\bear \label{horizon}
r_+=M+\sqrt{M^2-\beta}> 2M\ .
\ear
In this paper we consider the latter case, $\beta<0$. Therefore, in order to guarantee its negativity we introduce new notation $\beta=-Q^*$, where $Q^*$ is always positive, $Q^*>0$, and is called \textit{tidal charge (brane tension) parameter}~\cite{DadhichPLB1.2000}.
Black hole entropy is determined by horizon area as
\bear \label{entropy}
S=\frac{A}{4\pi}=r_+^2=\left(M+\sqrt{M^2+Q^*}\right)^2\ .
\ear
where $A$ is horizon area. The Hawking temperature is given by
\bear \label{temperature}
T=\frac{\sqrt{M^2+Q^*}}{2\pi\left(M+\sqrt{M^2+Q^*}\right)^2}\ .
\ear
As one see from (\ref{temperature}), unlike in the case of the Reissner-Nordstr\"{o}m black hole, that the nonvanishing tidal charge parameter leads to an increase of the black hole entropy and decrease of the Hawking temperature.
\section{Perturbation equations in braneworld spacetime}\label{sec-pert-eqs}
\subsection{Scalar and electromagnetic perturbations}
By considering perturbation terms dependent on time as $\sim exp(i\omega t)$ and separating angular and radial perturbations by introducing the tortoise coordinate $dx=dr/f$, we obtain a Schr\"{o}dinger-like wave equation as follows:
\bear \label{ax-5}
\left(\frac{d^2}{dx^2}+\omega^2\right)Z_s=V_s Z_s\ ,
\ear
where $s=0,1$ represent scalar and electromagnetic perturbations, respectively. For the massive scalar perturbations with mass $m$ of the black hole on the brane we have the potential
\bear \label{scalar-potential}
V_0&=&f\left[\frac{l(l+1)}{r^2}+\frac{f'}{r}+m^2\right]\nn
&=&f\left[\frac{l(l+1)}{r^2}+\frac{2(Mr+Q^*)}{r^4}+m^2\right]\ ,
\ear
where prime ("$'$") denotes the derivative with respect to $r$. Here, $l$ is the multipole number which represents the spherical harmonic index and takes only nonnegative integers for scalar perturbations.
For the electromagnetic perturbations the potential reads
\bear \label{em-potential}
V_1=f\frac{l(l+1)}{r^2}
\ear
where the multipole number $l$ for electromagnetic perturbations takes only natural numbers.
The potentials for the scalar~(\ref{scalar-potential}) and electromagnetic~(\ref{em-potential}) perturbations can be written in a compact form as
\bear \label{scalar-em}
V_s=f\left[\frac{l(l+1)}{r^2}+(1-s)\left(\frac{f'}{r}+m^2\right)\right]\ .
\ear
where $s=0$ and $s=1$ correspond to the scalar and electromagnetic perturbations, respectively. Of course, in the case of the electromagnetic perturbations in the standard situations we assume massless photons, i.e., $m=0$.
\subsection{Gravitational perturbations}
It is well known that the simplest way of studying the gravitational perturbations around black holes is to introduce the first order perturbations. If the considered black hole is not a vacuum solution of the Einstein equations,
the perturbation equations are governed by the equation $\delta R_{\mu\nu}=-\delta E_{\mu\nu}$. In the case of a black hole localized on a three-brane in five-dimensional gravity,
$R_{\mu\nu}$ and $E_{\mu\nu}$ are the Ricci tensor and
the projection of the five-dimensional Weyl tensor on the brane, respectively. The gravitational perturbations of the higher dimensional
black holes can be easily obtained
by using the master equations presented in~\cite{KodamaPTP701.2003}. However, the black hole solution in the
brane~\cite{DadhichPLB1.2000} was found without finding the 5-dimensional bulk metric. Therefore, we adopt the simplifying assumption
$\delta E_{\mu\nu}=0$ that can be
justified at least in a region where the perturbation
energy does not exceed the
threshold of the Kaluza-Klein massive modes~\cite{AbdallaNPB40.2006}. Moreover, the fact that the
gravitational perturbative field cannot travel deep into the bulk~\cite{AbdallaPRD083512.2002}
supports the above assumption.
In this subsection we present a general formalism for gravitational perturbations in a static, spherically symmetric background, following Chandrasekhar's method~\cite{Chandra.1983}. With the notation $(x^0, x^1, x^2, x^3)\equiv (t, \phi, r, \theta)$, the spherically symmetric, time-independent metric with small perturbations can be written in the form
\bear \label{pert-metric}
ds^2= -e^{2(\nu+\delta\nu)}dt^2&+&e^{2(\psi+\delta\psi)}(d\phi-q_0 dt-q_2 dr-q_3 d\theta)^2\nn
&+&e^{2(\mu_2+\delta\mu_2)} dr^2+ e^{2(\mu_3+\delta\mu_3)} d\theta^2 ,
\ear
where
\bear
&&e^{2\nu}=f(r), \qquad e^{2\psi}=r^2\sin^2\theta, \nn
&&e^{-2\mu_2}=f(r), \qquad e^{2\mu_3}=r^2,
\ear
and $q_0$, $q_2$, $q_3$, $\delta\nu$, $\delta\psi$, $\delta\mu_2$ and $\delta\mu_3$ are nonvanishing small perturbation terms. The first three small quantities ($q_0$, $q_2$, $q_3$) characterize axial tensor perturbations with odd parity, while the small quantities ($\delta\nu$, $\delta\psi$, $\delta\mu_2$, $\delta\mu_3$) correspond to the polar tensor perturbations with even parity. Below we briefly present these perturbations for the braneworld black hole~\cite{DadhichPLB1.2000}.
\subsubsection{Axial perturbations}
The axial perturbations with perturbation terms $q_0$, $q_2$ and $q_3$ are governed by the relation
\beq \label{ax-11}
\delta R_{\mu\nu}=0\ .
\eeq
We write the linearized axial gravitational perturbation equations from~(\ref{ax-11}) in the form
\bear\label{ax-01}
(r^2f Q_{23}\sin^3\theta)_{,3}=-r^4\sin^3\theta Q_{02,0}, \quad (\delta R_{12}=0),
\ear
\bear\label{ax-02}
(r^2f Q_{23}\sin^3\theta)_{,2}=\frac{r^2\sin^3\theta}{f}Q_{03,0}, \quad (\delta R_{13}=0),
\ear
where $Q_{\alpha\beta}\equiv q_{\alpha,\beta}-q_{\beta,\alpha}$. By considering the perturbation terms depending on time as $\sim exp(i\omega t)$ and introducing the new notation $Q$ by the relation
\beq \label{ax-2}
Q=r^2f Q_{23}\sin^3\theta=r^2f (q_{2,3}-q_{3,2})\sin^3\theta,
\eeq
we can write Eqs.~(\ref{ax-01}) and~(\ref{ax-02}) as follows:
\beq \label{ax-3.1}
\frac{1}{r^4\sin^3\theta}\frac{\partial Q}{\partial\theta}
=-i \omega q_{0, 2}+\omega^2q_{2}\ ,
\eeq
\beq \label{ax-3.2}
\frac{f}{r^2\sin^3 \theta}\frac{\partial Q}{\partial r}
=i\omega q_{0,3}+\omega^2q_{3}\ .
\eeq
In order to eliminate $q_{0, 3}$ from these equations, we differentiate Eqs.~(\ref{ax-3.1}) and~(\ref{ax-3.2}) with respect to the coordinates $\theta$ ($x^3$) and $r$ ($x^2$), respectively. Then, we obtain
\bear \label{ax-4}
r^4\frac{\partial}{\partial r}\left(\frac{f}{r^2}\frac{\partial Q}{\partial r}\right)&+&\sin^3\theta\frac{\partial}{\partial\theta}\left(\frac{1}{\sin^3\theta} \frac{\partial Q}{\partial\theta}\right)\nn &&+\omega^2\frac{r^2}{f}Q=0\ .
\ear
Eq.~(\ref{ax-4}) can be separated to the radial and angular variable differential equations by choosing the function $Q$ as $Q(r,\theta)=Q(r)C_{l+2}^{-3/2}(\theta)$, where $C_{l+2}^{-3/2}(\theta)$ are the Gegenbauer polynomials and related to the Legandre function $P_{l}(\theta)$ as
\bear \label{gegenbauer}
C_{l+2}^{-3/2}(\theta)=(P_{l,\theta,\theta}-P_{l,\theta}\cot\theta)\sin^2\theta\ ,
\ear
By substituting $Q(r)=rZ_2^{(-)}$ and introducing the tortoise coordinate $dr_\ast=dr/f$, we obtain a Schr\"{o}dinger-like wave equation
\bear \label{ax-1}
\left(\frac{d^2}{dx^2}+\omega^2\right)Z_2^{(\pm)}=V_2^{(\pm)}Z_2^{(\pm)}\ ,
\ear
where $-$ and $+$ denote axial and polar gravitational perturbations, respectively. Potential for the axial gravitational perturbations reads
\bear \label{ax-potential1}
V_2^{(-)}=f\left[\frac{l(l+1)}{r^2}+\frac{3(f-1)}{r^2}+4\pi(\rho-p_r)\right],
\ear
where $\rho=-T_t^t$ and $p_r=T_r^r$ are energy density and radial pressure of the fluid, respectively, with $G_\mu^\nu/8\pi=T_\mu^\nu=diag\{\rho,p_r,p_\theta,p_\phi\}$. Here, $\rho=-p_r=-Q^*/(8\pi r^4)$. Then, Eq.~(\ref{ax-potential1}) for the axial gravitational perturbations of the black hole on the brane takes the form:
\bear \label{ax-potential2}
V_2^{(-)}=f\left[\frac{l(l+1)}{r^2}-\frac{2(3Mr+2Q^*)}{r^4}\right].
\ear
In the limit $Q^*=0$ this potential coincides with the Regge-Wheeler potential for the Schwarzschild black hole~\cite{Regge-Wheeler1957}.
\subsubsection{Polar perturbations}
In the case of polar perturbations with even parity, $\delta\nu$, $\delta\psi$, $\delta\mu_2$ and $\delta\mu_3$ in the metric functions are considered nonvanishing. Polar perturbations are examined by vanishing the following Ricci and Einstein tensors: $\delta R_{02}$, $\delta R_{03}$, $\delta R_{23}$, $\delta R_{11}$ and $\delta G_{22}$ (see Ref.~\cite{Chandra.1983} for details). Moreover, in order to separate $r$ and $\theta$ variables one may introduce the following notations:
\bear \label{polar-1}
&&\delta\nu=N(r)P_l(\cos\theta)\ ,\\
&&\delta\mu_2=L(r)P_l(\cos\theta)\ ,\\
&&\delta\mu_3=[T(r)P_l+V(r)P_{l,\theta,\theta}]\ ,\\
&&\delta\psi=[T(r)P_l+V(r)P_{l,\theta}\cot\theta]\ .
\ear
Then, we obtain the following perturbation equations in terms of the new radial functions $N(r)$, $L(r)$, $T(r)$ and $V(r)$ in the form
\bear \label{polar-2}
&&N'=a N + b L +c X\ ,\\
&&L'=(a-\frac{1}{r}+\nu')N+(b-\frac{1}{r}-\nu')L+c X\ ,\\
&&X'=-(a-\frac{1}{r}+\nu')N-(b+\frac{1}{r}-2\nu')L\nn
&&-(c+\frac{1}{r}-\nu')X\ ,
\ear
where prime ("$'$") denotes the derivative with respect to $r$ and $X=nV$ and $n=(l-1)(l+2)/2$. Furthermore,
\bear \label{polar-coefs}
&&a=\frac{n+1}{rf}\ ,\\
&&b=-\frac{1}{r}-\frac{n}{rf}+\frac{f'}{2f}+\frac{nrf'^2}{4f^2}+\frac{r\omega^2}{f^2}\ ,\\
&&c=-\frac{1}{r}+\frac{1}{rf}+\frac{rf'^2}{4f^2}+\frac{r\omega^2}{f^2}\ .
\ear
After making some simplifications, we obtain the Schr\"{o}dinger-like equation for $"+"$ family in~(\ref{ax-1}), where the wave function $Z_2^{(+)}$ reads
\bear \label{polar-4}
Z_2^{(+)}=rV_2^{(+)}-K(L+nV_2^{(+)})\ ,
\ear
with
\bear \label{polar-5}
K=\frac{2r}{2(n+1)-2f+rf'}\ .
\ear
The potential for the polar perturbations reads
\bear \label{polar-potential}
V_2^{(+)}=&&\frac{f}{2r^2K}\left[2r\left(1+2rf'K'+f(rK''-4K')\right)\right.\nn
&&\left.+K(2\lambda+12f-7rf'+r^2f'')\right]\ ,
\ear
where
\bear \label{polar-1}
K=\frac{2r}{2(\lambda+1)-2f+rf'}\ ,
\ear
and $\lambda=(l-1)(l+2)/2$. For the gravitational perturbations, the multipole number $l$ takes all natural numbers starting from $2$. For the braneworld black hole the potential of the polar gravitational perturbations reads
\bear \label{polar-potential2}
V_2^{(+)}=\frac{2f}{r^4[2Q^*+r(3M+\lambda r)]^2}\left[r^3A+Q^*B\right]\ ,
\ear
where
\bear \label{coefr}
A=9M^3+9M^2\lambda r+3M\lambda^2r^2+\lambda^2(\lambda+1)r^3,
\ear
\bear \label{coefQ}
B&=&8Q^{*2}+10Q^*(3M+\lambda r)r\nn
&+&r^2[36M^2+2M(11\lambda-3)r+\lambda(\lambda-6)r^2].
\ear
In the limiting case $Q^*=0$ one recovers the Zerilli potential for the polar gravitational perturbations of the Schwarzschild black hole~\cite{Zerilli1970PRL}.
\begin{figure}[th!.]
\centering
\includegraphics[width=0.43\textwidth]{Veff.eps}
\caption{\label{fig_veff} Dependence of effective potentials for the scalar (red), electromagnetic (cyan), axial (blue) and polar (black) gravitational perturbations on the tortoise coordinate $x$ ($dx=dr/f(r)$) for the fixed values of the tidal charge parameter $Q^*/M^2=0.5$ and multipole number $l=2$.}
\end{figure}
One can see from Fig.~\ref{fig_veff} that the effective potential for the scalar perturbations is dominant in comparison with the other ones, while axial and polar gravitational perturbations have effective potentials with the smallest height among them. It is well known that an increase in the value of the multipole number $l$ increases the height of the potential~\cite{KokkotasLRR2.1999}.
\section{Quasinormal modes}\label{results}
\subsection{Method and numerical results}
The Schr\"{o}dinger-like wave equations~(\ref{ax-5}) and~(\ref{ax-1}) are solved as usual, imposing appropriate boundary conditions. Considering the wave is purely incoming at the event horizon and outgoing at the spatial infinity:
\bear \label{boundary_condition}
&&Z(r)\sim e^{-i\omega x}\ ,\quad at \quad x\rightarrow 0\ (r\rightarrow r_+)\ ,\nn
&&Z(r)\sim e^{i\omega x}\ ,\quad at \quad x\rightarrow \infty\ (r\rightarrow \infty)\ ,
\ear
\begin{table*}[ht!]
\begin{tabular}{llccccc}\label{tab-1}
& & ~~~& \textit{Scalar perturbations}~~~& ~~~& ~~~& \\
\hline
$n$ ~~~ & $l$ ~~~& $Q^*/M^2=0.1$ ~~~& $Q^*/M^2=0.4$ ~~~& $Q^*/M^2=0.7$ ~~~& $Q^*/M^2=1.0$ ~~~& $Q^*/M^2=2.0$ \\
\hline
0 & 0 & $0.1086-\imath~0.1003$ ~~~& $0.1037-\imath~0.0984$~~~& $0.0996-\imath~0.0965$~~~& $0.0960-\imath~0.0945$ ~~~& $0.0868-\imath~0.0888$ \\
& 1 & $0.2881-\imath~0.0972$ ~~~& $0.2755-\imath~0.0956$~~~& $0.2649-\imath~0.0940$~~~& $0.2557-\imath~0.0924$ ~~~& $0.2322-\imath~0.0875$ \\
& 2 & $0.4758-\imath~0.0962$ ~~~& $0.4552-\imath~0.0945$~~~& $0.4378-\imath~0.0928$~~~& $0.4229-\imath~0.0912$ ~~~& $0.3843-\imath~0.0862$ \\
\hline
1 & 1 & $0.2592-\imath~0.3053$ ~~~& $0.2455-\imath~0.3011$~~~& $0.2341-\imath~0.2967$~~~& $0.2245-\imath~0.2922$ ~~~& $0.2003-\imath~0.2783$ \\
& 2 & $0.4557-\imath~0.2941$ ~~~& $0.4342-\imath~0.2893$~~~& $0.4163-\imath~0.2843$ ~~~& $0.4010-\imath~0.2795$ ~~~& $0.3619-\imath~0.2647$ \\
& 3 & $0.6495-\imath~0.2907$ ~~~& $0.6202-\imath~0.2857$ ~~~& $0.5956-\imath~0.2807$ ~~~& $0.5744-\imath~0.2757$ ~~~& $0.5203-\imath~0.2609$ \\
\hline
2 & 2 & $0.4216-\imath~0.5065$ ~~~& $0.3988-\imath~0.4993$~~~& $0.3799-\imath~0.4916$~~~& $0.3640-\imath~0.4839$ ~~~& $0.3241-\imath~0.4602$ \\
& 3 & $0.6219-\imath~0.4936$ ~~~& $0.5915-\imath~0.4856$~~~& $0.5661-\imath~0.4775$ ~~~& $0.5445-\imath~0.4695$ ~~~& $0.4897-\imath~0.4451$ \\
& 4 & $0.8191-\imath~0.4878$ ~~~& $0.7808-\imath~0.4796$~~~& $0.7489-\imath~0.4713$~~~& $0.7215-\imath~0.4631$ ~~~& $0.6517-\imath~0.4385$ \\
\hline
& & ~~~& \textit{EM perturbations}~~~& ~~~& ~~~& \\
\hline
0 & 1 & $0.2436-\imath~0.0920$ ~~~& $0.2318-\imath~0.0900$~~~& $0.2219-\imath~0.0881$~~~& $0.2135-\imath~0.0863$ ~~~& $0.1921-\imath~0.0810$ \\
& 2 & $0.4499-\imath~0.0944$ ~~~& $0.4297-\imath~0.0927$~~~& $0.4128-\imath~0.0909$~~~& $0.3982-\imath~0.0892$ ~~~& $0.3609-\imath~0.0841$ \\
& 3 & $0.6461-\imath~0.0951$ ~~~& $0.6177-\imath~0.0933$~~~& $0.5938-\imath~0.0916$~~~& $0.5732-\imath~0.0899$ ~~~& $0.5204-\imath~0.0849$ \\
\hline
1 & 1 & $0.2091-\imath~0.2925$ ~~~& $0.1959-\imath~0.2874$~~~& $0.1851-\imath~0.2822$~~~& $0.1760-\imath~0.2771$ ~~~& $0.1539-\imath~0.2618$ \\
& 2 & $0.4285-\imath~0.2891$ ~~~& $0.4074-\imath~0.2841$ ~~~& $0.3899-\imath~0.2789$ ~~~& $0.3750-\imath~0.2740$ ~~~& $0.3372-\imath~0.2590$ \\
& 3 & $0.6307-\imath~0.2881$ ~~~& $0.6016-\imath~0.2830$ ~~~& $0.5773-\imath~0.2779$ ~~~& $0.5565-\imath~0.2729$ ~~~& $0.5033-\imath~0.2579$ \\
\hline
2 & 2 & $0.3922-\imath~0.4994$ ~~~& $0.3697-\imath~0.4918$ ~~~& $0.3511-\imath~0.4839$ ~~~& $0.3355-\imath~0.4761$ ~~~& $0.2968-\imath~0.4519$ \\
& 3 & $0.6022-\imath~0.4895$ ~~~& $0.5720-\imath~0.4815$ ~~~& $0.5470-\imath~0.4732$ ~~~& $0.5256-\imath~0.4651$ ~~~& $0.4717-\imath~0.4405$ \\
& 4 & $0.8042-\imath~0.4852$ ~~~& $0.7662-\imath~0.4770$ ~~~& $0.7344-\imath~0.4686$ ~~~& $0.7073-\imath~0.4604$ ~~~& $0.6381-\imath~0.4356$ \\
\hline
\end{tabular}
\caption{Quasinormal frequencies for the scalar and electromagnetic (EM) perturbations for the black hole on the brane for several values of the tidal charge parameter $Q^*/M^2$.}
\end{table*}
\begin{table*}[ht!]
\begin{tabular}{llccccc} \label{tab-2}
& & ~~~& \textit{Axial perturbations}~~~& ~~~& ~~~& \\
\hline
$n$ ~~~ & $l$ ~~~& $Q^*/M^2=0.1$ ~~~& $Q^*/M^2=0.4$ ~~~& $Q^*/M^2=0.7$ ~~~& $Q^*/M^2=1.0$ ~~~& $Q^*/M^2=2.0$ \\
\hline
0 & 2 & $0.3673-\imath~0.0883$ ~~~& $0.3508-\imath~0.0867$ ~~~& $0.3370-\imath~0.0850$ ~~~& $0.3250-\imath~0.0835$ ~~~& $0.2944-\imath~0.0788$ \\
& 3 & $0.5896-\imath~0.0921$ ~~~& $0.5637-\imath~0.0904$ ~~~& $0.5419-\imath~0.0887$ ~~~& $0.5231-\imath~0.0870$ ~~~& $0.4750-\imath~0.0821$ \\
& 4 & $0.7960-\imath~0.0936$ ~~~& $0.7613-\imath~0.0919$ ~~~& $0.7320-\imath~0.0901$~~~& $0.7069-\imath~0.0885$ ~~~& $0.6421-\imath~0.0835$ \\
\hline
1 & 2 & $0.3394-\imath~0.2720$ ~~~& $0.3215-\imath~0.2673$ ~~~& $0.3065-\imath~0.2627$~~~& $0.2938-\imath~0.2582$ ~~~& $0.2619-\imath~0.2448$ \\
& 3 & $0.5725-\imath~0.2797$ ~~~& $0.5459-\imath~0.2746$~~~& $0.5237-\imath~0.2695$ ~~~& $0.5047-\imath~0.2646$ ~~~& $0.4561-\imath~0.2498$ \\
& 4 & $0.7832-\imath~0.2827$ ~~~& $0.7480-\imath~0.2776$ ~~~& $0.7185-\imath~0.2724$ ~~~& $0.6931-\imath~0.2675$ ~~~& $0.6281-\imath~0.2526$ \\
\hline
2 & 2 & $0.2904-\imath~0.4756$ ~~~& $0.2696-\imath~0.4692$ ~~~& $0.2526-\imath~0.4626$ ~~~& $0.2385-\imath~0.4562$ ~~~& $0.2040-\imath~0.4363$ \\
& 3 & $0.5409-\imath~0.4765$ ~~~& $0.5130-\imath~0.4685$ ~~~& $0.4900-\imath~0.4602$ ~~~& $0.4703-\imath~0.4522$ ~~~& $0.4209-\imath~0.4280$ \\
& 4 & $0.7589-\imath~0.4773$ ~~~& $0.7227-\imath~0.4690$ ~~~& $0.6925-\imath~0.4606$ ~~~& $0.6667-\imath~0.4525$ ~~~& $0.6011-\imath~0.4279$ \\
\hline
& & ~~~& \textit{Polar perturbations}~~~& ~~~& ~~~& \\
\hline
0 & 2 & $0.3665-\imath~0.0881$ ~~~& $0.3477-\imath~0.0855$ ~~~& $0.3323-\imath~0.0830$ ~~~& $0.3193-\imath~0.0806$ ~~~& $0.2875-\imath~0.0732$ \\
& 3 & $0.5889-\imath~0.0921$ ~~~& $0.5614-\imath~0.0901$ ~~~& $0.5384-\imath~0.0882$ ~~~& $0.5188-\imath~0.0864$ ~~~& $0.4687-\imath~0.0811$ \\
& 4 & $0.7955-\imath~0.0936$ ~~~& $0.7596-\imath~0.0917$ ~~~& $0.7295-\imath~0.0899$ ~~~& $0.7037-\imath~0.0882$ ~~~& $0.6375-\imath~0.0831$ \\
\hline
1 & 2 & $0.3392-\imath~0.2712$ ~~~& $0.3202-\imath~0.2628$ ~~~& $0.3051-\imath~0.2539$ ~~~& $0.2931-\imath~0.2444$ ~~~& $0.2697-\imath~0.2072$ \\
& 3 & $0.5719-\imath~0.2794$ ~~~& $0.5438-\imath~0.2737$ ~~~& $0.5204-\imath~0.2681$ ~~~& $0.5005-\imath~0.2628$ ~~~& $0.4502-\imath~0.2470$ \\
& 4 & $0.7827-\imath~0.2826$ ~~~& $0.7463-\imath~0.2772$ ~~~& $0.7159-\imath~0.2718$ ~~~& $0.6899-\imath~0.2667$ ~~~& $0.6235-\imath~0.2513$ \\
\hline
2 & 2 & $0.2920-\imath~0.4745$ ~~~& $0.2717-\imath~0.4598$ ~~~& $0.2561-\imath~0.4418$ ~~~& $0.2439-\imath~0.4201$ ~~~& $0.2297-\imath~0.3085$ \\
& 3 & $0.5403-\imath~0.4761$ ~~~& $0.5111-\imath~0.4671$ ~~~& $0.4870-\imath~0.4580$ ~~~& $0.4667-\imath~0.4493$ ~~~& $0.4158-\imath~0.4234$ \\
& 4 & $0.7584-\imath~0.4771$ ~~~& $0.7211-\imath~0.4684$ ~~~& $0.6900-\imath~0.4596$ ~~~& $0.6636-\imath~0.4511$ ~~~& $0.5967-\imath~0.4258$ \\
\hline
\end{tabular}
\caption{Same as Table I but for the axial and polar gravitational perturbations.}
\end{table*}
\begin{figure*}[th!.]
\centering
\includegraphics[width=0.32\textwidth]{QNM_axial_polar.eps}
\includegraphics[width=0.32\textwidth]{QNM_axial_polar1.eps}
\includegraphics[width=0.32\textwidth]{QNM_axial_polar2.eps}
\caption{\label{fig_QNM} Here, $l=2$, $n=0$ quasinormal modes of the scalar ($\blacksquare$), electromagnetic ($\bigstar$), axial ($\circ$) and polar ($\bullet$) gravitational perturbations of the black hole on the brane with the change of the tidal charge parameter $Q^*/M^2$, where the red spot corresponds to the ones of the Schwarzschild black holes.}
\end{figure*}
Solving the Schr\"{o}dinger-like wave equations with the effective potentials (\ref{scalar-potential}), (\ref{em-potential}), (\ref{ax-potential2}) and (\ref{polar-potential2}) analytically is impossible. Therefore, we use the WKB method that was applied for
the first time for calculation of
quasinormal modes of black holes
by Schutz and Will~\cite{SchutzAJL33.1985}. Afterwards, to increase its accuracy of
the method was extended up to the
third order by Iyer and Will~\cite{IyerPRD3621.1987} and
up to the sixth order by Konoplya~\cite{KonoplyaPRD024018.2003}. The sixth order WKB method for solving
the Schr\"{o}dinger-like wave equation governing the quasinormal modes of black holes implies the relation
\bear \label{wkb}
\frac{i\left(\omega^2-V(r_0)\right)}{\sqrt{-2V''(r_0)}}+ \sum_{j=2}^6\Lambda_j=n+\frac{1}{2}\ ,
\ear
where $r_0$ is the value of the radial coordinate $r$ corresponding to the maximum of the potential $V(r)$, $j$ is the order of the WKB corrections and $\Lambda_j$ is the correction term corresponding to the $j$th order. One can find the expressions of $\Lambda_j$ in~\cite{IyerPRD3621.1987,KonoplyaPRD024018.2003}. The prime $"'"$ stands for the derivative with respect to the tortoise coordinate $x$, and $n$ is the overtone number.
In Tables I and II quasinormal frequencies of the scalar, electromagnetic, axial and polar gravitational perturbations of the braneworld black hole are given.
One can see from the numerical results presented in Tabs I, II and Fig.~\ref{fig_QNM} that an increase in the value of the tidal charge parameter decreases the frequency of the real oscillations and the damping rate. Moreover, with increasing multipole number $l$ the frequency of the oscillations increases while the damping rate decreases. This means that quasinormal frequencies with higher multipole numbers are longer lived. However, from the general characteristics of the overtone number $n$, we know that an increase of the value of the overtone number situation completely changes to the contrary, i.e. it implies the wave with lower oscillation frequency and higher damping rate.
\textit{\textbf{Evolution of the wave.}} The temporal evolution of the gravitational perturbations is governed by the equation
\bear \label{bwe1}
-\frac{\partial^2\Phi}{\partial t^2}+\frac{\partial^2\Phi}{\partial x^2}=V \Phi\ .
\ear
By turning into the null coordinates $u\equiv t-x$ and $v\equiv t+x$, one can rewrite the wave equation~(\ref{bwe1}) in the form
\bear \label{bwe2}
-4\frac{\partial^2\Phi}{\partial u\partial v}=V(x)\Phi\ ,
\ear
Discretization of Eq.~(\ref{bwe2}) allows us to calculate the values of wave function $\Phi$~\cite{WangPRD064025.2004}. We follow the method for numerical calculations presented in~\cite{ChirentiCQG4191.2007}.
\begin{figure}[h!.]
\centering
\includegraphics[width=0.43\textwidth]{time_domain1.eps}
\includegraphics[width=0.43\textwidth]{time_domain2.eps}
\caption{\label{time_domain} Semilog graphs of the absolute value of the wave function for the axial gravitational perturbations of the braneworld black holes evaluated at $x_0=10$ and $M=1$. Top panel: Different values of the tidal charge parameter $Q^*=1$ (blue curve), and $Q^*=2$ (red curve) in comparison with the Schwarzschild one $Q^*=0$ (black curve) in the fundamental mode $l=2$. Bottom panel: Different values of the multipole number $l=2$ (black curve), $l=3$ (blue curve), and $l=4$ (red curve).}
\end{figure}
In Fig.~\ref{time_domain} the time-domain profiles for the evolution of the axial gravitational perturbations\footnote{Scalar, electromagnetic and polar gravitational perturbations give qualitatively the same results. Therefore, we have shown the behavior only for the axial gravitational perturbations.} in the fundamental mode are presented for the braneworld black hole with mass $M=1$ at the radius $x_0=10$ in comparison with the Schwarzschild black hole. One can see in these figures a monotonic decay of the signal; the quasinormal mode signal dominates after $t\approx40$. From the top panel of Fig. \ref{time_domain} we see that with increasing the tidal charge parameter damping rate of the signal decreases and the duration of dominance of the quasinormal mode signal decreases relative to the Schwarzschild black hole. Moreover, the bottom panel shows that an increasing multipole number $l$ increases the damping and the real oscillations time scales.
\textbf{\textit{Massive scalar field.}} In the case of a massive scalar field around the black hole the region of the values of the parameters allowing for occurrence of the quasinormal modes is restricted~\cite{OhashiCQG3973.2004}. For small enough values of the scalar field mass parameter $m$, the effective potential $V(r)$ is in the form of a barrier. However, with increasing the value of $m$, the asymptotical value of $V(r)$ (namely $m^2$) increases more rapidly than the peak of the potential [because of $V(r\rightarrow\infty)\rightarrow m^2$]. Consequently, in the case of $V(r_{0})\leq m^2$, the effective potential is not in the barrier form anymore. Therefore, in that case quasinormal modes do not occur. Below in Figs.~\ref{mass_limit} these region of parameters are shown.
\begin{figure}[h!.]
\centering
\includegraphics[width=0.43\textwidth]{mass_limit1.eps}
\includegraphics[width=0.43\textwidth]{mass_limit2.eps}
\caption{\label{mass_limit} Possible values of the scalar field mass $m$ that give the limit of existence of quasinormal modes (QNMs) in the field of black hole on the brane. Shaded region represents part of parameter space where the QNM can occur. In the white region there are no QNM. Top panel illustrates for the mass of the scalar field $m$ versus multipole number $l$ for the fixed value of the tidal charge parameter $Q^*/M^2=0.5$. In the bottom panel also limiting values of the scalar field mass $m$ enabling existence of QNMs in the field of the black hole in the braneworld are given in dependence on the $Q^*/M^2$ in the fundamental mode $l=0$.}
\end{figure}
Contrary to the case of the standard Reissner-Nordstr\"{o}m black hole, in the case of the braneworld black hole with increasing value of the tidal charge parameter, maximum possible value of the mass of the scalar field decreases. This is connected with the increase of the horizon radius. From Fig.~\ref{mass_limit} one can see that with increasing multipole number $l$ the range of the possible values of the mass parameter increases.
\subsection{Quasinormal frequencies in the large multipole number limit}
It is known that the WKB method has very good accuracy for large values of the multipole number $l$. In the large multipole number limit one can solve the wave equation analytically by using the first order WKB approximation. To do this, we expand the expression (\ref{wkb}) in powers of $1/l$, and quasinormal frequencies tend to finite values. Interestingly, for the both scalar and electromagnetic perturbations, as well as axial and polar gravitational perturbations, eikonal limits are the same and read
\bear \label{eikonal1}
&&\omega_r\approx\frac{\sqrt{r_0(r_0-2M)-Q^*}}{r_0^2}\left(l+\frac{1}{2}\right)\ ,\\
&&\omega_i\approx-\frac{1}{r_0^3}\left[r_0^2(-3r_0^2+20r_0M-30M^2)\right.\nn
&&\left.+3Q^*(5r_0-14M)r_0-14Q^{*2}\right]^{1/2}\times\left(n+\frac{1}{2}\right),
\ear
where
\bear \label{max_r}
r_0\approx\frac{3M+\sqrt{9M^2+8Q^*}}{2}\ .
\ear
Therefore, in the large multipole number $l$ case it is almost impossible to distinguish the types of perturbations. Notice correspondence of $r_0$ and the radius of photon circular orbit~\cite{StuchlikAPS363.2002,ScheeIJMPD983.2009}
\begin{figure}[h!.]
\centering
\includegraphics[width=0.43\textwidth]{Large_l.eps}
\caption{\label{large_l} Dependence of the real (black curve) and imaginary (blue curve) parts of the quasinormal frequencies on the tidal charge parameter $Q^*/M^2$ in the large multipole number $l$ limit.}
\end{figure}
One can see from Fig.~\ref{large_l} that in the limit of large multipole number $l$ real and imaginary parts of the quasinormal frequencies tend to finite values.\footnote{Real part of the quasinormal frequencies tends to zero slightly more rapidly than the imaginary part.} Since there is no upper limit on the value of the tidal charge $Q^*/M^2$, in the eikonal limit for large values of the tidal charge parameter quasinormal frequencies tend to zero.
\subsection{Stability}
Calculations have shown that imaginary part of the quasinormal frequencies is always negative, $Im(\omega)<0$. Moreover, effective potentials (\ref{scalar-potential}), (\ref{em-potential}), (\ref{ax-potential2}) and (\ref{polar-potential2}) are always positive-definite. Furthermore, from Fig. \ref{time_domain} of time-domain profiles for the evolution of axial gravitational perturbations one can see that there is no indication of instability. One of the most important results of this paper is that in calculations made by the WKB method and the form of the wave in time domain profile, we have not observed any unstable mode. From these we can conclude that black holes on the brane (\ref{metric}) are stable against scalar, electromagnetic and gravitational perturbations.
\section{Scattering}\label{scattering}
In this section we study the greybody factor for the braneworld black holes. The greybody factor is understood as the probability for an outgoing wave in $\omega$-mode to reach infinity or, equivalently, the absorption probability for an incoming wave in $\omega$-mode to be absorbed by the black hole~\cite{CardosoPRL071301.2006,KonoplyaPLB199.2010}. In other words, the greybody factor is the tunneling probability of the wave through the barrier determined by the effective potential in the given black hole spacetime.
In order to calculate the greybody factor, we write the boundary conditions for the ingoing and outgoing waves that are the solution of the Schr\"{o}dinger-like wave equations in the asymptotic form
\begin{eqnarray}\label{svi1}
&&Z_\omega=e^{-i\omega x}+R(\omega) e^{i\omega x}, \quad at \quad x\rightarrow+\infty,\nonumber\\
&&Z_\omega=T(\omega)e^{-i\omega x}, \quad \quad \quad \quad at \quad x\rightarrow-\infty,
\end{eqnarray}
where $R(\omega)$ and $T(\omega)$ are the reflection and transmission coefficients, respectively. We write the above given boundary conditions~(\ref{svi1}) for $\omega\rightarrow-\omega$:
\begin{eqnarray}\label{svi2}
&&Z_{-\omega}=e^{i\omega x}+R(-\omega) e^{-i\omega x}, \quad at \quad x\rightarrow+\infty,\nonumber\\
&&Z_{-\omega}=T(-\omega)e^{i\omega x}, \quad \quad \quad \quad at \quad x\rightarrow-\infty.
\end{eqnarray}
From these two boundary conditions we can write the expression for the flux~\cite{ChoudhuryPRD064033.2004}
\begin{eqnarray}\label{svi3}
J=\frac{1}{2i}\left[Z_{-\omega}\frac{dZ_{\omega}}{dx}- Z_{\omega}\frac{dZ_{-\omega}}{dx}\right].
\end{eqnarray}
Using the boundary conditions~(\ref{svi1}) and~(\ref{svi2}), we find the relation between the reflection and transmission coefficients from the flux conservation in the form
\begin{eqnarray}\label{svi4}
R(\omega)R(-\omega)+T(\omega)T(-\omega)=1\ .
\end{eqnarray}
If we consider $\omega$ real ($\omega\in \mathbb{R}$), we can write that $Z_{-\omega}=Z_{\omega}^\ast$, $R(-\omega)=R^\ast(\omega)$ and $T(-\omega)=T^\ast(\omega)$. Then the relation~(\ref{svi4}) can be written in the form
\begin{eqnarray}\label{svi5}
|R|^2+|T|^2=1\ .
\end{eqnarray}
A wave with frequency $\omega$ larger than the height of the potential barrier $V_0$ will not be (classically) reflected by the barrier. Therefore, in this case reflection coefficient is close to zero. The incoming wave with smaller frequency than the height of the potential barrier, $\omega^2 < V_0$, is reflected partly, while the rest part is transmitted through the barrier by tunneling effect, depending on the values of $\omega$ and $V_0$. This is why this case is more interesting to study.
First, we consider the small frequency case when $\omega^2$ is much less than the height of the potential barrier $V_0$ ($\omega^2\ll V_0$). For such small values of $\omega^2$, the WKB approximation has not high accuracy because of the large distance between two turning points. In this case transmission coefficient is given by the well-known formula
\begin{eqnarray}\label{svi6}
T=e^{-\int_{x_1}^{x_2}dx\sqrt{V(r_\ast)-\omega^2}}\ , \quad as \quad \omega^2\ll V_0\ ,
\end{eqnarray}
and the reflection coefficient
\begin{eqnarray}\label{svi7}
R=\left(1-e^{-2\int_{x_1}^{x_2}dx\sqrt{V(r_\ast)-\omega^2}}\right)^{1/2}, \quad as \quad \omega^2\ll V_0\ .\nonumber\\
\end{eqnarray}
The radii $x_1$ and $x_2$ are the classical turning points that are the solutions of the equation $V(x)-\omega^2=0$. At the small values of $\omega$, transmission coefficient is close to zero while the reflection coefficient is close to one~\cite{BoonsermPRD101502.2008}.
The second limit case is that the value of $\omega^2$ is of the same order as the maximum height of the potential barrier $V_0$, i.e. $\omega^2\approx V_0$. Here, we can calculate the greybody factor by using the sixth order WKB approximation because of the small distance between the turning points~\cite{KonoplyaPLB199.2010,ToshmatovPRD083008.2015}. Then we arrive at
\begin{figure*}[th!.]
\centering
\includegraphics[width=0.32\textwidth]{GBF_scalar.eps}
\includegraphics[width=0.32\textwidth]{GBF_em.eps}
\includegraphics[width=0.32\textwidth]{GBF_axial.eps}
\caption{\label{GBF} Reflection coefficients for (from left to right) the scalar, electromagnetic and gravitational fields $l=2$ mode of the black hole in the braneworld for several values of the tidal charge parameter $Q^*/M^2$: $Q^*/M^2=0.6$ - dashed, $Q^*/M^2=1.2$ - dot-dashed and $Q^*/M^2=2$ - dotted curves. Where, in order to compare the results we have shown the reflection coefficients of the Reissner-Nordstr\"{o}m (grey solid curve) and Schwarzschild (black solid curve) black holes.}
\end{figure*}
\begin{eqnarray}\label{svi8}
R=\left(1+e^{-2i\pi(n+1/2)}\right)^{-1/2},
\end{eqnarray}
where $(n+1/2)$ is given by the formula of the sixth order WKB method~(\ref{wkb}). From~(\ref{svi5}) with~(\ref{svi8}) one can write the expression for the transmission coefficient as
\begin{eqnarray}\label{svi9}
|T|^2=1-\left|\left(1+e^{-2i\pi(n+1/2)}\right)^{-1/2}\right|^2.
\end{eqnarray}
In Fig.~\ref{GBF}, we give the reflection coefficients for the scalar, electromagnetic and axial gravitational fields in the mode $l=2$ for the braneworld black holes. Reflection and transmission coefficients for the axial and polar gravitational perturbations are almost the same, i.e. it is impossible to distinguish difference in figures.
One can see from Fig.~\ref{GBF} that for the scalar perturbative fields, probability of the wave reflection by the potential barrier is larger than for the electromagnetic and gravitational ones. Moreover, unlike the case of the standard Reissner-Nordstr\"{o}m black hole, the presence of the non-vanishing tidal charge parameter decreases the reflection ability of the potential barrier, while it increases the transmission probabilities in comparison with the Reissner-Nordstr\"{o}m and Schwarzschild ones.
\section{Absorption cross section of planar massless scalar waves}\label{absorption}
Absorption and scattering of test particles and fields in the black hole backgrounds are very important since they are relevant for observations related to accretion processes, deflection of light, and etc. So far, absorption and scattering properties of waves by various static spherically symmetric \cite{DasPRL417.1997,AnderssonPRD1808.1995,MacedoPRD064001.2014,CrispinoPRD124038.2010, DecaniniPRD044032.2011,BenonePRD104053.2014} and axially symmetric black holes~\cite{GlampedakisCQG1939.2001} were studied. Here we extend these studies for the analysis of the absorption cross sections of massless scalar field by braneworld black holes. The massless scalar field is defined by the equation~(\ref{ax-5}) with potential~(\ref{scalar-potential}).
In order to demonstrate the effect of the tidal charge we compare the results of the braneworld black holes with the Reissner-Nordstr\"{o}m black holes calculated in~\cite{MacedoPRD064001.2014}.
\textbf{\textit{Partial wave approach.}} In the partial wave approach we should solve the Schr\"{o}dinger-like wave equation~(\ref{ax-5}) with the scalar field potential (\ref{scalar-potential}) by using the boundary conditions (\ref{svi1}). For such scalar fields the total absorption cross section is defined by the sum of the partial absorption cross sections of the planar massless scalar waves
\bear\label{abs1}
\sigma_{abs}(\omega)=\sum_{l=0}^{+\infty}\sigma_{l}(\omega),
\ear
where $\sigma_{l}(\omega)$ represents the partial absorption cross section corresponding to the wave with angular momentum $l$ and energy $\omega$, given by
\bear\label{abs2}
\sigma_{l}(\omega)=\frac{\pi}{\omega^2}(2l+1)|T_l(\omega)|^2.
\ear
$T_l(\omega)$ is the transmission (absorption) coefficient for the wave with energy $\omega$ and angular momentum $l$. It is known from the classical mechanics that $|T_l(\omega)|^2$ is the absorption probability. By using the higher order WKB method one can write (\ref{abs2}) as
\bear\label{abs3}
\sigma_{l}(\omega)=\frac{\pi}{\omega^2}(2l+1)\left[1-\left|\left(1+e^{-2i\pi(n+1/2)}\right)^{-1/2}\right|^2\right],
\ear
where $(n+1/2)$ is given by the expression of the higher order WKB formulas (\ref{wkb}).
\textbf{\textit{High-energy limit.}} In the high-energy scale, the wavelength is almost negligible relative to the horizon scale of the black hole. Therefore, in this regime massless scalar waves propagate along the null geodesics \cite{DecaniniPRD024031.2010}. Therefore, one may use the classical capture cross section of the light by the black holes as the geometric cross section of the light rays. To do so, we consider the motion of the massless particle (photon) around the black hole confined to the equatorial plane ($\theta=\pi/2$),
\bear\label{abs4}
\dot{r}^2=E^2-V_{eff}, \qquad V_{eff}=f\frac{L^2}{r^2}\ ,
\ear
where $E$ and $L$ are energy and angular momentum of the photon, respectively, which are conserved quantities due to symmetries of the spacetime. However, for the photon motion only the impact parameter $b=L/E$ is relevant~\cite{MTW1973,StuchlikCQG215017.2010}. The geometric cross section of the light rays is given by the expression $\sigma_{geo}=\pi b_{ps}^2$, where $b_{ps}$ is the critical impact parameter of the light defined by the ratio of angular momentum and energy of the photon moving along the circular photon orbit as $b_{ps}=L_{ps}/E_{ps}$. Therefore,
\bear\label{abs5}
\sigma_{geo}=\pi b_{ps}^2=2\pi\frac{r_{ps}}{f'(r_{ps})} ,
\ear
where "$'$" means the derivative with respect to $r$, and $r_{ps}$ is the radius of the photon sphere which is found from the equation
\bear\label{abs6}
2f-rf'=0\ .
\ear
The radius of the photon sphere for the braneworld black hole is given by the expression~\cite{ScheeIJMPD983.2009}
\bear\label{abs7}
r_{ps}=\frac{3M+\sqrt{9M^2+8Q^*}}{2}
\ear
\begin{figure}[h!.]
\centering
\includegraphics[width=0.43\textwidth]{ps.eps}
\caption{\label{ps} The loci of the event horizons (dashed) and photon spheres (solid) of the black hole in braneworld (black) and Reissner-Nordstr\"{o}m black holes (blue).}
\end{figure}
In Fig. \ref{ps} the loci of the event horizons and photon spheres of the braneworld and Reissner-Nordstr\"{o}m black holes are shown for comparison. One can see from Fig. \ref{ps} that event horizon and photon spheres of the braneworld black holes never vanish for any values of the tidal charge parameter.
By inserting (\ref{abs7}) into (\ref{abs5}) we obtain expression for the geometric absorption cross section of the massless scalar wave by the braneworld black hole in the form
\bear\label{abs8}
\sigma_{geo}=\pi \frac{\left(3M+\sqrt{9M^2+8Q^*}\right)^4}{8(3M^2+2Q^*+M\sqrt{9M^2+8Q^*})} .
\ear
\begin{figure}[h!.]
\centering
\includegraphics[width=0.43\textwidth]{geo_cs.eps}
\caption{\label{geo_cs} The geometric absorption cross section by black hole in braneworld (black) and the Reissner-Nordstr\"{o}m black hole (blue).}
\end{figure}
In Fig. \ref{geo_cs} the geometric absorption cross section of the massless scalar wave by the braneworld black hole and the Reissner-Nordstr\"{o}m black holes are shown. One can see that unlike the case of the Reissner-Nordstr\"{o}m black hole, with increasing value of the tidal charge parameter the geometric absorption cross section of the massless scalar wave increases.
In the paper from~\cite{DecaniniPRD044032.2011} it has been shown that there are fluctuations (regular oscillations) of the high-energy (frequency) absorption cross section around the limiting value of the geometric cross section; this is a universal property of the absorption cross section of scalar field in the high-energy regime in the field of spherically symmetric black holes. This oscillatory part of the absorption cross section of the massless scalar waves can be written as
\bear\label{abs9}
\sigma_{osc}(\omega)=-8\pi b_c\lambda\sigma_{geo}e^{-\pi b_c\lambda}\sinc(2\pi b_c\omega),
\ear
where $\sinc(x)=sin(x)/x$ and $\lambda$ is the Lyapunov exponent used for analysis of the instability of the null geodesics~\cite{CardosoPRD064016.2009}. The total absorption cross section of the massless scalar waves is the sum of the geometric (\ref{abs5}) and oscillatory (\ref{abs9}) cross sections
\bear\label{abs10}
\sigma_{abs}\approx \sigma_{geo}+\sigma_{osc}\ .
\ear
\begin{figure}[h!.]
\centering
\includegraphics[width=0.43\textwidth]{abs_cs.eps}
\caption{\label{abs_cs} Absorption cross section of massless scalar waves by the braneworld for the values of the tidal charge parameter $Q^*/M^2=0.5$ (cyan), $Q^*/M^2=1.0$ (red) and $Q^*/M^2=1.5$ (magenta). The Schwarzschild (black) and Reissner-Nordstr\"{o}m with electric charge $q/M=0.5$ (blue) black holes are shown for comparison, where horizontal lines represent the geometric absorption cross sections.}
\end{figure}
In Fig. \ref{abs_cs} we show the total absorption cross sections of the massless scalar waves by the braneworld black holes for several values of the tidal charge parameter. For comparison, total absorption cross section obtained for the Schwarzschild and Reissner-Nordstr\"{o}m black holes are also shown. One can see from Fig. \ref{abs_cs} that with increasing tidal charge parameter the absorption cross section increases. By comparing the absorption cross sections to those related to the the Schwarzschild and Reissner-Nordstr\"{o}m black holes, we can conclude $\sigma_{abs,RN}<\sigma_{abs,Schw}<\sigma_{abs,Br}$.
\textbf{\textit{Low-energy limit.}} In the papers from \cite{DasPRL417.1997,MacedoPRD064001.2014} it has been shown that for small values of the massless
scalar wave frequency the absorption cross section by a black hole tends to the horizon area of the black hole ($\sigma_0(\omega\rightarrow0)\rightarrow4\pi r_+^2$). We know from (\ref{horizon}) that the horizon of the braneworld black hole always increases with increasing tidal charge. Then, in the low-frequency regime the absorption cross section of the massless scalar waves also increases appropriately to the increasing of the horizon.
\textbf{\textit{Scalar particle emission by Hawking radiation.}} As already stated by Hawking~\cite{HawkingCMP199.1975}, the particle absorption cross section is very relevant to the particle emission by black holes~\cite{CasalsJHEP071.2008}. The particle emission rate by Hawking radiation is defined by the number of emitted particles by the black hole per unit time and per unit frequency. One should take into consideration also the spin of the emitted particles. Here, we are considering a massless scalar field. It is well known that scalar particles have spin-zero (bosons) and their emission probabilities are defined by the Bose-Einstein distribution. Then, the particle emission rate (or emission spectrum) by Hawking scalar radiation reads
\bear\label{hawking_emission}
\frac{d^2N(\omega)}{d\omega dt}=\frac{1}{2\pi}\sum_{l=0}^{+\infty}\frac{(2l+1)|T_l(\omega)|^2}{e^{\omega/T}-1}= \frac{\omega^2}{2\pi^2}\frac{\sigma_{abs}(\omega)}{e^{\omega/T}-1}
\ear
where $T$ is the Hawking temperature given by (\ref{temperature}). In Fig.~\ref{emission} we represent the massless scalar particle emission spectrum by the braneworld black holes in comparison with the Reissner-Nordstr\"{o}m and Schwarzschild black holes.
\begin{figure}[h!.]
\centering
\includegraphics[width=0.43\textwidth]{emission.eps}
\caption{\label{emission} Massless scalar particle emission spectrum by the braneworld for the values of the tidal charge parameter $Q^*/M^2=0.5$ (cyan), $Q^*/M^2=1.0$ (red), and $Q^*/M^2=1.5$ (magenta) in comparison with the Schwarzschild (black) and Reissner-Nordstr\"{o}m with electric charge $q/M=0.5$ (blue) black holes.}
\end{figure}
One can see from Fig.~\ref{emission} that unlike the case of the Reissner-Nordstr\"{o}m black holes, an increase of the tidal charge parameter implies increase of the particle emission rate. Furthermore, with increasing frequency the particle emission spectrum rises up to peak and falls to zero rapidly. At the frequencies $\omega M>0.3$, emission spectrum is almost zero regardless the value of the tidal charge parameter. From Fig.~\ref{emission} one can deduce that at the high-frequency regime distinction of types of black holes is impossible. From this viewpoint, the low-frequency regime is more favorite to distinguish various black holes.
\section{Summary}\label{summary}
We have studied the scalar, electromagnetic, axial and polar gravitational perturbations of the Reissner-Nordstr\"{o}m black hole (with tidal charge instead of electric charge) in the Randall-Sundrum braneworld. By using the sixth order WKB method, we have calculated the quasinormal frequencies of these perturbations. Results have shown that with increasing tidal charge parameter the frequency of the real oscillations decreases while damping rate increases, unlike the case of the standard Reissner-Nordstr\"{o}m black hole.
Moreover, it has been shown that the current black hole solution localized in the Randall-Sundrum braneworld is stable against scalar, electromagnetic, and gravitational perturbations.
The reflection probability of the wave by the scalar potential barrier is larger than for the electromagnetic and gravitational ones; i.e., scalar fields are the most favorite in terms of reflection of waves, while the gravitational fields are the most favorite in terms of transmission of waves. With increasing tidal charge parameter reflection abilities of the perturbative fields decrease, as the radii of the horizon and the photon sphere increase with the tidal charge.
However, we should note that because of no upper limit on the value of the tidal charge parameter, black holes on the brane always have event horizons and photon spheres which are located further away from the central object in comparison to the Schwarzschild ones. We have studied also the absorption cross section of the massless scalar waves by the braneworld black hole in the low- and high-frequency regimes. Calculations have shown that the braneworld black holes always have bigger absorption cross sections than the corresponding ones related to the Schwarzschild and Reissner-Nordstr\"{o}m black holes. One of the main results of this paper is that in the high-frequency regime distinction of the black holes from the particle emission spectrum is almost impossible. From this point of view the low-frequency regime is more significant.
\begin{acknowledgments}
Authors would like to thank the anonymous referee for his careful reading of our manuscript and many insightful comments and suggestions that have improved the paper. The authors would like to thank Naresh Dadhich for his useful discussions on the properties of the black hole solution on the brane and Luciano Rezzolla for noting a problem with our calculations in a previous version of the manuscript. B.T., Z.S., and J.S. would like to express their acknowledgments for the institutional support of the Faculty of Philosophy and Science of the Silesian University in Opava, the internal student grant of the Silesian University (Grant No. SGS/23/2013) and the Albert Einstein Centre for Gravitation and Astrophysics under the Czech Science Foundation (Grant No.~14-37086G). B.A. acknowledges the Faculty of Philosophy and Science, Silesian University in Opava, Czech Republic, and the Goethe University, Frankfurt am Main, Germany, for their warm hospitality. The research of B.A. is supported in part by Projects No. F2-FA-F113, No. EF2-FA-0-12477, and No. F2-FA-F029 of the UzAS, and by the ICTP through Grants No. OEA-PRJ-29 and No. OEA-NET-76 and by the Volkswagen Stiftung, Grant No.~86 866.
\end{acknowledgments}
|
\section{Introduction}
Materials with low symmetry exhibit a large variety of intriguing optical phenomena. As is well known, the absence of inversion symmetry leads to the occurrence of natural optical activity. In addition to this, inversion symmetry breaking strongly affects the electric dipole selection rule, and thus allows for direct excitation of local crystal-field excitations in linear spectroscopy.
In magnetic materials, the breaking of time-reversal symmetry leads to the magneto-optical Kerr and Faraday effect. Many of the optical phenomena originating from
a reduced symmetry have technological applications, examples of
which include optical insulators and
magneto-optical storage devices. \cite{zvezdin}
In recent years, materials lacking both spatial inversion and time-reversal symmetry have been a focal point of condensed matter research. Their combined absence may for instance lead to magneto-electric and multiferroic behavior,\cite{multiferroicreviewtokura2014, khomskii2009, Khomskii2006} the formation of chiral and skyrmion magnetic ground states,\cite{tokura2013} and the occurrence of toroidal order, \cite{spaldin2008,zimmermann2014ferroic} and excitations. \cite{Kaelberer2010,papasimakis2016toroidal} Effects on the optical properties have been studied extensively.\cite{krichevtsov1993,pimenov2006possible,rovillain2012} One intriguing optical phenomenon originating from the combined absence of time-reversal and space-inversion symmetry is the so-called nonreciprocal directional dichroism.\cite{szaller2013,kezsmarki2014,toyoda2015}
The high sensitivity of optical properties to symmetry breaking may be used to gain a better understanding of the underlying material properties. Here we focus on the cuprate material Cu$_{2}$OSeO$_{3}$\@\xspace, belonging to the class of non-centrosymmetric cubic crystal structures. These materials have recently triggered a great deal of research interest owing to the occurrence of topologically protected spin-vortex-like structures, known as skyrmions.\cite{tokura2013}
In these chiral crystal structures, the absence of inversion symmetry between the spin sites leads to a non-vanishing antisymmetric Dzyaloshinskii-Moriya (DM) exchange interaction, which competes with the energetically stronger isotropic Heisenberg exchange. This combination of exchange interactions stabilizes spin helices.\cite{bak1980} In the presence of an external magnetic field, the Zeeman interaction energy stabilizes the formation of a topologically robust hexagonal lattice of nanometer-sized skyrmions.\cite{roessler2011}
In this article we report on a variety of linear optical properties \cite{glazer} associated with the broken inversion symmetry, structural chirality, and magnetic order in the chiral magnet Cu$_{2}$OSeO$_{3}$\@\xspace.\cite{seki2012} The orbital aspect of Cu$_{2}$OSeO$_{3}$\@\xspace has only received minimal experimental attention so far.\cite{langner2014} Here, we reveal local crystal-field excitations below the charge-transfer gap by means of spectroscopic ellipsometry. These orbital excitations acquire a finite dipole matrix element due to the low crystal-field symmetry. Optical polarization rotation measurements are used to study the structural chirality and to probe the magnetic order. The natural optical activity, resulting from the chiral crystal structure, shows an abrupt change upon magnetic ordering. This observation is evidence of a finite magneto-electric coupling in the phase with helical magnetic order.
The large magneto-optical response is quantified by a magneto-optical susceptibility on the order of $\mathcal{V}$($540\,$nm)\,$\sim$\,10$^{4}$\,rad/T$\cdot$m in the helimagnetic phase, and maximum Faraday rotation of $\sim 165\,$deg/mm in the field-polarized ferrimagnetic phase. This
strong
response serves as an excellent probe for the various magnetic phase transitions. The magneto-optical data allow us to derive the phase diagram of Cu$_{2}$OSeO$_{3}$\@\xspace as a function of magnetic field and temperature. The optically determined phase diagram is in excellent agreement with results obtained by other techniques.\cite{longwavelength,seki2012,ruff2015}
\section{Structure and magnetism}
Cu$_{2}$OSeO$_{3}$\@\xspace has a complex chiral crystal structure with cubic space group $P2_{1}3$. The unit cell contains 16 Cu ions, all having a $2+$ valence state ($d^9$ configuration). These Cu ions are located within two crystallographically distinct oxygen ligand field geometries,\cite{jwbos2008} which can be approximated by a trigonal bipyramid ($D_{3h}$) for Cu-I ions and by a square pyramid ($C_{4v}$) for Cu-II ions, as depicted in Fig.\ \ref{fig:splittings}. The true site symmetries are lower, and are given by $C_{3v}$ for Cu-I and $C_{1}$ for the Cu-II ions.\cite{jwbos2008} These ligand fields lead to different crystal-field splittings of the $3d$ orbitals on the Cu-I and Cu-II sites, as shown in Fig.\ \ref{fig:splittings}. For Cu-I the hole is located in the ${z^2}$ orbital, whereas for Cu-II the hole is in the ${x^2-y^2}$ orbital. Note that these are no pure $3d$ orbitals due to the low site symmetry, i.\,e.\@\xspace, parity is not a good quantum number.
\begin{figure}[t]
\center
\includegraphics[width=3.375in]{splittingsU10.png}
\caption{The $3d$ crystal-field splittings for the Cu-I (trigonal bipyramid) and Cu-II (square pyramid) ions. The dipole-active local crystal-field (CF) excitations are indicated with red arrows, and the charge-transfer excitations with blue arrows. Intersite excitations between the Cu ions, illustrated with a gray arrow, lie outside the experimentally accessible energy range of this work.}
\label{fig:splittings}
\end{figure}
The 16 Cu ions in the unit cell are distributed over 4 Cu-I sites and 12 Cu-II sites and form a network of corner-sharing tetrahedra, where each Cu tetrahedron comprises 1 Cu-I and 3 Cu-II ions. Owing to the different bond lengths between the 16 Cu$^{2+}$ sites in the unit cell, it is possible to make a real space classification of the exchange energy scales into a "strong" and "weak" type.\cite{clustercoso} Through the strong Heisenberg and DM exchange interactions, tetrahedral 3-spin-up-1-down triplet clusters couple far above\cite{clustercoso} the macroscopic ordering temperature $T_C$\,$\approx$\,$58$\,K. These $S$\,=\,$1$ spin clusters turn out to be the relevant low-energy spin entities in Cu$_{2}$OSeO$_{3}$\@\xspace. The weaker inter-tetrahedral Heisenberg and DM exchange
couplings mediate the interactions between the $S$\,=\,$1$ entities, giving rise to long-range helimagnetic order below $T_C$. This separation into inter- and intra-tetrahedral energy scales is well supported by the splitting of the magnon spectra in two well separated energy bands.\cite{portnichenko2016}
Finally, weak cubic magnetic anisotropy terms pin the helimagnetic spirals along the six equivalent
crystallographic $\langle$100$\rangle$ directions, leading to domain formation.\cite{longwavelength,belesi2012,magnetoelectricnature}
In the presence of an external magnetic field, different metamagnetic phases are formed. Applied magnetic fields of a few tenths of mT are enough to fully lift the degeneracy of the helical domains, giving a conical type of order with the propagation vector $q$ along the applied field. Above a second critical field the system is driven into the field-polarized ferrimagnetic phase, where all tetrahedral $S$\,=\,$1$ entities are aligned with the magnetic field. The skyrmion lattice phase (SkL) is located within a narrow field-temperature window just below $T_C$\,$\approx$\,$58$\,K for moderate applied fields of the order of 20\,-\,50\,mT.\cite{longwavelength,seki2012}
\section{Experimental methods}
High-quality single crystals were grown using a standard chemical vapor transport method.\cite{belesi2010ferrimagnetism} The studied crystals were oriented by crystal morphology inspection after which the fine orientation was done by means of a Laue camera. For the ellipsometry study a (100) surface was prepared (sample surface dimensions approximately 2.7\,mm\,$\times$\,2.8\,mm), whereas for the polarimetry study a (111) oriented sample was chosen (surface area 3.6\,mm\,$\times$\,3.2\,mm, used thicknesses $d$\,$\approx$\,1.00\,mm and $d$\,$\approx$\,221\,$\pm$\,3\,$\mu$m).
The samples were polished with Al$_2$O$_3$ suspension ($\approx 1\,\mu$m grain size) in order to obtain optically smooth surfaces.
For the optical spectroscopy part, a Woollam VASE spectroscopic ellipsometer with an autoretarder between source and sample was used. Ellipsometry allows us to obtain the real and imaginary parts of the complex dielectric function in a self normalizing way.
The (100) sample was mounted in a UHV chamber with liquid-He flow cryostat and measured in the range $0.75$ to $5$\,eV at a fixed angle of incidence ($70^\circ$). For the analysis of the ellipsometric data, the surface roughness was estimated using the knowledge that absorption is small in the transparency windows below 1\,eV and around $2.3\,$eV.\@
For the polarization rotation measurements we used a homebuilt optical polarimetry setup based on the polarization modulation technique described in Refs.\ \onlinecite{satopem} and \onlinecite{polisetty2008}. Measurements are possible in the energy range of $1.1$\,eV to $3.5$\,eV in fields up to $5$\,T and temperatures down to $10$\,K.\@
The measurements are performed in Faraday (transmission) geometry, where the light propagates along the crystallographic [111] direction, with the magnetic field also applied along this direction.
\raggedbottom
\section{Zero-field optical properties}
\subsection{Optical excitations}
In Fig.\ \ref{fig:COSO-ellip-full} the diagonal component $\sigma_{xx,1}$ of the optical conductivity at $15$\,K and $300$\,K is shown. We find a clear electronic gap with an onset at about $2.5$\,eV at $15$\,K, and charge-transfer excitations peaking at $3.2$\,eV and $4.0$\,eV with an optical conductivity of about $400$ and $1200$\,$\Omega^{-1}$cm$^{-1}$ respectively. DFT+$U$ calculations find narrow Cu hole bands for both Cu sites with an energy difference of about $\Delta E$\,=\,$E_{II}-E_{I}$\,=\,0.2\,eV, while the valence band primarily consists of broad oxygen bands.\cite{yang2012} Based on this, we tentatively
attribute the structure in the charge-transfer region to both, the splitting $\Delta E$ and the structure in the O $2p$ valence band density of states.
With decreasing temperature, the charge-transfer excitations show a blueshift and a sharpening.
\begin{figure}[t]
\center
\includegraphics[scale=1]{ellipsometry28april.png}
\caption{Optical conductivity at $15$\,K and $300$\,K as obtained by ellipsometry. The onset of charge-transfer excitations is observed at about $2.5$\,eV at $15$\,K.\@
Between $1.0$ and $2.0$\,eV local crystal-field excitations are observed with conductivity values not surpassing $50$\,$\Omega^{-1}$cm$^{-1}$. At $15$\,K, pronounced peaks are located at $1.2$, $1.4$, and $1.5$\,eV.\@ The features broaden with increasing temperature, obscuring the $1.4$\,eV peak at $300$\,K, as seen in the inset.}
\label{fig:COSO-ellip-full}
\end{figure}
The first consequence of the low structural symmetry appears in the optical conductivity. Between $1.0$ and $2.0$\,eV we find a multi-peak absorption feature, with a conductivity maximum of about $50$\,$\Omega^{-1}$cm$^{-1}$, i.\,e.\@\xspace, much weaker than the charge-transfer excitations (see inset of Fig.\ \ref{fig:COSO-ellip-full}). At 15\,K, pronounced peaks are located at $1.2$, $1.4$, and $1.5$\,eV.\@ The peaks broaden with increasing temperature, obscuring the peak at $1.4$\,eV at $300$\,K.\@ We interpret these features as local crystal-field (CF) excitations.
Similar excitation energies were observed before for CF excitations of Cu$^{2+}$ ions in a trigonal bipyramidal crystal field.\cite{Kuroda81}
For materials with inversion symmetry at the transition-metal site, such crystal-field excitations would be parity forbidden within the dipole approximation. They only become weakly allowed via the simultaneous excitation of an inversion-symmetry-breaking odd-parity phonon. Typical values of $\sigma_1(\omega)$ for such phonon-assisted excitations are below $10$\,$\Omega^{-1}$cm$^{-1}$.\cite{ruckkamp} In contrast, the absence of inversion symmetry at the Cu-I and Cu-II sites in Cu$_{2}$OSeO$_{3}$\@\xspace allows for dipole-active CF excitations by symmetry, and thus naturally explains the
sizable spectral weight below the gap between $1.0$ and $2.0$\,eV.
According to group theory,\cite{POINT} dipole-active excitations are allowed from $xy$ and $x^{2}-y^{2}$ to $z^{2}$ for the Cu-I site (assuming $D_{3h}$ symmetry) and from $xz$ and $yz$ to $x^{2}-y^{2}$ for the Cu-II site (assuming $C_{4v}$ symmetry).
Both sets of transitions are schematically indicated in Fig.\ \ref{fig:splittings}.
The spectral weight in the low-energy region, however, has a richer structure with at least three peaks at $15$\,K.\@ In the above mentioned analysis we ignored the fact that the Cu sites show a slight distortion away from the ideal square pyramidal and trigonal bipyramidal
symmetries. Considering the correct, lower site symmetries, the
remaining crystal-field excitations also are dipole active and will also contribute to the total spectral weight. Note that the two peak energies observed at 300\,K \textit{soften} to lower energy with decreasing temperature
(see the appendix for a more detailed $T$ dependence).
This is opposite to the behavior of the charge-transfer peaks and can be attributed to the asymmetric line shape found in particular at low temperature. Such an asymmetric line shape can be described by the Franck-Condon line shape typical for crystal-field excitations.
\begin{figure}[t]
\center
\includegraphics[scale=1]{absorption28april.png}
\caption{(a) Temperature-dependent absorption spectra in the range of the transmission window as measured on the $1$\,mm thick sample.
(b) Temperature dependence of the onset $E_g$ of excitations across the gap, determined at $\alpha$\,=\,$70$\,cm$^{-1}$. The fit (red line) is based on the empirical Varshni equation. (c) Cu$_{2}$OSeO$_{3}$\@\xspace is transparent for green light. The (111) sample, with a thickness $d$\,$\approx$\,$221$\,$\pm3\,\mu$m, is glued on a copper sample holder with a 2.0\,mm diameter hole.}
\label{fig:absorptioncoeff}
\end{figure}
Our assignment of the onset of charge-transfer excitations to an energy of about $2.5$\,eV is corroborated by the observation of a narrow transmission window which is situated between the crystal-field excitations and the charge-transfer region. At $10$\,K the absorption coefficient becomes as low as $10$\,cm$^{-1}$ at $2.3$\,eV, see Fig.\ \ref{fig:absorptioncoeff}. On the low-energy side the absorption decreases with decreasing temperature due to the sharpening of the crystal-field excitations. On the high-energy side a more drastic change of the absorption coefficient is found, reflecting the temperature dependence of the onset $E_g$ of excitations across the gap. This temperature dependence is well described by the empirical Varshni equation,\cite{varshni1967} as seen in Fig.\ \ref{fig:absorptioncoeff}b.
Previously, Miller \textit{et al}.\cite{Miller2010} reported on the optical conductivity of Cu$_{2}$OSeO$_{3}$\@\xspace at 300\,K based on a Kramers-Kronig analysis of reflectivity data. Above 2.5\,eV, their data reasonably agree with our results, showing a dominant peak at 4\,eV.\@ However, a Kramers-Kronig analysis is not sensitive to weak absorption features, thus the crystal-field excitations between $1.0$ and $2.0$\,eV were not resolved.
Additionally, Miller \textit{et al}.\cite{Miller2010} reported on a transmission window in the frequency range above the phonons and below about 1\,eV, which agrees with the onset of crystal-field excitations observed in our data.
\subsection{Natural optical activity}
\begin{figure}[t]
\center
\includegraphics[scale=1]{alleenNCBfinal28april.png}
\caption{The natural optical activity measured along the crystallographic (111) axis with probe photon energy $E_{\rm probe}$\,=\,$2.30\pm0.03$\,eV.\@ The left axis gives the natural optical rotation $\theta_{\rm NOA}$ per sample thickness $d$ in deg/mm and the right axis the natural circular birefringence (NCB). Above the Curie temperature $T_c\approx58$\,K, the temperature dependence is weak. Below $T_c$ an enhancement of $\theta_{\rm NOA}$ is observed, which is attributed to the finite magneto-electric coupling in the helimagnetically ordered state.}
\label{fig:temperaturencb}
\end{figure}
Due to the chiral crystal structure, Cu$_{2}$OSeO$_{3}$\@\xspace is expected to show circular birefringence with a concomitant optical rotation for linearly polarized light, known as natural optical activity with rotation angle $\theta_{\rm NOA}$.\cite{glazer,BARRONbook2004} The low absorption around $2.3$\,eV allows us to measure the temperature dependence of $\theta_{\rm NOA}$ across the paramagnetic-helimagnetic phase transition ($T_C$\,$\approx$\,$58$\,K) in transmission geometry. As a probe wavelength we used $\lambda_{\rm probe}$\,=\,$540\pm5$\,nm ($E_{\rm probe}$\,=\,$2.30\pm0.03$\,eV), which corresponds to the low-temperature transmission maximum.
The result is given in Fig.\ \ref{fig:temperaturencb}, showing both $\theta_{\rm NOA}/d$
and the difference of the refractive indices for left and right circularly polarized light, Re($\Delta n_{\rm NCB}$), where NCB is natural circular birefringence. Above $T_C$ a finite $\theta_{\rm NOA}/d$ of around $21.5$\,deg/mm is found. The rotation sign shows that the studied (111) oriented Cu$_{2}$OSeO$_{3}$\@\xspace sample is dextrorotatory, as seen from the source.\cite{newnham}
The temperature dependence of $\theta_{\rm NOA}$ in the paramagnetic phase hints at an increase of the structural chirality upon lowering temperature, i.\,e.\@\xspace, a displacement of ions at general coordinates within the Cu$_{2}$OSeO$_{3}$\@\xspace unit cell, satisfying the threefold rotational symmetry of the structural helix.
In the helimagnetically ordered phase, the temperature dependence of $\theta_{\rm NOA}$ is enhanced. Since there is no net magnetization in the helimagnetic
phase for $B$\,=\,0,\cite{belesi2012} no Faraday rotation is expected. Different studies suggest that no significant magnetostrictive lattice contraction nor lattice deformations are present in Cu$_{2}$OSeO$_{3}$\@\xspace related to the transition into the
helimagnetic phase.\cite{kurnosov2012, jwbos2008}
Instead, the increasing $\theta_{\rm NOA}$ strongly hints to the presence of a finite magneto-electric coupling in the helimagnetic phase, as suggested in Refs. \onlinecite{ruff2015}, \onlinecite{jwbos2008}, \onlinecite{belesi2012} and \onlinecite{magnetoelectricnature}.
\section{Magneto-optical properties}
\subsection{Phase transitions}
In the presence of a magnetic field $B_a$\,$\parallel$\,[111], different meta\-magnetic phases with a finite magnetization $M(B_a)$ form in Cu$_{2}$OSeO$_{3}$\@\xspace. For these phases, the total polarization rotation is of the form
\begin{equation}
\theta_{\rm tot} = \theta_{\rm NOA} + \theta_{\rm F}(M(B_a)) + \theta_{\rm HO}(M^2(B_a^2)) \,\,\, .
\label{eq:thetatot}
\end{equation}
The last term gives a higher-order (HO) rotation in field which is discussed later in the text,
while $\theta_{\rm F}$ denotes the Faraday rotation, which is proportional to the magnetization
$M(B_a)$ in the [111] direction.\cite{pershan1967}
The field-dependent magnetization can be rewritten as $M(B_a)$\,=\,$\chi_m(B_a)$\,$\cdot$\,$B_a$, where $\chi_m(B_a)$ refers to the magnetic susceptibility. Note that the magnetic susceptibility $\chi_m$ itself is
a function of the external field. We thus probe a Faraday rotation per sample thickness $d$ of
\begin{equation}
\begin{split}
\frac{1}{d}\,\theta_{\rm F}(\omega,M(B_a)) & = \beta(\omega)\cdot M(B_a)\\
& = \beta(\omega)\cdot\chi_m(B_a)\cdot B_a \\
\end{split}
\label{eq:faradayrotation}
\end{equation}
where $\beta(\omega)$ captures the microscopic magneto-optical properties. The Faraday rotation is anti-symmetric in applied field. By making field sweeps between opposite polarity field values, and afterwards symmetrizing/anti-symmetrizing the rotation response, $\theta_{\rm F}(M(B_a))$ is obtained. The symmetric part gives $\theta_{\rm NOA}+\theta_{\rm HO}(M^2(B_a^2))$. The used probe wavelength is again $\lambda_{\rm probe}$\,=\,$540\pm5$\,nm.
Figure \ref{fig:faradayrotation} shows the Faraday rotation per sample thickness, $\theta_{\rm F}/d$, as a function of field $B_a$ for temperatures ranging from $15$\,K to $65$\,K.\@ At zero applied field, different helimagnetic domains exist. However, there is no net magnetization, even for a single helical domain. In this way, $\theta_{\rm F}$ is zero for $B_a$\,=\,$0$\,mT.\@ With increasing field the helimagnetic domains acquire a conical contribution, leading to a finite field-induced magnetization $M(B_a)$, and hence a Faraday rotation. The multi-$q$ helimagnetic domain structure however still persists. At a critical field of around $60$\,mT at $15$\,K, the reorientation transition from the multi-$q$ helical to the single-$q$ conical phase is observed.
With increasing field, the spin projection parallel to $q$ and $B_a$ increases. It is for this reason that in the conical phase the Faraday rotation still increases with field, until a plateau is reached, marking the phase transition from conical to field-polarized ferrimagnetic order. At $15$\,K this phase transition is induced
around an applied field of $B_a$\,$\approx$\,$195$\,mT.\@ At the plateau a rotation of around $165$\,deg/mm is found for $15$\,K. The Faraday rotation sense
was determined to be
levorotatory as seen from the source.
\begin{figure}[t]
\center
\includegraphics[scale=1]{FaradayRotationMCB28april.png}
\caption{The Faraday rotation per sample thickness, $\theta_{\rm F}/d$, as a function of field for different temperatures. The right axis gives the corresponding magnetic circular birefringence (MCB). With increasing field, $\theta_{\rm F}$ increases until a plateau is reached, marking the
phase transition between conical and ferrimagnetic order.
At $15$\,K this phase transition is induced around an applied field of $B_a$\,$\approx$\,195\,mT.\@ At lower fields, in the $15$\,K curve around $B_a$\,$\approx$ 60\,mT, the helical-conical phase transition becomes apparent as a kink in the field dependence
of $\theta_{\rm F}$.}
\label{fig:faradayrotation}
\end{figure}
The second derivative\cite{numdiffchartrand} of the Faraday rotation allows us to construct the phase diagram shown in Fig.\ \ref{fig:phasediaglarge}. The right panel gives a zoom-in around $T_C$. In this way the phase transitions to the skyrmion lattice phase (SkL) become apparent. With a field applied along the $\langle$111$\rangle$ hard axis, the phase transition from conical to ferrimagnetic order (indicated by triangles) is of second order,\cite{clustercoso}
and the phase transition can be identified with
a local maximum in the second derivative of the order parameter $M$\,$\propto$\,$\theta_{\rm F}$. The phase transitions from helical to conical order (diamonds) and between the conical phase and the SkL phase (indicated by circles) are of first order
and can be identified with
a zero-crossing in the second derivative of $\theta_{\rm F}$. The diamonds and circles in Fig.\ \ref{fig:phasediaglarge} indicate these zero crossings, marking the phase boundaries.
\begin{figure}[t]
\centering
\includegraphics[width=3.375in]{phases9.png}
\caption{The magnetic phase diagram of Cu$_{2}$OSeO$_{3}$\@\xspace as obtained from the Faraday rotation. $B_a$ is the applied magnetic field along the crystallographic (111) direction. The left panel shows the magnetic phase diagram for a large ($B_a$,$T$) range, whereas the right panel gives a zoom-in around the skyrmion lattice phase (SkL). The triangles indicate the conical-ferrimagnetic phase boundary. Diamonds give the helical-conical phase boundary. The SkL phase boundary is indicated by the circles. The phase boundary between ordered phases and the paramagnetic phase is indicated by the squares. The color mapping indicates the second derivative of the Faraday rotation.
For a quantitative comparison of the critical fields with results obtained by other techniques,
one needs to take into account the demagnetization factor (see main text).
}
\label{fig:phasediaglarge}
\end{figure}
\begin{figure}[htb]
\center
\includegraphics[scale=1]{FluctuationsChiMOfinal02may.png}
\caption{Faraday rotation across the paramagnetic-helimagnetic phase transition, normalized by applied field $B_a$.
The curves are offset by $-(\pi/1.8)\cdot 10^5$\,rad/T$\cdot$m with respect to each other. Within the helical phase the scaled rotation attains a value of $\theta_{\rm F}$\,/($B_a$\,$\cdot$\,$d$)\,$\approx$\,$-1.7\cdot10^5$\,rad/T$\cdot$m.
The weak first-order transition into the helical phase is marked by black circles. The fluctuation-disorder regime is located between the black dot and the red diamond.
}
\label{fig:fluctuationtemperature}
\end{figure}
The phase transitions from the paramagnetic phase to the ordered phases are best observed in temperature sweeps at constant field. Depending on the field strength, different phase transitions occur.
Figure \ref{fig:fluctuationtemperature} shows $\theta_{\rm F}$\,/($B_a$\,$\cdot$\,$d$) across the paramagnetic-helimagnetic phase transition for $\vert B_a\vert$\,\textless\,$2$\,mT.\@ At low temperatures, deep within the helimagnetic phase, $\theta_{\rm F}$\,/($B_a$\,$\cdot$\,$d$) does not show a significant temperature dependence. However,
an anomaly
is observed around the paramagnetic-helimagnetic transition. In fact, the
anomaly
marks $T_C$. In mean-field approximation this phase transition is expected to be of second order. However, the interaction between chiral fluctuations renormalizes the phase transition into a weak first-order transition,\cite{firstordergarst} as seen in the temperature dependence of $\theta_{\rm F}$\,/($B_a$\,$\cdot$\,$d$).
Just above the phase transition, we find the fluctuation-disordered region \cite{Bauer2016} where the magnetic susceptibility deviates from pure Curie-Weiss behavior. The onset of this region is identified by the inflection point, located about $1$\,K above $T_C$ (indicated by the red diamonds).
The high-field phase transitions from the paramagnetic to the ordered phases are best seen in the temperature dependence of the first derivative $\partial \theta_{\rm F}/\partial B_a$, at different values of $B_a$.
Figure \ref{fig:fluctuationsverdet} shows
\begin{equation}
\chi_{\rm MO}=\frac{1}{d}\frac{\partial \theta_{\rm F}}{\partial B_a}\,=\,\beta(\omega)\big[\chi_m(B_a)+ \frac{\partial \chi_m}{\partial B_a}B_a\big]
\label{eq:susceptibility}
\end{equation}
given in rad/T$\cdot$m.
In this way the first derivative
can be identified with the
magneto-optical (MO) susceptibility.
The temperature and field dependence
of $\chi_{\rm MO}$
show good agreement with the magnetic ac-susceptibility as reported by \v{Z}ivkovi\'{c} \textit{et al}. \cite{criticalscaling}
For $5$\,mT the characteristic
anomaly
for the para\-magnetic-helimagnetic transition is visible again. For applied fields of $15$, $25$, and $45$\,mT the first-order nature remains for the paramagnetic-SkL and paramagnetic-conical transitions. The broad maximum found at lower temperatures for these applied fields indicates the temperature-induced
conical-helical
phase transition. At higher fields such as $100$, $150$, and $225$\,mT, the paramagnetic-ferrimagnetic phase transition is crossed. The magneto-optical susceptibility for these field values shows that the character of the phase transition changes to second order. For Cu$_{2}$OSeO$_{3}$\@\xspace, this
change of the phase-transition type
has been discussed extensively by \v{Z}ivkovi\'{c} \textit{et al}.\cite{criticalscaling}
The increase in $\chi_{\rm MO}$ at lower temperatures for field values between $80$ and $190$\,mT shows
where the temperature-induced ferrimagnetic-conical phase transition is crossed. With the first derivative,
the phase diagram can be completed. The boundary between the paramagnetic phase and the ordered phases
is indicated by squares in the phase diagram.
Qualitatively, the phase diagram that we have determined by optical means is in excellent agreement with previously reported results based on other techniques.\cite{seki2012,longwavelength,ruff2015} For a quantitative comparison of the applied field strengths at which the phase transitions are observed, one has to take into account demagnetization effects related to the sample shape. For the optical measurements in transmission geometry, we used a thin plate for which demagnetization effects are substantial. The demagnetization factor can be estimated \cite{satoishii1989} to lie around $N_z$\,$\approx$\,$0.9$. Accordingly, the phase transitions occur at higher applied fields than for instance in a spherical sample with $N_z$\,$=$\,$1/3$.
\begin{figure}[t]
\center
\includegraphics[scale=1]{MOsusceptibility28april.png}
\caption{The magneto-optical susceptibility $\chi_{\rm MO}$ as a function of temperature at different applied field strengths. The phase transition from the paramagnetic phase to
one of
the ordered phases shows a rescaling from first
order
(for the para to helical/conical/SkL phases) to second order (para-ferri) with increasing field. The changes in $\chi_{\rm MO}$ below $T_C$\,$\approx$\,$58$\,K show the temperature-induced conical-helical and ferrimagnetic-conical phase transitions.}
\label{fig:fluctuationsverdet}
\end{figure}
A higher-order magneto-optical rotation $\theta_{\rm HO}$ is anticipated owing to the magneto-electric nature of Cu$_{2}$OSeO$_{3}$\@\xspace. Through the finite magneto-electric coupling, the natural optical activity is enhanced in the helimagnetic phase, as discussed above. With this coupling, a field-even optical rotation $\theta_{\rm HO}(T,(M^2(B_a^2))$ is expected, see Equ.\ \ref{eq:thetatot}.
We indeed find an appreciable polarization rotation symmetric in $B_a$, as shown in Fig.\ \ref{fig:HOrotation}
At $15\,$K, the even rotation is about a factor $60$ smaller than the Faraday rotation. With increasing magnetization, the field-symmetric rotation $\theta_{\rm HO}$ increases until the ferrimagnetic phase is reached, where the rotation levels off. At the helical-conical phase transition a marked zig-zag feature is present. This feature is not a genuine higher-order rotation, but instead reflects the hysteresis in the Faraday rotation across the helical-conical first-order phase transition.
\begin{figure}[t]
\center
\includegraphics[scale=1]{HOrotation02may.png}
\caption{A field-even rotation
$\theta_{\rm HO}$, shown here per sample thickness $d$,
is present due to the finite magneto-electric effect. The different curves are offset for clarity. The phase transition from conical to ferrimagnetic order again becomes apparent. The zig-zag feature at the helical-conical phase transition is due to hysteresis in $\theta_{\rm F}$.}
\label{fig:HOrotation}
\end{figure}
\subsection{Magnitude of the magneto-optical effect}
At $15\,$K, we find a polarization rotation of around $165\,$deg/mm in the field-polarized
ferrimagnetic phase above 195\,mT and a magneto-optical susceptibility of the order of $-10^4$\,rad/T$\cdot$m
for smaller magnetic fields.
In the following, we will show that both values are large but can be explained by the fact that the spin clusters are
fully polarized already for moderate magnetic fields of 195\,mT.
For a field-independent magnetic susceptibility, the size of the magneto-optical response can be
expressed by the Verdet coefficient $\mathcal{V}(\omega)$\,=\,$\beta(\omega)$\,$\cdot$\,$\chi_m$ (cf.\ Equ.\ \ref{eq:susceptibility}).
In Cu$_{2}$OSeO$_{3}$\@\xspace, it is thus possible to assign a Verdet coefficient of
$-1.5\cdot 10^{4}\,$rad/T$\cdot$m to the helical phase at low temperatures and small fields,
see Fig.\ \ref{fig:fluctuationsverdet}.
Taking into account the demagnetizing field correction,\cite{criticalscaling,satoishii1989}
we find $\mathcal{V}(540$\,nm)\,$\approx$\,-3$\cdot$10$^{4}\,$rad/T$\cdot$m, a value rivaling known strong magneto-optical rotators such as the paramagnets Tb$_3$Ga$_5$O$_{12}$ (with $\mathcal{V}$($1053\,$nm)\,$\approx$\,0.3$\cdot$10$^{4}\,$rad/T$\cdot$m
at $4.2\,$K)\cite{yasuhara2007} and Cd$_{1-x}$Mn$_{x}$Te ($\mathcal{V} \lesssim 9\cdot$10$^{4}\,$rad/T$\cdot$m at $77\,$K depending on $\omega$ and stoichiometry).\cite{gaj1978} For these paramagnets, the large size of $\mathcal{V}$ solely originates in $\beta(\omega)$.\cite{pershan1967} In contrast, the large magneto-optical response of Cu$_{2}$OSeO$_{3}$\@\xspace can be attributed to the large magnetic susceptibility
at low fields and the strong magnetization at larger fields.
The magnetization\cite{jwbos2008}, of Cu$_{2}$OSeO$_{3}$\@\xspace shows that the $S$\,=\,1 clusters are fully aligned along
the applied field in the ferrimagnetic phase, i.e., already for moderate magnetic fields of 0.2\,T.\@
This reflects that the magnetocrystalline anisotropy is weak in this cubic magnet with small spin-orbit coupling. As a result, the helical or conical domains can be easily reoriented,\cite{chizhikov2015} even when the magnetic field
is applied along the $\langle$111$\rangle$ hard axis. This in turn is reflected in a large magnetic susceptibility $\chi_{m}$, and hence a large magneto-optical susceptibility $\chi_{\rm MO}$.
For the field-polarized ferrimagnetic phase, strong magneto-optical effects are naturally expected.
In fact, roughly similar (volume) magnetizations and polarization rotations have been observed
in ferromagnetic cuprates such as the 2D compounds K$_2$CuF$_4$ [\onlinecite{laiho1976}] and (CH$_3$NH$_3$)$_2$CuCl$_4$ [\onlinecite{arend1975}], showing rotations of $36\,$deg/mm and $50\,$deg/mm, respectively.
This shows that the magneto-optical coupling $\beta(\omega)$ in Cu$_{2}$OSeO$_{3}$\@\xspace is of equal order of magnitude as in these materials.
The microscopic magneto-optical interaction $\beta(\omega)$ depends on the difference in dipole transition strength
for the $\sigma_{+}$ and $\sigma_{-}$ transitions at $\omega$,
where $\pm$ refers to right/left circularly polarized light, respectively.\cite{pershan1967}
A \textit{larger} difference in transition strength leads to a \textit{stronger} rotation.
In a ferromagnet with full spin polarization, only one of the transitions $\sigma_{\pm}$ is preferentially allowed
due to angular momentum conservation.
In contrast, a perfect antiferromagnet with zero magnetization shows equal transition strengths and
thus no polarization rotation.
In Cu$_{2}$OSeO$_{3}$\@\xspace, one finds a ferromagnetic alignment of the $S$\,=\,1 clusters but antiferromagnetic alignment between
Cu-I and Cu-II sites within a given cluster, see Fig.\ \ref{fig:splittings}.
However, the excitation energies are different for Cu-I and Cu-II sites, which holds for both, the crystal-field excitations and charge-transfer excitations. In addition, the Cu-I and Cu-II sites are present in a ratio of 1:3. Accordingly, the cancellation typical for a simple antiferromagnet
does not occur in Cu$_{2}$OSeO$_{3}$\@\xspace.
Moreover, the transition dipole strength of both types of excitations is relatively large, which magnifies the magneto-optical response. Whether the different optical transitions have the same rotational sense or not
cannot be answered based on a single-wavelength measurement and remains open for further studies.
\section{Conclusions}
We have shown how a variety of linear optical properties can be used to probe inversion and time-reversal breaking properties of the chiral magnetic cuprate Cu$_{2}$OSeO$_{3}$\@\xspace. The broken inversion symmetry at the two crystallographically inequivalent Cu sites leads to a finite dipole matrix element for local crystal-field excitations. These orbital excitations were observed below the charge-transfer gap, in the energy range from $1.0$ to $2.0$\,eV.\@ The transmission window found between
the crystal-field and the charge-transfer absorption regions allowed us to measure the optical polarization
rotation across the various magnetic phase transitions in Cu$_{2}$OSeO$_{3}$\@\xspace. The zero-field optical rotation, corresponding to the natural optical activity of the crystal, is enhanced upon entering the helimagnetically ordered phase. This observation provides evidence for a finite magneto-electric coupling in the helimagnetic phase of Cu$_{2}$OSeO$_{3}$\@\xspace.
In the presence of an external magnetic field, the finite magnetization leads to a sizable Faraday rotation. The ease of domain reorientation by the external magnetic field was discussed to be at the origin of the large magneto-optical susceptibility in the helical and conical phases.
The large Faraday rotation observed in the field-polarized ferrimagnetic phase agrees with
results obtained on ferromagnetic cuprates.
The Faraday rotation provides a sensitive tool to conveniently probe the nature of the various magnetic phase transitions in Cu$_{2}$OSeO$_{3}$\@\xspace, including the subtle first-order nature of the helimagnetic-paramagnetic phase transition. From the magneto-optical data we reconstructed the phase diagram of Cu$_{2}$OSeO$_{3}$\@\xspace as a function of magnetic field and temperature, including the skyrmion lattice phase.
\section{acknowledgement}
The authors are grateful to M. Garst, A. Rosch, P. Becker, L. Bohat\'{y}, and T. Lorenz for insightful discussions. We also thank P. Padmanabhan for critical reading of the manuscript. Part of this work was financially supported by the Bonn-Cologne Graduate School of Physics and Astronomy (BCGS).
\section*{Appendix}
\subsection*{Temperature dependence of ellipsometry data}
The temperature dependence of the optical conductivity in the region of the crystal-field excitations is shown in more detail in Fig.\ \ref{fig:COSO-ellip-Tdep}. The inset shows the temperature dependence of the spectral weight of this region, i.\,e.\@\xspace the integral of $\sigma_1(\omega)$ from $1$ to $2$\,eV.\@ We find a slight increase of the spectral weight with increasing temperature. The onset of magnetic order at 58\,K has no significant effect on the spectral weight of the crystal-field excitations.
\begin{figure}[b]
\center
\includegraphics[scale=1]{opticalconductivityCF1.png}
\caption{Optical conductivity in the region of the crystal-field excitations for different temperatures. The inset depicts the spectral weight (integral of the optical conductivity between $1.0$ and $2.0\,$eV) as a function of temperature.}
\label{fig:COSO-ellip-Tdep}
\end{figure}
\vspace{5 mm}
|
\section{{#1}}}
\newcommand{\qA}[1]{\subsection{{#1}}}
\newcommand{\qun}[1]{\subsubsection{{#1}}}
\newcommand{\qa}[1]{\paragraph{{#1}}}
\def\vskip 2mm plus 0.2mm minus 0.2mm{\vskip 2mm plus 0.2mm minus 0.2mm}
\def\hfill \break{\hfill \break}
\fi
\ifnum\count101=2
\newcommand{\qI}[1]{\parindent=0mm \vskip 8mm
{\centerline{\LARGE \color{red}#1}}\vskip 3mm}
\newcommand{\qA}[1]{\vskip 2.5mm \noindent
{{\bf\large\color{blue} #1}} \vskip 1mm \parindent=0mm}
\newcommand{\qun}[1]{\vskip 1mm \noindent {\sl #1 }\quad }
\def\hfill \break{\hfill \break}
\def\vskip 2mm plus 0.2mm minus 0.2mm{\vskip 2mm plus 0.2mm minus 0.2mm}
\def\qno#1{$ ^{\hbox{\scriptsize #1}} $}
\fi
\def\vrule height 2pt depth 5pt width 0pt{\vrule height 2pt depth 5pt width 0pt}
\def\vrule height 12pt depth 0pt width 0pt{\vrule height 12pt depth 0pt width 0pt}
\def\vrule height 0pt depth 5pt width 0pt{\vrule height 0pt depth 5pt width 0pt}
\def\vrule height 12pt depth 5pt width 0pt{\vrule height 12pt depth 5pt width 0pt}
\def\vrule height 20pt depth 0pt width 0pt{\vrule height 20pt depth 0pt width 0pt}
\def\vrule height 0pt depth 10pt width 0pt{\vrule height 0pt depth 10pt width 0pt}
\newcommand{\qfoot}[1]{\footnote{{#1}}}
\def\sl {\sl }
\def\vskip 2mm{\vskip 2mm}
\parindent=0pt
\ifnum\count101=1
\def\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm{\hfill \par \hskip 6mm $ \bullet $ \hskip 2mm}
\def\qee#1{\hfill \par \hskip 6mm (#1) \hskip 2 mm}
\fi
\ifnum\count101=2
\def\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm{\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm}
\def\qee#1{\hfill \par \hskip 4mm (#1) \hskip 2 mm}
\fi
\def\qcenterline#1{\vskip -2cm {\hfill {#1}\hfill} \vskip 1.8cm}
\def\qparr{ \vskip 1.0mm plus 0.2mm minus 0.2mm \hangindent=10mm
\hangafter=1}
\ifnum\count101=1
\def\qdec#1{\parindent=0mm\par {\leftskip=2cm {#1} \par}}
\fi
\ifnum\count101=2
\def\qdec#1{\parindent=0mm \par {\leftskip=1cm {#1} \par}}
\def\qcit#1{\parindent=0mm \par \parshape=1 1.2cm 9cm {\small #1} \par
\normalsize \parindent=4mm}
\def\qcitb#1{\noindent \hbox to 102mm{\hfill \small #1} \vskip 1mm}
\fi
\def\qpages#1{\count102=0{\loop\advance\count102 by 1
\null \vfill\eject \ifnum\count102<#1 \repeat}}
\def\partial_t{\partial_t}
\def\partial_x{\partial_x}
\def\partial^{2}_{t^2}{\partial^{2}_{t^2}}
\def\partial^{2}_{x^2}{\partial^{2}_{x^2}}
\def\qn#1{\eqno \hbox{(#1)}}
\def\displaystyle{\displaystyle}
\def\sqrt{1+\alpha ^2}{\sqrt{1+\alpha ^2}}
\def\widetilde{\widetilde}
\def\qmax{\mathop{\rm Max}}
\def\qb#1{\hbox{#1}}
\def^{\ \circ}\hbox{C}{^{\ \circ}\hbox{C}}
\def^{\ \circ}\hbox{K}{^{\ \circ}\hbox{K}}
\def\ \hbox{--}\ {\ \hbox{--}\ }
\def\qh#1#2{\buildrel{\displaystyle #2} \over {#1}}
\font\timesgb=ptmb at 30 pt
\font\timesg=ptmb at 16 pt
\def\vrule height 12pt depth 5pt width 0pt{\vrule height 12pt depth 5pt width 0pt}
\def\vrule height 2pt depth 5pt width 0pt{\vrule height 2pt depth 5pt width 0pt}
\def\vrule height 12pt depth 0pt width 0pt{\vrule height 12pt depth 0pt width 0pt}
\def\vrule height 0pt depth 5pt width 0pt{\vrule height 0pt depth 5pt width 0pt}
\def\vskip 0.1mm plus 0.05mm minus 0.05mm{\vskip 0.1mm plus 0.05mm minus 0.05mm}
\def\hskip 0.0mm{\hskip 0.6mm}
\def\hskip 3mm{\hskip 3mm}
\def\vskip 0.5mm plus 0.2mm minus 0.2mm{\vskip 0.5mm plus 0.2mm minus 0.2mm}
\def\hskip 1.5mm{\hskip 1.5mm}
\def\qleg#1#2#3{\noindent {\bf \small #1\hskip 1.5mm}{\small #2\hskip 1.5mm}{\it \small #3}\vskip 0.1mm plus 0.05mm minus 0.05mm }
\def\qit#1{{\it \color{blue} #1}}
\def\qch#1{ \vskip 1.0mm plus 0.2mm minus 0.2mm #1 \hskip 2mm
\hangindent=10mm \hangafter=1}
\def\qchr#1#2#3{ \vskip 2.0mm plus 0.2mm minus 0.2mm
{\color{blue} \bf #1}\hskip 3mm
{#2}
\hfill \break {(\small \it #3.)}
\hangindent=10mm \hangafter=1 \vskip 0.1mm }
\def\vskip 3mm plus 0.2mm minus 0.2mm{\vskip 3mm plus 0.2mm minus 0.2mm}
\def\boxit#1#2{\vbox{\hrule\hbox{\vrule
\vbox spread#1{\vfil\hbox spread#1{\hfil#2\hfil}\vfil}%
\vrule}\hrule}}
\fi
\begin{document}
\thispagestyle{empty}
\markboth{{\sl \hfill \hfill \protect\phantom{3}}}
{{\protect\phantom{3}\sl \hfill \hfill}}
\color{yellow}
\hrule height 20mm depth 10mm width 170mm
\color{black}
\vskip -1.8cm
\centerline{\bf \Large The wage transition in developed countries and}
\vskip 2mm
\centerline{\bf \Large its implications for China}
\vskip 20mm
\centerline{\large
Belal Baaquie$ ^1 $,
Bertrand M. Roehner$ ^2 $,
Qinghai Wang$ ^3 $
}
\vskip 8mm
\large
{\bf Abstract}\quad
The expression ``wage transition'' refers to the fact that
over the past two or three decades
in all developed economies wage increases have leveled off.
There has been a widening divergence and decoupling between
wages on the one hand and GDP per capita
on the other hand. Yet, in China
wages and GDP per capita climbed in sync (at least up to now).\hfill \break
In the first part of the paper we present comparative statistical
evidence which measures the extent of the wage
transition effect.\hfill \break
In a second part we consider the reasons of this
phenomenon, in particular we explain how
the transfers of labor from low productivity sectors
(such as agriculture) to high productivity sectors (such as
manufacturing) are the driver of productivity
growth, particularly through their synergetic effects.
Although rural flight represents only one of these
effects, it is certainly the most visible because of the
geographical relocation that it implies; it is also
the most well-defined statistically. Moreover,
it will be seen that it is
a good indicator of the overall productivity and attractivity of
the non-agricultural sector.\hfill \break
Because this model accounts fairly well
for the observed evolution in industrialized countries,
we use it to predict the rate of Chinese economic growth
in the coming decades. Our forecast for the average annual
growth of real wages ranges from 4\% to 6\% depending on
how well China will control the development of its healthcare
industry.
\vskip 10mm
\centerline{\it Provisional. Version of 29 April 2016.
Comments are welcome.}
\vskip 10mm
{\small Key-words: productivity, demographic transition,
economic growth,
wage, earnings, agriculture, industry, primary sector,
secondary sector, tertiary sector.
\vskip 4mm
{\normalsize
1: Physics Department, National University of Singapore.
Email: [email protected]\hfill \break
2: Institute for Theoretical and High Energy Physics (LPTHE),
University Pierre and Marie Curie, Paris, France.
Email: [email protected]\hfill \break
3: Physics Department, National University of Singapore.
Email: <EMAIL>
}
\vfill\eject
\large
\qI{Introduction}
\qA{Post-industrial societies}
Nowadays (i.e. in early 2016) the dominant conception is that the
post-industrial stage reached by most developed countries
and in which at least
75\% of employment is in the service sector is a superior
stage. It seems to realize the dream of clean, energy efficient
and highly productive economies. Clean and energy efficient they may be,
but are post-industrial societies also highly productive?
This is one of the questions that we try to answer in the present
paper.
\vskip 2mm plus 0.2mm minus 0.2mm
Perhaps it may be helpful to explain what lead us to question
the mainstream conception. Doubts were raised step-wise by
a number of observations among which one can highlight the following.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm In the past two or three decades, following in the footsteps of
New York and other American cities, home-delivery of pizzas
started in many west European cities. This marked a sharp
break with the period following World War II. In this period
labor had been considered as a rare and costly resource
which had to be economized as much as possible.
Home-delivery of pizzas was a low productivity
business which could be profitable only if served by poorly
paid employees. It was the harbinger of the creation of
a whole range of jobs similarly characterized by poor pay
and low productivity. As emphasized by economists such
as Jean Fourasti\'e, the fact that the
price of hair cuts or opera tickets remained unchanged
in the course of time strongly
suggests that the tertiary sector holds little promises
for productivity improvement. At the end of the paper
we will propose a broader explanation which relies on the
fact that most service goods have little or no
synergetic potential.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm The second shock was the observation that
in the United States the average hourly wage had peaked around
1975 and stagnated ever since%
\qfoot{As a matter of fact there was a sudden change in the
trend of many other economic (e.g. income inequality)
and social indicators (e.g. proportion of inmate population).
These changes are documented more fully in Roehner (2009,
chapters 9-11).}
.
True, the Gross Domestic
Product (GDP) per capita had continued to increase but
this was due to concomitant changes, for instance the
fact that an increasing number of house wives took a full
time job or the fact that an increasing share of
income comes from non-salary earnings.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm The third element which triggered our questioning
came from the observation of service-centered economies.
Switzerland and Singapore embody fairly well the dream
of an efficient, environment-friendly
post industrial economy%
\qfoot{Although very different in terms of territory and area,
the two countries
are in the same class in terms of population (8 and 5,4 million
respectively) and GDP per capita (US\$ 84,000 and 55,000); the data
are for 2013.}%
.
In addition both countries similarly benefited from
massive capital inflows and foreign direct investments.
Yet, in spite of such favorable conditions
they show the same syndrome of levelling off wages
as do other developed countries.
For Switzerland this can be seen in Fig. 3 and
for Singapore it can be observed that whereas wages
have been multiplied
by 5.5 in the 4 decades from 1965 to 2005, they have stagnated
over the last decade from 2005 to 2015 (Shanmugaratnam 2015).
In orther words, even under the
best circumstances the post-industrial, service-based economy
is not up to its promises.
\count101=0 \ifnum\count101=1
If indeed in terms of productivity
service-based economies lag behind, then this effect should
be most obvious in city-states which have little industry.
Is that the case?\hfill \break
At first sight it does not seem so. In Singapore, the
real (i.e. adjusted for inflation) average monthly wage
increased by 180\% from 1990 to 2015%
\qfoot{The data are from the ``Trading Economics'' website.}%
,
indeed a remarkable achievement;
This progression was not even slowed by the
crisis of 2008-2009.
.
However, this picture must be corrected by two observations.
First of all, despite being a city-state, Singapore
is not purely a service economy. In 2014 its manufacturing sector
represented 18\% of its GDP (down from 25\% in 2000).
In addition, its semi-industrial port activity represented 33\% of its
GDP. Thus, altogether the industrial and semi-industrial
activities represented 51\% of GDP, i.e. much more than the share
of $ 15\% $ represented by
manufacturing, transportation and warehousing in the United States%
\qfoot{Manufacturing represented $ 11.5\% $, transportation and
warehousing $ 3.5 \% $. The data are for 2007 and the source is:
Wikipedia: sectors of US economy as percent of
GDP (1947-2009).}%
.
\hfill \break
Secondly, it should be observed that
the Statistical office of Singapore publishes only
wage data for the whole labor force. This contrasts with
most western countries which
also publish wage data for workers, i.e. non-managerial
employees (see Appendix A in this respect).
However, through a declaration made by the Deputy Prime Minister
(Shanmugaratnam 2015) one learns that real wages for the lowest paid
workers have been multiplied by 5.5 between
1965 and 2005 but have stagnated in the last decade.\hfill \break
In short, even in a successful economy like Singapore
the wages of workers leveled off in the last 10 years.
\fi
\qA{Parallel with the demographic transition}
Before describing the wage transition it may be useful
to present a transition of same kind, namely the
well-known ``demographic transition''. The two transitions are of
same kind in the sense that they occur over a similar time span
and in almost all countries.
In spite of the fact that the causes of the demographic transition
are not yet well understood, this paradigm gives us a predictive
perspective which, although rough, is quite useful nevertheless.
\vskip 2mm plus 0.2mm minus 0.2mm
There are many factors which may affect women fertility, that is
to say the average number of children that a woman will have
during her life-time. Just as illustrations on can
mention the following:
the mean age at marriage, the infant mortality rate,
the attitude of the prevailing religion
with respect to birth control, the laws ruling inheritance,
the place of residence (whether rural or urban%
\qfoot{The role of this factor is demonstrated in a fairly dramatic way
by the fact that the cities of Hong Kong, Macao and Singapore
have fertility rates as low as 1.17, 0.93 and 0.80 respectively
(the data are for 2014).}%
),
whether or not the family needs a second salary, the extent to which
girls have access to education. Many other factors could
be added to this list. As a matter of fact the list
is boundless in the sense
that for each ``primary'' factor one can cite many ``secondary''
factors. For instance, the mean age at marriage can be
seen as a primary factor but it is affected
by countless social and economic circumstances.
\begin{figure}[htb]
\centerline{\psfig{width=17cm,figure=transdemo.eps}}
\qleg{Fig.\hskip 0.0mm 1\hskip 3mm Demographic transition in western Europe.}
{The thin lines are centered moving averages (over intervals of
5 years) of the differences between birth and death rates for
4 countries:
Germany (black),
Denmark (blue),
England and Wales (green).
Sweden (red).
The thick line
is the average of the 4 cases. The demographic transition
consisted in an almost simultaneous fall of death rates (mostly
due to a decrease in infant mortality) and birth rates; however,
because the second effect was predominant the net effect
was a decrease in the rate of natural increase.
The graph shows that this transition was
marked by several ups and
downs. In addition, for some countries (e.g. the Netherlands or France)
this rule did not apply at all.}
{Source: Flora et al. (1987, vol. 2, chapter 6)}
\end{figure}
\vskip 2mm plus 0.2mm minus 0.2mm
One may think that by focusing on a single country
the question will become more clearly defined
and hence easier to solve. This is an illusion, however.
As an illustration,
consider the case of France. Between 1850 and 1940, the average fertility
rate of French women and the resulting rate of natural increase of the
French population were
much lower than in other Western European countries%
\qfoot{Detailed comparative data can be found in Flora et al. (1987)
but a broad view can be obtained just by comparing the number of
immigrants to the United States or to Latin America. There is
a strong contrast between the massive flows of Germans and Italians
and the very small numbers of French immigrants}%
.
The explanation most commonly accepted by historians relies on the
observation that according to French inheritance laws
all children were entitled to an equal part of their father's landholding.
Thus, it is said, to avoid splitting their land, farmers did not
wish several children. However, can such
an explanation be sufficient
when one realizes that in 1906 only 43\%
of the labor force was working
in agriculture and that only one half of
this percentage, that is to say some 20\% of the labor force,
were landowners (Flora et al. 1987, p. 500)?
\vskip 2mm plus 0.2mm minus 0.2mm
On the contrary, if, instead of focusing on a single case, the
scope of the analysis is broadened to a whole range of
similar countries, then a fairly well defined pattern
emerges which is called the {\it demographic transition}.
It consists in the fact that between 1900 and 1940,
in almost all west European countries the
annual rate of natural increase (per 10,000 population)
fell from about 120 to less than 50%
\qfoot{For instance, in Germany the fall was from 141 to 54,
in Britain from 116 to 30, in Sweden from 111 to 25, in
Switzerland from 100 to 46 (Bunle 1954, p. 16).
There were exceptions however, the most conspicuous of which
were Italy and the Netherlands. In both cases, the rate of natural
increase remained around 100. France should not be considered
as an exception because with its initial increase rate of only 11
per 1,000, no transition could happen.}%
.
\vskip 2mm plus 0.2mm minus 0.2mm
We do not know exactly what causes
triggered the fall in births per women
and the correlative fall in natural increase. Was it better education,
urbanization, higher income, or some other factors?
Yet, despite possible exceptions, the observation of the
West European demographic transition gave us a predictive rule.
This rule displayed its full usefulness when
between 1950 and 2000 a similar
demographic transition unfolded
in many developing countries, e.g. Brazil,
India, Indonesia, Mexico, South Korea, Turkey, Vietnam.
This transition was even of greater magnitude than the
first in the sense that within 50 years
fertility rates fell from a high level of about 5 to
a level of 2 which is the long-term replacement rate.
As in the first transition, in this second transition there were also
a few countries which did not follow the rule (e.g. the Philippines
or Nigeria). However, it can hardly be denied that
the demographic transition pattern gives us a better understanding
and allows us to predict future trends.
\qA{The wage transition rule}
In the present paper, we wish to describe what can be called
a {\it wage transition}. It may be summarized by the following
statement:
\qdec{\it {\color{blue} Wage transition rule.}\quad
In developed economies, after an initial phase during
which real wages increased exponentially (with a doubling-time of
about 30 years), a second phase set in around 1980 during which average
wages leveled off or even fell slightly.
Most often this second phase set in after the share of
agriculture in the labor force had fallen under 10\%.}
\vskip 2mm plus 0.2mm minus 0.2mm
Before stating this rule
we discussed briefly the demographic
transition because the two transitions share
several important aspects.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm Although (as will be shown below)
the wage transition is confirmed by a large amount of statistical evidence,
there are a few exceptions (i.e. the UK).
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm The demographic transition could not
be attributed to a single factor but was rather correlated
with a whole bundle of variables; similarly the
wage transition comes along with a whole range of changes.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm Despite these limitations, the wage transition rule
provides a framework that gives a better understanding
and allows testable predictions.
\vskip 2mm plus 0.2mm minus 0.2mm
As an example of a better understanding one can mention the
case of the United States.
When this example is considered alone it might appear
surprising that wages increased sharply until
1975 and slumped thereafter up to present day.
The observation
becomes less puzzling when seen in the light of
the wage transition rule because then it just appears
as an early instance of a transition that took place
in most countries. This tells us that instead of trying
to explain it by purely American causes (e.g. the
Vietnam War or the decline of the stock market)
we should rather consider
factors that are common to most industrialized countries.
\qA{Growth of the Chinese economy}
As an example of a testable prediction, one can mention
the issue of the growth of the Chinese
economy, a question much debated currently (November 2015).
The wage transition rule and the observation that the share
of Chinese agricultural employment will not fall under 10\% until
at least 2025 suggest the following prediction.
\qdec{After increasing exponentially (with a doubling time of 8.4 years)
from 1990 to 2015, the average wage in China should
follow the same trend for at least another decade
that is to say until 2025.}
\vskip 2mm plus 0.2mm minus 0.2mm
During the past three decades the increase of the average wage
paralleled the growth of the Gross Domestic Product (GDP).
From 1990 to 2015 the GDP per capita (at current prices)
grew from 1,866 RMB to 49,300 RMB. With an increase
of the GDP deflator from 12 to 42 this gives a multiplication
by 7.5 for the real GDP per capita. The average annual growth rate
was 8.2\% which corresponds to a doubling time of 8.4 years.
In a country as large as China, domestic consumption plays
a crucial and increasing role which means that
a rising average wage is both the effect and engine
of economic growth.
\vskip 2mm plus 0.2mm minus 0.2mm
The following qualifications may apply as second order effects.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm If productivity progress slows down
because of a rapid development of the healthcare sector
(there is a longer discussion of this point below)
then one would expect a somewhat slower growth, say around 6\%.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm In addition, in
the coming decades it is likely that,
as in western countries, the share of non-salary earnings will
increase. As a result, salaries may grow slower
than the GDP/capita. This would lead to an annual growth forecast
for salaries in the range $ 4\%-5\% $.
\vskip 2mm plus 0.2mm minus 0.2mm
Actually, in the previous prediction
the key-point is not the exact value
of the rate but rather the statement that such a long-term
trend will not be derailed by one or two short-lived
recessions%
\qfoot{Just for the purpose of comparison, it can be of
interest to consider the case of Mexico.
Between 1980 and 2013 it experienced 3 major crises.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm In December 1982 the country
was bailed out after defaulting
on its sovereign debt. It was not until 1994 that the GDP/capita
resumed its pre-crisis level.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm Then,
in early 1995 a second bail out
(\$50 billion) was necessary to prevent a new default.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm Finally, the worldwide crisis of 2008 led to a 5\% fall of GDP.
\hfill \break
As nowadays for Greece, there were a succession of debt restructuring
plans: e.g. the Baker, Brady and Clinton plans.
Because of these recurrent crises, the GDP/capita expressed in
US dollars increased by only 31\% in the 33 years from 1980
to 2013. Expressed in constant dollars
the growth would be about zero. Thus, it
can be said that the ``North American Free Trade
Agreement'' (NAFTA) which started in 1994
has had but a dismal effect on the economic growth of
Mexico.}%
.
Nevertheless, it may not hold if there is an
upheaval similar to what happened in the Soviet Union in the
years following 1991.
\qA{Connected issues}
Before starting the discussion of the wage transition,
it may be helpful to keep in mind the following observation.
In economics all phenomena are inter-connected.
Therefore, it would be easy to list many effects
which may possibly play a role in the levelling off
of wages. For instance, the financialisation of the economy
contributes to the divergence between
GDP/capita and wages; growing inequality contributes
to the divergence between high and low earnings;
through expansion of the supply side, immigration
contributes to eroding wages; the gradual disappearance
of unions may have the same effect.
\vskip 2mm plus 0.2mm minus 0.2mm
Although all these effects may play a role, thanks to
our comparative perspective, we know that they do not play
the main role. Indeed, the great advantage of this approach
is that it allows us to filter out secondary factors.
In Japan
there is little immigration, in Sweden unions are still
important forces, in Germany, financialisation of the economy
is much less developed than in the US. Nevertheless, in
all these countries the wage transition effect can be observed.
In short, even if the previous effects exist, they are
only second-order effects in the sense given to this expression
in physics.
\qI{The wage transition in the United States}
\qA{Observation}
There are two good reasons for
starting this investigation with the case of the United States.
The first is that, as already said, it was the first country where
this transition occurred. Secondly,
it is probably the country where the wage transition was the sharpest.
This can be seen clearly by comparing Fig. 2 and Fig. 3.
\begin{figure}[htb]
\centerline{\psfig{width=17cm,figure=usa12.eps}}
\qleg{Fig.\hskip 0.0mm 2\hskip 3mm Wage transition in the United States.}
{In phase 1 the mean wage, $ w $, increased exponentially
at an average rate of $ 2.0\% $ per year
whereas in phase 2 it stagnated. Simultaneously
the proportion of the labor force employed in agriculture, $ f_a $
fell from $ 41\% $ in 1900 to $ 4.4\% $ in 1970. There
is a high correlation, namely 0.987, between the two changes.
The
relationship between the logarithms of the two variables
is: $ \log w =-a\log f_a+b,\quad a=0.67 \pm 0.09,\ b=1.87\pm 0.06 $.
However, because several other economic factors changed as well
during this time period,
this close connection does not necessarily mean that the wage effect
resulted from the shift in the labor force.}
{Sources: Historical Statistics of the United States 1975,
Liesner 1989.}
\end{figure}
Whereas in the US the transition from exponential increase to
stagnation occurred abruptly, elsewhere there was a smooth transition
involving a progressive reduction of the growth rate.
\qA{Possible ``explanations''}
If we limit our attention to the US case, many possible
``explanations'' may come to mind. For instance, one can mention
the following.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm The decades after 1945 were marked by a progressive erosion
of the fighting capability and bargaining power
of American unions. This trend
already started with the Taft-Hartley Act of 1947 but
it really gained momentum only in the 1970s. One of its clearest
manifestations was the dramatic fall in the frequency of strikes.
More details can be found in Roehner (2009). For instance,
in the private sector the unionization rate dropped
from about 30\% in the 1950s to 8\% in 2007.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm The suddenness of the transition may be ``explained'' by
a series of external shocks.
(i)
the confidence crisis which affected the dollar in the late 1960s
and which eventually led to President Nixon's
action in 1971 ending its convertibility to gold.
(ii) the slump of stock prices at the New York Stock Exchange,
(iii) the oil shock,
(iv) the aftershocks of the Vietnam War d\'eb\^acle.
\vskip 2mm plus 0.2mm minus 0.2mm
In order to narrow down the set of possible ``explanations''
we must consider a broader range of cases.
This is what is done in the next section. The fact that
the same wage transition can be observed in many countries will
ipso facto eliminate any specifically American explanation.
\qI{The wage transition in comparative perspective}
\qA{Average wage increases}
In February 2015, in his budget speech the Deputy Prime Minister
(and also Finance Minister) of Singapore, Tharman Shanmugaratnam,
warned that wage stagnation had set in for most
developed economies in the United States, Europe and Japan
(Shanmugaratnam 2015). As seen in the previous section
regarding the United States, and as will be seen in the present
section for the European countries and Japan, wage stagnation
is in fact a phenomenon which started over two decades ago.
However, as already mentioned,
in Singapore the phenomenon started only around 2005.
In other words, like the demographic
transition, the wage transition did not occur simultaneously
in all places but occurred gradually with
country-dependent time lags.
This is illustrated in Fig. 3.
\begin{figure}[htb]
\centerline{\psfig{width=17cm,figure=wcomp.eps}}
\qleg{Fig.\hskip 0.0mm 3\hskip 3mm Wage transition in several countries.}
{In phase 1 the average wage increased exponentially
whereas in phase 2 it stagnated or increased more slowly.
The names of the countries at the bottom of the graph
indicate the years when the transition occurred. To
the 5 countries shown one could add Singapore whose transition
occurred in 2005 as discussed in the text.}
{Sources:
USA: see Fig.1;
Japan: Historical Statistics of Japan, average monthly contractual
earnings;
Switzerland: The website of the ``Federal Office of Statistics''
gives an annual series of real wages starting in 1939;
UK: Until 1985: Liesner 1989, 1990-2015: Office of National Statistics,
average weekly earnings of manual workers;
Germany: Until 1985: Liesner 1989, 1990-2015: Federal
Statistical Office, gross hourly earnings for manual production
workers.}
\end{figure}
\qA{The case of Britain}
It can be observed that Britain stands apart both in Fig. 3 and
in Fig. 6. One may wonder why.
\vskip 2mm plus 0.2mm minus 0.2mm
In 1846, that is to say very early compared with other European
countries,
through the repeal of the Corn Laws which imposed heavy duties
on grain imports, Britain decided that grain supplies could
be obtained in a much cheaper way through overseas imports
particularly from the British Empire (e.g. from Canada or India%
\qfoot{To the point that India exported grains to Britain
even in years
when there were famines in some parts of the country.}%
).
This step was foreshadowed by recurrent land ``Enclosure Acts''
(e.g. those of 1801 and 1845)
which transformed into private property the
``common lands'' that hitherto had allowed
farming by landless farmers.
\vskip 2mm plus 0.2mm minus 0.2mm
Through these steps people were compelled to move from the
countryside to cities. In other European countries a substantial
rural flight started only much later.
\qA{Interpretation}
Table 1 shows that with respect to productivity
the only clear and long-term distinction
is the one between agriculture and manufacturing.
The tertiary service sector is a mixed bag involving
activities of very different productivity levels.
As a consequence the composition of this sector will
be different from one country to another and in each
country it will change in the course of time.
For instance, the software industry which nowadays is
an important component in the United States was
non-nonexistent some 30 years ago and is still under-represented
in European countries. Another example is the health care
industry; according to table 1, in 2012
this sector represented 12\% of the
total employment and 7.1\% of the total GDP%
\qfoot{Another indicator is the weight of health
care expenditures; according to World Bank data in 2012
it represented 17\% of the US GDP.}%
; some 30 years ago the weight of this sector was much
smaller and nowadays it is still much lower
in many other developed countries.
\vskip 2mm plus 0.2mm minus 0.2mm
It is this changing shape of the tertiary sector which
makes it difficult to set up a model which would have
a permanent validity. There is a similar
difficulty with the demographic transition in the sense
that the transition described by Fig. 1 was in fact
followed by a second one in the early 21th century that
brought down the fertility rates of many developed countries
well below the replacement rate of 2.1.
Although the two transitions may share some characteristics
their mechanisms may not be identical.
\begin{table}[htb]
\small
\centerline{\bf Table 1 \quad GDP per employee in various
economic sectors (USA, 2012)}
\vskip 5mm
\hrule
\vskip 0.7mm
\hrule
\vskip 2mm
$$ \matrix{
&\hbox{Sector}\hfill & \hbox{Contribution} & \hbox{Employment}
\hfill&
\hbox{GDP per employee} \cr
&\hbox{}\hfill & \hbox{to GDP} & \hbox{} \hfill &\cr
\vrule height 0pt depth 5pt width 0pt
&\hbox{}\hfill & \hbox{[billion \$]} & \hbox{[million]}
& \hbox{[thousand \$]}\cr
\noalign{\hrule}
\vrule height 12pt depth 0pt width 0pt
1&\hbox{Educational services}\hfill & \hfill 183 & 3.34 & 54.8\cr
2&\hbox{Retail trade}\hfill & \hfill 932 & 14.9 & 62.5 \cr
3&\hbox{\color{blue} Health care, social assistance}\hfill & \hfill 1,153
&16.9 & \color{blue} 68.2 \cr
4&\hbox{Art, entertainment}\hfill & \hfill 157 & 2.09 & 75.1\cr
5&\hbox{\color{green}Agriculture, forestry, fishing }\hfill & \hfill 186 & 2.11
& \color{green} 88.1\cr
6&\hbox{Government (federal and local)}\hfill & \hfill 2,197 &21.9 & 100\cr
7&\hbox{Transportation, warehousing}\hfill & \hfill 467 &4.41 & 106\cr
8&\hbox{Professional services}\hfill & \hfill 1,912 &17.9 & 106\cr
&\hbox{\color{magenta} All tertiary sectors}\hfill &
\hfill 11,929 & 97.6& \color{magenta} 122\cr
&\hbox{\color{red} All sectors}\hfill & \hfill 14,509 & 112&
\color{red} 129\cr
&\hbox{\color{blue} All non-agricultural sectors}\hfill & \hfill 14,323 & 110&
\color{blue} 130\cr
9&\hbox{\color{blue} Manufacturing}\hfill & \hfill 1,983 &11.9 &
\color{blue}166\cr
10&\hbox{Wholesale trade}\hfill & \hfill 962 & 5.67 & 169\cr
11&\hbox{Information}\hfill & \hfill 737 & 2.74& 269\cr
12&\hbox{Financial activities}\hfill & \hfill 3,229 & 7.78& 415\cr
\vrule height 0pt depth 5pt width 0pt
13&\hbox{Mining}\hfill & \hfill 411 & 0.80 & 514\cr
\noalign{\hrule}
} $$
\vskip 1.5mm
\small
Notes: The GDP per employee given in the last column can be seen
as a measure of employee productivity.
Despite the modernization of US agricultural production
there is still a 1:2 productivity gap between
agriculture and manufacturing.
This table suggests that what is usually called the service
sector is in fact a very heterogeneous category in
which productivity covers a scale from 1 to 10.
The ratio between the productivities in agriculture and
non-agriculture sectors is equal to $ 130/88=1.47 $.
The present table is not an exhaustive list of all sectors
but it represents 90\% of the total GDP.\hfill \break
{\it Sources: GDP: Bureau of Economic Analysis (interactive table);
Employment: Bureau of Labor Statistics (``Table 2.1
Employment by major industry sector''.}
\vskip 5mm
\hrule
\vskip 0.7mm
\hrule
\end{table}
The global productivity of the tertiary sector is a weighted
average%
\qfoot{In fact in the present case the weighted average
is identical to the GDP per employee:
$ \sum (G_i/p_i)(p_i/P)=(1/P)\sum G_i $.}
which in the case of the sectors listed
in Table 1 (i.e. all 13 sectors except
agriculture, manufacturing and mining) is equal to \${$ 122,000 $}
per employee, a figure which is mid-way between
the productivities of agriculture and manufacturing.
\vskip 2mm plus 0.2mm minus 0.2mm
In the next section we examine the implications for
economic growth.
\qA{Implications for economic growth}
For China in 2013 the weights and productivity
of the agricultural sector on the one side and of
all non-agricultural sectors on the other side
are given in Table 2.
\begin{table}[htb]
\small
\centerline{\bf Table 2 \quad Contribution of different
sectors to the Chinese GDP, 1990-2013}
\vskip 5mm
\hrule
\vskip 0.7mm
\hrule
\vskip 2mm
$$ \matrix{
\hbox{Sector}\hfill & \hbox{Share} & \hbox{Share}&
\hbox{GDP per} & \hbox{Productivity}\cr
\hbox{}\hfill & \hbox{of GDP} & \hbox{of employment} \hfill &
\hbox{employee}& \hbox{multiplier}\cr
\hbox{}\hfill & g_i & e_i & (g_i/e_i)g & k=(g_i/e_i)/(g_1/e_1)\cr
\vrule height 0pt depth 5pt width 0pt
\hbox{}\hfill & \hbox{[\% of GDP]} & \hbox{[\% of employment]}
& \hbox{}&\cr
\noalign{\hrule}
\vrule height 12pt depth 0pt width 0pt
\hbox{\color{blue} 1990}\hfill & & & &\cr
\hbox{Agriculture}\hfill & g_1=25 & e_1=60 & g_1/e_1=0.42g &1\cr
\hbox{Industry and tertiary sector}\hfill & g_2=75 & e_2=40 &
g_2/e_2=1.87g & 4.3 \cr
\hbox{\color{blue} 2013}\hfill & & & \cr
\hbox{Agriculture}\hfill & g_1=10 & e_1=34 & g_1/e_1=0.29g & 1 \cr
\vrule height 0pt depth 5pt width 0pt
\hbox{Industry and tertiary sector}\hfill & g_2=90 & e_2=66 & g_2/e_2=1.36g &
4.5 \cr
\noalign{\hrule}
} $$
\vskip 1.5mm
\small
Notes: $ g $ denotes the GDP per employee for the whole economy.
Between 1990 and 2013 the percentage of the total labor force
engaged in the agricultural sector fell from 60\% to 34\%.
In 2013 the productivity gap between
agriculture and the rest of the economy was 4.5; this was three times
the ratio of 1.47 found in Table 1 for the United States.
Needless to say, even in agriculture there was
a marked productivity increase due to the rapid growth of $ g $.
\hfill \break
{\it Sources: GDP data: Bajpai (2014); employment data: World Bank,
(Interactive table entitled ``Employment in agriculture (\% of total
employment)'').}
\vskip 5mm
\hrule
\vskip 0.7mm
\hrule
\end{table}
With respect to rural flight,
in the United States
the major transformation took place between 1900 and 1950.
In this time span the percentage of the labor force engaged
in agriculture fell from 41\% to 11\%,
that is to say at an annual rate of 0.6\%.
Based on the data of Table 1 and 2, Fig. 4 summarizes
the respective situations of the United States and China.
In the United States between 2000 and 2010 the percentage
of the labor force employed in agriculture fell
from 2.5\% to 1.5\%, that is to say at an annual rate of 0.1\%.
\begin{figure}[htb]
\centerline{\psfig{width=12cm,figure=uschine.eps}}
\qleg{Fig.\hskip 0.0mm 4\hskip 3mm Transfer of rural population
to the non-agricultural sector.}
{In China the transfer rate was over 10 times faster than
in the United States.The figure respects the order of magnitude
of the productivity gap within each country and also
the magnitude of the respective rates of transfer.
The scales of the two vertical productivity axes
are not the same; this allowed us to draw
two identical rectangles to represent the non-agricultural
sectors.}
{Sources: Tables 1 and 2.}
\end{figure}
In China during the same time interval, this percentage
fell from 50\% to 36\%, that is to say at an annual rate
of 1.4\%.
Whereas nowadays in the United States
this transfer of jobs from the agricultural sector to the
non-agricultural sector represents a very marginal
source of productivity progress, in the first half of the
20th century it was a major source of productivity growth.
However, the present transfer rate in China is twice
as fast as it was at that time in the US. It was between 1940 and
1950 that the US transfer rate reached its maximum
with an annual rate of 0.9\%, still lower than
seen in China in the past decades.
\qI{Model describing the transfer effect}
In order to test to what extent the mechanism described in
Fig. 4 can explain long-term wage increases we need to
set up a model along these lines.
\qA{Definition of key-variables}
In order to make it as simple as possible, we will
introduce only two sectors: agriculture (sector $ 1 $) on the one hand
and the rest of the economy (sector $ 2 $) on the other hand.
At first sight this may seem a rough approximation;
for instance, mining and health care will be together
in sector $ 2 $; however, mining is a fairly small
sector (it is 21 times smaller than health care).
The purpose of our two-sector model is to provide
a broad picture and to
explain (and predict) long-term trends.
\vskip 2mm plus 0.2mm minus 0.2mm
We introduce the following notations:\hfill \break
$ E $: total employment, $ E_i,\ i=1,2 $: employment in sectors 1,2
respectively. $ e_i=E_i/E $.\hfill \break
$ G $: total (real) GDP, $ G_i,\ i=1,2 $: GDP of sectors 1,2 respectively,
$ g_i=G_i/G $.\hfill \break
We denote by $ p=G/E $ the GDP per employee of the whole economy.
$ p_i=G_i/E_i=(g_i/e_i)(G/E)=(g_i/e_i)p $ denotes the GDP per employee
in each sector, that is to say their productivity.
\vskip 2mm plus 0.2mm minus 0.2mm
Now we introduce a variable that we call the {\it productivity
multiplier} that will characterize how much the productivity
of sector 2 is higher than the one of sector 1.
$$ \hbox{Productivity multiplier:}\quad k={ p_2 \over p_1 }=
{ g_2/e_2 \over g_1/e_1 } = { (1-g_1)/(1-e_1)\over g_1/e_1 } $$
The interesting point is that $ k $ can be computed as soon as
one knows the two numbers $ g_1,e_1 $.
\qA{Order of magnitude and changes of key-variables}
In order to give some substance to these definitions, let us
give some numerical values for the United States and China.
The data for $ p $ are expressed in dollars of 1982 and renminbi of 2013
respectively.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm {\color{blue} USA}: \hfill \break
{\small
$ 1900:\ p=\$9,593,\ e_1=41\%,\ g_1=23\%,\ g_1/e_1=0.56,\ k=2.32 $\hfill \break
$ 1950:\ p=\$20,424,\ e_1=11\%,\ g_1=6.9\%,\ g_1/e_1=0.62,\ k=1.66$\hfill \break
$ 1980:\ p=\$32,095,\ e_1=3.4\%,\ g_1=3.0\%,\ g_1/e_1=0.88,\ k=1.14$\hfill \break
}
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm {\color{blue} China}: \hfill \break
{\small
$ 1983:\ p=\hbox{RMB}\ 5,114,\ e_1=67\%,\ g_1=33\%,\ g_1/e_1=0.49,\ k=4.12$\hfill \break
$ 2000:\ p=\hbox{RMB}\ 24,184,\ e_1=50\%,\ g_1=50\%,\ g_1/e_1=0.30,\ k=5.67$\hfill \break
$ 2013:\ p=\hbox{RMB}\ 73,513,\ e_1=31\%,\ g_1=10\%,\ g_1/e_1=0.32,\ k=4.04$\hfill \break
}
A noteworthy point is the stability of $ g_1/e_1 $; it
changes much less than $ e_1 $ and $ g_1 $ separately and also much
less than $ p $. How should this be interpreted?
The fact that $ g_1/e_1=p_1/p $ is quasi-constant in spite of the
fact that $ p $ doubles, triples or quadruples
means that agricultural productivity
increases approximately at the same pace as global productivity.
Moreover, it can be observed that $ k $ changes fairly slowly
which for a structural variable is of course a welcome characteristic.
\vskip 2mm plus 0.2mm minus 0.2mm
Other data will be found in Table 3.
In addition, intermediate values of the GDP per employee
for 1995, 2000, 2005 and 2010 are as follows (expressed in 2015 RMB):\hfill \break
$ p_1: 7,434,\ 7,547,\ 10,150,\ 20,030 $ \hfill \break
$ p_2: 27,323,\ 42,189,\ 60,273,\ 75,394 $\hfill \break
During these 25 years the GDP per employee
has been increasing exponentially both in the agricultural sector
and in the rest of the economy. The coefficients of correlation
$ (t,\log p_i) $ are 0.97 and 0.99 respectively. The exponents
are given in the notes of Table 3.
\begin{table}[htb]
\small
\centerline{\bf Table 3:\quad China: GDP per employee in agriculture
and rest of economy, 1990-2015}
\vskip 5mm
\hrule
\vskip 0.7mm
\hrule
\vskip 2mm
$$ \matrix{
& \hbox{} & & 1990 & 2015 & \hbox{Ratio} \cr
\vrule height 0pt depth 5pt width 0pt
& \hbox{} & & & & 2015/1990 \cr
\noalign{\hrule}
\vrule height 12pt depth 0pt width 0pt
& \hbox{\bf \color{blue} Whole economy} \hfill& & & &
\hbox{} \cr
1 & \hbox{GDP (billion of current RMB)} \hfill& & 1,866 & 67,000 &
\hfill 36.00 \cr
2 & \hbox{GDP deflator} \hfill& & 12 & 42 & \hfill 3.50 \cr
3 & \hbox{GDP (billion of 2015 RMB)} \hfill& G & 6,531 & 67,000 &
\hfill 10.30 \cr
4 & \hbox{Labor force (billion)} \hfill& E & 0.64 & 0.81 & \hfill 1.26\cr
5 & \hbox{GDP per employee (2015 RMB)} \hfill& p & 10,204 & 82,716 &
\hfill 8.10\cr
& \hbox{\bf \color{blue} Agriculture} \hfill& & & & \cr
6 & \hbox{GDP (billion of 2015 RMB)} \hfill& G_1 & 1,763 & 6,030 &
\hfill 3.42\cr
7& \hbox{Labor force (billion)} \hfill& E_1 & 0.33 & 0.25 & \hfill 0.75 \cr
8& \hbox{GDP per employee (2015 RMB)} \hfill& p_1 & 5,342 & 24,210
&\hfill 4.51\cr
& \hbox{\bf \color{blue} Rest of economy} \hfill& & & & \cr
9& \hbox{GDP (billion of 2015 RMB)} \hfill& G_2 & 4,768& 60,970 &
\hfill 12.79 \cr
10& \hbox{Labor force (billion)} \hfill& E_2 & 0.31 & 0.56 & \hfill 1.81\cr
\vrule height 0pt depth 5pt width 0pt
11 &\hbox{GDP per employee (2015 RMB)} \hfill& p_2 &
15,380&108,875&\hfill 7.07 \cr
\noalign{\hrule}
} $$
\vskip 1.5mm
\small
Notes: It can be observed that
the rates of change of the labor force are an order of magnitude
smaller than the rates of change of the GDP. When intermediate
values are introduced one finds that both $ p_1 $ and $ p_2 $
increase as exponentials with exponents $ \alpha_1=0.062\pm 0.016 $
(doubling time of 11.1 years) and
$ \alpha_2=0.075\pm 0.011 $ (doubling time of 9.2 years).
As a result, in the long-term the productivity gap
increases as $ p_2-p_1 \sim \exp(\alpha_2t) $
\hfill \break
{\it Sources: World Bank, Trading Economics, http://www.investopedia.com.}
\vskip 5mm
\hrule
\vskip 0.7mm
\hrule
\end{table}
The exponential increase of $ p_1 $ and
$ p_2 $ has two interesting implications.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm As is well known, in a limited world
an exponential growth cannot last for ever which means that
a saturation effect will set in sooner or later. The
data shown in Fig. 6 suggest that it should not occur before
2025.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm Over past decades the
productivity gap between the agriculture sector and the rest
of the economy has increased exponentially:
$$ \hbox{gap}=p_2-p_1=p_2(1-p_1/p_2)=
p_2\left[1-(A_1/A_2)\exp(-d t)\right]\ d=\alpha_2-\alpha_1=0.013 $$
For large $ t $ the gap becomes equal to $ p_2 $
itself. Thus, any labor transfer from sector 1 to sector 2 will
trigger a productivity jump whose magnitude increases
exponentially in the course of time.
This effect may compensate for diminishing transfers as $ e_1 $
becomes smaller. In other words, the productivity effect of
rural flight will remain significant even when $ e_1 $ will
fall under 10\% as may happen around 2030 according to Fig. 6.
\qA{Describing the GDP change due to a transfer of labor force}
Now, in order to describe the mechanism of Fig. 4 we must assume that
$ e_1 $ decreases by a quantity $ \Delta e_1 $. In fact, in what follows
it will be more convenient to follow the variable $ e_2 $
because its variations $ \Delta e_2=-\Delta e_1 $ are positive
in the course of time.\hfill \break
The change
in $ G $ resulting from a transfer of population from sector 1 to
sector 2 will be:
$$ G=G_1+G_2=p_1e_1E + p_2e_2E \rightarrow
\Delta G=\left( p_1\Delta e_1 + p_2\Delta e_2\right)E
=(p_2-p_1)E\Delta e_2 \qn{1} $$
In the above expression of $ \Delta G $ the productivities
$ p_1, p_2 $ have been kept constant because we wish to
focus on the effect of a labor force transfer%
\qfoot{The huge changes which occur in the course of time for
$ p_1 $ and $ p_2 $ are a different question which will
be discussed later.}%
.
By introducing the factor $ k $, expression (1) becomes:
$$ \Delta G=
(k-1)p_1 E\Delta e_2 =(k-1)\left({ g_1\over e_1 }\right)G\Delta e_2
\rightarrow
\left( { 1\over \Delta e_2 }\right){ \Delta G \over G }=
(k-1)\left({ g_1\over e_1 }\right) \qn{2} $$
We can take advantage of our previous observation regarding the stability
of $ g_1/e_1 $ to simplify expression (2) into:
$$ \left( { 1\over \Delta e_2 }\right){ \Delta G \over G }=
(k-1)<{ g_1\over e_1 }> \qn{3} $$
where $ <g_1/e_1> $ represents the average value
of $ g_1/e_1 $ over a period of several decades.
For the United states, one would have $ <g_1/e_1> =0.69 $
whereas for China: $ <g_1/e_1> = 0.35 $.
\vskip 2mm plus 0.2mm minus 0.2mm
In the left-hand side the numerator and denominator can be
multiplied by 100 which allows to express both $ \Delta e_2 $
and $ \Delta G/G $ in percent. Thus the left-hand side represents
the percentage variation of $ \Delta G/G $ for a 1\% variation
of $ e_2 $.
\vskip 2mm plus 0.2mm minus 0.2mm
Note that equation (3) applies to any transfer of labor force.
In particular, it can also describe the {\it decline} in GDP
that results from the transfer of labor force
from manufacturing to healthcare, a key-feature
in industrialized countries over the past three decades%
\qfoot{In this case sector 1 would be the whole economy
except healthcare and sector 2 would designate the healthcare
sector.}%
.
This point may be developed in a subsequent paper.
\qI{Labor force transfers as an engine of economic growth}
\qA{Rural flight seen in the broader view of all labor force transfers}
In the present paper we focus on the effect of rural flight
on economic growth for two main reasons.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm This labor force transfer has a clear statistical
definition because of the fact that agriculture is a well defined
economic sector. In contrast, a sector such as healthcare
has a fairly fuzzy
statistical definition. For instance, do nursing homes belong
to healthcare or to the hospitality industry?
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm Historically rural flight has been of great importance
in all developed countries and nowadays it is still a major
factor in countries such as China and India.
\vskip 2mm plus 0.2mm minus 0.2mm
However, this focus should not hide the fact that rural flight
is only one case (albeit the most significant)
of a broad variety of labor force transfers which take place
continuously. For instance in developed countries two sectors
have had a rapid growth over the past 50 years: (i) the healthcare
and (ii) the information technology sector.
According to Table 1, around 2012
in the United States healthcare employed
about 15\% of the total work force whereas the share of the
IT industry was of the order of one or two percent.
What impact on global productivity has had the development
of these sectors? Table 1 shows that healthcare's
and IT's productivities were one half and twice the national
average respectively.
\qA{A case in point: healthcare versus infrastructure development}
Across nations there are great disparities in the extension
of the healthcare sector. For countries with an aging populations
(as is the case of all developed countries) it is of course
natural to have an expanding healthcare sector but
it must be recognized that its actual extent
is determined by a political decision.
Thus, in the United Kingdom which has a nationalized
medical system the weight of healthcare in the economy
is only one half of what it is in the US%
\qfoot{In spite of that, around 2009 life expectancy in the UK was
2.5 years longer than in the US.}%
.
\vskip 2mm plus 0.2mm minus 0.2mm
The fact that through its own poor productivity level,
the healthcare sector drags down overall
productivity is not its only liability. Because
most of its patients are elderly persons beyond
retirement age, health care has
little synergy with other activities.
\vskip 2mm plus 0.2mm minus 0.2mm
Completely different is the case of infrastructure
expenses. Many other sectors will benefit from
better transportation infrastructure. Clearly,
healthcare based development and infrastructure
oriented development constitute two different
models. As one knows, the infrastructure model
is promoted by the Chinese government not only
domestically but also for the
rest of the Asian continent. Regarding healthcare,
at this point it is too early to say
which system China will choose.
Will it be the US or the UK model%
\qfoot{One can be sure that this kind of decision will
be subject to lobbying efforts from many sides.}
?
Clearly this choice will affect the productivity
of the whole Chinese economy. Too much
resources and investments sunk in healthcare
will be a drag on the economy.
At the beginning of the paper we said that one
would expect a 8\% growth trend until 2025 but an
over extended healthcare industry
may well slow down this expectation to 6\% or 5\%.
\qI{What fraction of growth does rural flight explain?}
In this section we compare the changes in GDP per employee
actually observed to those predicted by the
rural flight effect in three different ways.
\qA{Correlation of relative variations}
Firstly, we wish to see if there is a connection
between the ups and downs in $ e_2 $ on the one hand and
the wage changes on the other hand. Obviously,
cases (such as China)
in which there were little ups and downs will not be appropriate
for this test. In contrast, the test can be done
in good conditions for the case of the US because its
wage changes display several ups and downs.
\vskip 2mm plus 0.2mm minus 0.2mm
The comparison between observed wage changes and those
predicted by the transfer model shows a significant
correlation (Fig. 5).
\begin{figure}[htb]
\centerline{\psfig{width=12cm,figure=model.eps}}
\qleg{Fig.\hskip 0.0mm 5\hskip 3mm Observed US wage changes versus
changes resulting from the transfer effect.}
{The correlation between the two series is
0.64 (95\% confidence interval is 0.08 to 0.90).
It can be observed that real wages did not fall
during the Great Depression. This is due to the fact that
the price consumer index fell faster than nominal wages.}
{Sources: Same sources as for Fig. 3.}
\end{figure}
Although one would like to
perform additional tests, it is
not easy to find appropriate cases. For instance,
the income of
countries such as France, Germany or Japan which have
experienced wars on their territory will be affected
by exogenous shocks which have nothing to do
with the transfer effect.
\qA{Percentage of growth directly due to rural flight in 1990 and
2015}
The correlation test performed in the previous subsection
was largely independent of the magnitude of the respective
changes. Now
we wish to determine what proportion of the increase in
GDP per employee
can be accounted for by the rural flight effect.
For this test one needs a case in which the change rate
is as large as possible. China is well suited for this
purpose. Before making a formal calculation we will
compute the expected changes at the two ends of the time
interval 1990-2015. For this purpose we will use
the data given above (Table 3 and additional data).
\vskip 2mm plus 0.2mm minus 0.2mm
As already explained the rural flight effect is only one
of a number of transfer effects which contribute to
GDP growth. One of the clearest manifestations of the
effects of these transfers are the productivity
increases {\it within} the two sectors. This effect cannot be
explained by rural flight. Indeed, rural flight can only
explain the global growth of $ G $ due to the interaction
between agriculture and the rest of the economy.
The productivity
increases within the two sectors are due to interactions
between their various sub-sectors.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm Tractors, trucks instead of horse driven carriages,
engines, pumps, motorcycles,
mobile phones and many other items that are produced in
sector 2 progressively
came into use in the country side, thereby lifting
$ p_1 $. These changes would happen independently of
any transfer of labor force.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm At the same time, in sector 2 outdated textile
factories were equipped with modern spinning and
weaving machines manufactured in other factories of sector 2,
newspaper printing machines were replaced by computerized
systems, supermarkets replaced small stores. As above,
these changes would happen independently of
any transfer of labor force.
Because sector 2
is more diversified than sector 1 it experiences
more synergy and that is probably
why the endogenous growth of sector 2 is faster than
the endogenous growth of sector 1.
\vskip 2mm plus 0.2mm minus 0.2mm
According to the comments following Table 3, we expect
a greater effect of rural flight in 2015 than in 1990.
Let us see if this is confirmed by observation.
\vskip 2mm plus 0.2mm minus 0.2mm
Over the whole period 1990-2015 the labor force in sector 1
decreased by 170 millions which gives an average
change of 34 millions over every 5-year interval.
\vskip 2mm plus 0.2mm minus 0.2mm
{\bf \color{blue} 1990-1995} \quad Over this time interval,
one gets (everywhere RMB means ``RMB of 2015'')
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm The average of $ p_1 $ is
$ \overline{p_1}=(5,231+7,434)/2=6,332 $ RMB
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm The average of $ p_2 $ is
$ \overline{p_2}=(15,380+27,323)/2=21,351 $ RMB
\hfill \break
Thus, rural flight from 1 to 2 will increase $ G $
by:
$ \Delta G' = 34\times 10^6(\overline{p_2}-\overline{p_1})=511\times
10^9 $ RMB.
\vskip 2mm plus 0.2mm minus 0.2mm
The actual increase in $ G $ was
$ \Delta G=G(1995)-G(1990)= 5,627\times 10^9 $ RMB. Thus, the variation
due to rural flight represents: $ 511/5627=9.1\% $.
\vskip 2mm plus 0.2mm minus 0.2mm
{\bf \color{blue} 2010-2015} \quad A similar computation
leads to: $ \Delta G' =2,383\times 10^9 $ RMB, and
$ \Delta G=G(2015)-G(2010)= 5,627\times 10^9 $ RMB. In this case
the variation due to rural flight represents: $ 11.4\% $.
\vskip 2mm plus 0.2mm minus 0.2mm
As expected we see that in 2015 the change due to rural flight is a
larger proportion of the total increase than in 1990. We also see
that the largest part of GDP growth must be attributed
to endogenous increases of $ G_1 $ and $ G_2 $ which in turn are
due to transfers occurring {\it within} these sectors.
\qA{General formulation}
In this subsection the previous computation is presented
in a broader and more formal way.
\vskip 2mm plus 0.2mm minus 0.2mm
Let us denote by $ p $ the GDP adjusted for
inflation and per employee. From 1985 to 2014, $ p $
has increased regularly at an average annual rate of
9.2\%; for the whole interval
this rate resulted in a multiplication by 10.3
\vskip 2mm plus 0.2mm minus 0.2mm
The productivity multiplier $ k =p_2/p_1 $ fluctuated between
4 and 6.
In order to get an upper bound of the growth
accounted for by rural flight, we will
give $ k $ its maximum value of 6. As a first step,
consider the two extreme situations: $ e_1=1 $
and $ e_1=0 $. Moving from the first to the second would
result in a multiplication of $ G $ by 6. In other words,
in this thought experiment
the transfer effect explains only $ 6/10.3=58\% $ of the growth.
\vskip 2mm plus 0.2mm minus 0.2mm
In order to get a more realistic estimate we write
the values of $ G $ in 1985 and 2014 in the following way
($ G' $ denotes the values of $ G $ predicted by rural flight):
$$ G'(i)=p_1(i)E_1+6p_1(i)E_2=\left[e_1(i)+6e_2(i)\right]p_1(i)E $$
With $ e_1(1)=67\% $ and $ e_1(2)=31\% $ one gets:
$ G'(2)/G'(1)=1.67\left( p_1(2)/p_1(1) \right) $.
In the rural flight model the productivity $ p_1 $ is supposed fixed,
that is $ p_1(1)=p_1(2) $. Under this assumption
the transfer of labor accounts for only
$ 1.67/11=15\% $ of the actual productivity growth.
\vskip 2mm plus 0.2mm minus 0.2mm
Equation (3) would lead to a similar conclusion.
For average annual changes it gives:\hfill \break
$ \Delta G'/G'=(k-1)<g_1/e_1>\Delta e_2=5\times 0.4\times 1.2=2.4\% $\hfill \break
whereas the actual annual GDP growth is $ \Delta G/G=9.2\% $.
\qA{Implications}
The fact that rural flight accounts for only a fraction of the
total growth
has two important implications.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm A recession may temporarily stop employment shifts
from sector 1 to sector 2. If this transfer would account for
a high percentage (say some 80\% or more)
of economic growth, then its discontinuation
would freeze further growth altogether
and transform the recession into a depression.
However, if employment transfer actually is
only a secondary factor
the synergy engine will continue to work and will pull
the economy out of the recession%
\qfoot{Needless to say, economic growth can be thwarted
for many other reasons, e.g. deflation, capital flight,
disruption of the banking sector and so on.}%
.
\hfill \break
This is indeed what was
observed in developed countries in the decades following World
War II. There were only mild recessions.
In other words, unless there is a political
upheaval no major interruption of
growth should be expected in China for
at least 15 years, that is to say until it has become a service
economy too.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm As seen earlier, in developed countries
the growth of wages stopped or slowed down in the 1980s.
This was due to the shift of their economies toward
a service sector economy which is dominated by low productivity
sectors of which healthcare is an important component.
The fact that rural flight almost came to an end in the 1980s
was due to two circumstances: (i) The labor force of sector 1
was already largely depleted.
(ii) Because productivity growth in sector 2 was slowed down,
this sector lost its attractivity in terms of higher
salaries. Indeed, the growth of sector 2 was largely due to
the introduction of so-called ``odd jobs'',
``second class jobs'' or ``deskilled casual work''.
\qI{Conclusion}
\qA{Chinese economic growth over 2015-2035 in comparative perspective}
Several interesting features are highlighted in Fig. 6.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm The
transfer of labor force from agriculture to the industrial
and service sectors has progressed unabated in all
developed countries. Economic crises or even the two world wars
may have slightly slowed down the evolution (as in the
case of Austria for instance) but they did not have
any lasting impact.
\begin{figure}[htb]
\centerline{\psfig{width=12cm,figure=agri.eps}}
\qleg{Fig.\hskip 0.0mm 6\hskip 3mm Labor force in agriculture.}
{Apart from the UK which stands apart, all developed
countries experienced more or less the same evolution.
The fact that the curves for Sweden and Germany are
almost the same shows that wars had almost no
impact. The curve for China was extrapolated based on the
rates seen in other countries and particularly in
South Korea. Regarding the special case of Britain see the
explanation given above in relation with Fig. 3.}
{Sources: Until 1970: Flora et al, (1987) and Historical
Statistics of the United States (1975).
After 1970: Trading Economics website and Saint Louis
Federal Reserve.}
\end{figure}
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm We have already suggested
that South Korea may be a good predictor of the future
of the Chinese economy.
The development of South Korea followed the pattern
set by developed countries.
We see that in terms of proportion of the labor force working
in the agricultural sector
Finland was about 20 years ahead of South Korea
which itself was 30 years ahead of China%
\qfoot{In this respect one should not forget that, as in Taiwan,
the Japanese occupation had resulted in a substantial
industrialization not only in the north of Korea but also
in the south; more evidence can be found in Roehner (2015).}%
.
If one uses the rate of South Korea to extrapolate the curve
for China one arrives to the conclusion that in 2035
the percentage of the labor force in agriculture will be
about 10\%.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm Among developed countries the UK is a case apart.
Its agricultural sector
fell to a level of 10\% some 60 years before the United
States.
In this sense, Britain was the first
post-industrial economy; this may explain the fact
(observed in Fig. 3) that its rate of wage increase was
substantially lower than in other countries.
\vskip 2mm plus 0.2mm minus 0.2mm
What was the evolution of agricultural employment in the USSR
and later on in the Russian Federation? By collecting
data from various sources (Gaidar 2012, W\"adekin 1982,
Lerman et al. 2003, Trading Economics website)
we got the following picture:
\begin{table}[htb]
\small
\centerline{\bf Table 4\quad Russia: Employment in agriculture
in percentage of total labor force}
\vskip 3mm
\hrule
\vskip 0.7mm
\hrule
\vskip 2mm
$$ \matrix{
\vrule height 0pt depth 5pt width 0pt
1930 & 1960 & 1978 & 1985 & 1990 & 2000 & 2010 & 2014 \cr
\noalign{\hrule}
\vrule height 12pt depth 0pt width 0pt \vrule height 0pt depth 5pt width 0pt
87\% & 39\% & 21\% & 15\% & 14\% & 15\% & 9\% & 7\% \cr
\noalign{\hrule}
} $$
\vskip 2mm
\hrule
\vskip 0.7mm
\hrule
\end{table}
The table shows that the interruption of the fall
of the share of agricultural employment coincided with the
crisis of 1988-2000. So, once again, we see that
this variable
is a good indicator of long-term economic growth.
We see also that the time interval 1930-1960 saw a massive
reduction in agricultural employment and we know that
it was accompanied by rapid economic development.
\qA{Rural flight seen as a good indicator of productivity growth}
In the first part of the paper
it was shown that a wage transition occurred in most
developed countries in the 1980s. After decades of rapid growth,
wages started to slump. Rural flight appeared as
a good candidate for
explaining not only this wage transition but also the rapid
growth observed in China in past decades.
Incidentally,
one should not think that there was no or little
growth before 1983. According to Chinese GDP statistics,
the real average annual GDP growth rate was $ 9.8\% $ from 1952 to
1982 and $ 9.9\% $ from 1983 to 2014. Even though
national accounting methods may have been different
before 1983,
the idea of a stagnant economy before 1983
seems inappropriate.
Actually, the most conspicuous difference
between the two periods is that in the first one growth was
much more irregular than in the second. The coefficients
of variation of the annual growth rates were 194\%
and 27\% respectively.
\vskip 2mm plus 0.2mm minus 0.2mm
In order to test the idea of a connection between rural flight
and economic growth
we set up a model which allowed us to predict the
change of income resulting from rural flight.
It was seen that this model is in
good qualitative agreement with
observation. As expected, rural flight accounts for only
a fraction of total growth but it is also a good indicator
of the dynamism of sector 2. If tomorrow the
Chinese economic growth starts to create a lot of ``second class
jobs'' (for instance by an over-extension of the healthcare
sector) then, those young people who presently still remain in
Chinese villages (see Appendix C) will certainly think
about it twice
before moving to cities. This, in turn, will make economic growth
more sluggish. So, the main issue is whether
in the coming decades China will be able to
avoid the ``odd job trap''.
\qA{Wage progression and labor inflow}
It may seem surprising that in this paper we did not
consider the question of immigration.
As for any other good, the cost of labor is affected by supply and
demand changes. Supply is controlled by labor force expansion
either endogenously or exogenously (i.e. through immigration)
whereas demand is conditioned by GDP growth. This
question would deserve a separate study but here we will limit
ourselves to a few observations.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm Due to the sheer size
of the Chinese population foreign immigration
will have little incidence in China.
The only country whose population
could match the population of China is India. However, whereas
there are important Indian
communities in the UK and US, immigrants to China
would have to learn Chinese, quite a formidable
challenge for them.
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm The case of Japan suggests
that at the level of a whole country (as
opposed to a specific sector) labor inflow through immigration
is rather a second-order effect. Indeed,
despite a much lower immigration rate
Japanese wages leveled off in the same way as in European countries
(see Fig. 3).
\hfill \par \hskip 4mm $ \bullet $ \hskip 2mm In contrast, at sectorial level, immigration
may play a role. In this respect, as a case study,
it would be interesting to
investigate the impact on labor cost
of the inflow of Filipino nurses into the US%
\qfoot{Their immigration
is facilitated by the fact that their curriculum
integrates the requirements and specifications of the US healthcare
industry which includes practising the English
language.}%
.
\vskip 4mm
{\bf Acknowledgments}\quad We wish to express our gratitude
to Ms. Corina Neuerer of the ``Statistisches Bundesamt''
(German Federal Statistical Office) and to Ms. Alyson Williams
of the British ``Office of National Statistics'' for
helping us to locate and retrieve some of the statistical series
used in this paper.
|
\section{\label{sec:level1}Introduction}
Since the establishment of transformation optics~\cite{Pendry2006Sci,Leonhardt2006Sci,Schurig2006OE,Chen2010NatMat,pendry2015transforming,Leonhardt2010Dover}, the methodology for engineering wave functional devices via coordinate transformation has spread to various wave systems due to their common character, {\it i.e.} the form invariance of wave equations in arbitrary curvilinear coordinates and curved spacetimes. Transformation acoustics (TA) is one of these developments, which offers a standard procedure to design the material parameters so that sound waves propagate along prospective trajectories or hold proleptic properties in these materials~\cite{cummer2008material,chen2010acoustic}. The most representative device of TA is acoustic cloak that guides sound waves traveling around the cloaked objects without scattering and thus make itself and the hidden objects invisible from acoustic detections~\cite{cummer2007one,chen2007acoustic,cummer2008scattering,Norris2411Acoustic,Hua2009Hua}. From the standard approach of TA, acoustic cloaks in fluid are made up of anisotropic materials with tensor mass densities. Though real fluids do not possess these properties, a variety of cloaks have been produced by using artificial acoustic metamaterials~\cite{farhat2008broadband,Zhang2011broadband,popa2011experimental,sanchis2013three,zigoneanu2014three}.
On the other hand, it has been realized long ago that moving media act as curved spacetimes and can be used to simulate cosmic phenomena for both electromagnetic waves~\cite{gordon1923lichtfortpflanzung,leonhardt1999optics,leonhardt2000relativistic,de2003optical} and sound waves~\cite{white1973acoustic,unruh1981experimental,unruh1995sonic,visser1998acoustic,fischer2002riemannian,schutzhold2002gravity,unruh2008dumb,Bergliaffa2004Wave}.
In the pioneering works of Unruh~\cite{unruh1981experimental,unruh1995sonic}, he discovered the dynamic analogy between the massless particles in the spacetime of Schwarzschild black hole and the sound waves dominated by the radially flowing fluid. This analogy is also valid for superfluid~\cite{garay2000sonic,garay2001sonic,barcelo2001analogue,leonhardt2003bogoliubov} in which the effects associated with acoustic black hole have been observed experimentally~\cite{lahav2010realization,horstmann2010hawking,nguyen2015acoustic}. In recent years, Garc{\'\i}a-Meca {\it et~al.} combined the ideas of moving media and TA, and purposed the method of analogue transformation acoustics (ATA)~\cite{garcia2013analogue,garcia2014space,garcia2014analogue,garcia2014transformational}. They made use of background flow, instead of anisotropic effective media, to mimic the metric of spacetime, and to fabricate the desired acoustic devices, such as spacetime cloak and time-dependent spatial compressor. Through forcing the effective metric to have a certain form, the parameters of background flow, {\it i.e.} the flow velocity, the speed of sound and the mass density, can be determined. However, Garc{\'\i}a-Meca {\it et~al.} never discussed the feasibility of these requisite parameters of background flow, although they are not unrestricted but submitted to the laws of fluid theoretically.
In this paper, we try to reconcile the ATA-determined parameters with the basic laws of fluid dynamics and show several constrains of the background flow.
First, the governing equation of sound waves in ATA is accurate only when the background flow is irrotational, or at least the vorticity of the flow $\bm{\Omega}$ should be much smaller than the frequency of sound $\omega$, in order that the analogy between the flow fields and curved spacetime is applicable~\cite{Bergliaffa2004Wave}.
Second, there should exist external force fields (momentum sources) and external mass sources in general to guarantee momentum and mass conservations of the background flow. Third, since the spatial distribution of pressure and mass density are fixed after ATA procedures, their relation restricts the state equation of the background fluid. As an example, we put forward a scheme for constructing spherical cloaks with background flow which complies with the realistic fluid dynamics. In addition, we investigate the propagation of sound in this cloaking system with both geometric acoustics approach and solving the wave equation analytically. Our results show that the major difference of our scheme from the standard TA cloak is the phase retardation (or lead) effect of the sound waves traveling around the cloak. For eliminating the singularities of the parameters at the cloak surface, we introduce the unideal cloaks and evaluate their cloaking abilities.
\section{Basic theory}
To begin with, we briefly review the framework of ATA. The motion of ideal fluid (with no viscosity and thermal conductivity) obeys Euler equation and the continuity equation which correspond to the momentum and mass conservations respectively. If the fluid is barotropic, \textit{i.e.} its density $\rho$ only depends on pressure $P$, and the flow velocity $\bm{v}$ is irrotational, the compressible fluid submits to the following equations~\cite{LANDAU1987192}
\begin{subequations}\label{fluid}
\begin{flalign}
&\frac{\partial \psi}{\partial t}+\frac{1}{2}\bm{v}^2+H(\rho)+U(\bm{r})=F(t),\\
&\frac{\partial \rho}{\partial t}+\nabla\cdot\left(\rho\bm{v}\right)=M(\bm{r},t),
\end{flalign}
\end{subequations}
where $\psi$ is the velocity potential ($\bm{v}=\nabla\psi$), $H(\rho)=\int \mathrm{d} P/\rho=\int \mathrm{d}\rho\,c^2/\rho$ is the enthalpy per unit mass, $c=\sqrt{\mathrm{d} P/\mathrm{d}\rho}$ is the speed of sound, $F(t)$ is an arbitrary integral constant with respect to time, and we have supposed the existence of the external potential of momentum source $U(\bm{r})$ and the external mass source $M(\bm{r},t)$. Indeed, there are two origins of $U(r)$, one is the external bulk force $\bm{f}=-\nabla U_0$, the other corresponds to the mass source, thus $U=U_0+\int M\bm{v}\cdot \mathrm{d}\bm{r}$ (suppose $M\bm{v}$ is integrable). As we will see afterward, the two external sources are necessary.
Considering sound waves traveling on a background flow with the parameters $\{\bm{u}, \psi, \rho,P\}$ obeying Eq.~(\ref{fluid}), the sound waves perform as the perturbations $\{\tilde{\bm{v}}, \tilde\psi, \tilde\rho,\tilde{P}\}$ of the flow. Taking the substitutions $\bm{v}\rightarrow\bm{u}+\tilde{\bm{v}}$, $\psi\rightarrow\psi+\tilde{\psi}$, $\rho\rightarrow\rho+\tilde{\rho}$ into Eq.~(\ref{fluid}), we can obtain the governing equations of sound wave in linear approximation~\cite{unruh1981experimental,unruh1995sonic,visser1998acoustic}
\begin{equation}\label{wave eq}
\begin{split}
\partial_t\left(\frac{\rho}{c^2}\,\partial_t\tilde{\psi}\right)+\partial_t\left(\frac{\rho}{c^2}\,\bm{u}\cdot\nabla\tilde{\psi}\right)
+\nabla\cdot\left(\frac{\rho}{c^2}\,\bm{u}\,\partial_t\tilde{\psi}\right)&\\
+\nabla\cdot\left[\frac{\rho}{c^2}\,\bm{u}\left(\bm{u}\cdot\nabla\tilde{\psi}\right)\right]
-\nabla\cdot\left(\rho\nabla\tilde{\psi}\right)=&0.
\end{split}
\end{equation}
And the sound pressure satisfies
\begin{equation}
\tilde{P}=c^2\tilde\rho=-\rho\left(\partial_t\tilde{\psi}+\bm{u}\cdot\nabla\tilde\psi\right).
\end{equation}
Eq.~(\ref{wave eq}) is identical with the D'Alembert equation in a curved spacetime
\begin{equation}\label{D'Alembert}
\Box\,\tilde{\psi}=\frac{1}{\sqrt{-g}}\partial_\mu\left(\sqrt{-g}g^{\mu\nu}\partial_\nu\tilde{\psi}\right)=0,
\end{equation}
with the effective metric
\begin{equation}\label{}
\left(g_{\mu\nu}\right)=\frac{\rho_1}{c_1}
\left(
\renewcommand\arraystretch{1.6}
\begin{array}{c|@{\hspace{18pt}}c@{\hspace{18pt}}}
-(c^2-u^2) & -u_{i} \\ \hline
-u_i & \gamma_{\,ij}
\end{array}
\right),
\end{equation}
where $\rho_1=\rho/\rho_0$, $c_1=c/c_0$, $\rho_0$ and $c_0$ are constants with dimension of density and velocity respectively, $\gamma_{ij}$ denotes the spatial metric given in an arbitrary coordinate system of flat space, and $\mathrm{d} x^0=\mathrm{d} t$. This form of metric can be regarded as the generalized Schwarzschild metric written in the Painlev\'{e}-Gullstrand type coordinate system (P-G system)~\cite{visser1998acoustic}. If $\bm{u}/(c^2-u^2)$ is integrable, we can obtain the Schwarzschild type coordinate system (Sch.~system), through the coordinate transformation
\begin{equation}\label{transformation}
\mathrm{d}\tau=\mathrm{d} t+\frac{u_i\mathrm{d} x^i}{c^2-u^2},\quad \mathrm{d} x'^i=\mathrm{d} x^i,
\end{equation}
where $\tau$ is the $x^0$ coordinate of Sch.~system. Then the metric turns into the standard form
\begin{equation}\label{metric in sch}
\left(g'_{\mu\nu}\right)=\frac{\rho_1}{c_1}
\left(
\renewcommand\arraystretch{1.6}
\begin{array}{c|c}
-(c^2-u^2) & \mathbf{0} \\\hline
\mathbf{0} & \gamma_{\,ij}+\frac{u_iu_j}{c^2-u^2}
\end{array}
\right).
\end{equation}
In Sch.~system, time dimension and space dimensions are decoupled, nevertheless, the spatial projection of geodesics in this system holds the same form as in P-G system, since the spatial coordinates do not change. As pointed in Ref.~\cite{garcia2014transformational}, a major distinction between ATA and ordinary TA is that one generally deals with the wave equation of sound pressure in TA, while the kernel equation of ATA used to draw the analogy with curved spacetime is the equation of velocity potential. As a result, TA is restricted to design isobaric systems, whereas ATA is demanded to be globally barotropic~\cite{garcia2014transformational}.
According to the procedure of ATA, if we let the effective metric either in P-G system or in Sch. system equal to the requisite form gained by TA, the expressions of $\bm{u}$, $\rho$, $c$, can be determined. However, these quantities should also compose a real solution of fluid dynamics. By plugging this quantities back into Eq.~(\ref{fluid}), the required external sources are fixed. Besides, it is notable that the governing Eq.~(\ref{D'Alembert}) of ATA and Eq.~(\ref{metric in sch}) (if used) are valid only if $\bm{u}$ and $\bm{u}/(c^2-u^2)$ are irrotatinal. All these matters restrict the selective freedom of the background flow.
\begin{figure}[t]
\includegraphics[width=\columnwidth,clip]{fig1}
\caption{\label{fig1}The profiles of (a) radial transformation $f(r)$, (b) velocity of background flow $u(r)$, (c) mass density $\rho$, (d) speed of sound $c(r)$, (e) pressure difference $\Delta P(r)=P-P_0$ in comparison with the pressure $P_0$ at infinite, (f) constitutive relation between $\rho$ and $\Delta P$, (g) external potential $U(r)$, (h) external mass source $M(r)$, where the red, green and blue curves correspond to the transformations given in Eq.~(\ref{f1}) with $n=2,\ n=3,\ \text{and}\ n=4 $ respectively. The quantity written in the square bracket beside each axis gives the unit of the corresponding coordinate axis. }
\end{figure}
Now we show how to construct a spherical cloak with background flow. For spherically symmetric systems, all parameters are merely functions of radius and the flow velocity $\bm{u}$ is radial. Consequently, $\bm{u}$ and $\bm{u}/(c^2-u^2)$ would always be integrable. It is convenient to discuss the problem in spherical coordinates $\{r,\theta,\phi\}$, thus we have $\left(\gamma_{ij}\right)=\mathrm{diag}(1,r^2,r^2\sin^2\theta)$ and $(u_i)=(u,0,0)$. In the light of TA, the effective metric for a spherical cloak reads~\cite{Leonhardt2010Dover}
\begin{equation}
\left(g'_{\mu\nu}\right)=\textrm{diag}\left(-c_0^2,\ f'(r)^2,\ f(r)^2,\ f(r)^2\sin^2\theta\right).
\end{equation}
Here $f(r)$ ia an arbitrary radial transformation satisfying the invisibility condition $f(a)=0$, where $a$ is the inner radius of the cloak. And we demand $f(r)$ obeys the limit
$
\lim_{r\rightarrow \infty}\big(f(r)-r\big)=0,
$
so that the background flow goes to a homogenous and hydrostatic state with $\rho\rightarrow \rho_0,\ u\rightarrow0,\ c\rightarrow c_0,\ P\rightarrow P_0$, when $r\rightarrow\infty$. Letting Eq.~(\ref{metric in sch}) equal to the effective metric of spherical cloak, we obtain
\begin{subequations}\label{relations}
\begin{align}
&\rho\ =\ \rho_0f'(r),\label{rho}\\
&c\ =\ c_0\frac{r^2f'(r)}{f(r)^2},\\
&\bm{u}\ =\ \pm\,c_0\frac{r}{f(r)^2}\sqrt{r^2f'(r)^2-f(r)^2}\ \bm{\hat{e}}_r.\label{u}
\end{align}
\end{subequations}
The signs $\pm$ of $\bm{u}$ indicate that the flow could be either emanative or convergent. In terms of the barotropic postulate, the pressure can be derived by
\begin{equation}\label{pressure}
P(r)=\int c^2\mathrm{d}\rho=c_0^2\rho_0\int_{\infty}^r \frac{r^4f'(r)^2f''(r)}{f(r)^4}\mathrm{d} r+P_0.
\end{equation}
We can further obtain the required state equation of the fluid, namely the constitutive relation between $P$ and $\rho$, in terms of Eq.~(\ref{rho}) and Eq.~(\ref{pressure}). Substituting Eq.~(\ref{relations}) into Eq.~(\ref{fluid}), we get the required external potential and mass source
\begin{subequations}
\begin{align}
\begin{split}
&U(r) = -\frac{1}{2}\bm{u}^2-H(\rho)\\
= &-\frac{c_0^2r^2}{2f(r)^4}\left(r^2f'(r)^2-f(r)^2\right)-c_0^2\int \frac{r^4f'(r)f''(r)}{f(r)^4}\mathrm{d} r,
\end{split}\\
\begin{split}
&M(r) = \nabla\cdot(\rho\bm{u})\\
= &\pm\frac{\rho_0c_0}{r^2}\frac{\mathrm{d}}{\mathrm{d} r}\left(\frac{r^3f'(r)}{f(r)^2}\sqrt{r^2f'(r)^2-f(r)^2}\right).
\end{split}
\end{align}
\end{subequations}
Obviously, the desired background flow would not generally satisfy the momentum and mass conservations unless the above external source terms are provided. In principle, the external potential could be made up with external electric fields acting on charged fluid. The mass sources could be realized via the mass exchange of chemical reactions or via phase transition of multiphase fluid~\cite{asfaw2010prandtl}, or even via discretely distributed jets and outlets.
A class of functions satisfying the conditions $f(a)=0$ and $f(r)\rightarrow r$ as $r\rightarrow\infty$ is
\begin{equation}\label{f1}
f(r)=r-\frac{a^n}{r^{n-1}},\qquad (n\geq2).
\end{equation}
We calculate the profiles of $u,\ \rho,\ c,\ P,\ M$ with respect to $r$ for $n=2,3,4$, and obtain the required state equation $P(\rho)$. The results are shown in Fig.~\ref{fig1}. The figures illustrate the case of outgoing flow from the center ($\bm{u}$ holds the sign of $``+''$ in Eq.~(\ref{u})). In this case, the external potential is attractive and the mass source is negative. If $\bm{u}$ is convergent to the center, the mass source should be positive while other variables do not change.
\section{Geometric acoustics}
In this section, we study the propagation of sound waves with geometric acoustics approximation in the shortwave limit. Under the eikonal hypothesis, the velocity potential takes the form $\tilde{\psi}\sim \mathrm{e}^{\mathrm{i}\Phi(x^\mu)}$, and the phase $\Phi$ satisfies the eikonal equation $H=h(x^\mu)g^{\mu\nu}k_\mu k_\nu=0$, where $k_\mu=\partial_\mu \Phi$, $h(x^\mu)$ can be an arbitrary non-degenerate function, and
\begin{equation}\label{}
\begin{split}
\left(g^{\mu\nu}\right)&=\frac{1}{c_0c\rho_1}
\left(
\renewcommand\arraystretch{1.6}
\begin{array}{@{\hspace{18pt}}c@{\hspace{18pt}}|c}
-1 & -u^{i} \\ \hline
-u^i & c^2\gamma^{ij}-u^iu^j
\end{array}
\right)\\
&=\frac{1}{c_0c\rho_1}\ \left(\tilde{g}^{\mu\nu}\right).
\end{split}
\end{equation}
Sound rays obey the following canonical equations
\begin{equation}\label{canonical equations}
\frac{\mathrm{d} x^\mu}{\mathrm{d} s}=\frac{\partial H}{\partial k_\mu}, \qquad \frac{\mathrm{d} k_\mu}{\mathrm{d} s}=-\frac{\partial H}{\partial x^\mu}.
\end{equation}
It can be verified that the formulas of sound rays are identical with the geodesics in the P-G system with the effective metric. If $g^{\mu\nu}$ is time-independent, $\mathrm{d} k_0/\mathrm{d} s=-\partial H/\partial t\equiv0$, thus the frequency $\omega=-k_0$ is conserved along a ray.
Letting $h(x^\mu)=c_0c\rho_1$, $H=\tilde{g}^{\mu\nu}k_\mu k_\nu=c^2k^2-(\omega-\bm{u}\cdot\bm{k})^2=0$, we obtain $\omega=\bm{u}\cdot\bm{k}+ck$,
which is exactly the dispersion relation of sound waves in moving media~\cite{LANDAU1987192}, and is always applicable in the limit of geometric acoustics no matter if the background flow are irrotational or not. Therefore, although the analogy between background flow and the curved spacetime is rigorous only for potential flow, nevertheless, this restriction can be drawn off in the limit of geometric acoustics. In fact, as proved in Ref.~\cite{Bergliaffa2004Wave}, the analogy is still valid as long as the frequency of the sound is larger enough than the vorticity of the flow $\bm{\Omega}=\nabla\times\bm{u}$: $\omega\gg|
\bm{\Omega}|$, but there is no requirement of the magnitude of the spatial inhomogeneity comparing with wave length.
Substituting the Hamiltonian into the first set of the canonical equations, we obtain
\[
\frac{\mathrm{d} t}{\mathrm{d} s} =2kc,\qquad \frac{\mathrm{d} x^i}{\mathrm{d} s} =2kc\left(u^i+ck^i/k\right).
\]
\begin{figure}
\begin{center}
\includegraphics[width=0.6\columnwidth,clip]{fig2}
\end{center}
\caption{\label{fig2}Sound rays in the acoustic cloaking system. The blue curves denote the sound rays, and the red curves denote wave fronts, where the background flow (the direction is emanative from the center) corresponds to Eq.~(\ref{f1}) with $n=4$. }
\end{figure}
Cancellation of $ds$ leads to
\begin{equation}\label{}
\frac{\mathrm{d} x^i}{\mathrm{d} t} =u^i+ ck^i/k.
\end{equation}
The result is identical with the group velocity of the sound wave, $\bm{v}_g=\nabla_{\bm{k}}\omega=\bm{u}+ c\bm{k}/k$. Actually, for arbitrary eikonal equations $H(-\omega,k_i,x^\mu)=0$,
\[
v_g^i= \frac{\partial \omega}{\partial k_i}=-\frac{\partial H/\partial k_i}{\partial H/\partial \omega}=\frac{\partial H/\partial k_i}{\partial H/\partial k_0}.
\]
Therefore, the group velocity is always along the solution of the canonical equations.
For spherically symmetric metric, $k_\phi$ is conserved along a ray, since $dk_\phi/ds=0$. If we let $k_\phi=0$, then $\phi\equiv\phi_0,$ and $dk_\theta/ds=0$. As a result, the angular momentum of phonon perpendicular to the orbit $\ k_\theta\equiv L$ is also conserved. We can further remove all the wave vector terms in the canonical equations, and obtain
\begin{subequations}
\begin{gather}
\left(1-\frac{u^2}{c^2}\right)\frac{\mathrm{d} t}{\mathrm{d}\sigma}+\frac{u}{c^2}\frac{\mathrm{d} r}{\mathrm{d}\sigma}=\omega,\label{t}\\
\left(\frac{\mathrm{d} r}{\mathrm{d}\sigma}\right)^2=\frac{\omega^2}{c^2}- \left(1-\frac{u^2}{c^2}\right)\frac{L^2}{r^2},\label{r}\\
\qquad\frac{\mathrm{d}\theta}{\mathrm{d}\sigma}=\frac{L}{r^2},\quad\ \phi\equiv\phi_0,\label{theta}
\end{gather}
\end{subequations}
with $\mathrm{d}\sigma=2c^2\mathrm{d} s$. Inserting Eq.~(\ref{theta}) into Eq.~(\ref{r}), we get the compact form of the geodesic equation
\begin{equation}\label{ray equation}
\begin{split}
\left(\frac{1}{r^2}\frac{\mathrm{d} r}{\mathrm{d}\theta}\right)^2 &=\frac{\omega^2}{L^2}\frac{1}{c^2}-\left(1-\frac{u^2}{c^2}\right)\frac{1}{r^2}.
\end{split}
\end{equation}
Substitution the parameters of spherical cloak yields
\begin{equation}
\mathrm{d}\theta=\pm \frac{\mathrm{d} f}{f\sqrt{\frac{\omega^2}{L^2c_0^2}f^2-1}}.
\end{equation}
By integration, we obtain the general analytic solution of the sound rays
\begin{equation}\label{analytic ray function}
f(r)\sin(\theta-\theta_0)=\pm Lc_0/\omega,\quad \phi\equiv\phi_0,
\end{equation}
where $Lc_0/\omega$ denotes the impact parameter of the phonon incident from infinity. Actually, Eq.~(\ref{analytic ray function}) is exactly transformed from the straight lines $r\sin(\theta-\theta_0)=\mathrm{const.}$ according to TA. Moreover, based on Eq.~(\ref{ray equation}), the sound rays only depend on $\bm{u}$ and $c$ but are independent of the density distribution $\rho$, therefore, we have a freedom to select $\rho$ in the limit of geometric acoustics.
Fig.~\ref{fig2} displays a set of rays (blue curves) incident with the same direction but different impact parameters. Despite the same trajectories as in TA cloak, the wave fronts $\Phi=\mathrm{const.}$, shown by the red curves in Fig.~\ref{fig2}, significantly differ from those in TA cloak. In our case, the wave fronts before bypassing the cloak are asymmetric to those after bypassing the cloak, however, this symmetry exists in TA cloak~\cite{Schurig2006OE,cummer2007one,chen2007acoustic,cummer2008scattering}. This effect presents the phase retardation of the waves when bypassing the cloak, and it is essentially induced by the background flow $\bm{u}$. In the next section, we will further investigate this effect.
\section{Analytic solution}
According to Eq.~(\ref{transformation}), by applying the rules
\[
\partial_t\rightarrow\partial_\tau,\quad \partial_i\rightarrow\partial_i+\frac{u_i}{c^2-u^2}\partial_\tau,
\]
the wave equation (\ref{wave eq}) in Sch. system takes the form
\begin{equation}\label{wave eq2}
\frac{\rho}{c^2-u^2}\,\partial_\tau^2\,\tilde{\psi}-\frac{1}{\sqrt{\gamma}}\,\partial_i\left[\sqrt{\gamma}\,\frac{\rho}{c^2}\left(c^2\gamma^{ij}-u^iu^j\right)\partial_j\,\tilde{\psi}\right]=0.
\end{equation}
Substituting Eqs.~(\ref{relations}) into Eq.~(\ref{wave eq2}), we have
\begin{equation}\label{wave eq3}
\begin{split}
\left\{\frac{1}{f^2}\frac{\partial}{\partial f}\left(f^2\frac{\partial \tilde{\psi}}{\partial f}\right)+\frac{1}{f^2\sin\theta}\frac{\partial}{\partial \theta}\left(\sin\theta\frac{\partial\tilde{\psi}}{\partial\theta}\right)\right.&\\
\left.+\frac{1}{f^2\sin^2\theta}\frac{\partial^2\tilde{\psi}}{\partial\phi^2}\right\}-\frac{1}{c_0^2}\,\frac{\partial^2}{\partial\tau^2}\,\tilde{\psi}&=0.
\end{split}
\end{equation}
Considering monofrequent wave $\tilde{\psi}(\bm{r},t)=\tilde{\psi}_0(\bm{r})\mathrm{e}^{-\mathrm{i}\omega t}$, since $t=\tau-\int^r_b\frac{u\mathrm{d} r}{c^2-u^2}$, the velocity potential can be expressed as
$
\tilde{\psi}(\bm{r},t
=\tilde{\psi}(\bm{r})\mathrm{e}^{-\mathrm{i}\omega \tau}
$
with $\tilde{\psi}(\bm{r})=\tilde{\psi}_0(\bm{r})\exp\left[\mathrm{i}\,\omega\int^r_b\frac{u\mathrm{d} r}{c^2-u^2}\right]$. Then Eq.~(\ref{wave eq3}) becomes
\begin{equation}\label{wave eq in sch}
\begin{split}
\frac{1}{f^2}\frac{\partial}{\partial f}\left(f^2\frac{\partial \tilde{\psi}}{\partial f}\right)+\frac{1}{f^2\sin\theta}\frac{\partial}{\partial \theta}\left(\sin\theta\frac{\partial\tilde{\psi}}{\partial\theta}\right)&\\
+\frac{1}{f^2\sin^2\theta}\frac{\partial^2\tilde{\psi}}{\partial\phi^2}+\tilde{k}_0^2\tilde{\psi}&=0,
\end{split}
\end{equation}
where $\tilde{k}_0=\omega/c_0$ and $\tilde{\psi}$ is the abbreviation of $\tilde{\psi}(\bm{r})$. By separation of variables, $\tilde{\psi}(\bm{r})=R(f)Y(\theta,\phi)$, the radial equation of $R(f)$ is spherical Bessel equation, and the angular equation is spherical harmonious equation. Therefore, the solutions can be expressed by
\begin{subequations}\label{solution}
\begin{flalign}
\tilde{\psi}^{\mathrm{in}}\ &=\ \sum_{lm}\left[A^{\mathrm{in}}_l\, j_l(\tilde{k}_0f)+B^{\mathrm{in}}_l\, n_l(\tilde{k}_0f)\right]Y_{lm}(\theta,\phi),\\
\tilde{\psi}^{\mathrm{s}}\ &=\ \sum_{lm}A^{\mathrm{s}}_l\, h^{(1)}_l(\tilde{k}_0f)Y_{lm}(\theta,\phi),
\end{flalign}
\end{subequations}
where $\tilde{\psi}^{\mathrm{in}}$ and $\tilde{\psi}^{\mathrm{s}}$ are the velocity potentials of incident wave and scattering wave respectively, $j_l$, $n_l$, $h^{(1)}_l$ denote the $l$ order spherical Bessel, Neumann, and first Hankel functions respectively, and $Y_{lm}(\theta,\phi)$ denotes the $(l,m)$ order spherical harmonics. We demand that the incident wave tends to plane wave at infinity:
\begin{equation}\label{}
\begin{split}
\lim_{r\rightarrow\infty}\tilde{\psi}^{\mathrm{in}}\mathrm{e}^{-\mathrm{i}\omega\tau} & \sim Ae^{i(\tilde{k}_0r\cos\theta-\omega t)}\\
& =\sum_{l=0}^\infty a_l j_l(\tilde{k}_0r)P_l(\cos\theta)\mathrm{e}^{-\mathrm{i}\omega t},
\end{split}
\end{equation}
with $a_l=A(2l+1)\mathrm{i}^l$. The incident wave thus can be written as
\begin{align}\label{boundary condition}
&\tilde{\psi}^{\mathrm{in}}(\bm{r},t) = \sum_{l=0}^\infty \tilde{a}_l j_l(\tilde{k}_0f)P_l(\cos\theta)\mathrm{e}^{-\mathrm{i}\omega\tau}\nonumber\\
= &A\exp\left[\mathrm{i}\tilde{k}_0f(r)\cos\theta-\mathrm{i}\omega\int^r_\infty\frac{u\mathrm{d} r}{c^2-u^2}\right]\mathrm{e}^{-\mathrm{i}\omega t}\\ \
= &A\exp\left[\mathrm{i}\tilde{k}_0\left(f(r)\cos\theta\mp\int^r_\infty\frac{\sqrt{r^2f'^2-f^2}}{r}\mathrm{d} r\right)\right]\mathrm{e}^{-\mathrm{i}\omega t},\nonumber
\end{align}
where $\tilde{a}_l=a_l\exp\left[\int^\infty_b\frac{u\mathrm{d} r}{c^2-u^2}\right]$, $P_l(\cos\theta)$ is the $l$ order Legendre polynomial, and the signs $\mp$ in the last line correspond to emanative and convergent flows respectively (the signs $\mp$ have the same meaning in all of the following equations). At the inner interface $r=a$, we suppose there is a hard boundary of the sound wave:
\begin{gather}
\left.\left(v^{\mathrm{in}}_r+v^{\mathrm{s}}_r\right)\right|_{r=a}=\left.\left(\frac{\partial\tilde{\psi}^{\mathrm{in}}_0}{\partial r}+\frac{\partial\tilde{\psi}^{\mathrm{s}}_0}{\partial r}\right)\right|_{r=a}=0.\\
\text{(P-G system)}\notag
\end{gather}
\begin{figure}[!t]
\centering
\includegraphics[width=0.9\columnwidth,clip]{fig3}
\caption{\label{fig3}Distributions of (a) velocity potential $\tilde{\psi}$, (b) sound pressure $\tilde{P}$, in $x=0$ plane, of the sound wave incident from the left side, where the background flow (emanative from the center) corresponds to Eq.~(\ref{f1}) with $n=4$.}
\end{figure}
In Sch. system, the boundary condition turns into
\begin{gather}
\left(\frac{\partial}{\partial r}+\frac{u}{c^2-u^2}\frac{\partial}{\partial\tau}\right)\left.\left(\tilde{\psi}^{\mathrm{in}}_r+\tilde{\psi}^{\mathrm{s}}_r\right)\right|_{r=a}=0.\\
\text{(Sch. system)}\notag
\end{gather}
Then the scattering wave is derived as
\begin{equation}\label{scattering}
\begin{split}
\tilde{\psi}^{\mathrm{s}}(\bm{r},t) = &\sum_{l=0}^\infty \frac{\tilde{j}'_l(x_a)}{\tilde{h}^{(1)\prime}_l(x_a)}\tilde{a}_l h^{(1)}_l(\tilde{k}_0f)P_l(\cos\theta)\mathrm{e}^{-\mathrm{i}\omega\tau}\\
= &\sum_{l=0}^\infty \frac{\tilde{j}'_l(x_a)}{\tilde{h}^{(1)\prime}_l(x_a)}a_l h^{(1)}_l(\tilde{k}_0f)P_l(\cos\theta)\\
&\cdot\exp\left[\mp\,\mathrm{i}\tilde{k}_0\int^r_\infty\frac{\sqrt{r^2f'^2-f^2}}{r}dr\right]\mathrm{e}^{-\mathrm{i}\omega t},
\end{split}
\end{equation}
with $x_a=\tilde{k}_0f(a)$, and
\begin{subequations}
\begin{flalign}
\tilde{j}'_l(x_a)&=j'_l(x_a)\mp i\textstyle{\sqrt{1-\frac{f(a)^2}{a^2f'(a)^2}}}j_l(x_a),\\
\tilde{h}^{(1)\prime}_l(x_a)&=h^{(1)\prime}_l(x_a)\mp i\textstyle{\sqrt{1-\frac{f(a)^2}{a^2f'(a)^2}}}h^{(1)}_l(x_a).
\end{flalign}
\end{subequations}
If $f(a)=0$, the scattering wave $\tilde{\psi}^{\mathrm{s}}\equiv0$. Consequently, we accomplish ideal acoustic invisibility cloaking by means of background flow. The sound pressure can be obtained from
\begin{equation}\label{sound pressure}
\tilde{P}(\bm{r},t) = \rho\left(i\omega-u\frac{\partial}{\partial r}\right)\left(\tilde{\psi}^{\mathrm{in}}(\bm{r},t)+\tilde{\psi}^{\mathrm{s}}(\bm{r},t)\right).
\end{equation}
For the ideal case, Eq.~(\ref{sound pressure}) reduces to
\begin{align}\label{sound pressure2}
&\tilde{P}(\bm{r},t)\nonumber\\
= &\mathrm{i}\omega\rho_0f'^2\left(f'\mp\frac{r}{f^2}\sqrt{r^2f'^2-f^2}\cos\theta\right)\\
\cdot & A\exp\left[\mathrm{i}\,\tilde{k}_0\left(f(r)\cos\theta\mp\int^r_\infty\frac{\sqrt{r^2f'^2-f^2}}{r}dr\right)\right]\mathrm{e}^{-\mathrm{i}\omega t}.\nonumber
\end{align}
\begin{figure}[!t]
\centering
\includegraphics[width=0.9\columnwidth,clip]{fig4}
\caption{\label{fig4}Distributions of (a) phase velocity $v_p$, (b) group velocity $v_g$, in $x=0$ plane, of the sound wave incident from the left side, where the background flow (emanative from the center) corresponds to Eq.~(\ref{f1}) with $n=4$.}
\end{figure}
Fig.~\ref{fig3} shows the field distributions of both velocity potential $\tilde\psi$ and sound pressure $\tilde{P}$ for the case of idea cloaking with emanative background flow. As shown in the figures, the fields are asymmetric with respect to the $z=0$ plane in contrast to the case of TA cloak. For the incident half place ($z<0$), the phases of the fields are compressed heavily, while the phases are stretched in the outgoing half place ($z>0$). According to the expression of the wave fronts
\begin{equation}\label{wave front}
\Phi=\tilde{k}_0\left[f(r)\cos\theta\mp\int^r_\infty\frac{\sqrt{r^2f'^2-f^2}}{r}\mathrm{d} r\right]=\text{const.},
\end{equation}
the phase retardation and lead originates exactly from the additional phase $\int^r_\infty\frac{u\mathrm{d} r}{c^2-u^2}=\int^r_\infty\frac{\sqrt{r^2f'^2-f^2}}{r}\mathrm{d} r$. Besides, if $u$ is convergent to the center, the effect turns into opposite, namely the phases are stretched in the incident half place while are compressed in the outgoing half, and the field distributions shown in Fig.~\ref{fig3} are identical with the situation of a wave incident form the right side.
This effect also can be viewed in the light of the sound velocities. According to Eq.~(\ref{sound pressure2}), the wave vector reads
\begin{equation}
\begin{split}
&\bm{k}=\nabla\Phi\\
=&\tilde{k}_0\left[\left(f'\cos\theta\mp\frac{1}{r}\sqrt{r^2f'^2-f^2}\right)\bm{\hat{e}}_r-\frac{f\sin\theta}{r}\,\bm{\hat{e}}_\theta\right].
\end{split}
\end{equation}
Hence, the phase velocity of the sound can be obtained by $v_p=\omega/k$ with $k=\tilde{k}_0\big(f'\mp\sqrt{r^2f'^2-f^2}\cos\theta/r\big)$. Note that $c$ is the phase velocity of sound wave in static fluid, but $v_p\neq c$ in moving media.
And the group velocity is
\begin{equation}
\bm{v}_g=\bm{u}+ c\bm{k}/k=v_p\left(\cos\theta\bm{\hat{e}}_r-\frac{rf'(r)}{f(r)}\bm{\hat{e}}_\theta\right).
\end{equation}
Fig.~\ref{fig4} exhibits the distributions of phase velocity and group velocity corresponding to the sound fields given in Fig.~\ref{fig3}. As we can see, the phase velocity and the group velocity slow down as the sound wave approaches the inner boundary $r=a$ from the left side, whereas the velocities have extremely large values at the right side of inner boundary. For the case of convergent back ground flow, the distributions in Fig.~\ref{fig4} should yet correspond to the wave incident form the right side. It is precisely the asymmetric distribution of phase velocity that causes the phase retardation and lead in corresponding regions.
Noteworthy, an alternative scheme of acoustic cloak based on velocity potential wave equation has been porposed in Ref.~\cite{garcia2014transformational}.
In spite of the remarkable differences between their results and the ordinary TA cloaks treated with sound pressure equation, their proposal is more similar to a standard TA device than ours, since their primary method is still to construct effective anisotropic media with the conventional procedure used in TA~\cite{Pendry2006Sci,Schurig2006OE}, but our key idea totally originates from the analogy between the moving media and curved spacetime.
\begin{figure}[!t]
\centering
\includegraphics[width=0.8\columnwidth,clip]{fig5}
\caption{\label{fig5}The angular distributions of differential scattering cross sections of unideal cloaks at the wave lengthes: (a) $\lambda=0.5a$, (b) $\lambda=a$, (c) $\lambda=2a$. (d) The total scattering cross section versus wave length. In each figure, the black curve denotes the scatter of a ball with unit radius and hard boundary, the red, blue, orange, green curves correspond to the unideal cloaks of $\delta=0.1,\,0.05,\,0.02,\,0.01$ respectively. }
\end{figure}
\section{Unideal cloaking}
To realize ideal cloaking, all parameters of the background flow should tend to infinity at the inner boundary $r=a$ as shown in Fig.~\ref{fig1}, however, it is impossible in practice. In this section, we would discuss the unideal case with simply cutting off the infinite parameters. We still use the radial transformation $f(r)$ to construct the background flow, but let its zero point be $a'=(1-\delta)a$ in order to escape the singularity at the inner boundary, when $\delta\rightarrow0$ the cloak approaches to an ideal one. In terms of Eq.~(\ref{scattering}), the scattering wave tends to
\begin{equation}
\tilde{\psi}^{\rm s}\rightarrow A\frac{\mathrm{e}^{\mathrm{i}\,kr}}{kr}\sum_{l=0}^\infty (2l+1)\mathrm{i}\, A_l P_l(\cos\theta), \quad \text{as }r\rightarrow\infty,
\end{equation}
with $A_l=\tilde{j}'_l(x_a)/\tilde{h}_l^{(1)\prime}(x_a)$. Therefore, we can evaluate the cloaking ability of the background flow by means of the differential scattering cross section (DSCS)
\begin{equation}
\frac{\mathrm{d}\sigma}{\mathrm{d}\Omega}=\left|\frac{1}{k}\sum_{l=0}^\infty(2l+1)\mathrm{i} A_lP_l(\cos\theta)\right|^2,
\end{equation}
and the total scattering cross section (TSCS) $\sigma=\oint\frac{\mathrm{d}\sigma}{\mathrm{d}\Omega}\sin\theta{\mathrm{d}\phi}{\mathrm{d}\theta}$. We have checked the angular distributions of DSCS, and TSCS changing with wave length for $\delta=0.1$, $\delta=0.05$, $\delta=0.02$, and $\delta=0.01$ in comparison with the uncloaking results, as shown in Fig.~\ref{fig5}. The results reveal that the unideal constructions still have good property of cloaking for nearly full band of wave length as $\delta\leq 0.05$. However, for the case of $\delta=0.1$, only when $\lambda>2a$, its invisibility effect is acceptable. In addition, the angular distributions of DSCS show that the wave scatters quite uniformly in all directions when the background flow exists, whereas there are usually large components of back scattering in the uncloaking cases.
\begin{figure}[!t]
\centering
\includegraphics[width=0.6\columnwidth,clip]{fig6}
\caption{\label{fig6}Constitutive relation between pressure $P$ and density $\rho$ fitted by 5 order Virial expansion (blue curve). The red dotted curve denotes the required constitutive relation corresponding to the unideal cloak of $\delta=0.1$ constructed by the transformation function Eq.~(\ref{f1}) with $n=4$.}
\end{figure}
Finally, we shall discuss the required state equation of the background fluid. Though the constitutive relation between $P$ and $\rho$ is determined by the coordinate transformation, we can use Virial expansion to approach our constitutive relation. The Virial expansion is a physical state equation that can be derived directly from statistical thermodynamics, which takes the form
\begin{equation}
P=k_BT\rho+B_2(T)\rho^2+B_3(T)\rho^3+\cdots,
\end{equation}
where $k_B$, $T$ denote Boltzmann constant and temperature respectively, and $B_l$ is the $l$th order Virial coefficient. In Fig.~\ref{fig6}, we fit the $P-\rho$ relation for the unideal cloak of $\delta=0.1$, corresponding to Eq.~(\ref{f1}) with $n=4$, by 5 order Virial expansion. The fitting coefficients are $ B_1= 1994.73$, $B_2=-2430.51$, $B_3=1447.21$, $B_4=-422.527$, $B_5=48.7926$, and the pressure at infinity is $P_0=624.479$ (all coefficients are nondimensionalized in terms of the unit quantities $\rho_0$, $c_0$). Note that the expansion is not unique, since we can choose different orders to fit.
\section{Conclusion}
To summarize, we design an acoustic cloaking scheme by using background flow based on the method of ATA, and give all of the required physical parameters of the flow complying with the dynamic laws of fluid. We reveal that it is not enough to merely give the background velocity as previous works just did~\cite{garcia2013analogue,garcia2014space,garcia2014analogue}, but some restrictions should be satisfied, and some other necessities, such as the external mass and momentum sources, should be supplied. We also provide a comprehensive investigation about the propagation of sound waves in our cloaking system. According to our results, the major distinctive effect in our system is the phase retardation and lead when sounds bypass the cloak. Furthermore, we analyze the cloaking ability of unideal constructions and use Virial equation of state to fit the required constitutive relation of the background fluid. Our research not only offers a novel path to achieve acoustic cloaking but also complements the framework of ATA.
|
\section{Introduction}
Feynman diagrammatic technique is a powerful tool of statistical mechanics. Among the hallmarks of the method are the ability to deal---both analytically and numerically---with the thermodynamic limit rather than a finite-size cluster, the possibility of partial summations up to infinite order, and
the fully self-consistent formulation in terms of renormalized (dressed) quantities.
The latter properties allow one to go beyond the Taylor expansion in terms of the coupling constant
or any other parameter.
Advantages of the diagrammatic technique come at a price. The most serious issue is the
divergence of expansions in powers of the coupling constant for systems prone to
Dyson's collapse\cite{dyson1952divergence} ({\it i.e.}, pathological system behavior when the
coupling constant is rotated in the complex plane). For partial summation techniques to work,
the non-perturbed part of the theory has to be Gaussian
(in terms of either real, or complex, or Grassmann variables) to ensure the validity of Wick's theorem.
These issues are often related: for example, Ising and XY models formulated in terms of original
spin variables do not suffer from Dyson's collapse but lack the Gaussian (non-interacting) limit,
while their classical (lattice) field counterparts with the well-defined Gaussian limit are subject
to Dyson's collapse. It would be a mistake, however, to think that meaningful
diagrammatic series are only possible for a very limited class of Hamiltonians, namely,
when the original system is that of interacting lattice fermions.
As already clearly explained by Samuel in a series of papers,\cite{Samuel_anticommuting1,Samuel_anticommuting2,Samuel_anticommuting3}
a broad class of classical spin and dimer models can be reformulated in terms of
familiar interacting fermions and studied with field-theoretical techniques.
Similarly, rather arbitrary quantum spin/boson lattice models can be rigorously
mapped onto fermionic field theories.\cite{PopFed,misha,PopovFedotovG}
As expected, grassmannian formulations of spin/link/boson models with local constraints
are generically strongly-coupled theories at low temperature, and even the most advanced
self-consistent treatments based on the lowest-order graphs are not supposed to provide
quantitatively (and often qualitatively) accurate answers.
Moreover, these theories may contain arbitrary multi-particle interaction vertexes, which
further complicate the structure of the diagrammatic expansion.
One of the promising numerical techniques currently under development for strongly
correlated systems is diagrammatic Monte Carlo (DiagMC).
It is based on the stochastic evaluation of irreducible Feynman graphs up to some high order
and can be implemented in a number of ways, from perturbative expansions in powers
of the coupling constant to various self-consistent skeleton schemes based on fully renormalized
one- or two-body propagators. In such contexts as resonant fermions,\cite{VanHoucke2012}
frustrated magnetism,\cite{kulagin2013bdm, kulagin2013bdm2} and out-of-equilibrium impurity-like
models\cite{Profumo2015, Cohen_out_of_eq_2015} the method was recently shown to be able to go significantly beyond the state of the art. Also, significant progress has been
made in understanding superfluid properties of the Hubbard-type models.\cite{gukelberger2014pss,Deng2015,Gukelberger2015}
Notably, the infamous sign-problem
preventing conventional Monte Carlo methods from simulating fermionic system with sizes
large enough for reliable extrapolation to the thermodynamic limit, is absent as such in
DiagMC. Instead, the computational complexity is now linked to the number of diagrams
growing factorially with their order. Nevertheless, millions of diagrams can be accounted
for and the approach is flexible enough to deal with an arbitrary interaction Hamiltonian/action.
The current paradigm for generic lattice gauge models, as they occur in lattice-QCD as well
as in solid state and ultra-cold atomic physics, is to work with finite-size systems and
to treat link variables separately from the fermionic sector. More precisely, link
variables are simulated using classical Monte Carlo techniques (with local updates),
and fermions (quarks) are described by determinants. This approach suffers from a severe
sign-problem for finite density of fermions (non-zero chemical potential).\cite{QCDsignproblem,QCDsignproblem2}
If link variables are straightforwardly represented by bosonic fields, then the thermodynamic
limit can be addressed within the diagrammatic approach that treats bosonic and fermionic degrees
of freedom on equal footing. However, in this formulation the bosonic fields pose a
fundamental problem, which manifests itself in a zero convergence radius. It is thus desirable
to have a generic scheme for replacing link variables with Grassmann fields to ensure that the
diagrammatic expansion has proper analytic properties around the Gaussian point.
In this paper, we introduce a general procedure of {\it grassmannization} for
classical lattice models. It is by no means a unique one, and in certain specific cases
more compact/simpler representations can be found. There is a strong connection to the anti-commuting variables approach introduced
by S. Samuel,\cite{Samuel_anticommuting1, Samuel_anticommuting2, Samuel_anticommuting3} which can solve
the 2D Ising model exactly (free fermion operators to solve the Ising model exactly were first found by Kaufman\cite{Kaufman1949} and
refined by Schultz, Mattis and Lieb\cite{SchultzMattisLieb1964}) and provides a good starting point for field-theoretic studies of the 3D Ising model.
For the latter system our approach amounts to an alternative but equally complicated field theory.
Our prime goal is to build on these ideas and develop a scheme that is flexible enough to apply to a broader class of link models with arbitrary multi-bond
interactions and local constraints.
The idea of grassmannization is to represent the partition function of the model as a Grassmann
integral from the exponential of a Grassmann functional. The Feynman rules then emerge by
Taylor-expanding the non-Gaussian part of the exponential and applying Wick's theorem to the Gaussian averages. Paradigmatic lattice systems are link and plaquette
models featuring discrete degrees of freedom---integer numbers---residing on links (plaquettes) of square lattices and subject to certain
local constraints in terms of the allowed values of the sum of all link (plaquette) variables
adjacent to a given site (edge). It turns out that it is these constraints that require
special tricks involving multiple Grassmann variables for each value of each discrete variable.
Link models often emerge as high-temperature expansions of
lattice systems\cite{Oitmaa} in Ising, XY, O(3), etc. universality classes no matter whether the original
degrees of freedom are discrete or continuous (e.g., classical vector-field variables).
Link models may also emerge as dual (low-temperature) expansions, and specific examples
are provided by the 2D Ising model\cite{mccoy1973} and the 3D $|\psi|^4$ model (the latter case leads to the
so-called J-current model with long-range interactions). Similarly, plaquette models emerge as a
high-temperature expansion of lattice gauge theories, but sometimes they represent the dual
(low-temperature) expansion, as in the case of the 3D Ising model. Finally, it is worth mentioning
how the models with the same general structure are generated by strong-coupling expansions
in lattice-QCD.\cite{Wilsonloop}
The paper is structured as follows. In Sec.~\ref{sec:II} we explain how a partition function
of a discrete link model can be written as a Grassmann integral. The equivalence between the two
formulations is readily proved through term-by-term comparison. Standard properties of Grassmann
variables then immediately allow one to express the Grassmann weight in the exponential form in order to
define the field-theory. In Sec.~\ref{sec:III} we discuss generalizations of the proposed
grassmannization scheme. We start by describing the procedure for a broad class of plaquette models.
Next we show a simple way to introduce Grassmann variables for non-local link models with pairwise
interactions between the link variables. The construction is further simplified when constraints
are replaced with statistical penalties for certain configurations of link (plaquette) variables.
We conclude this section with defining the meaning of the term ``order of expansion" for the
resulting field theory. In Sec.~\ref{sec:IV} we deliberately choose the most general
grassmannization scheme for the 2D Ising model to illustrate and test how our construction
works in practice. We stress that our goal is not to solve the 2D Ising model
exactly\cite{Onsager44, mccoy1973} or determine a series expansion
for it\cite{Oitmaa} but to develop a general framework---including numeric component---for applying Grassmann variables
to link and plaquette models and show that its evaluation can be done realistically. After determining all field-theoretic parameters, characterizing various
interaction terms and source operators for calculating correlation functions (with and without magnetic
field), and explaining Feynman rules for constructing the perturbative expansion, we proceed with
the description of algorithms to compute them (Monte Carlo and deterministic)
in Sec.~\ref{sec:V}. Results are presented and discussed in Sec.~\ref{sec:VI}.
By comparing with the exact solution we show that the critical exponent
$\gamma$ for magnetic susceptibility could be determined with an accuracy of about $5$\%, while
the critical point could be located with sub-percent accuracy. In Sec.~\ref{sec:VII} we discuss the
implementation of the self-consistent skeleton technique within the so-called $G^2W$-expansion\cite{Heidin}
which computes irreducible (skeleton) diagrams for the self-energy and ``polarization" function
and uses them in the Dyson equations in order to find the renormalized propagators and screened interactions.
We also present results that emerge when this technique is based on several low-order diagrams.
We briefly comment in Sec.~\ref{sec:VIII} that both bare-series and the $G^2W$-expansion methods readily
solve the 1D Ising model exactly. We conclude with prospects for future work in Sec.~\ref{sec:IX}.
\section{Grassmannization of local link models}
\label{sec:II}
\subsection{Local link models}
For the purposes of this article, we mean by a link model a classical statistical model
with states labeled by a set of {\it discrete} variables $\{ \alpha_b \}$ residing on
links (bonds) of a certain lattice. In addition, we require that the ground state
is unique. Without loss of generality, it can be chosen to be the state with $\alpha_b=0$
on each link $b$.
We further narrow the class of link models---to which we will refer to as {\it local} link models---by the requirement that the statistical weight of a state factors into a product of link and site weights
(to be referred to as link and site factors, respectively). A link factor, $f_b$, is a function
of the corresponding link variable, $f_b\equiv f(\alpha_b)$. The site factor, $g_j$, is a function that
depends on all variables residing on links attached to the site $j$, denoted as $\{ \alpha_b \}_j$. Then, $g_j \equiv g(\{ \alpha_b \}_j)$.
Solely for the purpose of avoiding heavy notations, we consider translational invariance when $f(\alpha_b) \equiv f_b$ is the same function on all links and $g_j$ site independent, $g_j \equiv g$.
Given that only the relative weights of the states matter,
we set $f(0)=1$ and $g(0_j)=1$, where $0_j$ stands for the $\{ \alpha_b=0 \}_j$ set.
The site factors play the key role in link models. They describe interactions between
(otherwise independent) link degrees of freedom. In particular, this interaction can take the extreme form
of a {\it constraint} on the allowed physical configurations of $\{ \alpha_b \}_j$ ({\it e.g.,} the zero-divergency constraint in J-current models,\cite{Villain}
or the even-number constraint in the high-temperature expansion
of $Z_2$ models), in which case $g_j(\{ \alpha_b \}_j)$ is identically {\it zero} for each non-physical state
of $\{ \alpha_b \}_j$.
\subsection{Grassmannization}
\begin{figure}[tbp]
\centering
\includegraphics[width=0.18\columnwidth, angle=-90]{fig_fields.pdf}
\caption{Assignment of Grassmann fields for link (left) and site (right) factors.
Upon integration, the labels of the Grassmann variables must be equal in order to
connect variables from all factors (see text).
}
\label{fig:fields}
\end{figure}
For each label $\alpha \neq 0$ of the link $b$, introduce four Grassmann variables: $\xi_{\alpha,b}$, $\xi_{\alpha,b}'$, $\bar{\xi}_{\alpha,b}$, and $\bar{\xi}_{\alpha,b}'$. For a textbook
introduction to Grassmann variables, we refer to Ref.~\onlinecite{negele1988quantum}.
For $\alpha = 0$ we assume that
$\xi_{0,b}=\xi_{0,b}'=\bar{\xi}_{0,b}=\bar{\xi}_{0,b}'=1$.
In terms of these variables, define the Grassmann weight---a product of link, $A_b$, and site, $B_j$, factors such that tracing over all degrees of freedom yields the partition function $Z = {\rm Tr} \prod A_b \prod B_j$---by the following rules,
\begin{eqnarray}
A_b & = & \exp \left\{ \sum_{\alpha \neq 0} \left[
\frac{\bar{\xi}_{\alpha, b}' \xi_{\alpha, b}'}{\sqrt{f(\alpha)}} +
\frac{\bar{\xi}_{\alpha, b} \xi_{\alpha, b}}{\sqrt{ f(\alpha)}} \, \right]\right\} \nonumber \\
& = & \prod_{\alpha \neq 0}
\, \exp \left\{ \frac{\bar{\xi}_{\alpha, b}' \xi_{\alpha, b}'}{\sqrt{f(\alpha)}} +
\frac{\bar{\xi}_{\alpha, b} \xi_{\alpha, b}}{\sqrt{ f(\alpha)}} \right\} \, ,
\label{link}
\end{eqnarray}
\begin{eqnarray}
B_j\, =\, \sum_{ \{ \alpha_b \}_j} g(\{ \alpha_b \}_j) \prod_{b \in \{ b \}_j} \breve{\xi}_{\alpha_{b}, b}\, \breve{\xi}_{\alpha_{b}, b}^*\nonumber \\
\, =\, 1 \, + \sum_{ \{ \alpha_b \}_j \neq 0_j} g(\{ \alpha_b \}_j) \prod_{b \in \{ b \}_j} \breve{\xi}_{\alpha_{b}, b}\, \breve{\xi}_{\alpha_{b}, b}^* \, .
\label{site}
\end{eqnarray}
Here $\{ b \}_j$ stands for the set of all links incident to the site $j$, and variables
$\breve{\xi}_{\alpha_{b}, b}$ and $\breve{\xi}_{\alpha_{b}, b}^*$ are defined differently
for different links. We first introduce the notion of direction (on each link)
so that one of the two link ends becomes ``incoming" and its counterpart ``outgoing"
(with respect to the site adjacent to the end). Next, we assign (see Fig.~\ref{fig:fields} for an illustration)
\begin{equation}
\begin{array}{*{5}{l}}
\breve{\xi}_{\alpha_b, b} =\xi_{\alpha_b, b}\, , \;\; \breve{\xi}_{\alpha_b, b}^*= \xi_{\alpha_b, b}' \; \; \mbox{(for incoming end)} ,\\
\breve{\xi}_{\alpha_b, b} =\bar{\xi}_{\alpha_b, b}'\, , \;\; \breve{\xi}_{\alpha_b, b}^*= \bar{\xi}_{\alpha_b, b} \; \; \mbox{(for outgoing end) .}
\end{array}
\label{convention}
\end{equation}
The claim is that the Grassmann integral of the weight over all variables reproduces the partition function of the original link model. For a link $b$ to yield a non-zero contribution to the integral the link labels in (\ref{site}) for the sites of the incoming ($j=1$) and outgoing ($j=2$) ends of the link should match each other:
$\alpha_1=\alpha_2$. Indeed, at $\alpha_1\neq \alpha_2$, it is not possible to find
an appropriate term in the expansion of the link exponential (\ref{link})
such that---upon multiplying by the site factors $\breve{\xi}_{\alpha_1, b} \, \breve{\xi}_{\alpha_1, b}^* $ and $\breve{\xi}_{\alpha_2, b} \, \breve{\xi}_{\alpha_2, b}^*$---all powers of the Grassmann variables $\xi_{\alpha_1,b}$, $\xi_{\alpha_1, b}'$, $\bar{\xi}_{\alpha_1, b}$, $\bar{\xi}_{\alpha_1, b}'$, $\xi_{\alpha_2, b}$, $\xi_{\alpha_2, b}'$, $\bar{\xi}_{\alpha_2, b}$, $\bar{\xi}_{\alpha_2, b}'$ are exactly equal to 1 to ensure that the Grassmann integral is non-zero.
For $\alpha_1=\alpha_2\equiv \alpha$,
we need to consider two cases: $\alpha =0$ and $\alpha \neq 0$.
In the first case, the non-zero contribution to the integral comes from the product of second terms in the expansion of the link exponentials (\ref{link}):
\begin{eqnarray}
&\!& \!\!\!\!\!\!\!\!
\prod_{\gamma \neq 0} \int \mathcal{D} [ \bar{\xi}' \xi' \bar{\xi} \xi ]_{\gamma }
\, \exp \left\{ \frac{\bar{\xi}_{\gamma}' \xi_{\gamma}'}{\sqrt{f(\gamma)}} +
\frac{\bar{\xi}_{\gamma} \xi_{\gamma}}{\sqrt{ f(\gamma)}} \right\}
\nonumber \\
&=&\prod_{\gamma \neq 0} \int \mathcal{D} [ \bar{\xi}' \xi' \bar{\xi} \xi ]_{\gamma }
\left[ 1 + \frac{\bar{\xi}_{\gamma}' \xi_{\gamma}'}{\sqrt{f(\gamma)}} \right] \,
\left[ 1 + \frac{\bar{\xi}_{\gamma} \xi_{\gamma}}{\sqrt{ f(\gamma)}} \right] \nonumber \\
&=& \prod_{\gamma \neq 0} \frac{1}{f(\gamma)} \equiv \frac{1}{f_*} \;,
\label{f_star}
\end{eqnarray}
where we defined $f_*$ in the last step.
In the second case, the two end sites contribute the factor $\xi_{\alpha, b}\, \xi_{\alpha, b}'\, \bar{\xi}_{\alpha, b}' \, \bar{\xi}_{\alpha, b}=\bar{\xi}_{\alpha, b}' \, \xi_{\alpha, b}' \, \bar{\xi}_{\alpha, b} \, \xi_{\alpha, b}$. Now we have to consider the first term in the expansion
of the link exponential for state $\alpha$, while for other variables the calculation is repeated
as in (\ref{f_star})
\begin{eqnarray}
&\!& \!\!\!\!\!\!\!
\prod_{\gamma \neq 0} \int \mathcal{D} [ \bar{\xi}' \xi' \bar{\xi} \xi ]_{\gamma }
\bar{\xi}'_{\alpha} \bar{\xi}_{\alpha}
\left[ 1 + \frac{\bar{\xi}_{\gamma}' \xi_{\gamma}'}{\sqrt{f(\gamma)}} \right] \,
\left[ 1 + \frac{\bar{\xi}_{\gamma} \xi_{\gamma}}{\sqrt{ f(\gamma)}} \right]
\xi_{\alpha} \xi'_{\alpha} \nonumber \\
& = & \prod_{\gamma \neq 0, \alpha} \frac{1}{f(\gamma)} = \frac{f(\alpha)}{f_*}.
\label{non_groundstate}
\end{eqnarray}
We see that, apart from the irrelevant global factor $\prod_b 1/f_*$,
we reproduce the configuration space and weight factors of the original link model.
\subsection{Field-theoretical formulation}
To generate the Feynman diagrammatic expansion, we need to represent the Grassmann weight factor in the exponential form. The link factors (\ref{link}) have the form of Gaussian exponentials
already. Hence, it is only the site factors that need to be rewritten identically as
\begin{equation}
B_j\, =\, \exp \left[ \sum_{ \{ \alpha_b \}_j } \lambda(\{ \alpha_b \}_j) \prod_{b \in \{ b \}_j} \breve{\xi}_{\alpha_b, b} \, \breve{\xi}_{\alpha_b, b}^*
\right] \, .
\label{site2}
\end{equation}
The constants $ \lambda(\{ \alpha_b \}_j)$ are readily related to the site
factors $g(\{ \alpha_b \}_j)$ by simple algebraic equations obtained by expanding
the exponential and equating sums of similar terms to their counterparts in the r.h.s. of Eq.~(\ref{site}).
By expanding the non-Gaussian part of the exponential (\ref{site2}) and applying Wick's theorem,
we arrive at Feynman rules for the diagrammatic series.
The reader should avoid confusion by thinking that an expansion of the exponential (\ref{site2})
takes us back to Eq.~(\ref{site}). Recall that connected Feynman diagrams are formulated
for the free energy density, not the partition function, and summation over
all lattice sites is done for a given set of interaction vertexes in the graph,
as opposite to the summation over all vertex types for a given set of lattice points.
Therefore, the ``coupling constants" in Feynman diagrams are $\lambda$'s, not $g$'s.
\subsection{Absorbing link factors into site factors}
The separation of the weight factors into link and site ones is
merely a convention. Indeed, each link factor can be ascribed
to one of the two site factors at its ends. This leads to a
slightly different Grassmannization protocol.
This trick may prove convenient for generalization to non-local models
considered below.
\section{Generalizations}
\label{sec:III}
\subsection{Plaquette models}
A plaquette model can be viewed as a certain generalization of the local link model.
States (configurations) of a plaquette model are indexed by a set of discrete labels
residing on (oriented) plaquettes of a hyper-cubic lattice.
The plaquette label $\alpha$ takes on either a finite or countably infinite number of values.
The statistical weight of each state factors into a product of plaquette and edge weights (to be referred to as plaquette and edge factors, respectively). A plaquette factor, $f$, is a function of the corresponding plaquette variable, $f\equiv f(\alpha)$. An edge factor, $g$, is a function which depends on the labels of all plaquettes sharing this edge (this set of labels will be denoted as
$\{ \alpha_p \}_j$ for the edge $j$); it encodes, if necessary, constraints on the allowed
sets of $\{ \alpha_p \}_j$.
Without loss of generality (up to a global normalization factor), we identify the
``ground state" as $\alpha_p = 0$ for all plaquettes, and set $f(0)=1$.
The orientation of the plaquette (for some models it is merely a matter of convenience)
is enforced by an ordered enumeration of sites at its boundary. For a plaquette $p$, the vertex label $\nu \equiv \nu_p = 0,\, 1,\, 2,\, 3$ enumerates four vertices in such a way that $\nu \pm 1$ modulo 4 stands for the next/previous vertex with respect to the vertex $\nu$ in the clockwise direction.
For each state $\alpha \neq 0$ of the plaquette $p$, we introduce eight Grassmann variables: $\xi_{\alpha,p, \nu_p}$, $\bar{\xi}_{\alpha,p, \nu_p}, \nu_p = 0,1,2,3$.
As before, for $\alpha=0$ the variables $\xi$ and $\bar{\xi}$ are not Grassmannian, $\xi_{0, p, \nu } =0$, $\bar{\xi}_{0,p,\nu }=1$.
The corresponding plaquette weight in the Grassmann partition function reads
\begin{equation}
A_p = \exp \left\{ \sum_{\alpha \neq 0}\, [-f(\alpha)]^{-1/4} \sum_{\nu_p=0}^{3}\, \bar{\xi}_{\alpha,p, \nu_p}\, \xi_{\alpha,p, \nu_p} \right\}\, .
\label{plaquette}
\end{equation}
Note a close analogy with Eq.~(\ref{link}).
Site weights Eq.~(\ref{site}) are now replaced with edge weights $B_j$.
Using the notation $\{ p \}_j$ for the set of all plaquettes sharing the
edge $j$, and $0_j$ for the state when all plaquettes in this set have
$\alpha_p=0$, we write
\begin{equation}
B_j\, =\, 1 + \sum_{ \{ \alpha_p \}_j \neq 0_j} g(\{ \alpha_p \}_j) \prod_{p \in \{ p \}_j} \xi_{\alpha,p, (\nu_p^{(j)}+1)} \, \bar{\xi}_{\alpha,p, \nu_p^{(j)}}\, ,
\label{edge}
\end{equation}
where $\nu_p^{(j)}$ is the site enumeration index within the plaquette $p$, with respect to which the edge $j$ is outgoing. [Accordingly, the edge $j$ is incoming with respect to site $(\nu_p^{(j)}+1)$.] In what follows, we will associate $\nu_p^{(j)}$ not only with the site, but also with the corresponding edge.
The proof that the classical and Grassmannian partition functions are identical (up to a global factor) is similar to the one for the link model after we notice that a non-zero contribution from
plaquette $p$ is possible only if the same plaquette label $\alpha_p$ is used in all edge weights.
The $\alpha =0$ contribution comes from the term
\begin{equation}
-\prod_{\gamma} {1\over f(\gamma)}\, \prod_{\nu_p=0}^{3}\, \bar{\xi}_{\gamma, p, \nu_p} \, \xi_{\gamma, p, \nu_p} \qquad (\mbox{at}~~\alpha =0)
\label{groundstate_p}
\end{equation}
in the expansion of the exponential (\ref{plaquette}). It contributes a factor $1/q_*$, where
\begin{equation}
q_* = \prod_{\gamma } (-1) f(\gamma) \, .
\label{f_star_p}
\end{equation}
The $\alpha \neq 0$ contribution comes from the plaquette term
\begin{equation}
\prod_{\gamma \neq \alpha} {1\over f(\gamma)}\, \prod_{\nu_p=0}^{3}\, \bar{\xi}_{\gamma, p, \nu_p} \, \xi_{\gamma, p, \nu_p} \qquad (\mbox{at}~~\alpha \neq 0)
\label{non_groundstate_p}
\end{equation}
multiplied by the product $\prod_{\nu_p=0}^{3}\, \bar{\xi}_{\alpha, p, \nu_p} \, \xi_{\alpha, p, \nu_p}$ originating from the boundary edge terms $ \xi_{\alpha,p, (\nu_p^{(j)}+1)} \, \bar{\xi}_{\alpha,p, \nu_p^{(j)}}$. Because of the Grassmann anticommutation rules, this four-edge factor yields an additional minus sign, explaining the use of the negative sign
in front of $f(\alpha)$ in Eq.~(\ref{plaquette}).
Upon Grassmann integration, the contribution to the partition function of the resulting term equals to $f(\alpha)/q_*$.
Feynman diagrammatics for the plaquette model is obtained by following the same basic steps as for
the link models. The Gaussian part is given by Eq.~(\ref{plaquette}) with four pairs of Grassmann
fields for every non-zero plaquette state.
The interaction part of the Grassmann action is contained in edge weights (\ref{edge})
after they are written in an exponential form
\begin{equation}
B_j\, =\, \exp \left[ \sum_{ \{ \alpha_p \}_j} \lambda(\{ \alpha_p \}_j) \prod_{p \in \{ p \}_j} \xi_{\alpha,p, (\nu_p^{(j)}+1)} \, \bar{\xi}_{\alpha,p, \nu_p^{(j)}} \right] \, ,
\label{edge2}
\end{equation}
with the constants $ \lambda(\{ \alpha_b \}_j)$ unambiguously related to the edge factors $g(\{ \alpha_b \}_j)$.
\subsection{Unconstrained discrete models with pair-wise interaction}
The hallmark of the considered link (plaquette) models is the non-trivial interaction introduced via site (edge) factors. It is due to this type of interaction---and, in particular, its extreme form of a constraint on allowed combinations of discrete variables---that we had to introduce multiple
Grassmann variables for each state of the link (plaquette). The situation simplifies dramatically
if we are dealing with unconstrained discrete degrees of freedom with pair interactions between them.
Consider a link model defined by the statistical weight
\begin{equation}
W(\{ \alpha_b \}) = \prod_{b_1 , b_2}
F(\, \alpha_{b_1}, b_1; \, \alpha_{b_2}, b_2 ) \;,
\label{Wlink22}
\end{equation}
based on products of two-link factors. Without loss of generality, these factors can be
cast into the exponential form
\begin{equation}
W(\{ \alpha_b \}) = \prod_{b_1 , b_2}\,
e^{-(1/2) \eta_{\, \alpha_{b_1}, b_1;\, \alpha_{b_2}, b_2}} \;,
\label{Wlink2}
\end{equation}
We assume that all factors in the product are bounded and properties of the $\eta$-matrix are well-conditioned.
Grassmannization of this model can be done by taking advantage of properties of Gaussian integrals
that allow one to express (\ref{Wlink2}) identically (up to normalization) as
\begin{equation}
W(\{ \alpha_b \}) = \int {\cal D} X \prod_{b} \, e^{iX_{\,\alpha_b, b}}
\, W_G( \{ X_{\, \alpha_b, b} \} ) \;.
\label{Wlink3}
\end{equation}
Here $\{ X_{\, \alpha_b, b} \}$ is a collection of auxiliary real continuous variables.
For briefness, we do not show explicitly the Gaussian weight $W_G$ that is
uniquely defined by the values of all pairwise averages performed with this weight
\begin{equation}
\eta_{\, \alpha_{b_1}, b_1; \,\alpha_{b_2}, b_2} \, =\, \langle X_{ \alpha_{b_1}, b_1}
X_{\alpha_{b_2}, b_2 } \rangle \, .
\label{expon1}
\end{equation}
What we achieve for a fixed set of $X$ variables is a link model that contains only single-link factors
\begin{equation}
\forall b: \qquad f_b(\alpha_b) = \, e^{iX_{\alpha_b,b }}.
\label{expon}
\end{equation}
For models with site constraints, link factors can be attributed to site factors
at the incoming (or outgoing) ends with subsequent Grassmannization of the latter
as discussed above. For unconstrained models, Grassmannization is accomplished by replacing sums over link variables with
\begin{equation}
\sum_{\alpha_b} f_b(\alpha_b) \, \to \,
{\cal W}^{(G)}_b \, =\, \exp \left[ \bar{\xi}_b \xi_b \left( \sum_{\alpha_b} \, e^{iX_{\, \alpha_b, b}} \right) \right] \;.
\label{weight_GL}
\end{equation}
Note that here Grassmann variables have nothing to do with the discrete index $\alpha_b$,
in contrast with previous considerations. The resulting formulation contains both
Grassmann and real-number integrations.
Clearly, all considerations can be repeated identically (up to a trivial change in notations)
for a model based on discrete variables $\alpha_s$ residing on lattice sites when
the configuration weight is given by
\begin{equation}
W(\{ \alpha_s \}) = \prod_{s_1 , s_2}\,
e^{-(1/2) \eta_{\, \alpha_{s_1}, s_1;\, \alpha_{s_2}, s_2}} \; .
\label{Q_1_Q_2}
\end{equation}
\subsection{Order of expansion}
\label{subsec:D}
The notion of the order of expansion is absolutely central for practical applications
when diagrammatic series are truncated. Normally, it is defined as an integer non-negative power of a certain dimensionless parameter $\zeta$ playing the role of a generalized coupling constant,
such that the diagrammatic expansion corresponds to a Taylor expansion in $\zeta$ about the point $\zeta=0$. Without loss of generality,
we can always select $\zeta$ (by an appropriate rescaling) in such a way that the physical value of $\zeta$ is 1. This is especially convenient in cases
when there is more than one interaction vertex, and ascribing different powers of
$\zeta$ to them results in (re-)grouping of different terms in the series.
A reasonable guiding principle behind such a (re-)grouping is the requirement to end up with
Taylor series having finite convergence radius around $\zeta=0$.
The latter is guaranteed if the theory is analytic in $\zeta$ at the origin; the necessary condition for this to be true is the absence of Dyson's collapse when changing
the sign (more generally, the phase) of $\zeta$.
As an illustration, consider the theory (\ref{Wlink22})-(\ref{Wlink2}) and its
Grassmann counterpart (\ref{weight_GL}). Introduce the $\zeta$-dependence by the replacement
\begin{equation}
e^{iX_{\, \alpha_b, b } } \, \to \, e^{i\zeta X_{\, \alpha_b, b}} .
\label{zeta}
\end{equation}
In terms of the original theory, the replacement (\ref{zeta})
means $\eta \to \zeta^2 \eta$, for all $\eta$'s in Eq.~(\ref{Wlink2}).
If amplitudes of all $\eta$ values in (\ref{Wlink2}) are bounded,
we expect that such a dependence on $\zeta$ is analytic not only for a finite system,
but also in the thermodynamic limit at finite temperature.
In the Grassmann action (\ref{weight_GL}), the expansion of the exponential $e^{i\zeta X_{s, \alpha_b}}$ in powers of $\zeta$ generates an infinite series of interaction vertexes (the zeroth-order term defines the harmonic action):
\begin{equation}
\bar{\xi}_b \xi_b \sum_{\alpha_b } \left( i \zeta X_{\, \alpha_b, b} - {1\over 2} \zeta^2
X_{\, \alpha_b, b}^2 - {i\over 3!} \zeta^3 X_{\, \alpha_b, b}^3 + \ldots \right).
\label{coupling}
\end{equation}
Higher-order vertexes in $X$ come with a higher power of $\zeta$ and this sets unambiguously the rules for defining the diagram order.
\section{Illustration for the 2D Ising model}
\label{sec:IV}
\begin{figure}[tbp]
\centering
\includegraphics[width=1.0\columnwidth]{prediagrams.pdf}
\caption{(Color online) Four classes of generic pre-diagrams for link models on a square lattice. The elements in the first and the third row can only occur at the end points of the spin correlator (indicated by the open circle), the elements in the second and fourth row are the generic basic vertexes of the theory ascribed to the sites of the underlying lattice. There are hence 4 $V_1$ vertexes with 1 leg (first row, $U, R, D,$ and $L$), 6 $V_2$ vertexes with 2 legs (second row, $RU, RD, LD, LU, UD$ and $LR$), 4 $V_3$ vertexes with 3 legs (third row, $LUR, URD, LDR,$ and $DLU$), and 1 $V_4$ vertex with 4 legs (fourth row, $RULD$). Connected to the legs of these vertexes are pairs of bi-Grassmann fields (thick dash lines (blue and red)) that reside on the links of the underlying 2D lattice. Thin dashed lines (showing lattice links adjacent to the site of the vertex) are to guide the eye and have no other meaning than showing the underlying 2D lattice. The generalization to other dimensions is straightforward.} \label{fig:prediagrams}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[trim = 0 0mm 0mm 0mm, clip, width=1.0\columnwidth]{rho_generic_a.pdf}
\caption{(Color online) The first- and third-order diagrams for $\rho_{(1,0)}$ (at $h=0$) based on expanding (\ref{sigma_sigma_correlator_1}). The contribution of these diagrams is $\zeta+ 2\zeta^3$.}
\label{fig:rho_generic_a}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[trim = 0 0mm 0mm 0, clip, width=1.0\columnwidth]{rho_generic_b1.pdf}
\caption{(Color online) Fifth-order diagrams for $\rho_{(1,0)}$ (at $h=0$) based on expanding (\ref{sigma_sigma_correlator_1}): These four diagrams involve a three-leg end vertex. Each diagram contributes $(-2)\zeta^5$.}
\label{fig:rho_generic_b1}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[trim = 0 0mm 0mm 0, clip, width=1.0\columnwidth]{rho_generic_b2.pdf}
\caption{(Color online) Fifth-order diagrams for $\rho_{(1,0)}$ (at $h=0$) based on expanding (\ref{sigma_sigma_correlator_1}): These four counterparts of the diagrams shown in
Fig.~\ref{fig:rho_generic_b1} are obtained by replacing a three-leg end vertex with a one-leg end vertex.
Each diagram contributes $\zeta^5$.}
\label{fig:rho_generic_b2}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[trim = 0 0mm 0mm 0, clip, width=1.0\columnwidth]{rho_generic_b.pdf}
\caption{(Color online) The four remaining counterparts (cf. Fig.~\ref{fig:rho_generic_b2}) to Fig.~\ref{fig:rho_generic_b1}. }
\label{fig:rho_generic_b}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[trim = 0 0mm 0mm 0mm, clip, width=1.0\columnwidth]{rho_generic_c.pdf}
\caption{(Color online) Additional fifth-order diagrams for $\rho_{(1,0)}$ (at $h=0$) involving two one-leg end vertexes. Each diagram contributes $\zeta^5$.}
\label{fig:rho_generic_c}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[trim = 0 0mm 0mm 0mm, clip, width=0.6\columnwidth]{rho_generic_d.pdf}
\caption{(Color online) Fifth-order diagrams for $\rho_{(1,0)}$ (at $h=0$) containing a link with multiple Grassmann pairs. The net sum of the shown diagrams is $- \zeta^5$, because there are three ways of associating the primed and non-primed propagators along the bottom link,
two of them contribute with the negative sign (upper right and lower left panel) and the third one is contributing with the positive sign (lower right panel). The remaining possibility
(shown in the upper left panel) is not allowed since it produces a disconnected diagram.}
\label{fig:rho_generic_d}
\end{figure}
\subsection{Model and observables}
Consider the 2D Ising model on the square lattice with the Hamiltonian
\begin{equation}
-H/T = \beta \sum_{\langle i,j \rangle} \sigma_i \sigma_j + \sum_i h_i \sigma_i.
\end{equation}
The Ising variables $\sigma = \pm 1$ live on the sites of the 2D square lattice and interact ferromagnetically with their nearest neighbors, as is represented by the first term in the Hamiltonian. We write the dimensionless coupling as $\beta$ in units of the temperature $T$.
Additionally, every spin feels a dimensionless magnetic field $h_i=h$, which can be taken
$h \ge 0$ without loss of generality.
The partition function of the Ising model reads
\begin{equation}
Z = \sum_{ \{ \sigma_i \} }
\prod_{\langle i,j \rangle} e^{\beta \sigma_i \sigma_j} \; \prod_{i} e^{h_i \sigma_i}.
\end{equation}
The most typical observable of the Ising model is the spin-spin correlation function $\rho_{ij}$,
\begin{equation}
\rho_{\ij} = \langle \sigma_i \sigma_j \rangle = \frac{1}{Z} \left.
\frac{ \partial^2 Z}{\partial h_i \partial h_j} \right|_{h_i = h_j= h} \,.
\label{sigma_sigma_correlator_1}
\end{equation}
\subsection{Grassmannization of the high-temperature expansion}
Using the well-known identities
\begin{eqnarray}
e^{\beta \sigma_i \sigma_j} & = & \cosh \beta \left( 1 + \sigma_i \sigma_j \tanh \beta \right) \nonumber \\
e^{h \sigma_i} & = & \cosh h \left( 1 + \sigma_i \tanh h \right),
\end{eqnarray}
the partition function can be written as $Z = Z_0 Z'$ with
$Z_0 = ( \cosh \beta)^{2N} ( \cosh h)^N$ for a lattice of $N$ sites and $2N$ links. With the notation
$\zeta = \tanh \beta$ and $\eta = \tanh h$ the remaining factor is given by
\begin{equation}
Z' = \sum_{ \{ \sigma_i \} } \, \prod_{\langle i,j \rangle} ( 1 + \sigma_i \sigma_j \zeta )
\, \prod_{i} ( 1 + \sigma_i \eta).
\end{equation}
Upon summation over spin variables
we are left with a link model, where link variables take only two values, 0 or 1, to
specify whether we are dealing with the first or the second term in the sum
$(1 + \sigma_i \sigma_j \zeta)$. In the partition function, terms with an odd power of $\sigma_i$ on any of the sites yield zero upon spin summation. The remaining terms depend on link
variables in a unique way.
The formalism of the previous section can be straightforwardly applied, and we obtain
\begin{eqnarray}
& f(0) = 1\,, \;\; f(1) = f_* = \zeta \,, \\
& g(0) = g(2) = g(4)=1 \;, \;\; g(1) = g(3)=\eta \,. \label{eq:factors_f_g}
\end{eqnarray}
Here we label site factors using the total sum of incident link variables,
$\sum_{b \in \{ b \}_j } \alpha_b$, to avoid unnecessary rank-4 tensor notations.
If we further redefine $Z_0 \to Z_0 2^N f_*^{2N}$, then the Grassmann representation
of the partition function $Z'$ is given by
\begin{eqnarray}
Z' &= & \int \mathcal{D}[ \bar{\xi}' \xi' \bar{\xi} \xi ]_{ \{ \alpha_b \} } \prod_{ \{ \alpha_b \} } \exp \left( \frac{1}{\sqrt{\zeta}} \bar{\xi}'_{\alpha_b} \xi_{\alpha_b} + \frac{1}{\sqrt{\zeta}} \bar{\xi}_{\alpha_b} \xi_{\alpha_b} \right) \nonumber \\
{} & {} & \times \exp \left( \sum_j \lambda_{ \alpha_{ \{b \} } } \prod_{b_j} \breve{\xi}_{\alpha_b} \breve{\xi}^*_{\alpha_b} \right) \,.
\end{eqnarray}
\subsection{Vertex coefficients}
We now compute the factors $\lambda$. To this end, we first introduce notations
(for a fixed site $j$ and suppressing the site index for clarity)
\begin{eqnarray}
V_1\!\! &=&\!\! \breve{\xi}_R \breve{\xi}_R^* + \breve{\xi}_U \breve{\xi}_U^* + \breve{\xi}_L \breve{\xi}_L^* + \breve{\xi}_D \breve{\xi}_D^* = n_R + n_U + n_L + n_D, \nonumber \\
V_2\!\! &=&\!\! n_R n_U + n_R n_L + n_R n_D + n_U n_L + n_U n_D + n_L n_D, \nonumber \\
V_3\!\! &=&\!\! n_R n_U n_L + n_R n_U n_D + n_R n_L n_D + n_U n_L n_D, \nonumber \\
V_4\! \!&=&\! \!n_R n_U n_L n_D ,
\end{eqnarray}
and then Taylor expand
\begin{equation}
\exp \left[ \lambda_1 V_1 + \lambda_2 V_2 + \lambda_3 V_3 + \lambda_4 V_4 \right].
\end{equation}
The only non-zero terms generated by this expansion are $V_1^2 = 2V_2, V_1^3 = 6V_3, V_1^4 = 24 V_4, V_1 V_2 = 3V_3, V_1 V_3 = 4 V_4$ and $V_2^2 = 6 V_4$. All other powers and multiplications
of operators yield zero. Note that operators from different sites commute and may be
excluded from consideration here. The final result is
\begin{eqnarray}
\lefteqn{\exp \left[ \lambda_1 V_1 + \lambda_2 V_2 + \lambda_3 V_3 + \lambda_4 V_4 \right] = } \nonumber \\
& & 1 + \lambda_1 V_1 + \lambda_2 V_2 + \lambda_3 V_3 + \lambda_4 V_4 \nonumber \\
& & + \frac{1}{2} \left( \lambda_1^2 2V_2 + \lambda_2^2 6V_4 + 2 \lambda_1 \lambda_2 3 V_3 + 2 \lambda_1 \lambda_3 4V_4 \right) \nonumber \\
& & + \frac{1}{6} \left( \lambda_1^3 6V_3 + 3 \lambda_1^2 \lambda_2 12 V_4 \right) + \frac{1}{24} \lambda_1^4 24 V_4 .
\end{eqnarray}
Term-by-term matching with Eq.~(\ref{eq:factors_f_g}) then leads to
\begin{eqnarray}
g_1 = \eta & = & \lambda_1\, , \\
g_2 = 1 & = & \lambda_2 + \lambda_1^2\, ,\\
g_3 = \eta & = & \lambda_3 + 3 \lambda_1 \lambda_2 + \lambda_1^3\, ,\\
g_4 = 1 & = & \lambda_4 + 3 \lambda_2^2 + 4 \lambda_1 \lambda_3 + 6 \lambda_2 \lambda_1^4 + \lambda_1^4 \, .
\end{eqnarray}
The solution is immediate
\begin{eqnarray}
\lambda_1 & = & \eta \, , \\
\lambda_2 & = & 1 - \eta^2\, , \\
\lambda_3 & = & -2 \eta + 2 \eta^3\, , \\
\lambda_4 & = & -2 + 8 \eta^2 - 6 \eta^4 \, .
\end{eqnarray}
In what follows we will discuss the $\eta=0$ case (zero external field)
when the only vertexes with non-zero coupling in the partition function
are $V_2$ and $V_4$,
\begin{equation}
\prod_j \exp(V_2^{(j)} + V_4^{(j)}) = \exp \left\{ \sum_j (V_2^{(j)} - 2 V_4^{(j)}) \right\}.
\end{equation}
The expansion of $Z'$ in powers of $\zeta$ then goes as
\begin{eqnarray}
Z' & = & \prod_b \int \mathcal{D} [...]_b \exp \left( \frac{1}{\sqrt{\zeta}} \bar{\xi}'_b \xi'_b + \frac{1}{\sqrt{\zeta}} \bar{\xi}_b \xi_b \right) \nonumber \\
{} & {} & \times \exp \left\{ \sum_j (V_2^{(j)} - 2 V_4^{(j)}) \right\}
\nonumber \\
& = & \left[ 1 + 4 \zeta^4 + 12 \zeta^6 + \ldots \right]^N.
\label{zetprim}
\end{eqnarray}
and the spin-spin correlation function is given by
\begin{eqnarray}
\rho_{ij} & = & \frac{1}{Z'}\prod_b \int \mathcal{D} [...]_b \exp \left( \frac{1}{\sqrt{\zeta}} \bar{\xi}'_b \xi'_b + \frac{1}{\sqrt{\zeta}} \bar{\xi}_b \xi_b \right) \nonumber \\
{} & {} & \times (V_1^{(i)} - 2 V_3^{(i)})(V_1^{(j)} - 2 V_3^{(j)}) .
\label{rhoprim}
\end{eqnarray}
\subsection{Feynman rules}
In order to arrive at the Feynman perturbative expansion we need to write the partition function in the form
\begin{equation}
Z' = Z_{\cal l} \left( \sum_{n=0}^{\infty} \sum_{x_1, \ldots, x_n} \frac{(+1)^n}{n!} \langle V(x_1) \ldots V(x_n) \rangle_0 \right) ,
\end{equation}
where $Z_{\cal l}$ is the partition function of the Gaussian part
(it is the product of local link contributions),
$Z_{\cal l} = \prod_b \int \mathcal{D} [...] \exp( \bar{\xi}'_b \xi'_b + \frac{1}{\zeta} \bar{\xi}_b \xi_b ) = (1 + \zeta)^{(2N)}$.
Feynman rules for the correlation function of the 2D Ising model now follow from the textbook
considerations:
\begin{enumerate}
\item The bare propagators $G^{(0)}=\sqrt{\zeta}$ for primed and non-primes variables are local and reside on the links of the original lattice.
In the correlation function they always occur in pairs of conjugate Grassmann variables
and each pair contributes a factor $\zeta$.
The propagation lines do not have arrows. The bare interaction vertexes (or pre-diagrams, see Fig.~\ref{fig:prediagrams}) are also local and live on the sites of the lattice. There are different types belonging to the $V_2$ and $V_4$ classes with weight $1$ and $-2$, respectively [see Eq.~(\ref{zetprim})]. On the first (and last) site of the correlator we have a vertex belonging to the class $V_1$ or $V_3$ (see Figs.~\ref{fig:rho_generic_a}-\ref{fig:rho_generic_c}) with weight $1$ and $-2$, respectively [see (\ref{rhoprim})].
\item Draw in order $n$ all topologically distinct connected diagrams with $n$ pairs of bi-grassmann variables living on the links of the lattice. The number of interaction vertexes, excluding the end points, is at most $n-1$.
\item For links with multiple occupancy, a minus sign occurs when swapping 2 Grassmann variables. The minus sign can also be found by counting all closed fermionic loops.
\item The total weight of the diagram in order $n$ is hence $(-1)^P (-2)^q \zeta^n$ with $P$ the signature of the exchange permutation and $q$ the sum of all type-3 and type-4 vertexes.
\end{enumerate}
Disconnected diagrams are defined with respect to both the primed and non-primed Grassmann variables simultaneously. Thus, a link can lead to a disconnected diagram only if the primed and non-primed variables simultaneously lead to disconnected pieces (such as the upper left panel in Fig.~\ref{fig:rho_generic_d}). We check the connectivity of a diagram by the breadth-first algorithm.\\
\subsection{Example: the first element of the spin correlation function}
Let us focus on the first element of the correlation function connecting the sites $(0,0)$ and $(1,0)$ (using translational invariance, any 2 neighboring sites $\left< \bm{r}_1, \bm{r}_2 \right>$ can be taken).
To first order, we put a $V_1$ vertex on the origin and target site. There is one way to combine them, thus the total contribution is $\zeta$.
By the symmetry of the lattice, even expansion orders do not contribute. In third order, we can construct a diagram by putting a $V_2$ (RD) vertex on the site $(0,1)$ and a $V_2$ vertex $(LD)$ on the site $(1,1)$. The mirror symmetry of this diagram about the x-axis is also a valid diagram. Hence, the contribution is $2 \zeta^3$. These diagrams contributing in first and third order are shown in Fig.~\ref{fig:rho_generic_a}.
In fifth order, there are 4 diagrams with a $V_3$ vertex on one of the endpoints, yielding a contribution $-8\zeta^5$ . There are 14 diagrams consisting of only $V_1$ and $V_2$ vertexes and single pair-lines, yielding a contribution $14 \zeta^5$. The contributions to fifth order are shown in
Figs.~\ref{fig:rho_generic_b1},~\ref{fig:rho_generic_b2},~\ref{fig:rho_generic_c}, and~\ref{fig:rho_generic_d}.
There are however additional diagrams with 2 pairs of Grassmann variables living on the same link, as is shown in Fig.~\ref{fig:rho_generic_d} (there are equivalent diagrams obtained by mirror symmetry around the x-axis which are not shown).
They all have on the origin a $V_1$ and a $V_2$ ($RU$) vertex, and on the target site $(1,0)$ a $V_1$ and a $V_2$ ($UL$) vertex. On the site $(1,1)$ there is a $V_2$ ($LD$) and on site $(0,1)$ a $V_2$ ($RD$) vertex. Let us look more carefully at the link between the origin and target site:
\begin{equation}
\frac{1}{\zeta^2} \int \mathcal{D} \left[ \bar{\xi}' \bar{\xi} \xi \xi' \right] \bar{\xi}' \bar{\xi} \xi \xi' \bar{\xi}' \bar{\xi} \xi \xi' .
\label{example_term}
\end{equation}
The origin is associated with $\bar{\xi}' \bar{\xi}$ and the target with $ \xi \xi'$ by our convention.
Applying Wick's theorem, there are 4 possible ways to pair the Grassmann variables:
\begin{enumerate}
\item The pairing combination
$
\overbracket{\bar{\xi}'\underbracket{\bar{\xi} \xi} \xi'} \overbracket{\bar{\xi}'
\underbracket{\bar{\xi} \xi} \xi'}
$
comes with the sign +1 and leads to a connected diagram (this is the lower right panel in Fig.~\ref{fig:rho_generic_d}).
\item The pairing combination
$
\mathrlap{\overbracket{\phantom{\bar{\xi}' \bar{\xi} \xi \xi'}}} \bar{\xi}'\underbracket{\bar{\xi} \underbracket{\xi \xi' \mathrlap{\overbracket{\phantom{ \bar{\xi}' \bar{\xi} \xi \xi'}}}
\bar{\xi}' \bar{\xi}} \xi} \xi'
$
comes with the sign -1 and leads to a connected diagram (this is the upper right panel in Fig.~\ref{fig:rho_generic_d})
\item The pairing combination
$
\overbracket{\bar{\xi}'\underbracket{\bar{\xi} \xi} \underbracket {\xi'\bar{\xi}'}
\underbracket{\bar{\xi} \xi} \xi'}
$
comes with the sign -1 and leads to a connected diagram (this is the lower left panel in Fig.~\ref{fig:rho_generic_d})
\item The pairing combination
$
\overbracket{\bar{\xi}'\overbracket{\bar{\xi}\underbracket{\xi \underbracket{\xi'\bar{\xi}'}
\bar{\xi}} \xi} \xi'}
$
leads to a disconnected diagram and does not contribute to the correlation function (this is the upper left panel in Fig.~\ref{fig:rho_generic_d}).
\end{enumerate}
The net contribution of these 4 distinct diagrams is hence $-1$ (also the diagrams obtained by mirror symmetry around the x-axis yield $-1$, so the total contribution to fifth order is $ (-8 + 14 - 2) \zeta^5 = 4 \zeta^5$.
It is instructive to notice that the sum of all diagrams in which multiple Grassmann pairs live on the same link always produces zero in case all diagrams are connected, in line with the nilpotency of Grassmann variables. Wick's theorem splits however these contributions in connected and disconnected diagrams, where the disconnected diagrams cancel against the denominator of the Feynman expansion.
It is this non-trivial regrouping imposed by Wick's theorem that can yield non-zero
contributions from terms like (\ref{example_term}); and, in particular, from arbitrarily high powers of one and the same interaction vertex.
\section{Implementation}
\label{sec:V}
We explored two ways of evaluating the (bare) series for the spin-correlator: a stochastic Monte Carlo approach and a deterministic full evaluation of all diagrams.
\subsection{Monte Carlo sampling }
In order to perform a Monte Carlo sampling over all Feynman diagrams, we introduce a head and a tail that represent the endpoints of the correlation function. By moving them around the lattice and changing the diagrammatic elements in between the head and tail, we are able to reach an ergodic sampling. The algorithm can be formulated as follows:
The tail remains stationary at the origin whereas the head can move around the lattice. When the head and tail are on the same site and the expansion order is 0, the value of the correlation function is 1 which can be used for normalization of the Monte Carlo process. A Monte Carlo measurement contributes $+1$ or $-1$ depending on the sign of the diagram weight. The simplest Monte Carlo procedure samples according to the absolute weights of the diagrams and consists of the following pairs of reciprocal updates:
\begin{enumerate}
\item MOVE--RETRACT. We choose one of the 4 directions randomly, and attempt to place the head on the site adjacent to the current head site according to this direction. In case this direction does not correspond to backtracking, the current $V_1$ type of the tail turns into a $V_2$, otherwise the head goes back and changes the previous $V_2$ into a $V_1$ type (unless the diagram order is 0 or 1, when only $V_1$ types are possible).
When moving forward, the way of pairing primed and non-primed variables is always unique which in turns implies that we can only retract when the head is connected via a ``straight pair connection" to the previous vertex (both primed and non-primed Grassmann variables of the head are connected to the same vertex on the previous site). We only allow the MOVE--RETRACT updates if the end vertex types are $V_1$.
\item SWAP VERTEX. Swaps between the vertexes $V_1 + V_2 \leftrightarrow V_3$ (for head and/or tail) and $V_2 + V_2 \leftrightarrow V_4$ (anywhere in the diagram). This update is its own reciprocal.
\item RELINK. On a given link, relink primed and non-primed Grassmann variables. This can change the sign of the weight only. This update is its own reciprocal.
\end{enumerate}
The second and third type of updates may lead to disconnected diagrams. In such cases, the configuration is unphysical. We opt to allow such configurations, but a Monte Carlo measurement is forbidden and type-1 updates remain impossible until the diagram is connected again.
For small values of $\zeta$ the sign problem is nearly absent, but only low expansion orders can be reached. For higher values of $\zeta$ (close to and above the critical one) an increasing number of orders contributes significantly, consequently more time is spent in higher orders and the sign problem significantly worsens.
\subsection{Deterministic full evaluation}
For the case of the 2D Ising model, a Monte Carlo approach offers no advantages over a full series expansion approach. With this we mean the explicit listing and evaluation of all possible diagrams as opposed to the stochastic sampling over all topologies. This is because all diagrams in a given expansion order contribute a number of order unity (times the same power of $\zeta$), often with alternating sign, leading to huge cancellations. Only the exact cancellation has physical information, and this requires that every diagram is evaluated multiple times before the correct convergence can be seen. A Monte Carlo approach makes much more sense if the dominant contributions to the total weight are coming from a narrow parameter region, which is usually the case if there are additional integrals over internal momenta.
We therefore wrote a code that evaluates all diagrams for the correlation function up to a maximum order. The construction is based on the fact that there is an easy way to construct all the ``easy" diagrams (the ones that formally look like originating from a high-temperature series expansion). These can serve as parent diagrams, from which further offspring diagrams can be constructed which have one or multiple $V_3$ and $V_4$ vertexes as well as possible fermionic exchanges. All diagrams in order $n$ can be found as follows:
\begin{enumerate}
\item Write down all possible words of the form $X_1 X_2 \ldots X_n$ with the alphabet $X_j \in \{0,3 \}$ corresponding to the 4 directions on the square lattice. Make sure that subsequent directions are not backtracking. For example, if $X_4$ is in the positive $+\hat{x}$ direction, then $X_5$ cannot be in the negative $-\hat{x}$ direction. From this word we also know all sites and links that are visited, as well as all type-1 and type-2 vertexes that are used to make this diagram.
\item Such a parent diagram is added to a list of different topologies only if it has a unique topology. To store the topological information of a bare vertex, we need to store a pair consisting of a site index and a vertex type. The diagram is then stored as an ordered map where the ``key" values are given first by the lattice site index and second by the vertex type (in binary format). The ordered map may have multiple entries with the same key if multiple vertexes reside on the same site and if they are of the same type ({\it e.g.}, two $RL$ vertexes on the same site).
\item We iterate over this configuration list and check if the tail and head sites can be merged into a type-3 vertex by combining them with type-2 vertexes that reside on the same lattice site. If so, and if the resulting topology is unique, the diagram is added to the list. This step is performed in three parts: first for the head and tail together (in order to find all diagrams with 2 $V_3$ ends), then for the head alone, and finally for the tail alone.
\item We iterate again over the full configuration list and check if 2 type-2 vertexes that live on the same site can be merged into a type-4 vertex. This last step has to be repeated until no further merges are possible (since it may happen that a diagram has multiple type-4 vertexes or even multiple type-4 vertexes on the same site). Diagrams thus created are also added the configuration list if their topology is unique. After completion of this step, all possible topologies have been generated.
\item We compute the product of all the vertex weights, according to the Feynman rules.
\item From this list of parent diagrams we need to generate all offspring diagrams which feature all possible fermionic permutations for multiply occupied links. This first requires that we know how the vertexes are connected in the parent diagram, which is stored in the configuration list. The parent diagram always has permutation sign +1 (because the connections of the primed and non-primed Grassmann variables are always the same). Next we generate all possible permutations by relinking the primed and/or the non-primed Grassmann variables using Heap's algorithm. If a link has occupation number $m$, then there are $(m!)^2$ combinations to be generated (and there may be more than one multiply occupied link). The permutation signature is also stored.
\item We check the connectivity of the diagram using the breadth-first algorithm. Disconnected diagrams contribute 0.
\item Finally, we compute the isomorphism factor: if $m$ identical vertexes on the same site are found, a factor $1/m!$ must be taken into account. This is a consequence of how we construct the diagrams: topology checks were only performed on the parent diagrams (and based on vertexes only), not on offsprings obtained by fermionic exchange. (It would be prohibitively expensive to add the offspring diagrams to the list of all possible diagrams.) Hence, just as we generate illegal disconnected diagrams, we also have a double counting problem when identical vertexes occur in the list.
\end{enumerate}
In order 14, there were about 140,000 parent diagrams contributing to the first entry on the diagonal of the correlator. The hugest number of permutations was $(4!)^4(3!)^4 \approx 10^8$. Since the sum of these permutations has a net contribution of order 1, Monte Carlo has roughly a sign problem of the order of $10^{-8}$ for these diagrams. The first time a nontrivial isomorphism factor is seen is in order 6 for the first element on the diagonal of the spin correlator: There are diagrams in which two links are doubly occupied, and those links are connected by an identical $V_2$ vertex, hence the isomorphism factor $1/2$. More efficient ways of evaluating and storing the diagrams can probably be devised and implemented, but the above scheme is sufficient to check the validity of the technique and study the transition. \\
\section{Results}
\label{sec:VI}
\subsection{Spin-spin correlation function}
Our results for the spin-spin correlation function are shown in Table.~\ref{table:rho}. The correlation function is known recursively from Refs.~\onlinecite{Perk1980, Guttmann_rho1, Guttmann_rho2}. It is also known as a Painlev{\'e}-VI nonlinear differential equation\cite{McCoy76} but this is not so well suited to obtain the series coefficients. Along the principal axes and the diagonal it can also be expressed as a Toeplitz determinant.
The first element along and the axis and the diagonal can be recast in terms of complete elliptic integrals (see pp. 200-201 in Ref.~\onlinecite{mccoy1973}), which are convenient for series expansions,
\begin{eqnarray}
\rho_{(1,0)} & = & \coth(2 \beta) \left[ \frac{1}{2} + \frac{\cosh^2 2\beta }{\pi} (2 \tanh^2 2\beta - 1) K(k_>) \right] \nonumber \\
{} & \to & \zeta + 2 \zeta^3 + 4 \zeta^5 + 12 \zeta^7 + 42 \zeta^9 + \ldots \\
\rho_{(1,1)} & = & \frac{2}{\pi k_>} \left[ K(k_>) + (k_>^2 - 1) K(k_>) \right] \\
{} & \to & 2\zeta^2 + 4\zeta^4 + 10 \zeta^6 + 32 \zeta^8 + 118 \zeta^{10} + \ldots \nonumber
\end{eqnarray}
with $k_> = \sinh^2(2\beta)$, $K(.)$ and $E(.)$ the complete elliptic K and E functions, respectively.
The above-cited recursion relations could be initialized with these expansions and shown to yield the same results as the top 2 rows in Table.~\ref{table:rho}.
\begin{table}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
site/order & $\zeta$ & $\zeta^2$ & $\zeta^3$ & $\zeta^4$ & $\zeta^5$ & $\zeta^6$ & $\zeta^7$ & $\zeta^8$ & $\zeta^9$ & $\zeta^{10}$ & $\zeta^{11}$ \\ \hline
(1,0) & 1 & 0 & 2 & 0 & 4 & 0 & 12 & 0 & 42 & 0 & 164 \\ \hline
(1,1) & 0 & 2 & 0 & 4 & 0 & 10 & 0 & 32 & 0 & 118 & 0 \\ \hline
(2,0) & 0 & 1 & 0 & 6 & 0 & 16 & 0 & 46 & 0 & 158 & 0 \\ \hline
(2,1) & 0 & 0 & 3 & 0 & 11& 0 & 31 & 0 & 97 & 0 & 351 \\ \hline
(2,2) & 0 & 0 & 0 & 0 & 6 & 0 & 24 & 76 & 0 & 248 & 0 \\ \hline
(3,0) & 0 & 0 & 1 & 0 & 12 & 0 & 48 & 0 &152 & 0 &506\\ \hline
(3,1) & 0 & 0 & 0 & 4 & 0 & 26 & 0 & 92 & 0 &298 & 0\\ \hline
(3,2) & 0 & 0 & 0 & 0 & 10 & 0 & 55 & 0 & 201 & 0 & 684\\ \hline
(3,3) & 0 & 0 & 0 & 0 & 0 & 20 & 0 & 120 & 0 & 480 & 0 \\ \hline
(4,0) & 0 & 0 & 0 & 1 & 0 & 20 & 0 & 118 & 0 &7 452 & 0\\ \hline
(4,1) & 0 & 0 & 0 & 0 & 5 & 0 & 52 & 0 & 244 & 0 & 885\\ \hline
(4,2) & 0 & 0 & 0 & 0 & 0 & 15 & 0 & 118 &0 & 521 & 0 \\ \hline
(4,3) & 0 & 0 & 0 & 0 & 0 & 0 & 25 & 0 &259 & 0 & 1176 \\ \hline
(4,4) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 70 & 0 & 560 & 0 \\ \hline
(5,0) & 0 & 0 & 0 & 0 & 1 & 0 & 30 & 0 & 250 & 0 & 1200 \\ \hline
(5,1) & 0 & 0 & 0 & 0 & 0 & 6 & 0 & 92 & 0 & 574 & 0 \\ \hline
(5,2) & 0 & 0 & 0 & 0 & 0 & 0 & 21 & 0 & 231 & 0 & 1266 \\ \hline
(5,3) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 56 & 0 & 532 & 0 \\ \hline
(5,4) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 126 & 0 & 1176 \\ \hline
(5,5) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 252 & 0 \\ \hline
(6,0) & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 42 & 0 & 474 & 0 \\ \hline
(6,1) & 0 & 0 & 0 & 0 & 0 & 0 & 7 & 0 & 149 & 0 & 1215\\ \hline
(6,2) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 28 & 0 & 416 & 0 \\ \hline
(6,3) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 84 & 0 & 1026 \\ \hline
(6,4) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 210 & 0 \\ \hline
(6,5) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 462 \\ \hline
(7,0) & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 56 & 0 & 826\\ \hline
(7,1) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 8 & 0 & 226 & 0 \\ \hline
(7,2) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 36 & 0 & 699 \\ \hline
(7,3) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 120 & 0 \\ \hline
(7,4) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 330 \\ \hline
(8,0) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 72 & 0 \\ \hline
(8,1) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 9 & 0 & 326 \\ \hline
(8,2) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 45 & 0 \\ \hline
(8,3) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 165 \\ \hline
(9,0) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 90 \\ \hline
(9,1) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 10 & 0 \\ \hline
(9,2) & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 55 \\ \hline
(10,0)& 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 \\ \hline
(10,1)& 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 11 \\ \hline
(11,0)& 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 \\ \hline
\hline
\end{tabular}
\caption{Expansion coefficients for the correlation function up to order 11.}
\label{table:rho}
\end{table}
\subsection{Magnetic susceptibility}
\begin{figure}[tbp]
\centering
\includegraphics[width=1.0\columnwidth]{chi.pdf}
\caption{(Color online) The magnetic susceptibility versus $\zeta$ for different expansion orders from 12 to 1 (top to bottom), compared
to the order 100 result---the converged answer over this plotting range---obtained from Ref.~\onlinecite{Guttmann},
which shows a divergence in good agreement with the critical exponent $\gamma = 7/4$ starting from $\beta \ge 0.38$}
\label{fig:chi}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[width=1.0\columnwidth]{ratio_chi.pdf}
\caption{(Color online) Ratio of consecutive coefficients $\chi[n-1]$ and $\chi[n]$ in the expansion of the susceptibility as a function of the inverse of the expansion order $1/n$. Linear regression according to Eq.~(\ref{eq:lin_reg}) allows to determine the critical temperature with an accuracy of $0.5\%$ and the critical exponent $\gamma$ with an accuracy of $5\%$. The fitting regime included orders 9 through 14.
}
\label{fig:chi_ratio}
\end{figure}
The spin susceptibility is related to the zero momentum value of the Green function by $\beta^{-1} \chi = 1 +\rho (\bm{p}=0)$.
We can hence sum over the entire lattice to obtain
\begin{eqnarray}
\beta^{-1} \chi & = & 1 + 4 \zeta + 12 \zeta^2 + 36\zeta^3 + 100 \zeta^4 + 276\zeta^5 \nonumber \\
{} & {} & +740 \zeta^6 +1972 \zeta^7 + 5172 \zeta^8 + 13492 \zeta^9 \nonumber \\
{} & {} & + 34876 \zeta^{10} + 89764 \zeta^{11} + 229628 \zeta^{12} \nonumber \\
{} & {} & + 585508 \zeta^{13} + 1486308 \zeta^{14} +\ldots
\end{eqnarray}
To this order the series expansion agrees with the ones from Ref.~\onlinecite{Sykes1972} and Ref.~\onlinecite{Guttmann}. For a library of high-temperature series expansions,
see Ref.~\onlinecite{Butera2002}. Currently, the series is known (at least) up to order 2000 and still topic of active research.\cite{Guttmann, Guttmann_rho2}
The series is convergent for any finite expansion order, {\it i.e.}, in the thermodynamic limit the infinite series will diverge first at the phase transition point.
It is hence possible to study the critical behavior of the susceptibility, which is governed by the critical exponent $\gamma = 7/4$.
We plot in Fig.~\ref{fig:chi} the susceptibility versus $\beta$ for different expansion orders, and also plot the asymptotic behavior for comparison.
The critical temperature and the exponent $\gamma$ can be found from a study of the convergence radius of the series. Since
\begin{eqnarray}
\beta^{-1} \chi & = & \sum_n \chi_n \zeta^n \propto (1 - \zeta/ \zeta_c)^{-\gamma} \\
{} & = & 1 + \sum_{n=1}^{\infty} \frac{ \gamma (\gamma - 1) \cdots (\gamma + n - 1)}{ n!} \left( \frac{\zeta}{\zeta_c} \right)^n \nonumber
\label{eq:lin_reg}
\end{eqnarray}
the ratio of coefficients asympotically behaves as
\begin{equation}
\frac{\chi_n}{\chi_{n-1}} = \frac{1}{\zeta_c} + \frac{\gamma - 1}{\zeta_c}\frac{1}{n}.
\end{equation}
In Fig.~\ref{fig:chi_ratio} we extract the critical point $\zeta_c$ from the intercept and the critical exponent $\gamma$ from a linear fit through the ratio of the coefficients. The critical point could be determined with an accuracy of $0.5\%$, whereas the error on $\gamma$ is of the order of $5\%$. However, according to more advanced extrapolation techniques discussed in
Ref.~\onlinecite{ZinnJustin1979}, $\gamma$ can be determined independently from $\zeta_c$ as $\gamma \approx 1.751949$ on the square lattice when the series is known up to 14th order, {\it i.e.}, an accuracy of $0.5\%$.
\section{The $G^2W$ skeleton scheme}
\label{sec:VII}
The expansion of susceptibility in terms of $\zeta$ is, of course, identical to the one found by the high-temperature series expansion method. To make the distinction between the high-temperature series formalism and Grassmannization approach clear, we discuss the skeleton formulation of the interacting fermionic field-theory based on dressed (or ``bold") one-body propagators ($G$) and bold interaction lines ($W$). This leads to the so-called $G^2W$ skeleton scheme (see for instance Refs.~\onlinecite{Heidin,Molinari2006} for the terminology): all lines in all diagrams
are assumed to be fully renormalized propagators and effective potentials, but vertex functions
remain bare. In Sec. \ref{sec:VIII} we show that the $G^2W$-expansion scheme offers a very simple
way to solve the 1D Ising model exactly.
\subsection{Objects and notation}
\begin{figure}[tbp]
\centering
\includegraphics[width=0.3\columnwidth, angle=-90]{fig_Psi}
\caption{Two low-order contributions to the generalized Luttinger-Ward functional $\Psi$.
Dashed lines denote bold Green functions for primed and non-primed Grassmann variables, and
wavy solid lines are effective potential lines.
}
\label{fig:psi_functional}
\end{figure}
The key objects in the standard skeleton scheme are the selfenergy ($\Sigma$) and the polarization function ($\Pi$). They are related to the Green function ($G$) and the
effective potential ($W$) by their respective Dyson equations.
The diagrams for $\Pi$ and $\Sigma$ are obtained by removing one $W$- or $G$-line, respectively,
from connected graphs for the generalized Luttinger-Ward functional $\Psi$, shown to second order in Fig.~\ref{fig:psi_functional}. In this setup, the expansion order is defined by
the number of $W$-lines (obviously, the discussion of Sec. \ref{subsec:D} does not apply
to the self-consistent skeleton sequence). All objects of interest are tensors; they have a coordinate (or momentum) dependence, as well as the legs orientation dependence for the incoming
and outgoing parts. This conventional scheme has to be supplemented with $\Psi$-graphs
involving $V_4$ vertexes to account for all contributions. We start with neglecting $V_4$ vertexes,
and discuss their role later.
In more detail, the formalism of the $G^2W$ expansion in the absence of $V_4$ vertexes
is as follows:
\begin{enumerate}
\item There are six bare two-body interaction vertexes $V_2$ $( RU, RL, RD, LU, UD, LD)$, see the second line in Fig.~\ref{fig:prediagrams}. They reside on the sites of the original square lattice and all have weight 1. Symbolically, we encode the tensor structure
of $V_2$ using a convenient short hand notation
$V_2=\sum_{\alpha , \gamma =1}^{4} V(\alpha, \gamma )n_{\alpha} n_{\gamma}$, where
\begin{equation}
V(\alpha, \gamma) = \begin{bmatrix} 0 & 1 &1 & 1 \\ 1 & 0 & 1 & 1 \\ 1 & 1 & 0 & 1 \\ 1 & 1 & 1 & 0 \end{bmatrix} .
\end{equation}
The row index represents the first leg enumerated according to the convention
$(R,U,L,D) \to (0,1,2,3)$, and the column index represents the second leg.
By doing so, we artificially double the number of vertexes from 6 to 12.
For example, the element $(0,2)$ corresponds to $n_Ln_R$ whereas $(2,0)$
corresponds to $n_Rn_L$, which is exactly the same term.
\item The selfenergies $\Sigma$ for the primed and non-primed Grassmann variables take the same value. Thus, we have to compute only one of them and we can suppress the index that distinguishes between the two Grassmann fields. The selfenergy defines the Green function
through the Dyson equation
\begin{equation}
G(\alpha, \gamma ) = G^{(0)} (\alpha, \gamma ) + \sum_{\mu, \nu}
G^{(0)}(\alpha, \mu ) \Sigma(\mu, \nu ) G(\nu, \gamma ) \,. \label{eq:Dyson}
\end{equation}
For a link going from site $i$ to site $j$, the first index $\alpha$ refers to site $i$ (in the above-defined sense), and the second index $\gamma$ refers to site $j$.
Note the absence of the momentum dependence in Eq.~(\ref{eq:Dyson}): The bold Green function remains local on the links in any order of renormalization. It means, in particular, that the only non-zero element for a link between sites $(0,0)$ and $(1,0)$ is $G_{02}$; it can be alternatively denoted as $G_{x}$ and, by $90^{o}$ rotation symmetry of the square lattice, is the same
for all links.
\item The matrix structure of polarization $\Pi$ is similar to that of $V$.
The 0th order expression based on bare Green functions is given by
\begin{equation}
\Pi^{(0)}_{\rm (x,y)}(\alpha, \gamma) = \zeta \begin{bmatrix}
0 & \!\! 0 & \!\! \delta_{x,1}\delta_{y,0} & \!\! 0 \\
0 & \!\! 0 & \!\! 0 & \!\! \delta_{x,0}\delta_{y,1} \\ \delta_{x,-1}\delta_{y,0} & 0 &0 & 0 \\ 0 & \delta_{x,0}\delta_{y,-1} & \!\! 0 & \!\! 0 \end{bmatrix} .
\end{equation}
\item The effective potential $W$ is defined through the Dyson equation
in momentum representation
\begin{equation}
W_{\bm q}(\alpha, \gamma ) = V(\alpha, \gamma ) + \sum_{\mu, \nu} V(\alpha, \mu)
\Pi_{\bm q}(\mu, \nu) W_{\bm q}(\nu, \gamma ) \,. \label{eq:W}
\end{equation}
We expect to see signatures of the ferromagnetic transition in matrix elements of $W_{{\bm q}=0}$
because they directly relate to the divergent uniform susceptibility $\chi$.
\end{enumerate}
\subsection{Zeroth order result}\label{sec:zero}
\begin{figure}[tbp]
\centering
\includegraphics[width=\columnwidth]{W_order0}
\caption{Divergence of the 0th order result for $W_{{\bm q}=0}$ at $\zeta_c = 1/3$ is compared
with the Frobenius norm and a reference line with power $-1$.
} \label{fig:order0}
\end{figure}
To obtain the 0th order result, we replace $\Pi$ with $\Pi^{(0)}$ in Eq.~(\ref{eq:W}).
For any $\zeta$ we compute $W_{\bm q=0}$ from Eq.~(\ref{eq:W}) by matrix inversion.
We find a divergence at $\zeta_c = 1/3$ (shown in Fig.~\ref{fig:order0}) that can be also
established analytically. We see that $W$ diverges as $(\zeta_c - \zeta)^{-1}$.
We get the same power law behavior for the $(0,1)$ matrix element as well as for the Frobenius norm---they just differ by a constant factor.
It is not surprising that our $\zeta_c$ is below the exact value for this model;
the skeleton approach at 0th order is based exclusively on simple ``bubble''
diagrams in terms of bare Green functions that are all positive, leaving to an overestimate of the critical temperature. Fermionic exchange cycles
and vertexes with negative weights do not contribute at this level of approximation.
\subsection{First order result}\label{sec:one}
We now include the diagrams with one $W$ line for the selfenergy and the polarization.
In real space we find
\begin{eqnarray}
\Sigma^{(1)}_x & = & \Sigma^{(1)} = - G_x W_{(1,0)}(2,0)= - G W_{(1,0)}(2,0) \nonumber \\
\Pi^{(1)} & = & G^4 W_{(0,0)}(0,0) + {\rm cycl.}
\label{eq:2dorder}
\end{eqnarray}
The matrix structure of $\Pi^{(1)}$ is identical to that of $\Pi^{(0)}$ and is not shown here
explicitly. Coupled self-consistent Eqs.~(\ref{eq:2dorder}), (\ref{eq:Dyson}), and (\ref{eq:W}) are
solved by iterations.
\subsection{Second order result} \label{sec:two}
\begin{figure}[tbp]
\centering
\includegraphics[width=\columnwidth]{gw_two_L64}
\caption{(Color online) Shown is the Frobenius norm of $W_{{\bm q}=0}$ (to second order)
on a lattice of size $64 \times 64$.
For comparison, the 0th order result is also shown.
The critical point is found to be at $\zeta_c \approx 0.35$ and the exponent is close to 1.1. } \label{fig:order2}
\end{figure}
As mentioned previously, to account for second-order terms in $\Sigma$,
one goes to the second order graphs for $\Psi$ and removes a $G$ line,
whereas the second-order terms for $\Pi$ are obtained by removing one $W$-line from the
third-order graphs for $\Psi$. The corresponding expressions in real space are
\begin{eqnarray}
\Sigma^{(2)} & = & - W_{(0,0)}(0,0) W_{(0,0)}(2,2) G^3 \nonumber \\
\Pi_{ (0,0)}^{(2)}(0,0) & = & G^6 W_{(0,0)}(2,2) W_{(1,0)}(0,2) \nonumber \\
\Pi_{ (1,0)}^{(2)}(0,2) & = & G^6 W^2_{(1,0)}(0,2) + \nonumber \\
{} & {} & G^6 W_{(0,0)}(0, 0) W_{(0,0)}(2,2).
\end{eqnarray}
The remaining non-zero contributions are obtained by invoking discrete lattice symmetries.
Note that to this order the polarization function is extremely local and contains only
same site and n.n. terms. Again, coupled self-consistent GW-equations are solved by
fixed-point iterations. The resulting behavior for $W$ is analyzed in Fig.~\ref{fig:order2}.
The transition point has slightly shifted to larger values of $\zeta$ compared to the zeroth-order result, and the exponent has also slightly increased.
\subsection{Relating ${\Pi}$ to the spin correlation function}
The $G^2W$-expansion scheme treats different bare vertexes (see Fig.~\ref{fig:prediagrams})
on unequal footing: the $V_2$ vertexes are fully dressed, but the $V_4$ vertexes are included perturbatively (we neglected them so far). These higher-rank vertexes have a weight of comparable magnitude to the $V_2$ vertexes (-2 for $V_4$ vs +1 for $V_2$). In addition, the difference in sign between the weights is expected to result in important cancellations between the diagrams and better
convergent series for the spin correlation function (this is how $\zeta_c$ increases towards
its exact value).
Formally, there is no valid reason for neglecting the $V_4$ vertexes altogether. Let us
show how they can be taken care of in the spirit of the shifted action approach.\cite{Rossi2015}
This discussion also gives us the opportunity to explain how the spin correlator is related
to the $G^2W$ skeleton expansion, which is most easily understood in the limit $\zeta \ll 1$.
By assuming that the skeleton sequence (without $V_4$) is solved, we introduce the full
polarization function $\bar{\Pi} (\hat{\alpha}, \hat{\beta} )$ through the Dyson equation
\begin{equation}
\bar{\Pi}_{\bm q}(\alpha, \gamma ) = \Pi_{\bm q}(\alpha, \gamma ) +
\sum_{\mu, \nu} \Pi_{\bm q} (\alpha, \mu) V(\mu, \nu) \bar{\Pi}_{\bm q}(\nu, \gamma ) \,. \label{eq:Pi_full}
\end{equation}
To be specific, we focus on the n.n. element $\rho_{(1,0)}$; similar manipulations hold
for any other distance. Now consider all diagrams for this correlator without the $V_4$ vertexes
within the $G^2W$ formulation (see Ref.~\onlinecite{Rossi2015}):
\begin{itemize}
\item Put one $V_1$ vertex on the origin site $(0,0)$ and the other $V_1$ vertex on the target site ${\mathbf r}=(1,0)$, see Eq.~(\ref{rhoprim}). There are $4 \times 4 = 16$ different ways of doing that depending on the directions of legs. Connect the legs with $\bar{\Pi}_{\bm r}(\alpha, \gamma)$. For example, in the limit of $\zeta \ll 1$, choosing the ($\alpha\! =\! 0$)-leg on site $(0,0)$ and the $\gamma=2$-leg on site $(1,0)$ results in the contributions $\zeta - 4 \zeta^5 + \ldots $. Similarly, the choice of $\alpha=1$ and $\gamma=1$ leads to the
contribution $\zeta^3$.
\item Put $V_3$ on $(0,0)$ and $V_1$ on $(1,0)$, and connect all legs with $\bar{\Pi}$ lines.
There are four ways to orient the $V_3$ vertex and for each one there are two choices for connecting legs with $\bar{\Pi}$ propagators. The leading contribution to $\rho_{(1,0)}$ goes hence as $-8 \zeta^5$.
\item Putting $V_1$ on $(0,0)$ and $V_3$ on $(1,0)$ gives the same contribution by symmetry.
\item Put one $V_3$ vertex on $(0,0)$ and the other $V_3$ vertex on $(1,0)$.
Now there are 16 ways of orienting both $V_3$ vertexes, and for each orientation there are 15 choices for connecting the legs. These contributions start at order $\propto \zeta^9$.
\end{itemize}
Next, we repeat the above procedure of connecting legs by adding one $V_4$ vertex,
which can be put on any site, after that we can add two $V_4$ vertexes etc. to generate a
perturbative expansion in the number of $V_4$ terms.
Compared to the original bare series in powers of $\zeta$, we have reordered the series:
the effective potential is summing up all $V_2$ vertexes, whereas we expand (and sample in a Monte Carlo framework) in powers of $\lambda_4$.
To illustrate this framework, let us take $\zeta = 0.01$ and recall that in the bare series $\rho_{(1,0)} = \zeta + 2 \zeta^3 + 4 \zeta^5 + 12 \zeta^7 + \ldots$.
The first 3 terms can be reproduced without $V_4$ vertexes and with only 1 $V_3$ on either the origin or the target site, see Figs.~\ref{fig:rho_generic_a}--\ref{fig:rho_generic_d}.
The fifth order coefficient originates from 16 ``simple" diagrams containing just $V_1$ and $V_2$ vertexes without any exchange. The diagrams containing a $V_3$ vertex yield a coefficient $-8$, and the exchange diagrams yield a coefficient $-4$.
On a $16 \times 16$ lattice, the propagators obtained in Sec.~\ref{sec:zero} ({\it i.e.}, to zeroth order) are
\begin{eqnarray}
\bar{\Pi}^{(0)}_{ (x,y)=(1,0) } (0,2) & = & 1.00000002 \times 10 ^{-02} , \\
\bar{\Pi}^{(0)}_{ (x,y)=(1,0) } (1,1) & = & 1.00010011 \times 10 ^{-06} ,\\
\bar{\Pi}^{(0)}_{ (x,y)=(1,0) } (1,2) & = & 1.00080057 \times 10 ^{-10} ,\\
\bar{\Pi}^{(0)}_{ (x,y)=(0,0) } (1,2) & = & 1.00020021 \times 10 ^{-08} .
\end{eqnarray}
We do not mention explicitly other symmetry-related elements.
The sum of all matrix elements for $\bar{\Pi}^{(0)}_{(x,y)=(1,0)}$ is $0.01000200160$. One clearly recognizes the coefficients $1$, $2$ and $16$ for the first-, third- and fifth-order contributions
to the bare series.
Contributions from the $V_3$ vertexes can be estimated from multiplying
$\bar{\Pi}^{(0)}_{ (x,y)=(0,0)}(1,2) \times \bar{\Pi}^{(0)}_{ (x,y)=(1,0)}(0,2)$ which yields $ \approx 10^{-10}$. There are four different diagrams, each with weight $-2$,
resulting in the above-mentioned
coefficient $-8$.
On a $16 \times 16$ lattice, the propagators obtained in Sec.~\ref{sec:two} ({\it i.e.}, to second order) are
\begin{eqnarray}
\bar{\Pi}_{ (x,y)=(1,0)}(0, 2) & = & 9.99999980\times 10 ^{-03} \\
\bar{\Pi}_{ (x,y)=(1,0)}(1, 1) & = & 1.00009999\times 10 ^{-06} \\
\bar{\Pi}_{ (x,y)=(1,0)}(1, 2) & = & 1.00120089\times 10 ^{-10} \\
\bar{\Pi}_{ (x,y)=(0,0)}(1, 2) & = & 1.00020005\times 10 ^{-08}
\end{eqnarray}
The sum of all matrix elements for $\bar{\Pi}^{(0)}_{(x,y)=(1,0)}$ is $0.01000200120$. One clearly recognizes the coefficients $1$, $2$ and $12$ for first, third and fifth order contributions to the bare series. For the fifth order contribution, we now obtain $12$ instead of $16$ thanks to the Grassmann exchange contribution that is accounted for properly at this level of approximation.
By adding the $V_3$ diagrams in the way described above we recover the correct result to this
order in $\zeta$ (which is +4).
The first instance of a $V_4$ vertex occurs in order $\zeta^6$ in the bare series. The relevant bare diagrams are the ones for $\rho_{(1,1)}$ with a $V_4$ vertex on site $(1,0)$ (and all cases related by the lattice symmetry). Our bold expansion can correctly account for this contribution if we put a $V_4$ vertex on this site and connect all unpaired legs with $\bar{\Pi}$ propagators. However, with the propagators obtained in Sec.~\ref{sec:two} we are not supposed to account for all possible diagrams in the bare series to order 6 because our bold expansion in Sec.~\ref{sec:two} is only accurate
up to order $\zeta^3$: Consider again $\rho_{(1,1)}$ and the bare diagrams where exchanges are possible on the links between the sites $(0,0) - (1,0)$ and $(1,0)-(1,1)$. Then there are irreducible non-local contributions that are not accounted for in Sec.~\ref{sec:two} with a positive weight that involves exchanges on both links in a correlated fashion. These contributions would obviously be accounted for in higher order corrections to $\Psi$, when $\Pi$ becomes non-local. This is also seen in the numerics: the $G^2W$ approach to second order yields a coefficient of $6$ for $\zeta^6$ contribution to $\rho_{(1,1)}$,
which is below the correct value of $10$.
\section{The Ising model in one dimension}
\label{sec:VIII}
Let us show that the proposed approach solves the 1D Ising model exactly,
both in the bare formulation as well as in the $G^2W$ skeleton formulation.
\subsection{Bare series}
In 1D, the only allowed vertex is $RL$ (the last one in the second line of Fig.~\ref{fig:prediagrams}). It has weight +1. The only allowed endpoints are $L$ and $R$ (the second and fourth vertexes
shown in the first line of Fig.~\ref{fig:prediagrams}). As expected, this means that there are no loops, no fermionic exchanges, and no minus signs in 1D. At order $n$ of the expansion for the spin correlator there is only one contributing diagram with weight $\zeta^n$
(up to the lattice symmetry). The susceptibility is hence
\begin{equation}
T \chi = 1 +2 ( \zeta + \zeta^2 + \ldots) = 1 + 2 \frac{\zeta}{1-\zeta},
\label{eq:1d_bare}
\end{equation}
reproducing the exact solution
with asymptotic behavior $\chi \propto \beta \exp(2 \beta)$ as $T \to 0$.
\subsection{$G^2W$ formulation}
The $G^2W$ skeleton expansion becomes exact already in 0th order,
\begin{eqnarray}
\Pi & = & \Pi^0 = \zeta \\
\Sigma &= & 0
\end{eqnarray}
which yields $G = G_0 =\sqrt{\zeta}$, $W = V / (1 - V \Pi) = 1/(1-\zeta)$, and also $\Pi = \zeta/(1- \zeta)$. This immediately leads to the same result as in Eq.~(\ref{eq:1d_bare}) when adding the end-point vertexes $L$ and $R$ to $\Pi$.
\section{Conclusion}
\label{sec:IX}
We have developed a general scheme for mapping a broad class of classical statistical
link (plaquette) models onto interacting Grassmann-field theories that can be studied
by taking full advantage of the diagrammatic technique. This mapping, in particular,
would allow to formulate an all-diagrammatic approach to $(d+1)$-dimensional lattice
gauge theories with finite density of fermions. The resulting field-theory looks very
complex because it contains a large number of Grassmann variables with numerous multi-point
interaction vertexes. Moreover, it is generically strongly-coupled at low temperature
meaning that an accurate solution using diagrammatic methods is only possible when
calculations are performed to high order and extrapolated to the infinite-order limit.
The complexity of the problem should not be taken as an indication that the entire idea
is hopeless. Monte Carlo methods were designed to deal with configuration spaces
of overwhelming size and complexity and arbitrary weights. In this sense, diagrammatic Monte Carlo methods simulating the configuration space
of irredicuble connected Feynman graphs are based on the same general principles
and one should not be surprised that they can evaluate the sum of millions of bare
(or skeleton) graphs, enough to attempt an extrapolation to the infinite-order limit.
What makes diagrammatic Monte Calro distinctly unique (apart from working with ever-changing number
of continuous variables without systematic errors) is the radical transformation of
the sign problem. It is completely eliminated in conventional sense because the
thermodynamic limit is taken first. Given that the number of diagrams increases
factorially with their order, finite convergence radius in $\zeta$ is only possible
if same-order diagrams cancel each other to such a degree that at high order
their combined contribution is not increasing factorially. In other words,
non-positive weights are {\it required} for the entire approach to work and we call
it the ``sign-blessing" phenomenon. Diagram weights for Grassmann/fermion fields
alternate in sign depending on the diagram topology; this leads to the sign-blessing
phenomenon for lattice models.
We illustrated the proposed approach by considering the 2D Ising model
as a prototypical example. We have deliberately chosen to work with the
generic formulation to avoid model specific simplifications because our
goal was not to solve the model but to demonstrate how one would proceed
in the general case. The ultimate goal is to explore how this field-theoretical
approach can help with understanding properties of lattice gauge models.
\section{Acknowledgement}
\label{sec:X}
We are grateful to A. J. Guttmann for providing us with references to the high-temperature series expansions and U. Wolff for drawing our attention to the work by S. Samuel. This work was supported by the National Science Foundation under the grant PHY-1314735, FP7/Marie-Curie Grant No. 321918 (``FDIAGMC"), and FP7/ERC Starting Grant No. 306897 (``QUSIMGAS").
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
\label{sec:1}
In recent times, various authors and research scholars have been exploring means to find suitable and reputed journals for publication of their research work. The drive behind this is to own appreciation or award for the quality work that they do. Also, institutional
assessment and evaluation depends heavily on peer-reviewed publications whether it be for academia or research labs. Generally, the trend observed among various faculties is to publish their research in journals with an 'international' tag attached to it. Thus, evaluating internationality is an open problem owing to the fact that such journals are vast in number; every such entity claims "internationality" but citation and influence are a bare minimum.\\
\par Data collected from IEEE Xplore in the year 2009 showed an exponential increase of 25\% in journal publications, when compared with previous years. A study
conducted by Buchandiran \cite{1} reveals an enormous increase in publication of journals between the years 2004 and 2009, whereby in the year 2009, 6,132 Indian institutions have contributed 23,745 papers out of which 15,880 were from academic institutions. This clearly shows that academic institutions contribute to the majority of such published works. Leonard Heilig and Stefan VoB
cite{2} observed a significant increase in the number of research publications (only in the area of cloud computing) from 2008 onwards. Elsevier's Scopus covered 15,376 publications till 2014 and Thomson Reuters Web of Science covered 8,262 publications in the same field.\\
\par On the flip side, there exists scam open access publishers who unethically and unprofessionally exploit the open access publishing model for financial rewards. They charge authors for publication fees and publish their work without providing true editorial services as well as other types of services associated with any legitimate journal. This shady publishing practice was first noticed by Jeffrey Beall,an academic librarian and a researcher at the University of Colorado in Denver. He scrutinized and investigated further and based on his findings, published his first compiled list of predatory publishers in 2010 \cite{19}. Continuing on the same line, Jeffrey Beall,\cite{20} regularly updated this list of fake publishers and had put forth certain criteria for categorizing such publications in order to prevent newer scholars from falling prey to such practices.\\
\par The phenomenon of predatory publishing (also known as the dark side of open access publishing) has exploded in recent years with the number of such publications expanding from 53,000 in 2010 to 420,000 in 2014. Accepting articles quickly without peer-review, pursuing academicians to submit articles or to serve on editorial boards, notifying authors of article processing fees only after paper acceptance, improper usage of ISSN and counterfeit Impact Factor values are some of the key indicators which have emerged from the observed working pattern of fake, predatory publishers. Till date, no fool-proof method has been devised to distinguish legitimate publishers from illegitimate ones. \\
\par An abundance of work has been done to evaluate the influence or prestige of scholarly articles and journals. Citation Index, a concept defined by Eugene Garfield (founder of Science Citation Index, SCI and the Institute for Scientific Information, ISI) serves as a link between similar scientific journals and literature. Citation pattern and citation frequency used by Garfield in his foundational work for Web of Science (Thompson Reuters Web of Science) initiated a vast spectrum of research and provided fully indexed and searchable research content. Thompson Reuters then initiated publishing Journal Citation Reports (JCR) to evaluate citation frequency of journals and down-the-line, Impact Factor. \\
\par Another initiative, Elsevier’s Scopus has a vast collection of peer reviewed scholarly journals and citations in scientific, medical, technical and social science domain. Scopus utilizes its database to provide another type of journal metric used for ranking for its journals through the SCImago Journal and Country Rank (SJR) portal \cite{8}. The SJR rank is a score evaluated from the past five years' data addressing a small number of journals. It is claimed that SCI, Thompson Reuters is a little more selective than Scopus.
The concept of citation index, Impact Factor and SJR ranking provide a limited respite to the above mentioned challenges of distinguishing and ranking legitimate publishers from the fake entities. This gives plenty of motivation and reason to work on proving a journal's credibility and integrity as well as ascertaining the quality, impact and influence of the publications. \\
\par Our initiative, ScientoBASE epitomizes a new approach for evaluating journals in a "height-weight" manner. The database, when complete, will help identify and bring adequate attention to quality journals, including industry practitioner domains, which otherwise would not be possible because certain journals namely Software Quality Professional (SQP) refuse to be indexed. \\
\par The remainder of the paper is organized as follows. Section 2 contains literature survey carried out on existing work related to measuring a journal's internationality vis-a-vis non-local influence. The section also brings forth systemic lags in establishing an unbiased score for ranking of journals. Section 3 defines internationality as perceived by the authors. It presents a schematic view of the approaches used to model internationality. Section 4 presents the algorithmic overflow for calculating internationality. Section 5 discusses in detail different techniques and algorithms used to collect scholastic parameters for the model described in section 6. These parameters are programatically scraped from multiple web sources such as Google Scholar, IEEEXplore, SCImago and Aminer \cite{25,26}. Once parametric data is acquired, these are then fed into the Cobb Douglas production function; an econometric model that is described in detail in section 6. Section 7 sheds some light on the merits of Source-Normalized Impact per Paper (SNIP) and shows why SNIP - and not Impact Factor - is a good albeit incomplete indicator for estimating non-local influence. Further, the section introduces new metrics, Non-Local Influence Quotient (NLIQ) and Other Citation Quotient and argues in favor of the usefulness of such metrics towards computing internationality. The paper concludes with a discussion on future work embodying ranking and clustering of journals according to internationality in their respective subject areas. The future work is commensurate with the current framework and model proposed in this paper.
\par
\section{Literature Survey}
\label{sec:2}
Neelam Jangid Snehanshu Saha, Siddhant Gupta, Mukunda Rao J \cite{6,7} in their work used a lightweight approach and introduced a new metric, Journal Influence Score (JIS), which is calculated by applying principal component analysis (PCA) and multiple linear regression (MLR) on citation parameters, extracted and processed from various scholarly articles in different domains, to obtain a score that gauges a journal’s impact. The higher the score, the more the journal is valued and accepted. Journal’s ranking results are compared with ranks of SJR, which internally uses Google's PageRank algorithm to calculate ranks. The results showed minimal error and the model performed reasonably well. Seyyed Mehdi et al. \cite{9} studied the scientific output of fifty countries in the past 12 years. In order to measure the 'quality' and 'quantity' of research output, a two-dimensional map is constructed and analyzed. Clusters are generated after analysis to represent country wise research output. There exists no ranking mechanism to rank countries with the maximum output in terms of quality and quantity of journals themselves.
\par Anup Kumar Das, Sanjaya Mishra \cite{10} discussed how research communities are preferring article-level metrics (ALM) over Journal Impact Factor (JIF) to assess the performance of individual scientists and their contributions. Gunther K. H. Zupanc \cite{11} also stressed on the unsuitability of using Journal Impact Factor to compare the influence of journals, especially when journals are from different areas. He claims that authors are tempted to publish their work in high-Impact Factor journals instead of journals that are best suited for their research work. A. Abrizah et al. \cite{16} compared the coverage, ranking, impact and subject
categorization of Library and Information Science journals, in which 79 titles were from Web of Science and 128 from Scopus. The prestige factor score of journals from JCR (Journal Citations Report 2010) and SJR (SCImago Journal Rank 2010) was extracted and the difference in ranks was noted. They observed a high degree of similarity in impact factor of titles in both Web of Science and Scopus. At the same time, authors also observed that the two databases differ in the number of journals covered.\\
\par Henk F. Moed \cite{29} introduced a different indicator of journal citation.
impact, Source Normalized Impact per Paper (SNIP). SNIP is defined as the ratio of the journal’s citation count per paper and the citation potential in its subject field. It aims to allow direct comparison of sources in different subject fields. There is no single ‘perfect’ indicator of journal performance. Delimitation of a journal’s subject field does not depend upon some predefined categorization of journals into subject categories but is entirely based on citation relationships. It is carried out on a paper-by-paper basis, rather than on a journal-by-journal basis. SNIP is based on citations from peer-reviewed papers to other peer-reviewed papers. Ludo Waltman et al. \cite{12} have discussed a number of modifications that were recently made to the SNIP indicator. The SNIP indicator considers a source normalized approach to correct the differences in citation practices between scientific fields. The key benefit of this approach is that it does not require the classification of subject fields, where the boundaries of fields are defined explicitly. There are some arguments around the original SNIP indicator’s properties that may be considered counter-intuitive. For instance, it is possible that additional citation has a negative correlation with journal’s SNIP value. \\
\par Gaby Haddow, Paul Genoni \cite{13} defined a new model - Excellence for Research for Australia (ERA) to determine the efficacy of citations measures in order to determine the quality of Australian social science journals. Chiang Kao \cite{33} investigated the contribution of different countries to international repositories of research in industrial engineering journals. After compiling journal data from ISI from 1996 to 2005, it was evident that the USA, UK and China are the top three countries to contribute articles to IE journals and six Asian countries are in the top ten. Yu Lipinga et al. \cite{34} classified common journal evaluation indicators into three categories, namely three first-level indicators. They are, respectively, the indicators on journal impact, on timeliness, and on journal characteristics. The three categories of indicators are correlated with one another, so a structural equation may be established. Then authors calculated the value of three first-level indicators and gave subjective weights to these indicators. This approach provides a new perspective for scientific and technological evaluation, in a general sense.
There are some limitations of this approach:
\begin{itemize}
\item The availability of basic data and the rationality of modeling bear much upon the evaluation results.
\item If there are too many indicators in a scientific and technological evaluation, data availability will be relatively difficult, and the evaluation cost will increase. If indicators are too few, they cannot provide adequate information.
\item If the data is inaccurate or wrong, no satisfactory results will be
obtained. In scientific and technological evaluation, sometimes certain data is very difficult to gather.
\end{itemize}
\par Gualberto Buela-Casal et al. \cite{14} performed a survey on existing measures of internationality and observed that a valid and quantitative internationality index should differentiate between “quality” and “internationality”. They suggested that in order to measure internationality, suitable weights should be assigned to certain identified parameters using a large-scale census of journal data. They proposed a neuro-fuzzy system to construct an unambiguous journal internationality index. \\
\par Chia-Lin Changa et al. \cite{15} examined the issue of coercive journal
citations and the practical usefulness of two recent journal performance metrics i.e. Eigenfactor Score, which may be interpreted as measuring “journal influence”, and the Article Influence Score, using Thomson Reuters Web of Science. Authors compared the two new bibliometric measures with the existing ISI metrics, total citations and the 5-Year Impact Factor (5Y-IF) of a journal. It is shown that the sciences and social sciences are different in terms of the strength of the relationship of journal performance metrics, although the actual relationships are very similar. Authors concluded that the Eigenfactor Score (measuring journal influence) and Article Influence performance metrics for journal are shown to be closely related empirically to the two existing ISI metrics, and hence add little in practical usefulness to what is already known, except for eliminating pressure arising from coercive journal self-citations. \\
\par Predatory publishing has earned a lot of attention (in terms of approval as well as criticism) from different sections of research communities across the globe \cite{21}. Beall's list of predatory journals has been welcomed by many open access supporters, whereas others have raised serious doubts about it's credibility. Walt Crawford \cite{24} in 2014 thoroughly investigated the list and called it a "One Man's List". He concluded that it should be ignored and suggested some steps to evaluate a journal's trustworthiness prior to submission.\\
\par Step 1: To make a pertinent decision whether "The International Journal of A" is a good target, one must look it up in the Directory of Open Access Journals (doaj.org). If the journal is not in the directory, look for another journal in a similar subject category.\\
\par Step 2: If the journal is in DOAJ, explore its site, its APC policy, quality of English used, its editorial board members - whether they are real people. Otherwise start from step 1.\\
\par Step 3: Check whether article title over the past issues makes sense within the journal's scope or if any author show up repeatedly within the past few issues. If so, go to step 1 again.\\
\par One can escape from predatory journals utilizing this technique. Nonetheless, it needs a lot of involvement in knowing how to assess journals as there is no scientific model which will do so for us.
Additionally, this algorithm is, to a greater extent, a manual investigation and hence ungainly and elaborate.
Therefore, there is a pressing need to build a complete, end to end web interface that also serves as repository and information visualization toolkit for scientometric evaluation, modeling and analysis. ScientoBASE is designed to serve this purpose and cater to internationality modeling and interface estimation of peer-reviewed journals in the fields of science and technology.
\section{Definition, Objective and Schematic View}
\label{sec:3}
This section defines internationality and presents an overview of the steps to achieve the end results. The authors would like to take this opportunity to stress that the "internationality of a journal" is defined here as a measure of influence beyond restricted boundaries. These boundaries may be geographical or regional or even cliques or networks of journals. It was observed during the course of this research that citations occur mostly within the journal from which the original citing article was published. The authors believe that in a community which is essentially international by nature, such trends don't bode well. A new metric which is an offspring of this realization, will be elaborated in due course. In order to remain clear about our objectives and dispel any confusion, we reiterate that internationality as defined and measured throughout this paper is a reflection of "non-local influence" and therefore does concur with the standard definition. The basic steps taken to achieve the end results are as follows:
\begin{itemize}
\item Defining and measuring internationality
\item Creating a suitable model
\item Validating the model
\item Generating granular clusters of “international” journals and conferences (part of future work)
\item Model the diffusion of internationality (part of future work)
\end{itemize}
\textbf{Definition}: Internationality of a journal, as proposed by the authors, is a holistic parameterization, of the international aspects of a journal's dimensions. These dimensions include - but are not limited to - quality of publications and measure international span of subscribing readers, authors and reviewers. These additionally evaluate the geographic source of a journal’s citations and the impact it spreads across nations. Authors explore these dimensions in succession and refer to International Collaboration Ratio; a parameter that indicates the ratio of articles whose author affiliations are from distinct nation, has the potential to be a suitable candidate for evaluating a journal's prestige. Likewise, extensive self-citation is a self promoting strategy which is unfairly used by authors to artificially boost their scientific influence, and thus indirectly inflate the publishing journal's impact factor. If used skillfully, this self-citation parameter can provide a good insight to judge a journal's credibility. It must be stated at this juncture, that some of the most common attributes of internationality such as ISSN number, constitution of editorial boards, country of publication and reputation of publishers as input factors are not considered. This is precisely because the authors do not view these attributes as entirely sufficient measures for internationality. Rather, use of such attributes as yardsticks in judging internationality is viewed as impediments towards objectively classifying journals. In recent times most journals have become international by structure (composition od editorial boards, reviewers, submitting authors etc) therefore in an essentially internatnational community of scholastic publication, \textbf{The authors define internationality as assimilation and evaluation of parameters that are \lq truely scholastically international\rq by including significant factors beyond the local manipulation of authors/editors.}
\\
\par There are well-accepted influence measurement parameters used by various web portals. Source-Normalized Impact per Paper (SNIP), allows comparison of sources across and within the same subject field by calculating their citation potential and normalizing their citation impact by dividing their RIP's (Raw Impact per Paper) with the calculated database citation potential. Integrity of academic publications would be at risk if editors coerce authors to cite their journals for enhancing their impact factor. With the intention to weaken the effects of this strategy, authors have introduced a new metric Non-Local Influence Quotient (NLIQ) which is a ratio of "non-local" citations of an article to the total number of citations. Larger the value of NLIQ, more "international" a journal is. \\
\par Taking all these factors into account, authors proposes a high-level design and methodology to model internationality index that would scrape and assemble the above mentioned parameters and generate journal clusters of high, moderate and low internationality. As already indicated, after computing "internationality", generating granular clusters of journals is a future plan of action. The current work embodies various algorithms for parameter acquisition and discusses the suitability of the new metric for influence calculation.\\
\par Empowered by data acquisition techniques, two approaches are put forth for modeling (Fig. 1). The first approach takes data from the Scopus and SJR portals and calculates a journal's score (JIS) \cite{6} using a multiple linear regression model on the scraped scientific indicators. Second approach, uses non-indexed, non-Scopus/non-Web of Science databases to acquire scientific parameters and evaluate a journal's internationality score generated from a Modeling Index (JIMI, Journal Internationalty Modeling Index) \cite{30, 31}. The approach uses Cobb-Douglas \cite{35,36} and Log Production Model on the parameters scraped from web. The algorithms and procedures are described in section 5 and 6.\\
\par The prestige/internationality of a journal is a convex combination of JIS [ please refer additional files on GitHub, \textbf{32} ] and Internationality Score, represented as-
Internationality of a journal,
\begin{center}
YI = ${\alpha}$ JIS + (1 - ${\alpha}$) JIMI ;
\\ $0 < {\alpha} < 1$
\end{center}
where YI refers to the internationality score as response
variable(to be sorted in decreasing order), JIS is the influence score obtained from metric JIS, JIMI is the score evaluated from work done using two parameters (JIMI) and ${\alpha}$ is a weight deduced from the cross correlation.
\begin{figure}[h]
\centering
\includegraphics[width=11cm,scale=0.8]{1.png}
\caption{Block diagram of Methodology.}
\end{figure}
For JIS, refer \cite{17} and Appendix I in\cite{32}.
\subsection{Journal Internationality Modeling Index(JIMI)}
The prestige of an academic journal is derived from quantifiable as well as non-quantifiable factors. Some commonly accepted factors that reflect a journal’s prestige are Impact Factor (IF), Eigenfactor, SCImago Journal Rank (SJR), Source-Normalized Impact per Paper (SNIP), Impact per Publication (IPP), internationality etc. Impact Factor, as per Thomson Reuter's definition \cite{22} is a measure of the frequency with which an article of a journal has been cited in a particular duration. The IPP (Impact per Publication) measures the ratio of citations in a year to scholarly papers published in the three previous years divided by the number of scholarly papers published in those same years. When normalized for the citations in the subject field, the Impact per Publication becomes the Source-Normalized Impact per Paper (SNIP). The SJR or SCImago Journal Rank is a measure of the scientific prestige of scholarly sources. SJR assigns relative scores to all of the sources in a citation network.\\
\par In this section, authors discuss a technique \cite{30} to quantify internationality by
exploiting a mathematical model, which determines the internationality of a journal by using two major metrics - Source-Normalized Impact per Paper (SNIP) and International Collaboration. Here author stresses on the efficacy of such a model and confirm the model theoretically. We prove that the model has a global maxima where a particular value of the inputs (SNIP and International Collaboration) would ensure some maximum value of internationality, subject to a constraint or set of constraints.
\begin{figure}
\centering
\includegraphics[width=11cm]{4.png}
\caption{Computation model of Journal Internationality Modelling Index (JIMI).}
\end{figure}
As shown in Fig. 2, the ,modeling approach uses web scraping technique to extract required features of various journals to generate CSV data files.
All these features from various data sources are processed further for only desired features of a journal which will contribute to the evaluation of internationality index. These features are specifically,
\begin{itemize}
\item International Collaboration Ratio
\item Source-Normalized Impact per Paper (SNIP)
\item Other-Citation Quotient
\item Non-Local Influence Quotient (NLIQ)
\end{itemize}
Using the above features as input parameters to Cobb-Douglas production function \cite{35,36}, the authors intend to measure the internationality index, proposing a score to gauge the influence of peer-reviewed journals.
\\
\\ \par \textbf{Sample Data}: This modeling approach uses a consolidated database generated by crawling the web (using software tools) to gather all recent (non-indexed) journals that are older than 3 years and younger than 5 years from Google Scholar. Reasons for selecting this are:
\begin{enumerate}
\item Elsevier considers a 3-year window for SNIP mainly due to the difference in the rates at which subject fields mature, whereas Thomson Reuters has a 2-year and 5-year window for Impact Factor (IF). As noted in section 7.1, one notable advantage of SNIP over IF is that SNIP’s 3-year citation window allows fields that move at a slower pace to be compared with those that advance fairly rapidly, in as fair a manner as possible. Whereas the 2-year IF and 5-year IF only favor one or the other.
Thus, authors have taken a window of 3-5 years in order to cater to journals in both categories.
\item Scraping the data which is indexed by SCOPUS, WOS is much easier, if not then, we will have to develop robust algorithm equivalent of Google scholar to fetch this data, which is out of scope for this research.
\item Another reason is many Journals shutdown due to various reasons in a very short span of time. Hence any journal needs minimum incubation time up to 3 years to prove its worth.
\end{enumerate}
\section{Algorithmic Overview:} This section discusses the basic steps taken to compute internationality and to generate granular clusters. \\
\textbf{Step 1:} Collect data (algorithms to extract data are shown in section 5.3)\\
\textbf{Step 2:} Pose internationality as a score: "y" as response variable.\\
\textbf{Step 3:} Model y = $f(x_1,x_2, x_3,.....x_i)$; i= 1,2,.....n
where $x_1, x_2.... x_i$ are the input variables as will be discussed in secton 6.\\
\textbf{Step 4:} (i)Perform down-selection, in case there are too many input variables, some of which could be highly correlated. Otherwise, go back to Step 3.\\
(ii) For simulation and visualization aid use a 3-D down-selection model.
\begin{center}
$\displaystyle y = A\prod _{i=1}^2 x_i {^{{\alpha}_i}} $
\end{center} , obtain "best" estimate of ${{\alpha}_i}$ ; use the best fit values.\\
\textbf{Step 5}: Compute "y" for each category. \\
\textbf{Step 6}: Observe the density and histogram plot. \\
\textbf{Step 7}: Decide on the granularity of internationality into several classes.\\
\textbf{Step 8}: Predict/visualize "variations" in "y" based on small perturbations in $x_1 and x_2 $. \\
The next section details the procedures and algorithms critical for data acquisition from the public domain. These include extracting data for the input parameters required for the model. Additional information such as journal name, country name etc is scraped for building a public repository.
\section{Data Acquisition from Google Scholar}
\label{sec:7}
\subsection{Collection} There are many advantages of using Google Scholar as a data source because it is free to access, easy to use and quick and comprehensive in its coverage.
Various studies \cite{5} have also shown that Google Scholar is a serious alternative data source for various reasons.
\begin{enumerate}
\item Not everything published on the internet is counted in Google Scholar:\\
Google Scholar indexes only scholarly publications. As their website indicates "we work with publishers of scholarly information to index peer-reviewed papers, theses, preprints, abstracts and technical reports from all disciplines of research". Some not scholarly citations, such as student handbooks, library guides or editorial notes slip through.
\par There might be some overestimation of the number of non-scholarly citations in Google Scholar, for many disciplines this is preferable to very significant and systematic under-estimation of scholarly citations in ISI or SCOPUS.
\item Non-ISI publications can be high-quality publications:\\
There is a misconception that ISI listing is a stamp of quality and one should ignore non-ISI listed publication and citations.
\par However, there are a few problems with this assumption. a) ISI has a bias towards Science and English language, b) ISI ignores the majority of publications in the social sciences and humanities as well as engineering and computer science fields.
\item Google Scholar flaws don’t impact citation analysis much:\\
There is no doubt that the Google Scholar’s automatic parsing occasionally provides us with nonsensical results. However, these errors do not appear to be frequent or important. They do not generally impact the results of author or journal queries much, if at all.
\par What is more important is that these errors are random than systematic. In contrast, the commercial databases such as ISI and Scopus have systematic errors that do not include many journals, nor have good coverage of conference proceedings, books or book chapters. Therefore, although it is always a good idea to use multiple data sources, rejecting Google Scholar out of hand because of presumed parsing errors is not rational.
\end{enumerate}
In spite of the fact that Google Scholar is an incomprehensible storehouse with uninhibited access, it does not provide an API. Moreover, Google obstructs any computerized web crawling. Subsequently we turned to web scraping. With occasional time delays included in the script, we could gather the required data from Google Scholar.
\subsection{Organization} Web scraping is the procedure of consequently gathering data from the World Wide Web.
Under this, we plan to develop completely robotized frameworks that can change over whole web website into organized data for further handling.
Fig. 4 demonstrates the essential segments of our methodology of web scraping Google Scholar. DOM Parsing is the philosophy which assists the system with retrieving element content created by client-side scripts utilizing undeniable web program controls, for example, the Internet Explorer browser or the Mozilla browser control.
These program controls likewise parse web pages into a DOM tree, in light of which program can recover parts of the pages.
\begin{figure}
\centering
\includegraphics[width=11cm]{scraper.png}
\caption{Block diagram of Web Scraping Methodology}
\end{figure}
\begin{enumerate}
\item Create Scraping Template: Inspect Element is a developer tool that allows to view the HTML, CSS, and JavaScript that is currently on a web page.
On nearly any web page, one can right click and select “inspect element”. This will pull up the developers console to view will the HTML and CSS of the web page.
Using this tool we explored the Document Object Model (DOM) tree of the Google Scholar Engineering and Computer Science Section.
Fig. 4 depicts the screen capture of the Inspect Element tool in use.\\
Left lower panel in the Fig. 4 is the Elements panel used to inspect all elements in the Google Scholar web page (top pane) in one DOM tree. Using this Elements panel one can select any element and inspect the styles applied to it.
The right lower panel is Styles pane. It shows the CSS rules that apply to the selected element, from highest priority to lowest.
Styles pane is used to view and change the CSS rules applied to any selected element in Elements panel.
Following are the various other information access provided in Inspect Element tool.
\begin{itemize}
\item Elements: Shows the HTML for the current page
\item Network: Shows all the GET and POST requests that are made while the developers console is open. One can also identify the requests that are taking the longest to process.
\item Sources: Allows to view the JavaScript files (and other files) associated with the page. This is most used for debugging as a web page is being developed, but can be helpful for coding your own JavaScript in Qualtrics as well.
\item Timeline: The timeline shows where time is invested when a web page is loaded/refreshed. It logs GETs, PUTs, calculations, parsing JavaScript, etc.
\item Profiles: Also helps see where time is being spent on a page. One can record time spent by function, by JavaScript Object, and by script
\item Resources: Allows to inspect the resources that are loaded onto a page. (i.e. cookies)
\item Audits: Analyzes a page as it is loading and then gives suggestions to decrease the load time
\item Console: This is a JavaScript console where one can try out code as if he/she were coding it for the web page. One can use it to log information about debugging, to test out code snippets, etc.
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=11cm]{ie.png}
\caption{Inspect Element tool usage to explore Document Object Model(DOM).}
\end{figure}
\item Explore Site Navigation: To further explore and understand the Google scholar site navigation for dynamic URL formulation, we used Beautiful Soul Parser.
BeautifulSoup parser is also called Elixir and Tonic “The Screen-Scraper’s Friend” \cite{18}
It uses a pluggable XML or HTML parser to parse a (possibly invalid) document into a tree representation. Beautiful Soup provides methods and pythonic idioms that make it easy to navigate, search, and modify the parse tree. \\
\item Automate Navigation and Extraction: Python is a scripting language which is easy to learn, powerful programming language. We have used the python interpreter library which is freely available in source and binary form for all major platforms.\\
\end{enumerate}
\subsection {Extraction}
\subsubsection{Algorithm for Feature Extraction - Algorithm 1} Algorithm 1 is for features extraction and internationality index computation for all the listed journals from Google Scholar (data source) under engineering and computer science field.
Features such as total citations, other-citation, International Collaboration Ratio and SNIP are computed for each one of these journals later on. Also, additional data was obtained from Aminer Citation Network data set\cite{26} based on a paper by Jie Tang et al. \cite{25}. Using web scraping we first extract all the journal names from the source: line 1. Then extract Total Citations count and all the Articles published in each one of these journals: line 3 and 4. Further on we compute the cumulative/averaged parameter values for that journal from the various values extracted for each article: line 5 to 8. The various function calls in these lines are explained ahead in the report under respective algorithms. the average value for the International Collaboration is computed: line 11. Finally, line 12 and 13 invoke the functions to compute the SNIP and Internationality Index.\\
\begin{algorithm}
\begin{algorithmic} [1]
\State \textbf{Input:} URL link of Google Scholar
\State \textbf{Output:} Features such as International Collaboration Ratio, SNIP, Other-Citations and Internationality Index
\State $JNames[]$ = $Fetch\_Journal\_Names\_from\_Google$(Engineering and Computer Science)
\For {every journal: $JNames[i]$}
\State TotalCites = Get the totalcites value
\State Get all the published articles/papers: $X[ ]$
\For {every article: $X[i]$}
\State $JNames[i]$.Selfcites += compute\_SelfCitations($X[i]$)
\EndFor
\State $x_1 = 1- JNames[i].Selfcites/TotalCites$; \textbf{compute other citation quotient} \State $x_2 = compute\_Intl\_Collaboration\_Ratio(JNames[i])/100$; \textbf{compute International Collaboration Ratio}
\State $x_3 = compute\_SNIP(JNames[i])/MaxSNIP$; \textbf{compute SNIP}
\State $x_4 = compute\_NonLocalIQ(JNames[i])$; \textbf{compute NLIQ}
\State $Internationlity\_index = CobbDouglasModel(JNames[i], x_1,x_2,x_3,x_4)$; \textbf{compute JIMI}
\Comment refer section 6 for Cobb-Douglas Model
\EndFor
\end{algorithmic}
\caption{Driver Algo: Algorithm to extract various features and to compute Internationality Index of Journals}
\label{algo0}
\end{algorithm}
\subsubsection{Journal Name Extraction - Algorithm 2}
This algorithm is to extract the journal names for Algorithm 1 to work upon. For the given source's URL, we perform web scraping to first extract all the subcategories of the Engineering and Computer Science field: line 1 and then in turn scrape the 20 journal names listed in each of the web links for these subcategories by dynamically generating the URL addresses using subcategory names: line 2
Then on we accumulate these scraped journal names in the spreadsheet: line 3
We successfully extracted about 1160 journal names from all the subcategories listed under Engineering and Computer Science category in Google Scholar. Fig. 5 shows the sample capture of the journal names extracted.
\begin{algorithm}
\begin{algorithmic} [1]
\State \textbf{Input:} A html file of Google Scholar web page: $HLINK$
\State \textbf{Output:} List of Journal Names
\For {every sub category link in $HLINK: SUBLINK$}
\For {every hyperlink in the $SUBLINK: JLINK$}
\State Print $JLINK.gs\_title$ from $ <td>$ tag to spreadsheets
\EndFor
\EndFor
\end{algorithmic}
\caption{$Fetch\_Journal\_Names\_from\_Google()$: Algorithm to Extract Journal Names from Google Scholar}
\label{algo1}
\end{algorithm}
\begin{figure}
\centering
\includegraphics[width=11cm]{7.png}
\caption{Sample list of Journals extracted into spreadsheet.}
\end{figure}
\subsubsection{\textbf{Other-Citation Quotient} Computation - Algorithm 3 }
Self-citation is defined as a citation where the citing and the cited paper share at least one author. Other-Citation is the complement of self-citation/total citations, i.e $ 1 - $ self\_citation/total citations. Algorithm 3 provides the skeleton of self-citation computation for an article in a journal. The denominator, total citations, is already computed by parsing web sources. The key to computing Other-Citations Quotient is to calculate self-citations. For this, we first scrape all the cited papers for the input article name (line 1). Then for each one of these cited papers check if it shares at least one common author name with the input article. If true then we increment the self-citation count (lines 3 and 4). Google Scholar lists a maximum of 1000 cited papers for any listed article. By adding all the individual self-citation counts for every article in a journal, we will get the total self-citations count for a journal (line 5 in Algorithm 1).
Fig. 6 shows the output from this algorithm which is a raw data extracted to spreadsheet.
Fig. 7 shows the processed raw data which provided the total self-citations for every listed journal.
\begin{algorithm}[H]
\begin{algorithmic} [1]
\State \textbf{Input:} article/paper name ($P$) from Google Scholar
\State \textbf{Output:} self-citation count for article / paper ($P$)
\State Get all citedPapers for article/paper($P$): $citedBy[]$
\For {Every cited paper: $citedBy[i]$}
\If {$P.author\_name$ \textbf{IN} $citedBy[i].author\_names$}
\State $IncrByOne(P.SelfCitationCount)$
\EndIf
\EndFor
\State \Return {$SelfCitationCount$}
\end{algorithmic}
\caption{$compute\_Self\_Citations()$: Algorithm to compute Self-Citation Count}
\label{algo2}
\end{algorithm}
\begin{figure}[H]
\centering
\includegraphics[width=11cm]{6.png}
\caption{Sample data extracted into spreadsheet.}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=6.5cm]{selfcitesTable.png}
\caption{Journals with corresponding self-citation counts.}
\end{figure}
\subsubsection{Non-Local Influence Quotient (\textbf{ NLIQ} ) - Algorithm 4 }
\par Influence is termed as a factor which causes a paper to be cited by other papers. Non-local refers to the fact that some citations originate from different journals; that is, not from the same journal in which the cited paper is published in. Thus, Non-Local Influence Quotient (NLIQ) is defined as follows,
\par Let \textbf{A} be the number of citations made from articles in one journal X to articles belonging to a number of different journals.
Let \textbf{B} be the number of citations from articles in journal X to articles in the same journal, X. Then, for a given journal, we have:
\par
\begin{equation*}
\textbf{Non-Local Influence Quotient} = \dfrac{A}{A+B}
\end{equation*}
\par It must be stressed that \textbf{other-citations} are uniquely different from \textbf{non-local influence}. Namely, an other-citation occurs when a paper cites another paper where no authors are in common. On the other hand, non-local influence is the number of citations made from one paper in a given journal, to a number of different journals - divided by the total number of citations. Section 7 outlines the implications of collaboration and NLIQ towards "internationality" and shows that SNIP values of journals are weakly correlated to their Non-Local Influence Quotient (NLIQ).
\par Weak non local influence is different from Zero non local influence by definition. So, when a journal has a low NLIQ, journal’s high SNIP will automatically compensate for bringing its Internationality score high enough to be ranked as a good international Journal.
For example: Journal of discrete mathematics has low NLIQ, but this low value will be compensated by its high SNIP value when we compute internationality index for it. So, it will still have a good score.
\par We score any Journal with zero non local influence with zero internationality index since it is Journal with no non local influence what so ever. Within a small community of niche journals; NLIQ =0 implies there is no diffusion. That can’t be good since that implies the articles published in that journal are not able to influence/generate idea strong enough to permeate even within the community, obliviate outside it.
For example : Astroinformatics, an emerging area; has quite a few journals; JAC, APJ, MNRAS, Astronomy and Space science to name a few. So the concern that only one journal may exist in a specific domain rendering NLIQ of that Journal = 0 is a misplaced appropriation.
\par So when the NLIQ is low (and not zero), the convex combination of JIS and JIMI balances out. Even if NLIQ is low, it doesn’t necessarily make JIMI low, because the presence of other factors will weigh in. Internationality is not geospatial measure of Journal. We define it as a measure which is devoid of local influence OR non local influence which compliments internationality.
In order to compensate for a low NLIQ in a valid case, the number of journals and articles published in the subject area need to be computed so that, normalization and total number of citation information is also available. This will help in rendering appropriate weights on NLIQ across different subject area.
\algnewcommand\algorithmicforeach{\textbf{for each}}
\algdef{S}[FOR]{ForEach}[1]{\algorithmicforeach\ #1\ \algorithmicdo}
\begin{algorithm}[H]
\caption{$compute\_NonLocalIQ()$: Algorithm to calculate Non-Local Influence Quotient}
\begin{algorithmic}[1]
\State \textbf{Input: }$journal\_name, citation\_database$
\State \textbf{Output: }$NLIQ \enspace of journal\_name$
\State $A\gets 0$ \Comment{external citation count}
\State $B \gets 0$ \Comment{internal citation count}
\State $J\_articles \gets \left[ \enspace \right]$ \Comment{used to store articles in a journal}
\State $count \gets 0$
\ForEach{$article \in citation\_database$} \Comment{get all articles in a journal}
\If{$article\left[journal\right] = journal\_name$}
\State $J\_articles[count++] \gets article$
\EndIf
\EndFor
\ForEach{$article \in J\_articles$} \Comment{get count of internal, external cites}
\ForEach{$reference \in article\left[references\right]$}
\If{$reference \in ARTICLE\_TYPE$} \Comment{reference is an article}
\If{$reference\left[journal\right] \enspace \textbf{!=} \enspace journal\_name$}
\State $A \gets A + 1$
\Else
\State $B \gets B + 1$
\EndIf
\EndIf
\EndFor
\EndFor
\State $NLIQ \gets A \enspace / \enspace \left( A + B \right)$
\State \Return $NLIQ$
\end{algorithmic}
\end{algorithm}
\subsubsection{International Collaboration Ratio - Algorithms 5, 6, 7} International collaboration accounts for the articles that have been produced by researchers from several countries.
In order to compute this parameter, we first extracted the country information of the journal and then the author affiliations for each one of the published articles in that journal. Every author's country is matched with the country of publishing journal. Ratio is calculated on the basis of weights assigned to different combination of authors affiliation and origin of the publishing journal.
Algorithm 5 is for collecting the country information of the journals. Fig. 8 shows the sample of country names of a few listed journals.
\par Algorithm 6 shows computation of International Collaboration Ratio of a journal. In international collaboration we look for collaboration between two or more scholars with affiliating institutes in different countries. A person with multiple affiliations then, we will pick primary/first listed affiliation during international collaboration computation.
\par We scrape the data of multiple affiliations (Algorithm 7) from the websites but consider only the primary (first listed) institute in the computation of international collaboration. Primary reason for this approach is the occurrences of nexus of dummy affiliations in middle-east countries \cite{35}.
As Bhattacharjee (2011) \cite{38} reported some years ago in Science, Saudi Arabian universities offer highly cited researchers contracts in which the researchers commit themselves to listing the Saudi Arabian University as a further institution in publications (or on highlycited.com). In return, the researchers receive an adjunct professorship which is connected with an attractive salary and a presence at the University of only one or two weeks per year (for teaching duties on site). Gingras (2014a) \cite{36, 37} names the added institutions as “dummy affiliations”, with no real impact on teaching and research in universities, allow marginal institutions to boost their position in the rankings of universities without having to develop any real scientific activities.”
\par Secondly, the best way could have been to have the author provide weights (or number of weeks) for each affiliation. Nonetheless this approach is out of scope in our research work. Therefore, we will consider the primary/first listed affiliation only for further calculation.
\par NOTE: Typically authors are affiliated to more than one institute in same country. A few examples include Harvard University and Center for Astrophysics. NewYork University and Courant Institute of Mathematical Sciences, University of Texas Arlington and Automation and Robotics research institute, Cambridge University and the Institute of Astrophysics. In the western world, land grant institutions under a bigger university setup is very common triggering multiple affiliations to a bunch of faculty associated with the larger university setup. In such cases multiple affiliations do not imply different institutions.
\begin{algorithm}
\begin{algorithmic} [1]
\State \textbf{Input: }List of Journal names from Algorithm 1: $J$
\State \textbf{Output:} Country information of the Journals
\For {Every journal\_name: $JNames[i]$}
\State $Fetch URL http://www.scimagojr.com/journalsearch.php?q="+journal\_name+"\&tip=jou" $
\State $write\_to\_spreadsheet(forall('div', {'id': 'derecha\_contenido'}) $
\EndFor
\end{algorithmic}
\caption{$Country\_info(Journal\_name)$: Algorithm to fetch country information}
\label{algo2}
\end{algorithm}
\begin{figure}[H]
\centering
\includegraphics[width=11cm, height=9cm]{Jcountry.png}
\caption{Sample list of the Journals with country name }
\end{figure}
\begin{algorithm}[H]
\begin{algorithmic} [1]
\State Input: Journal Name: $J$
\State URL to all the articles in that Journal : $J.all\_articles\_url[]$
\State Country information of the Journal: $J.contryName$
\State Output: \%international collaboration ratio of Journal: $J$
\Comment{ Compute the internationality weight of an article Based on the combination (Eg: out of 5 authors 2 are from same rest from other) deduce the weight of the article from a predefined values for a given combination,
Eg: For all authors from different countries weight=1,
For all authors from same country weight = 0,
For n/2 authors from one country and n/2 from others weight=0.5 }
\State $authAffs = []$
\For {Every article in $J.all\_articles\_url[i]$ }
\State $Authors\_Affiliation \leftarrow Fetch\_Author\_Affiliations(article)$ \Comment Algorithm 7
\State $authAffs.append(read\_author\_name\_and\_first\_affiliation information(Author\_Affiliation))$
\Comment {Generate 2D array of i:author name, j:country name}
\State $iNtrNationality\_wt[i]= compute\_wt(article) $
\EndFor
\State $J.iNtrNational[][]$
\Comment {Create one big matrix for a journal where i:country names, j:author names}
\For {every i in $authAffs$}
\If {$Country\_of(i['Affiliation']) == J.countryName) $ }
\Comment {if author's country same as Journal's country then make entry = 0}
\State ${J.iNtrNational[Country\_of(i['Affiliation'])][i['Author']] = 0}$
\Else
\State $J.iNtrNational[Country\_of(i['Affiliation'])][i['Author']] = 1$
\EndIf
\EndFor
\State x = Ratio of(Number of 0's and Number of 1's in $J.iNtrNational[][]$)
\State y = cumulative weights($iNtrNationality\_wt[i]$)
\State \Return $(\%international\_collaboration = \alpha x + (1-\alpha) y)$ \Comment $\alpha$ is a weight deduced from cross correlation
\end{algorithmic}
\caption{$Intl\_Collaboration\_Ratio(JNames[i])$: Algorithm to compute international collaboration ratio of a Journal}
\label{algo2}
\end{algorithm}
Algorithm 7 illustrates steps to fetch author affiliations of an article. An article URL is scraped to obtain Author name and respective affiliations. We extract all the affiliations in case of multiple affiliations for an author. results of one such article are in the as shown below
\begin{algorithm}[h]
\begin{algorithmic} [1]
\State \textbf{Input:} Link to the article from algorithm 5: $article\_URL$
\State \textbf{Output:} Author names and respective Affiliations
\State $authors[] \longleftarrow scraped\_author\_names(article\_URL)$
\State $list = []$ \Comment list of dictionaries
\For { every $author$ in $authors[]$ }
\State $dictionary\_element = \{'Author': author\} $
\State $count = 1$
\For {every $affiliation$ of $author$ }
\State $dictionary\_element.update\{'count':$affiliation$\} $
\State $count = count + 1 $ \Comment First, Second, Third Affiliations
\EndFor
\State $list.append(dictionary\_element)$
\EndFor
\State \Return list
\end{algorithmic}
\caption{$Fetch\_Author\_Affiliations(article)$: Algorithm to fetch author affiliations information for the article}
\label{algo2}
\end{algorithm}
\textbf{Data cleansing:} The scraped raw data is not ready for use until it is cleansed further and pre processed. We did following operations to extract the useful data while scraping and post scraping.
\begin{itemize}
\item Remove extra spaces such as tab, newline etc.
\item Strip/truncate unwanted characters suffixed and postfixed such as \lq \# \rq for the reference ID of the Institution in the html tags.
\item Encode text using UTF-8 encoder in order to take care of the Unicode characters in the extracted raw data while displaying.
\item The Institution Name is a complete address so we pre-process this address string to extract just the country name (in Algorithm 6) for International Collaboration ratio computation, which is solely based on just the country information of the Affiliated Institution.
\end{itemize}
\fbox{
\begin{minipage}{40em}
\{'1': u'Institute of Space Sciences (CSIC-IEEC), UAB, Barcelona 08193, Spain',
\\'2': u'Institute for Sciences of the Cosmos (ICC), University of Barcelona, Barcelona 08028, Spain',
\\'3': u'Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA',
'author': u'Beth A. Reid'\}
\\ \{'1': u'Institute of Cosmology and Gravitation, University of Portsmouth, Portsmouth P01 2EG',
'author': u'Will J. Percival'\}
\\ \{'1': u'Steward Observatory, University of Arizona, 933 N. Cherry Ave., Tucson, AZ 85121, USA',
\\ 'author': u'Daniel J. Eisenstein'\}
\\ \{'1': u'Institute of Space Sciences (CSIC-IEEC), UAB, Barcelona 08193, Spain',
\\ '2': u'Institute for Sciences of the Cosmos (ICC), University of Barcelona, Barcelona 08028, Spain',
\\ '3': u'ICREA (Institucio Catalana de Recerca i Estudis Avancats), Passeig Llus Companys 23, 08010 Barcelona, Spain',
\\ 'author': u'Licia Verde' \}
\\ \{'1': u'Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA',
\\ '2': u'Princeton Center for Theoretical Science, Princeton University, Jadwin Hall, Princeton, NJ 08542, USA',
\\ 'author': u'David N. Spergel' \}
\\ \{'1': u'Max-Planck-Institute for Astronomy, nigstuhl 17, D-69117 Heidelberg, Germany',
\\ 'author': u'Ramin A. Skibba' \}
\\ \{'1': u'Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA',
\\ 'author': u'Neta A. Bahcall' \}
\\ \{'1': u'Department of Physics and Astronomy, The Johns Hopkins University, 3701 San Martin Drive, Baltimore, MD 21218, USA',
\\'author': u'Tamas Budavari' \}
\\ \{'1': u'Particle Astrophysics Center, Fermilab, PO Box 500, Batavia, IL 60510, USA',
\\ '2': u'Kavli Institute for Cosmological Physics, Department of Astronomy \& Astrophysics, University of Chicago, Chicago, IL 60637,USA',
\\ 'author': u'Joshua A. Frieman' \}
\\ \{'1': u'Institute for Cosmic Ray Research, University of Tokyo, Kashiwa 277-8582, Japan',
\\ 'author': u'Masataka Fukugita' \}
\\ \{'1': u'Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA',
\\ 'author': u'J. Richard Gott' \}
\\ \{'1': u'Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA',
\\'author': u'James E. Gunn'\}
\\ \{'1': u'Department of Astronomy, University of Washington, Box 351580, Seattle, WA 98195, USA',
\\'author': u'deljko Ivezi' \}
\\ \{'1': u'Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA',
\\ 'author': u'Gillian R. Knapp' \}
\\ \{'1': u'Department of Astronomy and Astrophysics, The University of Chicago, 5640 South Ellis Avenue, Chicago, IL 60615, USA',
\\ 'author': u'Richard G. Kron' \}
\\ \{'1': u'Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA',
\\ 'author': u'Robert H. Lupton'\}
\\ \{'1': u'Departments of Physics and Astronomy, University of Michigan, Ann Arbor, MI, 48109, USA',
\\ 'author': u'Timothy A. McKay' \}
\\ \{'1': u'SUPA; Institute for Astronomy, University of Edinburgh, Royal Observatory, Blackford Hill, Edinburgh EH9 3HJ',
\\ 'author': u'Avery Meiksin' \}
\\ \{'1': u'Institute of Cosmology and Gravitation, University of Portsmouth, Portsmouth P01 2EG',
\\ 'author': u'Robert C. Nichol' \}
\\ \{'1': u'Los Alamos National Laboratory, PO Box 1663, Los Alamos, NM 87545, USA',
\\ 'author': u'Adrian C. Pope' \}
\end{minipage}}
\fbox{
\begin{minipage}{40em}
\{'1': u'Lawrence Berkeley National Lab, 1 Cyclotron Road, MS 50R5032, Berkeley, CA 94720, USA',
\\ 'author': u'David J. Schlegel' \}
\\ \{'1': u'Department of Astronomy and Astrophysics, The Pennsylvania State University, University Park, PA 16802, USA',
\\ 'author': u'Donald P. Schneider' \}
\\ \{'1': u'Fermilab, PO Box 500, Batavia, IL 60510, USA',
\\ 'author': u'Chris Stoughton' \}
\\ \{'1': u'Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA',
\\ 'author': u'Michael A. Strauss' \}
\\ \{'1': u'Department of Physics and Astronomy, The Johns Hopkins University, 3701 San Martin Drive, Baltimore, MD 21218, USA',
\\ 'author': u'Alexander S. Szalay' \}
\\ \{'1': u'Department of Physics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA',
\\ 'author': u'Max Tegmark' \}
\\ \{'1': u'Department of Physics, Drexel University, Philadelphia, PA 19104, USA',
\\ 'author': u'Michael S. Vogeley' \}
\\ \{'1': u'Department of Astronomy, The Ohio State University, 140 West, 18th Avenue, Columbus, OH 43210, USA',
\\ 'author': u'David H. Weinberg' \}
\\ \{'1': u'Department of Astronomy and Astrophysics, The University of Chicago, 5640 South Ellis Avenue, Chicago, IL 60615, USA',
\\ '2': u'The Enrico Fermi Institute, The University of Chicago, 5640 South Ellis Avenue, Chicago, IL 60615, USA',
\\ 'author': u'Donald G. York' \}
\\ \{'1': u'Department of Astronomy, The Ohio State University, 140 West, 18th Avenue, Columbus, OH 43210, USA',
\\ 'author': u'David H. Weinberg' \}
\end{minipage}}
We consider the only the primary affiliation for computation on international collaboration ratio. \\
\fbox{\begin{minipage}{40em}
Author : Beth A. Reid
\\ Affiliation : Institute of Space Sciences (CSIC-IEEC), UAB, Barcelona 08193, Spain
Author: Will J. Percival
\\ Affiliation: Institute of Cosmology and Gravitation, University of Portsmouth, Portsmouth P01 2EG
Author: Daniel J. Eisenstein
\\ Affiliation: Steward Observatory, University of Arizona, 933 N. Cherry Ave., Tucson, AZ 85121, USA
Author: Licia Verde
\\Affiliation: Institute of Space Sciences (CSIC-IEEC), UAB, Barcelona 08193, Spain
Author: David N. Spergel
\\Affiliation: Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA
Author: Ramin A. Skibba
\\ Affiliation: Max-Planck-Institute for Astronomy, Königstuhl 17, D-69117 Heidelberg, Germany
Author: Neta A. Bahcall
\\ Affiliation: Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA
Author: Tamas Budavari
\\ Affiliation: Department of Physics and Astronomy, The Johns Hopkins University, 3701 San Martin Drive, Baltimore, MD 21218, USA
Author: Joshua A. Frieman
\\ Affiliation: Particle Astrophysics Center, Fermilab, PO Box 500, Batavia, IL 60510, USA
Author: Masataka Fukugita
\\ Affiliation: Institute for Cosmic Ray Research, University of Tokyo, Kashiwa 277-8582, Japan
Author: J. Richard Gott
\\ Affiliation: Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA
Author: James E. Gunn
\\ Affiliation: Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA
Author: Richard G. Kron
\\Affiliation: Department of Astronomy and Astrophysics, The University of Chicago, 5640 South Ellis Avenue, Chicago, IL 60615, USA
Author: Robert H. Lupton
\\Affiliation: Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA
Author: Timothy A. McKay
\\Affiliation: Departments of Physics and Astronomy, University of Michigan, Ann Arbor, MI, 48109, USA
Author: Avery Meiksin
\\Affiliation: SUPA; Institute for Astronomy, University of Edinburgh, Royal Observatory, Blackford Hill, Edinburgh EH9 3HJ
Author: Robert C. Nichol
\\Affiliation: Institute of Cosmology and Gravitation, University of Portsmouth, Portsmouth P01 2EG
Author: Adrian C. Pope
\\Affiliation: Los Alamos National Laboratory, PO Box 1663, Los Alamos, NM 87545, USA
Author: David J. Schlegel
\\Affiliation: Lawrence Berkeley National Lab, 1 Cyclotron Road, MS 50R5032, Berkeley, CA 94720, USA
Author: Donald P. Schneider
\\Affiliation: Department of Astronomy and Astrophysics, The Pennsylvania State University, University Park, PA 16802, USA
Author: Chris Stoughton
\\Affiliation: Fermilab, PO Box 500, Batavia, IL 60510, USA
Author: Michael A. Strauss
\\Affiliation: Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA
Author: Alexander S. Szalay
\\Affiliation: Department of Physics and Astronomy, The Johns Hopkins University, 3701 San Martin Drive, Baltimore, MD 21218, USA
\end{minipage}}
\fbox{\begin{minipage}{40em}
Author: Max Tegmark
\\Affiliation: Department of Physics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
Author: Michael S. Vogeley
\\Affiliation: Department of Physics, Drexel University, Philadelphia, PA 19104, USA
Author: David H. Weinberg
\\Affiliation: Department of Astronomy, The Ohio State University, 140 West, 18th Avenue, Columbus, OH 43210, USA
Author: Donald G. York
\\Affiliation: Department of Astronomy and Astrophysics, The University of Chicago, 5640 South Ellis Avenue, Chicago, IL 60615, USA
Author: Željko Ivezić
\\ Affiliation: Department of Astronomy, University of Washington, Box 351580, Seattle, WA 98195, USA
Author: Gillian R. Knapp
\\ Affiliation: Department of Astrophysical Sciences, Princeton University, Princeton, NJ 08544, USA
\end{minipage}}
\section{Cobb Douglas Model: Internationality Score function}
In economics, Cobb-Douglas production function \cite{35,37} is widely used to represent relationship of outputs to inputs. This is a technical relation which describes the Laws of Proportion, i.e., the transformation of factor inputs into outputs at any particular time period. This production function is used for the first time, to compute the internationality of a journal where the predictor/independent variables, $ x_i, i= 1,2,...n $ are algorithmically extracted from different sources as explained in the preceding section.
Internationality, $ y $ is defined as a multivariate function of $ x_i, i= 1,2,...n $. Internationality score varies over time and depends on scholastic parameters, subject to evaluations, constant scrutiny and ever changing patterns. \\
Cobb-Douglas function is given by
\begin{center}
$\displaystyle y = A\prod _{i=1}^n x_i {^{{\alpha}_i}} $
\end{center}
where y is the internationality score, \\
$x_i$ are the predictor variables/input parameters
and $\alpha_i$ are the elasticity
coefficients. The function has extremely useful properties such as convexity/concavity depending upon the elasticity's. The properties yield global extrema which are intended to be exploited in the computation of internationality.
\noindent A sample Cobb-Douglas production function for two inputs, $x_1$ and $x_2$ and internationality of journal as output, \(y\), is written as -
\[ y=A{x_1}^{\alpha}{x_2}^{\beta}\nonumber \]
where:
\begin{itemize}
\item $0<\alpha,\beta<1$
\item $y$: Internationality of journal
\item $x_1$: International Collaboration (percentage)
\item $x_2$: SNIP (Source-Normalized Impact per Paper)
\end{itemize}
As explained in the subsequent sections, the sample model is easily extended to accommodate all relevant input/predictor variables extracted during the acquisition process [ Please refer Section $5$ ].
\subsection{Functional Form}
Here, \(x_1\) and \(x_2\) values which are modified values of international collaboration and SNIP respectively, are taken into consideration for different journals and using these optimal values for \(\alpha\) and \(\beta\) is computed. In order to have data lying between 0-1, transformations on the data set is performed, which give the final input values for Cobb-Douglas production function. Finally, the two variables along with the elasticity values of \(\alpha\) and \(\beta\) are placed in Cobb-Douglas equation to compute \textit{y}. Following is the algorithm used:
\noindent Algorithm to find optimum values of \(\alpha\) and \(\beta\):\\
1. Input values of \(x_1,x_2\)\\
2. Vary \(\alpha\) for \(x_1\) such that the corresponding \textit{y} reaches it’s maximum value. Similarly compute \(\beta\) values by varying \(x_2\).\\
\noindent In the figure (Fig. 9) below it can be seen that \textit{y} is maximum for \(\alpha\)=0.1 and \(\beta\)=0.1
\begin{figure}[ht]
\begin{center}
\includegraphics[width=7cm]{alphaBeta.png}
\end{center}
\caption{Optimum values of \(\alpha\) and \(\beta\)}\label{optimum}
\end{figure}
\noindent Hence, the optimum values of \(\alpha\) and \(\beta\) are 0.1.
Using these values of \(\alpha\) and \(\beta\) in the Cobb-Douglas production function \begin{center}y = $A{x_1}^{\alpha}{x_2}^{\beta}\nonumber$ \end{center} \textit{y} is computed which represents internationality of journal. In the next subsection, regression method is used to compute \(\alpha\) and \(\beta\). The general model is endowed to accommodate any number of factors as proved in section 6.2. However in section 6 and 6.1, two factors are considered as illustration for the 3-D plot. As observed, \(\alpha\) and \(\beta\) are varied to capture the maximum value of "y", internationality of the journal. If the number of input parameters are increased, visualization becomes untenable.
\subsection{Proof of Concept} The section proves the efficacy of the model for n number of variables/inputs, n being countably finite\\
As a first exercise, authors have done the simulation for 2 variables and extended to 3. This can be extended to n variables as shown below.\\
Consider the following production function:
\begin{align*}
&y=\prod_{i=1}^{n}kx_i^{\alpha{_i}}\end{align*} $n = 4$, $x_1$ to $x_4$ are the input parameters as described below:
\begin{itemize}
\item $x_1$ : Other-Citations Quotient = $1$ - ( self-citations / total citations )
\item $x_2$ : International Collaboration / 100
\item $x_3$ : SNIP value / maximum SNIP value
\item $x_4$ : Non-Local Influence Quotient
\end{itemize}
\par \noindent
\textbf{Note: }Due to the fact that these input parameters are not just raw numbers but instead defined as quotients having values between 0 and 1 - that is, they are normalized - the Cobb-Douglas model allows for a fair comparison between different subject fields where collaboration and citation trends differ dramatically (such as in Computer Science, the Social Sciences and Mathematics).\\
We have discussed the parameters, $x_1$, $x_2$ in section\textbf{ 5 - Data Acquisition}. This section following the analytical explanation of the Cobb-Douglas model, section $7$ contains discussion on $x_3$ and $x_4$, since both are not merely acquisition oriented but deserve discussion by their own merit. \\
\newline
\textbf{\underline{Lemma I}}: Maximum internationality score can be obtained at decreasing returns to scale which is true when -
\begin{align*}
\sum_{i=1}^{n}\alpha_i<1
\end{align*}
where $\alpha_i$ is the $i^{th}$ elasticity of the input variable $x_i$. Consider the following production function:
\begin{align*}
&y=\prod_{i=1}^{n}kx_i^{\alpha{_i}}\end{align*}
To prove:
\begin{align*}
\sum_{i=1}^{n}\alpha_i<1
\end{align*}
Consider the profit function:
\begin{align*}
\pi_n=\prod_{i=1}^{n}kx_i^{\
alpha_i}-\sum_{i=1}^{n}{w_i}{x_i}
\end{align*}
\(w_i\): Unit cost of inputs\\
Profit maximization is achieved when: \(p\frac{\partial f}{\partial x_i}=w_i\). Deriving the condition for optimization:
\begin{align}
pk\frac{\alpha_1}{x_1}&\prod_{i=1}^{n}x_i^{\alpha _i}=w_1\\
pk\frac{\alpha_2}{x_2}&\prod_{i=1}^{n}x_i^{\alpha _i}=w_2\\&.\nonumber \\&. \nonumber \\&. \nonumber \\&. \nonumber\\
pk\frac{\alpha_n}{x_n}&\prod_{i=1}^{n}x_i^{\alpha _i}=w_n
\end{align}
Multiplying these equations with \(x_i\), respectively-
\begin{align}p{\alpha _1}\prod_{i=1}^{n}kx_i^{{\alpha _i}}&=w_1x_1\Rightarrow p\alpha_1y=w_1x_1\\
p{\alpha _2}\prod_{i=1}^{n}kx_i^{{\alpha _i}}&=w_2x_2\Rightarrow p\alpha_2y=w_2x_2\\&. \nonumber \\&. \nonumber \\&. \nonumber \\&. \nonumber\\
p{\alpha _n}\prod_{i=1}^{n}kx_i^{{\alpha _i}}&={w_n}x_n\Rightarrow p{\alpha_n}y={w_n}x_n
\end{align}
Dividing equations (5) .. (6) by (4), following equations are obtained:
\begin{align*}
x_2&=\frac{\alpha_2}{\alpha_1}\frac{w_1}{w_2}x_1\\
x_3&=\frac{\alpha_3}{\alpha_1}\frac{w_1}{w_3}x_1\\ &.\\ &.\\
x_{n-1}&=\frac{\alpha_{n-1}}{\alpha_1}\frac{w_1}{w_{n-1}}x_1\\
x_n&=\frac{\alpha_n}{\alpha_1}\frac{w_1}{w_n}x_1
\end{align*}
Substituting these values of\(x_i\) in equation (1),
\begin{align*}
&pk\frac{\alpha_1}{x_1}\prod_{i=1}^{n}x_i^{\alpha _i}=w_1\\
\Rightarrow &pk\alpha_1x_1^{\alpha_1-1}\left(\frac{\alpha_2}{\alpha_1}\frac{w_1}{w_2}x_1\right)^{\alpha_2}\left(\frac{\alpha_3}{\alpha_1}\frac{w_1}{w_3}x_1\right)^{\alpha_3}....\left(\frac{\alpha_{n-1}}{\alpha_1}\frac{w_1}{w_{n-1}}x_1\right)^{\alpha_{n-1}}\left(\frac{\alpha_{n}}{\alpha_1}\frac{w_1}{w_n}x_1\right)^{\alpha_{n}}=w_1\\
\Rightarrow &pk{x_1}^{\left(\alpha_1+\alpha_2+....+\alpha_n\right)-1}\alpha_1^{1-\left(\alpha_2+\alpha_3+....+\alpha_n\right)}\alpha_2^{\alpha_2}....\alpha_n^{\alpha_n}w_1^{-1+\left(\alpha_2+\alpha_3+....+\alpha_n\right)}w_2^{-\alpha_2}....w_n^{-\alpha_n }=1\\
\Rightarrow&x_1=\left(pk\alpha_1^{1-\left(\alpha_2+\alpha_3+....+\alpha_n\right)}\alpha_2^{\alpha_2}....\alpha_n^{\alpha_n}w_1^{-1+\left(\alpha_2+\alpha_3+....+\alpha_n\right)}w_2^{-\alpha_2}....w_n^{-\alpha_n}\right)^\frac{1}{1-\left(\alpha_1+\alpha_2+....+\alpha_n\right)}
\end{align*}
Performing similar calculations following values of \(x_i, \left(i>=2\right)\) are obtained,
\begin{align*}
x_2&=\left(pk\alpha_2^{1-\left(\alpha_1+\alpha_3+....+\alpha_n\right)}\alpha_1^{\alpha_1}....\alpha_n^{\alpha_n}w_2^{-1+\left(\alpha_1+\alpha_3+....+\alpha_n\right)}w_1^{-\alpha_1}....w_n^{-\alpha_n}\right)^\frac{1}{1-\left(\alpha_1+\alpha_2+....+\alpha_n\right)}\\ &.\\ &.\\
x_n&=\left(pk\alpha_n^{1-\left(\alpha_1+\alpha_2+....+\alpha_{n-1}\right)}\alpha_1^{\alpha_1}....\alpha_{n-1}^{\alpha_{n-1}}w_n^{-1+\left(\alpha_1+\alpha_2+....+\alpha_n\right)}w_2^{-\alpha_2}....w_{n-1}^{-\alpha_{n-1}}\right)^\frac{1}{1-\left(\alpha_1+\alpha_2+....+\alpha_n\right)}
\end{align*}
Substituting values of \(x_i\) in production function,
\begin{align*}
y=\left(kp^{\left(\alpha_1+\alpha_2+....+\alpha_n\right)}\alpha_1^{\alpha_1}\alpha_2^{\alpha_2}....\alpha_n^{\alpha_n}w_1^{-\alpha_1}w_2^{-\alpha_2}....w_n^{-\alpha_n}\right)^\frac{1}{1-\left(\alpha_1+\alpha_2+....+\alpha_n\right)}
\end{align*}
y increases in price of its output and decreases in price of its inputs iff:
\begin{align*}
1-\sum_{i=1}^{n}\alpha_i>0\\
\sum_{i=1}^{n}\alpha_i<1
\end{align*}
Therefore decreasing returns to scale, is validated.
\subsection{Proof of Concavity of Cobb-Douglas function using Hessian Matrix}\
This section proves that the Cobb-Douglas \cite{36,37} production model is concave in nature and hence a maximum internationality score can be found at a particular value of input factors which in this case are international collaboration and SNIP value. The concavity of the function is proved by showing that the Hessian Matrix of the function is negative semi-definite.\\
\noindent{\textbf{Definition:}}
\begin{enumerate}
\item Suppose \(f\in C^2\), U is an open curve set, then \(f:U\subset R^n \rightarrow R\) is concave/strictly concave iff the Hessian Matrix \(D^2f(x)=H\) is negative semi-definite/ negative definite \(\forall x\in U\).\\\(C^2\): Class of continuous and second order differential functions[11].
\item Let S be a convex set [12]; \(x_1,x_2\) be any two points in S; then a function \(f:S \subset R_n \rightarrow R\) is concave if,
\begin{align} (1-\lambda)f(x_1)+\lambda f(x_2)\leq f((1-\lambda)x_1+\lambda x_2); \hspace{1cm} \lambda \in [0,1]\nonumber \end{align}
\item Constant and Decreasing Returns to scale: \cite{14} In the phase of constant returns, an increase in one input may yield an increase in corresponding output in the same proportion. The 3D plots obtained are concave.\\
Whereas, In decreasing returns to scale the deployment of an additional input will result into increase in output but at a diminishing rate or lower ratio.
\end{enumerate}
\noindent{\textbf{Lemma II:} \(f\in C, U\subset R;U\) is a convex, open set, \(f:R \rightarrow R, f\) is a concave iff }
\begin{align} f(x+\theta)\leq f(x)+\nabla f(x)\theta; \;\; \;\; \forall\; \; \theta \in R^{N}; x+\theta \in A; \nonumber \end{align}
C: Class of continuous and first order differential functions, \\
\textbf{Proof:} Using the definition of concave functions;
\begin{align*} \hspace{5mm} &f(\alpha (x+\theta)+(1-\alpha)x)\geq \alpha f(x+\theta)+(1-\alpha)f(x) \nonumber \\
&\Rightarrow f(x+\alpha \theta)-f(x)\geq\alpha(f(x+\theta)-f(x)) \nonumber\\
&\Rightarrow f(x)+\frac{f(x+\alpha \theta)-f(x)}{\alpha}\geq f(x+\theta)\\
&\Rightarrow f(x)+\nabla f(x)\theta \geq f(x+\theta) \;\;\;\; as\;\;\; \alpha \rightarrow 0 \end{align*}
\\
\textbf{Theorem 1:}\(f\in C^2;x\in R; f:R^2 \rightarrow R\) is concave iff the Hessian Matrix, \(H\equiv D^{2}f(x)\) is negative semi-definite \(\forall x\in U\). [necessary and sufficient condition for concavity]\\ \\
\textbf{Proof:} \(f\) is concave, for some\(\;x\in U\) and some \(\theta \neq 0\), consider the Taylor expansion;
\begin{align}f(x+\alpha \theta)=f(x)+\nabla f(x)(\alpha \theta)+\frac{{(\alpha\theta)}^{2}}{2} D^2f(x+t\theta)\; \; for\; some\; 0<t<\theta \nonumber \end{align}
By lemma;
\begin{align}\frac{{(\alpha\theta)}^2}{2} D^2f(x+t\theta) \leq 0 \nonumber \end{align}
Consider an arbitrary \(\alpha \rightarrow 0,\)\; and\; \(t \rightarrow 0\)
\begin{align}\theta^2 D^2 f(x)\leq 0 \Rightarrow D^2f(x) \leq0 \Rightarrow H\; is\; negative\; semi-definite. \nonumber
\end{align}
\subsection{Implications of Theorem 1:}
Cobb-Douglas is concave for conditions on elasticity, thus for such values of elasticity, the Hessian Matrix of the function is negative semi-definite and therefore concave and attains a global maxima.\\
Now consider, the Cobb-Douglas function;$ f(x_1,x_2)=k x_1^{\alpha}x_2^{\beta}$ with $k,\alpha,\beta>0$ for the region $x_1>0$ and $ x_2>0 $
$$H=\begin{bmatrix}\alpha(\alpha-1)kx_1^{\alpha-2}x_2^{\beta}&\alpha \beta kx_1^{\alpha-1}x_2^{\beta-1}\\\alpha \beta kx_1^{\alpha-1}x_2^{\beta-1}&\beta(\beta-1)kx_1^\alpha x_2^{\beta-2}\end{bmatrix}$$
First order principal minors \cite{13} of H are:
$$M_1=\alpha(\alpha-1)k{x_1}^{\alpha-2}{x_2}^{\beta};\; \; \; \; \; \; M_1'=\beta(\beta-1)k{x_1}^{\alpha}{x_2}^{\beta-2}$$
Second order principal minor is:
$$M_2=k\alpha \beta x_1^{2\alpha-2}{x_2}^{2\beta-2}[1-(\alpha+\beta)]$$
H must be negative semi-definite, this implies \(f(x_1,x_2)\) is concave.\\ This will happen if \(M_1\leq0,\; \; M_1'\leq0 \;\;and\; \; M_2\geq0\)\\ For decreasing and constant returns to scale: \(\alpha+\beta\leq1\), therefore
\begin{align} & \alpha \leq 1, \beta < 1 \nonumber \\
& \Rightarrow (\alpha -1) \leq 0 \nonumber \\
& \Rightarrow M_1 \leq 0 \nonumber \\
& (1-(\alpha+\beta))\geq 0 \nonumber \\
& \Rightarrow M_2 \geq 0 \nonumber \end{align}
Both conditions for concave function are satisfied by decreasing and constant returns to scale. Therefore, \(f(x_1,x_2)\) is concave, if
\begin{align} \alpha \geq 0,\beta \geq 0, \alpha+\beta \leq 1\nonumber \end{align}
\textbf{Significance of concavity}:\\
The extrema of the function, \(f(x,y)\) used to model "internationality" is useful in finding a global maximal value of the "internationality" indicator. The modeling paradigm is based on the fact that, there exists a maximum internationality score and the score/values in the neighborhood could be classified as the levels of internationality. It is, in this context, we explore if the maxima given by the concave function, i.e.Cobb-Douglas is the global maxima.\\
\textbf{ Theorem 2: Global maxima result:} \\
Let \(f(x_1,x_2)=kx_1^{\alpha}x_2^{\beta}: U\subset R^2\rightarrow R\)
be concave function on U; U is an open convex set; the critical point, \(x^{\ast}\) is a global maximum.\\ \\
\textbf{Proof:} \(x^{\ast}\) is a critical point. Therefore;
\(Df(x^{\ast})=0\) [D: first order partial derivative]\\
Using a well known result about concave functions;\\
\(f:R^2\rightarrow R\) is concave iff \(f(x_2)-f(x_1)\leq Df(x_1)(x_2-x_1)\;\; \forall x_1,x_2\in U\);\\ Therefore; \\
\(f(x_2)-f(x_1) \leq \frac{\partial f(x_1)}{\partial (x_1)}(x_2'-x_1')+...\frac{\partial f(x_2)}{\partial (x_2)}(x_2^2-x_1^2)\) \\
Since,\\
\(Df(x^\ast)\equiv 0\) using the inequality\\
\(f(x_2)-f(x^\ast)\leq Df(x^\ast)(x_2-x^\ast)\)
\(\Rightarrow f(x_2)\leq f(x^\ast)\; \; \forall x_2\in U\)
\begin{figure}[htbp]
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.8\linewidth, height=4cm]{frame1.png}
\label{fig:subim1}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.8\linewidth, height=4cm]{frame2}
\label{fig:subim2}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.8\linewidth, height=4cm]{frame3}
\label{fig:subim1}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.8\linewidth, height=4cm]{frame4}
\label{fig:subim2}
\end{subfigure}
\caption{y, internationality values from Cobb-Douglas Production function at various instances \cite{27} }
\label{fig:image2}
\end{figure}
\textbf{Note:}
\begin{enumerate}
\item The functional modeling, \(f(x_1,x_2)=kx_1^{\alpha}x_2^{\beta}\) may be extended to \(f(x_1,x_2 \dots x_n)=k\prod_{i=1}^{n} x_i^{\alpha_i}\); in which case \(f:U\subset R^n\rightarrow R\) \& the global maxima holds.
\item U doesn't necessarily be open, the global maxima is guaranteed to be on the closed set as well, since the search for global maxima is allowed on the boundary.
\item Any Cobb-Douglas function is quasi concave.
\item The values of elasticity are computed by using \textbf{fmincon} command in Matlab. These elasticity values are the exponents in the expression, \(f(x_1,x_2 \dots x_n)=k\prod_{i=1}^{n} x_i^{\alpha_i}\); the function \textbf{fmincon} is a built-in convex optimization tool in MATLAB and corroborates \textbf{Lemma II} proved above.
\end{enumerate}
A 3-D graph of Cobb Douglas function with output measured along the vertical axis is shown in Fig 10. The graph is a part of an AVI file whose frame are created in MATLAB to demonstrates the quasiconcave nature of Cobb Douglas model and to show how y reaches its maximum value at certain input values of $x_1$, $x_2$ $\alpha$ and $\beta$.
The Matlab code can be viewed on GitHub, Appendix I \cite{32}. The input parameters ($x_1$, $x_2$) are SNIP and other-citations/total-citations and the elasticity coefficients ($\alpha$, $\beta$) are taken along X, Y coordinates. Lower values of y is indicated in blue region which increases and at certain values of elasticity coefficients, reaches to its maximum as marked in red. This is a sample representation and can't include more than two input parameters.\\
\section{SNIP and Non-Local Influence Quotient, a new metric definition}
\par The authors have taken a four-pronged approach to thoroughly validate the use of SNIP and NLIQ in the Cobb-Douglas model. First, we shall show the merits of Source-Normalized Impact per Paper (SNIP) over Impact Factor (IF) \cite{28}. Then, the algorithm used to compute SNIP is described and verified with a sample data set. The third sub-section looks at citation patterns with increased granularity; namely inter-journal and intra-journal collaboration which will help show that SNIP is a good indicator of collaboration at the journal level. Lastly, we will look at Non-Local Influence Quotient (NLIQ) described in section 5.3.4 and justify its usage in calculating the internationality of a journal and show why SNIP alone should not be used to determine the relative ranks of journals in academia.
\subsection{Comparison of SNIP and IF}
\par
\par Now, Source-Normalized Impact per Paper (SNIP) measures a source's contextual citation
impact. It takes into account characteristics of the source's subject field, especially the
frequency at which authors cite other papers in their reference lists, the speed at which
citation impact grows, and the extent to which the database used in the evaluation covers
the field’s documents. SNIP is the ratio of a source's average citation count per paper, and the
citation potential of its subject field. It aims to allow direct comparison of sources in different
subject fields due to the subject-field normalization that takes place in calculating it.
\par The impact factor (IF) of an academic journal is a measure reflecting the average number
of citations to recent articles published in that journal. In any given year, the impact factor of a
journal is the average number of citations received per paper published in that journal during
the two or five preceding years.
\par \textbf{SNIP offers several advantages:}
\begin{enumerate}
\item \textbf{Openly Available and Greater Coverage}
How IF is calculated and the source database for citations is known only to Thomson
Reuters (ISI Web of Science) which means journals not present in their database are not
assigned an IF value. Also, not all journals indexed by them are provided an IF. This
disallows researchers from comparing journals which are not indexed.
\par Scopus, on the other hand, provides journal metrics values to all peer-reviewed
journals
indexed in their database which is comparably larger. Furthermore, SNIP can be calculated
from any Open Access journal using the white paper describing the calculation of SNIP. This
allows one to compare journals, however, the types of citations taken into account from
Open
Access journals must be kept in mind to give as fair a comparison as possible.
\item \textbf{Subject Field Normalization}
Life Sciences have a much higher IF as compared to Mathematical journals due to the
differences in citation behavior between the two fields.The quality of a journal cannot be
derived from its Impact Factor. Due to the fact that SNIP inherently normalizes for
differences in citation practices across subject fields, comparison of the ‘prestige’ of two
journals belonging to difference subject fields is possible.
\item \textbf{Citation Window}
SNIP has an ideal citation window, in the authors’ opinion. A three-year citation window
allows fields that move at a slower pace to be compared with those that advance fairly
rapidly, in as fair a manner as possible. Whereas the 2-year IF and 5-year IF only favor
one or the other.
\item \textbf{More Difficult to Game the System}
A journal’s impact factor is derived from citations of all types of content - including non-peer
reviewed material such as editorials. On the other hand, SNIP is derived only from citations
of peer-reviewed content and directed to peer-reviewed content, which makes it much more
difficult to game the system as the content goes through some form of scrutiny vis-a-vis
editorials.
\par Further, given the dramatic increase in predatory journals in recent years who merely
charge a fee for publishing an author’s paper albeit with deceitful tactics; and their
the proportional increase in their IF values - it is clear that IF is not a suitable metric for
measuring the `prestige' of a journal.
\subsection{Algorithm to Compute SNIP}
\par As shown in the original SNIP indicator designed by Henk F. Moed \cite{29} the SNIP indicator is defined as the ratio of a journal's raw impact per paper (RIP) and a journal's database citation potential in its subject field (DCP), that is the RIP value of a journal equals the average number of times the publications of that journal were cited in the three years the year of analysis. For example, if 200 publications were present in a journal from 2009 to 2011 and if these publications were cited 400 times in 2012, the RIP value of the journal for 2012 would be 400 / 200 = 2. In calculating RIP, both citing and cited publications are included only if they have the Scopus document type article, conference paper or review - i.e peer reviewed material. RIP is similar to journal impact factor (IF), although RIP uses three instead of two years of cited publications and only includes citations to the previously mentioned document types. RIP does not account for differences in citation practices among different journals.
\par The DCP value of a journal is equal to the average number of references in the publications belonging to the journal's subject field, where the average is calculated as the arithmetic mean. By finding the ratio of a journal's RIP to it's DCP, we can compare journals belonging to two different fields in a more fair manner. Algorithm 8 shows how to calculate the same.
Although there are certain differences between the original SNIP indicator \cite{29} and the revised SNIP indicator, the authors decided to forgo the latter. This is because an empirical analysis was done between the two and Ludo et al. stated that "from an empirical point of view the differences between the original SNIP indicator and the revised one are relatively small" \cite{12}.
\begin{algorithm}
\caption{$compute\_SNIP( cites[][], Jpub[], Jsize )$ : Algorithm to calculate SNIP}\label{alg:SNIP}
\begin{algorithmic}[1]
\State \textbf{Input:} Database of cites (cites[][]) made to publications of journal J (Jpub[] with Jsize publications) in year X to all documents (article, conference paper or review) in the three years preceding X
\State \textbf{Output:} SNIP value for journal J in year X
\State $journal \gets \textit{Jname}$
\State $year \gets \textit{read year\_to\_be\_computed\_for}$
\State $citation\_count \gets \textit{0}$
\ForAll{paper in Jpub}
\State $citation\_count \gets \textit{citation\_count + count of papers published in year - 1, year - 2, year - 3}$
\State $num\_papers \gets \textit{num\_papers + 1}$
\EndFor
\State $RIP \gets \textit{citation\_count / num\_papers}$
\State $DCP \gets \textit{Average number of 1-3 year old cited references contained in papers in the dataset citing the target journal}$
\State $median \gets \textit{median DCP of all journals}$
\State $RDCP \gets \textit{DCP / median}$
\State $SNIP \gets \textit{RIP / RDCP}$
\State \textbf{return } $SNIP$
\end{algorithmic}
\end{algorithm}
\par A random set of journals in Computer Science and Mathematics were selected and Algorithm 8 was used to calculate SNIP for these journals. The SNIP values thus obtained for the year 2010 with citation data taken from the Aminer Citation Network Data Set \cite{26} were compared with their actual values for the same journals and same year provided by Journal Metrics \cite{3,4}. The values thus obtained were not on par in terms of sheer magnitude with the corresponding values provided by Journal Metrics due to two main reasons.
Firstly, our database is a fraction of the one used by Journal Metrics in terms of size, and further, some citations in our database may not have been included by Journal Metrics in calculating SNIP - and vice-versa.
\vspace{2mm}
\par \noindent However, on further analysis using regression, we were able to show that there does indeed exist a strong correlation between the two values - SNIP calculated by us and the actual SNIP values provided by Journal Metrics. Figure 11 shows a linear regression line, fitting calculated SNIP and actual SNIP. We obtain an R-squared value of 0.7363 meaning 73.63\% of the variance in actual SNIP is accounted for by the variance in calculated SNIP.
\begin{figure}[h]
\centering
\includegraphics[width=0.4\linewidth, height=6cm]{SNIP_regression.png}
\caption{Actual SNIP v/s Computed SNIP}
\label{fig:subim1}
\end{figure}
\vspace{2mm}
To further validate our algorithm, we calculated SNIP for 189 journals which were common with both the Aminer data set and Journal Metrics' data set (from Elsevier). In this case, simple linear regression would not suffice; support vector regression (SVR) was used instead \cite{23}. Consider a set of linearly separable points, then the support vectors are those points which are difficult to classify and have a direct influence on the optimal location of the decision boundary. SVR is designed to find an optimal hyperplane which divides the two sets of linearly separable points such that an $\epsilon$-margin from either of these support vectors is obtained.
\vspace{2mm}
\par \noindent Initially, an RMSE of 1.162693 was obtained for linear regression and 1.163683 for SVR, without any tuning. Further tuning was performed by changing the values of $\epsilon$ and cost. The range of $\epsilon$ was narrowed from (0,2) to (0.68,0.72) with the cost parameter narrowed from $2^9$ to $2^2$. As a result, an RMSE of 1.116192 was obtained. Figure 12(a) shows how performance varies with $\epsilon$ and cost, with darker blue areas indicating optimal performance.
\begin{figure}[h]
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.7\linewidth, height=6cm]{Gridsearch.png}
\caption{Performance of SVR}
\label{fig:subim1}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.7\linewidth, height=6cm]{Rplot.png}
\caption{The tuned SVR model}
\label{fig:subim1}
\end{subfigure}
\caption{Results of Support Vector Regression}
\end{figure}
The graph in figure 12(b) shows the linear regression model in blue, untuned SVR model in red and the tuned SVR model in green which gave us the best fit.
\par To corroborate the results from SVR, exponential and polynomial regression was also performed on the same data set. The exponential model shown in Fig.13(a) returned an $R^2$ value of 0.9868. The relationship between the Y (aSNIP) and X (cSNIP) along with the coefficients (with 95\% confidence bounds) is best represented by the following equation:
\begin{equation} Y=1.273 * e^{0.5626*X} \end{equation}
\par The polynomial regression model shown in Fig.13(b) also resulted in an $R^2$ value of 0.9866. The equation obtained from polynomial regression is:
\begin{equation} Y=1.741 * X+0.8641 \end{equation}
\begin{figure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.7\linewidth, height=6cm]{exponen.png}
\caption{Exponential Regression}
\label{fig:subim1}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\includegraphics[width=0.7\linewidth, height=6cm]{poly.png}
\caption{Polynomial Regression}
\label{fig:subim1}
\end{subfigure}
\caption{Curve fit of Exponential and Polynomial Regression}
\end{figure}
We can confidently conclude - with the help of the derived RMSE and R-squared values, graphs and equations - that support vector regression, polynomial regression and exponential regression are all suitable for predicting actual SNIP values from their calculated values, with a very high degree of certainty.
\vspace{1mm}
\subsection{Journal Collaboration}
\subsubsection{Inter-Journal Collaboration}
\par The same journals used in section 7.2 Fig. 11 are considered in this section as well. The citation data was taken from Aminer Citation Network Data Set \cite{26} and correlated with journals having SNIP values taken from Journal Metrics \cite{3,4}. The citation network was then constructed as follows - the nodes are taken as journals, the size of each node is relative to it’s SNIP value taken from Scopus. The edge between two journals is a citation between a paper in one journal to a
paper in another journal, where the color gradient of the edge is relative to the number of
citations in total using Algorithm 9, as shown below in Fig. 14.
\par Figure 15 shows the citation network when we separate the edges on the basis of the table is given in Fig. 16, namely, those originating from and ending at a journal having a SNIP value above the median, and the same for below the median, as well as those going from a journal having a SNIP value above the median to below and vice versa.
\begin{algorithm}
\caption{$InterJ\_Collaboration$ : Algorithm to show Inter-Journal Collaboration graph}\label{euclid}
\begin{algorithmic}[1]
\State \textbf{Input: }$databse \enspace of \enspace citations$
\State \textbf{Output: }$graph \enspace and \enspace adjacency \enspace matrix \enspace of \enspace inter-journal \enspace collaboration$
\State $data \gets \textit{read aminer\_cites}$
\State $journals\_low,journals\_high \gets \Call{SplitByMedian}{$data$}$
\State $G \gets \textit{DiGraph}$
\ForAll{publication in data}
\State $papers \gets \textit{data[publication]}$
\ForAll{paper in papers}
\State $cites \gets \textit{data[publication][paper]}$
\ForAll{cite in cites}
\State $src \gets \textit{publication}$
\State $dest \gets \textit{cite['publication']}$
\If{src = dest}
\Comment{self cite within publication}
\State continue
\EndIf
\If{src in journals\_low and dest in journals\_low}
\State $type \gets \textit{1}$
\EndIf
\If{src in journals\_low and dest in journals\_high}
\State $type \gets \textit{2}$
\EndIf
\If{src in journals\_high and dest in journals\_low}
\State $type \gets \textit{3}$
\EndIf
\If{src in journals\_high and dest in journals\_high}
\State $type \gets \textit{4}$
\EndIf
\ForAll{type $\leftarrow$ (1, 2, 3, 4)}
\If{edge(src, dest) in G }
\State $G[src][dest]['weight'] \gets \textit{G[src][dest]['weight'] + 1}$
\Else
\State $G[src][dest]['weight'] \gets \textit{1}$
\EndIf
\EndFor
\EndFor
\EndFor
\EndFor
\State $G1 \gets \textit{G}$
\ForAll{edge\_weight in G1 $\leftarrow$ edges}
\State $edge\_weight \gets \textit{log10(edge\_weight)}$
\Comment{Normalize weight}
\EndFor
\ForAll{type $\leftarrow$ (1, 2, 3, 4)}
\State \Call{PlotNodes}{$G1$}
\State \Call{PlotEdges}{$G1$}
\State \Call{PlotAdjacencyMatrix}{$G$}
\EndFor
\end{algorithmic}
\end{algorithm}
\vspace{2mm}
\par \noindent When we split the citations up into four groups (see Fig. 15, 16) - namely, out of 19,359 inter-journal citations, a very large majority - 57.962\% - are between journals having SNIP values above the median value versus only 5.103\% between journals below the median. The citation network between journals of low SNIP value is quite sparse while that between journals of high SNIP value is dense - this indicates there is far more collaboration among journals of higher prestige or ranking, and little to no collaboration between those journals having a lower SNIP value, despite the fact that half of the journals taken into account were those below the median. This justifies our use of SNIP as a metric for collaboration - higher the SNIP, more the collaboration.
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth, height=6cm]{SNIPgraph1.png}
\caption{Citation network between all journals}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=11cm]{SNIPgraph2.png}
\caption{Citation network between journals separated along median SNIP value}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth, height=4.5cm]{table1.png}
\caption{Inter-Journal Citation Distribution}
\label{fig:subim1}
\end{figure}
\vspace{2mm}
\par \noindent One must be careful to note, these numbers could imply that not only do authors tend to favor their papers being published in prestigious journals, but they also cite those papers present in journals of similar level, or papers of the same journal itself. In turn, a cycle is created - authors who publish in prestigious journals are cited more often than those who publish in less prestigious one - thereby increasing the apparent prestige of that journal due to the increased citation count. Whether these citations are genuine or simply reciprocal in nature is not known.
\par \noindent This factor also fuels the growth of predatory journals with nary an oversight in terms of authentic peer-review - less prestigious journals exist with minimal collaboration simply because there was a low bar for a paper to be accepted.
\begin{figure}[H]
\centering
\includegraphics[width=70mm]{SNIPadjacency1.png}
\caption{Adjacency matrix for citations between all journals}
\end{figure}
\subsubsection{Intra-Journal Collaboration}
\par The adjacency matrix between the journals shown in Fig. 14 and 15 are given in Fig. 17 and 18, with the color gradient being relative to the total number of citations between the journals. It is observed that there exist darker points along the principal diagonal indicating more collaboration within a journal as opposed to inter-journal collaboration.
\par Papers published in one journal cite papers from the same journal much more often than those from different journals, regardless of the journal's SNIP value. This, too, leads to a cycle wherein an individual journal's prestige is increased by virtue of increased citations from within. It should be noted that journals of higher SNIP value have a lower NLIQ value as shown in Fig. 19, compared to journals of lower SNIP value - meaning citations are mostly restricted to the same journal they originate from. This in no way implies that there is a correlation between the two (as shown in section 7.4); merely revealing that journals most people would consider to be highly ranked (i.e by having higher SNIP values) exhibit only a low level of non-local influence.
Evidently, information about the internationality of these journals is incomplete - whether the authors are from the same institution or the same country or merely citing their previous works or those of colleagues due to reciprocity, as previously mentioned - is not known.
\par These are all factors that can be heavily gamed to enhance the prestige and rank of an author as well as the journal their papers are published in. Hence, we proposed NLIQ in section 5, which favors inter-journal collaboration as opposed to intra-journal collaboration thereby accounting for non-local diffusion of influence and fortifying our definition of internationality.
\end{enumerate}
\begin{figure}[H]
\centering
\includegraphics[width=11cm]{SNIPadjacency2.png}
\caption{Adjacency matrix for citations between journals separated along median value}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=11.5cm, height = 4cm]{table2.png}
\caption{Non-Local Influence Quotient Statistics}
\end{figure}
\subsection{Non-Local Influence Quotient, NLIQ}
Reiterating the definition from section 5.3.4, NLIQ is the number of citations made by articles published in a journal X to articles published in different journals divided by the total number of citations made by all papers in that journal X. Clearly, higher the number of external citations made by articles in a journal, higher the NLIQ of that journal.
\vspace{2mm}
\par \noindent
In Fig. 20, we see a plot of SNIP on X-axis versus NLIQ on the Y-axis. Even though it appears at first that journals with low SNIP values tend to have higher NLIQ, once we look into the goodness of fit and correlation statistics, we see that there is an insignificant relationship between the two. A linear regression line is fit and the R-squared value obtained is 0.1681 and the cross-correlation coefficient as -0.41 at lag 0 and near 0 at lag -1 and +1. A cross correlation value not close to -1 or +1 indicates that there is little to no correlation between SNIP and NLIQ. Similar R-squared values were obtained for higher degree polynomial regression models.
\begin{figure}[H]
\centering
\includegraphics[width=65mm]{SNIPvsNLIQ.png}
\caption{SNIP vs NLIQ}
\end{figure}
\vspace{2mm}
\par \noindent
Further, even though we don't possess the entire database of citations, we have proven that the Aminer data set has sufficient and even coverage which allowed us to calculate SNIP to a high degree of accuracy - and thus calculating NLIQ with a complete database will not vary by much, either.
\vspace{2mm}
\par \noindent
Clearly, SNIP is a good indicator of impersonal influence and therefore is used as one of the parameters for computing JIMI, but does not distinguish between the type of collaboration; intra-journal or inter-journal. NLIQ on the other hand is able to differentiate between the two types. It is vital to differentiate between the two; take for example a set of authors who constantly publish papers in one particular journal. They are only collaborating with people in the same area. On the other hand, if papers in different journals and subject areas are cited, cross-collaboration is established. This form of collaboration is indicated by NLIQ, which will be used alongwith SNIP in the Cobb-Douglas model for computing the internationality of a journal.
\section{Discussion}
There are possibly several parameters which are not considered in the proposed model. We list a few of those which could embellish the metric and internationality score functions.
\begin{enumerate}
\item Turnaround time(amount of time from the time of submission to publication) and Acceptance Ratio ((Number Of Accepted Papers) / (Number Of Submitted Papers) are two parameters which should be additional measures in computing Internationality Index, However, barring a few journals, this information is not available for scraping. As a future endeavor, we have created a survey to reach out to the editors and hope that this information would be available to us in six to eight months time. Appendix contains the screen shot of the survey. Assuming lukewarm response from the Editor-in-Chief's, this task will have to be accomplished programmatically. The Cobb-Douglas model of scoring is endowed with handling these parameters as long as numerical values could be scraped and computed.
\item Elsevier considers a 3-year window for SNIP mainly due to the difference in the rates at which subject fields mature, whereas Thomson Reuters has a 2-year and a 5-year window for Impact Factor (IF). As noted in section 7.1, one unmistakable advantage of SNIP over IF is that SNIP’s 3-year citation window allows fields that move at a slower pace to be compared with those that advance fairly rapidly, in as fair a manner as possible. Whereas the 2-year IF and 5-year IF only favor one or the other. Thus, authors have considered a window of 3-5 years in order to cater to journals in both categories. Another reason is that many Journals shutdown due to various reasons in a very short span of time. Hence any journal needs minimum incubation time up of 3 years to prove its half-life.
\item An important additional factor, readership profile has to be included as input parameter to the internationality index computation model, which is article, downloads count per country. This could be a challenging task as IP addresses are often masked. The authors don't expect journals to have this feature or co-operate with the authors in order to provide the data, or even allow the authors to fetch the data automatically. Finding a workaround is challenging and the current exercise has no provisions to build on this. This is one weakness that needs to be resolved.
\item Currently we consider four predictor variables as input to Cobb Douglas model (Internationality model, JIMI). This model can scale up to any number of inputs, in theory. However, when the numbers of inputs grow in practice the complexity increases exponentially. As the number of input parameters increase, possible curvature violation of the Cobb-Douglas model may create a problem towards estimating the elasticities. However this problem may be resolved by using stochastic frontier analysis \cite{39}. In order to conclude on the exception handling of the model, scale up is needed in future. There are other econometric models more resistant towards curvature violations that the authors intend to explore. From the model perspective, there is a potential curse of dimensionality problem. In that case, use of dimensionality reduction methods become inevitable. These methods help in identifying most significant parameters with high impact on the output when used in the model.
\item Volumetric information: NLIQ may vary widely across domains and this may hurt some journals more than others. Normalization, not implemented yet in the computation of NLIQ, is a pertinent landmark to accomplish. In order to obtain normalized NLIQ, we could divide the NLIQ of a Journal with the total number of Journals belonging to a domain. The challenge lies in correctly identifying classification of journals for categorization and count as there is always some overlap across domains. SCOPUS, WOS and GS all have their own logic for segregating the Journals. ACM subject classification is useful and clear enough and might be used for this purpose. The good part of the overlap mentioned above is that journals in niche domains don't get isolated and subsequently NLIQ computation doesn't suffer abruptly. This should alleviate the concern of decent and good Journals having lower NLIQ.
\item In order to adequately reflect the effect of NLIQ on the internationality score the elasticities need to be adjusted accordingly. The way it should be done includes the choice of an appropriate exponent accompanying NLIQ in Cobb-Douglas Model.
If NLIQ is low but the SNIP is high, we need to choose the elasticities in such a way that it does not hurt the internationality score of the Journal. This is what we call as elasticity boosting, which is achieved through the design of experiment (DoE) study. This study involves computing the percentage contribution of each factor in the Cobb-Douglas model towards internationality score and if contribution of certain factor is low, we adjust elasticity accordingly, at the same time adhering to the constraints of the optimization problem solved during the process. (Theorem 2 of section 6.4 )
\item Estimation of the constant of proportionality, $A$, in the Cobb-Douglas Model: $A$ has been assumed to be $1$ for simplicity in an otherwise complicated computation. However, $A$ in the Cobb-Douglas formulation may be estimated from data by using sophisticated fitting models and constrained optimization techniques. Once $A$ is suitably estimated, elasticities may then be predicted/fitted accordingly.
\item Differentiating citations: Based on the type of article, it is possible to differentiate between number of survey and original research article citations of a journal. Survey or review articles, written tutorials and technical reports tend to receive a large number of citations. It is necessary to distinguish between journals which publish original research articles only, a mix of research and review articles and review articles only(ACM Computing Survey). Therefore, internationality score and parameter quantification need to be normalized accordingly to ensure fair comparison between those journals.
\item Cognizant Citations: It might be a possibility that Editor in Chief's (EiC) are in mutual cognizance and very skillfully suggesting authors to cite articles from journals edited by themselves. For clarity, assume there exists two journals A and B. EiC of A endeavors to boost citation count for B and EiC of B returns the favor. Such cases could be modeled and quantified for penalty in the Cobb Douglas Score function. Graph theoretic modeling might help.
\end{enumerate}
\par It is pertinent to articulate that SNIP alone is not sufficient to compute influence and hence the necessity of defining a metric, NLIQ arises, which disregards the local influence diffusion within a journal. Authors believe that any artificial enhancement of a journal’s influence through coercive citation can be effectively subsided by NLIQ. The paper also uses a parameter - other-citation/total citation which reflects a journal’s integrity owing to the fact that no legitimate journal will promote authors and allow them to indiscreetly cite their own work. Parameters are chosen to ensure that every induced boost to a journals influence through coercive citations, extensive self-citations, copious citations(Definition to follow in next paragraph) or through any other mechanism is negated by the author’s model of internationality and novel metric definitions and computation where ever "conspicuous-citations" are promoted. \\
\par Computing non-local influence and internationality of authors/countries is an important and useful exercise. However, a few questions remain. How wide is the reach of our model? Is the proposed model of internationality portable to authors? If yes, what would be the metrics for such computation and how will the computation be carried out? Is the process consistent with the approach adopted here? Affirmative responses to these questions would strengthen the merits of the model adopted in this paper. As an illustrative example, let us consider NLIQ for authors. The metric definition would vary significantly from the way it is defined for journals. For an author, a local network would include all collaborators, students and supervisor. Therefore, computation of author NLIQ would imply exploring the author's citation networks and genealogy tree. Moreover, it is important to calculate copious citations of authors, if any. "Copious citations" is defined as, if between two authors A and B, say; A cites all published papers of B and vice-versa. This is not difficult to compute but can't be fed directly to the scoring model as other metrics. The reason is straightforward. The proposed scoring model is a constrained growth model and therefore copious citations should be included in the model as penalty subtracted from the main score function each and every time there is an instance of such citation. The model needs to framed as a profit function where revenue function symbolizes the author internationality and penalty is measured as cost function. The current model does not accommodate such cost function definition and penalty estimation.\\
\par Exclusive algorithms meet the requirement for computation of International Collaboration, Other-Citations Quotient and Non-Local Influence Quotient (NLIQ). These algorithms are written to create and develop a platform for \textbf{ScientoBASE: http://pesitsouth.pes.edu/scibase/}, a repository, which will consist of international journals by subject category with ranks and scores of internationality and necessary metric information by using various web-scraping and parsing techniques. Various metrics discuss ed above are gathered,computed and updated in real-time. This is a major task, once it is ascertained that the proposed model is self-sufficient. It is clear from the discussion and the formulation in previous sections, any new metric or adjustments in the existing metric don't require alterations in the model. However complicated the data assimilation part of the exercise is, the suite of algorithms help in accomplishing the broader aim of our research in defining a yardstick of scientific contribution and international diffusion; especially in niche areas such as Astroinformatics, Computational Neuroscience, Industrial Mathematics and Data Science from India, as well as other countries across the globe. These are emerging areas and many new journals have come up and for obvious reasons, the metrics are not reported in Scopus and ISI WoS. The outcome of our research will pave way for data and model validation and construction of a data visualization and web interface tool (ScientoBASE Toolkit), an open source web interface, that will compute the scores and provide visualizations of all essential parameters of internationality, particularly for the journals in emerging areas as mentioned above. It is immensely beneficial from pedagogical and scholastic standpoint to be able to use a web-kit and understand the growth of Indian as well as global Scientometry in state of the art and emerging areas in Science and Technology.
\section{Conclusion and Future Work}
Internationality has thus been defined and perceived as the degree to which a journal transcends local communities and boundaries, with respect to the quality of publication and influence. The methods illustrated in the paper ensures disposition of any kind of local influence that unreasonably boosts scholarly impact of journals. The paper meticulously defines internationality of peer-reviewed journals as a measure of influence that spreads across boundaries and attempts to capture different and hitherto unperceived aspects of a journal for computing internationality.The current work utilizes parameters like "International Collaboration Ratio" which incorporates participation of authors from different demographic regions. Authors humbly submit that "true" internationality of a scholarly publication is necessarily contextual and must be devoid of local or community influence. Source-Normalized Impact per Paper (SNIP) is another parameter taken into account which normalizes the citation pattern within a subject field allowing the comparison of journals belonging to two different domains. This is one of the reasons why authors preferred SNIP over Thomson Reuters Impact Factor (details in section 7.1).\\
\par Exclusive algorithms are written to scrape metric information from the web. Algorithms to sweep journal, author and article level information needed scrutinization of web pages. However, thanks to the simplicity of Python, it made most of the scraping task hassle-free incurring lesser overhead. Acquisition of journal names, origin, article names, author names and their affiliations were the first few steps in tailoring the parameters of internationality index. These attributes don't contribute to the model directly but are pivotal to creating the indigenous database for subsequent computation. Data acquisition and build are therefore significant contributions of the paper and should not be overlooked.\\
\par The acquired data was validated using different regression techniques. It was observed that values obtained from SNIP algorithm were not very close to the original ones. This is because the native acquisition algorithms could not have scraped through all databases due to the prohibitory firewalls built in several of those. Predictive analytic techniques are used to overcome such barriers so that data recorded in our database is reasonably accurate, endowed with appreciable " goodness of fit " statistic. The derived RMSE and R-squared values from support vector regression (SVR), linear regression, polynomial and exponential regression were found to be satisfactory. This concludes that any of these regression methods can be used to predict original SNIP (for detailed analysis, refer section 7.2). \\
\par Cobb-Douglas Production Function is used for the first time to model internationality of journals. Appropriateness and adequacy of the model is evaluated in section 6. It is shown that a function is strictly concave if the Hessian Matrix of the second order partial derivatives is negative semi-definite (Theorem 1). The property holds true for the production function implying that the function is strictly concave in nature given that certain conditions on elasticity are met. The importance of concavity lies in validating that the maximum value obtained by the production function is actually a global maxima (Theorem 2 proves this) and the search for such maxima via the model is complete once the maxima is found,by simulation and otherwise (Fig. 10). This maxima is then used as an indicator of highest internationality score and the subsequent neighborhood values may define lower international levels for the same journal. This process is iterated for all journals sweeping through the database. \\
\par Painstaking care has been exercised in creating a knowledge base of citation pattern followed by authors when they publish their work. To investigate the pattern, inter-journal collaboration network was created from Aminer and Journal Metric dataset. Dense citation network between journals of high SNIP values validated the fact that authors are not only tempted to publish their work in prestigious journals but are also inclined to cite papers of journals having a higher SNIP value. By doing so, receiving citation in large numbers is assured. The intra-journal collaboration network, on the other hand, is a reflection of author’s tendency to cite the papers published in the same journal, suggesting signs of community behavior practiced within journals. In an attempt to disregard such publishing practice, authors, while computing internationality have considered parameters that precisely and unambiguously define, measure and render new meaning to the internationality of journals. Thus, Non-Local influence Quotient, \textbf{NLIQ} is a major contribution for computing internationality and could potentially be a metric to be used by peers, the authors believe! \\
\par Commensurate with the current work, author's research contributions may be summarized as follows.
\begin{itemize}
\item Quantification of `internationality' of peer-reviewed journals as a measure of influence, introducing a novel treatment by defining new parameters and acquiring new data.
\item Post-acquisition, extensive cleaning is performed and rigorous pre-processing was done to make the data readable and usable for the model.
\item Definition of Non-Local Influence Quotient (NLIQ): It is determined by computing the ratio of journal's "non-local" citations to its "local" citations. The parameter signifies the spread of a journal's influence outside its boundaries. It is not documented, but common knowledge that external and internal factors are at play to ramp up impact factors of journals, in the form of suggestions to cite articles from the same journal. This explains the importance of Non-Local Influence Quotient \textbf{NLIQ} as it could enunciate the \textbf{bias-corrected} impact of journals by boasting of greater number of non-local citations. Therefore the diffusion of a journal's internationality is not manipulated by local factors if it possesses greater NLIQ i.e closer to \textbf{$1$}. NLIQ is thus, a reasonably trustworthy indicator of internationality and a significant outcome of the manuscript.
\item Definition of Other-Citations Quotient: If a journal's self-citations/total citations ratio is high, then papers in a given journal more frequently cite other articles in the same journal, than articles in different journals. That is, a high level of intra-journal collaboration is exhibited as opposed to inter-journal collaboration. Hence, journals with a high self-citations/total citations value cannot be rewarded a high internationality score. In fact, such journals must be penalized. Self-citations/total citations needs to be low in order to appreciate influence diffusion. This prompted us to define "Other-Citations Quotient" as $ 1- $ (self-citations/total citations); if self-citations equal total citations for a journal, then a journal's internationality score shall be rendered \textbf{ZERO} since such a trend reflects closed-community behavior and not true internationality, as defined by the authors.
\item Definition of "internationality" as a metric that shows evidence of non-local diffusion. To effectively reduce the effect of localization, SNIP is considered as a parameter for influence calculation.
\item Novel Algorithms: Developing algorithms to compute International Collaboration, Other-Citations Quotient, NLIQ and SNIP as part of research carried out by the authors. The algorithms scrape and compute the required parameters to be fed into Cobb-Douglas model as a part of internationality computation.
\item Predictive Analytics: Extensive validation process is carried out on the values obtained from scraping algorithms particularly for SNIP. Regression analysis and support vector regression is performed to confirm these values and the results are found to meet the expected level. Elaborate simulation and testing support the validity of our results.
\item Normalization: The input parameters fed to the Cobb Douglas function for computing internationality score are normalized. For example, NLIQ takes into consideration the ratio of citations (external to total) and not raw numbers. This practice allows for a fair comparison between different subject fields such as Computer Science, the Social Sciences and Mathematics where collaboration and citation trends differ remarkably.
\end{itemize}
\par It is possible to consider parameters such as number of article downloads per country and average cites per country along with the ones already included in the model. The model and data acquisition methods may be extended further to visualize growth of a subject based on region (cartogram), author (geospatial influence), topic and journal (spatial diffusion temporal invariant model). Further, this may be extrapolated to include normalization of scientific contributions and diffusion of scientometric indices in niche areas. Once the normalization practice is put to place, ranking and clustering of journals based on internationality may be proposed. This is not very difficult but should be done carefully as just like any other metric, there could be unfair disparity in the internationality scores of journals. Normalization, thus, is central to this entire exercise.
\par Authors do realize that there exists a plethora of metrics for ranking and scoring mechanisms. A practical approach would be to propose one, supported by the two powerful models, Multiple Linear Regression (used in JIS) for general influence and Cobb Douglas Model (in JIMI) for international influence. Authors intend to compute a single score, \textbf{RAGIS -Reputation and Global Influence Score}, $y_{ragis}$ , as a convex combination of JIS and JIMI. JIS ( \cite{32}, Appendix II in the repository contains details of JIS) computes influence score for journals that are indexed in SJR, Scopus and Web of Science. Computation of JIMI brings many other journals under it's fold. \textbf{RAGIS} would facilitate clustering of journals as demonstrated by supplementary data provided in the authors repository (see note below). This is set as a future goal. The authors endeavor to pursue this line of reasoning, hoping for proliferation to a comprehensive set of journals and to cater to a much larger audience. \\
\textbf{{Note}:} Additional file on GitHub \cite{32} contains Matlab source code that generates an audio/video interface file. The file demonstrates frames of 3D plot of Cobb Douglas Production function. The file contains sample snapshots of the proposed toolkit, as well as other source code used in the course of this manuscript.
|
\section{Introduction}
In this paper we study the Dirichlet problem for the so-called $f$-minimal graph equation on a complete non-compact
$n$-dimensional Riemannian manifold $M$ with the Riemannian metric given by $ds^2 = \sigma_{ij}dx^i dx^j$ in local coordinates.
We equip $N=M\times\R$ with the product metric $ds^2 +dt^2$ and assume that $f\colon N\to\R$ is a smooth function. The Dirichlet
problem for $f$-minimal graphs is to find a solution $u$ to the equation
\begin{equation} \label{mingrapheq}
\begin{cases}
\dv \dfrac{\nabla u}{\sqrt{1+|\nabla u|^2}} = \ang{\nb f,\nu} \quad \text{in } \Omega \\
u|\p\Omega = \varphi,
\end{cases}
\end{equation}
where $\Omega \subset M$ is a bounded domain, $\bar{\n}f$ is the gradient of $f$ with respect to the product Riemannian metric, and $\nu$ denotes the downward unit normal to the graph of $u$, i.e.
\begin{equation}\label{unitnormal}
\nu = \frac{(\nabla u,-1)}{\sqrt{1+|\nabla u|^2}}.
\end{equation}
The regularity assumptions on $f,\ \p\Omega$, and on $\varphi$ will be specified in due course.
The equation \eqref{mingrapheq} can be written in non-divergence form as
\begin{equation}
\frac{1}{W} \left( \sigma^{ij} - \frac{u^i u^j}{W^2} \right) u_{i;j} = \ang{\nb f, \nu},
\end{equation}
where $W= \sqrt{1+|\nabla u|^2}$, $(\sigma^{ij})$ stands for the inverse matrix of $(\sigma_{ij}),\ u^i = \sigma^{ij}u_j$,
with $u_j =\partial u/\partial x^j$,
and $u_{i;j}=u_{ij}-\Gamma^{k}_{ij}u_k$ denotes the second order covariant derivative of $u$.
We recall that an immersed hypersurface $\Sigma$ of a Riemannian manifold $(N,g)$ is called an $f$-minimal hypersurface if its (scalar)
mean curvature $H$ satisfies an equation
\[
H= \ang{\nb f,\nu}
\]
at every point of $\Sigma$. Here, too, $\nu$ is a unit normal vector field along $\Sigma$, $f$ is a smooth function on $N$, and $\bar{\n}f$ denotes its
gradient with respect to the Riemannian metric $g$. Hence the graph of a solution $u$ of \eqref{mingrapheq} is an $f$-minimal hypersurface in
$M\times\R$. Note that we define the mean curvature as the trace of the second fundamental form. Other examples of $f$-minimal hypersurfaces are
\begin{itemize}
\item[(a)] minimal hypersurfaces if $f$ is identically constant,
\item[(b)] self-shrinkers in $\R^{n+1}$ if $f(x)=\abs{x}^2/4$,
\item[(c)] minimal hypersurfaces of weighted manifolds $M_f=\bigl(M,g,e^{-f}d\vol_{M}\bigr)$, where $(M,g)$ is a complete Riemannian manifold with the
Riemannian volume element $d\vol_{M}$.
\end{itemize}
We refer to \cite{ColMin_CMH}, \cite{ColMin_Annals}, \cite{cheng2012eigenvalue}, \cite{cheng2013simons}, \cite{cheng2012stability}, \cite{impera_rimoldi},
and references therein for recent studies on self-shrinkers and $f$-minimal hypersurfaces.
Let us just point out a recent result relevant to our paper. Wang in \cite{MR2780753} investigated graphical self-shrinkers in $\R^n$ by studying the
equation \eqref{mingrapheq} in the whole $\R^n$ when $f(x)= |x|^2 /4$. He proved that any smooth solution to this equation has to be a hyperplane
improving an ealier result of Ecker and Huisken \cite{MR1025164}, where they made the extra assumption that the solution has polynomial growth. We will show that the situation is quite different when $\R^n$ is replaced by a Cartan-Hadamard manifold with strictly negative sectional curvatures and for more general $f$ satisfying some suitable assumptions. In particular, we impose that $\sup_{\bar{\Omega}\times \R}\abs{\bar{\n}f}<\infty$ which is not valid for $f(x)=\abs{x}^2/4$.
In our existence results we always assume that $f\in C^{2}(\bar{\Omega}\times\R)$ is of
the form
\begin{equation}\label{f-form}
f(x,t)=m(x)+r(t).
\end{equation}
Our first result is the following:
\begin{thm}
\label{existence}
Let $\Omega\subset M$ be a bounded domain with $C^{2,\alpha}$ boundary $\partial \Omega$. Suppose that $f\in C^{2}(\bar{\Omega}\times\R)$ satisfies \eqref{f-form}, with
\[
F =\sup_{\bar{\Omega}\times \R}\abs{\bar{\n}f}<\infty,\quad
\Ric_{\Omega} \geq -\dfrac{F^2}{n-1},\quad\text{and}\quad
H_{\partial \Omega} \ge F,
\]
where $Ric_{\Omega}$ stands for the Ricci curvature of $\Omega$ and
$H_{\partial \Omega}$ for the inward mean curvature of $\partial \Omega$. Then, for all $\varphi \in C^{2,\alpha}(\partial \Omega)$,
there exists a solution
$u\in C^{2,\alpha}(\bar{\Omega})$ to the equation \eqref{mingrapheq} with boundary values $\varphi$.
\end{thm}
The proof of Theorem \ref{existence} is based on the Leray-Schauder method
(see \cite[Theorem 13.8]{GilTru}), and hence requires a priori height and gradient (both
interior and boundary) estimates for solutions.
It is worth noting already at this point that we cannot ask for the uniqueness of a
solution if the function $f\colon M\times\R\to\R$ depends on the $t$-variable since
comparison principles fail to hold. Indeed, an easy computation shows that for the open
disk $B(0,2)\subset\R^2$ and $f\colon \R^2\times\R\to\R,\ f(x,t)=|(x,t)|^2/4$, both
the upper and lower hemispheres and the disk $B(0,2)$ itself are $f$-minimal hypersurfaces
with zero boundary values on the circle $\partial B(0,2)$.
Thanks to the interior gradient estimate Lemma~\ref{globestim} we can weaken the regularity assumption on the boundary value function.
\begin{thm}
\label{cont_existence}
Let $\Omega\subset M$ be a bounded domain with $C^{2,\alpha}$ boundary $\partial \Omega$.
Suppose that $f\in C^{2}(\bar{\Omega}\times\R)$ satisfies \eqref{f-form}, with
\[
F =\sup_{\bar{\Omega}\times \R}\abs{\bar{\n}f}<\infty,\quad
\Ric_{\Omega} \geq -\dfrac{F^2}{n-1},\quad\text{and}\quad
H_{\partial \Omega} \ge F.
\]
Then, for all $\varphi \in C(\partial \Omega)$,
there exists a solution
$u\in C^{2,\alpha}(\Omega)\cap C(\bar{\Omega})$ to the equation \eqref{mingrapheq} with boundary values $\varphi$.
\end{thm}
Let us point out that the assumption $H_{\partial \Omega}\geq F$ is necessary. Indeed, Serrin \cite{MR0282058} has proved that the constant mean
curvature equation
\[
\dv \frac{\nabla u}{W}=H_0
\]
is solvable on a bounded domain $\Omega\subset\R^n$ if and only if $H_{\partial \Omega} \geq |H_0|$;
see also \cite{MR2373228} for a related result.
Finally in Section~\ref{DPat8}, we consider the Dirichlet problem at infinity. Here we suppose that $M$ is a
Cartan-Hadamard manifold, i.e. a complete, simply connected Riemannian manifold with non-positive sectional curvature.
We denote by $\bar{M}$ the compactification of $M$ in the cone topology
(see \cite{eberlein1973visibility}) and by $\partial_\infty M$ the asymptotic boundary of $M$. The
Dirichlet problem at infinity consists in finding solutions to \eqref{mingrapheq} in the case where $\Omega =M$ and $\partial \Omega =\partial_\infty M$.
In order to formulate the assumptions on sectional curvatures of $M$ and on the function $f\colon M\times\R\to\R$, we first
denote by $\rho(\cdot)=d(o,\cdot)$ the (Riemannian) distance to a fixed point $o\in M$. Then we assume that sectional curvatures of $M$ satisfy
\begin{equation}\label{curv_assump_gen}
-(b\circ\rho)^2(x)\le K(P_x)\le -(a\circ\rho)^2(x)
\end{equation}
for all $x\in M$ and all $2$-dimensional subspaces $P_x\subset T_xM$, where $a$ and $b$ are smooth functions subject to conditions
\eqref{A1}-\eqref{A7}; see Section~\ref{DPat8}. Given a smooth function
$k\colon [0,\infty)\to [0,\infty)$, we denote by $f_k \colon [0,\infty) \to \R$
the smooth non-negative solution to the initial value problem
\begin{equation}\label{Jacobi_eq}
\left\{
\begin{aligned}
f_k(0)&=0, \\
f_k'(0)&=1, \\
f_k''&=k^2f_k.
\end{aligned}
\right.
\end{equation}
To state the main result on the solvability of the asymptotic Dirichlet problem requires a number of definitions. First of all we assume that there exists an auxiliary smooth function $a_0\colon [0,\infty)\to (0,\infty)$ such that
\[
\int_1^\infty \left( \int_r^\infty \frac{ds}{f_a^{n-1}(s)} \right) a_0(r)f_a^{n-1}(r) dr <\infty.
\]
Then we define $g\colon [0,\infty)\to [0,\infty)$ by
\begin{equation}\label{gV}
g(r)=\frac{1}{f_a^{n-1}(r)}\int_0^r a_0(t)f_a^{n-1}(t) dt.
\end{equation}
The function $g$ was introduced in \cite{mastrolia2015elliptic} where they studied some elliptic and parabolic equations with asymptotic Dirichlet
boundary conditions on Cartan-Hadamard manifolds.
In addition to \eqref{f-form}, we assume that the function $f\in C^{2}(\bar{\Omega}\times\R)$
satisfies
\begin{equation}\label{f1}
\sup_{\partial B(o,r)\times\R}\abs{\nb f} \le
\min\left\lbrace \frac{a_0(r)+(n-1)\frac{f^\prime_a (r)}{f_a(r)}g^3(r)}{\bigl(1+g^2(r)\bigr)^{3/2}}, (n-1)\frac{f^\prime_a(r)}{f_a(r)}\right\rbrace,
\end{equation}
for every $r>0$,
and
\begin{equation}
\label{f2}
\sup_{\partial B(o,r)\times\R}\abs{\nb f}=o\left(\frac{f^{\prime}_a (r)}{f_a(r)}r^{-\ve-1}\right)
\end{equation}
for some $\epsilon>0$ as $r\to\infty$.
The general solvability result for the asymptotic Dirichlet problem is the following.
\begin{thm}\label{ThmMain}
Let $M$ be a Cartan-Hadamard manifold of dimension $n\ge 2$.
Assume that
\begin{equation*}
-(b\circ\rho)^2(x)\le K(P_x)\le -(a\circ\rho)^2(x)
\end{equation*}
for all $x\in M$ and all $2$-dimensional subspaces $P_x\subset T_xM$ where $a$ and $b$
satisfy assumptions \eqref{A1}-\eqref{A7} and that the function $f\in C^2(M\times\R)$ on the right side of \eqref{mingrapheq} satisfies \eqref{f-form}, \eqref{f1}, and \eqref{f2}.
Then the asymptotic Dirichlet problem for the equation \eqref{mingrapheq} is
solvable for any boundary data $\varphi\in C\bigl(\partial_{\infty}M\bigr)$.
\end{thm}
As a special case of the above theorem, we have:
\begin{cor}\label{thm1}
Let $M$ be a Cartan-Hadamard manifold of dimension $n\ge 2$. Suppose that there are constants
$\phi>1,\ \varepsilon>0$, and $R_0>0$ such that
\begin{equation}\label{curv_ass_minim}
-\rho(x)^{2\left( \phi-2\right) -\varepsilon}\leq K(P_x)\leq-\dfrac
{\phi(\phi-1)}{\rho(x)^{2}},
\end{equation}
for all $2$-dimensional subspaces $P_x\subset T_{x}M$ and for all $x\in M$, with $\rho(x)\ge R_0$. Assume, furthermore, that
$f\in C^2(M\times\R)$ satisfies \eqref{f-form}, \eqref{f1}, and \eqref{f2}, with $f_a(t)=t$ for small $t\ge 0$ and $f_a(t)=c_1t^\phi +c_2t^{1-\phi}$ for $t\ge R_0$.
Then the asymptotic Dirichlet problem for equation \eqref{mingrapheq} is
solvable for any boundary data $\varphi\in C\bigl(\partial_{\infty}M\bigr)$.
\end{cor}
In another special case we assume that sectional curvatures are bounded from above by a negative constant $-k^2$.
\begin{cor}\label{HVkor2_RT}
Let $M$ be a Cartan-Hadamard manifold of dimension $n\ge 2$.
Assume that
\begin{equation}\label{curv_assump_k}
-\rho(x)^{-2-\varepsilon}e^{2k\rho(x)}\le K(P_x)\le -k^2
\end{equation}
for some constants $k>0$ and $\varepsilon>0$ and for all $2$-dimensional subspaces $P_x\subset T_x M$, with $\rho(x)\ge R_0$. Assume, furthermore, that
$f\in C^2(M\times\R)$ satisfies \eqref{f-form}, \eqref{f1}, and \eqref{f2}, with $f_a(t)=t$ for small $t\ge 0$ and $f_a(t)=c_1\sinh (kt) +c_2\cosh (kt)$ for $t\ge R_0$.
Then the asymptotic Dirichlet problem for the equation \eqref{mingrapheq} is
solvable for any boundary data $\varphi\in C\bigl(\partial_{\infty}M\bigr)$.
\end{cor}
We refer to \cite[Ex. 2.1, Cor. 3.22]{HoVa} and to \cite[Cor. 3.23]{HoVa} for the verification of the assumptions \eqref{A1}-\eqref{A7} for the curvature bounds \eqref{curv_ass_minim} and \eqref{curv_assump_k}, respectively. We point out that, thanks to Examples \ref{a_0_exam1} and \ref{a_0_exam2}, the assumption \eqref{f1} in the above corollaries
is weaker than \eqref{f2}
when $r\to \infty$.
Let us discuss where the assumptions \eqref{f1} and \eqref{f2} will be used in our paper. First of all, we prove Theorem~\ref{ThmMain} by extending the boundary value function $\varphi$ to $M$, exhausting $M$ by geodesic balls and solving the Dirichlet problem \eqref{mingrapheq} in each ball. In this step, the assumption
\[
\sup_{\partial B(o,r)\times\R}\abs{\nb f} \le (n-1)\frac{f^\prime_a(r)}{f_a(r)}
\]
is used. Secondly, the other assumption in \eqref{f1},
\[
\sup_{\partial B(o,r)\times\R}\abs{\nb f} \le
\frac{a_0(r)+(n-1)\frac{f^\prime_a (r)}{f_a(r)}g^3(r)}{\bigl(1+g^2(r)\bigr)^{3/2}},
\]
is used to prove that the sequence of solutions above is uniformly bounded, thus allowing us to extract a subsequence converging towards a global solution.
Finally, we apply \eqref{f2} to prove that this global solution has proper boundary values at infinity.
Furthermore, concerning \eqref{f2}, let us mention a result of Pigola, Rigoli, and Setti in \cite{PigolaRigoliSetti}. There they considered the equation
\[
\dv \frac{\nabla u}{\sqrt{1+|\nabla u|^2}} =h(x),
\]
for a function $h\in C^\infty (M)$. They proved that if $\max_M |u|<\infty$, $h$ has a constant sign,
and $M$ satisfies one of the following growth assumptions:
\begin{equation}
\label{(i)}
\vol \bigl(\partial B(o,r)\bigr)\leq C r^\alpha,\text{ for some }\alpha\geq 0
\end{equation}
or
\begin{equation}
\label{(ii)}
\vol \bigl(\partial B(o,r)\bigr)\leq C e^{\alpha r},\text{ for some }\alpha\geq 0,
\end{equation}
then necessarily we have
\[
\liminf_{\rho(x)\to\infty} \frac{|h(x)|}{\rho^{-2}(x) \bigl(\log \rho(x)\bigr)^{-1}}=0,
\]
and
\[
\liminf_{\rho(x)\to \infty} \frac{|h(x)|}{\rho^{-1}(x)\bigl(\log r(x)\bigr)^{-1}}=0,
\]
respectively. We notice that condition \eqref{(i)} (resp. \eqref{(ii)}) is implied by \eqref{curv_ass_minim} (resp.
\eqref{curv_assump_k}). On the other hand, assuming \eqref{curv_ass_minim} (resp. \eqref{curv_assump_k}), we notice (using Examples \ref{a_0_exam1} and \ref{a_0_exam2}) that \eqref{f2} reduces to $\sup_{\partial B(o,r)\times \R}|\bar{\nabla} f|=o(r^{-2-\varepsilon})$ (resp. $\sup_{\partial B(o,r)\times \R}|\bar{\nabla} f|=o(r^{-1-\varepsilon})$) when $r\rightarrow \infty$. Therefore, in these cases, \eqref{f2} is almost sharp.
The paper is organised as follows: in Section~\ref{sec_estim}, we prove a priori height and gradient estimates that are needed in
Section~\ref{sec_ex} where we apply the Leray-Schauder method and prove Theorem~\ref{existence} and \ref{cont_existence}.
Section~\ref{DPat8} is devoted to the asymptotic Dirichlet problem and proofs of Theorem~\ref{ThmMain} and Corollaries~\ref{thm1}
and \ref{HVkor2_RT}.
\section{Height and gradient estimates}\label{sec_estim}
In this section we adapt
methods from \cite{dajczer2008killing},
\cite{dajczer2012interior}, \cite{Korevaar}, and \cite{MR2351645}
to obtain a priori height and gradient estimates.
\subsection{Height estimate}\label{subsec-height}
We begin by giving an a priori height estimate for solutions of the equation \eqref{mingrapheq} in a bounded
open set $\Omega\subset M$ with a $C^2$-smooth boundary assuming the estimate \eqref{Fxtra} on the function $f$.
First we construct an upper barrier for a solution $u$ of \eqref{mingrapheq} of the form
\begin{equation*}
\psi(x) = \sup_{\p\Omega} \varphi + h \big( d(x) \big),
\end{equation*}
where $d=\dist(\cdot,\p\Omega)$ is the distance from $\p\Omega$ and $h$ is a real valued function that will be determined later.
Denote by $\Omega_0$ the open set of all points $x\in \Omega$ that can be joined to $\p\Omega$ by a \define{unique} minimizing geodesic. It was shown in
\cite{LiNirenberg} that in $\Omega_0$ the distance function $d$ has the same regularity as $\p\Omega$.
In particular, now $d\in C^2(\Omega_0)$ and straightforward computations give
\begin{equation*}
\psi_i = h' d_i \quad \text{and} \quad \psi_{i;j} = h'' d_i d_j + h' d_{i;j}.
\end{equation*}
Moreover, $|\n d|^2 = d^i d_i = 1$ and hence $d^i d_{i;j} =0$. We also have that
\begin{equation*}
\sigma^{ij} d_{i;j} = \Delta d = - H,
\end{equation*}
where $H=H(x)$ is the (inward) mean curvature of the level set $\{y\in\Omega_0\colon d(y)=d(x)\}$.
Given a solution $u\in C^{2}(\Omega)$ of \eqref{mingrapheq},
\begin{equation*}
Q[u] = \frac{1}{W} \left( \sigma^{ij} - \frac{u^i u^j}{W^2} \right) u_{i;j} - \ang{\nb f,\nu} = 0,
\end{equation*}
we define $b\colon\Omega\to\R$ by
\begin{equation}\label{b-def}
b(x)=\left\langle \bar{\nabla}f\bigl(x,u(x)\bigr),\nu(x)\right\rangle,
\end{equation}
where $\nu(x)$ is the downward pointing unit normal to the graph of $u$ at $\big(x,u(x)\big)$.
Next we define an operator
\begin{equation*}
\tilde{Q}[v] = \frac{1}{W} \left( \sigma^{ij} - \frac{v^i v^j}{W^2} \right) v_{i;j} - b,
\end{equation*}
where $W = \sqrt{1+|\nabla v|^2}$ and $b$ does not depend on $v$. The reason to define such an operator is that
it allows us to use the comparison principle whereas the operator $Q$ need not satisfy the required assumptions,
see e.g. \cite[Theorem 10.1]{GilTru}.
Then for a point $x\in \Omega_0$ we obtain
\begin{align} \label{operQ}
\tilde{Q}[\psi] + b &= \frac{1}{W} \left( \sigma^{ij} - \frac{(h')^2 d^i d^j}{W^2} \right)
\left( h'' d_i d_j + h' d_{i;j} \right) \nonumber \\
&= \frac{1}{W} \left( h'' + h' \Delta d - \frac{(h')^2 h''}{W^2} \right) \nonumber \\
&= \frac{1}{W} \left( \frac{h''}{W^2} - h' H(x) \right) \nonumber \\
&= \frac{h''}{W^3} - \frac{h'}{W} H(x),
\end{align}
where we used that $W^2 = 1+(h')^2$.
Next we impose an extra condition on the function $f\colon M\times\R\to\R$ by assuming that
\begin{equation}\label{Fxtra}
\sup_{s\in\R}\abs{\bar{\nabla}f(x,s)}\le H(x)
\end{equation}
for all $x\in\Omega_0$. Hence $\abs{b(x)}\le H(x)$ for all $x\in\Omega_0$. By choosing
\[
h= \frac{e^{AC}}{C}\big(1-e^{-Cd}\big),
\]
where $A = \diam(\Omega)$ and
\[
C>\sup_{\Omega_0\times\R}\abs{\bar{\nabla}f}
\]
is a constant, we obtain
\[
h' = e^{C(A-d)} \ge 1 \quad \text{and} \quad h'' = -Ch',
\]
and so
\begin{align*}
\tilde{Q}[\psi] &= -\frac{Ch'}{W^3} - \frac{h'H}{W} - b \\
&< - \abs{b}\left(\frac{h'}{W^3} + \frac{h'}{W}-1 \right) \\
&\le 0.
\end{align*}
Therefore we have
\begin{equation*} \begin{cases}
\tilde{Q}[\psi] < 0 = \tilde{Q}[u] = Q[u] \quad \text{in } \Omega_0 \\
\psi|\p\Omega \ge u|\p\Omega = \varphi|\p\Omega. \end{cases}
\end{equation*}
Next we observe that $\psi \ge u$ in $\bar{\Omega}$. Assume on the contrary that the continuous function $u-\psi$ attains its positive maximum
at an interior point $x_0 \in \Omega$. As in \cite[p. 795]{MR2351645} (see also \cite[pp. 239-240]{dajczer2008killing}),
we conclude that, in fact, $x_0$ is an interior point of $\Omega_0$ that leads to a contradiction
with the comparison principle \cite[Theorem 10.1]{GilTru} which states that $u-\psi$ can not attain its maximum in
the open set $\Omega_0$.
Similarly we deduce that $\psi^{-}$,
\[
\psi^{-}(x)=\inf_{\p\Omega} \varphi - h \big( d(x) \big),
\]
is a lower barrier for $u$, i.e. $\psi^{-}\le u$ in $\bar{\Omega}$.
These barriers imply the following height estimate for $u$.
\begin{lem} \label{heightestim}
Let $\Omega\subset M$ be a bounded open set with a $C^2$-smooth boundary and suppose that
\begin{equation}\label{Fxtra-again}
\sup_{s\in\R}\abs{\bar{\nabla}f(x,s)}\le H(x)
\end{equation}
in $\Omega_0$. Let $u \in C^2(\Omega)\cap C(\bar{\Omega})$ be a solution of $Q[u] = 0$ with $u|\p\Omega = \varphi$.
Then there exists a constant $C=C(\Omega)$
such that
\[
\sup_{\Omega} |u| \le C + \sup_{\p\Omega} |\varphi|.
\]
\end{lem}
\subsection{Boundary gradient estimate}
In this subsection we will obtain an a priori boundary gradient estimate for the Dirichlet problem \eqref{mingrapheq}. We assume that $\Omega\subset M$ is a bounded open set with a $C^2$-smooth boundary and that $\Omega_\ve$
is a sufficiently small tubular neighborhood of $\p\Omega$ so that the distance function
$d$ from $\p\Omega$ is $C^2$ in $\Omega_\ve\cap \bar{\Omega}$. Furthermore, we assume that the (inward) mean curvature $H=H(x)$ of the level set $\{y\in\bar{\Omega}_0\colon d(y)=d(x)\}$
satisfies
\begin{equation}\label{H-f}
H(x)\ge\sup_{s\in\R}\abs{\bar{\nabla}f(x,s)}:=F(x)
\end{equation}
for all $x\in\Omega_\ve\cap\bar{\Omega}$.
Next we extend the boundary function $\varphi$, which is assumed to be $C^2$-smooth, to $\Omega_\ve$ by setting $\varphi\big(\exp_{y}t \n d(y) \big) = \varphi(y)$, for $y\in\p\Omega$, where $\n d(y)$ is the unit \define{inward} normal to $\p\Omega$ at $y\in\p\Omega$. We will construct barriers of the form $w+\varphi$,
where $w= \psi\circ d$ and $\psi$ is a real function that will be determined later.
We denote
\begin{equation}\label{aijdef}
a^{ij}=a^{ij}(x,\n v) = \frac{1}{W} \left( \sigma^{ij} - \frac{v^iv^j}{W^2} \right),\quad
W=\sqrt{1+\abs{\n v}^2},
\end{equation}
and, given a solution $u\in C^2(\Omega)\cap C^1(\bar{\Omega})$ of \eqref{mingrapheq}, we define an operator
\[
\tilde{Q}[v]=a^{ij}(x,\n v)v_{i;j}-b,
\]
with $b$ as in \eqref{b-def}.
The matrix $a^{ij}(x,\n v)$ is positive definite with eigenvalues
\begin{equation}\label{eigval}
\lambda = \frac{1}{W^3} \quad \text{and} \quad \Lambda = \frac{1}{W}
\end{equation}
with multiplicities $1$ and $n-1$ corresponding respectively to the directions parallel and orthogonal
to $\n v$. Hence a simple estimate gives
\begin{equation} \label{Qestim}
\tilde{Q}[w+\varphi] = a^{ij}( w_{i;j} + \varphi_{i;j}) - b
\le a^{ij} w_{i;j} + \Lambda \norm{\varphi}_{C^2} - b,
\end{equation}
where $a^{ij}=a^{ij}(x, \n w + \n \varphi),\
\Lambda=(1+\abs{\n w +\n \varphi}^2)^{-1/2}$,
and $\norm{\varphi}_{C^2}$ denotes the $C^{2}(\Omega_{\ve})$-norm of $\varphi$.
Since in $\Omega_\ve\cap\bar{\Omega}$ we have $|\n d|^2 = d^i d_i = 1$, $d^i d_{i;j} =0$, and $\ang{\n d,\n \varphi}=0$, straightforward computations give that
\begin{align*}
\Delta w = \psi'' &+ \psi' \Delta d, \\
w^i w^j w_{i;j} &= (\psi')^2 \psi'', \\
w^i \varphi^j w_{i;j} &= \psi' \psi'' \ang{\n d, \n \varphi} = 0,
\end{align*}
and also
\[
\varphi^i \varphi^j w_{i;j} =\psi''\ang{\n \varphi,\n d}^{2} +\psi' \varphi^i\varphi^j d_{i;j} =\psi' \varphi^i\varphi^j d_{i;j}.
\]
With these, and noticing that now $W^2= 1+ (\psi')^2 + |\n \varphi|^2$, we obtain
\begin{align} \label{aijweq}
a^{ij} w_{i;j} &= \frac{\psi' \Delta d}{W} + \frac{\psi''(1+ |\n \varphi|^2)}{W^3} -
\frac{\psi' \varphi^i \varphi^j d_{i;j}}{W^3}.
\end{align}
Putting \eqref{Qestim} and \eqref{aijweq} together, we arrive at
\begin{equation}\label{1stestim}
\tilde{Q}[w+\varphi] \le \frac{\psi' \Delta d}{W} + \frac{\psi''(1+ |\n \varphi|^2)}{W^3} -
\frac{\psi' \varphi^i \varphi^j d_{i;j}}{W^3} + \Lambda \norm{\varphi}_{C^2} +F.
\end{equation}
Next we define
\[
\psi(t) = \frac{C\log(1+Kt)}{\log(1+K)},
\]
where the constants
\[
C \ge 2\big(\max_{\bar{\Omega}}|u| +\max_{\bar{\Omega}}|\varphi|\big),
\]
$K\ge (1-2\ve)\ve^{-2}$, and $\ve\in (0,1/2)$ will be chosen later. Then
\[
\psi(\ve)=\frac{C\log(1+K\ve)}{\log(1+K)}\ge C/2
\]
and we have
\begin{equation}\label{inbdry}
(w+\varphi)|\Gamma_\ve = \psi(\ve) + \varphi|\Gamma_\ve
\ge u|\Gamma_\ve
\end{equation}
on the ``inner boundary'' $\Gamma_{\ve}=\{x\in\Omega\colon d(x)=\ve\}$ of $\Omega_\ve$.
On the other hand,
\begin{equation}\label{outbdry}
(w+\varphi)|\p\Omega = u|\p\Omega.
\end{equation}
We claim that $\tilde{Q}[w+\varphi]\le 0$ in $\Omega_\ve\cap\Omega$ if $C,\ K$, and $\ve$
are properly chosen. All the computations below will be done in $\Omega_\ve\cap\Omega$ without further notice.
We first observe that
\[
\psi'(t) = \frac{CK}{(1+Kt)\log(1+K)} \quad \text{and} \quad
\psi''(t) = -\frac{\log(1+K)\psi'(d)^2}{C},
\]
and therefore we have
\begin{align}\label{2ndestim}
W \tilde{Q}[w+\varphi] &\le \big(W-\psi'\big)H
- \frac{\log(1+K)}{C}\left(\frac{\psi'}{W}\right)^2\big(1+|\nabla\varphi|^2\big)\nonumber\\
&\quad + \norm{\varphi}_{C^2}
+ |\nabla\varphi|^2 H
\end{align}
by \eqref{H-f}, \eqref{eigval}, and \eqref{1stestim}. We estimate
\[
\psi'\ge\frac{CK}{(1+K\ve)\log(1+K)}=\frac{C}{(\ve+1/K)\log(1+K)} = 1
\]
and consequently,
\begin{align*}
\frac{\psi'}{W} &\ge c_1=c_1\big(\max_{\bar{\Omega}}|\nabla\varphi|\big)>0\\
\noalign{and}
W-\psi'&\le c_2=c_2\big(\max_{\bar{\Omega}}|\nabla\varphi|\big)
\end{align*}
by choosing $C= (\ve+1/K)\log(1+K)$.
The claim $\tilde{Q}[w+\varphi]\le 0$ now follows from \eqref{2ndestim} since
\[
\frac{\log(1+K)}{C}=\frac{1}{\ve +1/K}\ge \frac{c_2 H+\norm{\varphi}_{C^2}+|\nabla\varphi|^2 H}{c_1^2\big(1+|\nabla\varphi|^2\big)}
\]
by choosing sufficiently small $\ve$ and large $K$ depending only on $\max_{\bar{\Omega}}|u|,\ \norm{\varphi}_{C^2}$, and $H_{\p\Omega}$.
Hence
\[
\tilde{Q}[w+\varphi]\le 0=\tilde{Q}[u],
\]
and therefore
$w+\varphi$ is an upper barrier in $\Omega_\ve\cap\Omega$. Similarly, $-w+\varphi$ is a lower barrier.
Together these barriers imply that
\[
\abs{\n u}\le \abs{\n w} +\abs{\n \varphi}= \psi'(0)+\abs{\n \varphi}=\frac{CK}{\log(1+K)}+\abs{\n \varphi}
\]
on $\p\Omega$.
We have proven the following boundary gradient estimate.
\begin{lem}\label{boundestim}
Let $\Omega\subset M$ be a bounded open set with a $C^2$-smooth boundary and suppose that
\begin{equation}\label{H-f-again}
\sup_{s\in\R}\abs{\bar{\nabla}f(x,s)}\le H(x)
\end{equation}
in some tubular neighborhood of $\p\Omega$. Let $u\in C^2(\Omega) \cap C^1(\bar{\Omega})$ be a solution to $Q[u] =0$ with $u|\p\Omega =\varphi\in C^{2}(\p\Omega)$. Then
$$
\max_{\p\Omega} |\n u| \le C,
$$
where $C$ is a constant depending only on $\sup_{\bar{\Omega}} |u|,\ H_{\p\Omega}$,
and $\norm{\varphi}_{C^2(\p\Omega)}$.
\end{lem}
\subsection{Interior gradient estimate}
In this subsection we will assume that $u$ is a $C^3$ function. The
elliptic regularity theory will guarantee that the estimate holds also for $C^{2,\alpha}$ solutions. We also assume that $f\colon M\times\R\to\R$ is of the form
\[
f(x,t)=m(x)+r(t).
\]
In particular, all ``space'' derivatives
\[
f_i=\frac{\p f}{\p x_i},\quad i=1,\ldots,\dim M,
\]
are independent of $t$; $f_{it}=f_{ti}=0$.
For an open set $\Omega\subset M$, we denote $i(\Omega)=\inf_{x\in\Omega}i(x)$, where
$i(x)$ is the injectivity radius at $x$. Thus $i(\Omega)>0$ if $\Omega\Subset M$ is
relatively compact. Furthermore, we denote by $R_{\Omega}$ the Riemannian curvature
tensor in $\Omega$.
\begin{lem}\label{globestim}
Let $u\in C^3 (\Omega)$ be a solution of \eqref{mingrapheq} with $u<m_u$ for some constant $m_u<\infty$.
\begin{enumerate}
\item[(a)] For every ball $B(o,r)\subset\Omega$, there exists a constant
\[
L=L\big(u(o),m_u,r,R_{\Omega},\norm{f}_{C^2(\Omega\times (-\infty,m_u))}\big)
\]
such that $\abs{\nabla u(o)}\le L$.
\item[(b)] If, furthermore, $u\in C^1(\bar{\Omega})$, we have a global gradient bound
\[
\abs{\nabla u(o)}\le L
\]
for every $o\in\bar{\Omega}$, with
\[
L=L\big(u(o),m_u,i(\Omega),\diam(\Omega),R_{\Omega},\norm{f}_{C^2(\Omega\times (-\infty,m_u))},\max_{\partial\Omega}\abs{\nabla u})\big)<\infty.
\]
\end{enumerate}
\end{lem}
\begin{proof}
We apply the method due to Korevaar and Simon \cite{Korevaar}; see also \cite{dajczer2012interior}.
Let $0<r\le\min\{i(\Omega),\diam(\Omega)\},\ o\in\Omega$, and let $\eta$ be a continuous non-negative function on $M$, vanishing outside $B(o,r)$ and smooth
whenever positive. The function $\eta$ will be specified later. Define
\[
h=\eta W
\]
and assume first that $h$ attains its maximum at an interior point $p\in B(o,r)\cap\Omega$.
The case $p\in B(o,r)\cap\partial\Omega$ and $u\in C^1(\bar{\Omega})$ will be commented at the end of the proof.
We will first prove an upper bound for $\abs{\nabla u(p)}$. Therefore we may assume that $\abs{\nabla u(p)}\ne 0$.
We choose normal coordinates at $p$ so that $\partial_1=\nabla u/\abs{\nabla u}$ at $p$.
All the computations below will be made at $p$ without further notice. Thus we have $\sigma_{ij}=\sigma^{ij}=\delta^{ij},\ u_1=u^1=\abs{\nabla u}$,
and $u_{j}=u^{j}=0$ for $j>1$. Furthermore,
\[
a^{ij}=\frac{1}{W}\left(\delta^{ij}-\frac{\abs{\nabla u}^2\delta^{1i}\delta^{1j}}{W^2}\right),
\]
and therefore $a^{11}=W^{-3},\ a^{ii}=W^{-1}$ for $i>1$, and $a^{ij}=0$ if $i\ne j$. At the maximum point $p$, we have
$h_i=0$ and $h_{i;i}\le 0$ for all $i$. Hence
\begin{equation}\label{etaWsym}
\eta_i W=-\eta W_i
\end{equation}
and
\[
a^{ij}h_{i;j}=a^{ii}h_{i;i}=a^{ii}\left(W\eta_{i;i}+2\eta_{i}W_{i}+\eta W_{i;i}\right)\le 0.
\]
With \eqref{etaWsym} we can write this as
\begin{equation}\label{Waeta}
Wa^{ii}\eta_{i;i}+\frac{\eta a^{ii}}{W}\left(WW_{i;i}-2(W_{i})^2\right)\le 0.
\end{equation}
We have
\[
W_i=\frac{u^k u_{k;i}}{W}=\frac{\abs{\nabla u}u_{1;i}}{W}
\]
and from \eqref{unitnormal} we see that the $ k^{\rm th} $ component of the unit normal is
\[
\nu^k =\frac{u^k}{W}=\frac{\abs{\nabla u}\delta^{k1}}{W}.
\]
To scrutinize the second order differential inequality \eqref{Waeta}, we first compute
\begin{align*}
a^{ii}W_{i;i}&=a^{ii}\bigl(W^{-1}u^k u_{k;i}\bigr)_{;i}\\
&=-\frac{a^{ii}\abs{\nabla u}u_{1;i}W_i}{W^2}+
\frac{a^{ii}u^{k}_{;i}u_{k;i}}{W}
+\frac{a^{ii}\abs{\nabla u}u_{1;ii}}{W}\\
&=-\frac{a^{ii}\abs{\nabla u}^2(u_{1;i})^2}{W^3}+
\frac{a^{ii}u^{k}_{;i}u_{k;i}}{W}
+\frac{a^{ii}\abs{\nabla u}u_{1;ii}}{W}\\
&=\frac{a^{ii}(u_{1;i})^2}{W^3}+\frac{a^{ii}\sum_{k\ne 1}(u_{k;i})^2}{W}+
\frac{a^{ii}\abs{\nabla u}u_{1;ii}}{W}.
\end{align*}
Hence
\begin{equation}\label{W2deriv}
Wa^{ii}W_{i;i}=A+a^{ii}\abs{\nabla u}u_{1;ii},
\end{equation}
where
\[
A=a^{ii}(u_{1;i})^2W^{-2}+a^{ii}\sum_{k\ne 1}(u_{k;i})^2\ge 0.
\]
Using the Ricci identities for the Hessian of $u$ we get
\[
u_{k;ij}=u_{i;kj}=u_{i;jk}+R^{\ell}_{kji}u_{\ell},
\]
where $R$ is the curvature tensor in $M$. This yields
\begin{equation}\label{ricciaplic}
\abs{\nabla u}a^{ii}u_{1;ii}=\abs{\nabla u}a^{ii}u_{i;i1}+\abs{\nabla u}^2 a^{ii}R^{1}_{1ii}.
\end{equation}
To compute $\abs{\nabla u}a^{ii}u_{i;i1}$, we first observe that
\[
Wa^{ij}u_{i;j}=
Wa^{ii}u_{i;i}=\langle\bar{\nabla}f,(\nabla u,-1)\rangle=f_{i}u^{i}-f_t.
\]
Since
\begin{align*}
\nu^1\bigl(Wa^{ij}\bigr)_{;1}u_{i;j}&=\nu^1\bigl(\sigma^{ij}-u^iu^j W^{-2}\bigr)_{;1}u_{i;j}\\
&=-\frac{\abs{\nabla u}}{W}\left(\frac{2u^i u^{j}_{;1}}{W^2}-\frac{2u^i u^j W_1}{W^3}\right)u_{i;j}\\
&=-\frac{2\abs{\nabla u}^2}{W^3}u^{j}_{;1}u_{1;j}+\frac{2\abs{\nabla u}^4(u_{1;1})^2}{W^5}\\
&=-\frac{2\abs{\nabla u}^2}{W^3}\left(\sum_{i}(u_{1;i})^2-\frac{\abs{\nabla u}^2}{W^2}(u_{1;1})^2\right)\\
&=-\frac{2\abs{\nabla u}^2a^{ii}(u_{1;i})^2}{W^2}\\
&=-2a^{ii}(W_i)^2,
\end{align*}
we obtain
\begin{align}\label{WaW2deriv}
\abs{\nabla u}a^{ii}u_{i;i1} &=\abs{\nabla u}a^{ij}u_{i;j1}
=\nu^{1}Wa^{ij}u_{i;j1}\nonumber\\
&=\nu^1\bigl(Wa^{ij}u_{i;j}\bigr)_{;1}-\nu^1\bigl(Wa^{ij}\bigr)_{;1}u_{i;j}\nonumber\\
&=\nu^1\bigl(f_{i}u^{i}-f_t\bigr)_{;1}+2a^{ii}(W_i)^2\nonumber\\
&=\nu^1\bigl(f_i u^{i}_{;1}+(f_\ell)_{;1}u^{\ell}-(f_t)_{;1}\bigr)+2a^{ii}(W_i)^2\\
&=\frac{\abs{\nabla u}}{W}\bigl(f_i u^{i}_{;1}+(f_1)_{;1}u^1-f_{tt}u^1\bigr)+2a^{ii}(W_i)^2\nonumber\\
&=W_{i}f^{i}+\frac{f_{11}\abs{\nabla u}^2}{W}-\frac{f_{tt}\abs{\nabla u}^2}{W}+2a^{ii}(W_i)^2,\nonumber
\end{align}
where we have denoted $(f_j)_{;1}=\bigl(x\mapsto f_j(x,u(x))\bigr)_{;1}$ and used the assumption $f_{it}=f_{ti}=0$.
Putting together \eqref{etaWsym}, \eqref{W2deriv}, \eqref{ricciaplic}, and \eqref{WaW2deriv} we can estimate the inequality \eqref{Waeta} as
\begin{align}\label{WaetaEstim}
0&\ge Wa^{ii}\eta_{i;i}+\frac{\eta a^{ii}}{W}\left(WW_{i;i}-2(W_{i})^2\right)\nonumber\\
&=Wa^{ii}\eta_{i;i} + \frac{\eta}{W}\left( A+\abs{\nabla u}a^{ii}u_{1;ii}-\abs{\nabla u}a^{ii}u_{i;i1} +W_if^i
+\frac{\abs{\nabla u}^2 \left(f_{11}-f_{tt}\right)}{W}\right)\nonumber\\
&=Wa^{ii}\eta_{i;i}+\eta\left(\frac{A}{W}+\frac{\abs{\nabla u}^2 a^{ii}R^{1}_{1ii}}{W}
+\frac{\abs{\nabla u}^2(f_{11}-f_{tt})}{W^2}\right)-f^{i}\eta_{i}\\
&\ge Wa^{ii}\eta_{i;i}-f^{i}\eta_{i} -N\eta,\nonumber
\end{align}
where $N$ is a positive constant depending only on the curvature tensor in $\Omega$ and
the $C^2$-norm of $f$ in the cylinder $\Omega\times (-\infty,m_u)$. Note that
$A\ge 0$, $a^{11}=W^{-3}$, and $a^{ii}=W^{-1}$ for $i\ne 1$.
Now we are ready to choose the function $\eta$ as
\[
\eta(x)=g\bigl(\phi(x)\bigr),
\]
where
\[
g(t)=e^{C_1 t}-1
\]
with a positive constant $C_1$ to be specified later and
\[
\phi(x)=\left(1-r^{-2}d^2 (x)+ C\bigl(u(x)-m_u\bigr)\right)^{+}.
\]
Here $d(x)=d(x,o)$ is the geodesic distance to $o$ and
\[
C=\frac{-1}{2\bigl(u(o)-m_u\bigr)}>0.
\]
It follows that $\eta$ fulfils the requirements and, moreover, $\eta(o)=e^{C_1/2}-1>0$.
We have
\begin{equation}\label{etai}
\eta_i=\left(-r^{-2}(d^2)_i +Cu_i\right)g'
\end{equation}
and
\begin{equation}\label{etaij}
\eta_{i;j}=\left(-r^{-2}(d^2)_{i;j} +Cu_{i;j}\right)g'
+\left(-r^{-2}(d^2)_i +Cu_i\right)\left(-r^{-2}(d^2)_j +Cu_j\right)g''.
\end{equation}
A straightforward computation gives the estimate
\begin{align}\label{Waprodestim}
Wa^{ii}&\bigl(r^{-2}(d^2)_i -Cu_i\bigr)^2
= Wa^{ii}\bigl(r^{-4}(d^2)^2_i -2Cr^{-2}(d^2)_i u_i +C^2(u_i)^2\bigr) \nonumber\\
&= r^{-4}\abs{\nabla d^2}^2 - 2Cr^{-2}\langle\nabla d^2,\nabla u\rangle
+C^2\abs{\nabla u}^2 \nonumber\\
&\qquad -\frac{\langle\nabla d^2,\nabla u\rangle^2}{r^4 W^2}
+\frac{2C\abs{\nabla u}^2\langle\nabla d^2,\nabla u\rangle}{r^2 W^2}
-\frac{C^2\abs{\nabla u}^4}{W^2} \nonumber\\
&=\frac{C^2\abs{\nabla u}^2}{W^2}-\frac{2C\langle\nabla d^2,\nabla u\rangle}{r^2 W^2}
+\frac{1}{r^4}\left(\abs{\nabla d^2}^2-\frac{\langle\nabla d^2,\nabla u\rangle^2}{W^2}\right) \nonumber\\
&\ge \frac{C^2\abs{\nabla u}^2}{W^2}-\frac{2C\langle\nabla d^2,\nabla u\rangle}{r^2 W^2}.
\end{align}
Next we observe that
\begin{align}\label{waiiest}
Wa^{ii}\bigl(-r^{-2}(d^2)_{i;i}+Cu_{i;i}\bigr)&=-r^{-2}Wa^{ii}(d^2)_{i;i}+CWa^{ii}u_{i;i}\\
&=-r^{-2}\Delta d^2+\frac{\abs{\nabla u}^2}{r^2 W^2 }(d^2)_{1;1}+CW\langle\bar{\nabla}f,\nu\rangle\nonumber\\
&=-r^{-2}\Delta d^2+\frac{\abs{\nabla u}^2}{r^2 W^2 }\Hess d^2(\partial_1,\partial_1)
+CW\langle\bar{\nabla}f,\nu\rangle.\nonumber
\end{align}
Putting together \eqref{etai}, \eqref{etaij}, \eqref{Waprodestim}, and \eqref{waiiest} we
obtain
\begin{align*}
Wa^{ii}\eta_{i;i}&\ge g'\left(-r^{-2}\Delta d^2 +\frac{\abs{\nabla u}^2}{r^2W^2}\Hess d^2 (\partial_1,\partial_1)
+CW\langle\bar{\nabla}f,\nu\rangle\right)\\
&\qquad +g''\left(\frac{C^2\abs{\nabla u}^2}{W^2}-\frac{2C}{r^2W^2}\langle\nabla u,\nabla d^2\rangle\right).
\end{align*}
Hence, by \eqref{WaetaEstim}, we have
\begin{equation}\label{g2nd}
g''\left(\frac{C^2\abs{\nabla u}^2}{W^2}-\frac{2C}{r^2W^2}\langle\nabla u,\nabla d^2\rangle\right)
+ g'P-Ng\le 0,
\end{equation}
where
\[
P=\frac{\abs{\nabla u}^2}{r^2W^2}\Hess d^2 (\partial_1,\partial_1)- \frac{\Delta d^2}{r^2}+
\frac{f^i(d^2)_i}{r^2} - Cf_t.
\]
It is easy to see that
\[
\abs{P}\le
\frac{\abs{\Hess d^2(\partial_1,\partial_1)}+\abs{\Delta d^2}}{r^2}+\frac{2d\abs{f^i d_i}}{r^2}+C\abs{f_t}\le C_0,
\]
with a constant $C_0=C_0(u(o)-m_u,r,R_{\Omega},\norm{f}_{C^1})$.
In order to obtain an upper bound for $\abs{\nabla u(p)}$, we suppose that
\[
\abs{\nabla u(p)}\ge\frac{16\bigl(m_u-u(o)\bigr)}{r}
\]
and derive a contradiction.
Since $\abs{\nabla d^2(p)}\le 2r$, we see that
\[
\abs{\nabla u(p)}\ge\frac{4\abs{\nabla d^2(p)}}{Cr^2}
\]
and hence we have
\[
\abs{\nabla u}^2-\frac{2}{Cr^2}\langle\nabla u,\nabla d^2\rangle\ge\frac{1}{2}\abs{\nabla u}^2
\]
at $p$. Therefore there exists a constant $D$ depending only on $m_u-u(o)$ and $r$ such that
\[
\frac{C^2}{W^2}\left(\abs{\nabla u}^2-\frac{2}{Cr^2}\langle\nabla u,\nabla d^2\rangle\right)\ge D>0.
\]
But now, taking $C_1=C_1(C_0,D,N)$ large enough, we obtain
\[
Dg''(\phi(p))-C_0 g'(\phi(p))-Ng(\phi(p))=(DC_1^2-C_1C_0-N)e^{C_1\phi(p)}+N>0
\]
which is a contradiction with \eqref{g2nd}. Hence we have
\[
\abs{\nabla u(p)}<\frac{16\bigl(m_u-u(o)\bigr)}{r}
\]
which implies
\[
W(p)\le C_2=1+\frac{16\bigl(m_u-u(o)\bigr)}{r}.
\]
Since $p$ is a maximum point of $h=\eta W$, we have
\[
\left(e^{C_1/2}-1\right)W(o)=\eta(o)W(o)\le\eta(p)W(p)\le C_2\left(e^{C_1}-1\right).
\]
This proves the case (a).
For the case (b), we assume, in addition, that $u\in C^1(\bar{\Omega})$ and we fix $r=\min\{i(\Omega),\diam(\Omega)\}>0$.
Let $o\in\bar{\Omega}$ and $h=\eta W$ be as above with the same constant $C_1$.
If a maximum point $p$ of $h$ is an interior point of $\Omega$, the proof for the case (a) applies and we have a desired upper bound
for $\abs{\nabla u(o)}$. On the other hand, if $p\in\partial\Omega$ we have an upper bound
\[
\abs{\nabla u(p)}\le \max_{\partial\Omega}\abs{\nabla u}
\]
and again we are done.
\end{proof}
\section{Existence of $f$-minimal graphs}\label{sec_ex}
In this section we will prove Theorem~\ref{existence} and \ref{cont_existence}. Throughout this section we assume that
$\Omega\subset M$ is a bounded open set with $C^{2,\alpha}$ boundary $\p\Omega$. As in
Subsection~\ref{subsec-height} we denote by $\Omega_0$ the open
set of all those points of $\Omega$ that can be joined to $\p\Omega$ by a unique minimizing geodesic. We start with the following lemma from
\cite[Lemma 4.2]{MR2351645}; see also \cite[Lemma 5]{dajczer2008killing}. Since our definition of the mean curvature differs by a multiple constant from
the one used in \cite{MR2351645} and \cite{dajczer2008killing}, we sketch the proof.
\begin{lem} \label{meancurvlemma}
Let $F=\sup\{\abs{\bar{\n}f(x,s)}\colon (x,s)\in\bar{\Omega}\times\R\}<\infty$ and suppose that
$\Ric_{\Omega} \geq -F^2/(n-1)$ and $H_{\partial \Omega} \ge F$.
Then for all $x_0\in \Omega_0$ the inward mean curvature $H(x_0)$ of the level set $\{x\in\Omega \colon d(x) = d(x_0)\}$
passing through $x_0$ has a lower bound $H(x_0)\geq F$.
\end{lem}
\begin{proof}
Denote by $H(t)$ the inward mean curvature of the level set $\Gamma_t=\{x\in\Omega \colon d(x) = t\}$ at the point
which lies on the unit speed minimizing geodesic $\gamma$ joining $\gamma(0) \in \p\Omega$ to $x_0$. Denote by $N=\dot{\gamma}_{t}$ the
inward unit normal to $\Gamma_t$ and by $S_t$ the shape operator, $S_t(X) = -\nabla_{X} N$, of the level set $\Gamma_t$. As in \cite{dajczer2008killing}
we obtain the Riccati equation
\[
S^{\prime}_{t} = S^{2}_{t} + R_t,
\]
where $R_t=R(\cdot,\dot{\gamma}_t)\dot{\gamma}_t$.
Trace and derivative commute, but because of the term $S^{2}_{t}$, we need to substitute $s = \tr S_t/(n-1)$
in order to get similar differential equation for the traces. Hence we have
\[
s' = s^2 + r,
\]
where $r$ satisfies $r \ge \Ric(\dot{\gamma}_t,\dot{\gamma}_t)/(n-1)$. In other words,
\[
\frac{\tr S^{\prime}_{t}}{n-1} \ge \left(\frac{\tr S_t}{n-1} \right)^2 + \frac{1}{n-1} \Ric(\dot{\gamma}_t,\dot{\gamma}_t).
\]
Since $H(t) = \tr S_t$, we obtain the estimate
\[
\frac{H'(t)}{n-1} \ge \left(\frac{H(t)}{n-1}\right)^2 +
\frac{1}{n-1} \Ric\big(\dot{\gamma}_t,\dot{\gamma}_t\big) \ge
\frac{H^2(t)}{(n-1)^2} - \frac{F^2}{(n-1)^2}.
\]
On the boundary we have $H(0) = H_{\p\Omega} \ge F$
which implies that $H'(t) \ge 0$ and hence the claim follows.
\end{proof}
\begin{proof}[Proof of Theorem \ref{existence}]
In order to prove Theorem~\ref{existence} we assume that the given boundary value function is extended to a function
$\varphi\in C^{2,\alpha}(\bar{\Omega})$ and we consider a family of Dirichlet problems
\begin{equation} \label{eqfamily}
\begin{cases}
\dv \dfrac{\nabla u}{\sqrt{1+|\nabla u|^2}} - \tau\ang{\nb f,\nu}=0 \quad \text{in } \Omega, \\
u = \tau\varphi \quad \text{in } \p\Omega,\ 0\le \tau\le 1.
\end{cases}
\end{equation}
By Lemma~\ref{meancurvlemma},
\[
H(x)\ge F\ge \sup_{\bar{\Omega}\times\R}\abs{\bar{\n}(\tau f)}
\]
for all $x\in\Omega_0$ and for all $\tau\in [0,1]$.
Hence \emph{if} $u\in C^{2,\alpha}(\bar{\Omega})$ is a solution of \eqref{eqfamily} for some $\tau\in [0,1]$, it follows from
Lemmata~\ref{heightestim}, \ref{boundestim}, and \ref{globestim} that
\[
\norm{u}_{C^1(\bar{\Omega})}\le C
\]
with a constant $C$ that is independent of $\tau$. The Leray-Schauder method \cite[Theorem 13.8]{GilTru} then yields a solution to the Dirichlet
problem \eqref{eqfamily} for all $\tau\in [0,1]$. In particular, with $\tau=1$ we obtain a solution to the original Dirichlet problem.
\end{proof}
\begin{proof}[Proof of Theorem \ref{cont_existence}]
Let $\varphi \in C(\p\Omega)$ and let $\varphi^{\pm}_k\in C^{2,\alpha}(\p\Omega)$ be two monotonic
sequence converging uniformly on $\p\Omega$ to $\varphi$ from above and from below, respectively. Denote
$$
F^+ = \sup_{\bar{\Omega}\times\R} |\nb f| \quad \text{and} \quad F^- = -F^+.
$$
By Theorem~\ref{existence} there are functions $u^{\pm}_{k},v^{\pm}_{k}\in C^{2,\alpha}(\bar{\Omega})$ such that
$u^{\pm}_{k}|\p\Omega=v^{\pm}_{k}|\p\Omega=\varphi_k^\pm$ and
\begin{align*}
&a^{ij}(x,\n u^{\pm}_{k})(u^{\pm}_{k})_{i;j}-\langle\nb f,\nu^{\pm}_{k}\rangle=0\\
&a^{ij}(x,\n v^{\pm}_{k})(v^{\pm}_{k})_{i;j}+F^{\pm}=0
\end{align*}
in $\Omega$, where $a^{ij}$ is as in \eqref{aijdef} and $\nu^{\pm}_{k}$ is the downward unit normal to the graph of $u^{\pm}_{k}$.
Since
\begin{align*}
a^{ij}(x,\n v^{+}_{k})(v^{+}_{k})_{i;j} + F^{-} & \le a^{ij}(x,\n v^{+}_{k})(v^{+}_{k})_{i;j} + F^{+} = 0 \\
&=
a^{ij}(x,\n v^{-}_{\ell})(v^{-}_{\ell})_{i;j} + F^{-}
\end{align*}
and $v^{+}_{k}|\p\Omega \ge v^{-}_{\ell}|\p\Omega$ for all $k,\ell$, we obtain from the comparison principle \cite[Theorem 10.1]{GilTru} applied to the
operator $a^{ij}+F^{-}$ that
\[
v^{-}_{\ell}\le v^{+}_{k}\quad\text{in }\bar{\Omega}.
\]
On the other hand, since $\varphi^{+}_{k+1}\le \varphi^{+}_{k}$ and $\varphi^{-}_{\ell}\le\varphi^{-}_{\ell+1}$ on $\p\Omega$, we have again by the comparison principle that
\begin{equation}\label{v_kOrder}
v_1^- \le \cdots \le v_{\ell}^- \le v_{\ell+1}^- \cdots\le v_{k+1}^+ \le v_{k}^+ \cdots \le v_1^+.
\end{equation}
Similarly, since
\begin{align*}
a^{ij}(x,\n v^{+}_{k})(v^{+}_{k})_{i;j} - \langle \nb f,\nu^{+}_k\rangle &\le
a^{ij}(x,\n v^{+}_{k})(v^{+}_{k})_{i;j} - F^{-} = 0 \\
&=
a^{ij}(x,\n u^{+}_{k})(u^{+}_{k})_{i;j} - \langle \nb f,\nu^{+}_k\rangle
\end{align*}
and $v^{+}_k|\p\Omega =u^{+}_k|\p\Omega$, we get
\[
u^{+}_{k}\le v^{+}_{k}\quad\text{in }\bar{\Omega}.
\]
Similar reasoning implies that $v^{-}_k\le u^{-}_k$, and therefore
\begin{equation}\label{v_ku_kOrder}
v^{-}_{k}\le u^{\pm}_{k}\le v^{+}_k\quad\text{in }\bar{\Omega}.
\end{equation}
Hence the sequences $u_k^\pm,v_k^\pm$ have uniformly bounded $C^0$ norms and
the local interior gradient estimate (Lemma~\ref{globestim}) together with \cite[Corollary 6.3]{GilTru} imply that
the sequences $u_k^\pm,v_k^\pm$ have equicontinuous $C^{2,\alpha}$
norms on compact subsets $K\subset \Omega$.
Taking an exhaustion of $\Omega$ by
compact sets we obtain, with a diagonal argument, that $u_k^\pm$ and $v_k^\pm$
contain subsequences that converge uniformly in compact subsets to functions
$u,v^\pm \in C^2(\Omega)$ with respect to the $C^2$ norm. Moreover, we have
\[
a^{ij}(x,\nabla u)u_{i;j} - \ang{\nb f,\nu}=0
\quad \text{and} \quad a^{ij}(x,\nabla v^{\pm})v^\pm_{i;j} + F^\pm =0.
\]
Since $v_k^\pm |\p\Omega = \varphi_k^\pm$ convergences to $\varphi$, \eqref{v_kOrder}
implies that $v^\pm$ extends continuously to the boundary $\p\Omega$ and $v^\pm |\p\Omega=\varphi$.
In turn, this and \eqref{v_ku_kOrder} give that $u$ extends continuously to $\p\Omega$ with
$u|\p\Omega = \varphi$. Furthermore, because $f\in C^2 (M\times\R)$, it follows that $u\in C^{2,\alpha}(\Omega)
\cap C(\bar{\Omega})$ (\cite[Theorem 6.17]{GilTru}).
\end{proof}
\section{Dirichlet problem at infinity}\label{DPat8}
In this section we assume that $M$ is a Cartan-Hadamard manifold of dimension $n\ge 2$, $\partial_{\infty}M$ is the asymptotic boundary of $M$,
and $\bar{M}=M\cup\partial_{\infty}M$ the compactification of $M$ in the cone topology.
Recall that the asymptotic boundary is defined as the set of all equivalence classes of unit speed
geodesic rays in $M$; two such rays $\gamma_{1}$ and $\gamma_{2}$ are
equivalent if $\sup_{t\ge0}d\bigl(\gamma_{1}(t),\gamma_{2}(t)\bigr)< \infty$. The equivalence class of $\gamma$ is denoted by $\gamma(\infty)$.
For each $x\in M$ and $y\in\bar{M}\setminus\{x\}$ there exists a unique unit
speed geodesic $\gamma^{x,y}\colon\mathbb{R}\to M$ such that $\gamma^{x,y}_{0}=x$
and $\gamma^{x,y}_{t}=y$ for some $t\in(0,\infty]$. If $v\in
T_{x}M\setminus\{0\}$, $\alpha>0$, and $r>0$, we define a cone
\[
C(v,\alpha)=\{y\in\bar M\setminus\{x\}:\sphericalangle(v,\dot\gamma^{x,y}_{0})<\alpha\}
\]
and a truncated cone
\[
T(v,\alpha,r)=C(v,\alpha)\setminus\bar B(x,r),
\]
where $\sphericalangle(v,\dot\gamma^{x,y}_{0})$ is the angle between vectors
$v$ and $\dot\gamma^{x,y}_{0}$ in $T_{x} M$. All cones and open balls in $M$
form a basis for the cone topology on $\bar M$.
Throughout this section, we assume that the sectional curvatures of $M$ are bounded from below and above
by
\begin{equation}
\label{curv-bound-gen}
-(b\circ \rho)^2(x) \le K(P_x) \le -(a\circ \rho)^2 (x)
\end{equation}
for all $x\in M$,
where $\rho(x) = d(o,x)$ is the distance to a fixed point $o\in M$ and $P_x$ is any 2-dimensional subspace of $T_xM$. The functions
$a,b\colon [0,\infty) \to [0,\infty)$ are assumed to be smooth such that $a(t)=0$ and $b(t)$ is constant for $t\in [0,T_0]$ for some $T_0>0$,
and $b\ge a$. Furthermore, we assume that $b$ is monotonic and that there exist positive constants
$T_1, C_1, C_2, C_3$, and $Q\in (0,1)$ such that
\begin{align}
\tag{A1}\label{A1}
a(t)\begin{cases}=C_1t^{-1}&\text{if $b$ is decreasing,}\\
\ge C_1t^{-1}&\text{if $b$ is increasing}\\ \end{cases}
\end{align}
for all $t\ge T_1$ and
\begin{align}
\tag{A2}\label{A2}
a(t)&\le C_2, \\
\tag{A3}\label{A3}
b(t+1)&\le C_2b(t), \\
\tag{A4}\label{A4}
b(t/2)&\le C_2b(t), \\
\tag{A5}\label{A5}
b(t)&\ge C_3(1+t)^{-Q}
\end{align}
for all $t\ge 0$.
In addition, we assume that
\begin{align}
\tag{A6}\label{A6}
&\lim_{t\to\infty}\frac{b'(t)}{b(t)^2}=0
\end{align}
and that there exists a constant $C_4>0$ such that
\begin{align}
\tag{A7}\label{A7}
&\lim_{t\to\infty}\frac{t^{1+C_4}b(t)}{f_a'(t)}=0.
\end{align}
It can be checked from \cite{HoVa} or from \cite{casterasholopainenripoll1} that the curvature
bounds in Corollary \ref{thm1} and Corollary \ref{HVkor2_RT} satisfy the assumptions \eqref{A1}-\eqref{A7}.
\subsection{Construction of a barrier}\label{subsec_barrier_constr}
The curvature bounds \eqref{curv-bound-gen} are needed to control the first two derivatives of the ``barrier'' functions that we will construct in this subsection. Recall from the introduction that for a smooth function
$k\colon [0,\infty)\to [0,\infty)$, we denote by $f_k \colon [0,\infty) \to \R$
the smooth non-negative solution to the initial value problem
\begin{equation*}
\left\{
\begin{aligned}
f_k(0)&=0, \\
f_k'(0)&=1, \\
f_k''&=k^2f_k.
\end{aligned}
\right.
\end{equation*}
Following \cite{HoVa}, we construct a barrier function for each boundary point $x_0\in\partial_{\infty}M$.
Towards this end let $v_{0}=\dot\gamma^{o,x_{0}}_{0}$ be the initial (unit) vector of
the geodesic ray $\gamma^{o,x_{0}}$ from a fixed point $o\in M$ and define a
function $h:\partial_{\infty}M\to\mathbb{R}$,
\begin{equation}
\label{eq:hoodef}h(x)=\min\bigl(1,L\sphericalangle(v_{0},\dot\gamma^{o,x}_{0})\bigr),
\end{equation}
where $L\in(8/\pi,\infty)$ is a constant. Then we define a crude extension
$\tilde h\in C(\bar{M})$, with $\tilde h|\partial_{\infty}M=h$, by setting
\begin{equation}
\label{eq:hoodeftilde}\tilde h(x)=\min\Bigl(1,\max\bigl(2-2\rho
(x),L\sphericalangle(v_{0},\dot\gamma^{o,x}_{0})\bigr)\Bigr).
\end{equation}
Finally, we smooth out $\tilde{h}$ to get an extension
$h\in C^{\infty}(M)\cap C(\bar{M})$ with controlled first and second order
derivatives. For that purpose, we fix $\chi\in C^{\infty}(\R)$ such that
$0\le\chi\le 1$, $\spt\chi\subset[-2,2]$, and
$\chi\vert[-1,1]\equiv1$. Then for any function $\varphi\in C(M)$ we define
functions $F_{\varphi}\colon M\times M\to\mathbb{R},\ {\mathcal{R}}
(\varphi)\colon M\to M$, and ${\mathcal{P}}(\varphi)\colon M\to\mathbb{R}$ by
\begin{align*}
F_{\varphi}(x,y) & =\chi\bigl(b(\rho(y))d(x,y)\bigr)\varphi(y),\\
{\mathcal{R}}(\varphi)(x) & =\int_{M}F_{\varphi}(x,y) dm(y),\ \text{ and}\\
{\mathcal{P}}(\varphi) & =\frac{{\mathcal{R}}(\varphi)}{{\mathcal{R}}(1)},
\end{align*}
where
\[
{\mathcal{R}}(1)=\int_{M}\chi\bigl(b(\rho(y))d(x,y)\bigr)dm(y)>0.
\]
If $\varphi\in C(\bar M)$, we extend ${\mathcal{P}}(\varphi)\colon
M\to\mathbb{R}$ to a function $\bar{M}\to\mathbb{R}$ by setting
${\mathcal{P}}(\varphi)(x)=\varphi(x)$ whenever $x\in M(\infty)$. Then the extended
function ${\mathcal{P}}(\varphi)$ is $C^{\infty}$-smooth in $M$ and continuous
in $\bar{M}$; see \cite[Lemma 3.13]{HoVa}. In particular, applying
${\mathcal{P}}$ to the function $\tilde{h}$ yields an appropriate smooth
extension
\begin{equation}
\label{extend_h}
h:={\mathcal{P}}(\tilde{h})
\end{equation}
of the original function $h\in C\bigl(\partial_{\infty}M\bigr)$ that was defined in \eqref{eq:hoodef}.
We denote
\[
\Omega=C(v_{0},1/L)\cap M \ \text{ and }\ \ell\Omega=C(v_{0},\ell/L)\cap M
\]
for $\ell>0$ and collect various constants and functions together to a data
\[
C=(a,b,T_{1},C_{1},C_{2},C_{3},C_{4},Q,n,L).
\]
Furthermore, we denote by $\|\Hess_{x} u\|$ the norm of the Hessian of a
smooth function $u$ at $x$, that is
\[
\|\Hess_{x} u\|=\sup_{ \overset{ \mbox{\scriptsize$X\in T_xM$} }{\lvert X
\rvert\le1}}\lvert\Hess u(X,X) \rvert.
\]
The following lemma gives the desired estimates for derivatives of $h$.
We refer to \cite{HoVa} for the proofs of these estimates; see also \cite{CHR3}.
\begin{lem}\cite[Lemma 3.16]{HoVa}\label{arvio_lause}
There exist constants $R_1=R_1(C)$ and $c_1=c_1(C)$ such that the extended function
$h\in C^\infty(M)\cap C(\bar M)$ in \eqref{extend_h}
satisfies
\begin{equation}\label{arvio1}
\begin{split}
|\nabla h(x)|&\le c_1\frac{1}{(f_a\circ\rho)(x)}, \\
\|\Hess_x h\|&\le c_1\frac{(b\circ\rho)(x)}{(f_a\circ\rho)(x)}, \\
\end{split}
\end{equation}
for all $x\in 3\Omega\setminus B(o,R_1)$.
In addition,
\[h(x)=1
\]
for every $x\in M\setminus\bigl(2\Omega\cup B(o,R_1)\bigr)$.
\end{lem}
We define a function $F\colon M\to [0,\infty)$ and an elliptic operator $\tilde{Q}$ by setting
\begin{equation}\label{defF}
F(x)=\sup_{t\in\R}\abs{\bar{\nabla}f(x,t)}
\end{equation}
and
\begin{equation}\label{defqtilde}
\tilde{Q}[v]=\dv\frac{\nabla v}{\sqrt{1+\abs{\n v}^2}}+F(x).
\end{equation}
Let then $A>0$ be a fixed constant. We aim to show that
\begin{equation}\label{varphi_def}
\psi=A(R_{3}^{\delta}\rho^{-\delta}+h)
\end{equation}
is a supersolution $\tilde{Q}[\psi]< 0$
in the set $3\Omega\setminus\bar{B}(o,R_{3})$, where $\delta>0$ and $R_{3}>0$ are constants that will be specified later and $h$ is the extended function defined in \eqref{extend_h}.
We shall make use of the following estimates obtained in \cite{HoVa}:
\begin{lem}\cite[Lemma 3.17]{HoVa}\label{perusta}
There exist constants $R_2=R_2(C)$ and $c_2=c_2(C)$ with the following property.
If $\delta\in(0,1)$, then
\[\begin{split}
|\nabla h|&\le c_2/(f_a\circ\rho), \\
\|\Hess h\|&\le c_2\rho^{-C_4-1}(f_a'\circ\rho)/(f_a\circ\rho), \\
|\nabla\langle\nabla h,\nabla h\rangle|&\le c_2\rho^{-C_4-2}(f_a'\circ\rho)/(f_a\circ\rho), \\
|\nabla\langle\nabla h,\nabla(\rho^{-\delta})\rangle|&\le c_2\rho^{-C_4-2}(f_a'\circ\rho)/(f_a\circ\rho), \\
\nabla\bigl\langle\nabla(\rho^{-\delta}),\nabla(\rho^{-\delta})\bigr\rangle
&=-2\delta^2(\delta+1)\rho^{-2\delta-3}\nabla\rho
\end{split}\]
in the set $3\Omega\setminus B(o,R_2)$.
\end{lem}
As in \cite{HoVa} we denote
\[
\phi_{1}=\frac{1+\sqrt{1+4C_{1}^{2}}}{2}>1,\quad\text{and}\quad\delta_{1}
=\min\left\lbrace C_{4},\frac{-1+(n-1)\phi_{1}}{1+(n-1)\phi_{1}}\right\rbrace
\in(0,1),
\]
where $C_{1}$ and $C_{4}$ are constants defined in \eqref{A1} and \eqref{A7},
respectively.
\begin{lem}\label{PsiBarrierLemma}
Let $A>0$ be a fixed constant and $h$ the
function defined in \eqref{extend_h}. Assume that the function $F$ defined in \eqref{defF} satisfies
\begin{equation}\label{Fcond}
\sup_{\rho(x)=t}F(x)=o\left(\frac{f^{\prime}_a (t)}{f_a(t)}t^{-\ve-1}\right)
\end{equation}
for some $\epsilon>0$ as $t\to\infty$. Then there
exist two positive constants $\delta\in(0,\min(\delta_1,\ve))$ and $R_3$
depending on $C$ and $\varepsilon$ such that the function
$\psi=A(R_3^{\delta} \rho^{-\delta} + h)$ satisfies $\tilde{Q}[\psi]< 0$
in the set $3\Omega\setminus \bar{B}(o,R_3)$.
\end{lem}
\begin{proof}
In the proof $c$ will denote a positive constant whose actual value may vary even within a line.
Since
\begin{align*}
\tilde{Q}[\psi] &= \frac{\Delta \psi}{\sqrt{1+|\nabla\psi|^2}} - \frac{1}{2} \frac{\ang{\nabla|\nabla\psi|^2,
\nabla\psi}}{(1+|\nabla\psi|^2)^{3/2}} +F(x) \\
&= \frac{(1+|\nabla \psi|^2) \Delta\psi + (1+|\nabla\psi|^2)^{3/2} F(x) - \frac{1}{2}
\ang{\nabla|\nabla\psi|^2,\nabla\psi}}{(1+|\nabla\psi|^2)^{3/2}},
\end{align*}
it is enough to show that there exist $\delta >0$ and $R_3$
such that
\begin{equation} \label{PsiBarrierIneq}
(1+|\nabla \psi|^2) \Delta\psi + (1+|\nabla\psi|^2)^{3/2} F(x) - \frac{1}{2}
\ang{\nabla|\nabla\psi|^2,\nabla\psi} < 0
\end{equation}
in the set $3\Omega \setminus \bar{B}(o,R_3)$.
First we notice that $\psi$ is $C^{\infty}$-smooth and
\[
\nabla\psi=A\bigl(-R_3^{\delta}\delta\rho^{-\delta-1}\nabla\rho +\nabla h\bigr)
\]
in $M\setminus\{o\}$.
Lemma \ref{perusta} and our curvature assumption imply that
$|\nabla h| \le c/\rho$ for $\rho$ large enough, and therefore
\begin{equation*}
|\nabla \psi|^2 = (AR_3^\delta)^2 \delta^2 \rho^{-2\delta-2} + A^2|\nabla h|^2 - 2A^2R_3^\delta \delta
\rho^{-\delta-1} \ang{\nabla \rho,\nabla h}
\le c\rho^{-2}
\end{equation*}
in $3\Omega \setminus \bar{B}(o,R_3)$ for sufficiently large $R_3$. Then, to estimate the term with $\Delta \psi$ in \eqref{PsiBarrierIneq}, we first note that
\[
\Delta \psi = AR_3^\delta \big( \delta(\delta+1)\rho^{-\delta-2} -\delta \rho^{-\delta-1}\Delta \rho \big)
+ A\Delta h.
\]
Furthermore, for every $\delta\in (0,\delta_1)$, there exists $R_3=R_3(C,\delta)$ such that
\[
\Delta \rho \ge (n-1) \frac{f_a' \circ \rho}{f_a\circ \rho}
\ge \frac{(n-1)(1-\delta)\phi_1}{\rho}
>0
\]
whenever $\rho\ge R_3$; see \cite[(3.25)]{HoVa}.
Therefore, using Lemma \ref{perusta}, we obtain
\begin{align*}
(1+|\nabla \psi|^2)\Delta\psi &\le (1+|\nabla \psi|^2)AR_3^\delta \delta\left( \delta+1
- (n-1) \frac{\rho f_a' \circ \rho}{f_a\circ \rho} \right)\rho^{-\delta-2} \\
&\qquad + (1+|\nabla \psi|^2) Anc_2 \left(\frac{f_a' \circ \rho}{f_a\circ \rho} \right)
\rho^{-C_4-1} \\
&\le AR_3^\delta \delta \left( \delta+1 -(n-1) \frac{\rho f_a' \circ \rho}
{f_a\circ \rho} \right) \rho^{-\delta-2} \\
&\qquad + \left(1+ c\rho^{-2}\right) Anc_2 \left(\frac{\rho f_a' \circ \rho}{f_a\circ \rho}\right)\rho^{-C_4 -2}\\
&= -\left(\frac{\rho f_a' \circ \rho}{f_a\circ \rho}\right)\rho^{-\delta-2}
\left(AR_3^{\delta}\delta(n-1)-(1+c\rho^{-2})Anc_2\rho^{\delta-C_4}\right)\\
&\qquad + AR_3^{\delta}\delta(\delta+1)\rho^{-\delta-2}\\
&\le -c\left(\frac{\rho f_a^{\prime}\circ \rho}{f_a \circ\rho}\right)\rho^{-\delta-2}
\end{align*}
whenever $\delta\in (0,\delta_1)$ is small enough and $\rho\ge R_3(C,\delta)$.
These estimates hold since
\[
\delta +1 -(n-1)\frac{\rho f_a^{\prime}\circ\rho}{f_a\circ\rho}\le \delta +1 -(n-1)(1-\delta)\phi_1\le 0
\]
for a sufficiently small $\delta\in (0,\delta_1)$.
Now taking into account our assumption \eqref{Fcond} we obtain
\begin{align}
\begin{split}
(1+\abs{\nabla\psi}^2) \Delta\psi + (1+\abs{\nabla\psi}^2)^{3/2}F &
\le -c\left(\frac{\rho f_a^{\prime}\circ \rho}{f_a \circ\rho}\right)\rho^{-\delta-2} +
(1+c\rho^{-2})F \label{two-terms-est}\\
&\le -c\left(\frac{\rho f_a^{\prime}\circ \rho}{f_a \circ\rho}\right)\rho^{-\delta-2}
\end{split}
\end{align}
whenever $\delta\in (0,\min(\varepsilon,\delta_1))$ is small enough and $\rho\ge R_3(C,\delta)$.
It remains to estimate $\abs{\ang{\nabla|\nabla\psi|^2,\nabla\psi}}$ from above. Since
\[
\nabla\psi=AR_{3}^{\delta}\nabla(\rho^{-\delta})+A\nabla h,
\]
we have
\begin{align*}
\nabla|\nabla\psi|^2 &= A^2\nabla\ang{R_3^{\delta}\nabla(\rho^{-\delta})+\nabla h,R_3^{\delta}\nabla(\rho^{-\delta})+\nabla h}\\
&=(AR_3^{\delta})^2\nabla\ang{\nabla(\rho^{-\delta}),\nabla(\rho^{-\delta})}
+2A^2 R_3^{\delta}\nabla\ang{\nabla(\rho^{-\delta}),\nabla h}
+A^2 \nabla\ang{\nabla h,\nabla h}.
\end{align*}
By Lemma \ref{perusta} we then get
\begin{align}\label{3rd-term-est}
\abs{\ang{\nabla|\nabla\psi|^2,\nabla\psi}} &\le c\rho^{-1}\left(2(\delta AR_3^{\delta})^2(\delta+1)\rho^{-2\delta-3}
+A^2c_2(2R_3^\delta +1)\left(\frac{f_a^\prime \circ\rho}{f_a \circ\rho}\right)\rho^{-C_4 -2}\right)\nonumber\\
&\le c\delta^2 (\delta+1)\rho^{-2\delta-4} + c\left(\frac{\rho f_a^{\prime}\circ\rho}{f_a\circ\rho}\right)\rho^{-C_4 -4}\\
&\le c\left(\rho^{-2\delta-4}+\rho^{-C_4 -4}\right)\frac{\rho f_a^{\prime}\circ\rho}{f_a\circ\rho}.\nonumber
\end{align}
Putting together \eqref{two-terms-est} and \eqref{3rd-term-est} we finally obtain
\[
(1+|\nabla \psi|^2) \Delta\psi + (1+|\nabla\psi|^2)^{3/2} F(x) - \frac{1}{2}
\ang{\nabla|\nabla\psi|^2,\nabla\psi}
\le
-c\left(\frac{\rho f_a^{\prime}\circ \rho}{f_a \circ\rho}\right)\rho^{-\delta-2}<0
\]
in $3\Omega \setminus \bar{B}(o,R_3)$ for a sufficiently small $\delta>0$ and large $R_3$.
\end{proof}
Similarly, we have
\begin{equation}
\label{subsol}
\dv\frac{\nabla(-\psi)}{\sqrt{1+\abs{\nabla(-\psi)}^2}} - F(x)>0
\end{equation}
in $3\Omega\setminus \bar{B}(o,R_3)$.
\subsection{Uniform height estimate}
We will solve the asymptotic Dirichlet problem by solving the problem first in a sequence of
balls with increasing radii. In order to obtain a converging subsequence of solutions, we need
to have a uniform height estimate. This subsection is devoted to the construction of a barrier function
that will guarantee the height estimate.
Since $f_a''-a^2f_a=0$, where $a(t)=0$ for $t\in [0,T_0]$ and
\[
a(t)\ge \frac{\sqrt{\phi(\phi-1)}}{t}
\]
for $t\ge T_1$ and some $\phi>1$, we have $f_a(t)\ge ct^\phi$ for $t\ge T_1$. Therefore
\begin{equation}\label{int_psi_finite}
\int_1^\infty \frac{dr}{f_a^{n-1}(r)} < \infty.
\end{equation}
Let $\varphi \colon M \to \R$ be a bounded function. We aim to show the existence of a barrier function
$V$ such that $\tilde{Q}[V] \le 0$ and $V(x)>||\varphi||_\infty$ in $M$.
In order to define such a function $V$, we need an auxiliary function $a_0>0$, so that
\begin{equation}\label{HP1}
\int_1^\infty \left( \int_r^\infty \frac{ds}{f_a^{n-1}(s)} \right) a_0(r)f_a^{n-1}(r) dr <\infty.
\end{equation}
We will discuss about the choice of $a_0$ in Examples \ref{a_0_exam1} and \ref{a_0_exam2}.
Now, following \cite{mastrolia2015elliptic}, we can define
\begin{equation} \label{Vdefinition}\begin{split}
V(x) = V\big(\rho(x)\big) &= \left(\int_{\rho(x)}^\infty \frac{ds}{f_a^{n-1}(s)}\right) \left(\int_0^{\rho(x)} a_0(t)
f_a^{n-1}(t) dt \right) \\ &\quad - \int_0^{\rho(x)} \left( \int_t^\infty \frac{ds}{f_a^{n-1}(s)}\right) a_0(t)
f_a^{n-1}(t) dt - H + ||\varphi||_\infty,\end{split}
\end{equation}
where
\begin{equation}\label{Hdefinition}\begin{split}
H\coloneqq \limsup_{r\to\infty} \Big\{ \int_{r}^\infty &\frac{ds}{f_a^{n-1}(s)} \int_0^{r} a_0(t)
f_a^{n-1}(t) dt \\ &- \int_0^{r} \int_t^\infty \frac{ds}{f_a^{n-1}(s)} a_0(t)
f_a^{n-1}(t) dt \Big\} \le 0;\end{split}
\end{equation}
see \cite[(4.5)]{mastrolia2015elliptic}.
From \eqref{int_psi_finite} and \eqref{HP1} we see that $H$ is finite and hence $V$ is well defined.
As in the proof of Lemma \ref{PsiBarrierLemma}, we write
\begin{equation}\label{Qfraction}
\tilde{Q}[V] = \frac{(1+|\nabla V|^2) \Delta V + (1+|\nabla V|^2)^{3/2}F(x) - \frac{1}{2}
\ang{\nabla |\nabla V|^2, \nabla V}}{(1+|\nabla V|^2)^{3/2}},
\end{equation}
where $F(x)$ is as in \eqref{defF}, and estimate the terms of the numerator. To begin, we notice that
\begin{align*}
V'(r) &= - \frac{1}{f_a^{n-1}(r)} \int_0^r a_0(t) f_a^{n-1}(t) dt < 0,\\
V''(r)&= (n-1) \frac{f_a^\prime(r)}{f_a^n(r)} \int_0^r a_0(t) f_a^{n-1}(t) dt - a_0(r),\\
\noalign{and}
\left|\nabla V\big(\rho(x)\big)\right| &= \left| V'\big(\rho(x)\big) \nabla \rho(x)\right|
= \left|V'\big(\rho(x)\big)\right|.
\end{align*}
Note that $-V(r)=g(r)$, the function \eqref{gV} in Introduction.
The Laplace comparison theorem implies that
\[
\Delta \rho \ge (n-1)\frac{f_a^\prime\circ \rho}{f_a \circ \rho}.
\]
Hence we can estimate the Laplacian of $V$ as
\begin{align*}
\Delta V &= V''\big(\rho\big) + \Delta \rho V'\big(\rho\big) \\
&\le V''(\rho) + (n-1)\frac{f_a^\prime(\rho)}{f_a(\rho)}V'(\rho) \\
&= (n-1) \frac{f_a^\prime(\rho)}{f_a^n(\rho)} \int_0^\rho a_0(t) f_a^{n-1}(t) dt - a_0(\rho)
- (n-1)\frac{f_a^\prime(\rho)}{f_a^n(\rho)}
\int_0^\rho a_0(t) f_a^{n-1}(t) dt \\
&= -a_0(\rho),
\end{align*}
and thus the first term of \eqref{Qfraction} can be estimated as
\[
\big(1 + |\nabla V|^2\big) \Delta V \le -\big(1 + |\nabla V|^2\big)a_0(\rho)
\le -\big(1+V'(\rho)^2\big) a_0(\rho).
\]
Then, for the last term of \eqref{Qfraction} we have
\begin{align*}
-\frac{1}{2}\ang{\nabla|\nabla V|^2,\nabla V} &= -\frac{1}{2}\ang{\nabla (V'(\rho))^2, V'(\rho) \nabla \rho} = -\frac{1}{2}\ang{2V'(\rho) V''(\rho) \nabla \rho,
V'(\rho)\nabla \rho} \\
&= -\big(V'(\rho)\big)^2 V''(\rho) \\
&= \frac{-1}{f_a^{2n-2}(\rho)} \left(\int_0^\rho a_0(t) f_a^{n-1}(t) dt \right)^2 \\
&\qquad \cdot \left((n-1) \frac{f_a^\prime(\rho)}{f_a^n(\rho)}
\int_0^\rho a_0(t) f_a^{n-1}(t) dt - a_0(\rho) \right) \\
&= \frac{a_0(\rho)}{f_a^{2n-2}(\rho)} \left(\int_0^\rho a_0(t) f_a^{n-1}(t) dt \right)^2 \\
&\qquad -\frac{(n-1)f_a^\prime(\rho)}{f_a^{3n-2}(\rho)} \left(\int_0^\rho a_0(t) f_a^{n-1}(t) dt \right)^3\\
&= a_0(\rho) V'(\rho)^2 - (n-1) \frac{f_a^\prime(\rho)}{f_a(\rho)} \big( -V'(\rho) \big)^3.
\end{align*}
Collecting everything together, we obtain that $\tilde{Q}[V] \le 0$ if
\begin{align*}
\sup_{\p B(o,r) \times\R} |\bar\nabla f| \le \frac{a_0(r) + (n-1)
\frac{f_a^\prime(r)}{f_a(r)}\big( -V'(r) \big)^3}{\big( 1+V'(r)^2 \big)^{3/2}}.
\end{align*}
Finally it is easy to check that,
since $H$ is finite and $V$ is decreasing, we have
$V(x)>||\varphi||_\infty$ for all $x\in M$ and $V(x) \to ||\varphi||_\infty$ as $\rho(x) \to \infty$.
Altogether, we have obtained the following.
\begin{lem}\label{UniformBoundLem}
Let $\varphi \colon M \to \R$ be a bounded function and assume that the function $V$ defined in \eqref{Vdefinition} satisfies
\begin{align}\label{grad_fCond}
\sup_{\p B(o,r) \times\R} |\bar\nabla f| \le \frac{a_0(r) + (n-1)
\frac{f_a^\prime(r)}{f_a(r)}\big( -V'(r) \big)^3}{\big( 1+V'(r)^2 \big)^{3/2}}.
\end{align}
Then the function $V$ is an upper barrier for the Dirichlet problem such that
\begin{equation}\label{Vsuper}
\tilde{Q}[V] =\dv\frac{\nabla V}{\sqrt{1+\abs{\nabla V}^2}} + F(x)\le 0 \quad \text{in } M,
\end{equation}
\begin{equation}\label{Vheight}
V(x) > ||\varphi||_\infty \quad \text{for all } x\in M
\end{equation}
and
\begin{equation}\label{Vlimes}
\lim_{r(x)\to\infty} V(x) = ||\varphi||_\infty.
\end{equation}
Furthermore,
\begin{equation}
\label{-Vsubsol}
\dv\frac{\nabla(-V)}{\sqrt{1+\abs{\nabla(-V)}^2}} - F(x)\ge 0\quad \text{in } M.
\end{equation}
\end{lem}
Next we show by examples that in the situation of Corollaries \ref{thm1} and \ref{HVkor2_RT} the
condition \eqref{grad_fCond} is not a stronger restriction than the assumption \eqref{Fcond} in Lemma
\ref{PsiBarrierLemma}. First note that $V'(r) \to 0$ as $r\to\infty$, and hence the upper
bound \eqref{grad_fCond} for $|\nb f|$ is asymptotically the function $a_0$.
\begin{exa}\label{a_0_exam1}
Assume that the sectional curvatures of $M$ satisfy
\[
K(P_x)\le -a\big(\rho(x)\big)^2 =- \frac{\phi(\phi-1)}{\rho(x)^2}, \quad \phi>1,
\]
for $\rho(x)\ge T_1$.
We need to choose the function $a_0$ such that \eqref{HP1} holds, and since this is a
question about its asymptotical behaviour, it is enough to consider the integral
\[
\int_{T_1}^\infty \left( \int_r^\infty \frac{ds}{f_a^{n-1}(s)} \right) a_0(r)f_a^{n-1}(r) dr.
\]
For $t\ge T_1$, $f_a(t) = c_1 t^\phi + c_2 t^{1-\phi}$, and hence, by a straightforward computation, we have \eqref{HP1} if
\[
\int_{T_1}^\infty a_0(r)r \, dr < \infty.
\]
So it is enough to choose for example
\[
a_0(r) = O\left(\frac{1}{r^2(\log r)^{\alpha}}\right)
\]
as $r\to\infty$ for some $\alpha>1$. On the other hand, with this curvature upper bound, the assumption \eqref{Fcond}
requires decreasing of order $o\bigl(r^{-2-\ve}\bigr)$.
\end{exa}
\begin{exa}\label{a_0_exam2}
Assume that the sectional curvatures of $M$ satisfy
\[
K\le-k^2,
\]
for $\rho(x)\ge T_1$ and some constant $k>0$. Then, for large $t$,
$f_a(t) = c_1 \sinh kt+ c_2\cosh kt\approx e^{kt}$. Therefore
it is straightforward to see that we have \eqref{HP1} if
\[
\int_{T_1}^\infty a_0(r)\,dr <\infty,
\]
which holds by choosing, for example,
\[
a_0(r) = O\left(\frac{1}{r(\log r)^\alpha}\right),\ \alpha>1,
\]
as $r\to\infty$. On the other hand, with this curvature upper bound, the assumption \eqref{Fcond}
requires decreasing of order $o\bigl(r^{-1-\ve}\bigr)$.
\end{exa}
\subsection{Proof of Theorem \ref{ThmMain}}
We start with solving the Dirichlet problem in geodesic balls $B(o,R)$.
\begin{lem}\label{adp-gb}
Suppose that $f\in C^2(M\times\R)$ is of the form $f(x,t)=m(x)+r(t)$ and
satisfies
\[
\sup_{\partial B(o,r)\times\R}\abs{\bar{\nabla}f} \le (n-1)\frac{f_a^\prime (r)}{f_a(r)}
\]
for all $r>0$. Then for every $R>0$ and $\varphi\in C(\partial B(o,R))$ there exists a solution
$u\in C^{2,\alpha}(B(o,R))\cap C(\bar{B}(o,R))$ of the Dirichlet problem
\begin{equation*}
\begin{cases}
\dv \dfrac{\nabla u}{\sqrt{1+|\nabla u|^2}} = \ang{\nb f,\nu} \quad \text{in } B(o,R) \\
u|\p B(o,R) = \varphi.
\end{cases}
\end{equation*}
\end{lem}
\begin{proof}
Assuming first that $\varphi\in C^{2,\alpha}(\partial B(o,R))$ the claim follows by the Leray-Schauder method. Indeed, for each
$x\in \bar{B}(o,R)\setminus\{o\}$ the inward mean curvature $H(x)$ of the level set
$\{y\in \bar{B}(o,R)\colon d(y)=d(x)\}=\partial B(o,\rho(x))$ satisfies
\[
H(x)=\Delta\rho(x)\ge (n-1)\frac{f_a^\prime \big(\rho(x)\big)}{f_a\big(\rho(x)\big)}
\ge \sup_{\partial B(o,\rho(x))\times\R}\abs{\bar{\nabla}f}.
\]
In other words, \eqref{Fxtra-again} and \eqref{H-f-again} hold and therefore we can apply
the Leray-Schauder method as in the proof of Theorem~\ref{existence}. The general case
$\varphi\in C(\partial B(o,R))$ follows by approximation as in the proof of
Theorem~\ref{cont_existence}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{ThmMain}]
We extend the boundary data function $\varphi \in C(\p_\infty M)$ to a function $\varphi \in C(\bar M)$.
Let $\Omega_k = B(o,k), \, k\in\N$, be an exhaustion of $M$.
By Lemma~\ref{adp-gb}, there exist solutions $u_k\in C^{2,\alpha}(\Omega_k)\cap C(\bar\Omega_k)$ to
\begin{equation*}
\begin{cases}
Q[u_k]= \dv \dfrac{\nabla u_k}{\sqrt{1+|\nabla u_k|^2}} - \ang{\nb f,\nu_k} \quad \text{in } \Omega_k \\
u_k|\p\Omega_k = \varphi,
\end{cases}
\end{equation*}
where $\nu_k$ is the downward pointing unit normal to the graph of $u_k$.
Applying the uniform height estimate, Lemma~\ref{UniformBoundLem},
we see that the sequence $(u_k)$ is uniformly bounded and hence the
interior gradient estimate (Lemma \ref{globestim}), together with the diagonal argument, implies that
there exists a subsequence, still denoted by $u_k$, that converges locally uniformly
with respect to $C^2$-norm to a solution $u$. Therefore we are left to prove that $u$ extends continuously
to $\pinf M$ and satisfies $u|\pinf M = \varphi$.
Towards that end let us fix $x_0\in\pinf M$ and $\ve>0$. Since the boundary data function $\varphi$ is continuous, we find
$L\in(8/\pi,\infty)$ such that
\[
|\varphi(y) - \varphi(x_0)| < \ve/2
\]
for all $y \in C(v_0,4/L)\cap \pinf M$, where $v_0=\dot\gamma_0^{o,x_0}$ is the initial vector of the geodesic
ray representing $x_0$. Moreover, by \eqref{Vlimes} we can choose $R_3$ in Lemma~\ref{PsiBarrierLemma} so large that
$V(r)\le \max_{\bar{M}} |\varphi| + \ve/2$ for $r\ge R_3$.
We claim that
\begin{equation}\label{squeeze_ineq}
w^-(x) \coloneqq -\psi(x) + \varphi(x_0) - \ve \le u(x) \le w^+(x) \coloneqq \psi(x) + \varphi(x_0) + \ve
\end{equation}
in the set $U\coloneqq 3\Omega \setminus \bar{B}(o,R_3)$, where $\psi=A(R_3^\delta \rho^{-\delta} + h)$ is the
supersolution $\tilde{Q}[\psi]<0$ in Lemma \ref{PsiBarrierLemma} and $A= 2\max_{\bar M}|\tilde\varphi|$.
Recall the notation $\Omega = C(v_0,1/L) \cap M$ and $\ell\Omega = C(v_0,\ell/L) \cap M,\ \ell>0$,
from Subsection \ref{subsec_barrier_constr}.
The function $\varphi$ is continuous in $\bar{M}$ so there exists $k_0$ such
that $\p\Omega_{k_0} \cap U \neq \emptyset$, and
\begin{equation}\label{tilde_phi-phi}
|\varphi(x) - \varphi(x_0)| < \ve/2
\end{equation}
for all $x \in \p\Omega_k \cap U$ when $k\ge k_0$. Denote $V_k = \Omega_k \cap U$ for $k\ge k_0$. We will conclude that
\begin{equation}\label{ineq_in_Vk}
w^- \le u_k \le w^+
\end{equation}
in $V_k$ by using the comparison principle for the operator $\tilde Q_k$,
\[
\tilde{Q}_k[v]=\dv \dfrac{\nabla v}{\sqrt{1+|\nabla v|^2}} - \ang{\nb f,\nu_k},
\]
where $\nu_k$ is the downward pointing unit normal to the graph of the solution $u_k$. Notice that
\[
\p V_k = (\p\Omega_k \cap \bar U) \cup (\p U \cap \bar\Omega_k).
\]
Let $x \in \p\Omega_k \cap \bar U$ and $k\ge k_0$. Then \eqref{tilde_phi-phi} and $u_k|\p\Omega_k = \varphi|\p\Omega_k$ imply that
\[
w^-(x) \le \varphi(x_0) - \ve/2 \le \varphi(x) = u_k(x) \le \varphi(x_0) + \ve/2 \le w^+(x).
\]
Moreover, by Lemma \ref{arvio_lause}, we have
\[
h|M \setminus \big(2\Omega \cup B(o,R_1)\big) = 1
\]
and $R_3^\delta \rho^{-\delta} = 1$ on $\p B(o,R_3)$, so
\[
\psi \ge A = 2\max_{\bar M} |\varphi|
\]
on $\p U \cap \bar\Omega_k$.
By Lemma \ref{UniformBoundLem}, $V$ is a supersolution $\tilde Q[V]\le 0$ and hence
\begin{align*}
\dv \frac{\nabla V}{\sqrt{1+|\nabla V|^2}} - \ang{\nb f,\nu_k} &\le
\dv \frac{\nabla V}{\sqrt{1+|\nabla V|^2}} + F(x)\\
& = \tilde Q[V] \le 0\\
& =\dv \frac{\nabla u_k}{\sqrt{1+|\nabla u_k|^2}} - \ang{\nb f,\nu_{k}}.
\end{align*}
Since $V\ge \max_{\bar M} |\varphi|$ on $\p\Omega_k$, the comparison principle yields $u_k|\Omega_k \le V|\Omega_k,$
and by the choice of $R_3$, we have
\[
u_k \le \max_{\bar M} |\varphi| + \ve/2
\]
in $\Omega_k \setminus B(o,R_3)$.
Altogether, it follows that
\[
w^+= \psi + \varphi(x_0) + \ve \ge 2\max_{\bar M} |\varphi| + \varphi(x_0) + \ve
\ge \max_{\bar M}|\varphi| + \ve \ge u_k
\]
on $\p U\cap \bar\Omega_k$, and similarly $u_k \ge w^-$ on $\p U\cap \bar\Omega_k$. Consequently $w^-\le u_k \le w^+ $
on $\p V_k$.
By Lemma~\ref{PsiBarrierLemma}, $\tilde{Q}[\psi]<0$, and therefore
\begin{align*}
\tilde{Q}_k[w^+]&=\dv \frac{\nabla w^+}{\sqrt{1+|\nabla w^+|^2}} - \ang{\nb f,\nu_{k}}\\
&=
\dv \frac{\nabla\psi}{\sqrt{1+|\nabla \psi|^2}} - \ang{\nb f,\nu_{k}}\\
&\le \dv \frac{\nabla\psi}{\sqrt{1+|\nabla \psi|^2}} + F(x)\\
& = \tilde Q[\psi] < 0\\
& =\dv \frac{\nabla u_k}{\sqrt{1+|\nabla u_k|^2}} - \ang{\nb f,\nu_{k}}
\end{align*}
in $U$. By the comparison principle, $u_k\le w^+$ in $U$. Similarly, using \eqref{subsol} we conclude that
\[
\dv \frac{\nabla w^-}{\sqrt{1+|\nabla w^-|^2}} - \ang{\nb f,\nu_{k}} >
\dv \frac{\nabla u_k}{\sqrt{1+|\nabla u_k|^2}} - \ang{\nb f,\nu_{k}}
\]
in $U$. Hence $u_k\ge w^-$ in $U$ and we obtain \eqref{ineq_in_Vk}. This holds for every $k\ge k_0$ and hence \eqref{squeeze_ineq} follows.
Finally,
\[
\limsup_{x\to x_0} |u(x) - \varphi(x_0)| \le \ve
\]
since $\lim_{x\to x_0} \psi(x) =0$. Because $x_0\in\pinf M$ and $\ve>0$ were arbitrary, this shows that $u$
extends continuously to $C(\bar M)$ and $u|\pinf M = \varphi$.
\end{proof}
|
\section{Introduction}
In this paper we deal with the distributed solution of the following
optimization problem:
\begin{equation}\label{eq:Problem}
\underset{x\in\Re^n}{\minimize} \ \sum_{i=1}^{N} f_i(x) + g_i(C_i x)
\end{equation}
where for $i=1,\ldots,N$, $C_i$ is a linear operator, $f_i$ and $g_i$ are proper closed convex and possibly nonsmooth functions. We further assume that the \emph{proximal mappings} associated with $f_i$ and $g_i$
are efficiently computable \cite{combettes2011proximal}. In a more general case we can include another continuously differentiable term with Lipschitz-continuous gradient in
\eqref{eq:Problem} and use \cite[Algorithm 3]{AFBA2016} that includes the algorithm of V\~u and Condat \cite{vu2013splitting,condat2013primal} as special case. We do not pursue this here for clarity of exposition.
Problems of this form appear in several application fields. In a distributed model predictive control setting, $f_i$ can represent individual finite-horizon costs for each agent, $C_i$ model the linear dynamics of each agent and possibly coupling constraints that are split through the introduction of extra variables, and $g_i$ model state and input constraints.
In machine learning
and statistics the $C_i$ are feature matrices and functions $g_i$ measure the \emph{fitting}
of a predicted model with the observed data, while the $f_i$ is \emph{regularization} terms that enforces
some prior knowledge in the solution (such as sparsity, or belonging to a certain constraint set).
For example if $g_i$ is the so-called hinge loss and $f_i = \tfrac{\lambda}{2}\|\cdot\|_2^2$, for some $\lambda > 0$,
then one recovers the standard SVM model. If instead $f_i = \lambda\|\cdot\|_1$ then
one recovers the $\ell_1$-norm SVM problem \cite{zhu20041}.
Clearly problem \eqref{eq:Problem} can be solved in a centralized fashion, when all the
data of the problem (functions $f_i$, $g_i$ and matrices $C_i$, for all $i\in\{1,\ldots,N\}$) are available at one
computing node. When this is the case one might formulate and solve the aggregated problem
\begin{equation*
\underset{x\in\Re^n}{\minimize} \ f(x) + g(C x),
\end{equation*}
for which algorithms are available \cite{chambolle2011first,AFBA2016,briceno2011monotone+}.
However, such a centralized approach is not realistic in many scenarios. For example, suppose
that $g_i(C_i x)$ models least-squares terms and $C_1,\ldots,C_N$ are very large features matrices.
Then collecting $C_1,\ldots,C_N$ into a single computer may be infeasible due to communication costs,
or even worse they may not fit into the computer's memory. Furthermore, the exchange of such information
may not be possible at all due to privacy issues.
Our goal is therefore to solve problem \eqref{eq:Problem} in a distributed
fashion. Specifically, we consider a connected network of $N$ computing agents,
where the $i$-th agent is able to compute proximal mappings of $f_i$, $g_i$, and matrix vector products with $C_i$ (and its adjoint operator). We want all the agents
to iteratively converge to a \emph{consensus} solution to \eqref{eq:Problem}, and to do so
by only exchanging variables among neighbouring nodes, i.e, no centralized computations (\emph{i.e.}, existence of a fusion center) are needed
during the iterations.
To do so, we will propose a solution based on the recently introduced \emph{Asymmetric Forward-Backward-Adjoint} (AFBA) splitting method \cite{AFBA2016}. This new splitting technique solves monotone inclusion problems involving three operators, however, in this work we will focus on a special case that involves two terms. Specifically, we develop a distributed algorithm which is based on a special case of AFBA applied to the monotone inclusion corresponding to the primal-dual optimality conditions of a suitable \emph{graph splitting} of~\eqref{eq:Problem}. Our algorithm involves a nonnegative parameter $\theta$ which serves as a tuning knob that allows to recover different algorithms. In particular, the algorithm of \cite{chambolle2011first} is recovered in the special case when $\theta=2$. We demonstrate how tuning this parameter affects the stepsizes and ultimately the convergence rate of the algorithm.
Other algorithms have been proposed for solving problems similar to \eqref{eq:Problem} in a distributed way.
As a reference framework, all algorithms aim at solving in a distributed way the problem
\begin{equation*}
\underset{x\in\Re^n}{\minimize} \ \sum_{i=1}^N F_i(x).
\end{equation*}
In \cite{nedic2009distributed} a distributed subgradient method is proposed, and in \cite{duchi2012dual} this
idea is extended to the projected subgradient method.
More recently, several works focused on the use of
ADMM for distributed optimization.
In \cite{Boyd2010a} the generic ADMM for consensus-type problems is illustrated.
A drawback of this approach is that at every iteration the agents must solve a complicated subproblem that might
require an inner iterative procedure. In \cite{Parikh2014} another formulation is given for the
case where $F_i = f_i+g_i$, and only proximal mappings with respect to $f_i$ and $g_i$ are separately computed in each node.
Still, when either $f_i$ or $g_i$ is not separable (such as when they are composed with linear operators) these are
not trivial to compute and may require inner iterative procedures, or factorization of the data matrices involved.
Moreover, in both \cite{Boyd2010a, Parikh2014} a central node is required for accumulating each agents variables at every iteration,
therefore these formulations lead to \emph{parallel} algorithms rather than distributed. In \cite{teixeira2013optimal} the optimal
parameter selection for ADMM is discussed in the case of distributed quadratic programming problems.
In \cite{wei2012distributed, wei20131, makhdoumi2014broadcast}, fully distributed algorithms based on ADMM proposed,
assuming that the proximal mapping of $F_i$ is computable, which is impractical in many cases. In \cite{bianchi2014primal} the authors propose a variation of the
V{\~u}-Condat algorithm \cite{condat2013primal,vu2013splitting}, having
ADMM as a special case, and show its application to distributed optimization where $F_i = f_i+g_i$, but no composition with a linear operator is involved.
Only proximal operations with respect to $f_i$ and $g_i$ and local exchange of variables (\emph{i.e.}, among neighboring nodes)
is required, and the method is analyzed in an asynchronous setting.
In this paper we deal with the more general scenario of problem \eqref{eq:Problem}.
The main features of our approach, that distinguish it from the related works mentioned above, are:
\begin{enumerate}[(i)]
\item We deal with $F_i$ that is the sum of two possibly nonsmooth functions one of which is composed with a linear operator.
\item Our algorithm only require local exchange of information, i.e., only neighboring
nodes need to exchange local variables for the algorithms to proceed.
\item The iterations involve \emph{direct} operations on the objective terms. Only evaluations of
$\mathop{\rm prox}\nolimits_{f_i}$, $\mathop{\rm prox}\nolimits_{g^\star_i}$ and matrix-vector products with $C_i$ and $C_i^T$ are involved.
In particular, no inner subproblem needs to be solved iteratively by the computing agents, and no matrix inversions are required.
\end{enumerate}
The paper is organized as follows.
\Cref{sec:Formulation} is devoted to a formulation of problem
\eqref{eq:Problem} which is amenable to be solved in a distributed fashion by the proposed methods. In \Cref{sec:Algorithms} we detail how the primal-dual algorithm in \cite[Algorithm 6]{AFBA2016} together with an intelligent change of variables gives rise to distributed iterations. We then discuss implementation considerations and convergence properties. In \Cref{sec:Simulations} we illustrate some numerical results for several values of the constant $\theta$, highlighting the improved performance for $\theta=1.5$.
\section{Problem Formulation} \label{sec:Formulation}
Consider problem~\eqref{eq:Problem}
under the following assumptions:
\begin{ass}
For $i=1,\ldots,N$:
\begin{enumerate}[(i)]
\item $C_i:\Re^n\to\Re^{r_i}$ are linear operators.
\item $f_i:\Re^n\to\overline{\rm I\!R}$, $g_i:\Re^{r_{i}}\to\overline{\rm I\!R}$ are proper closed convex functions, where $\overline{\rm I\!R}=\Re\cup\{\infty\}$.
\item The set of minimizers of~\eqref{eq:Problem}, denoted by $S^\star$, is nonempty.
\end{enumerate}
\end{ass}
We are interested in solving problem~\eqref{eq:Problem} in a distributed fashion. Specifically, let $G=(V,E)$ be an undirected graph over the vertex set $V=\{1,\ldots,N\}$ with edge set $E\subset V\times V$.
It is assumed that each node $i\in V$ is associated with a separate agent, and each agent maintains its own cost components $f_i$, $g_i$, $C_i$ which are assumed to be private, and its own opinion of the solution $x_i\in\Re^n$. The graph imposes communication constraints over agents. In particular, agent $i$ can communicate directly only with its neighbors $j\in\mathcal{N}_i=\{j\in V\ |\ (i,j)\in E\}$. We make the following assumption.
\begin{ass} \label{ass:conctd}
Graph $G$ is connected.
\end{ass}
With this assumption, we reformulate the problem as
\begin{align*}
\underset{\bm{x}\in\Re^{Nn}}{\minimize}&\quad \sum_{i=1}^N f_i(x_i)+g_i(C_ix_i)\\
\rm subject\ to&\quad x_i=x_j\qquad (i,j)\in E
\end{align*}
where $\bm{x}=(x_1,\ldots,x_N)$.
Associate any orientation to the unordered edge set $E$. Let $M=|E|$ and $B\in\Re^{N\times M}$ be the \emph{oriented node-arc incidence matrix}, where each column is associated with an edge $(i,j)\in E$ and has $+1$ and $-1$ in the $i$-th and $j$-th entry, respectively. Notice that the sum of each column of $B$ is equal to $0$. Let $d_i$ denote the degree of a given vertex, that is, the number of
vertices that are adjacent to it. We have $B B^\top=\mathcal{L}\in\Re^{N\times N}$, where $\mathcal{L}$ is the graph \emph{Laplacian} of $G$, i.e.,
$$\mathcal{L}_{ij}=\begin{cases}d_i&\textrm{ if } i=j,\\
-1&\textrm{ if } i\neq j \textrm{ and node $i$ is adjacent to node $j$},\\
0&\textrm{ otherwise}.
\end{cases}$$
Constraints $x_i=x_j$, $(i,j)\in E$ can be written in compact form as $A\bm{x}=0$, where $A=B^\top\otimes I_n\in\Re^{Mn\times Nn}$. Therefore, the problem is expressed as
\begin{equation} \label{eq:primal}
\underset{\bm{x}\in\Re^{Nn}}{\minimize}\quad \sum_{i=1}^N f_i(x_i)+g_i(C_ix_i)+\delta_{\{0\}}(A\bm{x}),
\end{equation}
where $\delta_{X}$ denotes the indicator function of a closed nonempty convex set, $X$. The dual problem is:
\begin{equation} \label{eq:dual}
\underset{\begin{subarray}{c}
y_i\in\Re^{r_i} \\
w\in\Re^{Mn}
\end{subarray}}{\minimize}\quad \sum_{i=1}^N f_i^*(-A^\top_i w-C_i^\top y_i)+g_i^*(y_i),
\end{equation}
where $q^*$ denotes the Fenchel conjugate of a function $q$ and $A_i\in\Re^{Mn\times n}$ are the block columns of $A$.
Let $\partial q$ denote the subdifferential of a convex function $q$. The primal-dual optimality conditions are
\begin{equation} \label{eq:-1}
\begin{cases}
0\in\partial f_{i}(x_{i})+C_{i}^{\top}y_{i}+A_{i}^{\top}w, & i=1,\ldots,N\\
C_{i}x_{i}\in\partial g_{i}^{*}(y_{i}) & i=1,\ldots,N\\
\sum_{i=1}^{N}A_{i}x_{i}=0,
\end{cases}
\end{equation}
where $w\in\Re^{Mn}$, $y_i\in\Re^{r_i}$, for $i=1,\ldots,N$.
The following condition will be assumed throughout the rest of the paper.
\begin{ass} \label{ass:exst-prmdul}
There exist $x_i\in \mathop{\rm ri}\nolimits \mathop{\rm dom}\nolimits f_i$ such that $C_i x_i\in \mathop{\rm ri}\nolimits \mathop{\rm dom}\nolimits g_i$, $i=1,\ldots,N$ and $\sum_{i=1}^{N}A_i x_i=0$\footnote{$\mathop{\rm dom}\nolimits f$ denotes the domain of function $f$ and $\mathop{\rm ri}\nolimits C$ is the relative interior of the set $C$.}.
\end{ass}
This assumption guarantees that the set of solutions to~\eqref{eq:-1} is nonempty (see \cite[Proposition 4.3(iii)]{combettes2012primal}). If $(\boldsymbol{x}^\star,\boldsymbol{y}^\star,w^\star)$ is a solution to~\eqref{eq:-1}, then $\boldsymbol{x}^\star$ is a solution to the primal problem~\eqref{eq:primal} and $(\boldsymbol{y}^\star,w^\star)$ to its dual~\eqref{eq:dual}.
\iffalse
\section{Asymmetric Forward-Backward-Adjoint Splitting} \label{sec:AFBA}
In this section we summarize the main results concerning AFBA~\cite{AFBA2016}. To facilitate exposition we first introduce some notation and definitions regarding monotone operators.
\subsection{Monotone Operators}
Let $2^\mathcal{X}$ be the power set of a nonempty set $\mathcal{X}$. An operator $T$ on $\mathcal{X}$ is a set valued mapping $T:\mathcal{X}\to2^{\mathcal{X}}$, i.e., it maps every point $x\in\mathcal{X}$ to a subset $Tx$ of $\mathcal{X}$. In this paper we consider the space $\mathcal{X}=\Re^n$, for definitions in general Hilbert spaces see \cite{bauschke2011convex}. The domain of $A$ is denoted by $\mathop{\rm dom}\nolimits (A)=\{x\in\Re^n|Ax\neq\textrm{Ø}\}$, its graph by $\mathop{\rm gra}\nolimits (A)=\{(x,u)\in\Re^n\times\Re^n|u\in Ax\}$ and the set of zeros of $A$ is $\mathop{\rm zer}\nolimits (A)=\{x\in\Re^n|0\in Ax\}$. The inverse of $A$ is defined through its graph
$\mathop{\rm gra}\nolimits (A^{-1})=\{(u,x)|(x,u)\in \mathop{\rm gra}\nolimits (A)\}$. Furthermore, $A$ is monotone if
\[
\langle x-y,u-v\rangle \geq 0, \quad \forall (x,u),(y,v)\in \mathop{\rm gra}\nolimits (A)
\]
and maximally monotone if it is monotone and there exists no monotone operator $B:\Re^n\to2^{\Re^n}$ such that $\mathop{\rm gra}\nolimits (A) \subset \mathop{\rm gra}\nolimits (B)$ and $A\neq B$.
\subsection{Asymmetric Forward-Backward-Adjoint Splitting}
Recently, a new splitting algorithm was proposed in \cite{AFBA2016} that can solve the monotone inclusion problem of finding $z\in\Re^n$ such that
\begin{equation}\label{eq:mntnincl}
0\in {D}{z}+{M}{z}+{F}{z},
\end{equation}
where $D$ is a maximally monotone operator, $M$ is a bounded linear operator and $F$ is cocoercive. Here we briefly review recent results concerning AFBA in $\Re^n$ and refer the reader to \cite{AFBA2016} for a complete analysis in general Hilbert spaces. In~\Cref{sec:Algorithms} we will see that the optimality conditions~\eqref{eq:-1} have the form of monotone inclusion~\eqref{eq:mntnincl}. By properly choosing the parameters in AFBA we are able to derive \Cref{Algorithm-Vu-Condat,Algorithm-4,Algorithm-5} for the distributed problem.
We make the following assumptions:
\begin{ass} \label{assumption-1}
Let $S$ and $P$ be symmetric positive definite linear operators
, $K$
skew-symmetric linear operator, \emph{i.e.}, $K^{\top}=-K$ and $H=P+K$.
\end{ass}
\begin{ass} \label{assumption-2}
$D$ is a maximally monotone operator, $M$ is a monotone linear operator, and $C$ is $\beta_C$-cocoercive with respect to the $P$ norm, where
$\beta_C\in]1/4,+\infty[$, i.e.,
\[
\beta_C\|Cz-Cz^{\prime}\|_{P^{-1}}^{2}\leq\langle Cz-Cz^{\prime},z-z^{\prime}\rangle\quad\forall z,z^{\prime}\in\Re^n.
\]
\end{ass}
\begin{algorithm}[H]
\caption{Asymmetric Forward-Backward-Adjoint Splitting (AFBA)}
\label{Algorithm-1}
\begin{algorithmic}
\item\textbf{Inputs:} $z^0\in\Re^n$
\For{$k=0,\ldots,$}
\State \hypertarget{alg:line1}{$\bar{z}^{k} =[H+D]^{-1}[H-M-F]z^{k}$}
\State $\tilde{z}^{k} =\bar{z}^{k}-z^{k}$
\State \hypertarget{alg:line3}{
$
\displaystyle
\alpha_k
=
\frac%
{
\lambda_k
\|\tilde z^k\|_P^2
}
{
\|
\left(
H+M^*
\right)
\tilde z^k
\|_{S^{-1}}^2
}}
$
\State \hypertarget{alg:line4}{ $z^{k+1}=z^{k}+\alpha_{k}S^{-1}(H+M^*)\tilde{z}^{k}$ \label{eq:line4}}
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{thm} \label{thm:conv}
Let \Cref{assumption-1,assumption-2} hold true. Suppose the set of solutions to~\eqref{eq:mntnincl} is nonempty. Assume the parameters of \Cref{Algorithm-1} satisfy the following:
\begin{enumerate}
\item $(\lambda_{k})_{k\in{\rm{I\!N}}}\subseteq[0,\delta]$, with $\delta=2-\frac{1}{2\kappa}$, $\delta>0$
\item $\underset{k\rightarrow\infty}{\lim\inf} \,\lambda_{k}(\delta-\lambda_{k})>0$.
\end{enumerate}
Then both sequences $(\bar{z}^{k})_{k\in{\rm{I\!N}}}$ and $({z}^{k})_{k\in{\rm{I\!N}}}$ converge to a solution of~\eqref{eq:mntnincl}. Furthermore, when $F\equiv0$ all of the above hold with $\delta=2$.
\end{thm}
\begin{thm}[Convergence rate] \label{Thm: conv-rates}
Consider \Cref{Algorithm-1} under the assumptions of~\Cref{thm:conv}.
Let $c_1$ and $c_2$ be two positive constants satisfying
\begin{equation*}
c_1 P \preceq R \preceq c_2 P,
\end{equation*}
where $R=(H+M^*)^*S^{-1}(H+M^*)$ and
\begin{equation*}
(\lambda_{k})_{k\in{\rm{I\!N}}}\subseteq\left[0,c_1\delta/c_2\right].
\end{equation*}
Then
\begin{enumerate}
\item $(\|\tilde{z}^k\|_R^2)_{k\in{\rm{I\!N}}}$ is monotonically nonincreasing.
\item If in addition, $(\lambda_{k})_{k\in{\rm{I\!N}}}\subseteq[\epsilon,\delta-\epsilon]$, for some $\epsilon>0$, then
\begin{equation*}
\|\tilde{z}^{k}\|_{R}^{2}\leq\frac{c_2^2}{\epsilon^2(k+1)}\|z_{0}-z^{\star}\|_{S}^{2},
\end{equation*}
and $\|\tilde{z}^{k}\|_{R}^{2}=o(1/(k+1))$.
\end{enumerate}
Furthermore, when $C\equiv 0$ all of the above statements hold with $\delta=2$.
\end{thm}
The linear operators $P,K,S$ are independent of each other. This opens the possibility of deriving many interesting algorithms by intelligently choosing these matrices. In \cite{AFBA2016}, many recent primal-dual algorithms were generalized by appropriate choices of parameters. Here we follow the formulation used in \cite[Section 5]{AFBA2016} and derive distributed primal-dual versions of Algorithms 4 to 6 of the aforementioned article. Refer to \cite{AFBA2016} for more detailed study of these algorithms in general Hilbert spaces.
\fi
\section{Distributed Primal-Dual Algorithms} \label{sec:Algorithms}
In this section we provide the main distributed algorithm that is based on \emph{Asymmetric Forward-Backward-Adjoint} (AFBA), a new operator splitting technique introduced recently \cite{AFBA2016}. This special case belongs to the class of primal-dual algorithms. The convergence results include both the primal and dual variables and are based on \cite[Propositions 5.4]{AFBA2016}). However, the convergence analysis here focuses on the primal variables for clarity of exposition, with the understanding that a similar error measure holds for the dual variables.
Our distributed algorithm
consists of two phases, a local phase and the phase in which each agent interacts with its neighbors according to the constraints imposed by the communication graph. Each iteration has the advantage of only requiring local matrix-vector products and proximal updates. Specifically, each agent performs $2$ matrix-vector products per iteration and transmits a vector of dimension $n$ to its neighbors.
Before continuing we recall the definition of Moreau's proximal mapping. Let $U$ be a symmetric positive-definite matrix.
The \emph{proximal mapping} of a proper closed convex function $f:\Re^n\to\overline{\rm I\!R}$ relative to $\|\cdot\|_U$ is defined by
\begin{equation*}
\mathop{\rm prox}\nolimits^U_f(x) = \operatornamewithlimits{argmin}_{z\in \Re^n} f(z)+\frac{1}{2} \|x-z\|_U^2,
\end{equation*}
and when the superscript $U$ is omitted the same definition applies with respect to the canonical norm.
Let $\boldsymbol{u}=(\boldsymbol{x},\boldsymbol{v})$ where
$\boldsymbol{v}=(\boldsymbol{y},w)$ and
$\boldsymbol{y}=(y_{1},\ldots,y_{N})$.
The optimality conditions in~\eqref{eq:-1}, can be written in the form of the following monotone inclusion:
\begin{equation}\label{eq:mon-inclusion}
0\in D\boldsymbol{u} + M\boldsymbol{u}
\end{equation}
with
\begin{align}
{D}(\boldsymbol{x},\boldsymbol{y},w)&=(\partial\boldsymbol{f}(\boldsymbol{x}), \partial\boldsymbol{g}^{*}(\boldsymbol{y}),0),\label{eq:Aopt
\end{align}
and
\begin{equation*}
{M}=\left[\begin{array}{ccc}
0 & \boldsymbol{C}^{\top} & {A}^{\top}\\
-\boldsymbol{C} & 0 & 0\\
-{A} & 0 & 0
\end{array}\right],
\end{equation*}
where $\boldsymbol{f}(\boldsymbol{x})=\sum_{i=1}^{N}f_{i}(x_{i})$
,
$\boldsymbol{g}^{*}(\boldsymbol{y})=\sum_{i=1}^{N}g_{i}^*(y_{i})$, $\boldsymbol{C}=\mathop{\rm blkdiag}\nolimits(C_1,\ldots,C_N)$. Notice that ${A}x=\sum_{i=1}^{N}A_{i}x_{i}$, ${A}^\top w=(A_1^\top w,\ldots,A_N^\top w)$.
The operator $D+M$ is maximally monotone \cite[Proposition 20.23, Corollary 24.4(i)]{bauschke2011convex}
Monotone inclusion~\eqref{eq:mon-inclusion}, \emph{i.e.}, the primal-dual optimality conditions~\eqref{eq:-1}, is solved by applying \cite[Algorithm 6]{AFBA2016}. This results in the following iteration:
\begin{subequations} \label{alg:5}
\begin{align}
{\boldsymbol{x}}^{k+1} & =\mathop{\rm prox}\nolimits^{\Sigma^{-1}}_{ f}(\boldsymbol{x}^{k}-\Sigma C^{\top}\boldsymbol{y}^k-\Sigma A^{\top}w^{k}) \label{eq:a}\\
\bar{\boldsymbol{y}}^{k} & =\mathop{\rm prox}\nolimits^{\Gamma^{-1}}_{ g^{*}}(\boldsymbol{y}^{k}+\Gamma C(\theta{\boldsymbol{x}}^{k+1}+(1-\theta){\boldsymbol{x}}^{k}))\\
\bar{w}^{k} & =w^{k}+\Pi A(\theta{\boldsymbol{x}}^{k+1}+(1-\theta){\boldsymbol{x}}^{k}) \label{eq:c}\\
\boldsymbol{y}^{k+1} & =\bar{\boldsymbol{y}}^{k}+(2-\theta)\Gamma C({\boldsymbol{x}}^{k+1}-{\boldsymbol{x}}^{k}) \label{eq:d}\\
w^{k+1} &= \bar{w}^k + (2-\theta)\Pi A({\boldsymbol{x}}^{k+1}-{\boldsymbol{x}}^{k}) \label{eq:e}
\end{align}
\end{subequations}
where matrices $\Sigma,\Gamma,\Pi$ play the rule of stepsizes and are assumed to be positive definite. The iteration~\eqref{alg:5} can not be implemented in a distributed fashion because the dual vector $w$ consists of $M$ blocks corresponding to the edges. The key idea that allows distributed computations is to introduce the sequence
\begin{equation}\label{eq:rho}
(\rho^k_{i})_{k\in{\rm{I\!N}}}=(A_{i}^{\top}w^{k})_{k\in{\rm{I\!N}}}, \quad \textrm{for} \quad i=1,\ldots,N.
\end{equation}
This transformation replaces the stacked edge vector $w^k$ with corresponding node vectors $\rho_i$. More compactly, letting $\boldsymbol{\rho}^k=(\rho^k_{1},\ldots,\rho^k_{N})$, it follows from~\eqref{eq:c} and \eqref{eq:e} that
\begin{equation}
\boldsymbol{\rho}^{k+1}=\boldsymbol{\rho}^{k}+ A^\top \Pi A(2{\boldsymbol{x}}^{k+1}-{\boldsymbol{x}}^{k}),
\end{equation}
where $A^\top \Pi A$ is the \emph{weighted graph Laplacian}. Since $w^k$ in~\eqref{eq:a} appear as $A^\top w^k$ we can rewrite the iteration:
\begin{align*}
{\boldsymbol{x}}^{k+1} & =\mathop{\rm prox}\nolimits^{\Sigma^{-1}}_{ f}(\boldsymbol{x}^{k}-\Sigma C^{\top}\boldsymbol{y}^k-\Sigma \boldsymbol{\rho}^{k})\\
\bar{\boldsymbol{y}}^{k} & =\mathop{\rm prox}\nolimits^{\Gamma^{-1}}_{ g^{*}}(\boldsymbol{y}^{k}+\Gamma C(\theta{\boldsymbol{x}}^{k+1}+(1-\theta){\boldsymbol{x}}^{k})) \\
\boldsymbol{y}^{k+1} & =\bar{\boldsymbol{y}}^{k}+(2-\theta)\Gamma C({\boldsymbol{x}}^{k+1}-{\boldsymbol{x}}^{k}) \\
\boldsymbol{\rho}^{k+1}&=\boldsymbol{\rho}^{k}+ A^\top \Pi A(2{\boldsymbol{x}}^{k+1}-{\boldsymbol{x}}^{k})
\end{align*}
Set
\begin{align*}
\Sigma&=\mathop{\rm blkdiag}\nolimits\left(\sigma_1 I_n,\ldots,\sigma_N I_n\right),\\
\Gamma&=\mathop{\rm blkdiag}\nolimits\left(\tau_1 I_{r_1},\ldots,\tau_N I_{r_N}\right),\\
\Pi&=\mathop{\rm blkdiag}\nolimits\left(\pi_1 I_n,\ldots,\pi_M I_n\right),
\end{align*}
where $\sigma_i>0,\tau_i>0$ for $i=1,\ldots,N$ and $\pi_l>0$ for $l=1,\ldots,M$. Consider a bijective mapping between $l=1,\ldots,M$ and unordered pairs $(i,j)\in E$ such that $\kappa_{i,j}=\kappa_{j,i}=\pi_l$.
Notice that $\pi_l$ for $l=1,\ldots,M$ are step sizes to be selected by the algorithm and can be viewed as weights for the edges. Thus, iteration~\eqref{alg:5} gives rise to our distributed algorithm:
\begin{algorithm}[H]
\caption{
}
\label{Algorithm-5}
\begin{algorithmic}
\item\textbf{Inputs:} $\sigma_i>0$, $\tau_i>0$, $\kappa_{i,j}>0$ for $j\in\mathcal{N}_{i}$, $i=1,\ldots,N$, $\theta\in[0,\infty[$, \textrm{initial values $x_i^0\in\Re^n$, $y_i^0\in\Re^{r_i}$, $\rho_i^0\in\Re^n$}.
\For{$k=1,\ldots$}
\For {each agent $i=1,\ldots,N$}
\item{\hspace{0.9cm} Local steps:}
\State ${x}_{i}^{k+1}=\mathop{\rm prox}\nolimits_{\sigma_{i}f_{i}}(x_{i}^{k}-\sigma_{i}\rho_{i}^{k}-\sigma_{i}C_i^\top y_{i}^{k})$
\State $\bar{y}_{i}^{k}=\mathop{\rm prox}\nolimits_{\tau_{i}g_{i}^{*}}(y_{i}^{k}+\tau_{i}C_{i}(\theta x_i^{k+1}+(1-\theta)x_i^{k}))$
\State $y_i^{k+1} = \bar{y}_{i}^{k} + \tau_i(2-\theta)C_i (x_i^{k+1}-x_i^k)$
\State $u_i^k=2x_i^{k+1}-x_i^k$
\item {\hspace{0.9cm} Exchange of information with neighbors:}
\State $\rho_{i}^{k+1}=\rho_{i}^{k}+\sum_{j\in\mathcal{N}_{i}}\kappa_{i,j}(u_{i}^{k}-u_{j}^{k})$
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
Notice that each agent $i$ only requires $u_j^k\in\Re^n$ for $j\in\mathcal{N}_{i}$ during the communication phase.
Before proceeding with convergence results, we define the following for simplicity of notation:
\begin{equation*}
\begin{split}
\bar{\sigma}&=\max\{\sigma_1,\ldots,\sigma_N\},\\
\bar{\tau}&=\max\{\tau_1,\ldots,\tau_N,\pi_1,\ldots,\pi_M\},\\
L&= \mathcal{L}\otimes I_n+\boldsymbol{C}^\top \boldsymbol{C},\ \textrm{where }\mathcal{L}\ \textrm{is the graph Laplacian.}
\end{split}
\end{equation*}
It must be noted that the results in this section only provide choices of parameters that are sufficient for convergence. They can be selected much less conservatively by formulating and solving sufficient conditions that they must satisfy as \emph{linear matrix inequalities} (LMIs). Due to lack of space we do not pursue this direction further, instead we plan to consider it in an extended version of this work.
\begin{thm} \label{thm:Algorithm-5}
Let \Cref{ass:conctd,ass:exst-prmdul} hold true. Consider the sequence $(\boldsymbol{x}^{k})_{k\in{\rm{I\!N}}}=(x_1^k,\ldots,x_N^k)_{k\in{\rm{I\!N}}}$ generated by \Cref{Algorithm-5}. Assume the maximum stepsizes, \emph{i.e.}, $\bar{\sigma}$ and $\bar{\tau}$ defined above, are positive and satisfy
\begin{equation} \label{eq:alg5}
\bar{\sigma}^{-1}-\bar{\tau}(\theta^2-3\theta+3)\|L\|>0,
\end{equation}
for a fixed value of $\theta\in[0,\infty[$.
Then
the sequence $(\boldsymbol{x}^{k})_{k\in{\rm{I\!N}}}$ converges to $(x^{\star},\ldots,x^{\star})$ for some $x^\star\in S^\star$. Furthermore, if $\theta=2$ the strict inequality~\eqref{eq:alg5} is replaced with $\bar{\sigma}^{-1}-\bar{\tau}\|L\|\geq0$.
\end{thm}
\begin{proof}
\Cref{Algorithm-5} is an implementation of \cite[Algorithm 6]{AFBA2016}.
Thus convergence of $(\bld{x}^k)_{k\in{\rm{I\!N}}}$ to a solution of~\eqref{eq:primal} is implied by \cite[Proposition 5.4]{AFBA2016}. Combining this with \Cref{ass:conctd} yields the result.
Notice that in that work the step sizes are assumed to be scalars for simplicity. It is straightforward to adapt the result to the case of diagonal matrices.
\end{proof}
In \Cref{Algorithm-5} when $\theta=2$, we recover the algorithm of Chambolle and Pock \cite{chambolle2011first}.
One important observation is that the term $\theta^2-3\theta+3$ in~\eqref{eq:alg5} is always positive and achieves its minimum at $\theta=1.5$. This is a choice of interest for us since it results in larger stepsizes, ${\sigma}_i,\tau_i, \kappa_{i,j}$, and consequently better performance as we observe in numerical simulations.
Next, we provide easily verifiable conditions for $f_i$ and $g_i$, under which linear convergence of the iterates can be established. We remark that these are just sufficient and certainly less conservative conditions can be provided but are omitted for clarity of exposition. Let us first recall the following definitions from \cite{rockafellar2009variational,dontchev2009implicit}:
\begin{deff}[Piecewise Linear-Quadratic]
A function $f:\Re^n\to\overline{\rm I\!R}$ is called piecewise linear-quadratic (PLQ) if it's domain can be represented as union of finitely many polyhedral sets, relative to each of which $f(x)$ is given by an expression of the form $\frac{1}{2}x^\top Qx+d^\top x+c$,
for some $c\in\Re$, $d\in\Re^n$, and $D\in\Re^{n\times n}$.
\end{deff}
The class of piecewise linear-quadratic functions has been much studied and has many desirable properties (see \cite[Chapter 10 and 11]{rockafellar2009variational}).
Many practical applications involve PLQ functions such as quadratic function, $\|\cdot\|_1$, indicator of polyhedral sets, hinge loss, \emph{etc}. Thus, the $R$-linear convergence rate that we establish in \Cref{thm:conv-rate} holds for a wide range of problems encountered in control, machine learning and signal processing.
\begin{deff}[Metric subregularity] \label{metricsub}
A set-valued mapping $F:\Re^n\rightrightarrows\Re^n$ is metrically subregular at $z$ for $z^\prime$ if $(z,z^\prime)\in\mathop{\rm gra}\nolimits F$ and there exists $\eta\in[0,\infty[$,
a neighborhood $\mathcal{U}$ of $z$ and $\mathcal{V}$ of $z^\prime$ such that
\begin{equation}\label{eqnn:metricsubregularity}
d(x,F^{-1}z^\prime)\leq \eta d(z^\prime,Fx\cap \mathcal{V}) \;\; \textrm{for all} \; x\in \mathcal{U}
\end{equation}
where $\mathop{\rm gra}\nolimits F=\{(x,u)|u\in Fx\}$ and $d(\cdot,X)$ denotes the distance from set $X$.
\end{deff}
\begin{thm}\label{thm:conv-rate}
Consider \Cref{Algorithm-5} under the assumptions of \Cref{thm:Algorithm-5}. Assume $f_i$ and $g_i$ for $i=1,\ldots,N$, are piecewise linear-quadratic functions. Then the set valued mapping $T=D+M$ is \emph{metrically subregular} at any $z$ for any $z^\prime$ provided that $(z,z^\prime)\in\mathop{\rm gra}\nolimits T$.
Furthermore, the sequence $(\boldsymbol{x}^k)_{k\in{\rm{I\!N}}}$ converges R-linearly\footnote{The sequence $(x_{n})_{n\in{\rm{I\!N}}}$ converges to $x^{\star}$ $R$-linearly if there is a sequence of nonnegative scalars $({v_n})_{n\in{\rm{I\!N}}}$ such that $\|x_{n}-x^{\star}\|\leq v_n$ and $(v_n)_{n\in{\rm{I\!N}}}$ converges $Q$-linearly\footnotemark{} to zero.} \footnotetext{The sequence $(x_{n})_{n\in{\rm{I\!N}}}$ converges to $x^{\star}$ $Q$-linearly with $Q$-factor given by $\sigma\in]0,1[$, if for $n$ sufficiently large $\|x_{n+1}-x^{\star}\|\leq\sigma\|x_{n}-x^{\star}\|$ holds.}to $(x^{\star},\ldots,x^{\star})$ for some $x^\star\in S^\star$.
\end{thm}
\begin{proof}
Function $\boldsymbol{f}(\boldsymbol{x})=\sum_{i=1}^{N}f_{i}(x_{i})$ is piecewise linear-quadratic since $f_i$ for $i=1,\ldots,N$, are assumed to be PLQ. Similarly, it follows from \cite[Theorem 11.14 (b)]{rockafellar2009variational} that $\boldsymbol{g}^*$ is piecewise linear-quadratic. The subgradient mapping of a proper closed convex PLQ function is piecewise polyhedral, \emph{i.e.} its graph is the union of finitely many polyhedral sets \cite[Proposition 12.30 (b)]{rockafellar2009variational}. This shows that $D$ defined in~\eqref{eq:Aopt} is piecewise polyhedral. Since the image of a polyhedral under affine transformation remains piecewise polyhedral, and $M$ is a linear operator, graph of $T=D+M$ is piecewise polyhedral. Consequently, its inverse $T^{-1}$ is piecewise polyhedral. Thus by \cite[Proposition 3H.1]{dontchev2009implicit} the mapping $T^{-1}$ is calm at any $z^\prime$ for any $z$ satisfying $(z^\prime,z)\in \mathop{\rm gra}\nolimits T^{-1}$. This is equivalent to the metric subregularity characterization of the operator $T$ \cite[Theorem 3H.3]{dontchev2009implicit}. The second part of the proof follows directly by noting that \cite[Algorithm 6]{AFBA2016} used to derive \Cref{Algorithm-5} is a special case of \cite[Algorithm 1]{AFBA2016}. Therefore, linear convergence follows from first part of the proof together with \cite[Theorem 3.3]{AFBA2016}. The aforementioned theorem guarantees linear convergence for the stacked vector $\boldsymbol{u}$ in \eqref{eq:mon-inclusion}, however, here we consider the primal variables only.
\end{proof}
\subsection{Special Case} Consider the following problem
\begin{equation}\label{eq:specialprob2}
\underset{x\in\Re^n}{\minimize} \ \sum_{i=1}^{N} f_i(x),
\end{equation}
where $f_i:\Re^n\to\overline{\rm I\!R}$ for $i=1,\ldots,N$ are proper closed convex functions. This is a special case of~\eqref{eq:Problem} when $g_i\circ C_i\equiv0$. Since functions $g_i$ are absent, the dual variables $y_i$ in \Cref{Algorithm-5} vanish and for any choice of $\theta$ the algorithm reduces to:
\begin{align*}
x_i^{k+1}&=\mathop{\rm prox}\nolimits_{\sigma_{i}f_{i}}(x_{i}^{k}-\sigma_{i}\rho_{i}^{k})\\
u_i^k&= 2x_i^{k+1}-x_i^k\\
\rho_{i}^{k+1}&=\rho_{i}^{k}+\sum_{j\in\mathcal{N}_{i}}\kappa_{i,j}({u}_{i}^{k}-{u}_{j}^{k}).
\end{align*}
Thus setting $\theta=1.5$ in \eqref{eq:alg5} to maximize the stepsizes yields $\bar{\sigma}^{-1}-\frac{3\bar{\tau}}{4}\|\mathcal{L}\|>0$, where $\mathcal{L}$ is the graph Laplacian.
\iffalse
\begin{algorithm}[H]
\caption{
}
\label{Algorithm-5}
\begin{algorithmic}
\item\textbf{Inputs:} $\sigma_i>0$, $\tau_i>0$, $\kappa_{i,j}>0$ for $j\in\mathcal{N}_{i}$, $i=1,\ldots,N$, \textrm{initial values $x_i^0\in\Re^n$, $y_i^0\in\Re^{r_i}$.}.
\For{$k=1,\ldots$}
\For {each agent $i=1,\ldots,N$}
\item{\hspace{0.9cm} Local steps:}
\State $\bar{x}_{i}^{k}=\mathop{\rm prox}\nolimits_{\sigma_{i}f_{i}}(x_{i}^{k}-\sigma_{i}\rho_{i}^{k})$
\item {\hspace{0.9cm} Exchange of information with neighbors:}
\State $\rho_{i}^{k+1}=\rho_{i}^{k}+\sum_{j\in\mathcal{N}_{i}}\kappa_{i,j}(\bar{x}_{i}^{k}-\bar{x}_{j}^{k})$
\item{\hspace{0.9cm} Local steps:}
\State $x_{i}^{k+1}=\bar{x}_{i}^{k}-\sigma_{i}(\rho_i^{k+1}-\rho_i^{k})$
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
\begin{lem}
\label{lem:strongly positive p-1}Consider $P$ defined by \eqref{eq:P}.
Assume $\bar{\sigma}^{-1}-\frac{\bar{\tau}}{4}\theta^{2}\|L\|>0$
holds. Then $P$ is symmetric positive definite.
\end{lem}
\begin{proof}
It follows from Schur-complement.
\end{proof}
\begin{lem} \label{lem: cocoercivity lemma}
Let $P$ be defined by \eqref{eq:P}. Then $F$ given by~\eqref{eq:Copt} is $\beta_F$-cocoercive with respect to the $P$ norm, where
$\beta_F=\beta^{-1}\left(\bar{\sigma}^{-1}-\frac{\bar{\tau}}{4}\theta^{2}\|L\|\right)$.
\end{lem}
\begin{proof}
Define the linear
operator $Q:(\boldsymbol{x},\boldsymbol{v})\rightarrow(\boldsymbol{x},0)$, then,
\begin{align*}
\|Q\boldsymbol{u}\|_{P^{-1}}^{2} & =\langle \boldsymbol{x},(\Sigma^{-1}-\tfrac{1}{4}\theta^{2}C^\top T C-\tfrac{1}{4}\theta^{2}A^\top \Pi A)^{-1}\boldsymbol{x}\rangle\\
& \leq\|\boldsymbol{x}\|^{2}\|(\Sigma^{-1}-\tfrac{1}{4}\theta^{2}C^\top T C-\tfrac{1}{4}\theta^{2}A^\top \Pi A)^{-1}\|
\\
& \leq\|Q\boldsymbol{u}\|^{2}(\bar{\sigma}^{-1}-\tfrac{\bar{\tau}}{4}\theta^{2}\|L\|)^{-1},
\end{align*}
where the last inequality follows from \cite[Lemma 3.1(iii)]{combettes2014variable}. Considering the above inequality and~\eqref{eq:Copt}, we obtain for all $\bld{u}=(\bld{x},\bld{v})$, $\bld{u}'=(\bld{x}',\bld{v}')$
\begin{align*}
\|F\bld{u}-F\bld{u}^{\prime}\|_{P^{-1}}^{2} & \leq\eta\|\nabla h(\bld{x})-\nabla h(\bld{x}^{\prime})\|^{2}\\
& \leq\beta\eta\langle \nabla h(\bld{x})-\nabla h(\bld{x}^{\prime}),\bld{x}-\bld{x}^\prime \rangle\\
& =\beta\eta\langle F\bld{u}-F\bld{u}^{\prime},\bld{u}-\bld{u}^{\prime} \rangle,
\end{align*} \normalsize
where $\eta=(\bar{\sigma}^{-1}-\frac{\bar{\tau}}{4}\theta^{2}\|L\|)^{-1}$ and the last inequality follows from the Baillon-Haddad theorem \cite[Theorem 1.1]{bauschke2009baillon}.
\end{proof}
Next we turn our attention to another distributed primal-dual algorithm for solving~\eqref{eq:Problem}.
The reasoning and the change of variable~\eqref{eq:rho} can be used to derive a distributed algorithm that is based on the V\~u-Condat algorithm \cite{vu2013splitting,condat2013primal}. We use the convergence results of \cite[Proposition 5.1]{AFBA2016}. They are less conservative than the ones of \cite[Theorem 3.1]{vu2013splitting} and \cite[Theorem 3.1]{condat2013primal} that are based on theory of averaged operators (cf. \cite[Section 5.1 and Remark 5.7]{AFBA2016}).
\begin{algorithm}[H]
\caption{
}
\label{Algorithm-Vu-Condat}
\begin{algorithmic}
\item\textbf{Inputs:} $\sigma_i>0$, $\tau_i>0$, $\kappa_{i,j}>0$ for $i=1,\ldots,N$, $j\in\mathcal{N}_{i}$ \textrm{and the initial values $x_i^0\in\Re^n$, $y_i^0\in\Re^{r_i}$, $\rho_i^0\in\Re^n$}.
\For{$k=1,\ldots$}
\For {each agent $i=1,\ldots,N$}
\item{\hspace{0.9cm} Local steps:}
\State $\bar{x}_{i}^{k}=\mathop{\rm prox}\nolimits_{\sigma_{i}f_{i}}(x_{i}^{k}-\sigma_{i}(C_{i}^{\top}y_{i}^{k}+\rho_{i}^{k})-\sigma_{i}\nabla h_{i}(x_{i}^{k}))$
\State $u_{i}^{k}=2\bar{x}_{i}^{k}-x_{i}^{k}$
\State $\bar{y}_{i}^{k}=\mathop{\rm prox}\nolimits_{\tau_{i}g_{i}^{*}}(y_{i}^{k}+\tau_{i}C_{i}u_{i}^{k})$
\item {\hspace{0.9cm} Exchange of information with neighbors:}
\State $\bar{\rho}_{i}^{k}=\rho_{i}^{k}+\sum_{j\in\mathcal{N}_{i}}\kappa_{i,j}(u_{i}^{k}-u_{j}^{k})$
\item{\hspace{0.9cm} Local steps:}
\State $x_{i}^{k+1}=\lambda_{k}\bar{x}_{i}^{k}+(1-\lambda_{k})x_{i}^{k}$
\State $y_{i}^{k+1}=\lambda_{k}\bar{y}_{i}^{k}+(1-\lambda_{k})y_{i}^{k}$
\State $\rho_{i}^{k+1}=\lambda_{k}\bar{\rho}_{i}^{k}+(1-\lambda_{k})\rho_{i}^{k}$
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
Note that in this algorithm during the exchange of information each agent $i$, requires ${u}_j^k=2\bar{x}_j^k-x_j^k\in\Re^n$ for $j\in\mathcal{N}_{i}$. The proof of next theorem is similar to that of \Cref{thm:Algorithm-5} and follows by combining \Cref{ass:conctd} and \cite[Proposition 5.1(1-ii) and (2-i)]{AFBA2016}.
\begin{thm} \label{thm:vu-condat}
Let \Cref{ass:conctd,ass:exst-prmdul} hold true. Assume the parameters of \Cref{Algorithm-Vu-Condat} satisfy the following:
\begin{enumerate}
\item $\bar{\sigma}^{-1}-\bar{\tau}\|L\|>\frac{\beta}{4}$.
\item $(\lambda_{k})_{k\in{\rm{I\!N}}}\subseteq[\epsilon,\delta-\epsilon]$, for some $\epsilon>0$, and $\delta=2-\tfrac{\beta}{2}\left(\bar{\sigma}^{-1}-\bar{\tau}\|L\|\right)^{-1}$.
\end{enumerate}
Then there exists a solution $x^{\star}$ to~\eqref{eq:Problem}, such that the sequence $(\boldsymbol{x}_{k})_{k\in{\rm{I\!N}}}$ converges to $(x^{\star},\ldots,x^{\star})$. Furthermore, $\|\boldsymbol{x}^{k}-\boldsymbol{\bar{x}}^{k}\|^{2}=o(1/(k+1))$
\end{thm}
Next distributed algorithm is based on \cite[Algorithm 4]{AFBA2016}. Using the parameters defined in this section and algebraic manipulation similar to \Cref{Algorithm-5} we derive the following:
\begin{algorithm}[H]
\caption{
}
\label{Algorithm-4}
\begin{algorithmic}
\item\textbf{Inputs:} $\sigma_i>0$, $\tau_i>0$, $\kappa_{i,j}>0$ for $i=1,\ldots,N$, $j\in\mathcal{N}_{i}$ \textrm{and the initial values $x_i^0\in\Re^n$, $y_i^0\in\Re^{r_i}$, $\rho_i^0\in\Re^n$}.
\For{$k=1,\ldots$}
\For {each agent $i=1,\ldots,N$}
\item{\hspace{0.9cm} Local steps:}
\State $\bar{x}_{i}^{k}=\mathop{\rm prox}\nolimits_{\sigma_{i}f_{i}}(x_{i}^{k}-\sigma_{i}(C_{i}^{\top}y_{i}^{k}+\rho_{i}^{k})-\sigma_{i}\nabla h_{i}(x_{i}^{k}))$
\State $\bar{y}_{i}^{k}=\mathop{\rm prox}\nolimits_{\tau_{i}g_{i}^{*}}(y_{i}^{k}+\tau_{i}C_{i}x_{i}^{k})$
\State $\tilde{x}_{i}^{k}=\bar{x}_{i}^{k}-x_{i}^{k},\tilde{y}_{i}^{k}=\bar{y}_{i}^{k}-y_{i}^{k}$
\State $y_{i}^{k+1}=\bar{y}_{i}^{k}+\tau_{i}C_{i}\tilde{x}_{i}^{k}$
\item {\hspace{0.9cm} Exchange of information with neighbors:}
\State $x_{i}^{k+1}=\bar{x}_{i}^{k}-\sigma_{i}(C_{i}^{\top }\tilde{y}_{i}^k+\sum_{j\in\mathcal{N}_{i}}\kappa_{i,j}(x_{i}^{k}-x_{j}^{k}))$
\State $\rho_{i}^{k+1}=\rho_{i}^{k}+\sum_{j\in\mathcal{N}_{i}}\kappa_{i,j}(\bar{x}_{i}^{k}-\bar{x}_{j}^{k})$
\EndFor
\EndFor
\end{algorithmic}
\end{algorithm}
In this algorithm the local proximal steps for each agent can be performed in parallel given that $\bar{y}_i^k$ does not depend on $\bar{x}_i^k$.
The drawback of this algorithm is that each agent $i$, requires $\bar{x}_j^k\in\Re^n$ and $x_j^k\in\Re^n$ for $j\in\mathcal{N}_{i}$ during the communication phase which is double the transmission required by \Cref{Algorithm-5,Algorithm-Vu-Condat}. Furthermore, \Cref{Algorithm-4} requires $4$ matrix vector products in place of $2$. Next theorem provides convergence results based on \cite[Proposition 5.2(ii)]{AFBA2016} following the same reasoning used in the proof of \Cref{thm:Algorithm-5}. See \cite[Remark 5.8]{AFBA2016} for convergence rates.
\begin{thm} \label{thm:Algorithm-4}
Let \Cref{ass:conctd,ass:exst-prmdul} hold true. Assume the parameters of \Cref{Algorithm-4} satisfy
\begin{equation}\label{eq:alg4}
\beta\bar{\sigma}<2-2\bar{\sigma}\bar{\tau}\|L\|.
\end{equation}
Then there exists a solution $x^{\star}$ to~\eqref{eq:Problem}, such that the sequence $(\boldsymbol{x}^{k})_{k\in{\rm{I\!N}}}$ converges to $(x^{\star},\ldots,x^{\star})$. If in addition, \[\left(\bar{\sigma}^{-1}-\bar{\tau}\|L\|\right)^2\leq 2-\frac{\beta\bar{\sigma}}{2} \] holds, then $\|\boldsymbol{x}^{k}-\boldsymbol{\bar{x}}^{k}\|^{2}=o(1/(k+1))$.
\end{thm}
\fi
\section{Numerical Simulations}\label{sec:Simulations}
We now illustrate experimental results obtained by applying the proposed algorithm to the following problem:
\begin{equation}\label{eq:Example}\begin{split}
\minimize \ & \lambda\|x\|_1 + \sum_{i=1}^N \tfrac{1}{2}\|D_ix - d_i\|_2^2
\end{split}\end{equation}
for a positive parameter $\lambda$. This is the ${\ell}_1$ regularized least-squares problem. Problem \eqref{eq:Example} is of the form \eqref{eq:Problem} if we set for $i=1,\ldots,N$
\begin{equation}\begin{split}
f_i(x) &= \tfrac{\lambda}{N}\|x\|_1,\\
g_i(z) &= \tfrac{1}{2}\|z - d_i\|_2^2,\\
C_i &= D_i
\end{split}\end{equation}
where $D_i\in\Re^{m_i\times n}$, $d_i\in\Re^{m_i}$.
For the experiments we used graphs of $N=50$ computing agents, generated randomly according to the Erd\H{o}s-Renyi model, with parameter $p=0.05$.
In the experiments we used $n=500$ and generated
$D_i$ randomly with normally distributed entries,
with $m_i = 50$ for all $i=1,\ldots,N$. Then we generated vector $d_i$
starting from a known solution for the problem and ensuring {
$\lambda<0.1 \|\sum_i^N D_i^\top d_i\|_{\infty}$}.
For the stepsize parameters we set
$\sigma_i = \bar{\sigma}$, $\tau_i = \bar{\tau}$, for all $i=1,\dots,N$, and $\kappa_{i,j} = \kappa_{j,i} = \bar{\tau}$ for all edges $(i,j)\in E$, such that \eqref{eq:alg5} is satisfied. In order to have a fair comparison we selected $\bar{\sigma}=\alpha/\|L\|$ and $\bar{\tau}=0.99/(\alpha(\theta^2-3\theta+3))$ with $\alpha=20$ which was set empirically based on better performance of all the algorithms.
The results are illustrated in \Cref{fig1}, for several values of $\theta$, where the distribution of the number of communication rounds required by the algorithms to reach a relative error of $10^{-6}$ is reported. In \Cref{fig2} the convergence of algorithms is illustrated in one of the instances. It
should be noted that the algorithm of Chambolle and Pock, that corresponds to $\theta=2$, is generally slower than the case $\theta=1.5$. This is mainly due to the larger stepsize parameters guaranteed by \Cref{thm:Algorithm-5}.
\begin{figure}
\begin{minipage}{\columnwidth}
\hspace*{0.2cm}{\centering\includegraphics{fig3.pdf}}
\end{minipage}
\caption{Distribution of the number of communication rounds required by the algorithms to achieve a relative error of $10^{-6}$, for fixed data and $200$ randomly generated Erd\H{o}s-Renyi graphs, with parameter $p=0.05$.}
\label{fig1}
\end{figure}
\input{figures/fig2}
\section{Conclusions}
In this paper we illustrated how the recently proposed Asymmetric Forward-Backward-Adjoint
splitting method (AFBA) can be used for
solving distributed optimization problems where a set of $N$ computing agents, connected
in a graph, need to minimize the sum of $N$ functions. The resulting
\Cref{Algorithm-5} only involves
local exchange of variables (i.e., among neighboring nodes) and therefore no central
authority is required to coordinate them. Moreover, the single nodes only require
direct computations of the objective terms, and do not need to perform inner iterations
and matrix inversions. Numerical experiments highlight that \Cref{Algorithm-5} performs generally better
when the parameter $\theta$ is set equal to $1.5$ in order to achieve the largest stepsizes.
Future investigations on this topic include the study of how the topological structure
of the graph underlying the problem affects the convergence rate of the proposed methods, as well as problem preconditioning in a distributed fashion. Developing asynchronous versions of the algorithm
is another important future research direction.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
According to the predictions of general relativity, binary systems of
compact objects are the most efficient emitters of gravitational waves
(GWs). Indeed, Advanced LIGO has recently detected the GW signal from a
black-hole (BH) binary with masses $M_1\approx36\,M_\odot$ and
$M_2\approx29\,M_\odot$~\citep{Abbott:2016blz}, at a (luminosity)
distance of $\sim 410$ Mpc. In general, Advanced LIGO and other
terrestrial interferometers, such as Advanced Virgo and KAGRA, target BH
binaries with a variety of masses~\citep[up to a few hundred $M_\odot$,
if they exist;][]{2014ApJ...789..120B,2016ApJ...819..108B}. More
massive BH binaries are targeted by existing pulsar-timing arrays
\citep[in the mass range $10^8-10^{10}\,M_\odot$;][]{2013CQGra..30v4010M}
and by future space-borne interferometers such as eLISA \citep[in the
mass range $10^4-10^7\,M_\odot$;][]{elisa}.
One of the obvious difficulties of observing BH binaries with terrestrial
interferometers is that only the final part of the inspiral and the
merger/ringdown are in band. This is where the perturbative
post-Newtonian (PN) techniques valid earlier in the inspiral become
inaccurate, preventing the extraction the source's physical
parameters. Hence, to obtain the full gravitational waveforms, it is
necessary to resort to numerical-relativity (NR) simulations. In
practice, even under the reasonable assumption that BH binaries near the
merger have been circularized by earlier GW emission, the space of
parameters to be probed (the mass ratio $q$ and the spin vectors
$\boldsymbol{S}_1,~\boldsymbol{S}_2$, \textit{i.e.,}~seven parameters) is too large
to be handled by NR simulations alone.
To ensure a sufficient coverage of the parameter space, semi-analytical
techniques allowing faster waveform production are employed, \textit{e.g.,}~the
spin-effective-one-body (sEOB)
model~\citep{Buonanno:1998gg,2001PhRvD..64l4013D,Barausse:2009xi} or
``hybrid'' waveforms~\citep{2008PhRvD..77j4017A,2015arXiv150807253K},
which combine results from NR simulations with PN and quasinormal-mode
calculations. These techniques are faster, but require great care when
modeling the merger and the transition to the ringdown. Indeed, although
the ringdown can be modeled via a linear superposition of quasi-normal
modes, their frequencies depend on the remnant BH's mass and
spin, which, in turn, depend on the initial binary parameters.
This relation between the binary's initial and final states is highly
non-trivial because it encodes the details of the strong-field, highly
relativistic merger, which is only accessible via NR calculations. Yet, a
number of approaches to predict analytically or semi-analytically the
remnant's final spin magnitude and direction have been proposed. These
range from modeling the GW fluxes throughout the binary's evolution
within the EOB model~\citep[\textit{e.g.,}~by][for nonspinning
BHs]{2007PhRvD..76d4003D} to approaches that combine information from
PN theory, the extreme mass-ratio limit (EMRL), symmetry arguments, and
fits to NR data, to provide ``formulae'' for the final
spin~\citep{Rezzolla:2007a,kesden_can_2008,Rezzolla:2007b,tichy_final_2008,Rezzolla:2007c,buonanno_estimating_2008,barausse_predicting_2009,healy_remnant_2014}. Similar
formulae have also been derived for the remnant's final
mass~\citep{tichy_final_2008,kesden_can_2008,barausse_mass_2012,healy_remnant_2014},
which differs from the binary's total mass by the energy emitted in GWs.
Again, a common problem in these attempts is the difficulty to cover with
sufficient accuracy the seven-dimensional parameter space of
quasi-circular BH binaries. Indeed, while most of these formulae formally
cover the whole parameter space, they can be rather inaccurate,
especially for BHs with almost extremal spins.
By combining results from NR and information from the EMRL and PN theory,
we here derive a new formula for the spin magnitude and direction for the
merger remnant from quasi-circular BH binaries with arbitrary masses and
spins. We calibrate our formula against a catalog of 619
recently published NR
simulations~\citep{chu_high_2009,hannam_simulations_2010,nakano_intermediate-mass-ratio_2011,sperhake_numerical_2011,pollney_gravitational_2011,kelly_mergers_2011,buchman_simulations_2012,lovelace_high-accuracy_2012,hemberger_final_2013,hinder_error-analysis_2013,kelly_decoding_2013,pekowsky_comparing_2013,healy_remnant_2014,Lousto_2014b,lovelace_nearly_2015,scheel_improved_2015,szilagyi_approaching_2015,zlochower_modeling_2015,husa_frequency-domain_2016,SXS}\footnote{Note
that for the simulations of \citet{zlochower_modeling_2015}, we only
consider the horizon-extracted data, and \textit{not} the
radiation-based ones, which may be imprecise~\citep{lousto_private}.},
and validate it by comparing its results to self-force calculations and
plunge-merger-ringdown fluxes for nonspinning binaries with small mass
ratios, as well as to a set of 248 NR simulations not included in the
calibration dataset~\citep{gatech_2016}.
Our new formula builds upon \citet{barausse_predicting_2009}, who
introduced a final-spin formula that is widely used both in the
production of semi-analytical waveforms (\textit{e.g.,}~in sEOB and phenomenological
waveforms) and in cosmological studies of massive BH
evolution~\citep[see, \textit{e.g.,}][]{bertispin,fanidakis,baraussespin,sikora,dubois,Sesana:2014bea}. Our novel prescription
especially improves the accuracy of the formula by
\citet{barausse_predicting_2009} for extreme mass ratios and for
near-extremal spins. This is important since near-extremal spins are
expected, at least in some cases, for supermassive BHs
\citep{bertispin,fanidakis,baraussespin,sikora,dubois,Sesana:2014bea} and
possibly also for stellar-mass BHs~\citep{2011CQGra..28k4009M}. We assume
$G=1=c$ hereafter.
\section{Modeling the final spin}
Let us first consider a BH binary with spins parallel (\textit{i.e.,}~aligned or
anti-aligned) to the orbital angular momentum $\boldsymbol{L}$, and
denote the masses by $M_{1,2}$ (with $q\equiv M_2/M_1\leq1$) and the spin projections on the
angular-momentum direction by $S_{1,2}\equiv a_{1,2}\,M_{1,2}^2$
($a_{1,2}$ being the dimensionless spin-parameter projections). In the
EMRL $q\ll 1$, the final-spin projection on the
angular-momentum direction must be
\begin{equation}\label{emrl}
a_{\rm fin}=a_1+\nu\left(L_{_{\rm ISCO}}(a_1)-2a_{1}E_{_{\rm ISCO}}(a_1)\right)+\mathcal{O}(\nu^2)\,,
\end{equation}
with $\nu\equiv q/(1+q)^2$ the symmetric mass ratio, and $L_{_{\rm
ISCO}}(a)$, $E_{_{\rm ISCO}}(a)$, respectively the specific (dimensionless) angular
momentum and energy for a test particle at the innermost stable circular
orbit (ISCO) of a Kerr BH with spin parameter $a$~\citep{Bardeen:1972fi}
\begin{eqnarray}
\label{eisco}
&&{E}_{_{\rm ISCO}}(a)=\sqrt{1-\frac{2}{3{r}_{_{\rm ISCO}}(a)}}\,,\\
&&{L}_{_{\rm ISCO}}(a)=\frac{2}{3\sqrt{3}}\left[1+2\sqrt{3{r}_{_{\rm ISCO}}(a)-2}\right]\,,
\label{lisco}
\end{eqnarray}
\begin{eqnarray}
&&{r}_{_{\rm ISCO}}(a)=3+Z_{2}-\frac{a}{|a|}\sqrt{(3-Z_{1})(3+Z_{1}+2Z_{2})}\,,\\
&&Z_{1}=1+(1-a^2)^{1/3}\left[(1+{a})^{1/3}+(1-{a})^{1/3}\right]\,,\\
&&Z_{2}=\sqrt{3{a}^{2}+Z_{1}^{2}}\,.
\end{eqnarray}
The final-spin expression of \citet{Rezzolla:2007c} and
\citet{barausse_predicting_2009} reproduces Eq.~\eqref{emrl} only in the
special case $a_1=0$, when $a_{\rm
fin}=2\sqrt{3}\nu+\mathcal{O}(\nu^2)$. Indeed, one of the drawbacks of
those early expressions is that they may yield spins $a_{\rm fin}>1$ for
small mass ratios $\nu\ll 1$, in clear disagreement with
Eq.~\eqref{emrl}, which predicts $a_{\rm fin}\leq 1$, the equality
holding for $a_1=1$.
To enforce the EMRL exactly, we consider the following ansatz for the
final-spin projection:
\begin{multline}
a_{\rm fin}={a}_{\rm tot}+\nu[L_{_{\rm ISCO}}({a}_{\rm eff})-2{a}_{\rm tot}(E_{_{\rm
ISCO}}({a}_{\rm eff})-1)]\\
+\sum_{i=0}^{n_{_M}}\sum_{j=0}^{n_{_J}}k_{ij}\nu^{2+i}\,{a}_{\rm eff}^j\,,
\label{new_ansatz}
\end{multline}
where $k_{ij}$ are free coefficients to be determined from the NR data,
${a}_{\rm tot}\equiv(S_1+S_2)/(M_1+M_2)^2=({a_1}+{a_2}q^2)/(1+q)^2$ is
the ``total'' spin parameter used in \citet{barausse_mass_2012}, while
$a_{\rm eff}\equiv S_{\rm eff}/(M_1+M_2)^2$ is an ``effective'' spin
parameter. In more detail, we assume $S_{\rm eff}=(1 + \xi M_2/M_1) S_1 +
(1 + \xi M_1/M_2) S_2$, which yields $a_{\rm eff}=a_{\rm tot}+\xi \nu
(a_1+a_2)$. This choice is inspired by \citet{2001PhRvD..64l4013D}, who
finds that the leading-order conservative spin-orbit dynamics depends on
the spin only through $S_{\rm eff}$ with $\xi=3/4$, while the
leading-order conservative spin-spin dynamics depends on $S_{\rm eff}$
with $\xi=1$ \citep[see also][]{racine,kesden1,kesden2}. In the
following, we will keep $\xi$ as a free parameter and determine it from
the NR simulations.\footnote{Setting $\xi=3/4$ or $\xi=1$ yields a much
larger reduced $\chi^2$ (see below for how we compute it). For
$n_{_M}=1,\,n_{_J}=2~(n_{_M}=3,\,n_{_J}=4)$ we obtain $\chi^2_{\rm red}
\approx5~(1.4)$ for $\xi=3/4$, and $\chi^2_{\rm red}\approx51$ (10) for
$\xi=1$. This strong statistical evidence that $\xi\neq3/4,\,1$ is not
surprising, as one indeed expects the leading-order spin-orbit and
spin-spin couplings to be ``deformed'' for highly relativistic
binaries~\citep{Barausse:2009xi}.}
Note that Eq.~\eqref{new_ansatz} matches Eq.~\eqref{emrl} for $\nu\ll1$,
since $a_{\rm tot}=a_1(1-2\nu)+\mathcal{O}(\nu^2)$. Moreover, by singling
out ${a}_{\rm tot}$ as the first term in Eq.~\eqref{new_ansatz}, we have
isolated the ``direct'' contribution of the progenitor spins to the
remnant's spin. However, this does not mean that all leading-order
effects of the smaller BH's spin $a_2$ are already included. For
instance, the specific energy and angular momentum at the ISCO receive
corrections of ${\cal O}(a_2\,\nu)$ \citep[see \textit{e.g.,}][]{Barausse:2009xi},
which propagate into a term of ${\cal O}(a_2\,\nu^2)$ in the final
spin, c.f. Eq.~\eqref{emrl}. This effect, together with other ones,
is captured by the coefficient $k_{01}$.
The coefficients $k_{0j}$ of the $\nu^2$ terms in Eq.~(\ref{new_ansatz})
also encode the information about the self-force dynamics (both
dissipative and conservative) and the leading-order (in mass ratio)
plunge-merger-ringdown emission. More specifically, the conservative
self-force produces shifts $\nu\,\Delta E_{_{\rm ISCO}}$ and
$\nu\,\Delta L_{_{\rm ISCO}}$ in the ISCO specific energy
and angular momentum away from the geodesic values of
Eqs.~(\ref{eisco}) and (\ref{lisco}). For a nonspinning binary
($a_1=a_2=0$) with $\nu\ll 1$
\begin{equation}
\Delta L_{_{\rm ISCO}}\approx-0.802\,.
\end{equation}
This follows from evaluating Eq. (3c) of \cite{2012PhRvL.108m1103L} at
the ISCO frequency, which should include conservative self-force effects
as in Eq. (5) of the same reference. The plunge-merger-ringdown
angular-momentum flux is instead given
by~\citep{2010PhRvD..81h4056B}
\begin{equation}
\Delta J_{_{\rm MR}}\approx3.46\,\nu^2\,.
\end{equation}
Therefore, for a nonspinning binary one expects
\begin{align}
a_{\rm fin}\approx&\,\nu L_{_{\rm ISCO}}(0)-2\nu^{2}[E_{_{\rm ISCO}}(0)-1]L_{_{\rm ISCO}}(0)\nonumber\\&+\Delta L_{_{\rm
ISCO}}\nu^2-\Delta J_{_{\rm MR}}+{\cal O}(\nu^3)\\\approx&\,2\sqrt{3}\nu-3.87\,\nu^2+{\cal O}(\nu^3)\,,\nonumber
\end{align}
hence $k_{00}\approx-3.87$. (Note that at ${\cal O}(\nu)$ and after
setting $a_1=a_2=0$, this equation reduces to Eq.~\eqref{emrl}.) However,
since the transition from inspiral to plunge does not happen exactly at
the ISCO when accounting for deviations from adiabaticity, but takes
place smoothly around the ISCO~\citep{2000PhRvD..62l4022O}, and since the
the plunge-merger-ringdown fluxes are intrinsically approximate (as it is
difficult to define unambiguously the plunge-merger-ringdown as separate
from the late inspiral), we keep $k_{00}$ as a free parameter. As it
happens, at least for $n_{_M}=1$, $n_{_J}=2$, the fitted value is
$k_{00}\approx-3.82$, which is reasonably close to the one predicted by
the considerations above\footnote{For the cases $n_{_M}=3$, $n_{_J}=3$
and $n_{_M}=3$, $n_{_J}=4$, also considered in the following,
$k_{00}\approx-5.9$. However, we will show that unlike $n_{_M}=1$,
$n_{_J}=2$, those cases are probably overfitting the data.}.
In principle, we could fit all the coefficients $k_{ij}$ (as well as
$\xi)$ to the NR results. However, since simulations for equal-mass
non-spinning BH binaries have determined the final remnant's spin with
accuracy far better than for other configurations, we impose that
Eq.~(\ref{new_ansatz}) with $q=1$ and $a_1=a_2=0$ yields exactly the
final spin $a_{\rm fin}=0.68646\pm0.00004$ measured by the NR simulations
of \citet{scheel_high-accuracy_2009}. This gives the relation
\begin{equation}
\frac{\sqrt{3}}{2}+\sum_{i=0}^{n_{_M}}\frac{k_{i0}}{4^{2+i}}=0.68646\pm0.00004\,.
\end{equation}
With this constraint, we fit Eq.~(\ref{new_ansatz}) to the 246
simulations for parallel-spin binaries in our calibration dataset.
\begin{table}
\begin{center}
\begin{tabular*}{\columnwidth}{cccccc}
\hline
\hline
\!\!\!$k_{01}$ & \!\!\!$k_{02}$ & \!\!\!$k_{10}$ & \!\!\!$k_{11}$ & \!\!\!$k_{12}$ &\!\!\! $\xi$ \\
\!\!\!$-1.2019$ & \!\!\!$-1.20764$ & \!\!\!$3.79245$ & \!\!\!$1.18385$ & \!\!\!$4.90494$& \!\!\!$0.41616$ \\
\hline
\hline
& & & & & \\
\hline
\hline
\!\!\!$k_{01}$ & \!\!\!$k_{02}$ & \!\!\!$k_{03}$ & \!\!\!$k_{10}$ & \!\!\!$k_{11}$ & \!\!\!$k_{12}$ \\
\!\!\!$2.87025$ & \!\!\!$ -1.53315$ & \!\!\!$ -3.78893$ & \!\!\!$ 32.9127$ & \!\!\!$ -62.9901$ & \!\!\!$ 10.0068$ \\
\hline
\!\!\!$k_{13}$ & \!\!\!$k_{20}$ & \!\!\!$k_{21}$ & \!\!\!$k_{22}$ & \!\!\!$k_{23}$ & \!\!\!$k_{30}$ \\
\!\!\!$ 56.1926$ & \!\!\!$-136.832$ & \!\!\!$ 329.32$ & \!\!\!$ -13.2034$ & \!\!\!$ -252.27$ & \!\!\!$ 210.075$ \\
\hline
\!\!\!$k_{31}$ & \!\!\!$k_{32}$ & \!\!\!$k_{33}$& \!\!\!$\xi$ \\
\!\!\!$ -545.35$ & \!\!\!$ -3.97509$ & \!\!\!$368.405$ & \!\!\!$ 0.463926$ \\
\hline
\hline
& & & & & \\
\hline
\hline
\!\!\!$k_{01} $ & \!\!\!$k_{02} $ &\!\!\! $k_{03} $ & \!\!\!$k_{04} $ & \!\!\!$k_{10} $ & \!\!\!$k_{11} $ \\
\!\!\!$3.39221$ &\!\!\!$4.48865$ &\!\!\!$-5.77101$ &\!\!\!$-13.0459$ &\!\!\!$35.1278$ &\!\!\! $-72.9336$ \\
\hline
\!\!\!$k_{12} $ & \!\!\!$k_{13} $ &\!\!\! $k_{14} $ &\!\!\! $k_{20} $ & \!\!\!$k_{21} $ &\!\!\! $k_{22} $ \\
\!\!\!$-86.0036$ &\!\!\!$93.7371$ &\!\!\!$200.975$ &\!\!\!$-146.822$ &\!\!\!$387.184$ &\!\!\!$447.009$ \\
\hline
\!\!\!$k_{23} $ & \!\!\!$k_{24} $ &\!\!\!$k_{30} $ & \!\!\!$k_{31} $ & \!\!\!$k_{32} $ & \!\!\!$k_{33} $ \\
\!\!\!$-467.383$ &\!\!\!$-884.339$ &\!\!\!$223.911$ &\!\!\!$-648.502$ &\!\!\!$-697.177$ &\!\!\!$753.738$ \\
\hline
\!\!\!$k_{34}$ &\!\!\!$\xi$ \\
\!\!\!$1166.89$ &\!\!\!$0.474046$ \\
\hline
\hline
\end{tabular*}
\caption{The coefficients of our formula, for $n_{_M}=1,~n_{_J}=2$ (top
block), $n_{_M}=3,~n_{_J}=3$ (middle block) and $n_{_M}=3,~n_{_J}=4$
(bottom block).}
\label{table_coeffs}
\end{center}
\end{table}
\begin{figure*}
\begin{center}
\includegraphics[width=\columnwidth]{PDF_different_set_of_coefficients.pdf}
\includegraphics[width=\columnwidth]{PDF_comparison_aligned.pdf}
\caption{Left panel: probability distribution functions (PDFs),
obtained as Gaussian fits, for the residuals of our formula with
increasing number of coefficients
(\textit{i.e.,}~$n_{_M}=1,\,n_{_J}=2;~n_{_M}=3,\,n_{_J}=3;~n_{_M}=3,\,n_{_J}=4$)
and for that of \citet{barausse_predicting_2009}. Right panel: same
as the left panel, but for our formula with $n_{_M}=3,\,n_{_J}=4$ and
for the formulae of \citet{barausse_predicting_2009},
\citet{healy_remnant_2014}, and \citet{husa_frequency-domain_2016};
the inset shows the actual distribution for our formula.}
\label{PDFaligned}
\end{center}
\end{figure*}
However, before performing the fit, it is useful to quantify the average
error of the final spins calculated from NR simulations. This is possible
because our calibration dataset contains simulations by different groups with
the same initial data. More precisely, 71 parallel-spin simulations have
one or more ``twins'', \textit{i.e.,}~binaries with exactly the same
initial properties, so that the \textit{mean} of the absolute differences
between twin NR simulations can be measured to be
$\delta a_{\rm fin}\approx0.002$. This estimate allows not only
performing a fit, but also computing its reduced chi-squared $\chi^2_{\rm
red}$, thus gauging whether we are overfitting the data, which would
correspond to $\chi^2_{\rm red}<1$.
\begin{table}
\label{table_goodness}
\begin{center}
\begin{tabular}{lcccc}
\hline
\hline
Model &coeffs. & $\mu$ & $\sigma$ & $\chi_{\mathrm{red}}^2$\\
\hline
$n_{_M}=1,\ n_{_J}=2$ & $\phantom{0}6$ & $-0.000215$ & $0.00198$ & $0.985$ \\
$n_{_M}=3,\ n_{_J}=3$ & $16$ & $-0.000066$ & $0.00168$ & $0.712$ \\
$n_{_M}=3,\ n_{_J}=4$ & $20$ & $-0.000029$ & $0.00166$ & $0.694$ \\
Barausse \& Rezzolla (2009) & $\phantom{0}4$ & $-0.002310$ & $0.00564$ & $9.313$ \\
Husa {et al.} (2016) & $11$ & $-0.000240$ & $0.00453$ & $5.150$ \\
Healy {et al.} (2014) & $19$ & $\phantom{-}0.000014$ & $0.00170$ & $0.718$ \\
\hline
\hline
\end{tabular}
\end{center}
\caption{The mean and rms ($\mu$ and $\sigma$) of the residuals
$a_{\mathrm{fin}}^{\mathrm{num}}-a_{\mathrm{fin}}^{\mathrm{fit}}$
from the numerical data, as well as $\chi^2_{\rm red}$, for our
formula and those of \citet{barausse_predicting_2009},
\citet{husa_frequency-domain_2016}, and \citet{healy_remnant_2014}; also
displayed is the number of coefficients in the various cases.}
\end{table}
Since Eq. \eqref{new_ansatz} can be expanded to arbitrary order via its
last term, we have performed fits of the parallel-spin calibration
dataset with $n_{_M}=1,\ n_{_J}=2$ (6 coefficients),
$n_{_M}=3,\ n_{_J}=3$ (16 coefficients) and $n_{_M}=3,\ n_{_J}=4$ (20
coefficients). The fitted coefficients are given in Table
\ref{table_coeffs}. Table \ref{table_goodness} reports the mean ($\mu$)
and root-mean-square (rms, $\sigma$) of the residuals from the NR data,
as well as $\chi^2_{\rm red}$, for the three aforementioned sets of
coefficients, and for the formulae of \citet{barausse_predicting_2009},
\citet{husa_frequency-domain_2016} and \citet{healy_remnant_2014} (which
use 4, 11 and 19 coefficients, respectively). Table \ref{table_goodness}
shows that our new formula converges when increasing the number of
coefficients, although the optimal choice to avoid overfitting appears to
be $n_{_M}=1,\ n_{_J}=2$. The convergence of our formula is also
displayed in the left panel of Fig.~\ref{PDFaligned}, which shows the
probability distribution functions (PDFs) of the residuals, obtained
as Gaussian fits. The right panel shows instead the (fitted)
PDFs for our formula (for $n_{_M}=3,\ n_{_J}=4$) and for the formulae of
\citet{barausse_predicting_2009}, \citet{husa_frequency-domain_2016} and
\citet{healy_remnant_2014}; the inset shows the actual residual
distribution for our formula. Note that already with 16 coefficients our
new formula has a slightly smaller rms than \citet{healy_remnant_2014},
with the important advantage that it can be used also for generic
binaries (see below), unlike the formulae of
\citet{husa_frequency-domain_2016} and \citet{healy_remnant_2014}.
To generalize Eq.~(\ref{new_ansatz}) to generic spins, we write the
remnant's spin as the total spin
$\boldsymbol{S}=\boldsymbol{S}_1+\boldsymbol{S}_2$ plus an angular
momentum contribution (\textit{i.e.,}~the angular momentum at the binary's
``effective'' ISCO), \textit{i.e.,}~$\boldsymbol{S}_{\rm
fin}=\boldsymbol{S}+\Delta\boldsymbol{L}$. Since the final mass is
$M_{\rm fin}=(M_1+M_2) (1-E_{\rm rad})$ \citep[with $E_{\rm rad}\lesssim
0.1$ the mass radiated in GWs;][]{barausse_mass_2012}, the final spin
parameter is
\begin{equation}\label{afin} \boldsymbol{a}_{\rm
fin}=\boldsymbol{a}_{\rm tot}+\boldsymbol{\ell}\nu\,,\quad
\boldsymbol{a}_{\rm
tot}=\frac{1}{(1+q)^2}\left(\boldsymbol{a}_1+\boldsymbol{a}_2
q^2\right)\,,
\end{equation}
where we have reabsorbed the radiated energy $E_{\rm rad}$ in
$\boldsymbol{\ell}\equiv\Delta\boldsymbol{L}/[M_{1}M_{2}(1-E_{\rm
rad})^2]+\boldsymbol{S}[2 E_{\rm rad}+3E_{\rm rad}^2+{\cal O}(E_{\rm
rad})^3]/(M_1 M_2)$ (note that $\boldsymbol{\ell}$ remains finite in
the test-particle limit because
$|\Delta\boldsymbol{L}|=\mathcal{O}(\nu)=E_{\rm rad}$ as $\nu\to0$). By
evaluating Eq.~(\ref{afin}) for parallel spins and comparing it to
(\ref{new_ansatz}), we obtain
\begin{multline}
|\boldsymbol{\ell}|=\Bigg\vert L_{_{\rm ISCO}}({a}_{\rm eff})-2{a}_{\rm
tot}\left(E_{_{\rm ISCO}}({a}_{\rm eff})-1\right)\\
+\sum_{i=0}^{n_{_M}}\sum_{j=0}^{n_S}k_{ij}\nu^{1+i}\,{a}_{\rm eff}^j\Bigg\vert\,,
\label{ell}
\end{multline}
which can be generalized to precessing spins by following
\citet{barausse_mass_2012} \citep[see
also][]{Rezzolla:2007c,barausse_predicting_2009} and replacing
\begin{align}
\label{atot}
a_{\rm tot}&\to a_{\rm
tot}(\beta,\gamma,q)\equiv\frac{|\boldsymbol{a}_1|\cos\beta+|\boldsymbol{a}_2|\cos\gamma\,q^2}{(1+q)^2}\,,\\
\label{aeff}
{a}_{\rm eff}&\to a_{\rm eff}(\beta,\gamma,q)\equiv a_{\rm
tot}(\beta,\gamma)+\xi\nu(a_1\cos\beta+a_2\cos\gamma)\,,
\end{align}
with $\beta$ ($\gamma$) being the angle between $\boldsymbol{a}_1$
($\boldsymbol{a}_2$) and the orbital angular momentum. Clearly, with
this choice Eq.~\eqref{afin} matches Eq.~\eqref{new_ansatz} for parallel
spins (\textit{i.e.,}~for $\beta=0,\pi$ and $\gamma=0,\pi$).
Moreover, for equal masses ($q=1$), the leading-order PN spin effects in
the conservative sector (\textit{i.e.,}~the leading-order spin-orbit coupling) enter
the dynamics only through the combination
${\boldsymbol{\hat{L}}}\cdot\boldsymbol{S}/M^2=a_{\rm
tot}(\beta,\gamma,1)\propto a_{\rm eff}(\beta,\gamma,1)$ \citep[see
\textit{e.g.,}][]{2001PhRvD..64l4013D,Barausse:2009xi}, where a ``hat'' denotes a
unit-norm vector. Therefore, at this approximation order, the binding
energy and angular momentum at the effective ISCO depend on the
spins only through $a_{\rm tot}(\beta,\gamma,1)$ (or equivalently $a_{\rm
eff}(\beta,\gamma,1)$), as reflected in
Eqs.~\eqref{ell}--\eqref{aeff}. Similarly, in the EMRL, the leading
contributions to $|\boldsymbol{\ell}|$ come from the ISCO energy and
angular momentum of a test particle in Kerr. By construction,
$|\boldsymbol{\ell}|$ has the correct EMRL for parallel spins, but the
EMRL is also recovered approximately for generic-spin configurations, at
least at leading order in the primary-BH spin. Indeed, this happens
because the ISCO angular momentum and energy for a test particle in a
non-equatorial orbit in a Kerr spacetime are $L_{_{\rm ISCO}}(a_{\rm
tot}(\beta,\gamma,0))$ and $E_{_{\rm ISCO}}(a_{\rm
tot}(\beta,\gamma,0))$, at leading order in the spin~\citep[see
discussion in][]{barausse_mass_2012}.
Putting things together, the final-spin magnitude reads
\begin{eqnarray}
&|\boldsymbol{a}_{\rm fin}|=\frac{1}{(1+q)^2}\big[|\boldsymbol{a}_1|^2+|\boldsymbol{a}_2|^2q^4+2|\boldsymbol{a}_1||\boldsymbol{a}_2|q^2\cos\alpha\notag\\
&+2(|\boldsymbol{a}_1|\cos{\beta}+|\boldsymbol{a}_2|q^2\cos{\gamma})|\boldsymbol{\ell}|q+|\boldsymbol{\ell}|^2q^2\big]^{1/2}\,, \label{afinnorm}
\end{eqnarray}
where $\alpha$ is the angle between the two spins. In principle, the
angles $\alpha$, $\beta$ and $\gamma$ depend on the binary
separation. However, $\beta$ and $\gamma$ enter in our formulae only
through $a_{\rm tot}(\beta,\gamma,q)$ and $a_{\rm
eff}(\beta,\gamma,q)$. These combinations remain constant during the
adiabatic inspiral~\citep{acst-94}, if only the leading PN order in the
spins (\textit{i.e.,}~the leading-order spin-orbit coupling) is included, and either
\textit{(i)} the masses are equal; or \textit{(ii)} only one BH is
spinning; or \textit{(iii)} the mass ratio is extreme
(\textit{i.e.,}~$\nu\approx0$). Similarly, under the same assumptions, we can safely
assume that $\alpha$ remains constant during the adiabatic
inspiral~\citep{acst-94}, \textit{i.e.,}~the angle between the two spins is
preserved by the leading-order spin-orbit coupling for equal masses,
while it does enter the final-spin prediction if only one BH is spinning,
or when $\nu\approx 0$ (indeed, the effect of the smaller BH's spin
vanishes at leading order in $\nu$, because
$\vert\boldsymbol{S}_2\vert={\cal O}(\nu)^2$). Outside these special
cases, $\alpha$, $\beta$ and $\gamma$ are not exactly constant. For
instance, in general $\alpha$ oscillates and the oscillations may even
become ``flip-flop''-unstable between separations $r_{\rm ud\pm}=
(\sqrt{a_1}\pm\sqrt{q\,a_2})^4(M_1+M_2)/(1-q)^{2}$ for certain
unequal-mass configurations where the primary-BH spin is aligned with the
orbital angular momentum and the spin of the secondary is anti-aligned
with it~\citep{2016arXiv160105086L,gerosa_precessional_2015}. These
configurations, however, are unlikely if the spins are isotropically
distributed, or if the spins are almost aligned with the angular momentum
of a circumbinary disk due to the Bardeen-Petterson
effect~\citep{Bardeen:1975zz}.
Therefore, we follow~\citet{barausse_mass_2012,barausse_predicting_2009,Rezzolla:2007c} and
define $\alpha,\,\beta$ and $\gamma$ at the initial binary separation
$r_{\rm in}$
\begin{align}
\cos\alpha &\equiv{\boldsymbol{\hat{a}}}_1\cdot{\boldsymbol{\hat{a}}}_2\vert_{r_{\rm
in}}\,,\nonumber \\
\label{angles}
\cos\beta &\equiv{\boldsymbol{\hat{L}}}\cdot
{\boldsymbol{\hat{a}}}_1\vert_{r_{\rm in}}\,,\\
\cos\gamma &\equiv
{\boldsymbol{\hat{L}}}\cdot{\boldsymbol{\hat{a}}}_2\vert_{r_{\rm
in}}\,.\nonumber
\end{align}
Indeed, \citet{barausse_predicting_2009} and \citet{2010PhRvD..81h4054K}
have verified that the final-spin predictions are robust against the
initial separation $r_{\rm in}$, \textit{i.e.,}~in most cases the definitions
(\ref{angles}) are justified.
A comparison between our new formula with $n_{_J}=2$, $n_{_M}=1$ and the
generic-spin simulations in our calibration dataset yields the residuals
displayed in Fig.~\ref{PDFgeneric}. Also shown is the corresponding PDF
with mean $\mu\approx-0.005$ and rms $\sigma\approx0.007$. Note that in
this case we cannot reliably estimate $\chi^2_{\rm red}$, as none of the
generic-spin configurations have ``twins'' in our calibration dataset,
and the NR error is expected to be larger than in parallel-spin binaries
because of precession. Also shown by Fig.~\ref{PDFgeneric} is an
unattractive feature of our formula, namely, that the distribution of
residuals is biased toward negative values (\textit{i.e.,}~our formula
systematically overpredicts the final spin for generic
binaries). Although this bias is small, and because it follows from
assuming that $\alpha, \beta, \gamma$ are constant, we can amend it by
replacing the angles $\alpha,\,\beta,\,\gamma$ by ``effective'' angles
$\alpha^*,\,\beta^*,\,\gamma^*$ defined as
\begin{equation}
\label{angle_map}
\Theta^*\equiv2\arctan\left[(1+\epsilon_\Theta)\tan\frac{\Theta}{2}\right]\approx\Theta+\epsilon_\Theta\sin{\Theta}\,,
\end{equation}
where $\Theta=\alpha,\,\beta,\,\gamma$, $\epsilon_\Theta$ are free
coefficients to be fixed by the data, and we impose
$\epsilon_\beta=\epsilon_\gamma$ to make our formula symmetric under
exchange of the two BHs. Clearly, for parallel spins $\alpha^*=\alpha$,
$\beta^*=\beta$ and $\gamma^*=\gamma$. A comparison with the NR data
gives $\epsilon_\alpha\approx 0$ and
$\epsilon_\beta=\epsilon_\gamma\approx 0.024$, where we have used the
second equality of Eq.~\eqref{angle_map} (the first equality gives
similar results). The corresponding residual distribution has a smaller
bias and is shown in the inset of Fig.~\ref{PDFgeneric}, together with a
PDF with $\mu\approx-0.001$, $\sigma\approx0.006$.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{PDF_GU_generic.pdf}
\caption{The residual distribution for the remnant spin magnitude, for
binaries with generic spins; the inset shows how the modest bias of
the distribution can be reduced by adjusting the angles
$\beta,\,\gamma$.
\label{PDFgeneric}}
\end{center}
\end{figure}
As a further ``blind'' test of our formula, we consider data from the
recently published catalogue of \citet{gatech_2016} that is \textit{not}
already included in our calibration dataset (\textit{i.e.,}~83 parallel-spin and 165
precessing-spin simulations). Already when using only $n_{_M}=1$ and
$n_{_J}=2$, the comparison yields mean and rms residuals $\mu\approx
-5\times 10^{-5}$ and $\sigma\approx1.4\times 10^{-4}$ for parallel
spins, and $\mu\approx -0.004$ ($\mu\approx -0.0005$), and
$\sigma\approx3.3\times 10^{-4}$ ($\sigma\approx3.5\times 10^{-4}$) for
precessing spins with unadjusted (adjusted) angles $\beta,\gamma$.
Finally, for the final-spin direction we follow
\citet{barausse_predicting_2009,acst-94} and note that at leading PN
order in the spins (\textit{i.e.,}~including the leading-order spin-orbit coupling
alone), the GW-driven evolution in the adiabatic inspiral approximately
preserves the direction of the total angular momentum
$\boldsymbol{J}\equiv\boldsymbol{L}+\boldsymbol{S}$.
\citet{barausse_predicting_2009}, and later \citet{2014PhRvD..89b1501L},
verified that $\boldsymbol{\hat{J}}$ is approximately preserved (to
within a few degrees) also in the plunge, merger and ringdown. The only
exception to this ``simple-precession'' picture are binaries with spins
almost anti-aligned with the orbital angular momentum at large
separations~\citep{acst-94,2010PhRvD..81h4054K}. Indeed, when the GW
emission sheds enough angular momentum that $\boldsymbol{L}\approx
-\boldsymbol{S}$, these binaries undergo ``transitional
precession''~\citep{acst-94}, whereby the direction of
${\boldsymbol{\hat{J}}}$ changes significantly on short timescales. Note
that among the configurations that give rise to ``simple precession'' are
also the ``flip-flop'' binaries of~\citet{2016arXiv160105086L} and
\citet{gerosa_precessional_2015}. Since transitional-precession
configurations comprise a small portion of the parameter
space~\citep{2010PhRvD..81h4054K}, we follow
\citet{barausse_predicting_2009} and assume that the final-spin direction
is simply given by ${\boldsymbol{\hat{J}}}(r_{\rm in})$, \textit{i.e.,}~the
final-spin angle $\theta_{\rm fin}$ relative to the initial angular
momentum is simply
\begin{equation}
\label{eq:dir}
\cos\theta_{\rm fin}={\boldsymbol{\hat{J}}}(r_{\rm in})\cdot{\boldsymbol{\hat{L}}}(r_{\rm in})\,.
\end{equation}
Indeed, the 157 simulations~\citep{zlochower_modeling_2015,Lousto_2014b}
in our dataset that report the final-spin direction confirm that the
final spin is almost aligned with ${\boldsymbol{\hat{J}}}(r_{\rm in})$,
to within $\sim 18^\circ$ in the worst case, and to within $4^\circ$
($6^\circ$) in $64\%$ ($78\%$) of the cases. The distribution of the
angle between the final spin and ${\boldsymbol{\hat{J}}}(r_{\rm in})$ is
shown in Fig.~\ref{fig_inclination}; it is unclear whether the small
counts for $\theta_{\rm fin} \gtrsim 10^\circ$ are due to imprecisions in the
formula or in the numerical simulations.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{angle_distribution.pdf}
\caption{The distribution of the angle between the final spin and the
initial direction of the total angular momentum.
\label{fig_inclination}}
\end{center}
\end{figure}
Finally, we note that unlike other formulae for the final-spin direction
\citep{buonanno_estimating_2008,tichy_final_2008,Rezzolla:2007c},
Eq.~(\ref{eq:dir}) is valid also when $r_{\rm in}\gg M_1+M_2$. (This is also the
case for our formula for the final-spin magnitude.) This is
particularly important to predict the final spin in massive BH
mergers. Indeed, cosmological simulations (both numerical and
semi-analytical ones) cannot follow the evolution of massive BH binaries
below the separation $r_{_{\rm GW}}$ at which the GW dynamics starts
driving the orbital evolution. For a binary with
$M_1+M_2\sim10^8\,M_\odot$ in a gas-poor environment, $r_{_{\rm
GW}}\sim10^{-2}\,{\rm pc}\sim2\times10^{3}\,(M_1+M_2)$, a separation
at which other prescriptions for the final-spin direction become
significantly inaccurate \citep[see discussion
in][]{barausse_predicting_2009,2010JPhCS.228a2050B}.\footnote{In a gas-rich environment,
the separation $r_{_{\rm GW}}$ below which GWs dominate
the binary evolution and our formulae can be applied
is smaller \citep{Armitage_2012}, while for ``flip-flop'' binaries our
formula for the final-spin magnitude might be applicable only below $r_{\rm ud\pm}$.}
\section{conclusion}
By combining information from the test-particle limit,
perturbative/self-force calculations, the PN dynamics, and an extensive
set of NR simulations collected from the literature, we have constructed
a novel formula for the final spin from the merger of quasi-circular BH
binaries with arbitrary mass ratios and spins. When applied to
parallel-spin configurations, our novel formula performs better than
other expressions in the literature, and we have also tested its validity
for precessing-spin binaries, which other formulae are not able to model
accurately. Also, unlike models such as that of
\citet{healy_remnant_2014}, our formula is purely algebraic. Finally, we
have used our collected NR dataset to confirm that the final-spin
direction is almost parallel to the initial total angular-momentum
direction, as first suggested by \citet{barausse_predicting_2009}.
\acknowledgements
We thank Nathan Johnson-McDaniel for useful comments and Davide Gerosa
for clarifications on flip-flop binaries. We acknowledge
support from the European Union's Seventh Framework Programme
(FP7/PEOPLE-2011-CIG) through the Marie Curie Career Integration Grant
GALFORMBHS PCIG11-GA-2012-321608, from the H2020-MSCA-RISE-2015 Grant
No. StronGrHEP-690904, and from the ERC Synergy Grant ``BlackHoleCam -
Imaging the Event Horizon of Black Holes'' (Grant 610058).
\newpage
\bibliographystyle{apj}
|
\section{Introduction}
We consider the parabolic PDE\footnote{We use the Einstein summation convention: there is a sum on indices appearing twice in a formula: $\displaystyle A_\alpha B^\alpha = \sum_{\alpha = 1}^d A_\alpha B^\alpha $. We also use the Kronecker delta $\delta^{ab}$ which is 1 if $a=b$ and 0 if $a \neq b$.}
\begin{equation}
\label{eq:PDE}
\partial_t u + \mu^\alpha \partial_\alpha u + \frac{1}{2} C^{\alpha\beta} \partial_\alpha \partial_\beta u = r u
\end{equation}
with terminal condition $u_T(x) = h(x)$ in a $(1+d)$-dimensional space with time $t$ and space variables $x$ with coordinates $x_\alpha$, $\alpha = 1\ldots d$. All coefficients may depend on $t$ and $x$.
$C$ is a positive semi-definite symmetric matrix. It can be rewritten using Cholesky decomposition as $C = \sigma \sigma^\intercal$ o
$C^{\alpha\beta} = \delta^{ab} \sigma^\alpha{}_a \sigma^\beta{}_b$.
Feynman-Kac theorem states that the solution of this PDE is given by the expected value
\begin{equation}
u(t,x) = \mathbb{E}\!\left[e^{-\int_t^T r(s,X_s) \mathrm{d} s} h(X_T) \mid X_t = x \right] \rlap{\ .}
\label{eq:expected-value}
\end{equation}
under a multidimensional diffusion process
\begin{equation*}
\mathrm{d} X_t = \mu \mathrm{d} t + \sigma \mathrm{d} W_t \rlap{\ .}
\end{equation*}
$W_t$ is a $d$-dimensional vector of independent standard Brownian process.
Drifts $\mu$ and volatilities $\sigma$ are stochastic variables which depend on time $t$ and state variable $X$.
The expected value \eqref{eq:expected-value} can be numerically estimated using Monte Carlo simulation.
However common schemes exhibit some bias due to the discretization of time. This happens for instance with the simplest one, the Euler scheme. One usually controls this bias through smaller timestepping, which increases the computation time.
\cite{bally2014probabilistic} introduce schemes without such bias (in the case where $r=0$). This is achieved using random time steps, where the time discretization is given by jump times of an independent Poisson process. In addition, one has to multiply paths contributions by some weights which are functions of the time discretization and of the path. However this can correspond to integrating random variables which are not always integrable or do not have finite variance, which leads to poor convergence.
\cite{henry2015exact} enhance this algorithm in two different ways. First, the weight functions are obtained for all diffusion processes in term of Malliavin weights. They also manage to keep the variance of the integrated variable finite in two cases: constant volatility and correlation (constant $C^{\alpha\beta}$) or one-dimensional processes without drift ($\mu^\alpha = 0$). The variance is controlled by the intensity of the auxilliary Poisson process but it is not monotonous: it increases when the Poisson intensity becomes too small or too high. In particular, it is not possible to increase the average number of time steps beyond some value without increasing the variance.
\newline
Our first contribution is to introduce a different Monte Carlo scheme, with smaller variance. In particular the variance becomes asymptotically a decreasing function of the Poisson intensity.
As a second contribution, we show how to make the variance finite in the general case of a process with drift and nonconstant volatility, in any dimension.
Our third contribution is to handle the case where the discount rate $r$ is stochastic.
\newline
We introduce the basic principles of this scheme in section \ref{sec:poisson}. We then detail it in the one-dimensional case in section \ref{sec:one-dimension}, with numerical evidence in section \ref{sec:numerical}. We finally explain the general mutidimensional case with stochastic discount rate in section \ref{sec:multidimensional}.
\section{Unbiased scheme}
\label{sec:poisson}
\subsection{Poisson process}
Let us consider the parabolic PDE \eqref{eq:PDE}
with terminal condition
$
u_T(x) = h(x)
$.
We can rewrite this PDE as
$$
\partial_t u_t = - \mathcal{H}_t u_t
$$
where $\mathcal{H}_t$ is the elliptic differential operator
$$
\mathcal{H}_t(x) = - r(t,x) + \mu^\alpha(t,x) \partial_\alpha + \frac{1}{2} C^{\alpha\beta}(t,x) \partial_\alpha \partial_\beta \rlap{\ .}
$$
The evolution operator can be expressed using a time ordered exponential
$$
U_{t,T} = \mathcal{P} e^{\textstyle \int_t^T \!\mathcal{H}_s \,\mathrm{d} s}
$$
which is a notation for the infinite product
$$
U_{t,T} = \lim_{n \rightarrow \infty} \prod_{k=0}^{n-1} e^{\textstyle \mathcal{H}_{t+k(T-t)/n} \frac{(T-t)}{n} }
\rlap{\ .}
$$
This allows to write the solution of equation \eqref{eq:PDE} with terminal condition $u_T(x) = h(x)$ as
$$
u_t = U_{t,T} u_T = U_{t,T} h \rlap{\ .}
$$
Expliciting variables, this means
$$
u_t(x) = \int U_{t,T}(x,y) h(y) \, \mathrm{d} y \rlap{\ .}
$$
In some cases, for instance when the stochastic process is a pure Brownian motion or a lognormal diffusion, the resulting marginal probability measure can be computed explicitely or can be simulated exactly\footnote{
On an infinitesimal time $\delta t$, $U_{t,t+\delta t} = e^{\textstyle \mathcal{H}_t \delta t}$ acting on a test function $\phi$ can be represented by the convolution with a Gaussian kernel:
$$
\left(e^{\textstyle \mathcal{H}_t \delta t} \phi\right)\!(X_t) = e^{\textstyle -r \delta t } \frac{1}{\sqrt{(2 \pi \delta t)^d | \det(C) | }} \int \mathrm{d}( \delta X)^d
\ e^{\textstyle -\frac{1}{2} \left(\frac{\delta X}{\delta t} - \mu \right)^T C^{-1} \left(\frac{\delta X}{\delta t} - \mu \right) \delta t} \phi(X_t+\delta X) \ .
$$
Defining
$$
\mathcal{L}_t = \frac{1}{2} (\dot X_t-\mu_t)^T C_t^{-1} (\dot X_t-\mu_t) + r_t
$$
with $\dot X_t = \partial_t X_t = \frac{\delta X}{\delta t}$, this convolution reads
$$
\left(e^{\textstyle \mathcal{H}_t \delta t} \phi\right)\!(X_t) = \frac{1}{\sqrt{2 \pi | \det(C) | \delta t }} \int {\textstyle \prod\limits_i} \mathrm{d}( \delta X^{(i)})
\ e^{\textstyle -\mathcal{L}_t \delta t} \phi(X_t+\delta X) \ .
$$
Integrating over infinitesimal times between $t$ and $T$, $u_t$ can thus be expressed as a Feynman path integral
$$
u_t = \mathcal{N} \int \! \mathcal{D}X_s \, e^{\textstyle - \int_t^T \!\mathcal{L}_s \mathrm{d} s} h(X_T) \rlap{\ .}
$$
The integral is taken over all paths starting from $X_0$ with a proper normalization factor $\mathcal{N}$.
}.
In other cases, we suppose that we have a different elliptic operator
$$
\widehat{\mathcal{H}}_t = -\hat r + \hat\mu^\alpha \partial_\alpha + \frac{1}{2} \hat C^{\alpha\beta} \partial_\alpha \partial_\beta
$$
with coefficients $\hat r$, $\hat \mu$ and $\hat C$ chosen such that the PDE $\partial_t u_t = -\widehat{\mathcal{H}}_t u_t$ can be solved exactly by Monte Carlo simulation. We denote the evolution operator by $\widehat{U}_{t,t'} = \mathcal{P} e^{\textstyle \int_t^{t'} \!\widehat{\mathcal{H}}_s \,\mathrm{d} s} $ and decompose $\hat C$ as $\hat C^{\alpha\beta} = \delta^{ab} \hat\sigma^\alpha{}_a\hat\sigma^\beta{}_b$. This means that for any function $\phi(x)$ we can get the value of
$$
\left(\widehat{U}_{t,t'} \phi\right)(x) = \widehat{\mathbb{E}}\!\left[e^{-\int_t^{t'} \hat{r}(s,X_s) \mathrm{d} s} \phi(X_{t'}) \mid X_t = x \right]
$$
under the stochastic process
\begin{equation*}
\mathrm{d} X_t = \hat\mu \mathrm{d} t + \hat\sigma \mathrm{d} W_t \rlap{\ .}
\end{equation*}
With explicit coordinates this process follows the equation
\begin{equation}
\label{eq:processHat}
\mathrm{d} X_t^\alpha = \hat\mu^\alpha(t,X_t) \mathrm{d} t + \hat\sigma^\alpha{}_a(t,X_t) \mathrm{d} W^a_t
\end{equation}
where $W^a$ are independent standard Brownian motions.
We decompose the original operator as
$$
\mathcal{H}_t = \widehat{\mathcal{H}}_t + \Delta \mathcal{H}_t
$$
with a corrective term
\begin{equation}
\label{eq:deltaH}
\Delta \mathcal{H}_t = - \Delta r + \Delta \mu^\alpha \partial_\alpha + \frac{1}{2} \Delta C^{\alpha\beta} \partial_\alpha \partial_\beta
\end{equation}
with
\begin{eqnarray*}
\Delta r &=& r - \hat r
\\
\Delta \mu &=& \mu - \hat\mu
\\
\Delta C &=& C - \hat C \rlap{\ .}
\end{eqnarray*}
Using this split, the evolution operator becomes
$$
U_{t,T} = \mathcal{P} e^{\textstyle \int_t^T (\widehat{\mathcal{H}}_s + \Delta \mathcal{H}_s)\,\mathrm{d} s} \rlap{\ .}
$$
On each infinitesimal time $\delta t$ we have a term
\begin{equation}
\label{eq:deltaHdt}
e^{\textstyle \Delta \mathcal{H}_t \delta t} = 1+\Delta \mathcal{H}_t \delta t = 1 - \delta t \Delta r + \delta t \Delta\mu^\alpha \partial_\alpha + \frac{1}{2} \delta t \Delta C^{\alpha\beta}\partial_\alpha \partial_\beta \rlap{\ .}
\end{equation}
The process with drift $\hat\mu$ and covariance matrix $\hat C$ is chosen so that it can be simulated without any bias. However the term $\Delta \mathcal{H}_t \delta t$ given in equation \eqref{eq:deltaHdt} should be computed and taken into account at any infinitesimal time, which is numerically impossible.
Instead, this contribution $\Delta \mathcal{H}_t \delta t$ is kept only with probability $\lambda_t \delta t$ over infinitesimal time $\delta t$, compensated by a factor of $\frac{1}{\lambda_t \delta t}$, such that its expected value is unchanged.
In other words, as in \cite{bally2014probabilistic} and \cite{henry2015exact} we consider a Poisson process $N_t$ with intensity $\lambda_t$. We then replace $1+\Delta \mathcal{H}_t \delta t$ b
\begin{equation}
\label{eq:jump}
1+ \delta N_t \frac{\Delta \mathcal{H}_t}{\lambda_t} \rlap{\ .}
\end{equation}
Applied to a test function $\phi$, the expected value over the Poisson process gives the factor we want to take into account:
$$
\mathbb{E}^P\!\left[\left(1+ \delta N_t \frac{\Delta \mathcal{H}_t}{\lambda_t} \right) \phi \right] = \left(1 + \lambda_t \delta t \frac{\Delta \mathcal{H}_t}{\lambda_t}\right) \phi = e^{\textstyle \Delta \mathcal{H}_t \delta t} \phi \rlap{\ .}
$$
In addition, the intensity $\lambda_t$ could also be a stochastic process and depend on $t$ and $X_t$.
Let $p$ be the number of Poisson jumps between times $t_0$ and $T$ and $t_k$, $k \geq 1$ the jump times. Between two Poisson jumps, the evolution operator corresponds to the diffusion process \eqref{eq:processHat}:
$$
\widehat{U}_{t_k,t_{k+1}} = \mathcal{P} e^{\textstyle \int_{t_k}^{t_{k+1}} \!\widehat{\mathcal{H}}_s \,\mathrm{d} s} \rlap{\ .}
$$
The processes $\hat\mu_t$ and $\hat C_t$ can depend on the Poisson jump times and also on the value of $X_t$ given by the diffusion equation \eqref{eq:processHat}. Our explicit choice will be described in sections \ref{sec:one-dimension} and \ref{sec:multidimensional}.
Integrating over all times and taking the expected value on the Poisson process, we have
\begin{multline*}
u_{t_0} = \mathbb{E}^P_{t_0}\left[ \widehat{U}_{t_0,t_1} \left( 1+\frac{\Delta \mathcal{H}_{t_1}}{\lambda_{t_1}} \right)
\widehat{U}_{t_1,t_2} \left( 1+\frac{\Delta \mathcal{H}_{t_2}}{\lambda_{t_2}} \right)
\cdots
\right.
\\
\left.
\cdots
\widehat{U}_{t_{p-1},t_p} \left( 1+\frac{\Delta \mathcal{H}_{t_p}}{\lambda_{t_p}} \right)
\widehat{U}_{t_p,T} h(X_T)
\right]\ .
\end{multline*}
Operators $\widehat{U}$ act on functions by integration, which can be explicited as
\begin{multline}
\label{eq:integral}
u_{t_0}(X_{t_0}) =
\mathbb{E}^P_{t_0}\left[ \int\!\mathrm{d} X_{t_1} \, \widehat{U}_{t_0,t_1}(X_t,X_{t_1})
\left( 1+\frac{\Delta \mathcal{H}_{t_1}}{\lambda_{t_1}} \right)
\right.
\\
\int\!\mathrm{d} X_{t_2} \, \widehat{U}_{t_1,t_2}(X_{t_1},X_{t_2})
\left( 1+\frac{\Delta \mathcal{H}_{t_2}}{\lambda_{t_2}} \right)
\cdots
\\
\left.
\int\!\mathrm{d} X_{t_p} \, \widehat{U}_{t_{p-1},t_p}(X_{t_{p-1}},X_{t_p})
\left( 1+\frac{\Delta \mathcal{H}_{t_p}}{\lambda_{t_p}} \right)
\int\!\mathrm{d} X_T \, \widehat{U}_{t_p,T}(X_{t_p},X_T) h(X_T)
\right] \rlap{\ .}
\end{multline}
\subsection{Monte Carlo simulation}
In a Monte Carlo simulation, the integrals on $X_{t_1}, \cdots, X_T$ and the evolution operators $\widehat{U}$ are handled by averaging over simulated paths generated according to the law given by $\widehat{U}$. In order to get an unbiased Monte Carlo scheme, we generate random sampling for Poisson jump times $t_k$ and the values of $X_{t_k}$ at those times according to process \eqref{eq:processHat}. This process, with drift $\hat\mu$ and covariance matrix $\hat C$ is chosen so that it can be simulated exactly, \emph{i.e.} such that the Monte Carlo distribution of $X_{t_{k+1}}$ discounted at rate $\hat r$, conditional to $X_{t_k}$, tends to $\widehat{U}_{t_k,t_{k+1}}(X_{t_k},X_{t_{k+1}})$ when the number of samplings goes to infinity.
Operators $\Delta \mathcal{H}_{t_k}$ are differential operators acting on all the factors which follow in formula \eqref{eq:integral}. The first term to depend on the differentiation variable $X_{t_k}$ is in fact $\widehat{U}_{t_k,t_{k+1}}$. If we know the explicit form of this evolution kernel, we can differentiate it explicitely. This defines weights $\widehat{\mathcal{W}}_\alpha$ and $\widehat{\mathcal{W}}_{\alpha\beta}$:
\begin{eqnarray*}
\partial_{X_{t_k}^\alpha} \widehat{U}_{t_k,t_{k+1}} &=& \widehat{\mathcal{W}}_\alpha\, \widehat{U}_{t_k,t_{k+1}}
\\
\partial_{\partial X_{t_k}^\alpha}\partial_{X_{t_k}^\beta} \widehat{U}_{t_k,t_{k+1}} &=& \widehat{\mathcal{W}}_{\alpha\beta} \, \widehat{U}_{t_k,t_{k+1}} \rlap{\ .}
\end{eqnarray*}
These are similar to Malliavin weights, except that here the kernel $\widehat{U}$ includes the discount factor. With null or deterministic discount rates, $\widehat{\mathcal{W}}_\alpha$ and $\widehat{\mathcal{W}}_{\alpha\beta}$ are exactly the Malliavin weights, as introduced in this context by \cite{henry2015exact}. When the discount factor depends on $X_{t_k}$ they also incorporate the derivative of these discount factors.
However multiplying by these weights at this step would result in computing the expected value of a quantity with infinite variance. We will explain this issue and how to solve it in the following sections.
The other terms which can be functions of $X_{t_k}$ are $\hat\mu$, $\hat C$, $\hat r$ and $\lambda$, depending on the choice which is made for those functions. However the variations of these variables will not contribute to the final result.
To make this clear, we rewrite formula \eqref{eq:integral} in a recursive way for intermediate Poisson jump times:
\begin{equation}
\label{eq:recursive}
u_{t_{k-1}}(X_t) =
\mathbb{E}_{t_{k-1}}\!\left[ \int\!\mathrm{d} X_{t_{k}} \, \widehat{U}_{t_{k-1},t_k}(X_{t_{k-1}},X_{t_{k}})
\left( 1+\frac{\Delta \mathcal{H}_{t_{k}}}{\lambda_{t_{k}}} \right)
u_{t_{k}}(X_{t_{k}})
\right] \rlap{\ .}
\end{equation}
In this formula, $\Delta\mathcal{H}_{t_k}$ acts by differentiation with respect to $X_{t_k}$ on $u_{t_{k}}(X_{t_{k}})$. The formula for $u_{t_{k}}$ involves $\hat\mu$, $\hat C$ and $\lambda$. However these functions and the process \eqref{eq:processHat} are only intermediate objects used in the computation: the final value of $u_{t_{k}}$ does not depend on the values chosen for these functions. As a consequence, their variations do not contribute to the derivatives of $u_{t_{k}}$. In the following, differentiations in operators $\Delta\mathcal{H}_{t_k}$ are therefore computed with frozen $\hat\mu$, $\hat C$ and $\lambda$. In order to make this clear, we will denote by $X_t^*$ those frozen values which should not be variated when differentiating.
\subsection{Infinite variance}
Let us consider the simple case of a pure Brownian process. The evolution operator is the Gaussian kernel
$$
U_{s,t}(W_s,W_t) = \frac{1}{\sqrt{2\pi (t-s)}} e^{\textstyle -\frac{(W_t^a-W_s^a)(W_t^a-W_s^a)}{2 (t-s)}} \rlap{\ .}
$$
The corresponding Malliavin weights are therefore
\begin{eqnarray*}
\mathcal{W}_a &=& \frac{1}{U_{s,t}}\partial_{W_s^a} U_{s,t} = \frac{\Delta W^a}{\Delta t} \\
\mathcal{W}_{ab} &=& \frac{1}{U_{s,t}}\partial_{W_s^a}\partial_{W_s^b} U_{s,t} = \frac{\Delta W^a\Delta W^b}{\Delta t^2} -\frac{\delta_{ab}}{\Delta t}
\end{eqnarray*}
with
$\Delta W = W_t-W_s$ and $\Delta t = t-s$.
The time $\Delta t$ between two jumps is given by a Poisson law with density $\lambda e^{-\lambda \Delta t}$ which behaves as $O(1)$ at small $\Delta t$. As $\Delta W = O(\sqrt{\Delta t})$, Malliavin weights are of orders
$
\mathcal{W}_a = O\!\left( \frac{1}{\sqrt{\Delta t}}\right)
$ and $
\mathcal{W}_{ab} =O\!\left( \frac{1}{\Delta t}\right)
$. This remains true for a large class of diffusion processes. Direct multiplication by weights $\widehat{\mathcal{W}}_\alpha$ and $\widehat{\mathcal{W}}_{\alpha\beta}$ in \eqref{eq:integral} would therefore give a quantity of infinite variance. This would result in poor Monte Carlo convergence. In order for the expected value over jump time to be well defined and the integrand to have finite variance, the random variable we integrate should be of order $O\!\left(\frac{1}{\Delta t^\gamma}\right)$ with $\gamma < \frac{1}{2} $, so that its square is of order $O\!\left(\frac{1}{\Delta t^{2\gamma}}\right)$ with $2\gamma < 1 $.
In order to have a variable which behaves as $O(1)$ when $\Delta t$ is small, the weights should be multiplied by quantity of order $O(\Delta t)$. In equation \eqref{eq:integral}, $\Delta \mathcal{H}_{t_k} \widehat{U}_{t_k,t_{k+1}}(X_{t_k},\allowbreak X_{t_{k+1}})$ is multiplied by $1+\frac{\Delta \mathcal{H}_{t_{k+1}}}{\lambda}$, unless $t_k$ is the last Poisson time ($k=p$).
Let us consider first the term $\Delta \mathcal{H}_{t_{k+1}}$. According to its definition \eqref{eq:deltaH}, it is composed of terms proportional to $r(t_{k+1},X_{t_{k+1}})-\hat r(t_{k+1},X_{t_{k+1}})$, $\mu(t_{k+1},X_{t_{k+1}}) - \hat \mu(t_{k+1},\allowbreak X_{t_{k+1}})$ and $C(t_{k+1},X_{t_{k+1}})-\hat C(t_{k+1},X_{t_{k+1}})$. Our strategy is therefore to make all these terms of order $O(\Delta t_k) = O(t_{k+1} - t_k)$.
The other multiplicative term is $1$ which is not in $O(\delta t)$ and is multiplied by $\widehat{U}_{t_{k+1},t_{k+2}}(X_{t_{k+1}},X_{t_{k+2}})$. For this term, we will exploit the specific form of $\widehat{U}_{t_k,t_{k+1}}(X_{t_k},X_{t_{k+1}})$ to transfer the derivative on $X_{t_k}$ to a derivative on $X_{t_{k+1}}$. This derivative is then transferred to $\widehat{U}_{t_{k+1},t_{k+2}}(X_{t_{k+1}},X_{t_{k+2}})$ by integration by parts.
The last piece to handle is the last Poisson time, when the weights $\mathcal{W}$ must be multiplied by the final payoff. In this case, we use antithetic sampling as in \cite{henry2015exact} to have a final term in $O(\Delta t_p) = O(T-t_p)$.
Doing so, the random variable to integrate remains of order $O(1)$ and thus has finite variance under the Poisson law. Then the strong law of large numbers applies and the Monte Carlo convergence remains in $O\!\left(\frac{1}{\sqrt{N}}\right)$.
We now detail this scheme in the one-dimensional case with deterministic discount rate in section \ref{sec:one-dimension} and in the general multidimensional case with stochastic discount rate in section \ref{sec:multidimensional}.
\section{One-dimensional Monte Carlo scheme}
\label{sec:one-dimension}
Let us consider first the case of a one-dimensional process
$$
\mathrm{d} S = \mu(t,S) \mathrm{d} t + \sigma(t,S) \mathrm{d} W \rlap{\ .}
$$
We take a deterministic discount rate $r(t)$ and we want to price a European option of maturity $T$ with payoff $h(S_T)$. In other words, we want to solve the parabolic PDE
$$
\partial_t u_t(S) + \mu(t,S) \partial_S u_t(S) + \frac{1}{2} \sigma(t,S)^2 \partial_S^2 u_t(S) = r(t) u_t(S)
$$
with terminal condition $u_T(S) = h(S)$.
\subsection{Monte Carlo path}
We take a constant Poisson process intensity $\lambda$. Starting at date $t_0$, we draw at random the first Poisson time $t_1$. For instance, we draw a random uniform number $q$ between 0 and 1 and invert the cumulative law: $t_1 = t_0 -\frac{\log(q)}{\lambda}$. We can thus iteratively draw times $t_k$ until we get a date larger than $T$. We will denote by $p$ the last Poisson time before $T$.
Between two succesive dates $t_k$ and $t_{k+1}$ we will simulate a process with drift $\hat\mu(t,S)$ and volatility $\hat\sigma(t,S)$. More precisely, we consider a Brownian process $W_t$ and we choose a function $f_{(k)}(\Delta t, \Delta W)$ at each time $t_k$, such that $f_{(k)}(0,0) = 0$ and which can depend on $t_k$ and $S_{t_k}$. We then define
\begin{equation}
\label{eq:Sfromf}
S_t = S_{t_k} + f_{(k)}(t-t_k, W_t-W_{t_k}) \rlap{\ .}
\end{equation}
This corresponds to the It\^o process
$$
\mathrm{d} S_t = \left( \partial_{\Delta t} f_{(k)} + \frac{1}{2} \partial_{\Delta W}^2 f_{(k)} \right) \mathrm{d} t + \partial_{\Delta W} f_{(k)} \, \mathrm{d} W_t \rlap{\ .}
$$
Identifying with
$$
\mathrm{d} S_t = \hat\mu \mathrm{d} t + \hat\sigma \mathrm{d} W_t
$$
we define
\begin{eqnarray}
\label{eq:hatMuDef}
\hat\mu &=& \partial_{\Delta t} f_{(k)} + \frac{1}{2} \partial_{\Delta W}^2 f_{(k)}
\\
\label{eq:hatSigmaDef}
\hat \sigma &=& \partial_{\Delta W} f_{(k)} \rlap{\ .}
\end{eqnarray}
We suppose that we can write $f_{(k)}$ as a power series
$$
f_{(k)}(\Delta t, \Delta W) = \sum_{i,j} \frac{1}{i! j!}f_{ij} \Delta t^i \Delta W^j \rlap{\ .}
$$
($f_{ij}$ are coefficients which depend on $k$ but we do not make this explicit in order to simplify the notation.)
Then for $\Delta t = t_{k+1}-t_k$ and $\Delta W = W_{t_{k+1}} - W_{t_k}$ we have
\begin{eqnarray}
\label{eq:TaylorHatMu}
\hat\mu(t_{k+1},S_{t_{k+1}}) &=& f_{10} + f_{11} \, \Delta W + \frac{1}{2} f_{02} + \frac{1}{2} f_{03} \,\Delta W + O(\Delta t)
\\
\label{eq:TaylorHatSigma}
\hat \sigma(t_{k+1},S_{t_{k+1}}) &=& f_{01}+f_{02} \,\Delta W + O(\Delta t)
\end{eqnarray}
where we use $\Delta W = O(\sqrt{\Delta t})$.
On the other hand, a Taylor expansion of functions $\mu\big(t_k + \Delta t,S_{t_k} + f_{(k)}(\Delta t, \Delta W)\big)$ and $\sigma\big(t_k + \Delta t,S_{t_k} + f_{(k)}(\Delta t, \Delta W)\big)$ gives
\begin{eqnarray}
\label{eq:TaylorMu}
\mu(t_{k+1},S_{t_{k+1}}) &=& \mu(t_k,S_{t_k}) + \partial_S \mu(t_k,S_{t_k}) f_{01} \, \Delta W + O(\Delta t)
\\
\label{eq:TaylorSigma}
\sigma(t_{k+1},S_{t_{k+1}}) &=& \sigma(t_k,S_{t_k}) + \partial_S \sigma(t_k,S_{t_k}) f_{01} \, \Delta W + O(\Delta t) \rlap{\ .}
\end{eqnarray}
We have $\sigma(t_{k+1},S_{t_{k+1}}) = \hat \sigma(t_{k+1},S_{t_{k+1}}) + O(\Delta t)$ and therefore $C(t_{k+1},S_{t_{k+1}})-\hat C(t_{k+1},S_{t_{k+1}}) = \sigma(t_{k+1},S_{t_{k+1}})^2 - \hat \sigma(t_{k+1},S_{t_{k+1}})^2 = O(\Delta t)$ if and only if coefficients in equations \eqref{eq:TaylorHatSigma} and \eqref{eq:TaylorSigma} are equal. This means
\begin{eqnarray}
\label{eq:f01}
f_{01} &=& \sigma(t_k,S_{t_k})
\\
\label{eq:f02}
f_{02} &=& \partial_S \sigma(t_k,S_{t_k}) f_{01} = \sigma(t_k,S_{t_k}) \partial_S \sigma(t_k,S_{t_k}) \rlap{\ .}
\end{eqnarray}
Similarly, we have $\mu(t_{k+1},S_{t_{k+1}}) - \hat \mu(t_{k+1},S_{t_{k+1}}) = O(\Delta t)$ if and only if
\begin{eqnarray*}
f_{10} + \frac{1}{2} f_{02} &=& \mu(t_k,S_{t_k})
\\
f_{11} + \frac{1}{2} f_{03} &=& \partial_S \mu(t_k,S_{t_k}) f_{01} \rlap{\ .}
\end{eqnarray*}
Using the expressions for $f_{01}$ and $f_{02}$ in these equations, this reads
\begin{eqnarray}
\label{eq:f10}
f_{10} &=& \mu(t_k,S_{t_k}) - \frac{1}{2} \sigma(t_k,S_{t_k}) \partial_S \sigma(t_k,S_{t_k})
\\
\label{eq:f11}
f_{11} &=& \sigma(t_k,S_{t_k}) \partial_S \mu(t_k,S_{t_k}) - \frac{1}{2} f_{03}\rlap{\ .}
\end{eqnarray}
In addition, the continuity of $S_t$ at $t=t_k$ in \eqref{eq:Sfromf}, equivalent to $f(0,0)=0$, gives
\begin{equation}
\label{eq:f00}
f_{00} = 0 \rlap{\ .}
\end{equation}
Except expressions \eqref{eq:f01}, \eqref{eq:f02}, \eqref{eq:f10}, \eqref{eq:f11} and \eqref{eq:f00} we have the freedom to choose all other $f_{ij}$ coefficients. One simple choice is to set them to 0: $f_{ij} = 0$ for $(i=0, j\geq 3)$, $(i=1, j \geq 2)$ and $(i \geq 2)$. Between $t_k$ and $t_{k+1}$ we thus choose the function $f$ to be
\begin{multline}
f_{(k)}(\Delta t, \Delta W) = \mu(t_k,S_{t_k}) \Delta t + \sigma(t_k,S_{t_k}) \Delta W
\\
+ \frac{1}{2} \sigma(t_k,S_{t_k}) \partial_S \sigma(t_k,S_{t_k}) (\Delta W^2 - \Delta t)
+ \sigma(t_k,S_{t_k}) \partial_S \mu(t_k,S_{t_k}) \Delta t \Delta W \ .
\label{eq:f1d}
\end{multline}
At time $t_k$ we draw a Gaussian variable $\Delta W_k = W_{t_{k+1}} - W_{t_k}$ with variance $\Delta t_k = t_{k+1}-t_k$ and then get recursively the Monte Carlo path
\begin{equation}
S_{t_{k+1}} = S_{t_k} + f_{(k)}(\Delta t_k, \Delta W_k) \rlap{\ .}
\label{eq:MCpath}
\end{equation}
\subsection{Corrective terms}
\label{sec:1dcorrective}
Applying equations \eqref{eq:hatMuDef} and \eqref{eq:hatSigmaDef},
our choice \eqref{eq:f1d} corresponds to a process with drift and volatility
\begin{eqnarray*}
\hat\mu\big(t_k+\Delta t,S_{t_k}+f_{(k)}(\Delta t, \Delta W) \big)
&=&\mu(t_k,S_{t_k}) + \sigma(t_k,S_{t_k}) \partial_S \mu(t_k,S_{t_k}) \Delta W
\\
\hat\sigma\big(t_k+\Delta t,S_{t_k}+f_{(k)}(\Delta t, \Delta W)\big)
&=& \sigma(t_k,S_{t_k}) \big[ 1 + \partial_S \sigma(t_k,S_{t_k}) \Delta W
\\
&&\qquad\qquad\qquad\qquad\qquad
+ \partial_S \mu(t_k,S_{t_k}) \Delta t \big] \rlap{\ .}
\end{eqnarray*}
As the discount rate $r(t)$ is deterministic, we also take
$$
\hat r(t,S_t) = r(t) \rlap{\ .}
$$
Using these three function for $k-1$ with $\Delta t_k = \Delta t_{k-1} = t_k - t_{k-1}$ and $\Delta W = \Delta W_{k-1} = W_k - W_{k-1}$
and according to the definition of $\Delta \mathcal{H}_t$ we have
$$
1+\frac{\Delta\mathcal{H}_{t_k}}{\lambda} = 1+ \frac{\Delta \mu_k}{\lambda} \partial_{S_{t_k}} + \frac{1}{2}\frac{\Delta C_k}{\lambda} \partial_{S_{t_k}}^2
$$
with
\begin{eqnarray*}
\Delta \mu_k &=& \mu(t_k,S_{t_k} )
- \hat\mu\big(t_{k-1}+\Delta t_{k-1},S_{t_{k-1}}+f(\Delta t_{k-1}, \Delta W_{k-1}) \big)
\\
\Delta C_k &=& \sigma(t_k,S_{t_k})^2
- \hat\sigma\big(t_{k-1}+\Delta t_{k-1},S_{t_{k-1}}+f(\Delta t_{k-1}, \Delta W_{k-1}) \big)^2 \rlap{\ .}
\end{eqnarray*}
With our explicit choice of functions $f_{(k)}$, we get
\begin{eqnarray}
\nonumber
\Delta \mu_k
&=& \mu(t_k,S_{t_k}) - \big[ \mu(t_{k-1},S_{t_{k-1}}) + \sigma(t_{k-1},S_{t_k}) \partial_S \mu(t_{k-1},S_{t_{k-1}}) \Delta W_{k-1} \big]
\\
\label{eq:deltaMudeltaC}
\Delta C_k
&=& \sigma(t_k,S_{t_k})^2 - \sigma(t_{k-1},S_{t_{k-1}})^2 \big[ 1+ \partial_S \sigma(t_{k-1},S_{t_{k-1}}) \Delta W_{k-1}
\\
&& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad
+ \partial_S \mu(t_{k-1},S_{t_{k-1}}) \Delta t_{k-1} \big]^2 \rlap{\ .}
\nonumber
\end{eqnarray}
\subsubsection{Intermediate times}
\label{sec:1dintermediate}
If $t_k$ is not the last Poisson time, \emph{i.e} $k<p$, $\Delta \mathcal{H}_k$ acts on
\begin{equation}
\label{eq:intUcorrective}
u_k(t_k,S_{t_k}) = \int \mathrm{d} S_{t_{k+1}} \widehat{U}_{t_k,t_{k+1}}(S_{t_k},S_{t_{k+1}}) \left(1+\frac{\Delta \mathcal{H}_{k+1}}{\lambda}\right) u_{t_{k+1}}(S_{t_{k+1}}) \rlap{\ .}
\end{equation}
More generally we will consider the action on this expression of a second order differential operator
$$
\mathcal{A}_k = 1 + A_k^S \partial_{S_{t_k}} + \frac{1}{2}A_k^{SS} \partial_{S_{t_k}}^2 \rlap{\ .}
$$
In order to compute the derivatives of expression \eqref{eq:intUcorrective} with respect to $S_{t_k}$ we consider the change of variable from $(t,S)$ to $(t,W)$ defined as in equation \eqref{eq:MCpath}:
\begin{equation*}
S = S_* + f_{(k)}(t-t_k,W-W_*) \rlap{\ .}
\end{equation*}
with $S_* = S_{t_k}$ and $W_* = W_{t_k}$. We introduce $S_*$ and $W_*$ notations to make clear that once the function $f$ is chosen, $S_*$ and $W_*$ are constant and should not be differentiated.
This change of variable induces for the first derivative
\begin{equation}
\label{eq:changederW}
\partial_W = ( \partial_W S ) \partial_S = \partial_{\Delta W} f_{(k)}(t-t_k,W-W_{t_k}) \, \partial_S = \hat\sigma \partial_S \rlap{\ .}
\end{equation}
Differentiating a second time with respect to $W$ we have
\begin{equation}
\label{eq:changederWW}
\partial_W^2 = \hat\sigma^2 \partial_S^2 + (\partial_W \hat \sigma) \partial_S = \hat\sigma^2 \partial_S^2 + (\partial_{\Delta W}^2 f_{(k)}) \partial_S \rlap{\ .}
\end{equation}
With our particular choice \eqref{eq:f1d} for the function $f$ this is
$$
\partial_W^2 = \hat\sigma^2 \partial_S^2 + \sigma(t_k,S_{t_k}) \partial_S \sigma(t_k,S_{t_k}) \partial_S \rlap{\ .}
$$
Inverting these equations we have
\begin{eqnarray*}
\partial_S &=& \frac{1}{\hat \sigma} \partial_W
\\
\partial_S^2 &=& \frac{1}{\hat \sigma^2} \partial_W^2 - \frac{\partial_W \hat\sigma}{\hat\sigma^3}\partial_W \rlap{\ .}
\end{eqnarray*}
Thus we can write in term of variable $W$
$$
\mathcal{A}_k = 1 + \frac{A_k^S}{ \hat \sigma(t_k^+,W_{t_k})} \partial_W + \frac{1}{2} \frac{A_k^{SS}}{\hat\sigma(t_k^+,W_{t_k})^2}\left[ \partial_W^2 - \frac{\partial_W \hat\sigma(t_k^+,W_{t_k})}{\hat\sigma(t_k^+,W_{t_k})} \partial_W \right] \rlap{\ .}
$$
Using again
\begin{eqnarray*}
\hat\sigma(t_k^+,W_{t_k}) &=& \sigma(t_k,S_{t_k})
\\
\partial_W \hat\sigma(t_k^+,W_{t_k}) &=& \sigma(t_k,S_{t_k}) \partial_S\sigma(t_k,S_{t_k}) \rlap{\ .}
\end{eqnarray*}
this becomes
\begin{equation}
\label{eq:deltaHoverLambda}
\mathcal{A}_k = 1 + \left[\frac{A_k^S}{\sigma(t_k,S_{t_k})} - \frac{1}{2} \frac{A_k^{SS} \partial_S\sigma(t_k,S_{t_k})}{\sigma(t_k,S_{t_k})^2}\right] \partial_W + \frac{1}{2} \frac{A_k^{SS}}{\sigma(t_k,S_{t_k})^2} \partial_W^2 \rlap{\ .}
\end{equation}
After the change of variables from $(t,S)$ to $(t,W)$, equation \eqref{eq:intUcorrective} becomes
\begin{equation}
\label{eq:intUcorrectiveW}
u_k(t_k,W_{t_k}) = \int \mathrm{d} W_{t_{k+1}} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) \left(1+\frac{\Delta \mathcal{H}_{k+1}}{\lambda}\right) u_{t_{k+1}}(W_{t_{k+1}}) \rlap{\ .}
\end{equation}
As $W$ is a Brownian motion, in term of the new variable, the evolution operator $\widehat{U}^{(W)}$ is the product of the discount factor by a Gaussian kernel:
$$
\widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) = e^{\textstyle -\int_{t_k}^{t_{k+1}} r(s) \mathrm{d} s} \varphi(t_{k+1}-t_k, W_{t_{k+1}}-W_{t_k})
$$
with
\begin{equation}
\label{eq:GaussKernel}
\varphi(\Delta t, \Delta W) = \frac{1}{\sqrt{2\pi \Delta t}} e^{\textstyle - \frac{\Delta W^2}{2 \Delta t}} \rlap{\ .}
\end{equation}
We have to compute the effect of the differential operator $ \mathcal{A}_k $ acting on $u_k$ given in formula \eqref{eq:intUcorrectiveW}, which we expand as
\begin{multline}
\label{eq:intUcorrectiveWexpanded}
u_k(t_k,W_{t_k}) = \int \mathrm{d} W_{t_{k+1}} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) u_{t_{k+1}}(W_{t_{k+1}})
\\
+ \int \mathrm{d} W_{t_{k+1}} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}})\frac{\Delta \mathcal{H}_{k+1}}{\lambda} u_{t_{k+1}}(W_{t_{k+1}})
\rlap{\ .}
\end{multline}
When acting on the second term in this formula, which contains $\frac{\Delta \mathcal{H}_{k+1}}{\lambda}$, the derivatives on $W_{t_k}$ are replaced by Malliavin weights:
\begin{eqnarray*}
\partial_{W_{t_k}} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) &=& \mathcal{W}_W(\Delta t_k, \Delta W_k) \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}})
\\
\partial_{W_{t_k}}^2 \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) &=& \mathcal{W}_{WW}(\Delta t_k, \Delta W_k) \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}})
\end{eqnarray*}
where the explicit form of the Gaussian kernel \eqref{eq:GaussKernel} gives
\begin{eqnarray*}
\mathcal{W}_W(\Delta t_k, \Delta W_k) &=& \frac{\Delta W_k}{\Delta t_k}
\\
\mathcal{W}_{WW}(\Delta t_k, \Delta W_k) &=& \frac{\Delta W_k^2 - \Delta t_k}{\Delta t_k^2} \rlap{\ .}
\end{eqnarray*}
For the derivatives of the first term of expression \eqref{eq:intUcorrectiveWexpanded} with respect to $W_{t_k}$, we use the symmetry of the Gaussian kernel $\varphi$, therefore of $\widehat{U}^{(W)}$, with respect to $W_{t_k}$ and $W_{t_{k+1}}$ to write
$$
\partial_{W_{t_k}} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) = - \partial_{W_{t_{k+1}}} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) \rlap{\ .}
$$
Then we integrate by part in first term of \eqref{eq:intUcorrectiveWexpanded} to get
\begin{multline*}
\partial_{W_{t_k}} \int \mathrm{d} W_{t_{k+1}} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) u_{t_{k+1}}(W_{t_{k+1}}) =
\\
\int \mathrm{d} W_{t_{k+1}} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) \partial_{W_{t_{k+1}}}\! u_{t_{k+1}}(W_{t_{k+1}})
\end{multline*}
and similarly for the second derivative.
Then we come back from $W$ to $S$ variable at time $t_{k+1}$ using \eqref{eq:changederW} and \eqref{eq:changederWW} which in our case read
\begin{eqnarray*}
\partial_{W_{t_{k+1}}} &=& \sigma(t_k,S_{t_k}) \big[ 1 + \partial_S \sigma(t_k,S_{t_k}) \Delta W _k
+ \partial_S \mu(t_k,S_{t_k}) \Delta t_k \big] \partial_{S_{t_{k+1}}}
\\
\partial_{W_{t_{k+1}}}^2 &=& \sigma(t_k,S_{t_k})^2 \big[ 1 + \partial_S \sigma(t_k,S_{t_k}) \Delta W _k
+ \partial_S \mu(t_k,S_{t_k}) \Delta t_k \big]^2 \partial_{S_{t_{k+1}}}^2
\\
&& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad
+ \sigma(t_k,S_{t_k}) \partial_S \sigma(t_k,S_{t_k}) \partial_{S_{t_{k+1}}} \rlap{\ .}
\end{eqnarray*}
Combining all terms, we finally obtain
\begin{multline*}
\mathcal{A}_k u_k = \int \mathrm{d} S_{t_{k+1}} \bigg( 1 + \bigg[\frac{A_k^S}{\sigma(t_k,S_{t_k})} - \frac{1}{2}\frac{A_k^{SS} \partial_S\sigma(t_k,S_{t_k})}{\sigma(t_k,S_{t_k})^2}\bigg] \frac{\Delta W_k}{\Delta t_k}
\\
\qquad\qquad\qquad
+ \frac{1}{2}\frac{A_k^{SS}}{\sigma(t_k,S_{t_k})^2} \frac{\Delta W_k^2 - \Delta t_k}{\Delta t_k^2} \bigg) \widehat{U}_{t_k,t_{k+1}}(S_{t_k},S_{t_{k+1}}) \frac{\Delta \mathcal{H}_{k+1}}{\lambda} u_{t_{k+1}}(S_{t_{k+1}})
\\
+ \int \mathrm{d} S_{t_{k+1}} \widehat{U}_{t_k,t_{k+1}}(S_{t_k},S_{t_{k+1}})
\bigg( 1 + \bigg[A_k^S - \frac{1}{2}\frac{A_k^{SS} \partial_S\sigma(t_k,S_{t_k})}{\sigma(t_k,S_{t_k})}\bigg]
\qquad\qquad
\\
\big[ 1 + \partial_S \sigma(t_k,S_{t_k}) \Delta W _k
+ \partial_S \mu(t_k,S_{t_k}) \Delta t_k \big] \partial_{S_{t_{k+1}}}
\\
+ \frac{1}{2}A_k^{SS} \big[ 1 + \partial_S \sigma(t_k,S_{t_k}) \Delta W _k + \partial_S \mu(t_k,S_{t_k}) \Delta t_k \big]^2 \partial_{S_{t_{k+1}}} ^2
\\
+ \frac{1}{2}\frac{A_k^{SS} \partial_S\sigma(t_k,S_{t_k})}{\sigma(t_k,S_{t_k})} \partial_{S_{t_{k+1}}}
\bigg)
u_{t_{k+1}
}(S_{t_{k+1}}) \rlap{\ .}
\end{multline*}
We can rewrite this as
$$
\mathcal{A}_k u_k = \int \mathrm{d} S_{t_{k+1}} \widehat{U}_{t_k,t_{k+1}}(S_{t_k},S_{t_{k+1}}) \mathcal{A}_{k+1} u_{t_{k+1}}(S_{t_{k+1}})
$$
with
$$
\mathcal{A}_{k+1}= 1 + A_{k+1}^S \partial_{S_{t_{k+1}}} + \frac{1}{2} A_{k+1}^{SS} \partial_{S_{t_{k+1}}}^2
$$
\begin{eqnarray}
\label{eq:recursiveA}
A_{k+1}^S &=& \big[ 1 + b_k\big] A_k^S - \frac{1}{2} b_k \frac{ \partial_S\sigma(t_k,S_{t_k})}{\sigma(t_k,S_{t_k})} A_k^{SS}
+ d_k(\Delta t_k, \Delta W_k) \frac{\Delta \mu_{k+1}}{\lambda}
\\
A_{k+1}^{SS}&=& \big[ 1 + b_k\big]^2 A_k^{SS} + d_k(\Delta t_k, \Delta W_k) \frac{\Delta C_{k+1}}{\lambda}
\nonumber
\end{eqnarray}
and
$$
b_k = \partial_S \sigma(t_k,S_{t_k}) \Delta W_k + \partial_S \mu(t_k,S_{t_k}) \Delta t_k
$$
\begin{multline}
d_k(\Delta t,\Delta W) = 1 + \bigg[\frac{A_k^S}{\sigma(t_k,S_{t_k})} - \frac{1}{2}\frac{A_k^{SS} \partial_S\sigma(t_k,S_{t_k})}{\sigma(t_k,S_{t_k})^2}\bigg] \frac{\Delta W}{\Delta t}
\\
+ \frac{1}{2}\frac{A_k^{SS}}{\sigma(t_k,S_{t_k})^2} \frac{\Delta W^2 - \Delta t}{\Delta t^2} \ .
\label{eq:dk}
\end{multline}
Using formulas \eqref{eq:recursiveA}, we can recursively accumulate corrective terms. We start with $\mathcal{A}_0 = 1$ at $t=t_0$. Then starting from each date $t_k$ we simulate $S_{t_{k+1}}$ as given by equation \eqref{eq:MCpath} and we compute $\mathcal{A}_{k+1}$ as defined above, up to the last Poisson time $t_p$.
\subsubsection{Payoff}
\label{sec:payoff}
On the final Poisson time $t_p$, $\mathcal{A}_p$ should act on
$$
u_{t_p}(S_{t_p}) = \int \mathrm{d} S_T\, \widehat{U}_{t_p,T}(S_{t_p},S_T) h(S_T)
$$
where $h$ is the payoff function.
A naive approach would be to the following. One simulates $S_T$ from $S_{t_p}$ as given in equation \eqref{eq:MCpath}:
$$
S_T = S_{t_p} + f_{(p)}(\Delta t_p, \Delta W_p)
$$
with $\Delta t_p = T-t_p$ and $\Delta W_p = W_T - W_{t_p}$ a Gaussian variable of variance $\Delta t_p$. Then one computes the payoff $h(S_T)$ and computes the derivatives using Malliavin weights. This means multiplying the discounted payoff by $d_p(\Delta t_p, \Delta W_p)$ given in equation \eqref{eq:dk}:
$$
P_T = d_p(\Delta t_p, \Delta W_p) h\big(S_T\big) \rlap{\ .}
$$
However this term behaves as $O\!\left(\frac{1}{\Delta t_p}\right)$ at small $\Delta t_p$ and not $O(1)$ as we want.
Instead, we will use antithetic sampling on this last time step, as in \cite{henry2015exact}. More precisely, we compute
\begin{eqnarray}
\nonumber
S_T^{(+)} &=& S_{t_p} + f_{(p)}(\Delta t_p, \Delta W_p)
\\
S_T^{(0)} &=& \hat{\mathbb{E}}_{t_p}\big[ S_T \big] = S_{t_p} + \mu(t_p,S_{t_p}) \Delta t_p
\label{eq:S+S-S0}
\\
S_T^{(-)} &=& S_{t_p} + f_{(p)}(\Delta t_p, -\Delta W_p)
\nonumber
\end{eqnarray}
and then
\begin{multline*}
P_T = \frac{1}{2} d_p(\Delta t_p, \Delta W_p) h\big(S_T^{(+)}\big)
+ \frac{1}{2} d_p(\Delta t_p, -\Delta W_p) h\big(S_T^{(-)}\big)
\\
+ \left[1-\frac{1}{2} d_p(\Delta t_p, \Delta W_p) - \frac{1}{2} d_p(\Delta t_p, -\Delta W_p)\right] h\big(S_T^{(0)}\big)
\end{multline*}
which simplifies to
\begin{multline}
\label{eq:pathContrib}
P_T = \frac{1}{2} d_p(\Delta t_p, \Delta W_p) h\big(S_T^{(+)}\big)
+ \frac{1}{2} d_p(\Delta t_p, -\Delta W_p) h\big(S_T^{(-)}\big)
\\
- \frac{1}{2} \frac{A_p^{SS}}{\sigma(t_p,S_{t_p})^2} \frac{\Delta W_p^2 - \Delta t_p}{\Delta t_p^2} h\big(S_T^{(0)}\big) \ .
\end{multline}
The last term in $h\big(S_T^{(0)}\big)$ has expected value $0$ and the two first terms are antithetic contributions to the option price.
If the payoff function $h$ is smooth and has a Taylor expansion, one can check that $P_T$ is of order $O(1)$. This is not true for a Call or Put option payoff in the vicinity of the strike. In this case we have $P_T = O\!\left(\frac{1}{\sqrt{\Delta t_p}}\right)$. However, the probability to have the strike between $S_T^{(-)}$ and $S_T^{(+)}$ at the last time step scales as $O(\sqrt{\Delta t_p})$. As a consequence, the variance remains finite.
We finally discount on all time steps. As we take a deterministic discount rate in this first case, this factors out as a multiplication by
$
e^{\textstyle - \int_{t_0}^T r(t) \mathrm{d} t} \rlap{\ .}
$
\subsection{Monte Carlo scheme summary}
We now sum up the whole Monte Carlo scheme in this one-dimensional case with deterministic discount factor. We consider a European option with maturity $T$ and payoff $h(S_T)$.
\newline
For each path, we do the following:
\begin{enumerate}
\item Start from $S = S_0$ at time $t=t_0$. Define
\begin{eqnarray*}
A_0^S &=& 0
\\
A_0^{SS} &=& 0 \rlap{\ .}
\end{eqnarray*}
\item On each date $t_k$
\begin{enumerate}
\item
Draw the next Poisson time $t_{k+1} = t_k + \Delta t_k$ with intensity $\lambda$. For example draw a random uniform variable $q$ between 0 and 1 and take $\Delta t_k = -\frac{\log(q)}{\lambda}$.
\\
If $t_{k+1} > T$, set $p=k$ and go to step 3.
\item
Draw a Gaussian variable $\Delta W_k$ with variance $\Delta t_k$. Get $S_{t_{k+1}}$ by equations \eqref{eq:f1d} and \eqref{eq:MCpath}:
$$
S_{t_{k+1}} = S_{t_k} + f_{(k)}(\Delta t_k, \Delta W_k) \rlap{\ .}
$$
with
\begin{multline*}
f_{(k)}(\Delta t, \Delta W) = \mu(t_k,S_{t_k}) \Delta t + \sigma(t_k,S_{t_k}) \Delta W
\\
+ \frac{1}{2} \sigma(t_k,S_{t_k}) \partial_S \sigma(t_k,S_{t_k}) (\Delta W^2 - \Delta t)
\\
+ \sigma(t_k,S_{t_k}) \partial_S \mu(t_k,S_{t_k}) \Delta t \Delta W \ .
\label{eq:f1d}
\end{multline*}
\item Compute $\Delta \mu_k$ and $\Delta C_k$ according to equation \eqref{eq:deltaMudeltaC}:
\begin{eqnarray*}
\Delta \mu_k
&=& \mu(t_k,S_{t_k})
\\&& \qquad
- \big[ \mu(t_{k-1},S_{t_{k-1}}) + \sigma(t_{k-1},S_{t_k}) \partial_S \mu(t_{k-1},S_{t_{k-1}}) \Delta W_{k-1} \big]
\\
\Delta C_k
&=& \sigma(t_k,S_{t_k})^2
- \sigma(t_{k-1},S_{t_{k-1}})^2 \big[ 1+ \partial_S \sigma(t_{k-1},S_{t_{k-1}}) \Delta W_{k-1}
\\
&& \qquad\qquad\qquad\qquad\qquad\qquad\qquad
+ \partial_S \mu(t_{k-1},S_{t_{k-1}}) \Delta t_{k-1} \big]^2 \rlap{\ .}
\end{eqnarray*}
\item
Compute $A_{k+1}^S$ and $A_{k+1}^{SS}$ from $A_k^S$ and $A_k^{SS}$ as in equation \eqref{eq:recursiveA}:
\begin{eqnarray*}
A_{k+1}^S &=& \big[ 1 + b_k\big] A_k^S - b_k \frac{ \partial_S\sigma(t_k,S_{t_k})}{\sigma(t_k,S_{t_k})} A_k^{SS}
+ d_k(\Delta t_k, \Delta W_k) \frac{\Delta \mu_{k+1}}{\lambda}
\\
A_{k+1}^{SS}&=& \big[ 1 + b_k\big]^2 A_k^{SS} + \frac{1}{2} d_k(\Delta t_k, \Delta W_k) \frac{\Delta C_{k+1}}{\lambda}
\end{eqnarray*}
with
$$
b_k = \partial_S \sigma(t_k,S_{t_k}) \Delta W_k + \partial_S \mu(t_k,S_{t_k}) \Delta t_k
$$
\begin{multline*}
d_k(\Delta t,\Delta W) = 1 + \bigg[\frac{A_k^S}{\sigma(t_k,S_{t_k})} - \frac{A_k^{SS} \partial_S\sigma(t_k,S_{t_k})}{\sigma(t_k,S_{t_k})^2}\bigg] \frac{\Delta W}{\Delta t}
\\
+ \frac{A_k^{SS}}{\sigma(t_k,S_{t_k})^2} \frac{\Delta W^2 - \Delta t}{\Delta t^2} \ .
\end{multline*}
\end{enumerate}
\item On time $t_p$
\begin{enumerate}
\item
Draw a Gaussian variable $\Delta W_p$ with variance $\Delta t_p = T-t_p$.
Compute $S_T^{(+)}$, $S_T^{(0)}$ and $S_T^{(-)}$ as in equation \eqref{eq:S+S-S0}:
\begin{eqnarray*}
S_T^{(+)} &=& S_{t_p} + f_{(p)}(\Delta t_p, \Delta W_p)
\\
S_T^{(0)} &=& S_{t_p} + \mu(t_p,S_{t_p}) \Delta t_p
\\
S_T^{(-)} &=& S_{t_p} + f_{(p)}(\Delta t_p, -\Delta W_p) \rlap{\ .}
\end{eqnarray*}
\item
Get the undiscounted path contribution \eqref{eq:pathContrib} for the payoff $h(S_T)$:
\begin{multline*}
P_T = \frac{1}{2} d_p(\Delta t_p, \Delta W_p) h\big(S_T^{(+)}\big)
+ \frac{1}{2} d_p(\Delta t_p, -\Delta W_p) h\big(S_T^{(-)}\big)
\\
- \frac{A_p^{SS}}{\sigma(t_p,S_{t_p})^2} \frac{\Delta W_p^2 - \Delta t_p}{\Delta t_p^2} h\big(S_T^{(0)}\big) \ .
\end{multline*}
\end{enumerate}
\item Multiply by the discount factor:
$$
e^{\textstyle - \int_{t_0}^T r(t) \mathrm{d} t} P_T \rlap{\ .}
$$
\end{enumerate}
We finally average over all path to get the unbiased Monte Carlo estimation of the option price.
\section{Numerical results}
\label{sec:numerical}
\subsection{Convergence}
\label{sec:convergence}
In order to test numerically our Monte Carlo scheme, we take a model for which we can have a closed formula as a reference value. We therefore choose the Black-Scholes model
$$
\mathrm{d} S = \mu_0 S \mathrm{d} t + \sigma_0 S \mathrm{d} W_t
$$
which corresponds to local drift an volatility
\begin{eqnarray*}
\mu(t,S) &=& \mu_0 S
\\
\sigma(t,S) &=& \sigma_0 S \rlap{\ .}
\end{eqnarray*}
There already exist an unbiased Monte Carlo scheme for this model, using $\log(S)$ as a variable. For the purpose of ours tests we ignore this and we naively apply our scheme and compare it to a Euler scheme.
We consider an underlying with spot $S_0 = 100$ and volatility $\sigma_0 = 50\%$. We suppose the interest rate and the drift are $r = \mu_0 = 5\%$. We price a Put option of maturity $T=1$~year at strike $K=80$. The Black-Scholes formula gives an option price of 7.8909.
Figure \ref{fig:convergence} shows the simulated value as a function of the number of paths for the unbiased scheme with $\lambda=3$.
\begin{figure}[!htbp]
\begin{center}
\includegraphics[trim = 20mm 70mm 20mm 70mm, clip, width=.75\textwidth]{Images/MC-100-5-50-80-200k}
\caption{Convergence of the simulated option price with respect to the number of paths for the unbiased scheme and the Euler scheme.}
\label{fig:convergence}
\end{center}
\end{figure}
For comparison, it also shows the same convergence graph for a Euler scheme with the same average number of time steps $n=4$. We see that this latter scheme converges to a biased value.
\subsection{Comparison with Euler and Milstein scheme}
When using the Euler scheme, the bias can be decreased using more time steps. This in turn increases the computation time. On the other hand, using the unbiased scheme is done at the cost of increasing the variance for small values of $\lambda$. Increasing the value of $\lambda$ makes the variance smaller but also increases the computation time, as the average number of time steps is higher.
In order to assess the performance gain which can be achieved, we consider the 80\% Put option described in the section \ref{sec:convergence} and we price it with both schemes. As the path generation has similarities with the Milstein scheme, we also compare with it. This allows to better isolate the effect of corrective terms in the unbiased scheme.
For the unbiased scheme, we use increasing values of $\lambda$ from .01 to 29. For Euler and Milstein scheme, we take increasing numbers of time steps from 1 to 300. In all cases we draw $N=1$~million paths and compute numerically the estimated option price, the empirical standard deviation and the computation time.
The results are shown in figure \ref{fig:unbiased-Euler}.
The estimated prices are plotted with $99\%$ confidence interval against the computation time in logarithmic scale.
\begin{figure}[!htbp]
\begin{center}
\includegraphics[width=\textwidth]{Images/UnbiasedVsEuler}
\caption{Simulated price of a European Put obtained using the unbiased scheme, compared to Euler and Milstein schemes with the same number of Monte Carlo paths. Prices are plotted with 99\% confidence interval against the computation time in logarithimic scale.}
\label{fig:unbiased-Euler}
\end{center}
\end{figure}
At very low Poisson intensity ($\lambda = 0.01$), the Monte Carlo variance of the unbiased scheme is higher than the Euler and Milstein schemes. However it quickly decreases when $\lambda$ increases. For $\lambda \geq 0.3$ it becomes even lower than noise of the Euler and Milstein schemes. Beyond $\lambda =1$ the Monte Carlo noise is almost stationary: the corrective terms add negligible variance compared to the basic variance of the payoff. For all values of $\lambda$, we check that the final estimate is consistent with the theoretical value, up to the Monte Carlo statistical error. The value $\lambda=1$ which appears to be optimal with respect to computation time corresponds to an average number of time steps $n=2$ over the period of 1 year.
On the opposite, the Euler scheme exhibits a large bias when the number of time steps is small. This bias decreases linearly with $\Delta T = T/n$. Performing a weighted least squares regression, we estimate the bias to behave asymptotically as $\frac{2.99}{n}$. In order to have a bias equal to the Euler Monte Carlo standard deviation $0.13$, we thus need $n \sim 230$ time steps. This corresponds to a computation time 43 times longer than the unbiased scheme with $\lambda = 1$.
The Milstein scheme has a smaller bias, asymptotically $\frac{-0.648}{n}$ in this example. We thus need $n=50$ time steps to get a bias of the same magnitude as the Monte Carlo noise. This means a computation 10 times slower than the unbiased scheme.
\section{Multidimensional process}
\label{sec:multidimensional}
We consider now the general multi-dimensional case, where all parameters can depend on $t$ and $X$ in the parabolic PDE \eqref{eq:PDE}:
\begin{equation}
\partial_t u_t(X_t) + \mu^\alpha(t,X_t) \partial_\alpha u_t(X_t) + \frac{1}{2} C^{\alpha\beta}(t,X_t) \partial_\alpha \partial_\beta u_t(X_t) = r(t,X_t) u_t(X_t) \rlap{\ .}
\label{eq:PDEmulti}
\end{equation}
According to Feynman-Kac theorem, it corresponds to the multi-dimensional process
$$
\mathrm{d} X_t^\alpha = \mu^\alpha(t,X_t) \mathrm{d} t + \sigma^\alpha{}_a(t,X_t) \mathrm{d} W_t^a
$$
with stochastic discount rate $r(t,X_t)$. $\sigma(t,X_t) $ is a volatility matrix such that
$$
C^{\alpha\beta}(t,X_t) = \sigma^\alpha{}_a(t,X_t) \sigma^\beta{}_a(t,X_t)
$$
which can be obtained by Cholesky decomposition.
$W_t^a$ are independent standard Brownian processes.
\subsection{Monte Carlo path}
We draw Poisson times $t_k$ with intensity $\lambda$. Between these times, we simulate a $d$-dimensional Brownian process $W_t$.
This corresponds to the parabolic PDE
\begin{equation}
\label{eq:PDEBRownian}
\partial_t v_t(W_t) + \frac{1}{2} \delta^{ab} \partial_a \partial_b v_t(W_t) = 0 \rlap{\ .}
\end{equation}
Between two Poisson times $t_k$ and $t_{k_+1}$, we will consider two change of variables\footnote{If we see the total space made of space variables and prices as a fiber bundle, this corresponds to a change of variables on the base space and the fiber.}:
\begin{itemize}
\item \textbf{Space variables.} We go from $(t,W)$ to $(t,X)$, using functions $f^\alpha_{(k)}(\Delta t, \Delta W)$:
$$
X_t^\alpha = X_{t_k}^\alpha + f^\alpha_{(k)}(t-t_k,W_t-W_{t_k}) \rlap{\ .}
$$
\item \textbf{Num\'eraire.} We change the num\'eraire, using a function $g_{(k)}(\Delta t, \Delta W)$
$$
u_t = v_t e^{\textstyle g_{(k)}(t-t_k,W_t-W_{t_k})} \rlap{\ .}
$$
\end{itemize}
The change of num\'eraire transfers to derivatives as\footnote{We drop $(k)$ indices when they are not needed in order to simplify the notation.}
\begin{eqnarray*}
\partial_t v_t &=& e^{-g} \partial_t u_t - (\partial_t g) \, e^{-g} u_t
\\
\partial_a v_t &=& e^{-g} \partial_a u_t - (\partial_a g) \, e^{-g} u_t
\\
\partial_a \partial_b v_t &=& e^{-g} \partial_a \partial_b u_t - (\partial_a\partial_b g) \, e^{-g} u_t
+ (\partial_a g)(\partial_b g) \, e^{-g} u_t
\\ && \qquad\qquad\qquad\qquad\qquad
- (\partial_a g) \, e^{-g} \partial_b u_t
- (\partial_b g) \, e^{-g} \partial_a u_t \rlap{\ .}
\end{eqnarray*}
Using the notations
$$
\partial_a = \partial_{W^a} \qquad \partial_\alpha = \partial_{X^\alpha}
$$
we define
$$
e^\alpha{}_a = \partial_a f^\alpha
$$
and
$$
c^\alpha{}_{ab} = \partial_a \partial_b f^\alpha \rlap{\ .}
$$
We will also use the inverse matrix $e^a{}_\alpha$, with defining properties
$$
e^\alpha{}_a e^a{}_\beta = \delta^\alpha_\beta \qquad e^a{}_\alpha e^\alpha{}_b = \delta^a_b \rlap{\ .}
$$
The change of space variables induces the following transformations on derivatives:
\begin{eqnarray*}
\partial_a &=& e^\alpha{}_a \, \partial_\alpha
\\
\partial_a \partial_b &=& e^\alpha{}_a e^\beta{}_b \, \partial_\alpha \partial_\beta + c^\alpha{}_{ab} \, \partial_\alpha
\\
\partial_t\!\!\mid_W &=& \partial_t\!\!\mid_X + \partial_t f^\alpha \, \partial_\alpha \rlap{\ .}
\end{eqnarray*}
In term of the new variables $t$, $X$ and $u_t$, PDE \eqref{eq:PDEBRownian} thus becomes
\begin{equation}
\partial_t u_t + \hat\mu^\alpha \partial_\alpha u_t + \frac{1}{2} \hat C^{\alpha\beta} \partial_\alpha \partial_\beta u_t = \hat r u_t
\label{eq:PDEhat}
\end{equation}
with
\begin{eqnarray}
\hat\mu^\alpha &=& \partial_t f^\alpha - (\partial_a g) e^\alpha{}_a + \frac{1}{2} c^\alpha{}_{aa}
\nonumber
\\
\hat C^{\alpha\beta} &=& e^\alpha{}_a e^\beta{}_a
\label{eq:hatMuCr}
\\
\hat r &=& \partial_t g + \frac{1}{2}\partial_a\partial_a g - \frac{1}{2} (\partial_a g) (\partial_a g) \rlap{\ .}
\nonumber
\end{eqnarray}
We want to find functions $f^\alpha_{(k)}(\Delta t, \Delta W)$ and $g_{(k)}(\Delta t, \Delta W)$ such that $\mu - \hat\mu$, $C - \hat C$ and $r - \hat r$ behaves as $O(\Delta t)$ at small $\Delta t$.
We suppose that we can expand $f^\alpha_{(k)}$ and $g_{(k)}$ as power series in $\Delta t$ and $\Delta W$:
\begin{eqnarray*}
f^\alpha_{(k)}(\Delta t, \Delta W) &=& f^\alpha_{00} + (f^\alpha_{01})_a \Delta W^a + f^\alpha_{10} \Delta t + \frac{1}{2!} (f^\alpha_{02})_{ab} \Delta W^a \Delta W^b
\\ && \qquad\qquad
+ (f^\alpha_{11})_a \Delta t\Delta W^a + \frac{1}{3!} (f^\alpha_{03})_{abc} \Delta W^a \Delta W^b \Delta W^c + \cdots
\\
g_{(k)}(\Delta t, \Delta W) &=& g_{00} + (g_{01})_a \Delta W^a + g_{10} \Delta t + \frac{1}{2!} (g_{02})_{ab} \Delta W^a \Delta W^b
\\ && \qquad\qquad
+ (g_{11})_a \Delta t\Delta W^a + \frac{1}{3!} (g_{03})_{abc} \Delta W^a \Delta W^b \Delta W^c + \cdots
\end{eqnarray*}
In this expression, $(f^\alpha_{ij})_{abc\cdots}$ or $(g_{ij})_{abc\cdots}$ are tensors which are symmetric in space indices $a,b,c, \cdots$.
The continuity contraint at $t=t_k$ is $f^\alpha(0,0) = 0$. It translates to
$$
f^\alpha_{00} = 0 \rlap{\ .}
$$
In addition, we choose num\'eraires such that $u_t$ and $v_t$ coincide at the beginning of the period, $t=t_k$. Mathematically this is
$g(0,\Delta W) = 0$, which gives for all $j$
$$
(g_{0j}) = 0 \rlap{\ .}
$$
Taking into account these constraints, equations \eqref{eq:hatMuCr} have Taylor expansions
\begin{eqnarray}
\hat \mu^\alpha &=& f^\alpha_{10} + (f^\alpha_{11})_a \Delta W^a + \frac{1}{2} (f^\alpha_{02})_{bb} + \frac{1}{2} (f^\alpha_{03})_{bba} \Delta W^a + O(\Delta t)
\label{eq:order1muhat}
\\
\hat C^{\alpha\beta} &=& (f^\alpha_{01})_b (f^\beta_{01})_b + (f^\alpha_{01})_b (f^\beta_{02})_{ba} \Delta W^a
+ (f^\beta_{01})_b (f^\alpha_{02})_{ba} \Delta W^a + O(\Delta t)
\label{eq:order1Chat}
\\
\hat r &=& g_{10} + (g_{11})_a \Delta W^a + O(\Delta t) \rlap{\ .}
\label{eq:order1rhat}
\end{eqnarray}
On the other hand, a Taylor expansion of the parameters of PDE \eqref {eq:PDEmulti} gives
\begin{eqnarray}
\mu^\alpha &=& \mu^\alpha(t_k,X_{t_k}) + e^\gamma{}_a(t_k,W_{t_k}) \partial_\gamma \mu^\alpha(t_k,X_{t_k}) \Delta W^a + O(\Delta t)
\label{eq:order1mu}
\\
C^{\alpha\beta} &=& C^{\alpha\beta}(t_k,X_{t_k}) + e^\gamma{}_a(t_k,W_{t_k}) \partial_\gamma C^{\alpha\beta}(t_k,X_{t_k}) \Delta W^a + O(\Delta t)
\label{eq:order1C}
\\
r &=& r(t_k,X_{t_k}) + e^\gamma{}_a(t_k,W_{t_k}) \partial_\gamma r(t_k,X_{t_k}) \Delta W^a + O(\Delta t) \rlap{\ .}
\label{eq:order1r}
\end{eqnarray}
From its definition, $e^\gamma{}_a$ is
$$
e^\gamma{}_a = \partial_a f^\alpha = (f^\gamma_{01})_a + (f^\gamma_{02})_{ab} \Delta W^b + O(\Delta t) \rlap{\ .}
$$
At the beginning of the period, with $\Delta t = 0$ and $\Delta W = 0$, this is
$$
e^\gamma{}_a(t_k,W_{t_k}) = (f^\gamma_{01})_a \rlap{\ .}
$$
We want to make $\mu-\hat\mu$, $C-\hat C$ and $r-\hat r$ vanish up to $O(\Delta t)$ terms.
Let us start with the constant term in equations \eqref{eq:order1Chat} and \eqref{eq:order1C}. From a Cholesky decomposition
$$
C^{\alpha\beta}(t_k,X_{t_k}) = \sigma^\alpha{}_b(t_k,X_{t_k}) \sigma^\beta{}_b(t_k,X_{t_k})
$$
we get a solution
\begin{equation}
\label{eq:fa01}
(f^\alpha_{01})_b = \sigma^\alpha{}_b(t_k,X_{t_k}) \rlap{\ .}
\end{equation}
This also gives
$$
e^\alpha{}_a(t_k,W_{t_k}) = \sigma^\alpha{}_a(t_k,X_{t_k}) \rlap{\ .}
$$
Its inverse is
$$
e^a{}_\alpha(t_k,W_{t_k}) = \big[\sigma(t_k,W_{t_k})^{-1}\big]^a{}_\alpha \rlap{\ .}
$$
Equating the first order terms of equations \eqref{eq:order1Chat} and \eqref{eq:order1C}, we have
\begin{equation*}
e^\alpha{}_b(t_k,W_{t_k}) (f^\beta_{02})_{ba}
+ e^\beta{}_b(t_k,W_{t_k}) (f^\alpha_{02})_{ba} = e^\gamma{}_a(t_k,W_{t_k}) \partial_\gamma C^{\alpha\beta}(t_k,X_{t_k})
\end{equation*}
where $C^{\alpha\beta}$ is symmetric in indices $\alpha$ and $\beta$ and $(f^\beta_{02})_{ab}$ in indices $a$ and $b$.
We multiply this equation by $e^c{}_\alpha(t_k,W_{t_k})$ and $e^d{}_\beta(t_k,W_{t_k})$, also using $e^a{}_\alpha e^\alpha{}_b = \delta^a_b$:
\begin{equation}
f^d{}_{ca} + f^c{}_{da} = C_a{}^{cd}
\label{eq:efefedC}
\end{equation}
where we introduced the notations
\begin{eqnarray*}
f^a{}_{bc} &=& e^a{}_\alpha(t_k,W_{t_k}) (f^\alpha_{02})_{bc}
\\
C_a{}^{bc} &=& e^\alpha{}_a(t_k,W_{t_k}) e^b{}_\beta(t_k,W_{t_k})e^c{}_\gamma(t_k,W_{t_k}) \partial_\alpha C^{\beta\gamma}(t_k,X_{t_k}) \rlap{\ .}
\end{eqnarray*}
$f^a{}_{bc}$ and $C_a{}^{bc}$ are tensors symmetric in the two last indices.
We write equation \eqref{eq:efefedC} for the three cyclic permutations of the indices and get
\begin{eqnarray}
f^b{}_{ca} + f^c{}_{ba} &=& C_a{}^{bc}
\label{eq:ffC1}
\\
f^c{}_{ab} + f^a{}_{cb} &=& C_b{}^{ca}
\label{eq:ffC2}
\\
f^a{}_{bc} + f^b{}_{ac} &=& C_c{}^{ab} \rlap{\ .}
\label{eq:ffC3}
\end{eqnarray}
The linear combination \eqref{eq:ffC3} + \eqref{eq:ffC2} - \eqref{eq:ffC1} then gives
$$
f^a{}_{bc} = \frac{1}{2} ( C_c{}^{ab} + C_b{}^{ca} - C_a{}^{bc} ) \rlap{\ .}
$$
Inverting the definition of $f^a{}_{bc}$ in term of $(f^\alpha_{02})_{bc}$ we get
\begin{eqnarray}
(f^\alpha_{02})_{bc} &=& \frac{1}{2} \sigma^\alpha{}_a(t_k,X_{t_k}) ( C_c{}^{ab} + C_b{}^{ca} - C_a{}^{bc} )
\nonumber
\\
&=& \frac{1}{2} \Big[ \phantom{+}\ \sigma^\gamma{}_c(t_k,X_{t_k}) e^b{}_\beta(t_k,X_{t_k}) \partial_\gamma C^{\alpha\beta}(t_k,X_{t_k})
\nonumber
\\
\label{eq:fa02}
&& \phantom{\frac{1}{2} \Big[}
+ \sigma^\beta{}_b(t_k,X_{t_k}) e^c{}_\gamma(t_k,X_{t_k}) \partial_\beta C^{\alpha\gamma}(t_k,X_{t_k})
\\
\nonumber
&& \phantom{\frac{1}{2} \big[}
- e^b{}_\beta(t_k,X_{t_k}) e^c{}_\gamma(t_k,X_{t_k}) C^{\alpha\delta}(t_k,X_{t_k}) \partial_\delta C^{\beta\gamma}(t_k,X_{t_k}) \Big]
\rlap{\ .}
\end{eqnarray}
Equating terms in equations \eqref{eq:order1muhat} and \eqref{eq:order1mu} we get
\begin{eqnarray*}
f^\alpha_{10} &=& \mu^\alpha(t_k,W_{t_k}) - \frac{1}{2}(f^\alpha_{02})_{bb}
\\
(f^\alpha_{11})_a &=& \sigma^\gamma{}_a(t_k,X_{t_k}) \partial_\gamma \mu^\alpha(t_k,W_{t_k}) - \frac{1}{2}(f^\alpha_{03})_{bba} \rlap{\ .}
\end{eqnarray*}
Using the property $\Tr ( M^{-1} \partial_\alpha M) = \partial_\alpha\! \log(\det(M))$ we rewrite $(f^\alpha_{02})_{bb}$ as
$$
(f^\alpha_{02})_{bb} = \partial_\beta C^{\alpha\beta}(t_k,W_{t_k}) - \frac{1}{2} C^{\alpha\gamma}(t_k,W_{t_k}) \partial_\gamma \!\log(\det(C))(t_k,W_{t_k})
$$
and get
\begin{equation}
\label{eq:fa10}
f^\alpha_{10} = \mu^\alpha(t_k,W_{t_k}) - \frac{1}{2} \partial_\beta C^{\alpha\beta}(t_k,W_{t_k})
+\frac{1}{4} C^{\alpha\gamma}(t_k,W_{t_k}) \partial_\gamma \!\log(\det(C))(t_k,W_{t_k}) \ .
\end{equation}
Making the additional choice $f^\alpha_{ij} = 0$ for $f_{ij} = 0$ for $(i=0, j\geq 3)$, $(i=1, j \geq 2)$ and $(i \geq 2)$, the expression for $(f^\alpha_{11})_a$ simplifies to
\begin{equation}
\label{eq:fa11}
(f^\alpha_{11})_a = \sigma^\gamma{}_a(t_k,X_{t_k}) \partial_\gamma \mu^\alpha(t_k,W_{t_k}) \rlap{\ .}
\end{equation}
This completes the definition of $f_{(k)}^\alpha$ as a polynomial in $\Delta t$ and $\Delta W$
$$
f^\alpha_{(k)}(\Delta t, \Delta W) = (f^\alpha_{01})_a \Delta W^a + f^\alpha_{10} \Delta t + \frac{1}{2} (f^\alpha_{02})_{ab} \Delta W^a \Delta W^b
+ (f^\alpha_{11})_a \Delta t\Delta W^a
$$
with all coefficients defined in equations \eqref{eq:fa01}, \eqref{eq:fa10}, \eqref{eq:fa02} and \eqref{eq:fa11}.
Finally, equating equations \eqref{eq:order1rhat} and \eqref{eq:order1r} we get
\begin{eqnarray}
\label{eq:g10}
g_{10} &=& r(t_k,X_{t_k})
\\
\label{eq:g11}
(g_{11})_a &=& \sigma^\gamma{}_a(t_k,X_{t_k}) \partial_\gamma r(t_k,X_{t_k}) \rlap{\ .}
\end{eqnarray}
Choosing all other $g_{ij}$ coefficients to be 0, we also get $g_{(k)}$ as a polynomial in $\Delta t$ and $\Delta W$
$$
g_{(k)}(\Delta t, \Delta W) = g_{10} \Delta t + (g_{11})_a \Delta t \Delta W^a \rlap{\ .}
$$
Using function $f^\alpha_{(k)}$ and $g_{(k)}$ we are able to solve the PDE \eqref{eq:PDEhat} by Monte Carlo simulation. At time $t_k$ we draw $d$ independent Gaussian variable $\Delta W^a_k, 1 \geq a \geq d$ with variance $\Delta t_k = t_{k+1}-t_k$. Then we compute the value of the system at the following date
$$
X^\alpha_{t_{k+1}} = X^\alpha_{t_k} + f^\alpha_{(k)}(\Delta t_k,\Delta W_k) \rlap{\ .}
$$
The (stochastic) discount factor between $t_k$ and $t_{k+1}$ is
$$
D(t_k,t_{k+1}) = e^{\textstyle -g_{(k)}(\Delta t_k,\Delta W_k)} \rlap{\ .}
$$
We reccursively multply it to get the discount factor up from $t_0$ to $t_{k+1}$ as
$$
D_{k+1} =D_k D(t_k,t_{k+1}) = D_k e^{\textstyle -g_{(k)}(\Delta t_k,\Delta W_k)} \rlap{\ ,}
$$
starting from $D_0 = 1$.
\subsection{Corrective terms}
From the definition of $f^\alpha_{(k)}$ and $g_{(k)}$, between $t_k$ ad $t_{k+1}$,
we have
$$
e^\alpha_{(k)a}(\Delta t, \Delta W) = (f^\alpha_{01})_a + (f^\alpha_{02})_{ab} \Delta W^b + (f^\alpha_{11})_a \Delta t
$$
Using the definitions of coefficients, this is
\begin{equation}
\label{eq:edtdw}
e^\alpha_{(k)a}(\Delta t, \Delta W) = \sigma^\alpha{}_a(t_k,X_{t_k})
+ c^\alpha_{(k)ab} \Delta W^b
+\sigma^\gamma{}_a(t_k,X_{t_k}) \partial_\gamma \mu^\alpha(t_k,W_{t_k}) \Delta t
\end{equation}
with
\begin{multline*}
c^\alpha_{(k)ab} =
\frac{1}{2} \big[ \sigma^\gamma{}_b(t_k,X_{t_k}) e^a{}_\beta(t_k,X_{t_k}) \partial_\gamma C^{\alpha\beta}(t_k,X_{t_k})
\\
+ \sigma^\beta{}_a(t_k,X_{t_k}) e^b{}_\gamma(t_k,X_{t_k}) \partial_\beta C^{\alpha\gamma}(t_k,X_{t_k})
\\
- e^a{}_\beta(t_k,X_{t_k}) e^b{}_\gamma(t_k,X_{t_k}) C^{\alpha\delta}(t_k,X_{t_k}) \partial_\delta C^{\beta\gamma}(t_k,X_{t_k}) \big] \rlap{\ .}
\end{multline*}
Equations \eqref{eq:hatMuCr} read
\begin{eqnarray*}
\hat \mu_k^\alpha(\Delta t, \Delta W) &=& f^\alpha_{10} + \frac{1}{2} (f^\alpha_{02})_{bb} + (f^\alpha_{11})_a \Delta W^a
- (g_{11})_a e^\alpha{}_a(\Delta t, \Delta W) \Delta t
\\
\hat C_k^{\alpha\beta}(\Delta t, \Delta W) &=& e^\alpha_{(k)a}(\Delta t, \Delta W) e^\beta_{(k)a}(\Delta t, \Delta W)
\\
\hat r_k(\Delta t, \Delta W) &=& g_{10} + (g_{11})_a \Delta W^a -\frac{1}{2} (g_{11})_a (g_{11})_a \Delta t^2 \rlap{\ .}
\end{eqnarray*}
Using the definitions of coefficients or the fact these quantities should coincide with $\mu^\alpha$, $C^{\alpha\beta}$ and $r$ up to $O(\Delta t)$ terms we can rewrite this as
\begin{eqnarray*}
\hat \mu_k^\alpha(\Delta t, \Delta W) &=& \mu^\alpha(t_k,X_{t_k})
+ \sigma^\beta{}_a(t_k,X_{t_k})
\\ && \qquad\qquad
\big[ \partial_\beta \mu^\alpha(t_k,X_{t_k}) \Delta W^a
- \partial_\beta r(t_k,X_{t_k}) e^\alpha_{(k)a}(\Delta t, \Delta W) \Delta t \big]
\\
\hat C_k^{\alpha\beta}(\Delta t, \Delta W) &=& e^\alpha_{(k)a}(\Delta t, \Delta W) e^\beta_{(k)a}(\Delta t, \Delta W)
\\
\hat r_k(\Delta t, \Delta W) &=& r(t_k,X_{t_k}) + \sigma^\alpha{}_a(t_k,X_{t_k}) \partial_\alpha r(t_k,X_{t_k}) \Delta W^a
\\ &&\qquad\qquad\qquad
-\frac{1}{2} C^{\alpha\beta}(t_k,X_{t_k}) \big( \partial_\alpha r(t_k,X_{t_k}) \big)\big(\partial_\beta r(t_k,X_{t_k})\big) \Delta t^2 \rlap{\ .}
\end{eqnarray*}
Using these expressions for $k-1$ we compute
\begin{eqnarray*}
\Delta \mu_k^\alpha &=& \mu^\alpha(t_k,X_{t_k}) - \hat \mu_{k-1}^\alpha(\Delta t_{k-1}, \Delta W_{k-1})
\\
\Delta C_k^{\alpha\beta} &=& C^{\alpha\beta}(t_k,X_{t_k}) - \hat C_{k-1}^{\alpha\beta}(\Delta t_{k-1}, \Delta W_{k-1})
\\
\Delta r_k &=& r(t_k,X_{t_k}) - \hat r_{k-1}(\Delta t_{k-1}, \Delta W_{k-1})
\end{eqnarray*}
and we get
$$
1 + \frac{\Delta \mathcal{H}_k}{\lambda} = 1 - \frac{\Delta r_k}{\lambda} + \frac{\Delta \mu^\alpha_k}{\lambda} \partial_{X^\alpha_{t_k}} + \frac{1}{2} \frac{\Delta C^{\alpha\beta}_k}{\lambda} \partial_{X^\alpha_{t_k}}\partial_{X^\beta_{t_k}}
$$
Except on the last time $t_p$, $\Delta \mathcal{H}_k$ acts on
\begin{equation}
\label{eq:intUcorrectiveMulti}
u_k(t_k,X_{t_k}) = \int \mathrm{d} X_{t_{k+1}} \widehat{U}_{t_k,t_{k+1}}(X_{t_k},X_{t_{k+1}}) \left(1+\frac{\Delta \mathcal{H}_{k+1}}{\lambda}\right) u_{t_{k+1}}(X_{t_{k+1}}) \rlap{\ .}
\end{equation}
More generally we will consider the action on this expression of a second order differential operator
$$
\mathcal{A}_k = A_k + A_k^\alpha \partial_{X^\alpha_{t_k}} + \frac{1}{2} A_k^{\alpha\beta} \partial_{X^\alpha_{t_k}}\partial_{X^\beta_{t_k}} \rlap{\ .}
$$
As in section \ref{sec:1dintermediate},e consider the change of variable between $X_t$ and $W_t$ defined by
$$
X^\alpha_t = X^\alpha_{t_k} + f_{(k)}^\alpha(t-t_k,W_t-W_{t_k}) \rlap{\ .}
$$
On derivatives it induces
\begin{eqnarray}
\label{eq:multiChangeDer}
\partial_a &=& e^{\alpha}_{(k)a} \partial_\alpha
\\
\nonumber
\partial_a \partial_b &=& e^{\alpha}_{(k)a} e^{\beta}_{(k)b} \partial_\alpha \partial_\beta + c^\alpha_{(k)ab} \partial_\alpha
\end{eqnarray}
and their inverse relations
\begin{eqnarray*}
\partial_\alpha &=& e^{a}_{(k)\alpha} \partial_a
\\
\partial_\alpha \partial_\beta &=& e^{a}_{(k)\alpha} e^{b}_{(k)\beta} \partial_a \partial_b - e^{a}_{(k)\alpha} e^{b}_{(k)\beta} e^{c}_{(k)\gamma} c^\gamma_{(k)ab} \partial_c \rlap{\ .}
\end{eqnarray*}
In the new variables, the differential operator $\mathcal{A}_k$ is
\begin{multline*}
\mathcal{A}_k = A_k + A_k^\alpha e^{a}_{(k)\alpha}(0,0) \partial_{W_{t_k}^a}
- \frac{1}{2} A_k^{\alpha\beta} e^{a}_{(k)\alpha}(0,0) e^{b}_{(k)\beta}(0,0) e^{c}_{(k)\gamma}(0,0) c^\gamma_{(k)ab} \partial_{W^c_{t_k}}
\\
+ \frac{1}{2} A_k^{\alpha\beta} e^{a}_{(k)\alpha}(0,0) e^{b}_{(k)\beta}(0,0) \partial_{W^a_{t_k}}\partial_{W^b_{t_k}} \rlap{\ .}
\end{multline*}
This operator acts on expression \eqref{eq:intUcorrectiveMulti}. In the new variables, the evolution operator becomes
\begin{equation}
\label{eq:hatUmulti}
\widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) = e^{\textstyle -g_{(k)}(t_{k+1}-t_k,W_{t_{k+1}}-W_{t_k})} \varphi(t_{k+1}-t_k, W_{t_{k+1}}-W_{t_k})
\end{equation}
where $\varphi$ is the $d$-dimensional Gaussian kernel
$$
\varphi(\Delta t, \Delta W) = \frac{1}{(2\pi \Delta t)^{d/2}} e^{\textstyle - \frac{1}{2} \frac{\Delta W^a \Delta W^a}{\Delta t}} \rlap{\ .}
$$
When acting on the term in $\frac{\Delta \mathcal{H}_{k+1}}{\lambda}$, we differentiate $\widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}})$ with respect to $W_{t_k}$, which means multiplying by the weights $\widehat{\mathcal{W}}_{(k)a}$ and $\widehat{\mathcal{W}}_{(k)ab}$, according to their definition
\begin{eqnarray*}
\partial_{W_{t_k}^a} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) &=& \widehat{\mathcal{W}}_{(k)a}(\Delta t_k, \Delta W_k)
\widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}})
\\
\partial_{W_{t_k}^a}\partial_{W_{t_k}^b} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) &=& \widehat{\mathcal{W}}_{(k)ab}(\Delta t_k, \Delta W_k)
\widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}})\rlap{\ .}
\end{eqnarray*}
From expression \eqref{eq:hatUmulti} we have
\begin{eqnarray*}
\widehat{\mathcal{W}}_{(k)a(\Delta t, \Delta W)} &=& \frac{\Delta W^a}{\Delta t} + \partial_a g_{(k)}(\Delta t, \Delta W)
\\
&=& \frac{\Delta W^a}{\Delta t} + \sigma^\alpha{}_a(t_k,X_{t_k}) \partial_\alpha r(t_k,X_{t_k}) \Delta t
\end{eqnarray*}
and
\begin{eqnarray*}
\widehat{\mathcal{W}}_{(k)ab}(\Delta t, \Delta W) &=&
\left[ \frac{\Delta W^a}{\Delta t} + \partial_a g_{(k)}(\Delta t, \Delta W) \right]
\left[ \frac{\Delta W^b}{\Delta t} + \partial_b g_{(k)}(\Delta t, \Delta W) \right]
\\ && \qquad\qquad\qquad\qquad\qquad\qquad
- \frac{\delta_{ab}}{\Delta t} - \partial_a\partial_b g_{(k)}(\Delta t, \Delta W)
\\
&=&
\left[ \frac{\Delta W^a}{\Delta t} + \sigma^\alpha{}_a(t_k,X_{t_k}) \partial_\alpha r(t_k,X_{t_k}) \Delta t \right]
\\ && \qquad\qquad\quad
\left[ \frac{\Delta W^b}{\Delta t} + \sigma^\beta{}_b(t_k,X_{t_k}) \partial_\beta r(t_k,X_{t_k}) \Delta t \right]
- \frac{\delta_{ab}}{\Delta t} \rlap{\ .}
\end{eqnarray*}
For the term where $\widehat{U}{t_k,t{k+1}}$ directly acts on $u_{t_{k+1}}$, we use the fact that
$\widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}})$ given in equation \eqref{eq:hatUmulti} depends only on $W_{t_{k+1}} -W_{t_k} $ to transfer the derivatives from the first variable to the second one:
$$
\partial_{W_{t_k}^a} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) = -
\partial_{W_{t_{k+1}}^a} \! \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) \rlap{\ .}
$$
Then we integrate by part on $W_{t_{k+1}}$ to transfer the derivative on $u_{t_{k+1}}$ so that
\begin{multline*}
\int \mathrm{d} W_{t_{k+1}} \partial_{W_{t_k}^a} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) u_{t_{k+1}}(W_{t_{k+1}})
\\
= - \int \mathrm{d} W_{t_{k+1}} \Big( \partial_{W_{t_{k+1}}^a}\! \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) \Big) u_{t_{k+1}}(W_{t_{k+1}})
\\
= \int \mathrm{d} W_{t_{k+1}} \widehat{U}^{(W)}_{t_k,t_{k+1}}(W_{t_k},W_{t_{k+1}}) \partial_{W_{t_{k+1}}^a} \!u_{t_{k+1}}(W_{t_{k+1}})
\end{multline*}
and similarly for the second derivative. We then go back to the original variable $X$ at date $t_{k+1}$ using equations \eqref{eq:multiChangeDer}.
Assembling all terms, we finally get
$$
\mathcal{A}_k u_{t_k}(X_{t_k}) = \int \mathrm{d} X_{t_{k+1}} \widehat{U}_{t_k,t_{k+1}}(X_{t_k},X_{t_{k+1}}) \mathcal{A}_{k+1} u_{t_{k+1}}(X_{t_{k+1}})
$$
where we define
$$
\mathcal{A}_{k+1} = A_{k+1} + A_{k+1}^\alpha \partial_{X^\alpha_{t_{k+1}}} + \frac{1}{2} A_{k+1}^{\alpha\beta} \partial_{X^\alpha_{t_{k+1}}}\partial_{X^\beta_{t_{k+1}}}
$$
with
\begin{eqnarray*}
A_{k+1} &=& A_k - d_k(\Delta t_k,\Delta W_k) \frac{\Delta r_{k+1}}{\lambda}
\\
A_{k+1}^\alpha &=& [\delta^\alpha_\gamma+b_{(k)\gamma}^\alpha] A_k^\gamma
- e^b{}_{\beta}(t_k,X_{t_k}) e^c{}_{\gamma}(t_k,X_{t_k}) c_{(k)bc}^\delta b_{(k)\delta}^\alpha A_k^{\gamma\delta}
\\&& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad
+ d_k(\Delta t_k,\Delta W_k) \frac{\Delta \mu^{\alpha}_{k+1}}{\lambda}
\\
A_{k+1}^{\alpha\beta} &=& [\delta^\alpha_\gamma+b_{(k)\gamma}^\alpha][\delta^\beta_\delta+b_{(k)\delta}^\beta] A_{k+1}^{\gamma\delta}
+ d_k(\Delta t_k,\Delta W_k) \frac{\Delta C^{\alpha\beta}_{k+1}}{\lambda} \rlap{\ .}
\end{eqnarray*}
We define $b_{(k)\gamma}^\alpha$ by
$$
e^a_{(k)\gamma}(0,0) e^\alpha_{(k)a}(\Delta_k,\Delta W_k)
=
\delta^\alpha_\gamma + b_{(k)\gamma}^\alpha
$$
which in our case, using \eqref{eq:edtdw}
and
$$
e^a_{(k)\gamma}(0,0) = e^a{}_{\gamma}(t_k,X_{t_k}) =\big[\sigma(t_k,W_{t_k})^{-1}\big]^a{}_\gamma \rlap{\ ,}
$$
gives
$$
b_{(k)\gamma}^\alpha = e^a{}_{\gamma}(t_k,X_{t_k}) c^\alpha_{(k)ab} \Delta W_k^b + \partial_\gamma \mu(t_k,X_{t_k}) \Delta t_k \rlap{\ .}
$$
We also define the effect of operator $\mathcal{A}_p$ acting by weights multiplication as
\begin{multline*}
d_k(\Delta t,\Delta W) = A_k + A_k^\alpha e^a{}_\alpha(t_k,X_{t_k}) \widehat{\mathcal{W}}_{(k)a}(\Delta t, \Delta W)
\\
+\frac{1}{2} A_k^{\alpha\beta} e^a{}_{\alpha}(t_k,X_{t_k}) e^b{}_{\beta}(t_k,X_{t_k}) \qquad
\\
\Big[ \widehat{\mathcal{W}}_{(k)ab}(\Delta t, \Delta W)
- e^c{}_{\gamma}(t_k,X_{t_k}) c^\gamma_{(k)ab} \widehat{\mathcal{W}}_{(k)c}(\Delta t, \Delta W)
\Big] \ .
\end{multline*}
Finally, on the last date, we keep the variance finite by antithetic sampling as in section \ref{sec:payoff}.
We compute
\begin{eqnarray*}
X_T^{\alpha(+)} &=& X^\alpha_{t_p} + f^\alpha_{(p)}(\Delta t_p, \Delta W_p)
\\
X_T^{\alpha(0)} &=& \hat{\mathbb{E}}_{t_p}\big[ X^\alpha_T \big] = X^\alpha_{t_p} + \mu^\alpha(t_p,S_{t_p}) \Delta t_p
\\
X_T^{\alpha(-)} &=& X^\alpha_{t_p} + f^\alpha_{(p)}(\Delta t_p, -\Delta W_p)
\end{eqnarray*}
and the corresponding discount factors on the last time step
\begin{eqnarray*}
D_{p,T}^{(+)} &=& e^{\textstyle - g_{(p)}(\Delta t_p, \Delta W_p)}
\\
D_{p,T}^{(0)} &=& e^{\textstyle - g_{(p)}(\Delta t_p, 0)}
\\
D_{p,T}^{(-)} &=& e^{\textstyle - g_{(p)}(\Delta t_p, -\Delta W_p)} \rlap{\ .}
\end{eqnarray*}
Then we can get the contribution from the path to the Monte Carlo estimate as
\begin{multline*}
P_T = D_p \bigg( \frac{1}{2} d_p(\Delta t_p, \Delta W_p) D_{p,T}^{(+)} h\big(S_T^{(+)}\big)
+ \frac{1}{2} d_p(\Delta t_p, -\Delta W_p) D_{p,T}^{(-)} h\big(S_T^{(-)}\big)
\\
- d_p^{(0)} D_{p,T}^{(0)} h\big(S_T^{(0)}\big)
\bigg)
\end{multline*}
with
\begin{eqnarray*}
d_p^{(0)} &=& \frac{1}{2} \Big[ d_p(\Delta t_p, \Delta W_p) + d_p(\Delta t_p, -\Delta W_p) \Big]
- \widehat{\mathbb{E}}_{t_p}\Big[ d_p(\Delta t_p, \Delta W_p) \Big]
\\
&=& \frac{1}{2} A_p^{\alpha\beta} e^a{}_{\alpha}(t_k,X_{t_p}) e^b{}_{\beta}(t_p,X_{t_p}) \bigg( \frac{\Delta W_p^a \Delta W_p^b}{\Delta t_p^2} - \frac{\delta_{ab}}{\Delta t_p} \bigg) \rlap{\ .}
\end{eqnarray*}
By construction $d_p^{(0)}$ has a null expected value. Thus $P_T$ is the average of two antithetic contributions to the option price, minus a term of null expected value.
Finally, the average of $P_T$ over all paths gives the Monte Carlo estimate of $u_{t_0}(X_0)$. It converges to it when the number of paths goes to infinity without any bias, as explained in section \ref{sec:poisson}.
\subsection{Monte Carlo scheme summary}
In a Monte Carlo simulation, on each path we start by initializing operator $\mathcal{A}_0$ by the identity
$
\mathcal{A}_0 = 1
$
and the disccount factor $D_0$ by 1
$
D_0 = 1
$. We start from $X_{t_0} = X_0$.
We draw Poisson times $t_k$. When going from date $t_k$ to date $t_{k+1}$ on a Monte Carlo path, we do the following:
\begin{enumerate}
\item Compute all coefficients $f_{ij}^\alpha$ and $g_{ij}$ in order to get functions $f_{(k)}^\alpha$ and $g_{ij}$.
\item Draw $d$ independent Gaussian variables $\Delta W_k^a$ with variance $\Delta t_k$ and get the next state $X_{t_{k+1}}$ as
$
X^\alpha_{t_{k+1}} = X^\alpha_{t_k} + f^\alpha_{(k)}(\Delta t_k,\Delta W_k)
$.
\item Accumulate the discount factor as
$
D_{k+1} = D_k e^{\textstyle -g_{(k)}(\Delta t_k,\Delta W_k)}
$.
\item Compute $\mathcal{A}_{k+1}$ from $\mathcal{A}_k$ as explained above.
\end{enumerate}
When we reach the last time $t_p$ before maturity $T$, we perform steps 1 and 2 with $t_{p+1} = T$, then we compute the corrected, discounted payoff $P_T$ as explained above. The final value is obtained by averaging over all Monte Carlo paths.
\section{Final comments}
In this article, we introduce a Monte Carlo scheme which converges to the theoretical value without any bias, while keeping a finite variance. It applies to multidimensional diffusion processes and it can also handle stochastic interest rates. It allows to decrease the average number of time steps needed to reach a given precision, which can save a lot of computation time.
\subsection{Related work}
We leverage some interesting work presented in \cite{henry2015exact}. However our Monte Carlo scheme differs in several ways.
One of the main differences is that in their scheme, paths which take into account corrective terms do not take into account the basic payoff contribution $h(S_T)$.
In other words, their choice is equivalent to keeping the constant unit term in equation \eqref{eq:jump} only when there is no jump at time $t$. This occurs with probability $1-\lambda_t \delta t$ and this is compensated by a factor $\frac{1}{1-\lambda_t \delta t} \sim e^{\lambda \delta t}$. Equation \eqref{eq:jump} is thus replaced by
$$
(1 - \delta N_t)e^{\textstyle \lambda_t \delta t} + \delta N_t \frac{\Delta \mathcal{H}_t}{\lambda_t} \rlap{\ .}
$$
The probability of not having any Poisson jump over a maturity $T$ is $e^{-\lambda T}$. Thus only this proportion of the Monte Carlo paths contains the basic payoff contribution. This is compensated by a weight $e^{\lambda T}$ coming from the product of $e^{\lambda \delta t}$ on all infinitesimal times.
However this increases the total Monte Carlo noise, especially for large values of the Poisson intensity $\lambda$. In this case $e^{-\lambda T}$ is close to zero and very few paths, if any, include the payoff contribution $h(S_T)$. In addition all paths contributions include a factor $e^{\lambda T}$ which can become very large.
In our scheme, the payoff contribution is kept for all paths, whether they include corrective terms or not. In addition, there is no such factor $e^{\lambda T}$. This makes the scheme usable for any value of $\lambda$, even when it becomes large.
A second difference is that our simulation scheme can handle simultaneously non-zero drift and nonconstant volatilities, in any dimension.
We also show how to take into account stochastic interest rates.
\subsection{Possible improvements}
The Monte Carlo scheme presented here can be enhanced in several ways. In particular, one can make different choices for the precise form of functions $f^\alpha_{(k)}$ and $g_{(k)}$. Depending on the specific choice, this can allow to have a simulated path closer to the original process and thus the corrective terms would be smaller. A a simple case, one can factor in some time dependence in parameters.
In addition, the Poisson intensity $\lambda$ can depend on time $t$ and on the stochastic variables $X_t^\alpha$. One could increase it in the regions where the corrective terms are higher and decrease it when they are smaller.
\subsection*{Acknowledgment}
We thank Calypso Herrera, Martial Millet and Arnaud Rivoira for useful comments.
\newpage
\bibliographystyle{apalike}
|
\section*{Results}
\textbf{Decoherence-free subspace.}
Before describing the CQ qubit in detail, we first
recall a DFS for spins.
Three spins can encode a DFS that protects against arbitrary uniform magnetic field fluctuations\cite{DiVincenzo2000,DeFilippo2000,Yang2001}, $\delta\bf B$.
The DFS consists of two states with the same values of the total spin along the quantization axis, $S_z=\sum_i s_{zi}$, and of the total spin $S^2=S_x^2+S_y^2+S_z^2$.
The DFS has two important properties: first, the difference in the energies of the two qubit states is independent of magnetic field, and second, changing a spin-independent Hamiltonian causes the system to evolve only between the qubit states; non-qubit (``leakage") states cannot be accessed because they are not coupled to the qubit states by the Hamiltonian.
Here we construct a similar arrangement for charge states that protects against uniform electric field variations.
All linear superpositions of the logical states must have the same total charge
and also the same
center of mass, so that the contribution to the energy from a uniform field,
$\sum_ie\, {\bf E}\cdot{\bf r}_i$, is the same for all qubit states.
In addition, it is important that the system Hamiltonian does not couple the qubit states to the other states in the full Hilbert space.
These
conditions are satisfied
if the Hamiltonian conserves charge and has an appropriate symmetry:
the qubit logical states should have the same total charge and be eigenstates of a symmetry operator with the same eigenvalue.
An appropriate candidate geometry is a central dot that is symmetrically coupled to a set of outer dots having the same center of mass as the center dot, even under permutation.
Analogous to the situation for a spin DFS\cite{DeFilippo2000,Lidar2003}, the symmetry constraints cannot be satisfied in a two-dimensional (double-dot) code space;
the smallest device that can support a charge DFS is a triple dot.
\textbf{Charge quadrupole qubit.}
Here we consider a linear triple dot geometry, where the
symmetry operation is the permutation operator between the outer two dots, $p_{1,3}$, which is equivalent to reflection about the center.
It is convenient to adopt the basis states $\{\ket{C}=\ket{010}, \ket{E}=(\ket{100}+\ket{001})/\sqrt{2}\}$, where $C$ and $E$ refer to ``center" and ``even."
The resulting $p_{1,3}$ eigenvalue is $+1$, corresponding to even symmetry.
The third, orthogonal state, $\ket{L}=(\ket{100}-\ket{001})/\sqrt{2}$, has eigenvalue $-1$, corresponding to odd symmetry, and generates a dipole that couples to charge fluctuations when superposed with $\ket{E}$.
When the symmetry constraint is satisfied, the even and odd manifolds decouple.
\begin{figure}[t]
\includegraphics[width=3.in]{Fig1}
\caption{
\textbf{\red{Basis states and gate operations of the charge quadrupole qubit.}}
\textbf{(a)} A triple-dot charge qubit can be described in terms of a position basis, corresponding to the single-electron occupations of dots 1, 2, or 3.
For a symmetrized triple dot, it is preferable to adopt, instead, an even-odd basis $\{\ket{C},\ket{E},\ket{L}\}$, referring to center, even, and leakage states, respectively.
Here, $\ket{C}$ and $\ket{E}$ have even symmetry, $\ket{L}$ has odd symmetry, and the half-filled circles represent average occupations of 1/2.
\textbf{(b)} The \red{charge quadrupole} eigenstates, $\ket{\tilde 0}$, $\ket{\tilde 1}$, and $\ket{\tilde L}$, obtained by solving \red{the system Hamiltonian,} equation~(\ref{eq:HCQ}), as a function of quadrupolar detuning, $\epsilon_\text{q}$.
Here, we set \red{the tunnel couplings,} $t_\text{A}=t_\text{B}\, (\equiv t/\sqrt{2})=2.5$~GHz and \red{the dipolar detuning,} $\epsilon_\text{d}=0$.
The insets depict the effect of $\epsilon_\text{q}$ on the triple-dot confinement potential.
\textbf{(c)} A cartoon depiction of \red{microwave (AC)} and \red{pulsed (DC)} gate sequences useful for qubit manipulation.
Initialization and readout are implemented in the far-detuned regime, $\epsilon_\text{q} \ll 0$, with $t\gtrsim 0$.
In the DC scheme, $\epsilon_\text{q}$ is suddenly pulsed to the double sweet spot, $\epsilon_\text{q}=0$.
Free evolution then yields an $X$ rotation in the logical basis $\{\ket{ C},\ket{E}\}$.
For a $Z$ rotation, $t$ is suddenly pulsed to 0, while $\epsilon_\text{q}$ is pulsed away from zero (either positive or negative).
In the AC scheme, $t\gtrsim0$ is held fixed, while an adiabatic ramp of $\epsilon_\text{q}$ to its sweet spot leaves the qubit in its logical ground state.
$X$ and $Y$ rotations in the rotating frame are implemented by applying resonant microwave bursts with appropriate phases to $\epsilon_\text{q}$, centered at the sweet spot.
Alternatively, microwaves may be applied to $t$, although we do not consider that possibility here.}
\label{fig:main}
\end{figure}
The logical charge states of a CQ qubit are protected from uniform electric field fluctuations because their charge distributions have the same center of mass
(in other words, no dipole moment).
It is interesting to note that several related systems also propose to use dipole-free geometries, including
the zero-detuning sweet spot of a conventional charge qubit\cite{Petersson2010,Shi2013,Kim2015}, which we analyze below, a three-island transmon qubit\cite{Gambetta2011}, and
quantum cellular automata\cite{Oi2005,Hentschel2007,Bayat2015}.
We now examine the CQ qubit in more detail.
We consider a triple dot with one electron, as illustrated in Fig.~\ref{fig:main}.
The full Hamiltonian in the position basis is given by
\begin{equation}
H_\text{CQ}\!=\!\begin{pmatrix}
U_1 & t_\text{A} & 0 \\
t_\text{A} & U_2 & t_\text{B} \\
0 & t_\text{B} & U_3
\end{pmatrix}\!=\!\begin{pmatrix}
\epsilon_\text{d} & t_\text{A} & 0 \\
t_\text{A} & \epsilon_\text{q} & t_\text{B} \\
0 & t_\text{B} & -\epsilon_\text{d}
\end{pmatrix}\! +\frac{U_1+U_3}{2} , \label{eq:HCQ}
\end{equation}
where $t_\text{A}$ and $t_\text{B}$ are the tunneling amplitudes between neighboring dots, and $U_1$, $U_2$, and $U_3$ are site potentials.
We have also defined the dipolar and quadrupolar detuning parameters, $\epsilon_\text{d}$
and $\epsilon_\text{q}$, as
\begin{equation}
\epsilon_\text{d} = (U_1-U_3)/2 \quad\text{and}\quad
\epsilon_\text{q}=U_2-(U_1+U_3)/2 . \label{eq:epsq}
\end{equation}
The eigenvalues of $H_\text{CQ}$ are plotted as a function of
$\epsilon_\text{q}$ in Fig.~\ref{fig:main}b, where the lowest and highest energy levels correspond to the logical eigenstates $\ket{\tilde 0}$ and $\ket{\tilde 1}$, respectively, and the middle level is the leakage state $\ket{\tilde L}$.
This ordering is an uncommon but benign feature of the CQ qubit, as shown below.
We note that, away from $\epsilon_\text{q} = 0$, the eigenstate $\ket{\tilde L}$ differs slightly from the basis state $\ket{L}$ due to mixing terms in equation~(\ref{eq:HCQ}).
Below, we show that under ideal conditions, the mixing terms are small, so that $\ket{\tilde L}\simeq\ket{L}$.
It is instructive to compare $H_\text{CQ}$ to a conventional, one-electron charge qubit formed in a double dot, which we refer to as a charge dipole (CD):
\begin{equation}
H_\text{CD}=\begin{pmatrix}
\epsilon_\text{d}/2 & t \\
t & -\epsilon_\text{d}/2 \end{pmatrix} . \label{eq:HCD}
\end{equation}
In this case, $\epsilon_\text{d}=U_1-U_2$ is the dipole detuning, and there is no quadrupole detuning.
In what follows, we express the detuning parameters in terms of their average ($\bar\epsilon$) and fluctuating ($\delta\epsilon$) components.
Uniform electric field fluctuations are then associated with $\delta\epsilon_\text{d}$, while fluctuations of the field gradient are associated with $\delta\epsilon_\text{q}$.
\textbf{Charge noise.}
It has been shown that phonon decoherence processes can be classified based on multipole moments\cite{Storcz2005}. Here, we consider charge noise decoherence processes for the leading order (dipole and quadrupole) moments in the noise, which by construction we will expect to behave very differently for CD and CQ qubits.
Fluctuations in $\delta\epsilon_\text{d}$ are dangerous for single-qubit operations since they cause fluctuations of the energy splitting between the qubit levels, $E_{01}$, resulting in phase fluctuations.
The success of the DFS depends on our ability to engineer a triple-dot in which the dephasing effects of $\delta\epsilon_\text{d}$ fluctuations are suppressed.
The next-leading source of fluctuations, $\delta\epsilon_\text{q}$, is much weaker, and we show in Supplementary Note 1 that
\begin{equation}
\delta\epsilon_\text{q}/\delta\epsilon_\text{d}\simeq d/R, \label{eq:small}
\end{equation}
where $d$ is the interdot spacing and $R$ is the characteristic distance between the qubit and the charge fluctuators that cause $\delta\epsilon_\text{d}$.
We also estimate that $d/R\simeq 0.1$ in recent
devices used for double dot qubit experiments\cite{Petersson2010,Shi2013}.
In Supplementary Note 2, we further show that $\delta \epsilon_\text{d}$ is related to the
more fundamental noise parameter $\delta\! E$ (the fluctuating electric field) as $\delta \epsilon_\text{d} \propto d\, \delta\! E$, so that $\delta \epsilon_\text{q} \propto d^2 \delta\! E$. Therefore, the effects of charge noise can be suppressed by making $d$ smaller through engineering, by reducing the lithographic feature size and the interdot spacing.
This is one of the key attractions of the quadrupole qubit: it provides a straightforward path for systematically improving the qubit fidelity, by reducing the device size.
The Hamiltonian $H_\text{CQ}$ has four independently tunable parameters.
We now determine the control settings consistent with DFS operation.
Our goal is to block diagonalize $H_\text{CQ}$ so that it decomposes into a two-dimensional (2D) logical subspace, and a 1D leakage space.
Any coupling to the leakage space would result in energy-level repulsions as a function of the tuning parameters.
We can therefore suppress such coupling by requiring that $\partial E_\text{L}/\partial\epsilon_\text{q}=\partial E_\text{L}/\partial\epsilon_\text{d}=0$, where $E_\text{L}$ is the leakage state energy, yielding
the desired tunings: $t_\text{A}=t_\text{B}$ and $\bar\epsilon_\text{d}=0$.
These are the same conditions obtained by requiring that $[H_\text{CQ},p_{1,3}]=0$, to obtain simultaneous eigenstates of $H_\text{CQ}$ and $p_{1,3}$.
The even-symmetry states $\ket{C}$ and $\ket{E}$ are good choices for basis states in the 2D manifold.
With the basis set $\{\ket{C},\ket{E},\ket{L}\}$, and the parameter tunings $t_\text{A}=t_\text{B}$ and $\bar\epsilon_\text{d}=0$, we find that $H_\text{CQ}$ block diagonalizes as desired.
(For notational convenience, we define $t/\sqrt{2}\equiv t_\text{A}=t_\text{B}$.)
In the logical subspace $\{\ket{C},\ket{E}\}$, the reduced Hamiltonian is then given by $H_\text{CQ,ideal}=(\bar\epsilon_\text{q}/2)(1+\sigma_z)+t\sigma_x$, where $\sigma_x$ and $\sigma_z$ are Pauli matrices.
\begin{figure}[t]
\includegraphics[width=3in]{Fig2}
\caption{
\textbf{\red{Preserving the symmetry of the quadrupole qubit with external couplings.}}
Schematic of the geometry of \textbf{(a)} a microwave stripline resonator coupled to \textbf{(b)} a quadrupole qubit.
The stripline geometry shown is similar to those suggested in refs~\onlinecite{Frey2012,Petersson2012,Stockklauser2015,Mi2017}.
Both accumulation-mode gates \red{that control the dot occupations} (with local potentials labeled $U_1$, $U_2$, and $U_3$) and depletion-mode gates \red{that control the tunnel couplings} (labeled $t_L$, $t_\text{A}$, $t_\text{B}$, and $t_R$) are included here (see refs~\onlinecite{Wu2014,Eng2015,Zajac2015,VeldhorstUnp} for a discussion).
The corresponding dots are labeled 1, 2, and 3.
The coupling occurs through the middle gate 2, which is connected to the resonator.
The qubit can be coupled, similarly, to other qubits or charge sensors.}
\label{fig:Coupling}
\end{figure}
We now compare the effects of fluctuations on the energy levels of CD and CQ qubits.
(In the following sections, we explore the effect of fluctuations on gate operations.)
The energy splitting of CD qubits is obtained from equation~(\ref{eq:HCD}) as $E_{01,\text{CD}}=\sqrt{\epsilon_\text{d}^2+4t^2}$.
A fluctuation expansion in powers of $\delta \epsilon_\text{d}$ yields
\begin{multline}
E_{01,\text{CD}}
=\sqrt{\bar\epsilon_\text{d}^2+ 4t^2}+
\left[ \frac{\bar\epsilon_\text{d}}{(\bar\epsilon_\text{d}^2+4t^2)^{1/2}}\right] \delta\epsilon_\text{d}
\\ +\left[ \frac{2t^2}{(\bar\epsilon_\text{d}^2+4t^2)^{3/2}}\right] \delta\epsilon_\text{d}^2
+O[\delta\epsilon_\text{d}^3] . \label{eq:E01CD}
\end{multline}
The first term in equation~(\ref{eq:E01CD}) indicates that $\bar\epsilon_\text{d}$ and $t$ are the main control parameters.
The second term indicates that the qubit is only protected from fluctuations of $O[\delta\epsilon_\text{d}]$ at the sweet spot, $\bar\epsilon_\text{d}=0$.
In contrast, the CQ qubit has two detuning parameters.
In this case, we fix $\bar\epsilon_\text{d}=0$ and calculate the energy splitting $E_{01,\text{CQ}}$ by
writing $\epsilon_\text{d}\rightarrow \delta\epsilon_\text{d}$ and $\epsilon_\text{q}\rightarrow\bar\epsilon_\text{q}+\delta\epsilon_\text{q}$.
Expanding in $\delta \epsilon_\text{d}$ and $\delta \epsilon_\text{q}$
yields
\begin{multline}
E_{01,\text{CQ}}
=\sqrt{\bar\epsilon_\text{q}^2+ 4t^2}+
\left[ \frac{\bar\epsilon_\text{q}}{(\bar\epsilon_\text{q}^2+4t^2)^{1/2}}\right] \delta\epsilon_\text{q}
\\ +\left[ \frac{\bar\epsilon_\text{q}^2+2t^2}{t^2(\bar\epsilon_\text{q}^2+4t^2)^{1/2}}\right] \delta\epsilon_\text{d}^2
+O[\delta\epsilon_\text{q}^2,\delta\epsilon_\text{d}^3],
\label{eq:E01CQ}
\end{multline}
where we note that $\delta \epsilon_\text{q}^2 \ll \delta \epsilon_\text{d}^2$.
By construction, $E_{01,\text{CQ}}$ has no terms of $O[\delta\epsilon_\text{d}]$ when $\bar\epsilon_\text{d}=0$.
Moreover, we see that fluctuations of $O[\delta\epsilon_\text{q}]$ vanish when $\bar\epsilon_\text{q}=0$.
Hence, $\bar\epsilon_\text{d}=\bar\epsilon_\text{q}=0$ represents a double sweet spot.
Since $\delta \epsilon_\text{q} \ll \delta \epsilon_\text{d}$, dephasing is minimized when we set $\bar\epsilon_\text{d}=0$ and adopt $\bar\epsilon_\text{q}$ and $t$ as the control parameters for CQ gate operations.
Although $\delta\epsilon_\text{d}$ does not appear at linear order in $E_{01,\text{CQ}}$, its main effect is to cause leakage for CQ qubits, rather than dephasing --- a point that we return to below.
For both CD and CQ qubits, we note that increasing the tunnel coupling $t$ also suppresses the energy fluctuations, particularly near the sweet spot; this is consistent with recent results in a resonantly gated three-electron exchange-only qubit\cite{Medford2013,Taylor13}.
\textbf{Pulsed gates for the CQ qubit.}
Here we investigate pulsed (DC) gates, assuming that $\bar\epsilon_\text{q}$ and $t$ can be
independently tuned and set to zero.
We perform
rotations of angle $\alpha$ around the $\hat{x}$ axis of the Bloch sphere ($X_\alpha$) by
setting $\bar{\epsilon}_q =0$ and $t = t_x>0$.
Rotations of angle $\beta$ around the $\hat{z}$ axis ($Z_\beta$) are achieved by setting $\epsilon_\text{q} = \epsilon_z\neq 0$ and $t =0$.
Readout is performed by measuring the charge occupation of the center dot.
In fact, all external couplings to initialization and readout circuits or to other qubits
should be made through the center dot, to preserve the symmetries of the qubit,
as illustrated in Fig.~\ref{fig:Coupling}.
We compare gate operations in CQ qubits to those of CD qubits via simulations that include quasistatic charge noise.
Results for the infidelities of simple (``bare") $X_\pi$ rotations for CQ and CD qubits are presented in Fig.~\ref{fig:fidelity} (blue and orange curves, respectively), as functions of the standard deviation of charge noise fluctuations.
(See Methods for details.)
Both curves follow the same scaling behavior, which can be explained as follows.
The effect of quasistatic noise, $\delta\epsilon_\text{d}$, may be regarded as a gating error that induces underrotations, overrotations, or rotations about a misoriented axis.
For CD qubits, such errors occur within the logical Hilbert space, while for CQ qubits, the misrotation occurs primarily to the leakage state.
Defining the state fidelity as $F_\text{s} = |\braket{\psi_\text{actual}}{\psi_\text{ideal}}|^2$, where $\ket{\psi_\text{actual}}$ is the actual final state and $\ket{\psi_\text{ideal}}$ is the target state, the noise drives $F_\text{s}<1$.
For either qubit, the resulting infidelity of noisy rotations scales as $(1-F_\text{s}) \propto \delta\epsilon_\text{d}^2$.
For example, the probability of a CQ qubit being projected onto its leakage
state is $|\braket{\tilde{L}}{C}|^2 =\delta\epsilon_\text{d}^2/(\delta\epsilon_\text{d}^2+t^2)\sim \delta\epsilon_\text{d}^2/t^2$.
\textbf{Composite pulse sequences.}
Fortunately, time evolution remains largely coherent throughout a
gate operation, so that special pulse sequences can be used to undo the leakage and suppress the errors\cite{Wang2012,Kestner2013,GhoshUnp}.
In ref.~\onlinecite{GhoshUnp}, three-pulse sequences were constructed for the CQ qubit, following the same control constraints indicated in Fig.~\ref{fig:main}c, which are experimentally motivated: the control parameters $\epsilon_\text{q}$ and $t$ can be pulsed independently, but not simultaneously, between zero and a finite value.
There it was shown that special values of the control parameters can be used to cancel out the leading order effect of $\delta\epsilon_\text{d}$ noise, yielding a universal set of low-leakage, single-qubit gate operations.
In Fig.~\ref{fig:fidelity}, we compare one such composite sequence for the CQ qubit, $\tilde X_\uppi \equiv Z_{2\uppi}X_{3\uppi}Z_{-2\uppi}$, to bare, single-step sequences for $X_\uppi$ rotations in both CD and CQ qubits.
The results show that significant benefits can be achieved with composite sequences: for charge noise levels consistent with recent experiments\cite{Kim2014,Mi2017}, fidelity improvements are in the range of 10-1000.
\begin{figure}[t]
\centering
\includegraphics[width=2.in]{Fig3}
\caption{
\textbf{\red{Simulated process fidelities for charge qubit gate operations.}}
Simulations of bare $X_\uppi$ rotations and composite pulse sequences, $\tilde X_\uppi$, for \red{charge dipole (CD)} and \red{charge quadrupole (CQ)} qubits are performed in the presence of \red{dipolar ($\delta\epsilon_\text{d}$)}, and \red{quadrupolar ($\delta\epsilon_\text{q}=\delta\epsilon_\text{d}/40$) detuning fluctuations}.
Plots show the infidelity ($=1-$fidelity) as a function of the standard deviation $\sigma_\epsilon$ of $\delta\epsilon_\text{d}$.
(The simulations, charge noise averages, and process fidelity calculations are described in Methods.)
Here, the blue and orange curves correspond to bare, single-pulse $X_\uppi$ rotations of CQ and CD qubits, respectively.
The infidelity follows the same scaling in both cases, even though it arises from different mechanisms: pure dephasing for the CD qubit vs.\ leakage for the CQ qubit.
The gray curve corresponds to a composite, three-pulse sequence, \red{$\tilde X_\uppi \equiv Z_{2\uppi}X_{3\uppi}Z_{-2\uppi}$}, which removes the leading order $\delta \epsilon_\text{d}$ noise in the CQ qubit;
no comparable sequence exists for CD qubits.
The simple form of the CQ pulse sequence derives from the quadrupole geometry, which transfers some of the overhead for noise protection from the control pulse sequence to the qubit hardware.
All simulations assume the same tunnel couplings ($t$ for the CD qubit; $t_\text{A,B}$ for the CQ qubit) of \SI{10}{GHz}.
}
\label{fig:fidelity}
\end{figure}
While noise-cancelling pulse sequences have been proposed for quantum dot spin qubits\cite{Wang2012,Kestner2013}, they are significantly more complex than the three-pulse sequence used in Fig.~\ref{fig:fidelity}.
Those sequences are constructed by inserting identity operations into the pulse sequence and assuming a continuous range of rotation axes in some plane of the Bloch sphere.
The constraints assumed above, where $\epsilon_\text{q}$ and $t$ are not varied simultaneously, yield bare $X$ and $Z$ rotations, but no continuous range of rotation axes.
Under such conditions, no three-pulse sequence exists that can cancel out leading-order $\delta\epsilon_\text{d}$ noise in CD qubits.
By relaxing these constraints to allow simultaneous tuning of $\epsilon_\text{q}$ and $t$ -- a challenging but potentially achievable goal -- it becomes possible to construct a five-step sequence to cancel out the leading-order noise in CD qubits.
Thus, the three-step sequence described above for CQ qubits is truly ``minimal," in the sense that it has the same level of complexity as a conventional spin-echo sequence, which has been shown to be effective for preserving the coherence of a CD qubit\cite{Kim2015}.
\textbf{Microwave driven gates.}
While it is necessary to move away from the sweet spot
to perform certain microwave-driven (AC) gate operations, it is possible to \red{center} the AC
signal at the sweet spot, as indicated in Fig.~\ref{fig:main}b, which improves
the coherence of gate operations.
Below, we analyze the AC gate sequence shown in Fig.~\ref{fig:main}b.
Here, initialization and readout are performed in the far-detuned regime.
However, we now consider an adiabatic ramp to the sweet spot, so that the working point $\bar\epsilon_\text{q}=0$ defines the quantization axis $\hat{z}$ in the laboratory frame.
For AC gates, one typically moves to the frame rotating at the qubit frequency, where
$X$ and $Y$ rotations are obtained by driving the appropriate detuning parameter (dipolar for
a CD qubit, quadrupolar for a CQ qubit), with the appropriate phase, at the resonance frequency $\nu=E_{01}/h$.
In the rotating frame, the primary decoherence mechanism during $X$ or $Y$ rotations is longitudinal, with the corresponding decay time $T_{1\uprho}$\cite{Yan2013}.
In this case, the charge noise environment is nearly Markovian,
so that, on resonance, it is sufficient to use Bloch-Redfield theory, giving\cite{Jing2014}
\begin{multline}
1/T_{1\uprho} = 2 S_z(\epsilon_{\text{ac}}/\hbar) \\
+ S_x([\epsilon_{\text{ac}}+2t]/\hbar)+
S_x([\epsilon_{\text{ac}}-2t]/\hbar) , \label{eq:T1rho}
\end{multline}
where $\epsilon_\text{ac}$ is the amplitude of the resonant drive, and $S_z(\omega)$ and $S_x(\omega)$ are the longitudinal and transverse noise spectral densities in the lab frame, respectively.
These functions describe the noise in the detuning parameters used to drive the rotations ($\epsilon_\text{d}$ for CD qubits, or $\epsilon_\text{q}$ for CQ qubits).
In the weak driving regime, $\epsilon_\text{ac}\ll 2t$, the term $2 S_z(\epsilon_{\text{ac}}/\hbar)$ would normally dominate equation~(\ref{eq:T1rho}) because $S_{x,z}(\omega)\propto 1/\omega$ for charge noise.
However, at the sweet spot, the $\epsilon$ noise for either type of qubit is orthogonal to the quantization axis, so that $S_z(\omega)=0$.
The other terms in equation~(\ref{eq:T1rho}) are relatively small, since their arguments are large.
We can compare $T_{1\uprho}$ for CD and CQ qubits by assuming that the noise terms, $\delta\epsilon_\text{d}$ and $\delta\epsilon_\text{q}$, both arise from the same charge fluctuators.
In this case, the ratio of their amplitudes, $\delta\epsilon_\text{q}/\delta\epsilon_\text{d}$, is independent of the frequency and
the decoherence rates for resonant $X$ and $Y$ rotations in a CQ qubit are suppressed by this same ratio, as compared to a CD qubit.
After applying simple pulse sequences to suppress the leakage, CQ qubits are therefore protected from the dominant ($O[\delta\epsilon_\text{d}]$) noise source for all rotation axes, for both pulsed and resonant gates, while CD qubits require a more complex correction scheme.
\textbf{Spin quadrupole qubits.}
Up to this point, we have focused on charge qubits.
However, quadrupolar geometries can also be used to protect logical spin qubits from dipolar detuning fluctuations.
For example, the standard two-electron singlet-triplet ($S$-$T$) qubit formed in a double quantum dot\cite{Levy2002,Petta2005} is not protected from dipolar detuning fluctuations during implementation of an exchange gate.
But a singlet-triplet qubit formed in a triple dot could be protected by tuning the device, symmetrically, to one of the charging transitions, $(1,0,1)$-$(1/2,1,1/2)$ or $(0,2,0)$-$(1/2,1,1/2)$.
Here, the delocalized states with half-filled superpositions are analogous to those shown in Fig.~\ref{fig:main}.
The magnitudes of the local Overhauser fields on dots 1 and 3 should be equalized for $S$-$T$ qubits, to enforce the symmetry requirements and suppress leakage out of the logical subspace.
We note that a different type of symmetric sweet spot was recently employed for a singlet-triplet qubit in a double-dot geometry\cite{Reed2016,Martins2016}.
In those experiments, the resonant pulse was applied to the tunnel coupling, as suggested in ref.~\onlinecite{Koh2013}, while the detuning parameter was set to a sweet spot.
Three-electron logical spins, such as the quantum dot hybrid\cite{Shi2012,Kim2014,Koh2012,Kim2015b,Ferraro2014,Mehl2015} or exchange-only\cite{Medford2013a,Medford2013,DiVincenzo2000,Taylor13,Eng2015} qubits, can also be implemented using a quadrupolar triple dot.
In this case, we must work at one of the charging transitions $(1,1,1)$-$(3/2,0,3/2)$, $(1,1,1)$-$(1/2,2,1/2)$, or $(0,3,0)$-$(1/2,2,1/2)$.
When the qubit basis involves singlet- and triplet-like spin states\cite{Shi2012,Koh2012}, localized in dots 1 or 3 [e.g., at the $(1,1,1)$-$(3/2,0,3/2)$ transition], the $S$-$T$ splittings in those dots should be equalized.
We note that measuring exchange-only qubits, or performing capacitive two-qubit gate operations, requires accessing the charge sector of those devices.
The conventional charging transition used for this purpose is $(1,1,1)$-$(2,0,1)$\cite{Medford2013a}, which is not protected from $\delta\epsilon_\text{d}$ fluctuations.
A symmetric quadrupolar geometry could therefore benefit such operations.
\textbf{External couplings and two qubit gates.}
Two main types of couplings have been proposed for two-qubit gates in quantum dot qubits: classical electrostatic (capacitive) interactions\cite{Taylor2005} or quantum exchange interactions\cite{Loss1998}.
We only consider capacitive couplings here, since exchange couplings require the dots to be in very close proximity.
Capacitive couplings mediated by qubit proximity or floating top gates\cite{Trifunovic2012} are convenient for quadrupole qubits, provided that the device symmetries are preserved during gate operations.
This suggests that the coupling should occur through the gate above the middle dot.
Readout and charge-to-photon interconversions should also be performed in the same way.
Capacitive two-qubit interactions, which yield an effective coupling of form $J\sigma_{z1}\sigma_{z2}$ in the basis $\{\ket{C},\ket{E}\}_1\otimes\{\ket{C},\ket{E}\}_2$, have previously been demonstrated in CD qubits\cite{Shinkai2009} and in logical spin qubits\cite{Nichol2017}. Here, $J(\epsilon_\text{q1},\epsilon_\text{q2})$ represents the capacitive dipole-dipole coupling, and the indices 1 and 2 refer to the interacting qubits.
One advantage of this coupling is that no new leakage states are incurred, beyond the single-qubit states $\ket{L}_1$ and $\ket{L}_2$, in contrast with two-qubit gates in some other DFS\cite{DiVincenzo2000}.
We now describe a simple protocol for nonadiabatic, pulsed two-qubit gate operations for CQ qubits, based on schemes developed for Cooper-pair boxes\cite{Yamamoto2003}, which are superconducting versions of the CD qubit.
\red{(AC gating schemes based on state-dependent resonant frequencies are also candidates for two-qubit operations\cite{Chen2014}, although we do not consider them here.)}
Our DC scheme can be viewed as a shift of the degeneracy point $\epsilon_\text{q2}=0$ of qubit~2, depending on the state of qubit~1.
The qubits are first prepared in their ground states in the far-detuned regime, yielding $\ket{\tilde 0 \tilde 0}$.
Qubit~2 is then pulsed to its degeneracy point, where free evolution yields an $X_\pi$ rotation to state $\ket{\tilde 0\tilde 1}$.
On the other hand, if an $X_\pi$ rotation is first applied to qubit~1, so the system is in state $\ket{\tilde 1\tilde 0}$, there will be an effective shift in $\epsilon_\text{q2}$ due to the interaction term.
Now when $\epsilon_\text{q2}$ is pulsed, it does not reach its degeneracy point.
In this case, no $X_\pi$ is implemented on qubit~2, and the system remains in state $\ket{\tilde 1\tilde 0}$.
The net result is a controlled (C)-NOT gate.
We note that since the qubits spend most of their time away from sweet spots in this protocol, the special noise protection afforded by CQ qubits should significantly improve their coherence.
Other types of external couplings are also possible.
For example, a microwave stripline resonator could potentially enable two-qubit couplings, readout, and charge-to-photon conversions by techniques described in refs~\onlinecite{Frey2012,Petersson2012,Stockklauser2015,Mi2017}, when coupled to a CQ qubit, as illustrated in Fig.~\ref{fig:Coupling}.
Qubit-resonator coupling strengths in the range $g=5$-50~MHz have been reported for cavity quantum electrodynamic (cQED) systems employing CD qubits\cite{Frey2012,Petersson2012,Stockklauser2015,Mi2017}.
Strong coupling has been achieved in such devices\cite{Mi2017}, but it is challenging\cite{Wallraff2013}, due to short CD coherence times of order 1~ns.
Achieving strong coupling requires that both $g/\Gamma_\text{q} \gg 1$ and $g/\Gamma_\text{s} \gg 1$, where $\Gamma_\text{q}\sim 1/T_{1\uprho}$ is the main decoherence rate for the qubit, and $\Gamma_\text{s}$ is the decoherence rate for the superconducting stripline.
We expect that $g$ for CQ qubits should be similar to CD qubits, while $\Gamma_\text{q}$ should be reduced by a factor of $\sim$10, so $g/\Gamma_\text{q} $ should increase by a factor of $\sim$10, which would mitigate the difficulties in achieving strong coupling.
It should also be possible to couple microwave striplines to quadrupolar spin qubits, using spin-to-charge conversion\cite{Childress2004}.
\section*{Discussion}
In conclusion, we have shown that charge qubit dephasing can be suppressed by employing a quadrupolar geometry, because the
quadrupolar detuning fluctuations are much weaker than dipolar fluctuations.
On the other hand, the quadrupolar detuning parameter $\epsilon_\text{q}$ is readily controlled by applying voltages to the top gates, and we expect gate times for CQ qubits to be just as fast as CD qubits.
Since dephasing is suppressed for CQ qubits while gate times are unchanged, we expect noise suppression techniques to be more effective for CQ qubits than CD qubits.
We have confirmed this by simulating minimal composite pulse sequences, designed to cancel out the effects of leakage.
This is a promising result for charge qubits because the fidelities of pulsed\cite{Petersson2010} and microwave\cite{Kim2015} gating schemes are
not currently high enough to enable error correction during gate operations.
We have also shown that the coherence properties of CQ qubits improve as the devices shrink, and we expect future generations of small CQ qubits to achieve very high gate fidelities.
We have further shown that logical spin qubits in quantum dots should benefit from a quadrupolar geometry.
We expect a prominent application for quadrupolar qubits to be cQED, where improvements in coherence properties could enhance strong coupling.
\section*{Methods}
\textbf{\red{Simulations and fidelity calculations.}}
Gate operations on CQ and CD qubits were simulated using standard numerical techniques to solve $i\hbar\dot{\rho}=[H,\rho]$, where $\rho$ is the 2D (3D) density matrix for the CD (CQ) qubit, defined by Hamiltonian $H=H_\text{CD}$ ($H_\text{CQ}$).
For the simulations shown in Fig.~\ref{fig:fidelity}, $H$ implements either a simple $X_\uppi$ rotation, or a composite rotation $\tilde X_\uppi$, as defined in the main text and Supplementary Note 3.
Process tomography is performed using the Choi-Jamiolkowski representation\cite{Gilchrist2005} for process $\mathcal{E}(\rho)$, defined by the evolution of
$i\hbar\dot{\rho}_\mathcal{E}=[I\otimes H,\rho_\mathcal{E}]$, where $I$ is the identity matrix of an ancilla qubit with the same dimensions as $H$.
Here, the initial Jamiolkowski state is given by $\rho_\mathcal{E}(0)=\ket{\Phi}\bra{\Phi}$, where $\ket{\Phi}=(\ket{00}+\ket{11})/\sqrt{2}$ for the CD qubit, and $\ket{\Phi}=(\ket{CC}+\ket{EE})/\sqrt{2}$ for the CQ qubit.
Formed in this way, $\rho_\mathcal{E}$ is equivalent to the standard $\chi$ matrix\cite{Gilchrist2005}, and we compute $F=\text{Tr}[\chi_\text{ideal}\chi_\text{actual}]$, where $\chi_\text{ideal}$ is obtained by setting $\delta\epsilon_\text{d}=\delta\epsilon_\text{q}=0$.
\textbf{\red{Charge noise averages.}}
Averages over quasistatic charge noise were performed using a gaussian probability distribution $P$ sampled from 41 equally spaced points in the range $\delta\epsilon_\text{d}\in [-6\sigma_\epsilon,6\sigma_\epsilon]$, where
\begin{equation}
P(\delta\epsilon_\text{d})=\frac{1}{\sqrt{2\pi\sigma_\epsilon^2}} \text{exp}
\left(-\frac{\delta\epsilon_\text{d}^2}{2\sigma_\epsilon^2}\right) ,
\end{equation}
and $\sigma_\epsilon$ is the standard deviation of the distribution.
\section*{Acknowledgements}
We thank Robin Blume-Kohout, Adam Frees, and John Gamble for helpful conversations and information.
This work was supported by ARO under award no.\ W911NF-12-0607, and by NSF under award no.\ PHY-1104660.
The authors would also like to acknowledge support from the Vannevar Bush Faculty Fellowship program sponsored by the Basic Research Office of the Assistant Secretary of Defense for Research and Engineering and funded by the Office of Naval Research through grant no.\ N00014-15-1-0029.
\section*{Supplementary Information}
In these Supplementary Notes, we explore several issues related to the operation of a charge quadrupole qubit.
\vspace{0.5in}
\noindent\textbf{Supplementary Note 1: Estimated size of dipolar vs.\ quadrupolar detuning fluctuations.}
\vspace{0.25in}
The charge quadrupole (CQ) qubit is less susceptible to charge noise than a charge dipole (CD) qubit because
in solid state devices the dipolar component of the charge noise, $\delta\epsilon_\text{d}$, is typically
much larger than the quadrupolar component, $\delta\epsilon_\text{q}$.
Here, we estimate the relative strengths of these two components based on experimental measurements of charge noise in semiconducting qubit devices, assuming that
both types of electric field noise arise from the same remote charge fluctuators.
We begin by considering charge noise from remote charge traps in the semiconductor device\cite{Dial2013,Petersson2010,Buizert2008,Shi2013}.
As a simple model, we consider a charge trap with two possible states: occupied vs.\ empty.
Compared to a dipole fluctuator, in which the charge toggles between two configurations, the monopole fluctuator can be considered a worst-case scenario because the monopole potential decays as $1/R$ while the dipole potential decays as $1/R^2$, where $R$ is the dot-fluctuator separation.
Following ref.~\onlinecite{Gamble2012}, this monopole model can be used to estimate the characteristic separation $R$ between the fluctuator and the quantum dot, based on charge noise measurements in a double-dot charge qubit.
Experimental measurements of the dephasing of charge
qubits\cite{Dial2013,Petersson2010,Buizert2008,Shi2013} yield estimates for the standard deviation of the dipole detuning parameter, $\sigma_\epsilon$, which range between roughly
$3$ and \SI{8}{\micro eV} for double dots separated by \SI{200}{nm}, leading to estimates for the dot-fluctuator separation of $R\sim 1.1$-\SI{2.5}{\micro m}.
(Note that a significantly smaller $\sigma_\epsilon$ was recently reported in ref.~\onlinecite{Mi2017}, which would correspond to a much larger value of $R$.)
With this information, we can estimate the ratio $\delta\epsilon_\text{q}/\delta\epsilon_\text{d}$.
In a worst-case scenario, corresponding to the strongest quadrupolar fluctuations, the monopole fluctuator would be lined up along the same axis as the triple dot.
Adopting a point-charge approximation for the fluctuator potential, $V(r)=e^2/4\pi\varepsilon r$, where $e$ is the charge of the electron, $\varepsilon$ is the dieletric constant, and $r$ is distance from the point-charge, and assuming an interdot spacing $d\ll R$, equation~(2) of the main text yields
\begin{equation}
\frac{\delta\epsilon_\text{q}}{\delta\epsilon_\text{d}}\simeq \frac{d}{R} . \label{eq:Lscaling}
\end{equation}
Taking $d=$ \SI{200}{nm}, and $R\simeq 1$-\SI{3}{\mu m}, we estimate that $\delta\epsilon_\text{q}/\delta\epsilon_\text{d}\simeq 0.07$-0.2 for typical devices.
In other words, for current devices, the quadrupolar detuning fluctuations should be $\sim$10 times weaker than dipolar detuning fluctuations.
Moreover, new generations of quantum dots in heterostructures without modulation doping\cite{Wu2014,Veldhorst2014,BorselliUnp,Zajac2015} have the potential to achieve much smaller $d$,
which would further suppress $\delta\epsilon_\text{q}/\delta\epsilon_\text{d}$.
In summary, we have estimated the characteristic separation $R$ between a double dot and a charge fluctuator, based on measurements of charge noise in quantum dot devices.
Of course, fluctuators are randomly distributed in solid-state systems, and it is possible for a defect to be located much closer to the qubit than our estimate suggests.
Such noisy environments have a negative impact on both CD and CQ qubits.
Fortunately, the length scales $d$ and $R$ appear to be well separated, so that fluctuations in a given qubit are very likely to be dominated by dipolar detuning fluctuations rather than quadrupolar fluctuations.
In fact, the scaling expression in equation~(\ref{eq:Lscaling}) is one of the most appealing arguments for exploring CQ qubits, which couple primarily to gradient field fluctuations, because the dephasing effects of the quadrupolar fluctuations can always be suppressed by reducing the device size and shrinking the interdot distance.
Indeed, quantum devices with dot separations of $d\simeq$ \SI{50}{nm} have recently been reported\cite{VeldhorstUnp}, corresponding to a further reduction in $\delta\epsilon_\text{q}/\delta\epsilon_\text{d}$ by a factor of 4 compared to the estimate given above.
\vspace{0.5in}
\noindent\textbf{Supplementary Note 2: Quantum dot variability.}
\vspace{0.25in}
The combined requirements of $\bar\epsilon_\text{d}=0$ and $t_\text{A}=t_\text{B}\equiv t/\sqrt{2}$ indicate that the CQ dot geometry should be highly symmetric.
Other types of symmetric geometries have also been proposed for improving the operation of charge-based qubits in superconducting Cooper-pair boxes \cite{Zhou04,You05,Shaw07}, as well as an exchange-only logical spin qubit\cite{Medford2013,Taylor13}.
To achieve such symmetry in a triple-dot qubit, we must assume that $t_\text{A} $ and $t_\text{B}$ are independently tunable.
In the main text, we assume that uniform electric field fluctuations, $\delta E$, couple to $\epsilon_\text{d}$ but not to $\epsilon_\text{q}$.
However, this statement contains some hidden assumptions about the symmetries of a triple dot, which may not be valid when we account for dot variability.
Here, we show that if the triple-dot symmetry is imperfect, uniform field fluctuations could induce effective quadrupolar fluctuations $\delta\epsilon_\text{q}$ that potentially spoil the CQ noise protection, and we explain how to avoid this problem.
Quantum dots are confined in all three dimensions.
The vertical confinement is typically very strong, so we can
apply the usual subband approximation and treat the dot in two dimensions (2D)\cite{Ando1982}.
Let us begin with a 1D parabolic approximation for the lateral confinement potential in a single dot:
\begin{equation}
V_i(x)=\frac{m\omega_i^2}{2}(x-x_i)^2 + U_{0i} , \label{eq:Vi}
\end{equation}
where $i=1,2,3$ is the dot index, $m$ is the effective mass, $\hbar\omega_i$ is the splitting between the simple harmonic energy levels, $x_i$ is the center of the dot, and $U_{0i}$ is the local potential.
A more accurate description of $V_i(x)$ could include anharmonic terms, which would yield higher-order corrections to the results obtained here.
The parameters $\omega_i$, $x_i$, and $U_{0i}$ all depend on voltages applied to the top gates. We assume that the $U_{0i}$ terms are adjusted to satisfy the requirement that $\bar\epsilon_\text{d}=0$, and henceforth ignore them.
The dot positions $x_i$ can also be controlled electrostatically by tuning the gate voltages
near the dot.
The parameter $\omega_i$ is the most difficult to adjust after device fabrication, because it is mainly determined by the fixed top-gate geometry, or other fixed features in the electrostatic landscape.
Electrons in dots with different $\omega_i$ respond differently to $\delta E$, and can therefore potentially affect the symmetries of a CQ qubit.
However, we now show that dot-to-dot variations in $\omega_i$ do not couple to $\delta E$ fluctuations at linear order.
A uniform fluctuating field $\delta E$ introduces a term of form $-ex\delta E$ in the energy.
Adding this term to equation~(\ref{eq:Vi}) and rearranging yields
\begin{equation}
V_i(x)=\frac{m\omega_i^2}{2}(x-x'_i)^2 -ex_i\delta E-\frac{e^2\delta E^2}{2m\omega_i^2}, \label{eq:Vi2}
\end{equation}
where $x'_i=x_i+(e/m\omega_i^2)\delta E$ represents the shifted center of the dot.
Considering the first term on the right-hand side of equation~(\ref{eq:Vi2}), we note that the energy of a shifted harmonic oscillator does not depend on its position, $x'_i$.
Dot-to-dot variations in this term therefore do not depend on $\delta E$, and can be compensated by adjusting the potentials $U_{0i}$.
The leading order fluctuation term in equation~(\ref{eq:Vi2}) is therefore $-ex_i\delta E$, which does not depend on $\omega_i$.
The coupling between $\delta E$ and $\omega_i$ only arises at higher order, in the third term of equation~(\ref{eq:Vi2}).
The term $-ex_i\delta E$ in equation~(\ref{eq:Vi2}) can be viewed as a fluctuating site potential $\delta U_i$.
The CQ symmetric design strategy provides a mechanism for eliminating the leading order dipolar detuning fluctuations.
However, from the definition of the quadrupolar detuning in equation~(2) of the main text, we see that the quadrupolar detuning fluctuations are given by
\begin{equation}
\delta\epsilon_\text{q} = \delta U_2-\frac{\delta U_1+\delta U_3}{2}
=e\left(-x_2+\frac{x_1+x_3}{2}\right)\delta E .
\end{equation}
In other words, \emph{uniform} electric field fluctuations can also generate \emph{quadrupolar} detuning fluctuations in an asymmetric triple dot.
Fortunately, it is straightforward to suppress this effect by adjusting the dot separations to make them equal:
\begin{equation}
x_2-x_1=x_3-x_2=d_x .\label{eq:symx}
\end{equation}
Repeating this analysis for the dot confinement along the $y$ axis, we obtain the additional requirement that
\begin{equation}
y_2-y_1=y_3-y_2=d_y . \label{eq:symy}
\end{equation}
Hence, the three dots must be equally spaced along a line.
These new symmetry requirements are not oppressive, and can be achieved by simply including two top gates to fine-tune the $x$ and $y$ positions of one of the dots; such fine-tuning can even be accomplished via automated methods\cite{Kelly2014}.
Moreover, small errors in the dot position, $\delta x$, are tolerable since they only increase the detuning by a linear factor, $\delta\epsilon_\text{q}=(\delta x/d)\delta\epsilon_\text{d}$, where we have expressed the uniform field fluctuations in terms of the dipolar detuning parameter.
\vspace{0.5in}
\noindent\textbf{Supplementary Note 3: Details on the $\tilde X_\uppi$ pulse sequence.}
\vspace{0.25in}
We consider the specific pulse sequence $\tilde X_\uppi \equiv Z_{2\uppi}X_{3\uppi}Z_{-2\uppi}$.
A more general set of three-step sequences is discussed in ref.~\onlinecite{GhoshUnp}.
For the bare $Z_{2\uppi}$ gate, we choose $\epsilon_\text{q} = \epsilon_z>0$, with the corresponding gate time $\tau_z=h/\epsilon_z$.
For the $Z_{-2\uppi}$ gate, we replace $\epsilon_z\rightarrow (-\epsilon_z)$, but keep the same gate time.
For the $X_{3\uppi}$ gate, we set $t=t_x\equiv \epsilon_z/2\pi$, with gate time $\tau_x=3h/4t_x$.
|
\section{Introduction}
Most optical flow methods assume that there is only one imaging layer on the observed image with the brightness of scene objects, and use the brightness constancy constraint (BCC) to estimate the optical flow for scene objects. This single imaging layer assumption, however, can be often violated in real-world situations, especially in cases involving transparency or reflection. Transparencies and reflections are frequently met in imaging process, \eg, when one is looking at street scene from inside a car through a stained windscreen, or seeing through a thin layer of rain, looking into a window with semi-reflections on the window surface \etc. The BCC will generally not hold for the resultant double-layer images, even in ideal noise-free cases.
In all the above examples, the observed image $\m I$ can be modeled as a superposition of two constituting layers, denoted as $\m I =\m L_1\oplus\m L_2$, where $\oplus$ denotes some suitable layer combination operator. Without loss of generality, we call $\m L_1$ the background scene layer, which corresponds to the image of the desired scene that we intend to capture, and $\m L_2$ the foreground distracting layer, which corresponds to the semi-transparent media (\eg a glass window with dirt or reflections on it) or the semi-reflected image.
The main goal of this paper is to robustly estimate the optical flow field of the scene objects (\ie the background layer), which is of concern for vision systems. We consider two general cases: the foreground distracting layer is \emph{stationary}, or \emph{dynamically changing}.
Let $\m I$ and $\m I'$ be two time-consecutive frames of a scene containing the aforementioned two layers. In the presence of a dynamic foreground layer, there are two legitimate optical flow fields -- one for the foreground layer and one for the background layer. Denote the two flow fields generated by the movements of the two layers as $\m U$ and $\m V$ respectively. The relationships among the observed images, the image layers and the optical flow fields can be given as
\begin{center}
$\begin{CD}
\m I~=~\m L_1~~@.{\oplus} @.~~~\m L_2\\
~~~~~~@VV{\m U }V~~~ @. ~@VV{\m V }V \\
\m I'~=~\m L_1'~~@.{\oplus} @.~~~\m L_2'\\
\end{CD}$\\
\end{center}
When $\m V\equiv\m 0$ and $\m L_2\equiv\m L_2'$, \ie the foreground layer is static, our task is to estimate a single flow field $\m U$ for background layer, and also estimate the layers $\m L_1$, $\m L_1'$, $\m L_2$. Otherwise when a dynamic foreground layer exists, we will estimate two flow fields $\m U, \m V$ as well as the layers $\m L_1$, $\m L_1'$, $\m L_2$, $\m L_2'$. As we explicitly perform image layer separation (\ie estimating $\m L_1$, $\m L_1'$, $\m L_2$, $\m L_2'$), an appealing byproduct of our method is the restoration of the clear scene images.
For either of the two cases with a static or dynamic foreground layer, this is a highly ill-posed problem, especially considering optical flow estimation and image layer separation problems \emph{per se} are known to be ill-posed. From only two input images, our task is to recover one or two optical fields, as well as the two unknown layers.
Little work has been reported in the literature concerning this double-layer image optical flow estimation problem, with only a few exceptions in the early days of computer vision research, \eg \cite{shizawa1990simultaneous}\cite{shizawa1991unified}\cite{langley1992transparent}\cite{darrell1993nulling}. These works however often used over-simplified assumption and restrictive motion field models, such as assuming a constant flow field over time or over space (\eg. globally translating). Bergen \etal~\cite{bergen1992three} proposed a ``three-frame algorithm" to recover two constituting flow fields, assuming the flow field is constant over at least three frames.
In contrast, this paper removes these restrictive assumptions, and proposes a two-frame algorithm for robustly recovering the flow field(s). Our method works for generic motions, and is thus applicable to a much wider range of practical situations for robust optical flow estimation.
\subsection{Related Works}
This paper is concerned with optical flow estimation in double-layer images where both layers can possibly be moving. Despite that the phenomena of such multiple imaging layers and motions are frequently encountered in reality, few papers in the literature have been devoted to this topic. This is in a sharp contrast to the existence of vast amount of papers on the classic optical flow problems (an analysis of recent practices of optical flow can be found in \cite{Sun_review}).
One of the first work for multiple optical flow computation is possibly due to Shiwaza \etal\cite{shizawa1990simultaneous}\cite{shizawa1991unified}. By assuming the two underlying flow fields to be constant (\eg pure translating), they derived a generalized brightness constancy constraint for the multi-motion case. However, this constant motion assumption is restrictive, not applicable for general flow fields with complex motions. Nevertheless, their method, being one of the first, has inspired a number of variants and extensions \cite{pingault2002optical}\cite{auvray2009jointmotion}\cite{ramirez2006multi}\cite{toro2000multiple}. Some variants operate in the Fourier domain, \eg \cite{langley1992multiple}\cite{langley1992transparent}\cite{darrell1993nulling}.
The flow estimation problem for two-layer images in this paper should not be confused with those works concerning ``motion-layer segmentation", albeit the two do share some similarity and the boundary between them can sometimes be fuzzy. For example, Wang and Adelson~\cite{wang1994representing} proposed to segment the image layers based on a pre-computed optical flow field. Irani \etal \cite{irani1994computing} used temporal integration to track occluding or transparent moving objects with parametric motion. Black and others~\cite{black1996robust}\cite{ju1996skin}\cite{sun2010layered}\cite{wulff2014modeling} proposed a number of algorithms for multiple parametric motion estimation and segmentation.
Yang and Li~\cite{yang2015dense} fit a flow filed with piecewise parametric models.
Weiss~\cite{weiss1997smoothness} presented a nonparametric motion estimation and segmentation method to handle generic smooth motions, thus this method is more related to ours. However, the method of Weiss and most other aforementioned methods primarily focused on
image and motion segmentation,
while we decompose the whole image into two composite brightness layers, and compute one generic flow field on each layer.
The proposed method involves solving two tasks simultaneously: optical flow field estimation, and reflection/transparent layer separation. For the second task, many researches have been published previously.
For example,
Levin \etal~\cite{levin2002learning}\cite{levin2007user} proposed methods for separating an image into two transparent layers using local statistics priors of natural images.
Single image solutions are also investigated in \cite{li2014single} and \cite{yeung_CVPR08}.
To utilize multiple frames, layer separation methods have been proposed based on aligning the frames with one layer~\cite{Wexler_ECCV}\cite{Li_brown_ICCV13}\cite{guo2014robust} or multiple layers~\cite{szeliski2000layer}\cite{gai2012blind}.
Sarel and Irani~\cite{sarel2004separating} presented an information theory based approach for separating transparent layers by minimizing the correlation between the layers. Chen \etal~\cite{Chen_ICCV09} gave a gradient domain approach for moving layer separation which is also based on information theory. Schechner \etal~\cite{schechner2000separation} developed a method for layer separation using image focus as a cue.
By using independent component analysis, Farid and Adelson~\cite{farid1999separating} proposed a layer separation method which works on multiple observations under different mixing weights.
Techniques for image layer separation were also developed in the field of intrinsic image/video extraction~\cite{Tappen_intrinsic}\cite{Weiss_intrinsic}\cite{ye2014intrinsic}.
In the context of stereo matching with transparency, Szeliski and Golland~\cite{szeliski1998stereo} simultaneously recovered disparities, true colors, and opacity of visible surface elements. Tsin \etal~\cite{tsin2006stereo} estimated both depth and colors of the component layers.
Li \etal\cite{li2015simultaneous} proposed a simultaneous video defogging and stereo matching algorithm.
The recent work of Xue \etal~\cite{xue2015computational} has a very similar formulation compared to ours. However, the goal and motivation of obstruction-free photography from a video sequence in \cite{xue2015computational} are different from ours.
The underlying assumptions on the flow fields, the employed flow solvers and the initialization techniques are
dissimilar.
\section{Problem Setup}
\label{sec:problem}
For ease of presentation, in formulating the problem (Sec.~\ref{sec:problem} and Sec.~\ref{sec:priors}) and presenting the optimization (Sec.~\ref{sec:minimization}), we will focus on the dynamic foreground case (\ie double-layer flow estimation). The static foreground case (\ie single-layer flow estimation) is simpler and can be derived accordingly. Note that, the static foreground case, though relatively simpler, is also of interest and very challenging.
\subsection{Linear Additive Imaging Model}
In previous discussion, we simply used $\m I=\m L_1\oplus\m L_2$ to denote the layer superposition operation, but did not give its exact form.
To make the idea of this paper more concrete, we opt for the linear additive model $+$ as a concrete example for $\oplus$, \ie, $\m I=\m L_1 + \m L_2$.
The linear additive model itself, while simple, has been used successfully in the past in solving many vision problems involving transparency and reflection (\eg, in shadow removal \cite{yeung_CVPR08}, image matting \cite{szeliski1998stereo} and reflection separation \cite{li2014single}). Moreover, by applying logarithm operation, a multiplicative superposition model can also be converted to an additive one.
Taking two frames of observations, $\m I$ and $\m I'$, at two consecutive time steps $t$ and $t+1$, we have
\begin{eqnarray}
\m I(\m X) \!\!&=&\!\! \m L_1(\m X) +\m L_2 (\m X),\label{eq:linear_additive_model1}\\
\m I'(\m X) \!\!&=&\!\! \m L_1'(\m X) + \m L_2'(\m X), \label{eq:linear_additive_model2}
\end{eqnarray}
where $\m X$ is a matrix indexing all pixel coordinates.
\subsection{Double Layer Brightness Constancy}
In the presence of transparencies or reflections, it is important to note that the conventional BCC condition cannot be applied directly to the observed images. Below, we will derive a generalized BCC condition which is applicable to the double-layer case.
The basic assumption that we will base our method on is: any component layer of the observed image must satisfy the brightness constancy condition individually. This is a realistic and mild assumption which is applicable to a wide range of transparency and reflection phenomena encountered in natural images. Cases that violate this basic assumption are deemed beyond the scope of this current paper.
Suppose, during two small time steps, layer $\m L_1$ changed to $\m L_1'$ according to a motion field of $\m U$, and layer $\m L_2$ changed to $\m L_2'$ according to a different motion field $\m V$. Based on the assumption that the brightness of the objects in each individual layer is constant, we have
\begin{eqnarray}
\m L_1(\m X) \!\!&=&\!\! \m L_1'(\m X + \m U), \label{eq:double-layer-BCC-1}\\
\m L_2(\m X) \!\!&=&\!\! \m L_2'(\m X + \m V). \label{eq:double-layer-BCC-2}\end{eqnarray}
Together with the imaging model in \eqref{eq:linear_additive_model1} and \eqref{eq:linear_additive_model2}, we call the above constraints the \emph{generalized double-layer BCC condition} for an input double-layer image pair $(\m I, \m I')$.
\subsection{The Double Layer Optical Flow Problem}
Given the above linear additive imaging model as well as the generalized BCC conditions, we aim to recover both $\m L_1$, $\m L_1'$, $\m L_2$, $\m L_2'$ and $\m U$, $\m V$.
To make this severely ill-posed problem trackable, we adopt the energy minimization framework, and base it on the generalized BCC conditions as well as priors for optical flows and image layers. The energy function reads as
\begin{equation}
E = E_{B}+\lambda_L E_{L} +\lambda_F E_{F},\label{eq:energy_function}
\end{equation}
where $E_B$ corresponds to the double-layer BCC condition, $E_L$ and $E_F$ are the regularization terms (or prior terms) for the latent image layers, and the unknown optical flow fields, respectively. The $\lambda$s are trade-off parameters.
In energy \eqref{eq:energy_function}, we use $E_B=E_B(\m L_1,\m L_1',\m L_2,\m L_2',\m U,\m V)$ to represent the BCC condition in the following way\footnote{For brevity, hereafter we use a short-hand notation for functions defined on all pixel coordinates $\m X$: a function $f(\m X)$ should be understood as $\sum_{\m x \in \m X}f(\m x)$, unless otherwise specified.}:
\begin{align}
E_B=&\| \m L_1(\m X)-\m L_1'(\m X +\m U) \|+ \|\m L_2(\m X)- \m L_2'(\m X +\m V) \|.
\end{align}
We use $\|\cdot\|$ to denote the $\ell_1$-norm in this paper unless otherwise specified. We choose to use $\ell_1$-norm as the cost function mainly for its robustness~\cite{brox2004high,zach2007duality} and its convenience in optimization. The two regularization terms $E_L$ and $E_F$ will be detailed in the following section.
\section{Regularization}
\label{sec:priors}
Using prior information as regularization is a common practice for solving ill-posed problems. In this paper, the task is to separate the input frames into latent layers, and to recover the associated flow fields.
Priors are generally task-dependent. By enforcing different priors to latent layers and to optical flow fields, the algorithm can be adapted to solving different tasks. For example, if one knows the two latent layers are images of natural scenes, then the layers can be assumed to have sparse gradients (\ie, satisfying the well-known natural image priors). Moreover, for general optical flow fields, one can assume they are piecewise constant or piecewise smooth.
\subsection{Natural Image Prior: Sparse Gradient}
The research in natural image statistics shows that images of typical real-world scenes obey sparse spatial gradient distributions~\cite{Tappen_intrinsic,levin2007user}. The distribution of a natural image $\m L$ can often be modeled as a generalized Laplace distribution (\emph{a.k.a.}, generalized Gaussian distribution), \ie,
\begin{equation}
P(\m L)= \prod_{\m x\in\m X}\exp(-|\partial_x\m L(\m x)|^p-|\partial_y\m L(\m x)|^p),\label{eq:sparsegrad}
\end{equation}where the power $p$ is a parameter usually within $[0.0,1.0]$. A convenient choice is $p=1$, with which the energy is reduced to the $\ell_1$-norm of image spatial gradients. For ease exposition, we will let $p=1$ in this paper, though bear in mind that using other values of $p$ is possible and may be advantageous in particular applications.
Taking the negative logarithm, the prior in (\ref{eq:sparsegrad}) can be represented in the energy minimization form, \ie
\begin{equation}
\|\nabla\m L(\m X)\|\rightarrow \min,
\end{equation}
where $\nabla=(\partial_x,\partial_y)^\top.$ Therefore, the latent layer regularization term $E_L=E_L(\m L_1,\m L_1',\m L_2,\m L_2')$ can be written as
\begin{align}
E_L\!=\!\|\nabla \m L_1(\m X)\|\!+\!\|\nabla\m L_1'(\m X)\|\!+\!\|\nabla\m L_2(\m X)\|\!+\!\|\nabla\m L_2'(\m X)\|.\label{eq:sprasegradient}
\end{align}
\subsection{Optical Flow Priors: Spatial Smoothness}
Early methods for solving multi-layer optical flow problem often made restrictive assumption about the unknown flow fields. For example, \cite{bergen1992three} proposed a three-frame algorithm for recovering two component motion fields by assuming that the motion fields are constant over time, and \cite {shizawa1990simultaneous} was built upon a local constant motion assumption to derive its basic equation. In this paper, these restrictions are removed and the proposed method can handle more general and more complex motion fields.
We use a general assumption on flow field, namely, the optical flows are generally piecewise constant or piecewise smooth. To capture this prior, we adopt the total variation (TV) model~\cite{zach2007duality} or total generalized variation (TGV) model~\cite{bredies2010total}. Specifically, a flow field $\m U$ will be regularized by the following energy:
\begin{eqnarray}
\|\m U\|_{{\text{TGV}}^k} \rightarrow \min,
\end{eqnarray}
where $\|\m U\|_{{\text{TGV}}^k}\doteq{\textit{TGV}}^k(\m U_x)+{\textit{TGV}}^k(\m U_y)$, and \mbox{${\textit{TGV}}^k(~\!\!\cdot\!\!~)$} denotes the $k$-th order TGV measure for horizontal and vertical flow components ${\m U}_x$ and ${\m U}_y$.
In general, the $k$-th order TGV favors solutions that are piecewise composed of $(k\!-\!1)$-th order polynomials: with $k=1$, TGV$^1$ reduces to the TV model which favors piecewise constant fields; with $k=2$, TGV$^2$ favors piecewise affine fields. We will only consider TV and TGV$^2$ in this paper, and the resultant prior regularization term $E_F=E_F(\m U,\m V)$ for the flow fields can be written as
\begin{align}
E_F(\m U,\m V) = \|\m U\|_{{\text{TGV}}^k} +\|\m V\|_{{\text{TGV}}^k}.
\end{align}where $k=1$ (\ie TV) or $2$.
\section{Energy Minimization}
\label{sec:minimization}
\subsection{The Overall Objective Function}
By stacking all the constraints over both latent layers and flow fields, we reach an energy minimization problem as
\begin{align}
&\!\!\!\!\!\min E({\m L_1,\m L_1',\m L_2,\m L_2',\m U,\m V})= E_B+\lambda_L E_L +\lambda_F E_F & \nonumber \\
& ~~~~=\!\left(\| \m L_1(\m X)\!-\!\m L_1'(\m X\!+\!\m U) \|+ \|\m L_2(\m X)\!-\!\m L_2'(\m X\!+\!\m V) \|\right) & \nonumber \\
& ~~~~~~~~+ \lambda_L \left(\|\nabla\m L_1\|\!+\!\|\nabla\m L_1'\|\!+\!\|\nabla\m L_2\|\!+\!\|\nabla\m L_2'\|\right) & \nonumber\\
& ~~~~~~~~+ \lambda_F \left(\|\m U\|_{{\text{TGV}}^k} +\|\m V\|_{{\text{TGV}}^k}\right), & \label{eq:objfunction}\\
&\!\!\!\!\text{subject to} \nonumber & \\
& ~~~~~~~~~~~~~~~~~~\m I = \m L_1 +\m L_2 ,~~\m I' = \m L_1' + \m L_2',& \label{eq:ADD} \\%& ~~~~~~~~~~~~~~~~~~~~~~~~~\m I' = \m L_1' + \m L_2',& \label{eq:ADD-2} \\
& ~~~~~~~\m 0\le\m L_2\le\min(\m I, c),~~\m 0\le\m L_2'\le \min(\m I', c). &\label{eq:bound}
\end{align}where the $\m X$'s in the gradient terms of \eqref{eq:sprasegradient} are omitted for brevity.
Note that, to distinguish background and foreground layers, we introduce in \eqref{eq:bound} the element-wise bound constraints on the layers. We assume the foreground layer containing transparency or reflection has weaker signal, and use a small constant scalar $c$ (\eg $c=0.25$ for brightness values in the range of [0,1]) as its brightness upper bound. This can be understood as an additional bound prior for layer separation. Also note that, putting aside \eqref{eq:bound}, there is a global shift ambiguity for the layer values: adding an arbitrary scalar $s\in\mathbb{R}$ to $\m L_1,\m L_1'$ then $-s$ to $\m L_2,\m L_2'$ dose not change the energy in (\ref{eq:objfunction}), nor dose it affect (\ref{eq:ADD}). This is because all the terms in (\ref{eq:objfunction}) depend on value difference rather than absolute value. Nevertheless, both the lower and upper bounds in \eqref{eq:bound} help constrain the absolute values.
\subsection{Alternated Minimization}
To solve the above energy minimization problem, we first substitute the additive model constraints in \eqref{eq:ADD} as hard constraints to eliminate $\m L_1$ and $\m L_1'$ in (\ref{eq:objfunction}). Consequently, the energy function is now defined only on latent layers $\m L_2,\m L_2'$ and optical flows ${\m U, \m V}$.
Then, examining the energy form in \eqref{eq:objfunction}, we notice that: \emph{i}) the prior terms for optical flow field, \ie $E_F$, is independent of the prior term for latent layers $E_L$; and \emph{ii}) the BCC energy term $E_B$ is the only term that links the flow estimation with latent layer separation. Based on these observations, we solve the minimization problem via block coordinate descent in an alternating fashion.
Specifically, starting from a proper initialization, our algorithm alternately solves the following two sub-problems:
\begin{itemize}
\item {\bf(Layer Separation):} Given current flow field estimates $\{\m U,\m V\}$, solve for image layers \{$\m L_2, \m L_2'$\} via the following minimization:
\vspace{-1pt}
\begin{eqnarray}
\min_{\m L_2,\m L_2'}\left( E_B(\m L_2, \m L_2' ) + \lambda_L E_L(\m L_2, \m L_2') \right).
\end{eqnarray}
\item {\bf(Flow Computation):} Given current image layers $\{\m L_2,\m L_2'\}$, estimate $\{\m U,\m V\}~$ by solving the following two-layer optical flow problem:
\vspace{-1pt}
\begin{eqnarray}
\min_{\m U, \m V} \left( E_{B}(\m U, \m V) +\lambda_F E_{F}(\m U, \m V) \right).
\end{eqnarray}
\end{itemize}
\vspace{-1pt}
More details are given below.
\vspace{-4pt}
\subsubsection{Update the image layers}
Given current optical flow estimates $\m U$ and $\m V$, the latent image layers $\m L_2, \m L_2^{'}$ can be updated by solving the following optimization problem:
\begin{align}
&\!\!\!\!\!\!\min_{\m L_2,\m L_2'}\!\! \|(\m I \!-\! \m L_2)(\m X)\!-\! (\m I' \!\!-\!\! \m L_2')(\m X \!\!+\!\!\m U) \|\!\!+\!\!\| \m L_2(\m X)\! -\!\m L_2'(\m X\!\!+\!\!\m V) \| & \nonumber \\
& ~~~~+\!\lambda_L\! \left( \|\nabla(\m I \!-\! \m L_2)\|\!+\!\|\nabla(\m I' \!-\! \m L_2')\|\!+\!\|\nabla\m L_2\| \!+\!\|\nabla\m L_2'\|\right) & \nonumber \\
&\!\!\!\!\mbox{subject to~~} \m 0\!\le\!\m L_2\!\le\!\min(\m I, c),~\m 0\!\le\!\m L_2'\!\le\!\min(\m I', c),&
\end{align}
This is a convex optimization problem defined on $\m L_2$ and $\m L_2'$, and the cost function can be arranged into
\begin{eqnarray}
&\displaystyle\min_{\mathbf{l}} \| \m A \cdot \mathbf{l}- \mathbf{b}\|,\nonumber\\
&\mbox{subject to~~} lb_i \leq l_i \leq ub_i, \forall i
\end{eqnarray}where $\m A$ and $\mathbf{b}$ encode all the $\ell_1$ constraints on latent layers, which are extremely sparse (only a few elements in each row are non-zero). $\mathbf{l}$ is a column vector containing elements in $\m L_2$ and $\m L_2'$. $lb_i$ and $ub_i$ are constant bounds from (\ref{eq:bound}). The constraints are linear function of the latent layers $\m L_2$ and $\m L_2'$, thus this problem can be solved as a linear programming using off-the-self solvers.
Nevertheless, to utilize the sparse structure in the problem and speed up the implementation, we solve the problem by using a tailored version of Iteratively Reweighted Least Squares (IRLS)~\cite{chartrand2008iteratively}. With IRLS, one can also adapt the formulation to different priors readily, \eg, replacing $\ell_1$-norm with $\ell_p$-norm ($0\!<\!p\!<\!1$). Details of our IRLS variant can be found in the \emph{Supplementary Material}.
\paragraph{Use of color images.} The above formulations can be easily extended to color RGB images. With color images, the double-layer BCC term $E_B$ and layer regularization term $E_L$ will be evaluated at R-G-B channels separately. The flow fields $\m U$ and $\m V$ are shared by all three channels.
\subsubsection{Update the flow fields $\m U$ and $\m V$}
Given current layer estimates $\m L_2, \m L_2'$, and $\m L_1 = \m I - \m L_2$, $\m L_1' = \m I'- \m L_2'$, the next step is to update the associated two flow fields $\m U$ and $\m V$. This is done by solving the following optimization problem:
\begin{align}
& \min_{\m U, \m V} \| \m L_1(\m X) -\m L_1'(\m X +\m U) \|+ \|\m L_2(\m X)-\m L_2'(\m X +\m V) \| & \nonumber \\ &~~~~+\lambda_F \left(\|\m U\|_{{\text{TGV}}^k} +\|\m V\|_{{\text{TGV}}^k}\right).&
\end{align}
The computations for these two flow fields are in fact separable. This can be easily seen from the above optimization, as the cost function can be expressed as the sum of two terms, each of which can be solved in isolation, \ie, given $\{\m L_1,\m L_1',\m L_2,\m L_2'\},$ solve
\begin{align}
& ~~~~~\min_{\m U} \| \m L_1(\m X) -\m L_1'(\m X +\m U) \| +\lambda_F \|\m U\|_{{\text{TGV}}^k}, & \\
& ~~~~~\min_{\m V} \|\m L_2(\m X)- \m L_2'(\m X +\m V) \| +\lambda_F \|\m V\|_{{\text{TGV}}^k}. &
\end{align}
To solve the above optical flow problems, we use quadratic relaxation and introduce an auxiliary flow field to decouple the BCC term and regularization term, similar to \cite{zach2007duality,steinbrucker2009large}. To solve the resulting TV-$L_2$ (\emph{a.k.a.} the ROF model) and TGV$^2$-$L_2$ problem, we apply the primal-dual method of \cite{chambolle2011first} which is GPU-friendly. Details of our algorithm and implementation can be found in the \emph{Supplementary Material}.
\section{Experimental Results}
\begin{figure}[!tp]
\begin{center}
\includegraphics[width=0.48\textwidth]{Book1_cvpr16_cr.pdf}
\end{center}
\vspace{-13pt}
\caption{Convergence of the proposed method. Top: optical flow estimation error (EPE) \emph{w.r.t.} iterations. Bottom: energy and layer estimation errors \emph{w.r.t.} iterations. The layer error is evaluated as $1-NCC(\text{GT}~\m L_2, \text{estimated}~\m L_2$).\label{fig:convergence}}
\vspace{2pt}
\end{figure}
In this section, we validate the proposed model and framework, and evaluate the performance of our method. We report the experimental results on both synthetic data and real images (\eg Middlebury~\cite{middlebury} and Sintel~\cite{butler2012naturalistic} flow datasets, and the reflection dataset in \cite{Li_brown_ICCV13}).
\paragraph{Initialization.} Being an alternated method, the proposed algorithm requires an initialization to start the alternation. One can start from either an initial optical flow estimation or from an initial layer separation. The latter one is used in our experiments, and the initialization details will be given later in the experiments.
\paragraph{Parameters.} In the following experiments, the weights of the priors, \ie $\lambda_L$, $\lambda_F$, are roughly tuned according to the results. Both TV and TGV$^2$ flow regularizers worked well, consistently improving the accuracy upon initialization. Due to space limitation, in the following we report the results using TV (\ie $k=1$). The results using TGV$^2$ (\ie $k=2$) can be found in the \emph{Supplementary Material}.
\subsection{Static Foreground Cases}
\begin{figure*}[!htp]
\begin{center}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_gray_cvpr16/O1_observed_two_layers_.jpg}
\vspace{-17pt}\caption{\footnotesize Input $\m I$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_gray_cvpr16/I1_final_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m L_1$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_gray_cvpr16/I2_final_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m L_1'$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_gray_cvpr16/J_final_forground_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m L_2$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_gray_cvpr16/flow_final_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m U$ (epe 0.29)}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_gray_cvpr16/flow_init_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize Naive $\m U$ (epe 1.01)}
\end{subfigure}\\
\vspace{2pt}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_gray_cvpr16/O2_observed_two_layers_.jpg}
\vspace{-17pt}\caption{\footnotesize Input $\m I'$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_gray_cvpr16/I1_gt_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize GT $\m L_1$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_gray_cvpr16/I2_gt_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize GT $\m L_1'$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_gray_cvpr16/J_gt_forground_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize GT $\m L_2$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_gray_cvpr16/flow_gt_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize GT $\m U$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_gray_cvpr16/flow_oracle_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize Oracle $\m U$ (epe 0.16)}
\end{subfigure}
\end{center}
\vspace{-15pt}
\caption{Performance evaluation of the proposed method on a single flow case, where a rain image is superimposed on the Dimetrodon image pair. The estimated flow (e) is significantly better than the initialization (f), a naive optical flow estimate without layer separation. The error evolution curve is shown in Fig.~\ref{fig:convergence}. Oracle flow (l) is computed with clean background images (\ie with ground-truth layer separations). (\emph{\textbf{Best viewed on screen}})}
\label{fig:single_gray_cvpr16}
\end{figure*}
\begin{figure*}[!htp]
\begin{center}
\begin{subfigure}{0.195\linewidth}
\includegraphics[width=1\textwidth]{sintel/alley_1_0001_O1.jpg}
\vspace{-17pt}\caption{\footnotesize Input $\m I$}
\end{subfigure}
\begin{subfigure}{0.195\linewidth}
\includegraphics[width=1\textwidth]{sintel/alley_1_0001_I1.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m L_1$}
\end{subfigure}
\begin{subfigure}{0.195\linewidth}
\includegraphics[width=1\textwidth]{sintel/alley_1_0001_J.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m L_2$}
\end{subfigure}
\begin{subfigure}{0.195\linewidth}
\includegraphics[width=1\textwidth]{sintel/alley_1_0001_flow.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m U$ (epe 0.30)}
\end{subfigure}
\begin{subfigure}{0.195\linewidth}
\includegraphics[width=1\textwidth]{sintel/alley_1_0001_flow_init.jpg}
\vspace{-17pt}\caption{\footnotesize Naive $\m U$ (epe 0.61)}
\end{subfigure}\\
\begin{subfigure}{0.195\linewidth}
\includegraphics[width=1\textwidth]{sintel/sleeping_2_0046_O1.jpg}
\vspace{-17pt}\caption{\footnotesize Input $\m I$}
\end{subfigure}
\begin{subfigure}{0.195\linewidth}
\includegraphics[width=1\textwidth]{sintel/sleeping_2_0046_I1.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m L_1$}
\end{subfigure}
\begin{subfigure}{0.195\linewidth}
\includegraphics[width=1\textwidth]{sintel/sleeping_2_0046_J.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m L_2$}
\end{subfigure}
\begin{subfigure}{0.195\linewidth}
\includegraphics[width=1\textwidth]{sintel/sleeping_2_0046_flow.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m U$ (epe 0.21)}
\end{subfigure}
\begin{subfigure}{0.195\linewidth}
\includegraphics[width=1\textwidth]{sintel/sleeping_2_0046_flow_init.jpg}
\vspace{-17pt}\caption{\footnotesize Naive $\m U$ (epe 0.33)}
\end{subfigure}
\end{center}
\vspace{-15pt}
\caption{Typical results of our method on single-flow cases, where the rain drop image is superimposed on images from the Sintel dataset. For clarity, we only show here the first frame $\m I$ and its layer separation result. (\emph{\textbf{Best viewed on screen}})}
\label{fig:sintel}
\end{figure*}
\begin{table}
\centering
\caption{Mean flow EPE for three Sintel image sequences superimposed with the static rain image. Oracle flows are computed with clean background images.}\label{tab:sintelepe}
\vspace{-5pt}
\begin{tabular}{|c|c|c|c|}
\hline
Sequence & Naive flow & Our flow & Oracle \\
\hline
``alley1" & 0.49 & \textbf{0.35} & 0.22 \\
\hline
``sleeping1" & 0.80 & \textbf{0.33} & 0.12 \\
\hline
``sleeping2" & 0.26 & \textbf{0.21} & 0.07 \\
\hline
\end{tabular}
\end{table}
We start from the simpler case where only background layer $\m L_1$ is dynamically changing by an unknown motion field $\m U$, while the foreground layer is static (\ie $\m L_2 \equiv \m L_2'$ and $\m V \equiv \m 0$).
The task is to estimate flow field $\m U$ and component layers $\m L_1,\m L_1',\m L_2$. Again, we would like to emphasize that, even though we call it the ``simpler case'', to jointly estimate an accurate flow field and recover latent layers remains a challenging task. To the best of our knowledge, there was no previous method that recovers both a complex dense flow field under transparency/reflection, and separate the two constituting layers.
In the following tests, a rather conservative strategy is used to initialize the proposed method: we initiate the static foreground image $\m L_2$ to be all zeros. Consequently, in the beginning of the optimization we compute an initial optical flow field naively based on the two input images.
Seeing through rain is a practical situation where measures should be taken to avoid the rain ruining vision systems. In the first test, we first synthesized a scene by superimposing a static rain image over the pair of Dimetrodon in the Middlebury dataset. Gray images were used. As illustrated in Fig.~\ref{fig:convergence}, within about 25 iterations, the optical flow estimation error has been decreased from about 1.0 pixels to about 0.3 pixels. This demonstrates the advantage of our formulation for robust optical flow estimation. The qualitative results are demonstrated in Fig.~\ref{fig:single_gray_cvpr16}.
Additionally, we overlay the rain image with three color image sequences from the Sintel dataset. We evenly sampled 10 images from the ``alley 1", ``sleeping 1" , and ``sleeping 2" sequences respectively, and Table~\ref{tab:sintelepe} shows that the proposed method has clearly reduced the mean EPE of initial flows. Two typical results are shown in Fig.~\ref{fig:sintel}.
\begin{figure*}[!htp]
\vspace{0pt}
\begin{center}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_color_5_/O1_observed_two_layers_.jpg}
\vspace{-17pt}\caption{\footnotesize Input $\m I$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_color_5_/I1_final_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m L_1$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_color_5_/I2_final_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m L_1'$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_color_5_/J_final_forground_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m L_2$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_color_5_/flow_final_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize Output $\m U$ (epe 0.88)}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_color_5_/flow_init_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize Naive $\m U$ (epe 1.45)}
\end{subfigure}\\
\vspace{2pt}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_color_5_/O2_observed_two_layers_.jpg}
\vspace{-17pt}\caption{\footnotesize Input $\m I'$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_color_5_/I1_gt_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize GT $\m L_1$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_color_5_/I2_gt_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize GT $\m L_1'$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_color_5_/J_gt_forground_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize GT $\m L_2$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_color_5_/flow_gt_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize GT $\m U$}
\end{subfigure}
\begin{subfigure}{0.162\linewidth}
\includegraphics[width=1\textwidth]{single_color_5_/flow_oracle_background_layer_.jpg}
\vspace{-17pt}\caption{\footnotesize Oracle $\m U$ (epe 0.67)}
\end{subfigure}
\end{center}
\vspace{-15pt}
\caption{Performance evaluation of the proposed method on a single flow case, where the Lena image is superimposed on the Grove image pair. The estimated flow (e) is significantly better than the initialization (f), a naive optical flow estimate without layer separation. Oracle flow (l) is computed with ground-truth $\m L_2$. (\emph{\textbf{Best viewed on screen}})}
\label{fig:single_color_5}
\vspace{0pt}
\end{figure*}
To further test the performance of our method, we synthesized another pair by superimposing the Lena image with the Grove image in the Middlebury dataset. The results are demonstrated in Fig.~\ref{fig:single_color_5}. Again, we obtained a much better optical flow compared to the initial naive optical flow estimate. As for the layer separation results, the portrait of Lena can be hardly seen in the restored grove images.
In Fig.~\ref{fig:statistics} we show the image gradient statistics of the three foreground images used in the above experiments. The experimental results have shown that the proposed method works well on these images with the sparse gradient prior. Whenever available, other strong statistical priors can be incorporated into the optimization framework to further improve the performance.
\begin{figure}[!tp]
\begin{center}
\begin{subfigure}{0.314\linewidth}
\includegraphics[width=1\textwidth]{distribution/hist4__.pdf}
\vspace{-17pt}\caption{\footnotesize Dimetrodon}
\end{subfigure}
\begin{subfigure}{0.32\linewidth}
\includegraphics[width=1\textwidth]{distribution/hist2__.pdf}
\vspace{-17pt}\caption{\footnotesize Rain drop}
\end{subfigure}\!
\begin{subfigure}{0.33\linewidth}
\includegraphics[width=1\textwidth]{distribution/hist3__.pdf}
\vspace{-17pt}\caption{\footnotesize Lena}
\end{subfigure}
\end{center}
\vspace{-13pt}
\caption{Gradient statistics of three used images.}
\label{fig:statistics}
\vspace{0pt}
\end{figure}
\subsection{Dynamic Foreground Cases}
In this section, we test the proposed method in the dynamic foreground cases, where the task is that given two frames of input images $\m I$ and $\m I'$, recover four component layers $\m L_1, \m L_1', \m L_2, \m L_2'$, and two dense motion fields $\m U$, $\m V$. In the problem of reflection removal, both the background scene and the reflection can be dynamic, which can give rise to such a situation.
We use two pairs of dynamic reflection scenes from \cite{Li_brown_ICCV13} to test the proposed method on the double-layer optical flow problem. In previous single-flow experiments, we initialize the method with foreground layers being all zero. However, this simple strategy did not work for the double-flow case. No reasonably good flow field could be obtained with this strategy for the background or reflection layer, especially for the reflection layer as its signal is weak. Indeed, the fact that the background layer is much more prominent has been took advantage of by some layer separation methods~\cite{Li_brown_ICCV13}\cite{guo2014robust} which align the input images with respect to the background layer.
To obtain proper initialization, we first ran method of \cite{Li_brown_ICCV13} for initial layer separations\footnote{Method of \cite{Li_brown_ICCV13} takes multiple images as input, with one of them being the reference on which the reflection is to be removed. We apply this method on two images, and run it twice with each image as reference.}, then computed initial optical flows on them.
The initial and final results are presented in Fig.~\ref{fig:double_color}. Visually inspected, the final optical flow fields are smoother and more consistent (see e.g. the
results on the back wall in the first example, and results on the floor in the second example). As no ground truth optical flow is available, we use image warping error to quantitatively evaluate the estimated flows. The warping error for a pixel $\mathbf{x}$ in $\m L_1$ or $\m L_2$ is $\|\m L_1(\mathbf{x}\!+\!\m U(\mathbf{x}))\!-\!\m L_1'(\mathbf{x})\|_2$ or $\|\m L_2(\mathbf{x}\!+\!\m V(\mathbf{x}))\!-\!\m L_2'(\mathbf{x})\|_2$, respectively. We compute the mean warping errors for all pixels on $\m L_1$ and $\m L_2$. As shown in Table~\ref{tab:warperr}, our method has significantly reduced the warping error upon the initializations. Figure~\ref{fig:double_color} shows the improvements of the reflection removal results upon the initial estimates.
\begin{figure*}[!htp]
\vspace{-5pt}
\begin{center}
\captionsetup[subfigure]{labelformat=empty}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/42_.jpg}
\vspace{-18pt}\caption{\footnotesize Input $\m I$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/I1_i__.jpg}
\vspace{-18pt}\caption{\footnotesize Initial $\m L_1$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/I1_e__.jpg}
\vspace{-18pt}\caption{\footnotesize Final $\m L_1$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/J1_i.jpg}
\vspace{-18pt}\caption{\footnotesize Initial $\m L_2$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/J1_e.jpg}
\vspace{-18pt}\caption{\footnotesize Final $\m L_2$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/u_i.jpg}
\vspace{-18pt}\caption{\footnotesize Initial $\m U$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/u_e.jpg}
\vspace{-18pt}\caption{\footnotesize Final $\m U$}
\end{subfigure}\\
\vspace{2pt}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/43_.jpg}
\vspace{-18pt}\caption{\footnotesize Input $\m I'$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/I2_i__.jpg}
\vspace{-18pt}\caption{\footnotesize Initial $\m L_1'$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/I2_e__.jpg}
\vspace{-18pt}\caption{\footnotesize Final $\m L_1'$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/J2_i.jpg}
\vspace{-18pt}\caption{\footnotesize Initial $\m L_2'$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/J2_e.jpg}
\vspace{-18pt}\caption{\footnotesize Final $\m L_2'$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/v_i.jpg}
\vspace{-18pt}\caption{\footnotesize Initial $\m V$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/v_e.jpg}
\vspace{-18pt}\caption{\footnotesize Final $\m V$}
\end{subfigure}\\
\vspace{2pt}
\begin{subfigure}{0.243\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/1.jpg}
\vspace{-18pt}\caption{\footnotesize Close-up of initial $\m L_1$}
\end{subfigure}~
\begin{subfigure}{0.243\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/2.jpg}
\vspace{-18pt}\caption{\footnotesize Close-up of final $\m L_1$}
\end{subfigure}~~
\begin{subfigure}{0.243\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/3.jpg}
\vspace{-18pt}\caption{\footnotesize Close-up of initial $\m L_1'$}
\end{subfigure}~
\begin{subfigure}{0.243\linewidth}
\includegraphics[width=1\textwidth]{double_color_1/4.jpg}
\vspace{-18pt}\caption{\footnotesize Close-up of final $\m L_1'$}
\end{subfigure}
\vspace{15pt}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/199_.jpg}
\vspace{-18pt}\caption{\footnotesize Input $\m I$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/I1_i__.jpg}
\vspace{-18pt}\caption{\footnotesize Initial $\m L_1$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/I1_e__.jpg}
\vspace{-18pt}\caption{\footnotesize Final $\m L_1$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/J1_i.jpg}
\vspace{-18pt}\caption{\footnotesize Initial $\m L_2$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/J1_e.jpg}
\vspace{-18pt}\caption{\footnotesize Final $\m L_2$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/u_i.jpg}
\vspace{-18pt}\caption{\footnotesize Initial $\m U$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/u_e.jpg}
\vspace{-18pt}\caption{\footnotesize Final $\m U$}
\end{subfigure}\\
\vspace{2pt}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/200_.jpg}
\vspace{-18pt}\caption{\footnotesize Input $\m I'$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/I2_i__.jpg}
\vspace{-18pt}\caption{\footnotesize Initial $\m L_1'$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/I2_e__.jpg}
\vspace{-18pt}\caption{\footnotesize Final $\m L_1'$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/J2_i.jpg}
\vspace{-18pt}\caption{\footnotesize Initial $\m L_2'$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/J2_e.jpg}
\vspace{-18pt}\caption{\footnotesize Final $\m L_2'$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/v_i.jpg}
\vspace{-18pt}\caption{\footnotesize Initial $\m V$}
\end{subfigure}
\begin{subfigure}{0.138\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/v_e.jpg}
\vspace{-18pt}\caption{\footnotesize Final $\m V$}
\end{subfigure}\\
\vspace{2pt}
\begin{subfigure}{0.243\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/I1_i_.jpg}
\vspace{-18pt}\caption{\footnotesize Close-up of initial $\m L_1$}
\end{subfigure}~
\begin{subfigure}{0.243\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/I1_e_.jpg}
\vspace{-18pt}\caption{\footnotesize Close-up of final $\m L_1$}
\end{subfigure}~~
\begin{subfigure}{0.243\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/I2_i_.jpg}
\vspace{-18pt}\caption{\footnotesize Close-up of initial $\m L_1'$}
\end{subfigure}~
\begin{subfigure}{0.243\linewidth}
\includegraphics[width=1\textwidth]{double_color_2/I2_e_.jpg}
\vspace{-18pt}\caption{\footnotesize Close-up of final $\m L_1'$}
\end{subfigure}
\end{center}
\vspace{-13pt}
\caption{Experimental results on real reflection images. The initial layer separations are estimated by running method of \cite{Li_brown_ICCV13} on the two input images. Visually inspected, the final optical flow fields are smoother and more consistent (see \eg the results on the back wall in the first example, and results on the floor in the second example). The corresponding warping errors are presented in Table~\ref{tab:warperr}. The close-up images in the third rows show the improvements of the reflection removal results upon the initial estimates. (\emph{\textbf{Best viewed on screen}})
}
\label{fig:double_color}
\vspace{0pt}
\end{figure*}
\begin{table}
\centering
\caption{Mean image warping errors (in gray levels) from the double-flow estimation results.}\label{tab:warperr}
\vspace{-3pt}
\begin{tabular}{|c|c|c|c|}
\hline
Image pair & Initial results & Our final results \\
\hline
\#1 & 6.27& \textbf{2.55} \\
\hline
\#2 & 3.86 & \textbf{1.49} \\
\hline
\end{tabular}
\vspace{2pt}
\end{table}
\vspace{-5pt}
\paragraph{Discussion.} The dynamic foreground case with double-layer flow estimation is generally much harder than the single-flow case. This is not only because the former has more unknown variables to be solved for, but also due to the difficulties in obtaining a good initialization.
Nevertheless, our experiments show that the proposed method consistently improved the reasonable initializations given to it, for both the single-flow and double-flow cases.
\vspace{-5pt}
\paragraph{Limitation.} The proposed method is better suited for scenarios where the correlation between latent layers and their flow fields are relatively small. It will fail if both the two layers are textureless (as infinite numbers of possible motions exist satisfying the BCC constraints), or they undergo a same motion (thus the original BCC holds and only a single motion field can be extracted).
\section{Conclusions and Future Work}
This paper has defined the problem of robust optical flow estimation in the presence of possibly moving transparent or reflective layers. To our knowledge, the problem goes beyond the scope of conventional optical flow methods and was not properly investigated before.
We have presented a generalized double-layer brightness constancy condition as well as an optimization framework to solve this problem. The double-layer brightness constancy condition couples the flow fields and the brightness layers.
Encouraging experimental results of optical flow estimation and layer separation on challenging data have been obtained, even though we are using simple priors for them.
We hope that this paper can inspire future works to further address this challenging ill-posed problem.
Our current framework is based on a generative model,
which is applied uniformly to both the foreground and background layers.
In future, we plan to leverage discriminative models to exploit the differences between the two layers for better layer separation.
We also would like to explore some other optical flow priors. One possible strategy is to apply piecewise parametric motion model~\cite{ju1996skin,yang2015dense}, which provides stronger constraints than general
smoothness regularizers such as a TV, and is recently demonstrated to have advanced performances~\cite{yang2015dense}. Some other issues such as occlusion handling could also be considered.
\vspace{-5pt}
\paragraph{Acknowledgments}
{\small
H. Li's research is funded in part by Australian Research Council Grants of DP120103896, LP100100588, ARC Centre of Excellence on Robotic Vision (CE140100016) and NICTA (Data61). Y. Dai's contribution is funded in part by ARC Grants (DE140100180, LP100100588) and National Natural Science Foundation of China (61420106007).
}
{\small
\bibliographystyle{ieee}
|
\section{Introduction}
\label{sec:intro}
Ultrarelativistic heavy-ion collision
experiments at the Relativistic Heavy Ion Collider (RHIC)
and the Large Hadron Collider (LHC) study the quark-gluon plasma (QGP) using high energy nuclear collisions. The collective behavior seen in these experiments is quite successfully described by relativistic fluid dynamics. In early works, relativistic ideal hydrodynamics was applied assuming the QGP to behave like a perfect fluid \cite{Huovinen:2001cy,Hirano:2002ds,Kolb:2003dz}. Later on, to include the dissipative (viscous) effects, viscous hydrodynamics has been applied \cite{Muronga:2001zk,Muronga:2003ta,Muronga:2004sf,Heinz:2005bw,Baier:2006um,Romatschke:2007mq,Baier:2007ix,Dusling:2007gi,Luzum:2008cw,Song:2008hj,Heinz:2009xj,Bozek:2009ty,Bozek:2009dw,El:2009vj,PeraltaRamos:2009kg,PeraltaRamos:2010je,Denicol:2010tr,Denicol:2010xn,Schenke:2010rr,Schenke:2011tv,Bozek:2011wa,Bozek:2011ua,Niemi:2011ix,Niemi:2012ry,Bozek:2012qs,Denicol:2012cn,Denicol:2012es,PeraltaRamos:2012xk,Calzetta:2014hra,Denicol:2014vaa,Florkowski:2015lra,Ryu:2015vwa,Niemi:2015voa,Niemi:2015bpj}. Recently, due to the large momentum anisotropies generated during heavy-ion collisions, a new framework called anisotropic hydrodynamics has been developed \cite{Martinez:2010sc,Florkowski:2010cf,Ryblewski:2010bs,Martinez:2010sd,Ryblewski:2011aq,Florkowski:2011jg,Martinez:2012tu,Ryblewski:2012rr,Florkowski:2012as,Florkowski:2013uqa,Ryblewski:2013jsa,Bazow:2013ifa,Tinti:2013vba,Florkowski:2014bba,Florkowski:2014txa,Nopoush:2014pfa,Denicol:2014mca,Nopoush:2014pfa,Tinti:2015xra,Bazow:2015cha,Nopoush:2015yga,Bazow:2015zca,Alqahtani:2015qja,Florkowski:2015cba,Molnar:2016vvu} (for a recent review, see Ref.~\cite{Strickland:2014pga}). This new framework has been compared to traditional viscous hydrodynamics in many ways. For boost-invariant and transversely homogeneous systems, by comparing to exact solutions it has been shown that anisotropic hydrodynamics more accurately describes the dynamics in all cases considered \cite{Florkowski:2013lza,Florkowski:2013lya,Bazow:2013ifa,Florkowski:2014sfa,Denicol:2014mca,Tinti:2015xra}. In addition, it has been shown that anisotropic hydrodynamics best reproduces exact solutions of Boltzmann equation subject to 1+1d Gubser flow \cite{Nopoush:2014qba,Denicol:2014xca,Denicol:2014tha}. Finally, we also mention that it has been shown that anisotropic hydrodynamics shows better agreement with data from ultracold Fermi gases experiments than viscous hydrodynamics \cite{Bluhm:2015bzi,Bluhm:2015raa}.
The anisotropic hydrodynamics program is now focused on making phenomenological predictions for heavy-ion physics, including anisotropic freeze-out and a realistic lattice-based equation of state (EoS)~\cite{Nopoush:2015yga}. In a recent paper it was demonstrated how to impose a realistic EoS assuming approximate conformality of the QGP~\cite{Nopoush:2015yga}. In Ref.~\cite{Alqahtani:2015qja} a different method for imposing a realistic EoS was proposed in which the non-conformality of the QGP is taken into account by modeling the QGP as a gas of massive quasiparticles with temperature-dependent masses. This quasiparticle approach is motivated by perturbative results such as hard thermal loop (HTL) resummation, where the quarks and gluons can have temperature-dependent masses \cite{Weldon:1982aq,Weldon:1982bn,Braaten:1989mz,Blaizot:2001nr,Andersen:1999fw,Andersen:2003zk,Andersen:2004fp,Andersen:2011sf,Haque:2014rua}. In the quasiparticle anisotropic hydrodynamics framework one introduces a single-finite temperature mass which is fit to available lattice data for the QCD EoS. Once $m(T)$ is determined, the realistic EoS together with the non-equilibrium energy momentum tensor can be used to derive the dynamical equations for such a quasiparticle gas using Boltzmann equation~\cite{Alqahtani:2015qja}. In this work, we extend the previous 0+1d work of Ref.~\cite{Alqahtani:2015qja} to 1+1d and we use ``anisotropic Cooper-Frye freeze-out'' to compute the primordial particle spectra.
Here we compare results of quasiparticle anisotropic hydrodynamics to the standard anisotropic hydrodynamics \cite{Nopoush:2015yga} and second-order viscous hydrodynamics. We will refer to the three methods considered herein as ``aHydroQP'', ``aHydro'' and ``vHydro'', respectively. For our comparisons, we derive the dynamical equations necessary for all cases for a 1+1d system which is azimuthally-symmetric and boost invariant. We then specialize to a relaxation-time approximation for collisional kernel and solve the partial differential equations for each approach numerically. We then present comparisons of the total number of charged particles $N_{\rm chg}$, the average transverse momentum $\langle p_T \rangle$ for pions, kaons, and protons, and the differential pion, kaon, and proton spectra predicted by each approach. We find that the three methods agree well for small shear viscosity to entropy density ratio, $\eta/s$, but differ at large $\eta/s$. We find, in particular, that when using standard viscous hydrodynamics, the bulk-viscous correction can drive the primordial particle spectra negative at large $p_T$. Such a behavior is not seen in either anisotropic hydrodynamics approach, irrespective of the value of $\eta/s$.
The structure of the paper is as follows. In Sec.~\ref{sec:conventions}, we specify the notation and conventions used in the paper. In Sec.~\ref{sec:setup}, we review the necessary setup including basis vectors necessary in different cases, the anisotropic distribution function, and the lattice based equation of state. In Sec.~\ref{sec:boltzman}, the Boltzmann equation and its generalization to quasiparticles with temperature-dependent masses is discussed. In Sec.~\ref{sec:bulk-var}, we derive expressions for the particle four-current, energy density, and components of the pressure by taking different moments of distribution function. In Sec.~\ref{sec:dynamical-eqs}, the dynamical equations for massive anisotropic hydrodynamics are derived for azimuthally-symmetric boost-invariant systems. In Sec.~\ref{sec:freeze-out} we discuss anisotropic Cooper-Frye freeze-out in the context of leading-order anisotropic hydrodynamics. In Sec.~\ref{sec:results}, our numerical results obtained using the three methods for central Pb-Pb and p-Pb collisions at LHC energies are presented. Sec.~\ref{sec:conclusions} contains our conclusions and an outlook for the future. In App.~\ref{app:basis} the basis vectors used in the paper are presented. In App.~\ref{app:vhydro} we present details about second-order viscous hydrodynamics equations. Finally, all necessary identities and function definitions are collected in Apps.~\ref{app:identities} and \ref{app:h-functions}.
\section{Conventions and Notation}
\label{sec:conventions}
A parentheses in the indices indicates a symmetrized form, e.g. $A^{(\mu\nu)} \equiv (A^{\mu\nu} + A^{\nu\mu})/2$. The metric is taken to be in the ``east coast convention'' such that in Minkowski space with $x^\mu\equiv(t,x,y,z)$ the measure is $ds^2=g_{\mu\nu} dx^\mu dx^\nu=dt^2-dx^2-dy^2-dz^2$. We also use the standard transverse projector, $\Delta^{\mu\nu} \equiv g^{\mu\nu} - u^\mu u^\nu$. When studying relativistic heavy-ion collisions, it is convenient to transform to Milne coordinates defined by $\tau =\sqrt{t^2 - z^2}$, which is the longitudinal proper time, and $\varsigma ={\rm tanh}^{-1}(z/t)$, which is the longitudinal spacetime rapidity. For a system which is azimuthally symmetric with respect to the beam-line, it is convenient to transform to polar coordinates in the transverse plane with $r=\sqrt{x^{2}{+}y^{2}}$ and $\phi ={\rm tan}^{-1}(y/x)$. In this case, the new set of coordinates $x^\mu=(\tau,r,\phi,\varsigma)$ defines polar Milne coordinates. Finally, the invariant phase space integration measure is defined as
\begin{equation}
dP \equiv N_{\rm dof} \frac{d^3p}{(2\pi)^3} \frac{1}{E} = \tilde{N} \frac{d^3p}{E} \, ,
\end{equation}
where $\tilde{N} \equiv N_{\rm dof}/(2\pi)^3$ with $N_{\rm dof}$ being the number of degrees of freedom.
\section{Setup}
\label{sec:setup}
In this paper, we derive non-conformal anisotropic hydrodynamics equations for a system of quasiparticles with a temperature-dependent mass. To accomplish this goal, we take moments of Boltzmann equation, appropriate for the system of quasiparticles with thermal mass, to obtain the necessary anisotropic hydrodynamics equations. Using a general set of basis vectors and some simplifying assumptions relevant for 1+1d system, equations are expanded and simplified to the form appropriate for describing a boost-invariant and azimuthally-symmetric QGP. The obtained 1+1d equations are then solved numerically for our tests.
\subsection{Ellipsoidal form}
\label{subsec:distribution-func}
In non-conformal anisotropic hydrodynamics one introduces an anisotropy tensor of the form \cite{Nopoush:2014pfa,Martinez:2012tu}
\begin{equation}
\Xi^{\mu\nu} = u^\mu u^\nu + \xi^{\mu\nu} - \Delta^{\mu\nu} \Phi \, ,
\label{eq:aniso-tensor1}
\end{equation}
where $u^\mu$ is four-velocity, $\xi^{\mu\nu}$ is a symmetric and traceless tensor, and $\Phi$ is associated with the bulk degree of freedom. The quantities $u^\mu$, $\xi^{\mu\nu}$, and $\Phi$ are understood to be functions of spacetime and obey $u^\mu u_\mu = 1$, ${\xi^{\mu}}_\mu = 0$, ${\Delta^\mu}_\mu = 3$, and $u_\mu \xi^{\mu\nu} = 0$; therefore, one has ${\Xi^\mu}_\mu = 1 - 3 \Phi$.
At leading order in the anisotropic hydrodynamics expansion one assumes that the one-particle distribution function is of the form
\begin{equation}
f(x,p) = f_{\rm iso}\!\left(\frac{1}{\lambda}\sqrt{p_\mu \Xi^{\mu\nu} p_\nu}\right) ,
\label{eq:genf}
\end{equation}
where $\lambda$ has dimensions of energy and can be identified with the temperature only in the isotropic equilibrium limit ($\xi^{\mu\nu} = 0$ and $\Phi=0$).\footnote{Herein we assume that the chemical potential is zero.}
We note that, in practice, $f_{\rm iso}$ need not be a thermal equilibrium distribution. However, unless one expects there to be a non-thermal distribution at late times, it is appropriate to take $f_{\rm iso}$ to be a thermal equilibrium distribution function of the form $f_{\rm iso}(x) = f_{\rm eq}(x) = (e^x + a )^{-1}$, where $a= \pm 1$ gives Fermi-Dirac or Bose-Einstein statistics, respectively, and $a=0$ gives Boltzmann statistics. From here on, we assume that the distribution is of Boltzmann form, i.e. $a=0$.
Since the most important viscous corrections are the diagonal components of the energy-momentum tensor, to good approximation one can assume that \mbox{$\xi^{\mu\nu} = {\rm diag}(0,{\boldsymbol \xi})$} with ${\boldsymbol \xi} \equiv (\xi_x,\xi_y,\xi_z)$ and $\xi^i_i=0$.\footnote{For a 1+1d system this is exact since one only needs the $rr$, $\phi\phi$, and $\varsigma\varsigma$ components of the energy-momentum tensor.} In this case, expanding the argument of the square root appearing on the right-hand side of Eq.~(\ref{eq:genf}) in the LRF gives
\begin{equation}
f(x,p) = f_{\rm eq}\!\left(\frac{1}{\lambda}\sqrt{\sum_i \frac{p_i^2}{\alpha_i^2} + m^2}\right) ,
\label{eq:fform}
\end{equation}
where $i\in \{x,y,z\}$ and the scale parameters $\alpha_i$ are
\begin{equation}
\alpha_i \equiv (1 + \xi_i + \Phi)^{-1/2} \, .
\label{eq:alphadef}
\end{equation}
Note that, for compactness, one can collect the three anisotropy parameters into vector $\boldsymbol\alpha \equiv (\alpha_x,\alpha_y,\alpha_z)$. In the isotropic equilibrium limit, where $\xi_i = \Phi = 0$ and $\alpha_i =1$, one has $p_\mu \Xi^{\mu\nu} p_\nu = (p \cdot u)^2 = E^2$ and $\lambda\rightarrow T$ and, hence,
\begin{equation}
f(x,p)=f_{\rm eq}\!\left(\frac{E}{T(x)}\right) .
\label{eq:feqform}
\end{equation}
Out of the four anisotropy and bulk parameters there are only three independent ones. In practice, we use the three variables $\alpha_i$ as the dynamical anisotropy parameters since, by using Eq.~(\ref{eq:alphadef}) and the tracelessness of $\xi^{\mu\nu}$, one can write $\Phi$ in terms of the anisotropy parameters, $\Phi = \frac{1}{3} \sum_i \alpha_i^{-2} - 1$.
\subsection{Equation of state}
\label{subsec:eos}
Herein we consider a system at finite temperature and zero chemical potential. At asymptotically high temperatures, the pressure of a gas of quarks and gluons approaches the Stefan-Boltzmann (SB) limit. At the temperatures probed in heavy-ion collisions there are important corrections to the SB limit and at low temperatures the relevant degrees of freedom change from quarks and gluons to hadrons. The standard way to determine the QGP EoS is to use non-perturbative lattice calculations. For this purpose, we use an analytic parameterization of lattice data for the QCD interaction measure (trace anomaly), $I_{\rm eq} = {\cal E}_{\rm eq} - 3 {\cal P}_{\rm eq}$, taken from the Wuppertal-Budapest collaboration \cite{Borsanyi:2010cj}. We refer the reader to the reference \cite{Alqahtani:2015qja} for more details.
\subsubsection*{Method 1: Standard equation of state}
In the standard approach for imposing a realistic EoS in anisotropic hydrodynamics, one derives the necessary equations in the conformal limit and exploits the conformal multiplicative factorization of the components of the energy-momentum tensor~\cite{Martinez:2010sc,Florkowski:2010cf}. With this method, one relies on the assumption of factorization even in the non-conformal (massive) case. Such an approach is justified by the smallness of the corrections to factorization in the massive case in the near-equilibrium limit~\cite{Nopoush:2015yga}. For details concerning this method, we refer the reader to Refs.~\cite{Ryblewski:2012rr,Nopoush:2015yga}. Although this method is relatively straightforward to implement, it is only approximate since for non-conformal systems there is no longer exact multiplicative factorization of the components of the energy-momentum tensor. This introduces a theoretical uncertainty which is difficult to quantitatively estimate.
\subsubsection*{Method 2: Quasiparticle equation of state}
Since the standard method is only approximate, one would like to find an alternative method for imposing a realistic equation of state in an anisotropic system that can be applied for non-conformal systems. In order to accomplish this goal, we implement the realistic EoS detailed above by assuming that the QGP can be described as an ensemble of massive quasiparticles with temperature-dependent masses. As is well-known from the literature \cite{gorenstein1995gluon}, one cannot simply substitute temperature-dependent masses into the thermodynamic functions obtained with constant masses because this would violate thermodynamic consistency. For an equilibrium system, one can ensure thermodynamic consistency by adding a background contribution to the energy-momentum tensor, i.e.
\begin{equation}
T^{\mu\nu}_{\rm eq} = T^{\mu\nu}_{\rm kinetic,eq} + g^{\mu\nu} B_{\rm eq} \, ,
\end{equation}
with $B_{\rm eq}\equiv B_{\rm eq}(T)$ being the additional background contribution. The kinetic contribution to the energy momentum tensor is given by
\begin{equation}
T^{\mu\nu}_{\rm kinetic, eq} = \int\!dP \, p^\mu p^\nu f_{\rm eq}(x,p) \, .
\end{equation}
For an equilibrium Boltzmann gas, the number and entropy densities are unchanged, while, due to the additional background contribution, the energy density and pressure are shifted by $+B_{\rm eq}$ and $-B_{\rm eq}$, respectively, giving
\begin{eqnarray}
n_{\rm eq} &=& 4 \pi \tilde{N} T^3 \, \hat{m}_{\rm eq}^2 K_2\left( \hat{m}_{\rm eq}\right) , \label{eq:neq} \nonumber \\
{\cal S}_{\rm eq} &=&4 \pi \tilde{N} T^3 \, \hat{m}_{\rm eq}^2 \Big[4K_2\left( \hat{m}_{\rm eq}\right)+\hat{m}_{\rm eq}K_1\left( \hat{m}_{\rm eq}\right)\Big] ,
\label{eq:Seq} \nonumber \\
{\cal E}_{\rm eq} &=& 4 \pi \tilde{N} T^4 \, \hat{m}_{\rm eq}^2
\Big[ 3 K_{2}\left( \hat{m}_{\rm eq} \right) + \hat{m}_{\rm eq} K_{1} \left( \hat{m}_{\rm eq} \right) \Big]+B_{\rm eq} \, ,
\label{eq:Eeq} \nonumber \\
{\cal P}_{\rm eq} &=& 4 \pi \tilde{N} T^4 \, \hat{m}_{\rm eq}^2 K_2\left( \hat{m}_{\rm eq}\right)-B_{\rm eq} \, ,
\label{eq:Peq}
\end{eqnarray}
where $\hat{m}_{\rm eq} = m/T$ with $m$ implicitly depending on the temperature from here on. In order to fix $B_{\rm eq}$, one can require, for example, the thermodynamic identity
\begin{equation}
T {\cal S}_{\rm eq} = {\cal E}_{\rm eq} + {\cal P}_{\rm eq} = T \frac{\partial P_{\rm eq}}{\partial T} \, ,
\label{eq:thermoid}
\end{equation}
be satisfied. Using Eqs.~(\ref{eq:Peq}) and (\ref{eq:thermoid}) one obtains
\begin{eqnarray}
\frac{dB_{\rm eq}}{dT} &=& - \frac{1}{2} \frac{dm^2}{dT} \int\!dP \, f_{\rm eq}(x,p) \nonumber \\
&=& -4\pi \tilde{N}m^2 T K_1(\hat{m}_{\rm eq}) \frac{dm}{dT} \, .
\label{eq:BM-matching-eq-1}
\end{eqnarray}
If the temperature dependence of $m$ is known, then Eq.~(\ref{eq:BM-matching-eq-1}) can be used to determine $B_{\rm eq}$.
In practice, in order to determine $m$, one can use the thermodynamic identity
\begin{equation}
{\cal E}_{\rm eq}+{\cal P}_{\rm eq}=T{\cal S}_{\rm eq} = 4 \pi \tilde{N} T^4 \, \hat{m}_{\rm eq}^3 K_3\left( \hat{m}_{\rm eq}\right) .
\label{eq:meq}
\end{equation}
Using the lattice data parameterization to compute the equilibrium energy density and pressure, one can numerically solve for $m(T)$. We refer the reader to Ref.~\cite{Alqahtani:2015qja} for more details.
\section{Boltzmann equation and its moments}
\label{sec:boltzman}
In this paper, we derive the necessary hydrodynamical equations by taking the moments of Boltzmann equation. In this section, we introduce the Boltzmann equation and its different moments for the general case of quasiparticles with a temperature-dependent mass. Then we simplify them for the case of massless particles, when necessary, since the massless equations are used in the standard approach.
\subsection{Boltzmann Equation}
\label{subsec:thermal-boltzmann}
Generally, the Boltzmann equation for on-shell quasiparticles with temperature dependent mass can be written as ~\cite{Jeon:1995zm,Romatschke:2011qp,Alqahtani:2015qja}
\begin{equation}
p^\mu \partial_\mu f+\frac{1}{2}\partial_i m^2\partial^i_{(p)} f=-\mathcal{C}[f]\,,
\label{eq:boltz2}
\end{equation}
where $p^\mu\equiv(\sqrt{{\bf p}^2+m^2},\bf p)$ is the on-shell momentum four-vector, $i$ indexes the spatial coordinates, and
$\partial^i_{(p)}\equiv -\partial/\partial p^i$. In the constant mass limit, the above Boltzmann equation simplifies to
\begin{equation}
p^\mu \partial_\mu f = -{\cal C}[f]\, .
\label{eq:boltzmanneq}
\end{equation}
The function ${\cal C}[f]$ appearing above is the collisional kernel containing all interactions involved in the dynamics. In what follows, we specialize to the case that the collisional kernel is given by the relaxation-time approximation (RTA), however, the general methods presented here can be applied to any collisional kernel. In RTA, one has
\begin{equation}
{\cal C}[f]=\frac{p^\mu u_\mu}{\tau_{\rm eq}}(f-f_{\rm eq})\,.
\label{eq:RTA}
\end{equation}
In this relation, $f_{\rm eq}$ denotes the equilibrium one-particle distribution function (\ref{eq:feqform}) and $\tau_{\rm eq}$ is the relaxation time which can depend on spacetime but which we assume to be momentum-independent. To obtain a realistic model for $\tau_{\rm eq}$, which is valid for massive systems, one can relate $\tau_{\rm eq}$ to the shear viscosity to entropy density ratio as \cite{anderson1974relativistic,Czyz:1986mr}
\begin{equation}
\eta(T)=\frac{\tau_{\rm eq}(T) {\cal P}_{\rm eq}(T)}{15}\kappa(\hat{m}_{\rm eq})\,.
\end{equation}
In this formula the function $\kappa(x)$ is defined as
\begin{eqnarray}
\kappa(x) &\equiv& x^3 \bigg[\frac{3}{x^2}\frac{K_3(x)}{K_2(x)}-\frac{1}{x}+\frac{K_1(x)}{K_2(x)}
\nonumber \\ && \;\;\;\;
-\frac{\pi}{2}\frac{1-xK_0(x)L_{-1}(x)-xK_1(x)L_0(x)}{K_2(x)}\bigg] , \hspace{5mm}
\end{eqnarray}
where $K_n(x)$ are modified Bessel functions of second kind and $L_n(x)$ are modified Struve functions. Assuming that the ratio of the shear viscosity to entropy density, $\eta/{\cal S}_{\rm eq}\equiv \bar{\eta}$, is held fixed during the evolution and using the thermodynamic relation ${\cal E}_{\rm eq}+{\cal P}_{\rm eq}=T{\cal S}_{\rm eq}$ one obtains
\begin{equation}
\tau_{\rm eq}(T)=\frac{15 \bar{\eta}}{\kappa(\hat{m}_{\rm eq})T}\bigg(1+\frac{{\cal E}_{\rm eq}(T)}{{\cal P}_{\rm eq}(T)}\bigg) .
\label{eq:teq}
\end{equation}
Note that, $\lim_{m\rightarrow 0} \kappa(\hat{m}_{\rm eq})\rightarrow12$, giving
\begin{equation}
\lim_{m\rightarrow 0} \tau_{\rm eq}(T)= \frac{5 \eta}{4{\cal P}_{\rm eq}(T)} \, .
\label{eq:teq0}
\end{equation}
\subsection{Moments of Boltzmann Equation}
\label{subsec:boltzmann-moments}
By calculating moments of Boltzmann equation one obtains evolution equations for tensors of different ranks, with the zeroth moment giving the evolution of particle four-current, the first moment giving the evolution of the energy-momentum tensor, and the second-moment describing the evolution of a particular rank three tensor.
Taking the zeroth, first, and second moments of Boltzmann equation gives, respectively
\begin{eqnarray}
\partial_\mu J^\mu&=&-\int\!dP \, {\cal C}[f]\, , \label{eq:J-conservation} \\
\partial_\mu T^{\mu\nu}&=&-\int\!dP \, p^\nu {\cal C}[f]\, , \label{eq:T-conservation} \\
\partial_\mu {\cal I}^{\mu\nu\lambda}- J^{(\nu} \partial^{\lambda)} m^2 &=&-\int\!dP \, p^\nu p^\lambda{\cal C}[f]\, \label{eq:I-conservation},
\end{eqnarray}
where the particle four-current $J^\mu$, energy-momentum tensor $T^{\mu\nu}$, and the rank-three tensor ${\cal I}^{\mu\nu\lambda}$ are given by
\begin{eqnarray}
J^\mu &\equiv& \int\!dP \, p^\mu f(x,p)\, , \label{eq:J-int} \\
T^{\mu\nu}&\equiv& \int\!dP \, p^\mu p^\nu f(x,p)+B g^{\mu\nu}, \label{eq:T-int}\\
{\cal I}^{\mu\nu\lambda} &\equiv& \int\!dP \, p^\mu p^\nu p^\lambda f(x,p) \, .
\label{eq:I-int}
\end{eqnarray}
We note that we have introduced the non-equilibrium background field $B\equiv B({\boldsymbol\alpha},\lambda)$, which is the analogue of the equilibrium background $B_{\rm eq}$ in order to guarantee that the correct equilibrium limit of $T^{\mu\nu}$ is obtained. In the process of the derivation one finds that, in order to write the energy momentum conservation in the form given in Eq.~(\ref{eq:T-conservation}), there must be a differential equation relating $B$ and the thermal mass \cite{Alqahtani:2015qja}
\begin{equation}
\partial_\mu B = -\frac{1}{2} \partial_\mu m^2 \int\!dP f(x,p)\,.
\label{eq:BM-matching}
\end{equation}
In practice, one can use (\ref{eq:BM-matching}) to write the derivative of $B$ with respect to any variable in terms of the derivative of the thermal mass times the $E^{-1}$ moment of the non-equilibrium distribution function.
\section{Bulk variables}
\label{sec:bulk-var}
In this section, the bulk variables necessary (number density, energy density, and the pressures) are calculated by taking projections of $J^\mu$ and $T^{\mu\nu}$.
\subsection{Particle four-current}
\label{subsec:4current}
The particle four-current $J^\mu\equiv(n,\bf{J})$ is defined in Eq.~(\ref{eq:J-int}). Using Eqs.~(\ref{eq:fform}) and (\ref{eq:J-int}) one has
\begin{equation}
J^\mu=(n,{\bf 0})=nu^\mu \, ,
\label{eq:J-exp2}
\end{equation}
where $n=\alpha n_{\rm eq}(\lambda,m)$ and $\alpha\equiv\alpha_x \alpha_y \alpha_z$.
\subsection{Energy-momentum tensor}
\label{subsec:T-tensor}
The energy-momentum tensor $T^{\mu\nu}$ is defined in Eq.~(\ref{eq:T-int}). Expanding it using the basis vectors one obtains
\begin{equation}
T^{\mu\nu}={\cal E}u^\mu u^\nu+{\cal P}_x X^\mu X^\nu+{\cal P}_y Y^\mu Y^\nu+{\cal P}_z Z^\mu Z^\nu \, .
\label{eq:T-expan}
\end{equation}
\subsubsection{Quasiparticle method}
Using Eqs.~(\ref{eq:fform}), (\ref{eq:T-int}), and (\ref{eq:T-expan}) and taking projections of $T^{\mu\nu}$ one can obtain the energy density and the components of pressure
\begin{eqnarray}
{\cal E} &=& {\cal H}_3({\boldsymbol\alpha},\hat{m}) \, \lambda^4+B \, ,\nonumber \\
{\cal P}_x &=& {\cal H}_{3x}({\boldsymbol\alpha},\hat{m}) \, \lambda^4-B \, ,\nonumber \\
{\cal P}_y &=& {\cal H}_{3y}({\boldsymbol\alpha},\hat{m}) \, \lambda^4-B \, ,\nonumber \\
{\cal P}_z &=& {\cal H}_{3L}({\boldsymbol\alpha},\hat{m}) \, \lambda^4-B \, ,
\label{eq:bulk_var}
\end{eqnarray}
where $\hat{m} \equiv m/\lambda$.
\subsubsection{Standard method}
For a massless conformal Boltzmann gas, one has ${\cal E}_{\rm eq}(\lambda) = 24 \pi \tilde{N}\lambda^4$ and ${\cal P}_{\rm eq}(\lambda) = 8\pi \tilde{N}\lambda^4$. Using these relations, one can rewrite Eqs.~(\ref{eq:bulk_var}) in the standard case, by taking the massless limit, $m \rightarrow0$, and hence $B\rightarrow 0$
\begin{eqnarray}
{\cal E} &=& {\cal E}_{\rm eq}(\lambda) \, \hat{{\cal H}}_3({\boldsymbol\alpha}) \, \, ,\nonumber \\
{\cal P}_x &=& {\cal P}_{\rm eq}(\lambda) \,\hat{{\cal H}}_{3x}({\boldsymbol\alpha}) \, \, ,\nonumber \\
{\cal P}_y &=& {\cal P}_{\rm eq}(\lambda) \,\hat{{\cal H}}_{3y}({\boldsymbol\alpha}) \, \, ,\nonumber \\
{\cal P}_z &=& {\cal P}_{\rm eq}(\lambda) \,\hat{{\cal H}}_{3L}({\boldsymbol\alpha}) \, .
\label{eq:bulk_var_massless_fac}
\end{eqnarray}
These formulas suggest that, in order to impose a realistic EoS, one only has to replace ${\cal E}_{\rm eq}(\lambda)$ and ${\cal P}_{\rm eq}(\lambda)$ by the results obtained from lattice QCD calculations.
\section{Dynamical equations}
\label{sec:dynamical-eqs}
In order to obtain the dynamical equations from Eqs.~(\ref{eq:J-conservation})-(\ref{eq:I-conservation}), one needs to impose the RTA collisional kernel. Enforcing energy-momentum tensor conservation leads to an extra matching (constraint) equation. In total, one has the following four general dynamical equations
\begin{eqnarray}
\partial_\mu J^\mu &=&\frac{1}{\tau_{\rm eq}}(n_{\rm eq}-n) \, , \\
\partial_\mu T^{\mu\nu}&=&0\, , \\
\partial_\mu {\cal I}^{\mu\nu\lambda}-J^{(\nu} \partial^{\lambda)} m^2 &=& \frac{u_\mu}{\tau_{\rm eq}}( {\cal I}^{\mu\nu\lambda}_{\rm eq}- {\cal I}^{\mu\nu\lambda})\,, \label{eq:2ndmom} \\
{\cal E}_{\rm kinetic}&=&{\cal E}_{\rm kinetic,eq}\,.
\end{eqnarray}
Using the tensor decomposition of $J_\mu$, $T_{\mu\nu}$, and ${\cal I}_{\mu\nu\lambda}$ in the basis vectors appropriate for a boost-invariant and azimuthally symmetric 1+1d system, one can expand the expressions above to obtain the final form of the equations \cite{Nopoush:2014pfa,Alqahtani:2015qja}. Choosing two equations from the first moment, three from the second moment, together with the matching condition we end up with six equations for six independent variables $\boldsymbol\alpha$, $\lambda$, $T$, and $\theta_\perp$. The non-trivial equations from the first moment are
\begin{eqnarray}
&& D_u{\cal E}+{\cal E}\theta_u+ {\cal P}_x D_x \theta_{\perp}
\nonumber \\
&& \hspace{1cm} + \frac{1}{r}{\cal P}_y \sinh\theta_{\perp} +\frac{1}{\tau}{\cal P}_z\cosh\theta_{\perp} = 0 \, , \label{eq:1st-mom-1}\\
&& D_x {\cal P}_x+{\cal P}_x\theta_x +{\cal E} D_u\theta_{\perp} \nonumber \\
&& \hspace{1cm} -\frac{1}{r}{\cal P}_y\cosh\theta_{\perp}-\frac{1}{\tau}{\cal P}_z\sinh\theta_{\perp} = 0
\label{eq:1st-mom-2} \,.
\end{eqnarray}
The convective derivatives $D_\alpha$ and divergences $\theta_\alpha$, with $\alpha\in\{u,x,y,z\}$, are defined in App.~\ref{app:identities}. Depending on the model, one can replace the bulk variables from (\ref{eq:bulk_var}) or (\ref{eq:bulk_var_massless_fac}) in the above equations to obtain the final form of dynamical equations for massive quasiparticle or massless standard models, respectively.
Also, taking the $XX$-, $YY$-, and $ZZ$- projections of the second moment equation (\ref{eq:2ndmom}), one obtains
\begin{eqnarray}
\frac{D_u {\cal I}_x}{{\cal I}_x} + \theta_u + 2 D_x \theta_{\perp}
&=& \frac{1}{\tau_{\rm eq}} ( \frac{{\cal I}_{\rm eq}}{{\cal I}_x} - 1 ) \, , \\
\frac{D_u {\cal I}_y}{{\cal I}_y} + \theta_u + \frac{2}{r} \sinh \theta_{\perp}
&=& \frac{1}{\tau_{\rm eq}} ( \frac{{\cal I}_{\rm eq}}{{\cal I}_y} - 1 ) \, , \\
\frac{D_u {\cal I}_z}{{\cal I}_z} + \theta_u + \frac{2}{\tau} \cosh \theta_{\perp}
&=& \frac{1}{\tau_{\rm eq}} ( \frac{{\cal I}_{\rm eq}}{{\cal I}_z} - 1 ) \, .
\end{eqnarray}
Evaluating the necessary integrals using the distribution function (\ref{eq:fform}), one finds
\begin{eqnarray}
{\cal I}_u &=& \Big(\sum_i \alpha_i^2\Big) \alpha \, {\cal I}_{\rm eq}(\lambda,m) + \alpha m^2 n_{\rm eq}(\lambda,m) \, ,
\\
{\cal I}_i &=& \alpha \, \alpha_i^2 \, {\cal I}_{\rm eq}(\lambda,m) \, ,
\label{eq:I-i}
\end{eqnarray}
with ${\cal I}_{\rm eq}(\lambda,m) = 4 \pi {\tilde N} \lambda^5 \hat{m}^3 K_3(\hat{m})$. In the massless case, they simplify to
\begin{eqnarray}
{\cal I}_u &=& \Big(\sum_i \alpha_i^2\Big) \alpha \, {\cal I}_{\rm eq}(\lambda) \, ,
\\
{\cal I}_i &=& \alpha \, \alpha_i^2 \, {\cal I}_{\rm eq}(\lambda) \, ,
\end{eqnarray}
with ${\cal I}_{\rm eq}(\lambda) = 32 \pi {\tilde N} \lambda^5$. Finally, the matching condition is
\begin{eqnarray}
{\cal H}_3 ({\boldsymbol\alpha},\hat{m})\lambda^4 &=& {\cal H}_{3,\rm eq}(\hat{m}_{\rm eq}) T^4,
\end{eqnarray}
where
\begin{equation}
{\cal H}_{3,\rm eq}(\hat{m}_{\rm eq})\equiv \lim_{\substack{\lambda\rightarrow T \\ \boldsymbol\alpha\rightarrow 1 }} {\cal H}_3 ({\boldsymbol\alpha},\hat{m}).
\end{equation}
\section{Anisotropic Freeze-out}
\label{sec:freeze-out}
We now turn to the topic of hadronic freeze-out. Our technique will be to perform ``anisotropic Cooper-Frye freeze-out'' using Eq.~(\ref{eq:genf}) as the form for the one-particle distribution function. This is different than the typical freeze-out prescription used in viscous hydrodynamics in which one takes into account the dissipative correction to the equilibrium distribution function only at linear order in a Taylor expansion around equilibrium. One immediate benefit of performing anisotropic freeze-out using Eq.~(\ref{eq:genf}) is that, with this form, one is guaranteed that the one-particle distribution function is positive-definite in all regions in phase space.
In practice, we start from the standard freeze-out integral
\begin{equation}
N=\int_{\Sigma} d^3\Sigma_\mu J^\mu \, ,
\label{eq:particle-number1}
\end{equation}
where $\Sigma$ is the three-dimensional freeze-out hypersurface defining the boundary of the four-dimensional volume occupied by the fluid, $d^3\Sigma^\mu$ is the surface normal vector, and $J^\mu$ is the particle four-current. Due to the presence of momentum-space anisotropies, one cannot simply use the momentum scale $\lambda$ when defining the freeze-out hypersurface $\Sigma$. Instead, one should use the energy density, from which one can obtain the effective freeze-out temperature $T_{\rm FO} \equiv T_{\rm eff,FO}= T({\cal E}_{\rm FO})$ where $T({\cal E})$ is obtained using our realistic EoS. After identifying $\Sigma$, we follow the parametrization presented in \cite{Nopoush:2015yga}.
Unlike \cite{Nopoush:2015yga}, here we take into account the breaking of conformality. The only change required is in the distribution function itself.
Parameterizing the particle momentum in the lab frame as
\begin{eqnarray}
p^\mu\equiv(m_\perp \cosh y,p_\perp \cos\varphi,p_\perp \sin\varphi,m_\perp \sinh y)\, ,
\label{eq:ptl-mom}
\end{eqnarray}
where $m_\perp = \sqrt{p_\perp^2+m^2}$, $y = \tanh^{-1}(p^z/p^0)$ is the particle's rapidity, and $\varphi$ is the particle's azimuthal angle.
In order to set up the distribution function, having $p^\mu$ defined in Eq.~(\ref{eq:ptl-mom}), one can use Eqs.~(\ref{eq:aniso-tensor1}) and (\ref{eq:genf}) to find in both the quasiparticle and standard cases
\begin{widetext}
\begin{eqnarray}
p^\mu \Xi_{\mu\nu} p^\nu &=&
(1+\Phi)\Big[m_\perp \cosh \theta _\perp \cosh
(y-\varsigma)-p_\perp \sinh \theta _\perp \cos (\phi -\varphi )\Big]^2
\nonumber \\
&+& \xi _x\,\Big[m_\perp \sinh \theta _\perp \cosh (y-\varsigma)-p_\perp \cosh \theta
_\perp \cos (\phi -\varphi )\Big]^2
\nonumber \\
&+& \xi_z\, m_\perp^2 \sinh^2(y-\varsigma)
+\xi_y\,\, p_\perp^2 \sin^2(\phi -\varphi )
- \Phi m^2 \,.
\label{eq:pxipqu}
\end{eqnarray}
\end{widetext}
Note that $ p^\mu \Xi_{\mu\nu} p^\nu $ is Lorentz invariant, and by going to LRF, one can show that this is positive definite in any frame.
\begin{figure}[t]
\includegraphics[width=0.9\linewidth]{spectrum_1.pdf}
\caption{Comparison of the neutral pions, kaons ($K^+$), and protons spectra as a function of transverse momentum $ p_T$ obtained using aHydroQP, aHydro and vHydro. The shear viscosity to entropy density ratio is $ 4 \pi \eta / s =1$.}
\label{fig:spectrum1}
\end{figure}
\begin{figure}[t]
\includegraphics[width=0.9\linewidth]{spectrum_3.pdf}
\caption{Same as Fig.~\ref{fig:spectrum1}, except here the shear viscosity to entropy density ratio was taken to be $ 4 \pi \eta / s =3 $.}
\label{fig:spectrum3}
\end{figure}
\begin{figure}[t]
\includegraphics[width=0.9\linewidth]{spectrum_10.pdf}
\caption{Same as Fig.~\ref{fig:spectrum1}, except here the shear viscosity to entropy density ratio was taken to be $ 4 \pi \eta / s =10 $.}
\label{fig:spectrum10}
\end{figure}
\begin{figure}[t]
\vspace{-1mm}
\includegraphics[width=0.89\linewidth]{N_pi_k_p.pdf}
\caption{Comparison of the total number of pions ($\pi^0$), kaons ($K^0$), and protons as a function of $ 4 \pi \eta / s $ obtained using aHydroQP, aHydro, and vHydro. }
\label{fig:numberPKPr}
\end{figure}
\begin{figure}[t]
\vspace{1mm}
\includegraphics[width=0.89\linewidth]{pt_avg.pdf}
\caption{Comparison of the neutral pions, kaons ($K^+$), and protons average transverse momentum as a function of $ 4 \pi \eta / s $ obtained using aHydroQP, aHydro, and vHydro.}
\label{fig:ptavg}
\end{figure}
\begin{figure}[t]
\centerline{\includegraphics[width=0.9\linewidth]{N_chg.pdf}\hspace{5mm}}
\caption{The number of total charged particles as a function of $ 4 \pi \eta / s $ obtained using aHydroQP, aHydro, and vHydro.}
\label{fig:numberchg}
\end{figure}
\begin{figure}[t]
\includegraphics[width=0.9\linewidth]{spectrumpPb_1.pdf}
\caption{Comparisons of the p-Pb neutral pions, kaons ($K^+$), and protons spectra as a function of transverse momentum $ p_T$ obtained using aHydroQP, aHydro, and vHydro. The shear viscosity to entropy density ratio is $ 4 \pi \eta / s =1$. }
\label{fig:spectrumpPb1}
\end{figure}
\begin{figure}[t]
\includegraphics[width=0.9\linewidth]{spectrumpPb_3.pdf}
\caption{Same as Fig.~\ref{fig:spectrumpPb1}, except here the shear viscosity to entropy density ratio was taken to be $ 4 \pi \eta / s =3 $. }
\label{fig:spectrumpPb3}
\end{figure}
\section{Results}
\label{sec:results}
We now turn to our numerical results. We present comparisons of results obtained using the dynamical equations of anisotropic hydrodynamics presented in Sec.~\ref{sec:dynamical-eqs} and the second-order viscous hydrodynamics equations from Denicol et al. \cite{Denicol:2012cn,Denicol:2014vaa}. For details about the vHydro equations solved herein we refer the reader to App.~\ref{app:vhydro}.
\paragraph*{Pb-Pb collisions:}
For all results presented in this section we use smooth Glauber wounded-nucleon overlap to set the initial energy density. As our test case we consider Pb-Pb collisions with $\sqrt{s_{\rm NN}} = $ 2.76 GeV. The inelastic nucleon-nucleon scattering cross-section is taken to be $\sigma_{\rm NN} = $ 62 mb. For the aHydro results, we use 200 points in the radial direction with a lattice spacing of $\Delta r = 0.15$ fm and temporal step size of $\Delta \tau =$ 0.01 fm/c. For the vHydro results, we use 600 points in the radial direction with a lattice spacing of $\Delta r = 0.05$ fm and temporal step size of $\Delta \tau =$ 0.001 fm/c.\footnote{We found that the vHydro code was more sensitive to the spatial lattice spacing and required a smaller temporal step size for stability.} In all cases, we use fourth-order Runge-Kutta integration for the temporal updates and fourth-order centered differences for the evaluation of all spatial derivatives.\footnote{Since the initial conditions considered herein are smooth, naive centered differences generally suffice.} We take the central initial temperature to be $T_0 = 600$ MeV at $\tau_0 = 0.25$ fm/c and assume that the system is initially isotropic, i.e. $\alpha_x(\tau_0) =\alpha_y(\tau_0) = \alpha_z(\tau_0) = 1$ for anisotropic hydrodynamics and $\pi^{\mu\nu}(\tau_0)=\Pi(\tau_0)=0$ for second-order viscous hydrodynamics. We take the freeze-out temperature to be $T_{\rm eff} = T_{\rm FO} = $ 150 MeV in all cases shown. For the freeze-out we use 371 hadronic resonances ($M_{\rm hadron} \leq 2.6 \; {\rm GeV}$), with the masses, spins, etc. taken from the SHARE table of hadronic resonances \cite{Torrieri:2004zz,Torrieri:2006xi,Petran:2013dva}. We do not perform resonance feed down, hence all spectrum shown herein are primordial spectra.
In Figs.~\ref{fig:spectrum1}-\ref{fig:spectrum10} we present our results for the primordial pion, kaon, and proton spectra produced for $4\pi\eta/s = 1, 3,$ and 10, respectively. In each case, we have held the initial conditions fixed and only varied $\eta/s$. In each of these figures the solid black line is the result from standard second-order viscous hydrodynamics (vHydro), the red short-dashed line is the result from standard anisotropic hydrodynamics (aHydro), and the blue long-dashed line is the result from quasiparticle anisotropic hydrodynamics (aHydroQP). As can be seen from Fig.~\ref{fig:spectrum1}, for $4\pi\eta/s=1$, both aHydro approaches are in good agreement over the entire $p_T$ range shown with the largest differences occurring at low momentum. The second-order viscous hydrodynamics result, however, shows a significant downward curvature in the pion spectrum resulting in many fewer high-$p_T$ pions. The trend is the same for the kaon and proton spectra, however, for the larger mass hadrons the downturn is less severe. We note that although we plot only up to $p_T = 3$ GeV, these plots can be extended to larger $p_T$, in which case one finds that eventually the primordial pion spectrum predicted by vHydro becomes negative, which is clearly unphysical. This can be seen more clearly in Figs.~\ref{fig:spectrum3} and \ref{fig:spectrum10} which show the same results for $4\pi\eta/s = 3$ and 10, respectively. As these figures demonstrate, for larger $\eta/s$ the vHydro primordial particle spectra become unphysical at lower momenta. For example, for $4\pi\eta/s=3$ the differential pion spectrum goes negative at $p_T \sim 1.6$ GeV while for $4\pi\eta/s=10$ it goes negative at $p_T \sim 0.9$ GeV.
This behavior is a result of the bulk-viscous correction to the one-particle distribution function specified in Eq.~(\ref{eq:deltafbulk}). We have checked that if we neglect the bulk-viscous correction to the distribution function, then the resulting spectra are positive definite in the range of $p_T$ shown in Figs.~\ref{fig:spectrum1}-\ref{fig:spectrum10}. We have verified that this is a known issue with the bulk-viscous correction in the second-order viscous hydrodynamics approach. The same form for the bulk correction (\ref{eq:deltafbulk}) was used by the authors of Ref.~\cite{Ryu:2015vwa} and they also observed a downward curvature turning into negatively-valued spectra at large $p_T$ \cite{SchenkeDenicolPrivate}. This problem does not occur in either aHydro approach because the one-particle distribution function is positive-definite by construction in this framework.
Next, we turn to a discussion of Fig.~\ref{fig:numberPKPr}. In this figure we plot the total number of $\pi^0$'s (top), $K^0$'s (middle), and $p$'s (bottom) obtained by integrating the differential yields over transverse momentum as a function of $4\pi\eta/s$. The line styles are the same as in Figs.~\ref{fig:spectrum1}-\ref{fig:spectrum10}. From this figure we see that at small $\eta/s$ all approaches are in agreement, however, at large $\eta/s$ the three methods can give dramatically different results. We, in particular, note that the number of pions from vHydro drops much quicker than the two aHydro approaches. This is primarily due to the fact that in vHydro one sees a negative number of pions at large $p_T$. As shown in Fig.~\ref{fig:numberPKPr} (top), the total number of pions predicted by vHydro goes negative at $4\pi \eta/s \sim 22$. This is a signal of the complete breakdown of viscous hydrodynamics and should come as no surprise since this approach is intended to be applicable only in the case of small $\eta/s$ and $p_T$. Finally, we note that although aHydro and aHydroQP are in reasonably good agreement for the pion and kaon spectra, we see a rather strong dependence on the way the EoS is implemented in the proton spectra and, hence, the total number of primordial protons.
As an additional way to compare the three methods considered, in Fig.~\ref{fig:ptavg} we present the average $p_T$ for $\pi^0$'s (top), $K^+$'s (middle), and $p$'s (bottom). The line styles are the same as in Figs.~\ref{fig:spectrum1}-\ref{fig:spectrum10}. From this figure we see that both aHydro approaches predict a weak dependence of the pion and kaon $\langle p_T \rangle$ on the assumed value of $\eta/s$, whereas vHydro predicts a much more steep decrease in $\langle p_T \rangle$ for the pions and kaons. Once again, we see that the vHydro $\langle p_T \rangle$ for pions becomes negative for $4\pi\eta/s \gtrsim 5$. This rapid decrease stems directly from the negativity of the pion spectra at high $p_T$ which is more important in this case since the integrand is more sensitive to the high-$p_T$ part of the spectra. Finally, we note that once again the proton spectra and hence $\langle p_T \rangle$ for protons is sensitive to the way in which the EoS is implemented when comparing the two aHydro approaches.
Finally, in Fig.~\ref{fig:numberchg} we plot the total number of charged particles as a function of $4\pi\eta/s$ predicted by each of the three approaches considered. Once again the line styles are the same as in Figs.~\ref{fig:spectrum1}-\ref{fig:spectrum10}. As this figure demonstrates, for small $\eta/s$ all three frameworks are in agreement and approach the ideal result as $\eta/s$ tends to zero. All three frameworks predict that $N_{\rm chg}$ at first increases, then reaches a maximum, and then begins to decrease. The precise turnover point depends on the method with the lowest turnover seen using vHydro around $4 \pi \eta/s \sim 3$; however, this turnover is due in large part to the fact that the total pion number drops precipitously in vHydro, eventually becoming negative, as we pointed out in the discussion of Fig.~\ref{fig:numberPKPr}. As a result, one cannot trust the large $\eta/s$ predictions of vHydro.
\paragraph*{p-Pb collisions:}
We now turn to the case of an asymmetric collision between a proton and a nucleus. For this purpose, we use the same parameters as used for the Pb-Pb collisions considered previously, except for p-Pb we use a lower initial temperature of $T_0 = 400$ MeV . In addition, for the aHydro results, we use a lattice spacing of $\Delta r = 0.06$ fm and for vHydro results, we use a lattice spacing of $\Delta r = 0.02$ fm. The smaller lattice spacings simply reflect the smaller system size of the QGP created in a p-Pb collision.
In Figs.~\ref{fig:spectrumpPb1}-\ref{fig:spectrumpPb3} we present our results for the primordial pion, kaon, and proton spectra produced for $4\pi\eta/s = 1,$ and 3, respectively in p-Pb collisions. As we can see from Fig.~\ref{fig:spectrumpPb1}, both aHydro approaches are in a good agreement at high $p_T$, however, there are some quantitative differences at low $p_T$. Comparing the low $p_T$ difference with that seen in Pb-Pb collisions, we find that there is a larger variation in the p-Pb spectra comparing aHydro and aHydroQP. This variation is a bit worrisome since it indicates a kind of theoretical uncertainty in the aHydro approach. Importantly, however, we mention that the two aHydro results are quite different than the vHydro result. For p-Pb collisions, the vHydro result shows the same behavior seen in the Pb-Pb collisions, namely that the particle spectra goes negative at high $p_T$. In Fig.~\ref{fig:spectrumpPb3}, one can clearly see the unphysical behavior in the vHydro primordial spectra. As a result, there is even larger theoretical uncertainty associated with applications of vHydro to p-Pb collisions.
\section{Conclusions and Outlook}
\label{sec:conclusions}
In this paper, we compared three different viscous hydrodynamics approaches: aHydro, aHydroQP, and vHydro. For all three cases we included both shear and bulk-viscous effects using the relaxation time approximation scattering kernel. In the standard aHydro approach one uses the standard method for imposing an equation of state in anisotropic hydrodynamics, which is to obtain conformal equations and then break the conformality only when introducing the EoS itself. In aHydroQP, one takes into account the breaking of conformality at the outset by modeling the QGP as a quasiparticle gas with a single temperature-dependent mass $m$ which is fit to available lattice data for the EoS. Finally, for our comparisons with viscous hydrodynamics we used the formalism of Denicol et al \cite{Denicol:2012cn,Denicol:2014vaa} specialized to the case of relaxation time approximation.
For each method, we specialized to the case of 1+1d boost-invariant and azimuthally-symmetric collisions using smooth Glauber initial conditions. We specialized to this case because of the computational intensity of the aHydroQP approach which requires real-time evaluate of complicated multi-dimensional integrals which are functions of all three anisotropy parameters and the local temperature-dependent mass. Using the resulting numerical evolution, we then extracted fixed energy density freeze-out hypersurfaces in each case and implemented the scheme-appropriate freeze-out to hadrons allowing us to have an apples-to-apples comparison between the three different approaches. We found that the primordial particle spectra, total number of charged particles, and average transverse momentum predicted by the three methods agree well for small shear viscosity to entropy density ratio, $\eta/s$, but differ at large $\eta/s$. Our most important finding was that when using standard viscous hydrodynamics, the bulk-viscous correction can drive the primordial particle spectra negative at large $p_T$. Such a behavior is not seen in either anisotropic hydrodynamics approach, irrespective of the value of $\eta/s$.
Looking to the future, it is feasible to extend the aHydroQP approach to (3+1)d, however, this will be numerically intensive and require parallelization to implement fully. One possibility is to use polynomial fits to parametrize the various massive ${\cal H}$-functions necessary instead of evaluating them on-the-fly in the code. If this is possible, then aHydroQP could become a viable alternative to the standard method of implementing the EoS in aHydro and, since it takes into account the non-conformality of the system from the beginning, this could give us an idea of the theoretical uncertainty associated with the EoS method used in phenomenological applications. For now, this work will serve as a reference point for possible differences between the two approaches to imposing the aHydro EoS.
\acknowledgments{We thank G. Denicol and B. Schenke for useful discussions. M. Strickland and M. Nopoush were supported by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics under Awards No.~DE-SC0013470. M. Alqahtani was supported by a PhD fellowship from the University of Dammam.}
\begin{widetext}
|
\section{Introduction and Related Work}
\label{sec:Rlt_wk}
In the past few years, there has been an increased interest in millimeter wave (mmWave) technology to fulfill the data rate requirements foreseen for the fifth generation cellular communication (5G) \cite{KhanF_mmWave}.
However, frequencies in mmWave bands experience high path-loss, which in comparison to microwave bands may result in a significant coverage reduction when considering omnidirectional communication.
To overcome these coverage issues, beamforming at mmWave is an effective solution.
Due to the small wavelengths at mmWave frequencies, a large number of antennas can be packed in a small space, and this allows to generate high gains and highly directional beams.
Beamforming solutions proposed for LTE are quite different from what is required for mmWave communication.
In LTE, initial access is performed using omnidirectional communication, and beamforming is used only when directional information is available after initial physical layer access.
However, in mmWave communication beamforming may be required even during the initial access/cell search process to overcome the coverage problems of omnidirectional communication with mmWaves.
Recently, two different approaches have been considered for directional initial access/cell discovery.
Firstly, in \cite{CaponeFS15_CI_BF}, considering a HetNet scenario, context information (CI) regarding mobile station (MS) positioning is provided to the mmWave base station (BS) by the microwave BS.
Based on this, the mmWave BS points its beam (using analog beamforming) in the desired direction.
The authors also address the issue of erroneous CI,
proposing that the BS, in addition to searching in the CI based direction, also searches the rest of the angular space by forming beams in different directions and also with different beamwidth (to increase the coverage).
Results showed that the enhanced cell discovery, where in case of positioning error the BS searches the adjacent angular directions, outperforms the greedy search approach where the BS searches the angular space sequentially.
In addition, at the MS, omnidirectional reception is considered, which in comparison to a directional reception results in a reduced gain.
Recently, the authors of \cite{CaponeFS15_CI_BF} extended their work by considering a more complex channel model with multiple rays and obstacles \cite{Capone_ObsAv15}.
Secondly, a non context information based approach is used in \cite{Barati_IBF, Alkhateeb_IBF, Marco_IniAcc16}.
In \cite{Barati_IBF}, omnidirectional and random directional transmissions are considered from the mmWave BS, and the MS can perform either analog, digital or hybrid beamforming.
It is observed that using an omnidirectional transmission at the BS and digital beamforming at the MS outperforms the other schemes.
However, the use of omnidirectional transmission at the BS during initial synchronization/access will reduce coverage, while random directional transmission will result in large delays.
In \cite{Alkhateeb_IBF}, exhaustive and hierarchical searches are compared while considering analog, digital and hybrid beamforming at the BS and the MS.
In exhaustive search, the whole angular space is covered by sequentially transmitting beams in a time division multiplexing fashion, and initial beamforming is done by selecting the best combination of Tx-Rx beams.
The hierarchical search, instead, is a multiple step process.
In the first step, a MS initially utilizes fewer antennas to form a relatively small number of wide beams.
The received signal is combined with all the beams and the best combiner beam is selected as a reference for the next step, where several narrower beamwidth combiners are formed, within the initially selected beam.
Considering scenarios with limited mobility, the process finishes when the combiner beam is within the range of $5^\circ$ to $10^\circ$.
However, selecting an incorrect combiner in the initial stage can result in an initial access error in the following stage.
Recently, in \cite{Marco_IniAcc16}, iterative and exhaustive search schemes using analog beamforming have been studied and compared, and the authors showed that the optimal scheme depends on the target SNR regime.
\subsection{Our Approach}
\label{subsec:Our_app}
Analog beamforming (ABF) results in a lower power consumption compared to digital and hybrid beamforming (DBF and HBF) but has been shown to have comparable performance in terms of initial access probability when combined with an exhaustive search \cite{Alkhateeb_IBF}.
Although the exhaustive approach is not preferred in general due to its inherent search delay,
the availability of CI can greatly improve its delay performance.
To address this, in contrast to \cite{CaponeFS15_CI_BF}, we consider the availability of CI at the MS, so that the MS, rather than searching the whole angular space, will form its combining beam only in the direction provided by the CI\footnote{The initial cell search delay associated with the directional communication (for $N_{BS}=64$ and $N_{MS}=16$ with ABF) is roughly around 10 ms \cite{my_ICD16}, and is small enough to ingnore the rotational motion of the MS, which makes the directional signal reception at the MS a feasible option.}.
This will result in a reduced initial cell search delay and, in comparison to omnidirectional reception (as in \cite{CaponeFS15_CI_BF}), the ABF at the MS will provide a higher gain.
Moreover, the availability of CI at the BS does not necessarily minimize the delay, as if there are multiple MSs belonging to different beams the BS has to scan all of them, losing at least part of the delay savings.
On the contrary, in initial cell search the MS typically listens to a single (or a small number of) BS, and therefore the availability of CI allows it to form a single beam in the right direction, thereby avoiding the beam scanning delay.
We consider a HetNet scenario
for transferring the CI to the MS, where a microwave BS during the exchange of the initial control signals also transfers the location information of the mmWave BS (e.g., global positioning system (GPS) coordinates) to the MS.
Due to the recent increase in location based applications and GPS positioning accuracy, GPS based CI may be considered as a viable option.
We assume that the GPS coordinates of the MS are already available.
The MS, after acquiring the CI, figures out the expected angle of arrival (AoA) and aligns its beam in the required direction to receive the initial synchronization signals from the mmWave BS.
In this work, our results give a quantitative evaluation of the benefit of having CI, by measuring how much better a mmWave system can perform with the availability of CI.
A detailed assessment of the cost of obtaining the CI is beyond the scope of this paper, and is left as future work.
Specifically, in this paper we address the following issues:
\begin{itemize}
\item how ABF with CI performs in comparison to ABF with non CI based approaches (Random \cite{Barati_IBF} and Exhaustive \cite{Alkhateeb_IBF} search);
\item how the angular error in the provided CI will affect the performance of the initial access process;
\item how the optimal number of MS antennas that results in the best access performance varies with the angular error in the available CI.
\end{itemize}
Finally, we propose an analog beamforming based phase shifters network (PSN) architecture (see Figure \ref{fig:PSN}) to mitigate the effect of the angular error in the available CI, and also compare its performance with hybrid and digital beamforming in terms of initial access error and power consumption.
The rest of the paper is organized as follows.
In Section \ref{sec:Sys_mod}, we define the MIMO based system with beamforming and extend the idea to CI and also present our PSN architecture.
Next, we discuss the simulation results for the different mentioned beamforming schemes in Section \ref{sec:Sim_Res}, and finally conclude the paper in Section \ref{sec:Con}.
\section{System Model}
\label{sec:Sys_mod}
Consider a downlink mmWave MIMO communication system, with $N_{BS}$ BS antennas and $N_{MS}$ MS antennas.
A typical MIMO received signal model, without considering beamforming, is given as \cite{rappaport2014millimeter}
\begin{equation}
\textbf{y} = \sqrt{P}\textbf{Hs}\ +\ \textbf{n}
\end{equation}
\normalsize
where $\textbf{H}$ is $N_{MS} \times N_{BS}$ matrix which represents the channel between the BS and the MS, $\textbf{s}$ and $\textbf{y}$ are the transmitted and received symbol vectors, respectively, $P$ is the transmitted power, and $\bf n$ is the complex white Gaussian noise, $\textbf{n} \sim \mathcal{CN}(0,\sigma^2)$.
The mmWave MIMO channel can be modeled with a few scatterers ($L$) \cite{KhanF_mmWave} and is well represented by the following geometric model \cite{Alkhateeb14_ChanEst}
\begin{equation}
\textbf{H} = \sqrt{\dfrac{N_{BS}N_{MS}}{\rho L}}\sum_{l=1}^{L}\eta_l\textbf{a}_{MS}(\phi_{l}) \textbf{a}_{BS}^H(\theta_{l})
\end{equation}
\normalsize
where $H$ represents the conjugate transpose, $\rho$ is the path-loss, $\eta_l$ is the complex gain associated with the $l^{th}$ path, $\textbf{a}_{MS}$ and $\textbf{a}_{BS}$ are the spatial signatures of the MS and the BS, respectively, and $\phi_l$ and $\theta_l$ $\in [0,2\pi)$ represent the AoA and angle of departure (AoD) of the $l^{th}$ path at the MS and the BS, respectively.
In this paper, for simplicity we restrict our analysis to single path line of sight (LOS) scenarios and therefore will not consider the subscript $l$ in further analysis, whereas the study of a multi-path scenario is left as future work.
Moreover, both the MS and the BS are equipped with a uniform linear array (ULA).
Now, considering the ULA at the BS,
$\textbf{a}_{BS}$ is defined as
\begin{equation}
\textbf{a}_{BS} = \dfrac{1}{\sqrt{N_{BS}}}[1, e^{j(2\pi /\lambda) d\sin (\theta)}, ... , e^{j(N_{BS} - 1)(2\pi /\lambda)d\sin (\theta)}]^T
\end{equation}
where $T$ represents the transpose, $d$ is the spacing between antenna elements, and $\lambda$ is the wavelength of the transmitted signal. For $d = \lambda/2$, $\textbf{a}_{BS}$ becomes
\begin{equation}
\textbf{a}_{BS} = \dfrac{1}{\sqrt{N_{BS}}}[1, e^{j{\pi}\sin (\theta)}, ... , e^{j(N_{BS} - 1){\pi}\sin (\theta)}]^T
\end{equation}
\normalsize
and the spatial signature $\textbf{a}_{MS}$ can be defined similarly.
\subsection{Signal Model with Analog Beamforming}
\label{ABF}
The received signal, after applying beamforming and combining at the transmitter (the BS) and the receiver (the MS), can be written as
\begin{equation}
y = \sqrt{P}\textbf{w}_{MS}^H\textbf{H}\textbf{w}_{BS}{s}\ +\ \textbf{w}_{MS}^H\textbf{n}
\label{eq:Tx-Rx_BF}
\end{equation}
where $\textbf{w}_{BS}$ and $\textbf{w}_{MS}$ are the transmit beamforming and receiver combining vectors, respectively.
For simplicity in the notation, the subscripts for the BS and the MS are removed in our general description of beamforming vectors, however, we explicitly use them whenever necessary.
\begin{figure}
\centering
\psfrag{Access Error Probability}{$p$}
\scalebox{1}{\includegraphics[width=\columnwidth]{PSN}}\vspace{-4mm}
\caption{\protect\renewcommand{\baselinestretch}{1}\footnotesize A Phase Shifters Network based receiver architecture.}\vspace{-4mm}
\label{fig:PSN}
\end{figure}
Considering ABF, the beamforming vector $\textbf{w}$ models the analog phase shifters which apply successively progressive phases to the antenna elements.
These phase shifters are digitally controlled and thus result in a finite number of possible phases.
Considering $q$ quantization bits, $2^q$ different phases can be applied to each element of an antenna array.
The set of these quantized phases is represented as
\begin{equation}
{t} = \bigg{\{}0,\ 2\pi(\frac{1}{2^q}),\ 2\pi(\frac{2}{2^q}),\ ...,\ 2\pi(\frac{2^q-1}{2^q})\bigg{\}}
\label{eq:Quan_Ph}
\end{equation}
and with $N$ antennas, a codebook of $2^{qN}$ beamforming vectors can be generated.
However, in order to reduce the complexity, we only consider a subset of this codebook, and let $\textbf{W} = [\textbf{w}^1\ \textbf{w}^2\ \ldots\ \textbf{w}^{N_q}]$ represent an $N\times N_q$ reduced size codebook, where each column of $\textbf{W}$ represents a unique beamforming vector.
Now for a quantized phase $\vartheta_i \in {t}$, we generate the receiver combining vector $\textbf{w}^i$ as
\begin{equation}
\textbf{w}^i = \dfrac{1}{\sqrt{N}} [1,\ e^{j\vartheta_i},\ e^{j2\vartheta_i},\ ...,\ e^{j(N-1)\vartheta_i}]^T
\label{eq:BF_vec}
\end{equation}
where $i = 1,\ 2, \ldots\ N_q$.
Moreover, we set the cardinality of $\textbf{W}$ as $card(\textbf{W}) = N_q = 2N$, and therefore by definition of $\textbf{w}^i$ in (\ref{eq:BF_vec}), the required number of quantization bits is $q = \log _2 (2N)$.
This gives an acceptable performance, as shown in \cite{ABF_Codebook_Wang09}.
The transmitter and the receiver beamforming vectors
are assumed to be defined according to the above approach, and $\textbf{W}_{MS}$ and $\textbf{W}_{BS}$ represent the beamforming codebooks at the MS and the BS, respectively.
\subsection{MS Combining Vector Selection with the Available CI}
We consider a scenario where the BS transmits sequentially in different angular directions using ABF, and the MS combines the received signal with a combining vector selected based on the available CI.
Note that the quantized phase $\vartheta_i$ is related to the physical AoA ($\phi$) as $\phi = \sin^{-1}(\vartheta_i/\pi)$.
The MS first estimates the CI based AoA ($\phi_{CI}$) and then identifies $\vartheta_i$ that minimizes the angular distance with $\phi_{CI}$.
Finally, the combining vector that corresponds to the best $\vartheta_i$ is selected.
This combining vector maximizes the norm $|{\textbf{w}_{MS}^i}^H\textbf{a}_{MS}(\phi_{CI})|$, where $\textbf{a}_{MS}(\phi_{CI})$ represents the CI based estimated spatial signature at the MS.
This maximization problem can be expressed as
\begin{equation}
\begin{split}
\textbf{w}_{MS}^\star =\ &\operatorname*{arg\,max}_{\textbf{w}_{MS}^i} |{\textbf{w}_{MS}^i}^H\textbf{a}_{MS}(\phi_{CI})| \\
&s.t.\ {\textbf{w}_{MS}^i} \in \textbf{W}_{MS}
\end{split}
\label{eq:CI_dist}
\end{equation}
the formulation of (\ref{eq:CI_dist}) also maximizes the signal to noise ratio ($SNR$) of the received signal (Equation (\ref{eq:Tx-Rx_BF})), expressed a
\begin{equation}
SNR = \frac{|\textbf{w}_{MS}^H \textbf{H} \textbf{w}_{BS}|^2P}{||\textbf{w}_{MS}^H||^2\sigma^2}
\label{eq:SNR}
\end{equation}
\normalsize
where $P$ is the power of the transmitted signal, and $\sigma^2$ is the noise power.
We now extend this approach by considering an error in the available CI.
This will add an angular error $\phi_e$ to the estimated AoA, which may result in a wrong selection of $\vartheta_i$ and $\textbf{w}_{MS}$, which finally decreases the $SNR$ of the received signal.
We model $\phi_e$ as uniformly distributed in $[-\phi_e^{max}, \phi_e^{max}]$, where $\phi_e^{max}$ represents the maximum angular error\footnote{The general trends followed by different beamforming schemes (as shown in Section \ref{sec:Sim_Res}) will be maintained if different angular error distributions are used.}.
The amount by which $\phi_e$ affects the selection of $\vartheta_i$ is related to the 3 dB beamwidth $\theta_{BW}$ of the antenna array,
which for a ULA with $ d = \lambda/2$ and $N$ antennas can be computed as $\theta_{BW} = 2\sin^{-1}(0.891/N)$ \cite{oap-b1105852}.
Therefore, for a reliable performance it is desirable to have $\phi_e^{max} \leq \theta_{BW}$.
However, to address larger $\phi_e^{max}$, i.e., $\phi_e^{max} > \theta_{BW}$, one possible solution is to increase $\theta_{BW}$ by reducing the number of antennas at the cost of a reduced gain. The other option is to form multiple beams simultaneously, as discussed in the next section
\subsection{Phase Shifters Network to Combat $\phi_e$ }
\label{ssec:PSN}
The performance of ABF (which only results in a single beam) starts degrading as $\phi_e^{max}$ becomes comparable to $\theta_{BW}$.
A solution to mitigate the effect of large $\phi_e^{max}$ is to form multiple simultaneous beams and then identify among them the best combining vector.
HBF or DBF are attractive options to form multiple simultaneous beams\footnote{For details regarding HBF and DBF, see \cite{Alkhateeb14_ChanEst} and \cite{Litva:1996:DBW:547927}, respectively.}, but at the cost of higher power consumption.
Compared to a fully Digital architecture (DBF), HBF is preferred for mmWave communication due to its lower power consumption because of its reduced number of RF chains.
However, the required number of ADCs in HBF, which are considered as the main power hungry blocks in mmWave receiver design, increases proportionally to the number of RF chains.
Therefore, in comparison to ABF, HBF and DBF may result in significantly higher power consumption\footnote{However, a more detailed power consumption analysis that takes into account all components in addition to the ADCs reveals that there are interesting tradeoffs, and there exist regimes in which DBF may actually be a more convenient choice than HBF and in some cases even almost as good as ABF \cite{My_PCComp_EW16}.}.
To jointly address the issue of higher power consumption and large $\phi_e^{max}$, we propose a phase shifters network (PSN) architecture.
The architecture of PSN is similar to ABF but instead of forming a single beam with one combining vector, PSN allows the formation of multiple beams simultaneously by using multiple combining vectors.
The idea is to form multiple beams and identify the best combining vector which corresponds to the desired AoA all in the analog domain.
Therefore, PSN provides a power efficient design with only two ADCs (for the inphase and the quadrature phase signals), whereas in HBF the number of ADCs is directly proportional to the number of RF chains.
The PSN receiver architecture is shown in Figure \ref{fig:PSN}, where $N_C$ is the number of combiners.
The number of simultaneous beams is equal to $N_C$, where phase shifters connected to a particular combiner represent a unique receiver combining vector.
The received signal is combined with $N_C$ combining vectors and the output of each combiner is compared using a comparator and with the help of a switch only the output of the best combiner (the one with the strongest signal) is forwarded to the ADC for further digital processing.
The architecture of PSN is also similar to HBF, with the exception that it consists of a single RF chain and allows beamforming only in the analog domain.
Therefore, in comparison to HBF and DBF, PSN does not provide the advantages of digital beamforming.
However, during cell search the inherent advantages of digital beamforming like multiplexing, interference cancellation or multiuser communication are not required at the MS\footnote{ Note that, in order to get the highest energy efficiency in the first phase of initial access while enjoying the advantages of digital beamforming in subsequent phases and during data communication, the PSN architecture can be easily converted to HBF by switching the output of each combiner to a separate RF chain rather than to the comparator. A detailed study of such hybrid architecture is part of our future work.}.
Therefore, PSN, thanks to its lower power consumption receiver design, is a viable option for mmWave initial cell search.
To form multiple beams, the selection of the main combining vector for PSN follows a similar formulation as in Equation (\ref{eq:CI_dist}).
However, in contrast to ABF where only a single combining vector is required, to form multiple beams in PSN $N_C$ different combining vectors from codebook $\textbf{W}_{MS}$ must be used.
Among the $N_C$ different combining vector, the main combining vector is selected according to (\ref{eq:CI_dist}), while the other $N_C-1$ combining vectors are selected from $\textbf{W}_{MS}$ such that they have minimum angular distance from the main combining vector.
According to the formation of the combining matrix $\textbf{W}_{MS}$, this results in a selection of $N_C-1$ combining vectors that are adjacent to the main combining vector.
This is a suitable choice as $\textbf{W}_{MS}$ is generated in a way to ensure that the gain fluctuation among any two adjacent combining vectors is within 1 dB \cite{ABF_Codebook_Wang09}.
\section{Simulation Setup and Results}
\label{sec:Sim_Res}
In our performance evaluation, we assume that the CI of the mmWave BS is
already available at the MS and with this location information the MS can point its beam in the desired direction.
We also assume that the mmWave BS always transmits the signal in different directions sequentially using ABF,
while for a better comparison different beamforming schemes are considered at the MS.
A more comprehensive comparison considering different beamforming options also at the BS is left for future study.
In our simulations, we consider a single path LOS scenario $(L = 1)$, Rician fading with $k = 10$, carrier frequency of $28$ GHz, transmit power of $30$ dBm, noise figure of $5$ dB, thermal noise power of $-174$ dBm/Hz, path-loss exponent of $2.2$,
and 64 transmit antennas at the BS.
The simulation results are averaged over $10^5$ different channel realizations for each Tx-Rx distance.
We evaluate the performance in terms of access error probability $P_{AccEr}$, which is defined as the probability that the $SNR$ of the received signal at the MS (Eq. (\ref{eq:SNR})) is below a certain threshold $\Theta$
\begin{equation}
P_{AccEr} = \Pr(SNR < \Theta)
\end{equation}
In our simulations this threshold $\Theta$ is set to $-4$ dB \cite{Alkhateeb_IBF}
%
%
%
%
\begin{figure}
\centering
\psfrag{Access Error Probability }{\ \ \ \ \ \ \ \ $P_{AccEr}$}
\psfrag{Tx-Rx Distance }{\!\!\!\!\!\ $Tx-Rx Distance$}
\psfrag{Nbs = 64}{\!\!\!\!\!\back\!\!\!\!\! $N_{BS} = 64, \phi_e = 0^\circ$}
\scalebox{1}{\includegraphics[width=\columnwidth]{ABF_CI_vs_noCI_up2}}\vspace{-4mm}
\caption{\protect\renewcommand{\baselinestretch}{1.25}\footnotesize A comparison of Random Search, Exhaustive Search, and CI based schemes. }\vspace{-4mm}
\label{fig:ABF_CI_vs_noCI}
\end{figure}
We start with a comparison of our proposed CI (with ABF) based initial cell search scheme with two non CI (with ABF) approaches, namely Random search (RS) \cite{Barati_IBF} and Exhaustive search (ES) \cite{Alkhateeb_IBF}.
Figure \ref{fig:ABF_CI_vs_noCI} shows how $P_{AccEr}$ for CI and non CI based approaches varies for different values of Tx-Rx Distance and $N_{MS}$.
Results show that RS performs worst in comparison to the other two schemes.
This is because in RS a MS forms its beam in a randomly selected direction, and therefore it is less likely to form a beam in the required direction, which results in a higher initial access error.
Also in RS, as $N_{MS}$ increases, the beamwidth gets narrower and the MS has to form more beams to cover the complete angular space, and therefore the probability of finding the right combining beam decreases, which results in an increase in $P_{AccEr}$.
On the other hand ES, where a MS searches the whole angular space sequentially, and our proposed CI based initial cell search approach show similar performance.
However, considering ABF both at the BS and the MS, the exhaustive approach has a large search delay because $card(\textbf{W}_{BS}) \times card(\textbf{W}_{MS})$ joint angular directions need to be considered, whereas for our CI based approach this search delay is reduced to $card(\textbf{W}_{BS}) \times 1$, at the cost of obtaining CI.
For the rest of this section, the performance evaluation is only focused on the CI based approach.
We now discuss how the CI based approach performs in the presence of the angular error in the available CI, and also how this angular error is related to $N_{MS}$.
In Figure \ref{fig:ABFNms4vs16}, $P_{AccEr}$ is plotted for different values of $\phi_e^{max}$, while considering 4 and 16 antennas at the MS, respectively.
As expected, for $\phi_e^{max} = 0$, with an increase in $N_{MS}$ the beamforming gain increases, and therefore $P_{AccEr}$ decreases.
For both $N_{MS} = 4$ and $N_{MS} = 16$, $P_{AccEr}$ increases with an increase in the Tx-Rx distance or the angular error $\phi_e$.
The increase in $P_{AccEr}$ with an increase in $\phi_e^{max}$ for $N_{MS} = 4$ is not very significant due to its wider beamwidth ($\theta^4_{BW} = 25.7^\circ $).
However, in the case of $N_{MS} = 16$,
the beamwidth decreases ($\theta^{16}_{BW} = 6.38^\circ $), and
therefore with an increase in the angular error it is more likely that the incoming signal will not fall within $\theta_{BW}$, and this results in a significant increase in $P_{AccEr}$.
\begin{figure}[t]
\psfrag{Access Error Probability }{\ \ \ \ \ \ \ \ \ $P_{AccEr}$}
\psfrag{AnEr = 0A}{\scriptsize $\phi_e^{max} = 0^{\circ}$, $N_{MS} = 4$}
\psfrag{AnEr = 5A}{\scriptsize $\phi_e^{max} = 5^{\circ}$, $N_{MS} = 4$}
\psfrag{AnEr = 10A}{\scriptsize $\phi_e^{max} = 10^{\circ}$, $N_{MS} = 4$}
\psfrag{AnEr = 0P}{\scriptsize $\phi_e^{max} = 0^{\circ}$, $N_{MS} = 16$}
\psfrag{AnEr = 5P}{\scriptsize $\phi_e^{max} = 5^{\circ} $, $N_{MS} = 16$}
\psfrag{AnEr = 10P p p p p p p p p p p p p}{\scriptsize $\phi_e^{max} = 10^{\circ}$, $N_{MS} = 16$}
\psfrag{AnEr = 10 PSN Nc3}{\scriptsize $\phi_e^{max} = 10^{\circ}, N_{MS} = 16 $}
\scalebox{1}{\includegraphics[width=\columnwidth]{ABFCINms4vs16_up}}\vspace{-4mm}
\caption{\protect\renewcommand{\baselinestretch}{1.25}\footnotesize Access error probability vs Tx-Rx distance for ABF with CI for different values of $N_{MS}$ and with $\phi_e^{max}$.}\vspace{-4mm}
\label{fig:ABFNms4vs16}%
\end{figure}
Figure \ref{fig:ABFNms4vs16} also highlights how the angular error is related to the optimal number of receive antennas.
It is observed that when the angular error starts falling outside of $\theta_{BW}$ of the estimated beam, $P_{AccEr}$ also increases.
Therefore, an increase in $N_{MS}$ makes it more susceptible to angular errors and hence results in even worse performance.
It is obvious from Figure \ref{fig:ABFNms4vs16} that the optimal number of antennas varies with $\phi_e^{max}$.
For instance, with $\phi_e^{max} = 0^\circ$, $N_{MS} = 16$ is the better option, whereas for $\phi_e^{max} = 10^\circ$, $N_{MS} = 4$ performs better than $N_{MS} = 16$.
Moreover, for $\phi_e^{max} = 5^\circ$, $N_{MS} = 16$ has a lower $P_{AccEr}$ for Tx-Rx distance 50 m and above, whereas below 50 m $N_{MS} = 4$ is a preferable choice.
This shows a tradeoff between distance dependent path-loss and $\phi_e$.
At short Tx-Rx distance, $\phi_e$ has more significance on $P_{AccEr}$ and therefore $N_{MS} = 4$ performs better, however as the Tx-Rx distance increases the path-loss also increases and therefore $N_{MS} = 16$ with higher beamforming gain results in a lower $P_{AccEr}$.
We now compare the performance of PSN with ABF and show how PSN can mitigate the effect of $\phi_e$.
For PSN, we consider $N_C = 3$.
Figure \ref{fig:ABFvsPSN_Nms16} presents a comparison of ABF and PSN for different angular errors.
For both schemes, $P_{AccEr}$ increases with an increase in $\phi_e^{max}$ or Tx-Rx distance.
However, PSN performs better than ABF, and is less prone to a small increase in $\phi_e^{max}$.
For instance, the performance of PSN is unaffected for $\phi_e^{max} = 5^\circ$, whereas there is an increase in $P_{AccEr}$ for $\phi_e^{max} = 10^\circ$.
This is because the PSN architecture allows to form multiple simultaneous beams and is able to capture the signal energy even in the presence of angular errors.
The performance of PSN in case of angular error improves with $N_C$, as a
larger $N_C$ will result in more combiner beams covering more angular space and thus is less susceptible to angular errors.
As PSN outperforms ABF in case of erroneous CI, we now further evaluate its performance by comparing it with HBF and DBF. In simulations, for PSN and HBF $N_C = N_{RF} = 3$.
\begin{figure}
\centering
\psfrag{Access Error Probability }{\ \ \ \ \ \ \ \ \ $P_{AccEr}$}
\psfrag{Angular Error = 0 ABF}{\scriptsize$\phi_e^{max} = 0^\circ$ ABF}
\psfrag{Angular Error = 5 ABF}{\scriptsize$\phi_e^{max} = 5^\circ$ ABF}
\psfrag{Angular Error = 10 ABF}{\scriptsize$\phi_e^{max} = 10^\circ$ ABF}
\psfrag{Angular Error = 0 PSN Nc = 3}{\scriptsize$\phi_e^{max} = 0^\circ$ PSN}
\psfrag{Angular Error = 5 PSN Nc = 3}{\scriptsize$\phi_e^{max} = 5^\circ$ PSN}
\psfrag{Angular Error = 10 PSN Nc = 3}{\scriptsize$\phi_e^{max} = 10^\circ$ PSN}
\scalebox{1}{\includegraphics[width=\columnwidth]{ABFvsPSN}}\vspace{-4mm}
\caption{\protect\renewcommand{\baselinestretch}{1.25}\footnotesize Access error probability vs Tx-Rx distance for ABF and PSN with $N_C = 3$. }\vspace{-1mm}
\label{fig:ABFvsPSN_Nms16}
\end{figure}
Figure \ref{fig:DBFvsHBFvsPSN} presents a comparison of DBF and HBF with PSN.
It is depicted that DBF shows a stable performance with an increase in the angular error.
This is because DBF allows to form multiple beams which allows the MS to look in all angular directions simultaneously, and therefore the performance of DBF is not effected by angular errors (but at the cost of higher power consumption).
Note that the performance of PSN and HBF for small angular errors is similar to that of DBF.
\begin{figure}
\centering
\psfrag{Access Error Probability}{\ \ \ \ \ \ \ \ \ $P_{AccEr}$}
\psfrag{Angular Err = +/-0, PSN}{\scriptsize$\phi_e^{max} = 0^\circ$, PSN}
\psfrag{Angular Err = +/-10, PSN}{\scriptsize$\phi_e^{max} = 10^\circ$, PSN}
\psfrag{Angular Err = +/-0, DBF}{\scriptsize$\phi_e^{max} = 0^\circ$, DBF}
\psfrag{Angular Err = +/-10, DBF}{\scriptsize$\phi_e^{max} = 10^\circ$, DBF}
\psfrag{Angular Err = +/-0, HBF}{\scriptsize$\phi_e^{max} = 0^\circ$, HBF}
\psfrag{Angular Err = +/-10, HBF}{\scriptsize$\phi_e^{max} = 10^\circ$, HBF}
\scalebox{1}{\includegraphics[width=\columnwidth]{DBFvsHBFvsPSN}}\vspace{-4mm}
\caption{\protect\renewcommand{\baselinestretch}{1.25}\footnotesize Access error probability vs Tx-Rx distance for PSN, HBF and DBF.}\vspace{-5mm}
\label{fig:DBFvsHBFvsPSN}
\end{figure}
Moreover, for $\phi_e^{max} = 0^{\circ}$, HBF results in a slightly lower $P_{AccEr}$ than PSN due its additional digital baseband combiners (which further minimizes the distance between $\phi_{CI}$ and $\vartheta_{i}$), but the performance difference between PSN and HBF diminishes with an increase in $\phi_e^{max}$.
Also, note that the digital baseband combiner in HBF requires additional RF chains, which results in higher power consumption.
Therefore, PSN is a viable solution as it provides significant power savings at the cost of only a small performance degradation at lower $\phi_e^{max}$.
In order to make the above considerations about power consumption of the various beamforming schemes more precise, we study how the total receiver power consumption $P_{Tot}$, varies with an increase in the number of ADC bits for ABF, DBF, HBF and PSN.
In our evaluation, we consider $N_C = N_{RF} = 3$ and $N_{MS} = 16$.
$P_{Tot}$ for ABF and DBF can be calculated similar to \cite{OrhanER15_PowerCons}\footnote{For a detailed power consumption comparison among ABF, DBF and HBF, the reader is referred to \cite{My_PCComp_EW16}.}
\begin{equation}
P_{Tot}^{ABF} = N_{MS}(P_{LNA} + P_{PS}) + P_C + P_{RF} + 2P_{ADC}
\label{eq:ABF}
\end{equation}
\begin{equation}
P_{Tot}^{DBF} = N_{MS}(P_{LNA} + P_{RF} + 2P_{ADC})
\label{eq:DBF}
\end{equation}
and similarly $P_{Tot}$ for HBF \cite{Alkhateeb15_SwvsPS} and PSN can be calculated as
\begin{equation}
\begin{split}
P_{Tot}^{HBF} = &\ N_{MS}(P_{LNA} + P_{SP} + N_{RF}P_{PS}) \\ & + N_{RF}(P_C + P_{RF} + 2P_{ADC})
\end{split}
\label{eq:HBF}
\end{equation}
\begin{equation}
\begin{split}
P_{Tot}^{PSN} &=\ N_{MS}(P_{LNA} + P_{SP} + N_{C}P_{PS}) \\
& + N_{C}P_C + P_{RF} + P_{Comp} + P_{Sw} + 2P_{ADC}
\end{split}
\label{eq:PSN}
\end{equation}
where $P_{RF} = P_{M} + P_{LO} + P_{LPF} + P_{BB_{amp}}$ represents the power consumption of the RF chain, and
where $P_{LNA}$, $P_{PS}$, $P_C$, $P_M$, $P_{LO}$, $P_{LPF}$, $P_{BB_{amp}}$, $P_{ADC}$, $P_ {SP}$, $P_{Sw}$ and $P_{Comp}$ represent the power consumption of low noise amplifier (LNA), phase shifter, combiner, mixer, local oscillator, low pass filter, baseband amplifier, ADC, splitter, switch and comparator, respectively.
The power consumption of an ADC scales exponentially with the number of bits and linearly with the sampling rate \cite{ADC_b_B}. Therefore, considering Nyquist sampling rate, the ADC power consumption is modeled as
\begin{equation}
P_{ADC} = cB2^b
\label{eq:Padc}
\end{equation}
where $b$ is the number of ADC bits, $B$ is the sampling rate (Bandwidth) and $c$ is the energy consumption per conversion step.
For $B = 500$ MHz and $b = 5$, the power consumption of an ADC in \cite{Alkhateeb15_SwvsPS} is considered as 200 mW, and this correponds to $c = 12.5$ pJ.
For details regarding the power consumption of other components, the reader is referred to \cite{Alkhateeb15_SwvsPS}.
\begin{figure}
\centering
\psfrag{Total Power Consumption (Watts)}{\ \ \ \ \ \ \ \ \ \ $P_{Tot}\ (Watts)$}
\psfrag{Number of ADC bits}{\ \ \ \ \ \ \ \ \ \ $b$}
\scalebox{1}{\includegraphics[width=\columnwidth]{PowerCons_Nms16_ADC200_up2}}\vspace{-4mm}
\caption{\protect\renewcommand{\baselinestretch}{1.25}\footnotesize Power consumption comparison of ABF, DBF, HBF and PSN.}
\label{fig:PowerCons_Nms16}
\end{figure}
Figure \ref{fig:PowerCons_Nms16} shows that with an increase of $b$ there is an exponential increase in power consumption.
As expected, $P_{Tot}$ for ABF is always lower than for the other beamforming schemes, and hence ABF with CI is a suitable option for initial access considering small angular errors.
On the other hand, DBF always has higher power consumption than other beamforming schemes except for $b = 1$.
Finally, PSN consumes lower power than HBF irrespective of the number of ADC bits.
Moreover, the $P_{Tot}$ difference for PSN and HBF increases with an increase in $b$.
Therefore, PSN with $P_{Tot}$ less than DBF and HBF is a viable option even for high resolution (i.e., higher number of bits) ADC based mmWave receiver design.
\section{Conclusions}
\label{sec:Con}
In this work, we discussed how the availability of CI at the MS can reduce the inherent search delay of ABF.
However, the performance of ABF starts degrading with an increase of the angular error in the available CI.
We also showed that the optimal number of receiver antennas is related to the angular error.
In addition, considering small angular errors in CI, ABF with lower power consumption is the best option for initial access.
Moreover, we presented an analog beamforming based PSN architecture with a single RF chain to mitigate the effect of the angular error.
Simulation results validate that this solution has equivalent performance to HBF, while exhibiting lower power consumption.
This makes PSN a viable approach for initial cell search in mmWave 5G cellular networks.
In the future, we will extend this work to multipath scenarios, and study how different beamforming schemes perform in the presence of multiple transmitting BS.
We will also evaluate the optimal number of receiver antennas based on the statistics of the angular error.
Moreover, we will analyze the performance of PSN with respect to HBF and DBF without the availability of context information.
Finally, we will study the choice of an appropriate beamforming scheme which jointly minimizes the access error probability and the energy consumption.
\bibliographystyle{IEEEtran}
|
\section{Introduction}\label{sec:introduction}
\noindent In this note, we consider the Brownian ``spider,'' a process also known as the ``Walsh'' Brownian motion, due to \cite{Pitman} and \cite{Dubins}. The Brownian spider is constructed as a set of $n \ge 1$ half-lines, or ``ribs,'' meeting at a common point, $O$. A Brownian motion
on a spider starting at zero may be constructed from a standard reflecting
Brownian motion ($|W_t|,t \geq 0$) by assigning an integer
$i \in \{1,\ldots,n\}$ uniformly and independently to each excursion
which is then transferred to an excursion on rib $i$ (here, $i$ should be interpreted as the index of the rib on which the next excursion occurs). It is helpful to think about the Brownian spider as a bivariate process; the first coordinate of the process is reflecting Brownian motion and the second coordinate of the process is the rib index. Formally, we define the Brownian spider process $Z_t$ as
\begin{equation}
Z_t= \parens{|W_t|, R_t}, t \geq 0
\end{equation}
where $|W_t|$ is reflected Brownian motion and $R_t$ is the rib on which the process is located at time $t$. $|W_t|$ can be decomposed into excursions away from 0 with endpoints $t_k$ s.t. $|W_{t_k}| = 0$. $R_t$ is constant between $t_k$ and $t_{k+1}$ for all $i$, and $R_t = i$ means the excursion occurs on the rib $i$.
We define the supremum of reflected Brownian motion on each rib as
\vspace{-0.25cm}\begin{eqnarray*}
S_i(t)= \sup_{\{t:\,\, R_t=i\}}|W_t|, \,\, t \geq 0, \, i=1,...,n.
\end{eqnarray*}
\indent Below is a sample path realization of the Brownian spider for $n=3$. We use $W_i(t)$ to denote the process on the rib $i$:
\begin{equation}
W_i(t) = \left\{ \begin{array}{cc}
|W_t| & \text{ if } R_t = i \\
0 & \text{ if } R_t \neq i.
\end{array}
\right.
\end{equation}
\begin{figure}[H]
\centering
\includegraphics[width=0.5\linewidth]{figure2.jpg}
\caption{A sample path realization of the Brownian spider for $n=3$.}
\label{fig:marginals}
\end{figure}
\indent In an attempt to understand the unboundedness of Brownian motion on the spider up to time $t$, a natural question to ask is: what is $\expe{\sum_{i=1}^nS_i(t)}$?
However, Lester Dubins (personal communication with Larry Shepp) asked a different question. Dubins wished to design a \textit{stopping time} to maximize the coverage of Brownian motion on the spider for a given expected time. That is, he wished to find
\vspace{-0.25cm}\begin{eqnarray}
C_n := \sup_{\{\tau: \,\, \expe{\tau} =1\}} \expe{S_1(\tau)+\ldots+S_n(\tau)},
\end{eqnarray}
\noindent
where the supremum is calculated over all stopping times of mean one. Equivalently, Dubins wished to calculate the smallest $C = C_n$ such that for every stopping time $\tau$ the
following inequality holds
\vspace{-0.25cm}\begin{eqnarray} \label{eq2}
\expe{S_1(\tau)+S_2(\tau)+\ldots+S_n(\tau)} \le C_n \sqrt{\expe{\tau}}.
\end{eqnarray}
(note that for any stopping time $\tau$, $\expe{S_i(\tau)}$ scales with with $\sqrt{\tau}$). The left side of equation (\ref{eq2}) is the mean total
measure of space visited on the spider up to time $\tau$.
\indent For $n = 0$, we,
somewhat inconsistently, define $C_0$ in a similar way for ordinary Brownian
motion without a reflecting barrier at zero. We seek the smallest
constant $C_0$ for which the {\em one-sided maximum} satisfies
\vspace{-0.25cm}\begin{eqnarray}
\expe{\max_{\{0 \le t \le \tau\}} W(\tau)} \le C_0 \sqrt{\expe{\tau}}.
\end{eqnarray}
In this note, we will prove that the optimal bounds $C_n = \sqrt{n+1}$, for $n = 0,1,2$. Without further delay, the author notes that the \textit{solution} of the optimal bounds for $n=0,1,2$ is \textit{not new}. The cases $n=0,1$ were solved by \cite{Dubins}, and the case $n=1$ was also independently solved by \cite{Gilat} by a different method. The $n=2$ case was recently resolved by \cite{Dubins2}. What \textit{is new}, however, is the dynamic programming \textit{strategy} the author employs to find the bounds $C_n$ for $n=0,1,2$, which he believes to be the most tractable approach for solving for $C_n$ for all $n$ (despite much effort by many researchers, this problem remains open). The behavior of $C_n$ for large $n$ is interesting because when $n = \infty$, the total measure of space visited on the spider up to time $t > 0$ is also infinite. This is because it is the total variation of a Brownian motion on $[0,t]$ because at each return to the node, a fresh rib is chosen.
Larry Shepp saw dynamic programming to be the root of all optimal control problems. In general, there are two strategies that can be used to solve a dynamic programming problem.
(A) Guess a candidate for an optimal strategy, calculate the reward function for the strategy, then prove its excessiveness.
(B) Guess the optimal reward function and establish its optimality by proving its excessiveness.
Unlike \cite{Dubins2}, which employs strategy (A), our approach is that of (B), and to the best of our knowledge, we are the first to do so. In stochastic optimization, strategy (B) reduces to ``guessing'' the right optimal control function. If one can guess the right function, the supermartingale becomes a martingale, and It\^o calculus takes care of the rest. This approach appears prominently throughout Shepp's most seminal works, specifically on p.634 of \cite{Shepp2}, on p.207 of \cite{Shepp3}, p.1528 of \cite{Shepp4}, on p.335 of \cite{Ernst}, and most recently, on p.422 of \cite{Ernst2}.
The organization of this note is as follows: In Section 2, we formalize our guess for the optimal reward function. In Section 3, we establish the optimality of this function by proving its excessiveness, albeit only in the cases $n=0,1,2$. We conclude by arguing the viability of our strategy towards a solution of the general problem.
\section{Our guess of the optimal reward function}\label{sec:formal_model_approach}
Let $r = R_0$ be the index of the starting rib, $x$ be a fixed distance along the rib $r$, and $C$ and $M$ finite constants. In order to obtain the least upper bound $C_n$, we must solve
a more general optimal stopping problem. Let $s_1, s_2,...,s_n \geq 0$ be the distances that have already been covered on each of the respective ribs at time $0$. For {\it every value} of
$C > 0$, and \textit{every} choice of $r$ and $x$ such that $x \le s_r$, and $s_1,\ldots,s_n$, we must find the value of
\vspace{-0.25cm}\begin{eqnarray} \label{quantity}
V(x,r;s_1,\ldots,s_n;C) := \sup_{\{\tau: \,\, \expe{\tau} \leq M\}} \expesub{\{x,r,s_1,\ldots,s_n\}}{S_1(\tau)+\ldots + S_n(\tau) - C\tau}.
\end{eqnarray}
The subscript of the expectation, ${\{x,r,s_1,\ldots,s_n\}}$, denotes that the process is currently at a distance $x$ on rib $r$ at time $0$. By abuse of notation, $S_i(\tau)$ denotes the furthest point covered on rib $i$ up to time $\tau$. Note that we must find $V(x,r;s_1,\ldots,s_n;C)$ not only for $x=0$ and $s_1 = \ldots = s_n = 0$, but for
{\it every point} of the spider at $x$ on {\it every} rib $r$ as initial
point, and every starting position for $s_i, i=1,\ldots,n$.
In (\ref{quantity}), the supremum is taken over bounded stopping times $\tau$. Even though we
only need the case when the initial point is $O$ and when $s_i = 0,\,\,i=1,\ldots,n$, standard
martingale methods of solving optimal stopping problems do not work unless we
can find the formula $V$ for every starting position (see, for example: ~\cite{Brumelle},~\cite{Dynkin}, ~\cite{Shepp}, ~\cite{Talagrand}, and ~\cite{Walrand}).
We ``guess'' that ${\hat{V}}(x,r,s_1,\ldots,s_n,C)$ should have the following properties:
\medskip
(a) ${\hat{V}}(0,r,s_1,\ldots,s_n,C)$ does not depend on $r$ (if $x_i=0, r$ becomes irrelevant).
\medskip
(b) $\frac{d}{dx}{\hat{V}}(x,r,s_1,\ldots,s_n,C) = 0$ at $x=0$ $\forall r$.
\medskip
(c) $\frac{d}{ds_r}{\hat{V}}(x,r,s_1,\ldots,s_n,C) = 0$ at $x = s_r\,\, \forall \,r$.
\medskip
(d) $\frac{1}{2} \frac{d^2}{dx^2}{\hat{V}}(x,r,s_1,\ldots,s_n,C) \leq C, \, 0 \le x \le s_r,\,r=1,\ldots,n$.
\medskip
(e) ${\hat{V}}(x,r,s_1,\ldots,s_n,C) \ge s_1+\ldots+s_n$, for all $0 \le x \le s_r$ and $r = 1,\ldots,n$.
\medskip
(f) If strict inequality holds in property (e), $\frac{1}{2} \frac{d^2}{dx^2}{\hat{V}}(x,r,s_1,\ldots,s_n,C) = -C$, $0 \le x \le s_r$.\\
\noindent Intuitively, at a stopping place, we are far from any boundary point where an
$s$ would increase and thus we are willing to accept the reward
${\hat{V}}(x,r,s_1,\ldots,s_n,C) = s_1+\ldots+s_n$.
\section{Establishing the optimality of the reward function}\label{sec:formal_model_approach}
\begin{theorem}
If we have a function ${\hat{V}}$ satisfying properties (a)-(f) in Section 2, then
\vspace{-0.25cm}\begin{eqnarray}
V(x,r,s_1,\dots,s_n;C) \equiv {\hat{V}}(x,r,s_1,\dots,s_n,C).
\end{eqnarray}
\end{theorem}
\begin{proof}
Consider the process
\vspace{-0.25cm}\begin{eqnarray}
Y(t) = {\hat{V}}\parens{Z_t,{\bf S}(t),C} - Ct, \,\,\,t \ge 0
\end{eqnarray}
where $\mathbf{S}(t) = \parens{S_1(t),...,S_n(t)}$. $Y(t)$ is a continuous local supermartingale at
$x=0$ by properties (a) and (b), at $x = s_r$ by property (c), and at any $x$ by property (d). For any bounded stopping time $\tau$, it follows from the optional sampling theorem that $\expe{Y(\tau)} \leq Y(0)$.
Property (e) gives us that for any bounded $\tau$,
\vspace{-0.25cm}\begin{eqnarray} \label{coolone}
\expesub{\{x,r,s_1,\ldots,s_n\}}{S_1(\tau)+\ldots+S_n(\tau)- C \tau} \le {\hat{V}}\parens{x,r,s_1,\ldots,s_n,C}.
\end{eqnarray}
From the definition of $V$ in (\ref{quantity}),
\noindent
we must have $V \le \hat{V}$. \\
\indent We now consider the reverse inequality $V \ge {\hat{V}}$. By property (f), equality holds in the last argument for the ``right $\tau$.'' Although this ``right $\tau$'' does not always exist in such problems, it does for our problem; the ``right $\tau$'' is the first entry time of the underlying Markov process $(Z,S)$ in the set where equality holds in (e). Further, this ``right $\tau$'' is a particular stopping time that is ``approximable by uniformly bounded ones.'' Larry Shepp used the phrase ``approximable by uniformly bounded ones'' to denote that we can take the ``right $\tau$'' at which the equality is attained, approximate this ``right $\tau$'' with ``right $\tau$'' $\wedge\, \,n$ for $n \ge 1$, and then proceed to pass to the limit for $n$. This is valid in our setting since the ``right $\tau$'' has finite expectation. When property (f) holds, and when equality holds in (d), $Y$ will be a local martingale up to the first entry of the underlying Markov process $(Z,S)$ into the the set where equality holds in (e). Since the ``right $\tau$'' has a finite expectation, we may invoke the standard form of Doob's stopping theorem for bounded stopping times, as in \cite{Doob}. Thus,
\vspace{-0.25cm}\begin{eqnarray} \label{coolone}
\expesub{\{x,r,s_1,\ldots,s_n\}}{S_1(\tau)+\ldots+S_n(\tau)- C \tau} \geq {\hat{V}}\parens{x,r,s_1,\ldots,s_n,C}.
\end{eqnarray}
and one can optimize over $\tau$ on both sides. The reverse inequality $V \ge {\hat{V}}$ thus holds and thus
$V \equiv {\hat{V}}$, completing the proof.
\end{proof}
If we can find the right ${\hat{V}}$ satisfying properties (a)-(f), we then know that
\vspace{-0.25cm}\begin{eqnarray} \label{thetaeq}
A_n(C) := V\parens{O,r,0,\ldots,0;C} = \frac{\theta_n}{C} \label{eqn:first},
\end{eqnarray}
\noindent
where $\theta_n$ is a number independent of $C$. $V$ must be of the form
$\frac{\theta_n}{C}$ because a scaling argument allows us to reduce the
problem to any one value of $C$. This is because we will show that
\vspace{-0.25cm}\begin{eqnarray}
V\parens{x,r,s_1,\ldots,s_n;C} = \frac{1}{C}V\parens{Cx,r,Cs_1,\ldots,Cs_n;1}. \label{eqn:C}
\end{eqnarray}
Note that if we start at $x = O$ and $s_1 = \ldots = 0$ then above form for
$A_n(C)$ is obtained.
Let
\vspace{-0.25cm}\begin{eqnarray*}
S(\tau)\triangleq S_1(\tau)+...+S_n(\tau).
\end{eqnarray*} For any $C$ and any $\tau$, $\expe{S(\tau)} \le A_n(C)+C\expe{\tau}$. If we specify $m = \expe{\tau}$ for any fixed stopping time $\tau$, then
we will obtain the best upper bound by minimizing over $C$, which is
\vspace{-0.25cm}\begin{eqnarray*}
\expe{S(\tau)} \le \inf\parens{\frac{\theta}{C} + C m}.
\end{eqnarray*}
The infimum is attained at $C = \sqrt{\frac{\theta_n}{m}}$ and gives the bound
$C_n = 2\sqrt{\theta_n}$ for any $n$. Thus we need only find $V(O;C)$ for
any one value of $C$.
\subsection{Solution for n=0,1,2}
\begin{corollary}
$C_0=1$.
\end{corollary}
\begin{proof}
For $n=0$, consider the function
\vspace{-0.25cm}\begin{eqnarray*}
{\hat{V}}\parens{x,s,C} = C\parens{\parens{x-s+\frac{1}{2C}}^+}^2 + s.
\end{eqnarray*}
\noindent
We note that properties (a)-(f) hold, and so for $x = s = 0$, and for any $C > 0$
\vspace{-0.25cm}\begin{eqnarray}
\expe{S_\tau} \le C\expe{\tau} + \frac{1}{4C}.
\end{eqnarray}
\noindent
Minimizing over $C$, i.e., taking $C = \frac{1}{2\sqrt{\expe{\tau}} }$, as above
for any $\tau$, we obtain the inequality
\vspace{-0.25cm}\begin{eqnarray}
\expe{S(\tau)} \le \sqrt{\expe{\tau}}
\end{eqnarray}
for all
$\tau$, i.e., $C_0 = 1$.
\end{proof}
\begin{corollary}
$C_1=\sqrt{2}$.
\end{corollary}
\begin{proof}
For $n=1$, the right ${\hat{V}}$ is given by:
\vspace{-0.25cm}\begin{eqnarray}
{\hat{V}}\parens{x,s,C} = Cx^2+ \frac{1}{2C}, \quad 0 \le x \le s \le \frac{1}{2C};
\end{eqnarray}
\vspace{-0.25cm}\begin{eqnarray*}
{\hat{V}}\parens{x,s,C} = C\parens{\parens{x-s+\frac{1}{2C}}^+}^2 +s, \quad 0 \le x \le s, \,\,s > \frac{1}{2C}.
\end{eqnarray*}
We use the above argument to see that $A_2(C) = \frac{1}{2C}$
and $\theta_1 = \frac{1}{2}$ and so $C_1 = \sqrt{2}$.
\end{proof}
\begin{corollary}
$C_2=\sqrt{3}$.
\end{corollary}
\begin{proof}
For $n=2$, ${\hat{V}}$ is, for $i \ne j$ \textit{and} $s_1+s_2 \le \frac{1}{C}$,
\vspace{-0.25cm}\begin{eqnarray} \label{difficulteqn}
{\hat{V}}(x,r,s_1,s_2,C) = Cx^2 - Cx(s_i-s_j) + \frac{C(s_1^2+s_2^2)}{2} + \frac{3}{4C}, \quad 0 \le x \le s_i.
\end{eqnarray}
We further simplify equation (\ref{difficulteqn}) as follows:
\small
\vspace{-0.25cm}\begin{eqnarray}
{\hat{V}}(x,r,s_1,s_2,C) = C\parens{\parens{x -\parens{s_i-\frac{1}{2C}}}^+}^2 + C\parens{\parens{-x -\parens{s_j-\frac{1}{2C}}}^+}^2 + s_1+s_2,
\end{eqnarray}
\normalsize
where $0 \le x \le s_i, s_1+s_2 \ge \frac{1}{C}$. We can use the above argument to see that with $V(O;C) = \frac{3}{4C}$
we arrive at $C_2 = \sqrt{3}$.
\end{proof}
\section{$n=3$ and beyond}\label{sec:formal_statement_results}
At present, we possess a non-trivial but ultimately incomplete strategy for addressing the case $n=3$. Our strategy is to develop the ``correct'' nonlinear Fredholm equation in order that we may reduce the problem to that of a nonlinear integral recurrence. Based on simulation approaches, we \textit{conjecture} the following about the constant:
\begin{conjecture}
The $\sqrt{n+1}$ pattern for the spider constant does not hold for $n=3$.
\end{conjecture}
\noindent Further, it is likely that the spider constant for $n=3$ is not an elementary number.
\section{Final Remarks}We are hopeful of a solution to the general $n$ case for the Dubins spider and maintain that our proposed dynamic programming approach constitutes the most tractable direction for solving the problem, for the following reasons: 1) The use of linear programming would be infeasible because the approximate linear programming would be large and unwieldy, making accurate numerics impossible. 2) Bellman's dynamic programming method seems intractable for the same reason as that of using linear programming. 3) The more standard method of dynamic programming, namely that of guessing a candidate for an optimal strategy, calculating the reward function of the strategy, and proving its excessiveness (as most recently done by \cite{Dubins2}) was unsuccessful in obtaining the general solution.
\section*{Acknowledgments}
First and foremost, I am indebted to my mentor, Professor Larry Shepp, for his extraordinary support, for introducing me to this literature, and for his enormously insightful conversations about this problem. I am also indebted to my colleague Professor Goran Peskir for his excellent inspiration and insight, particularly regarding the proof of Theorem 3.1. I am grateful to Quan Zhou for his invaluable help in producing the figure in this note as well as for his careful reading of the manuscript. I thank Professor David Gilat and Professor Isaac Meilijson for their detailed input. I thank Professor Ton Dieker for his helpful comments. Finally, I am tremendously grateful to an anonymous referee whose very helpful comments enormously improved the quality of this work.
\bibliographystyle{plain}
|
\section{I. INTRODUCTION}
Low-dimensional electronic systems are of great interest in contemporary condensed-matter physics because of their susceptibility to charge density wave (CDW) instability~\cite{Carpinelli}, non-Fermi liquid behavior~\cite{Blumenstein}, spin ordering~\cite{Erwin,Li}, and superconductivity~\cite{Qin,TZhang} at low temperatures. Specifically, metal-atom adsorption on semiconductor surfaces provides a unique playground for the exploration of such exotic physical phenomena~\cite{Tejeda,Snijders}. We here focus on a prototypical example of quasi-one-dimensional (1D) systems, self-assembled indium (In) atom wires on the Si(111) surface~\cite{Bunk,Yeom1,Cho}. Each In wire consists of two zigzag chains of In atoms [see the left panel of Fig. 1(a)]~\cite{Bunk}. Below ${\sim}$120 K, this quasi-1D system undergoes a reversible phase transition initially from a 4${\times}$1 structure to a 4${\times}$2 one, then to an 8${\times}$2 structure~\cite{Yeom1,Kumpf}, showing a period doubling both parallel and perpendicular to the In wires. This (4${\times}$1)${\rightarrow}$(8${\times}$2) structural phase transition is accompanied by a metal-insulator (MI) transition~\cite{Yeom1,Ahn,Yeom2}. For the explanation of such a MI transition, the CDW mechanism due to a Peierls instability was initially proposed~\cite{Yeom1,Ahn,Yeom2,Park}, but subsequently other mechanisms based on an order-disorder transition~\cite{Gonzalez1,Gonzalez2} and many-body interactions~\cite{GLee} have been proposed. Despite such debates, the CDW mechanism has been most widely believed to drive the observed MI transition~\cite{Yeom1,Ahn,Yeom2,Park,Morikawa,SHUhm,DMOh}. It is noted that the CDW formation invokes the strong coupling between lattice vibrations and electrons near the Fermi level $E_F$, caused by Fermi surface nesting with a nesting vector 2$k_{F}$ = ${\pi}$/$a_{\rm x}$ ($a_{\rm x}$: the 4${\times}$1 lattice constant along the In wires)~\cite{Yeom1,Gruner}. The resulting Peierls dimerization was believed to occur on each chain, and the two dimerized chains further interact with each other, leading to a coupled double Peierls-dimerized chain model~\cite{Cheon} [see Fig. 1(b)].
\begin{figure}[ht]
\centering{ \includegraphics[width=7.0cm]{fig1.eps} }
\caption{(Color online) (a) Top view of the optimized 4${\times}$1 (left panel) and 4${\times}$2 (right panel) structures. The dark and gray circles represent In and Si atoms, respectively. For distinction, Si atoms below In chains are drawn with small circles. Each unit cell is indicated by the dashed line. The STM images of the 4${\times}$1 and 4${\times}$2 structures is displayed in (b), together with the overlap of the coupled double Peierls-dimerized chain model (from Ref.~\cite{Cheon}): Schematic of two dimerized phases, A (with a positive displacement ${\Delta}$) and B (with a negative ${\Delta}$), of a single Peierls chain is also given on the right side. }
\end{figure}
Regarding the phase transition of the In/Si(111) system, there are still some unsettled issues. Although it is well established that the structural model of the 8${\times}$2 phase is constituted by the basic building block of In ``hexagon" [see the right panel of Fig. 1(a)], the microscopic mechanism of the hexagon formation is so far unclear whether it is driven by the Peierls dimerization on the two chains~\cite{Yeom1,Cheon} or by a shear distortion where neighboring chains are displaced in opposite directions~\cite{Gonzalez1}. Recently, it was also reported that both the covalent bonding and van der Waals (vdW) interactions between the two chains play crucial roles in forming In hexagons~\cite{HJKim}. Moreover, the order of the phase transition has been controversial whether it belongs to first-order~\cite{Park2,Wall,Klasing} or second-order~\cite{Gonzalez1,Gonzalez2,Guo,GLee2}. According to the mean-field theory~\cite{Gruner}, the CDW or order-disorder mechanism can be classified as the second-order phase transition. However, existing scanning tunneling microscopy (STM) experiments have reached the conflicting conclusions between the first-order~\cite{Park2} and second-order~\cite{Guo,GLee2} transitions, whereas a recent high-resolution low-energy electron diffraction (HRLEED) study~\cite{Klasing} observed a robust hysteresis of diffraction spot intensities as the sample temperature slowly increases and decreases during the (4${\times}$1)${\leftrightarrow}$(8${\times}$2) phase transition. The latter observation~\cite{Klasing} obviously indicates the existence of energy barrier between the two phases, representing the first-order transition~\cite{Landau}. All of these controversies on the microscopic mechanism of the hexagon formation and the order of the phase transition reflect our incomplete understanding of the origin of the phase transition in the In/Si(111) system.
In this paper, using first-principles density-functional theory (DFT) calculations, we propose the atomistic picture for the phase transition of the In/Si(111) system to resolve the existing problems such as the microscopic mechanism of the hexagon formation, the order of the phase transition, as well as the origin of the MI transition. We find that the hexagon formation is driven by an exothermic reaction with the consecutive bond-breaking and bond-making processes, giving rise to an energy barrier as well as a gap opening. This atomistic picture not only reveals the first-order nature of the (4${\times}$1)${\leftrightarrow}$(8${\times}$2) phase transition, but also illustrates how the observed solitons~\cite{Zhang,THKim,Cheon} can be created at the boundary of two differently oriented hexagon structures. Our findings clarify that the MI transition of the In/Si(111) system is attributed to the ${\times}$2 periodic lattice reconstruction of In hexagons via the bond breakage and the new bond formation, rather than the CDW formation arising from Fermi surface nesting~\cite{Yeom1,Ahn,Yeom2,Park,Cheon}.
To properly predict the energetics of the 4${\times}$1, 4${\times}$2, and 8${\times}$2 structures~\cite{HJKim,Zhang2,SWKim}, we have performed the van der Waals energy corrected~\cite{vdW} hybrid DFT calculations using the FHI-aims~\cite{Aims} code that gives an accurate, all-electron description based on numeric atom-centered orbitals, with ``tight" computational settings. For the exchange-correlation energy, we employed the hybrid functional of HSE06~\cite{HSE1,HSE2}. The ${\bf k}$-space integrations in various unit-cell calculations were done equivalently with 64 ${\bf k}$ points in the surface Brillouin zone of the 4${\times}$1 unit cell. The Si(111) substrate (with the Si lattice constant $a_0$ = 5.418 {\AA}) below the In chains was modeled by a 6-layer slab with ${\sim}$30 {\AA} of vacuum in between the slabs. Each Si atom in the bottom layer was passivated by one H atom. All atoms except the bottom layer were allowed to relax along the calculated forces until all the residual force components were less than 0.01 eV/{\AA}.
\begin{figure}[ht]
\centering{ \includegraphics[width=7.0cm]{fig2.eps} }
\caption{(Color online) (a) Schematics of four degenerate CDW ground phases (denoted as AA, BA, BB, and AB) in a coupled double Peierls-dimerized chain model (from Ref.~\cite{Cheon}). (b) Atomistic picture for the formation of the four degenerate hexagon structures. The dashed lines in (b) indicate the 4${\times}$2 unit cell. For comparison with the Peierls dimerization in (a), the filled ellipses (or half-ellipses) are drawn in the outer In atoms of the 4${\times}$2 structure in (b). }
\end{figure}
\begin{figure*}[ht]
\centering{ \includegraphics[width=17.0cm]{fig3.eps} }
\caption{(Color online) (a) Calculated energy profile along the transition path from the 4${\times}$1 to the 4${\times}$2 structure. (b) Atomic geometries and surface band structures of the 4${\times}$1 structure, the $T$ state, the $I$ state, and the 4${\times}$2 structure. The surface Brillouin zones of the 4${\times}$1 and 4${\times}$2 structures are also drawn. The In-In interatomic distances (in {\AA}) are given. The energy zero in (b) represents the Fermi level.}
\end{figure*}
According to the previously proposed CDW mechanism~\cite{Yeom1,Ahn,Yeom2,Park,Cheon} of the structural phase transition, the ground state of the In wire was described by a double Peierls-dimerized chain model with a zigzag interchain coupling between two chains [see Fig. 1(b)]: that is, below ${\sim}$120 K, the In wire undergoes a period doubling CDW transition through the Peierls dimerization on both chains. Since there are the two dimerization directions that form ``A" phase with a positive displacement ${\Delta}$ and ``B" phase with a negative ${\Delta}$ [see Fig. 1(b)] on the two chains [i.e., chain ``1" and chain ``2" in Fig. 1(b)], the double chain can have four degenerate CDW ground phases, which were denoted as AA, BA, BB, and AB in Ref.~\cite{Cheon} [see Fig. 2(a)]. It is, however, noticeable that such a coupled double Peierls-dimerized atomic chain is unlikely to be fit well for the well-established hexagon structural model~\cite{Gonzalez1} which involves the bond formation between the two chains. Therefore, we here propose a new atomistic model of the hexagon formation, where the bonds on the chains 1 and 2 are broken to form the new bonds between both chains. For one example of such bond breakage and new bond formation, the bond $b_{12'}$ between In$_1$ and In$_{2'}$ atoms on chain 1 and the bond $b_{34'}$ between In$_3$ and In$_{4'}$ atoms on chain 2 are broken to form the new bonds $b_{2'3'}$ and $b_{23}$, respectively, leading to the formation of the 4${\times}$2 hexagon structure [see Fig. 1(a)]. Note that the bond breaking/forming of $b_{12'}$/$b_{2'3'}$ accompanies that of $b_{34'}$/$b_{23}$ and vice versa, as discussed below. Since there are four different bonds on chain 1 (or 2) within the 4${\times}$2 unit cell, four degenerate hexagon structures can be generated as shown in Fig. 2(b), which correspond to AA, BA, BB, and AB configurations in Ref.~\cite{Cheon}.
The present atomistic model of the hexagon formation is expected to have an energy barrier for the consecutive bond-breaking and bond-making processes. To find this energy barrier, we calculate the energy profile along the transition path between the 4${\times}$1 and 4${\times}$2 structures by using the nudged elastic-band method~\cite{NEB}. The calculated energy profile is displayed in Fig. 3(a). We find that the transition ($T$) state is higher in energy than the 4${\times}$1 structure by 5.4 meV per 4${\times}$1 unit cell, yielding an energy barrier ($E_b$) of ${\sim}$11 meV on going from the 4${\times}$1 to the 4${\times}$2 structure. Since the 4${\times}$2 structure is more stable than the 4${\times}$1 structure by 13.6 meV per 4${\times}$1 unit cell (see Table I), we can say that the hexagon formation occurs through an exothermic reaction with bond breakage and new bond formation. It is noted that the 8${\times}$2 structure is further stabilized over the 4${\times}$1 structure by 32.9 meV per 4${\times}$1 unit cell (see Table I), and $E_b$ is reduced to be ${\sim}$8 meV on going from 4${\times}$1 to 8${\times}$2, which is much smaller than that ($E_b$ = 40 meV) obtained by a previous DFT calculation with the local-density approximation~\cite{Wall}. The existence of the energy barrier for the (4${\times}$1)${\rightarrow}$(8${\times}$2) phase transition is consistent with a recent HRLEED experiment~\cite{Klasing} where the energy barrier was confirmed by the observation of a hysteresis of diffraction spot intensities upon slow increase and decrease of the sample temperature at the (4${\times}$1)${\leftrightarrow}$(8${\times}$2) phase transition. Thus, the present theory and the previous HRLEED experiment support the first-order nature of the phase transition, contrasting with the second-order nature deduced from the CDW~\cite{Guo,GLee2} or order-disorder mechanism~\cite{Gonzalez1,Gonzalez2}.
Figure 3(b) shows the atomic geometries of the $T$ state and an intermediate ($I$) state, together with those of the 4${\times}$1 and 4${\times}$2 structures. We find that, along the transition path from the 4${\times}$1 to the 4${\times}$2 phase, the In-In interatic distance $d_{12'}$ ($d_{34'}$) increases as 3.01 (3.01), 3.05 (3.06), 3.10 (3.12), and 3.16 (3.19) {\AA} for 4${\times}$1, $T$, $I$, and 4${\times}$2, respectively, while $d_{2'3'}$ ($d_{23}$) decreases as 3.09 (3.09), 3.01 (3.01), 3.00 (2.99), and 3.00 (2.99) {\AA} [see Fig. 3(b)]. These results indicate that during the structural phase transition the bond breakage of $b_{12'}$ proceeds the new bond formation of $b_{2'3'}$, simultaneously taking place with the bond breakage of $b_{34'}$ and the new bond formation of $b_{23}$. It is noticeable that such bond-breaking and bond-making processes leading to the hexagon formation were not taken into account in the coupled double Peierls-dimerized chain model~\cite{Cheon}. Thus, unlike the present atomistic model of the hexagon formation, the latter model~\cite{Cheon} does not properly describe the hexagon structure of the indium wire.
\begin{table}[ht]
\caption{Calculated total energies of the 4${\times}$2 and 8${\times}$2 structures relative to the 4${\times}$1 structure, together with the band gap $E_{\rm g}$. }
\begin{ruledtabular}
\begin{tabular}{lcc}
& 4${\times}$2 & 8${\times}$2 \\ \hline
${\Delta}E$ (meV per 4${\times}$1 unit cell) & --13.6 & --32.9 \\
$E_{\rm g}$ (eV) & 0.27 & 0.31 \\
\end{tabular}
\end{ruledtabular}
\end{table}
The calculated surface band structures along the transition path between the 4${\times}$1 and 4${\times}$2 structures are also displayed in Fig. 3(b). We find that the 4${\times}$1 structure has the three surface bands $S_1$, $S_2$, and $S_3$ crossing the Fermi level, showing a metallic feature. On the other hand, it is clearly seen that the band gap opens during the bond-breaking and bond-making processes. Specifically, the gap opening starts from the $I$ state, leading to a band gap ($E_{\rm g}$) of 0.27 (0.31) eV at the 4${\times}$2 (8${\times}$2) structure: see Fig. 3(b) (Fig. 5S of the Supplemental material~\cite{supp}). These values of $E_{\rm g}$ agree well with those (${\sim}$0.3 eV) obtained from surface transport measurements~\cite{Tanikawa} and scanning tunneling spectroscopy~\cite{Gonzalez2}. Based on a coupled double Peierls-dimerized chain model [see Fig. 1(b)], a tight-binding Hamiltonian analysis showed that the Peierls dimerization on both chains hybridizes the $S_1$ and $S_2$ states to produce a gap opening~\cite{Cheon}. However, this Peierls instability-driven gap opening is characteristically different from the present gap opening driven by the ${\times}$2 periodic lattice reconstruction of In hexagons that involves the bond-breaking and bond-making processes within the In wire. Since the former Peierls chain model~\cite{Cheon} is lacking in the atomic description of the hexagon structure as discussed above, we believe that the CDW mechanism would not be the origin of the MI transition of the In/Si(111) system.
\begin{figure}[ht]
\centering{ \includegraphics[width=7.0cm]{fig4.eps} }
\caption{(Color online) Schematic diagrams of the soliton formation. In the top panel, the initial bond-breaking and bond-making positions are marked (X), and the arrows represent the propagations of In hexagons. In the inset, the soliton configuration AA${\rightarrow}$BA between the AA and BA CDW phases is from Ref.~\cite{Cheon}. The corresponding soliton structure is sketched based on the hexagon model, together with the possible bond-breaking and bond-making processes for the soliton movement. The kink atom indicated by the small arrow in the inset is drawn with bright color. }
\end{figure}
Finally, it is noteworthy that the coupled double Peierls chain model classified three types of topological edge states at the domain boundary between two different CDW phases~\cite{Cheon}: i.e., right-chiral, left-chiral, and nonchiral solitons among the total twelve possible solitons arising from four degenerate CDW phases. However, there remain the question of how the multiple CDW phases and the resulting solitons can be created from the 4${\times}$1 phase along the In wires. Based on the present atomistic picture for the phase transition, we speculate that around the phase transition temperature the bond breakage takes place at random sites of the 4${\times}$1 phase, initiating the hexagon formation which then propagates along the In wire (see sketch in the upper panel of Fig. 4). This speculation is supported by our calculation with a larger 4${\times}$8 supercell, where after the formation of a single hexagon it is converged to the 4${\times}$2 hexagon structure without any barrier. Indeed, the propagation of the hexagon structure was observed by an ultrafast time-resolved reflection high energy electron diffraction experiment~\cite{Wall} where adsorbates trigger the propagation of the phase front of the 8${\times}$2 structure with a constant velocity of 82 m/s. Therefore, a soliton is expected to be created at the midpoint between two bond-breakage positions after the propagation of two different hexagon structures [see Fig. 2(b)] in the opposite directions. Moreover, the created solitons can move along the In wire by the activation of bond-breaking and bond-making processes. For instance, one soliton structure corresponding to the AA${\rightarrow}$BA configuration in Ref.~\cite{Cheon} is sketched in Fig. 4, together with its movement. Future theoretical and experimental works are anticipated to investigate the detailed kinetics of solitons such as the activation energy for the soliton motion, the propagation speed of solitons, the transformation of soliton configurations, and so on.
In summary, based on first-principles DFT calculations, we have proposed the atomistic picture for the phase transition of the In/Si(111) system, where the low-temperature hexagon structure is formed through an exothermic reaction with the consecutive bond-breaking and bond-making processes. During such bond breakage and new bond formation, we found the existence of an energy barrier as well as a gap opening. Therefore, we revealed not only the first-order nature of the (4${\times}$1)${\leftrightarrow}$(8${\times}$2) phase transition but also the origin of the MI transition in terms of the ${\times}$2 periodic lattice reconstruction of In hexagon, rather than the prevailing Peierls instability-driven CDW mechanism. We anticipate that our findings will stimulate future theoretical and experimental works to reinterpret many of the existing controversial issues in the prototypical 1D In/Si(111) system in the light of the presently proposed atomistic picture for the phase transition.
This work was supported by National Research Foundation of Korea (NRF) grant funded by the Korea Government (MSIP) (2015R1A2A2A01003248). The calculations were performed by KISTI supercomputing center through the strategic support program (KSC-2015-C3-044) for the supercomputing application research. S.W.K. acknowledges support from POSCO TJ Park Foundation.
\noindent $^{*}$ Corresponding author: <EMAIL>
|
\section{Introduction}
~~~~Recently a new class of theories modifying General Relativity (GR) has started to attract great attention mainly in order to explain observation data indicating the accelerated expansion of the late Universe \cite{observ}. It is based on the rather old theory attributed to Einstein himself \cite{Einstein}, however, forgotten for decades since it is in fact not a separate theory from the viewpoint of equation of motion, but rather a reformulation of GR. Instead of torsion-free Levi-Civita connections it uses curvature free Weitzenb\"{o}ck connections \cite{Weitzenbock}. It is well known that zero-curvature connections allow for the existence of a path-independent definition of vector parallel transport, so this theory has got the name of the Teleparallel Equivalent of General Relativity (TEGR) \cite{TEGR,Hayashi:1979qx,JGPereira,Maluf:2013gaa}. What, however, has become clear rather recently is that the well-known modifications of GR, such as a theory of a scalar field nonminimally coupled to gravity, when constructed on the base of TEGR, lead to different equations of motion (for the reason of this see \cite{Bahamonde}). Such a theory has been intensively investigated in many papers during several recent years.
Generalizations of TEGR are usually constructed in the same way as modifications of GR. It is possible to generalize the Lagrangian replacing in it the torsion scalar $T$ with a function $f(T)$ \cite{Ferraro:2006jd,Linder:2010py}. Different types of cosmological scenarios appear in scalar-torsion gravity, namely, in the class of models with nonminimal coupling between the torsion scalar and the scalar field $\phi$ of the form $\xi T F(\phi)$, where $\xi$ is a coupling constant, $F(\phi)$ is some function of a scalar field \cite{Geng:2011aj,Xu:2012jf,Otalora:2013tba,Geng:2013uga}. Recently some other modifications of TEGR including a nonminimal derivative coupling to torsion \cite{Kofinas} and analogs of the Gauss-Bonnet invariant have appeared \cite{Kofinas:2014owa}. There are also modifications of TEGR with no direct analogs of GR modification \cite{Bahamonde}.
In our recent paper \cite{we} we applied dynamical system methods to scaler-torsion theory in order to find some cosmological asymptotic regimes and describe the corresponding phase portraits, focusing mainly on growing scalar field potentials.
The main qualitative result of that paper is that scalar-torsion coupling leads to much less variety of possible dynamical regimes than scalar-curvature coupling \cite{wesami,wevernov}. In the present paper we give an heuristic explanation of this feature, as well as show that more possibilities appears for decreasing scalar fields potentials, usually used in quintessence models. As such models have been created for a description of the late Universe, when the usual matter is important as well, we add hydrodynamical matter to the scalar field. We will use units with $\hbar=c=1$.
The paper is organized as follows: in Sect. 2 we present equation of motion in initial and expansion normalized variables, in Sect. 3 the analysis of stationary points and corresponding asymptotic solutions is given, and Sect. 4 gives a brief summary of results obtained.
\section{Main equations}
~~~~We start with describing basic objects of teleparallel gravity.
In this theory the dynamical variables are the
tetrad fields (also called the vierbein fields) ${\mathbf{e}_A(x^\mu)}$; here Greek indices relate to space-time and capital Latin indices belong to the tangent space-time. The metric tensor is expressed in terms of the tetrad as
\begin{equation}
\label{gmn}
g_{\mu\nu}=\eta_{\mathrm{AB}}\, e^A_\mu \, e^B_\nu,
\end{equation}
where $\eta_{\mathrm{AB}}=\mathrm {diag} (1,-1,-1,-1)$. For the definition of parallel transport of a vector the Weitzenb\"{o}ck connection \cite{Weitzenbock} is used,
\begin{equation}
\label{GW}
\overset{\mathbf{w}}\Gamma^{\lambda}_{\mu\nu}\equiv e^{\lambda}_A\partial_{\mu}e^{A}_{\nu}.
\end{equation}
Then the torsion tensor and the torsion scalar are given by
\begin{equation}
\label{Tlmn}
T^{\lambda}_{\mu\nu}\equiv\overset{\mathbf{w}}\Gamma^{\lambda}_{\nu\mu}-\overset{\mathbf{w}}\Gamma^{\lambda}_{\mu\nu}=e^{\lambda}_A(\partial_\mu e^A_{\nu}-\partial_\nu e^A_{\mu}),
\end{equation}
\begin{equation}
\label{T}
T\equiv\frac{1}{4}T^{\rho\mu\nu}T_{\rho\mu\nu}+\frac{1}{2}T^{\rho\mu\nu}T_{\nu\mu\rho}-T^{\rho}_{\rho\mu}T^{\nu\mu}_{\nu}.
\end{equation}
On the other hand, the curvature of the Weitzenb\"{o}ck connection vanishes identically.
~~~~We consider the cosmological model with the following action:
\begin{equation}
S=\int e~ d^{4}x \left[\frac{T}{2 K}+\frac{1}{2}\,\partial_{\mu}\phi
\partial^{\mu}\phi-V(\phi)+\frac{\xi}{2}B(\phi)T\right]+\mathcal{S}_m,
\label{action}
\end{equation}
where $e=\sqrt{-g}$ is the determinant of the tetrad, ~~$K=8\pi G$, ~~$\phi$ is a canonical scalar field, ~~$V(\phi)$ its potential, and $B(\phi)$ its arbitrary nonminimal coupling with the torsion scalar $T$, ~~$\mathcal{S}_m$ is the matter action. In the classical scalar-curvature theory the action has the same form except for the curvature scalar $R$ replacing torsion scalar $T$.
For the spatially flat Friedmann-Robertson-Walker tetrad $e^A_\mu = \mathrm{diag}(1, a(t), a(t), a(t))$
\\ (the corresponding metric is $\mathrm{d}s^2=\mathrm{d}t^2-a^2(t)\mathrm{d}l^2$)
the system of field equations is \cite{we}
\begin{equation}
\begin{array}{l}
3H^2=K\left( \frac{{\dot{\phi}}^2}{2}+V(\phi)-3\xi H^2 B(\phi)+\rho\right) ,
\label{system1}
\end{array}
\end{equation}
\begin{equation}
\begin{array}{l}
2\dot{H}=-K( {\dot{\phi}}^2+2\xi H \dot{\phi}B'(\phi)+2\xi\dot{H}B(\phi)+\rho(1+\omega)),
\label{system2}
\end{array}
\end{equation}
\begin{equation}
\begin{array}{l}
\ddot{\phi}+3 H\dot{\phi}+3\xi H^2 B'(\phi)+V'(\phi)=0.
\label{system3}
\end{array}
\end{equation}
Here $a(t)$ is the scale factor, $H(t)\equiv\frac{\dot a}{a}$ is the Hubble parameter, the prime denotes the derivative with respect to $\phi$. We have used that in the chosen tetrad $T=-6H^2$. The matter equation of state is $p=\omega\rho$, where $\omega\in[-1;1]$.
\subsection{The effective potential}
~~~~In the standard scalar-curvature theory of a nonminimal coupling the conformal transformation to Einstein frame is often used. In the Einstein frame the theory is equivalent to GR with a scalar field as a source, so the evolution of the scalar field in an expanding Universe is, as usual, directed to the minimum of the potential, which is not true in the initial nonminimal formulation, called the Jordan frame. This conformal transformation is usually divided into two steps: a redifinition of the potential via $V_{\mathrm{eff}}(\phi)=V(\phi)/U^2(\phi)$, where $U(\phi)=1+K\xi B(\phi)$ in our notations, and a redifinition of the scalar field in order to get the canonical kinetic term. The combination denoted here as $V_{\mathrm{eff}}(\phi)$ is rather interesting by itself. It is a conformal invariant, so it is not changed by any conformal transformation \cite{Kuusk}. Moreover, it can give important information as regards the dynamics of the system without the second step to the Einstein frame (which is usually much more technically complicated than very easily calculated first step). Namely, de Sitter solutions corresponds to minima of the effective potential, their stability is determined by sign of the second derivative of the effective potential in a standard way \cite{wevernov, Kuusk}.
In the scalar-torsion theory the Einstein frame does not exist \cite{Yang}. However, it is possible to introduce some analog of effective potential with the same properties as in the scalar-curvature theory. Indeed, taking into account (\ref{system1}) we rewrite (\ref{system3}) for $\rho=0$
\begin{equation}
\ddot{\phi}+3 H\dot{\phi}+\frac{K\xi B'(\phi){\dot\phi}^2}{2(1+K\xi B(\phi))}+\frac{K\xi B'(\phi)V(\phi)+V'(\phi)(1+K\xi B(\phi))}{1+K\xi B(\phi)}=0.
\label{ddotphi}
\end{equation}
We see that the effective potential, of the form
\begin{equation}
V_{\mathrm{eff}}(\phi)=V(\phi)(1+K\xi B(\phi)),
\end{equation}
has a derivative with respect to $\phi$
\begin{equation}
\frac{\mathrm{d}V_{\mathrm{eff}}(\phi)}{\mathrm{d}\phi}=V'_{\mathrm{eff}}(\phi)=K\xi B'(\phi) V(\phi)+V'(\phi)(1+K\xi B(\phi)),
\end{equation}
which coincides with numerator of the last term in (\ref{ddotphi}).
The de Sitter solution $H=H_0$, $\phi=\phi_0$, in this model exists for
\begin{equation}
\begin{cases}
3{H_0}^2(1+K\xi B(\phi_0))=KV(\phi_0)
\\3\xi {H_0}^2 B'(\phi_0)=-V'(\phi_0)
\end{cases}
\end{equation}
From this system it is follows that
\begin{equation}
\begin{array}{l}
\frac{KV(\phi_0)}{1+K\xi B(\phi_0)}=-\frac{V'(\phi_0)}{\xi B'(\phi_0)}~~\Rightarrow\\
\\\Rightarrow K\xi B'(\phi_0) V(\phi_0)+V'(\phi_0)(1+K\xi B(\phi_0))=V'_{eff}(\phi_0)=0
\label{conditiondS}
\end{array}
\end{equation}
Now we add small perturbations to the de Sitter solution: $\phi(t)=\phi_0+\delta\phi$, $\dot\phi(t)=\delta\dot\phi$,
\\$H(t)=H_0+\delta H$. Substituting these perturbations to (\ref{ddotphi}) we get in the first order
\begin{equation}
\delta\ddot\phi+3 H_0\delta\dot{\phi}+\delta\phi\frac{V''_{\mathrm{eff}}(\phi_0)}{1+K\xi B(\phi_0)}=0.
\end{equation}
New variables $s=\delta\phi$, $r=\delta\dot\phi$ are introduced and the first-order system of differential equations is written in the form
\begin{equation}
\begin{array}{l}
\dot s=r,\\
\dot r=-3 H_0 r-s\frac{V''_{\mathrm{eff}}(\phi_0)}{1+K\xi B(\phi_0)}.
\label{systems1s2}
\end{array}
\end{equation}
We find eigenvalues for the system (\ref{systems1s2})
\begin{equation}
\begin{array}{l}
\begin{vmatrix}
-\lambda & 1\\
-\frac{V''_{\mathrm{eff}}(\phi_0)}{1+K\xi B(\phi_0)} & -3H_0-\lambda
\end{vmatrix}
=\lambda^2 +3H_0\lambda+\frac{V''_{\mathrm{eff}}(\phi_0)}{1+K\xi B(\phi_0)}=0 \Rightarrow\\
\\\Rightarrow \lambda_{1,2}=-\frac{3}{2}H_0\pm\frac{1}{2}\sqrt{9{H_0}^2-\frac{4 V''_{\mathrm{eff}}(\phi_0)}{1+K\xi B(\phi_0)}}<0 ~~~~\text{for ~~$V''_{\mathrm{eff}}(\phi_0)>0$, ~~$\xi>0$, ~~$B(\phi_0)>0$}.
\end{array}
\end{equation}
So, the de Sitter solution is stable in the minima of the effective potential, as expected. It is worth to note that the second equation in Eq. (\ref{conditiondS}) is just the ``balanced solution'' studied in \cite{Laur}, so from computational point of view the effective potential gives nothing new. However, from heuristic point of view often it is much easier to visualize the locations of the minima of some function instead of doing calculations. For example, for positive power-law potential and coupling functions $B(\phi) \sim \phi^N$ and $V(\phi) \sim \phi^n$ de Sitter solution exists only for negative $n$ with $0< -n < N$. This fact has been established computationally in \cite{we} and becomes now a trivial consequence of the form of $V_{\mathrm{eff}}(\phi)=U(\phi)V(\phi)$.
\subsection{Dimensionless variables}
~~~~We introduce new dimensionless variables
\begin{equation}
\begin{array}{l}
x=\frac{K{\dot{\phi}}^2}{6 H^2(1+K\xi B(\phi))},~~~~
y=\frac{K V(\phi)}{3 H^2(1+K\xi B(\phi))},~~~~
z=\frac{K \rho}{3 H^2(1+K\xi B(\phi))},\\
\\~~~~~~~~~~~~~~~~~~~~~~~~m=\frac{\dot{\phi}}{H \phi},~~~~
A=\frac{\phi B'(\phi)}{1+K\xi B(\phi)}
\label{variables}
\end{array}
\end{equation}
and also dimensionless parameters
\begin{equation}
\begin{array}{l}
b=\frac{\phi B''(\phi)}{B'(\phi)},~~~~
c=\frac{\phi V'(\phi)}{V(\phi)}.
\end{array}
\end{equation}
Choosing the power-law function $B(\phi)=\phi^N$ and the potential $V(\phi)=V_0 \phi^n$ we get $b=N-1$, $c=n$.
Note the useful relation between $A$, $x$, $m$, namely,
\begin{equation}
x=\frac{K}{6 N}m^2 A^{\frac{2}{N}}{(N-K\xi A)}^{\frac{N-2}{N}}.
\label{xmAN}
\end{equation}
It is useful to introduce auxiliary variables
\begin{equation}
\begin{array}{l}
X=\frac{\ddot{\phi}}{H\dot{\phi}},~~~~
Y=\frac{\dot{H}}{H^2},
\end{array}
\end{equation}
which are expressed through dimensionless variables and parameters from the system (\ref{system1})-(\ref{system3}),
\begin{equation}
\begin{array}{l}
X=-3-\frac{K\xi A m}{2 x}-\frac{c y m}{2 x},\\
\\Y=-3 x-K\xi A m-\frac{3}{2} z(1+\omega),
\label{XYN}
\end{array}
\end{equation}
From (Eq. \ref{system1}) using (\ref{variables}) we get
\begin{equation}
\begin{array}{l}
1=x+y+z,
\label{z}
\end{array}
\end{equation}
then $z=1-x-y$.
Taking the derivative of the variables $y$, $m$, $A$ with respect to $\ln (a)$ ($'=\frac{\mathrm{d}}{\mathrm{d}\ln a}$) we obtain the following first-order system of differential equations:
\begin{equation}
\begin{array}{l}
y'=y(cm-2 Y-K\xi A m),\\
\\m'=m(X-Y-m),\\
\\A'=A m(b+1-K\xi A),\
\end{array}
\end{equation}
and finally substituting (\ref{xmAN}), (\ref{XYN}), and (\ref{z}) we get
\begin{equation}
\begin{array}{l}
y'=y\left[ c m+\frac{K}{b+1} m^2 A^{\frac{2}{b+1}}{(b+1-K\xi A)}^{\frac{b-1}{b+1}}+K\xi m A+\right. \\
~~~~~~\left. +3\left( 1-\frac{K}{6(b+1)} m^2 A^{\frac{2}{b+1}}{(b+1-K\xi A)}^{\frac{b-1}{b+1}}-y\right) (1+\omega)\right] ,\\
\\m'=-3 m-3(b+1)\xi A^{\frac{b-1}{b+1}}{(b+1-K\xi A)}^{\frac{1-b}{b+1}}-\frac{3 c y (b+1)}{K A^{\frac{2}{b+1}}}{(b+1-K\xi A)}^{\frac{1-b}{b+1}}+\\
~~~~~~~~+\frac{K m^3}{2(b+1)} A^{\frac{2}{b+1}}{(b+1-K\xi A)}^{\frac{b-1}{b+1}}+K\xi A m^2-m^2+\\
~~~~~~~~+\frac{3}{2} m\left( 1-\frac{K m^2}{6(b+1)} A^{\frac{2}{b+1}}{(b+1-K\xi A)}^{\frac{b-1}{b+1}}-y\right) (1+\omega),\\
\\A'=A m(b+1-K\xi A).
\end{array}
\end{equation}
If $N=2 ~~(b=1)$, then instead of (\ref{xmAN}), and (\ref{XYN}) we have
\begin{equation}
x=\frac{K}{12}m^2 A,
\end{equation}
\begin{equation}
\begin{array}{l}
X=-3-6\frac{\xi}{m}-6\frac{K c y A}{m},\\
\\Y=-\frac{K}{4}m^2 A-K\xi A m-\frac{3}{2}(1-\frac{K}{12}m^2 A-y)(1+\omega),
\label{XYN2}
\end{array}
\end{equation}
and the first-order system of differential equations
\begin{equation}
\begin{array}{l}
y'=y\Big{(}c m+\frac{K}{2} m^2 A+K\xi m A+3\left( 1-\frac{K}{12} m^2 A-y\right) (1+\omega)\Big{)} ,\\
\\m'=-3 m-6\xi-\frac{6 c y}{K A}+\frac{K}{4}m^3 A+K\xi A m^2-m^2
+\\
~~~~~~~~+\frac{3}{2} m\left(1-\frac{K m^2}{12} A-y \right)(1+\omega),\\
\\A'=A m(2-K\xi A).
\label{systemyAN2}
\end{array}
\end{equation}
We consider only the case of $N=2$ in the present paper.
\section{Stationary points and corresponding regimes}
\subsection{Stationary points analysis}
~~~~Solving the system (\ref{systemyAN2}) with vanishing left-hand sides, we find
the following stationary points:
\\
\\
\textbf{1.}~~~~$x=0$, $y=1$, $z=0$, $m=0$, $A=-{\frac{c}{K \xi}}$.
\\We calculate the eigenvalues for the Jacobian matrix associated with the system (\ref{systemyAN2}) to find
\begin{equation}
\begin{array}{l}
\lambda_1 =-3(1+\omega),\\
\lambda_{2,3}=-\frac{3}{2}\pm\frac{1}{2}\sqrt{9-24\xi(c+2)},
\end{array}
\end{equation}
so this point is stable for non-phantom matter.
For this point the quantity $Y=\frac{\dot H}{H^2}=0$ is found using (\ref{XYN2}) and, therefore, $\dot H=0$. Then we find the time dependence of the scale factor,
\begin{equation}
\begin{array}{l}
a(t)=a_0 e^{H_0(t-t_0)}.
\label{dSa}
\end{array}
\end{equation}
As the coordinate of this fixed point is $A=\frac{N\phi^N}{1+K\xi\phi^N}\to-{\frac{n}{K \xi}}\neq0$, then the scalar field approaches a constant
\begin{equation}
\begin{array}{l}
\phi=\phi_0.
\label{dSphi}
\end{array}
\end{equation}
This is a de Sitter solution. The constants $\phi_0$ and $H_0$ are found with the substitution of this solution to the initial system of Eqs. (\ref{system1})-(\ref{system3}): ~~$\phi_0=\pm\sqrt{-\frac{n}{K\xi(2+n)}}$, ~~$H_0=\pm\sqrt{-\frac{nV_0{\phi_0}^{n-2}}{6\xi}}$. Clearly, this regime exists only for $0>n>-2$. Since it is stable it can be used for describing late-time acceleration of our Universe. Definitely, the unnaturally low value of $H_0$ (in natural units) should transform to very low value of some of the parameters of the theory. Note, however, that in the theory under investigation we have more possibilities to get a very low $H_0$ --- this may happen due to either very low $V_0$, a very low $\xi$ or a potential power index $n$ being very close to $-2$. For the last possibility, if we denote $n+2=\epsilon << 1$, we get $H_0 \sim \epsilon\sqrt{V_0 \xi}$, so small corrections to the $\phi^{-2}$
potential would do the job.
\\
\\
\textbf{2.}~~~~$x=1$, $y=0$, $z=0$, $m=\sqrt{6\xi}$, $A=\frac{2}{K\xi}$.
\\The corresponding eigenvalues are
\begin{equation}
\begin{array}{l}
\lambda_1=-2\sqrt{6\xi},\\
\lambda_2=3(1-\omega)+2\sqrt{6\xi},\\
\lambda_3=(c+2)\sqrt{6\xi}+6.
\end{array}
\end{equation}
Using (\ref{XYN2}) the quantity $Y$ at this stationary point is obtained to be $Y_{\mathrm{stat}}=-3-2\sqrt{6\xi}$ and now we can restore the time dependence of the Hubble parameter, $H(t)=-\frac{1}{Y_{\mathrm{stat}}(t-t_0)}$, and the scale factor
\begin{equation}
\begin{array}{l}
a(t)=a_0{|t-t_0|}^{-\frac{1}{Y_{\mathrm{stat}}}}=a_0{|t-t_0|}^{\frac{1}{3+2\sqrt{6\xi}}}.
\end{array}
\end{equation}
The corresponding behavior of the scalar field is obtained applying the definition of the variable $m=\frac{\dot\phi}{H\phi}$, then ~~$\frac{\dot\phi}{\phi}=m_{\mathrm{stat}}\frac{\dot a}{a}$, where $m_{\mathrm{stat}}$ -- the coordinate of a stationary point,
\begin{equation}
\begin{array}{l}
\phi(t)=\phi_0{|t-t_0|}^{-\frac{m_{\mathrm{stat}}}{Y_{\mathrm{stat}}}}=\phi_0{|t-t_0|}^{\frac{\sqrt{6\xi}}{3+2\sqrt{6\xi}}}.
\end{array}
\end{equation}
\\
\\\textbf{3.}~~~~$x=1$, $y=0$, $z=0$, $m=-\sqrt{6\xi}$, $A=\frac{2}{K\xi}$.\\
The eigenvalues for this fixed point are
\begin{equation}
\begin{array}{l}
\lambda_1 =2\sqrt{6\xi},\\
\lambda_2=3(1-\omega)-2\sqrt{6\xi},\\
\lambda_3=6-\sqrt{6\xi}(c+2).
\end{array}
\end{equation}
We get asymptotic behavior $a(t)$ and $\phi(t)$ calculating the quantity $Y_{\mathrm{stat}}=-3+2\sqrt{6\xi}$ and using $m=m_{\mathrm{stat}}=-\sqrt{6\xi}$ at this point,
\begin{equation}
\begin{array}{l}
a(t)=a_0{|t-t_0|}^{\frac{1}{3-2\sqrt{6\xi}}},\\
\\\phi(t)=\phi_0{|t-t_0|}^{\frac{\sqrt{6\xi}}{2\sqrt{6\xi}-3}}.
\label{aphi3}
\end{array}
\end{equation}
For $\xi=\frac{3}{8}$ instead of this solution (\ref{aphi3}) we have the exponential one (because in this case $Y=0$, ~~$H=H_0$, ~~$\frac{\dot\phi}{\phi}=m_{\mathrm{stat}}H_0$, ~~$m_{\mathrm{stat}}=-\frac{3}{2}$)
\begin{equation}
\begin{array}{l}
a(t)=a_0e^{H_0(t-t_0)},\\
\\\phi(t)=\phi_0 e^{-\frac{3}{2} H_0(t-t_0)}.
\label{exp3aphi}
\end{array}
\end{equation}
The points $2$ and $3$ represent a situation when kinetic term of the scalar field dominates.
\\
\\\textbf{4.}~~~~$x=\frac{8\xi}{3(\omega-1)^2}$, $y=0$, $z=\frac{3+3\omega^2-6\omega-8\xi}{3(\omega-1)^2}$, $m=\frac{4\xi}{\omega-1}$, $A=\frac{2}{K\xi}$.
\\We find eigenvalues for this point,
\begin{equation}
\begin{array}{l}
\lambda_1=\frac{3\omega^2-3+4 c\xi}{\omega-1}\\
\lambda_2=\frac{3{(\omega-1)}^2-8\xi}{2(\omega-1)}\\
\lambda_3=\frac{8\xi}{1-\omega}.
\end{array}
\end{equation}
As the quantity $Y$ at this stationary point is $Y_{\mathrm{stat}}=\frac{3-3\omega^2-8\xi}{2(\omega-1)}$, we have
\begin{equation}
\begin{array}{l}
a(t)=a_0{|t-t_0|}^{\frac{2(\omega-1)}{3\omega^2-3+8\xi}},\\
\\\phi(t)=\phi_0{|t-t_0|}^{\frac{8\xi}{3\omega^2-3+8\xi}}.
\label{aphi4}
\end{array}
\end{equation}
The corresponding behavior $\rho(t)$ is found using $\dot \rho+3H\rho(1
+\omega)=0$, then $\frac{\dot \rho}{\rho}=-3(1+\omega)\frac{\dot a}{a}$ and
\begin{equation}
\begin{array}{l}
\rho(t)=\rho_0{|t-t_0|}^{\frac{3(1+\omega)}{Y_{\mathrm{stat}}}}=\rho_0{|t-t_0|}^{\frac{6(1-\omega^2)}{3\omega^2-3+8\xi}}.
\label{rho4}
\end{array}
\end{equation}
This is a kinetic tracker regime when the potential term is negligible, and the ratio between the kinetic energy of the scalar field and the matter
density remains constant. When $\xi=\frac{3}{8}(1-\omega^2)$, the solution (\ref{aphi4}), (\ref{rho4}) does not exist and the exponential one appears ($m_{\mathrm{stat}}=-\frac{3}{2}(1+\omega)$),
\begin{equation}
\begin{array}{l}
a(t)=a_0 e^{H_0(t-t_0)},\\
\\\phi(t)=\phi_0 e^{-\frac{3}{2}(1+\omega)H_0(t-t_0)},\\
\\\rho(t)=\rho_0 e^{-3(1+\omega)H_0(t-t_0)},
\label{exp4aphirho}
\end{array}
\end{equation}
where ${H_0}^2=-\frac{4\rho_0}{9{\phi_0}^2\omega(1+\omega)}$ is found on substituting the exponential solution (\ref{exp4aphirho}) into the initial system of equations (\ref{system1})-(\ref{system3}).
\\
\\\textbf{5.}~~~~$x=\frac{\xi{(c+2)}^2}{6}$, $y=1-\frac{\xi{(c+2)}^2}{6}$, $z=0$, $m=-\xi(c+2)$, $A=\frac{2}{K\xi}$.
\\
The eigenvalues are calculated,
\begin{equation}
\begin{array}{l}
\lambda_1 =2\xi(c+2),\\
\lambda_2=-3+\frac{\xi{(c+2)}^2}{2},\\
\lambda_3=-3(1+\omega)+c\xi(c+2).
\end{array}
\end{equation}
Similar to the previous points, using ~~$Y_{\mathrm{stat}}=\frac{1}{2}\xi(4-c^2)$ and ~~$m_{\mathrm{stat}}=-\xi(c+2)$ we find $a(t)$, $\phi(t)$ ~~($c=n$),
\begin{equation}
\begin{array}{l}
a(t)=a_0{|t-t_0|}^{\frac{2}{\xi(n^2-4)}},\\
\\\phi(t)=\phi_0{|t-t_0|}^{{\frac{2}{2-n}}}.
\end{array}
\end{equation}
\\
\\\textbf{6.}~~~~$x=\frac{3{(1+\omega)}^2}{2\xi c^2}$, $y=\frac{3-3\omega^2-4 c\xi)}{2\xi c^2}$, $z=\frac{c^2\xi-3-3\omega+2 c\xi}{\xi c^2}$, $m=\frac{-3(1+\omega)}{c}$, $A=\frac{2}{K\xi}$.
\\We obtain the eigenvalues for this point,
\begin{equation}
\begin{array}{l}
\lambda_1=\frac{6(1+\omega)}{c},\\
\\\lambda_{2,3}=\frac{1}{4\xi c}\left( 3\xi(2-c+\omega(c+2))\pm\sqrt{-3\xi(72\omega^3+f_1(\xi,c)\omega^2+f_2(\xi,c)\omega+f_3(\xi,c))}\right) .
\label{lambda6}
\end{array}
\end{equation}
where $f_1(\xi,c)=-27 c^2\xi-12\xi-60 c\xi+72$,
\\\text{~~~~~~~~}$f_2(\xi,c)=-24\xi-72+96 c\xi+6 c^2\xi$, and
\\\text{~~~~~~~~}$f_3(\xi,c)=21 c^2\xi-72+156 c\xi-32\xi^2 c^3-12\xi-64 c^2\xi^2$.
The quantities $Y_{\mathrm{stat}}=\frac{3}{2 c}(1+w)(2-c))$ and $m_{\mathrm{stat}}=\frac{-3(1+\omega)}{c}$ give us the behavior $a(t)$, $\phi(t)$, and $\rho(t)$ ~~($c=n$)
\begin{equation}
\begin{array}{l}
a(t)=a_0{|t-t_0|}^{\frac{2 n}{3(1+\omega)(n-2)}},\\
\\\phi(t)=\phi_0{|t-t_0|}^{\frac{2}{2-n}},\\
\\\rho(t)=\rho_0{|t-t_0|}^{\frac{2 n}{2-n}},
\end{array}
\end{equation}
This is a tracker when the potential and kinetic terms of the scalar field as well as the density of the matter are constant with respect to each other.
\subsection{Conditions of existence and stability of obtained solutions}
~~~~~With the substitution of the power-law solution ~~$a(t)=a_0{(t-t_0)}^\alpha$, ~~$\phi(t)=\phi_0{(t-t_0)}^{\alpha\beta}$,
\\and $\rho(t)=\rho_0{(t-t_0)}^{-3(1+\omega)\alpha}$ to the initial system of equations (\ref{system1}), (\ref{system2}), (\ref{system3}), where $B(\phi)=\phi^2$, $V(\phi)=V_0\phi^n$, we find the conditions of the existence and stability of the six solutions obtained from the previous subsection. We consider only the cases with $n<0$, $\xi>0$, and only positive values of $\phi$. It is worth to recall that in general a fixed point does not necessary correspond to some solution of the initial system. First of all, all regimes found above (except the de Sitter solution, which is an exact, and not only an asymptotic solution) exist in the $\phi \to \infty$ limit where we can neglect the Einstein term in comparison with the term originating from a nonminimal coupling in the denominator of the expansion in normalized variables we use. When the field is small (which is realized near a cosmological singularity for decreasing power-law potentials) the Universe expands according to GR, since the correction terms are less important. Such regimes for quintessence potentials are well known and are not included in our analysis here. As for the large $\phi$ regimes, studied in the present paper, we require that the omitted terms in the equations of motion be negligible for large $\phi$. Otherwise the regime cannot be realized as an asymptotic solution. For example, suppose we have a vacuum asymptotic solution. If the influence of matter grows with growing $\phi$, this means that the asymptotic solution corresponding to this particular expansion dynamics is absent if we add any amount of matter. This does not mean that such a regime has no physical meaning at all --- the Universe may follow it as transient one, if the amount of matter is small enough. However, such a situation does not belong to the asymptotic regimes, and we will not list it in the following.
The analysis described above (with stability results got from the corresponding eigenvalues) leads to the following results for vacuum solutions summarized in Table 1 (points $4$ and $6$, being non-vacuum, are not included).
We can see from this table that in the vacuum case there are three possible future stable regimes:
\begin{itemize}
\item For $n>-2$ a de Sitter solution exists and is stable.
\item For $n<-2$ and $\xi < 6/{(n+2)}^2$ the regime of point $5$ is stable.
\item For $n<-2$ and $\xi>6/{(n+2)}^2$ the kinetic dominating regime of point $2$ is stable.
\end{itemize}
When the matter is taken into account, the situation is described in the Table 2 (note that the de Sitter solution in the presence of matter is an asymptotic solution).
\newpage
\begin{center}
Table 1: \textbf{Conditions of existence and stability of solutions in vacuum}
\begin{tabular}{|c|c|c|c|}
\hline
{} & \multicolumn{2}{c|}{} & {} \\
\textbf{\textnumero} & \multicolumn{2}{c|}{\textbf{Conditions}} & \textbf{Type}\\
{\textbf{of point}} & \multicolumn{2}{c|}{\textbf{of existence}} & ~~\textbf{of stability}~~\\
\hline
{} & \multicolumn{2}{c|}{} & {}\\
{} & \multicolumn{2}{c|}{} & $0<\xi\leqslant\frac{3}{8(n+2)}$ \\
\textbf{1.} & \multicolumn{2}{c|}{$-2<n<0$} & \textbf{Stable} node\\
\cline{4-4}
{} & \multicolumn{2}{c|}{} & {}\\
{} & \multicolumn{2}{c|}{} & $\xi>\frac{3}{8(n+2)}$\\
{} & \multicolumn{2}{c|}{} & \textbf{Stable} focus\\
\hline
{} & \multicolumn{2}{c|}{} & {}\\
\textbf{2.} & \multicolumn{2}{c|} {$n<-2$, ~~$\xi>\frac{6}{{(n+2)}^2}$, ~~$t\to\infty$} & \textbf{Stable} node\\
{} & \multicolumn{2}{c|}{} & {}\\
\hline
{} & {} & {} & {}\\
{} & {} & \textbf{1).} $0<\xi<\frac{3}{8}$, ~~$t\to t_0$ & {}\\
{} & $-2<n<0$, & ~~~~~~\textbf{2).} $\frac{3}{8}<\xi<\frac{6}{{(n+2)}^2}$, ~~$t\to\infty$ & {}\\
{} & {} & {} & {}\\
\cline{2-3}
\textbf{3.} & {} & {} & {}\\
{} & {} & \textbf{1).} $0<\xi<\frac{3}{8}$, ~~$t\to t_0$ & Unstable node\\
{} & $n\leqslant-2$, &\textbf{2).} $\xi>\frac{3}{8}$, ~~$t\to\infty$ & {}\\
{} & {} & {} & {}\\
\cline{2-3}
{} & \multicolumn{2}{c|}{} & {}\\
{} & \multicolumn{2}{c|}{$n<0$, ~~$\xi=\frac{3}{8}$, ~~$t\to-\infty$} & {}\\
{} & \multicolumn{2}{c|}{} & {}\\
\hline
{} & \multicolumn{2}{c|}{} & {}\\
{} & \multicolumn{2}{c|}{$-2<n<0$, ~~$0<\xi<\frac{6}{{(n+2)}^2}$, ~~$t\to\infty$} & Saddle\\
\textbf{5.} & \multicolumn{2}{c|}{} & {}\\
\cline{2-4}
{} & \multicolumn{2}{c|}{} & {}\\
{} & \multicolumn{2}{c|}{$n<-2$, ~~$0<\xi<\frac{6}{{(n+2)}^2}$, ~~$t\to\infty$} & \textbf{Stable} node\\
{} & \multicolumn{2}{c|}{} & {}\\
\hline
\end{tabular}
\end{center}
\newpage
\begin{center}
Table 2: \textbf{Conditions of existence and stability of solutions for $\rho\neq0$ }
\begin{tabular}{|c|c|c|c|}
\hline
{} & \multicolumn{2}{c|}{} & {} \\
\textbf{\textnumero} & \multicolumn{2}{c|}{\textbf{Conditions}} & \textbf{Type}\\
{\textbf{of point}} & \multicolumn{2}{c|}{\textbf{of existence}} & ~~\textbf{of stability}~~\\
\hline
{} & \multicolumn{2}{c|}{} & {}\\
{} & \multicolumn{2}{c|}{} & $0<\xi\leqslant\frac{3}{8(n+2)}$ \\
\textbf{1.} & \multicolumn{2}{c|}{$-2<n<0$, ~~$\omega\in(-1;1]$} & \textbf{Stable} node\\
\cline{4-4}
{} & \multicolumn{2}{c|}{} & {}\\
{} & \multicolumn{2}{c|}{} & $\xi>\frac{3}{8(n+2)}$\\
{} & \multicolumn{2}{c|}{} & \textbf{Stable} focus\\
\hline
{} & {} & {} & {}\\
{} & {} & \textbf{1).} $\omega\in[-1;0)$, ~~$0<\xi<\frac{3{(1-\omega)}^2}{8}$, ~~$t\to t_0$ & {}\\
\textbf{3.} & $n<0$, &\textbf{2).} $\omega\in[0;1)$, ~~$\frac{3}{8}<\xi<\frac{6}{{(n+2)}^2}$, ~~$t\to\infty$ & {}\\
{} & {} & {} & Unstable node\\
\cline{2-3}
{} & \multicolumn{2}{c|}{} & {}\\
{} & \multicolumn{2}{c|}{$n<0$, ~~$\xi=\frac{3}{8}$, ~~$\omega\in[-1;0)$, ~~$t\to-\infty$} & {}\\
{} & \multicolumn{2}{c|}{} & {}\\
\hline
{} & \multicolumn{2}{c|}{} & {}\\
{} & \multicolumn{2}{c|}{$n<0$, ~~$0<\xi<\frac{3(1-\omega^2)}{8}$, ~~$\omega\in(-1;1)$, ~~$t\to t_0$} & Unstable node\\
\textbf{4.} & \multicolumn{2}{c|}{} & {}\\
\cline{2-4}
{} & \multicolumn{2}{c|}{} & {}\\
{} & \multicolumn{2}{c|}{$n<0$, ~~$\xi=\frac{3}{8}(1-\omega^2)$, ~~$\omega\in(-1;0)$, ~~$t\to-\infty$} & Saddle\\
{} & \multicolumn{2}{c|}{} & {}\\
\hline
{} & \multicolumn{2}{c|}{} & {}\\
\textbf{5.} & \multicolumn{2}{c|}{$n<-2$, ~~$0<\xi<\frac{3(1+\omega)}{n(n+2)}$, ~~$\omega\in(-1;1]$, ~~$t\to\infty$} & \textbf{Stable} node\\
{} & \multicolumn{2}{c|}{} & {}\\
\hline
{} & \multicolumn{2}{c|}{} & {}\\
{} & \multicolumn{2}{c|}{} & $\frac{3(1+\omega)}{n(n+2)}<\xi\leqslant\xi_0$\\
\textbf{6.} & \multicolumn{2}{c|}{$n<-2$, ~~$\xi>\frac{3(1+\omega)}{n(n+2)}$, ~~$\omega\in(-1;1]$, ~~$t\to\infty$} & \textbf{Stable} node \\
\cline{4-4}
{} & \multicolumn{2}{c|}{} & {}\\
{} & \multicolumn{2}{c|}{} & $\xi>\xi_0$\\
{} & \multicolumn{2}{c|}{} & \textbf{Stable} focus \\
\hline
\end{tabular}
\end{center}
In the Table 2 ~~$\xi_0=-\frac{3}{64n^2(n+2)}(f_4(n,\omega)+\sqrt{f_5(n,\omega)f_6(n,\omega)})$,
\\where $f_4(n,\omega)=-52n-7n^2+4+\omega^2(9n^2+20n+4)+2\omega (4-16n-n^2)$
\\\text{~~~~~~~~}$f_5(n,\omega)=\omega^2(4+36n+81n^2)+\omega(126n^2-64n+8)+4-100n+49n^2$,
\\\text{~~~~~~~~}$f_6(n,\omega)={(2-n+\omega(n+2))}^2$.
For $\xi=\xi_0$ the square root equals zero in the eigenvalues $\lambda_{2,3}$ of point 6 (see \ref{lambda6}).
We see that point $2$ (a kinetic dominating regime) is absent --- this situation is described in the beginning of this section as an illustration of what we call the existing asymptotic regimes. In this particular case matter with $w<1$ would inevitably destroy a vacuum regime of the point $2$ in late time. The future stable regimes are:
\begin{itemize}
\item For $n>-2$, as for the vacuum case, the future stable point is de Sitter solution.
\item For $n<-2$ and $\xi<3(1+\omega)/(n^2+2n)$ the regime of point $5$ is stable, so the scalar field dominates in late time.
\item For $n<-2$ and $\xi>3(1+\omega)/(n^2+2n)$ the tracker (point $6$) is stable to the future.
\end{itemize}
An example of configurations of stability regions for $\omega=0$ is shown in Fig. \ref{Fig1}.
\begin{figure}[hbtp]
~~~~~~~~~~~~\includegraphics [scale=0.68] {Regionstabw0.eps}
\caption{Regions of a stability of solutions corresponding to the points 1 (red), 5 (green), 6 (cyan) for the case
$\rho\neq0$ with $\omega=0$.}
\label{Fig1}
\end{figure}
\section{Conclusions}
~~~~In the present paper we considered the evolution of the Universe in a teleparallel version of a nonminimally coupled scalar field theory, focusing mostly on quintessence scalar field potentials. In our previous paper \cite{we} we already pointed out the radical differences in the dynamics of a ``classical'' and teleparallel cases of this theory. Here we provided a qualitative explanation of this difference using the notion of an effective scalar field potential. It appears as a first step from a Jordan to an Einstein frame (to be followed by redefinition of the scalar field in order to transform the kinetic term to the canonical form) in the standard theory, though it already gives an intuitively clear picture of possible dynamical regimes without this second step. It is interesting that in the teleparallel version, where the Einstein frame does not exist at all \cite{Yang}, the effective potential nevertheless can be introduced in the same manner. The difference in the forms of this effective potential (being $V(\phi)/U^2(\phi)$ for the classical case and $V(\phi)U(\phi)$ for the teleparallel case, where $V(\phi)$ is the scalar field potential and $U(\phi)$ is the coupling function which includes the Einstein-Hilbert term) results in very different dynamics for the same scalar field potentials.
For the standard scalar-curvature coupling, the presence of $U(\phi)$ [which is usually taken in the form $U(\phi)=1+8\pi G\xi \phi^2$] in the denominator leads to run-away solution for growing $V(\phi)$, corresponding to $V_{\mathrm{eff}}(\phi)$ decreasing for large $\phi$. This may happen if $U^2(\phi)$ grows more rapidly than $V$ with $\phi$ \cite{wevernov}. In the boundary case of $U^2(\phi) \sim V(\phi)$ for large $\phi$ the effective potential is asymptotically flat, which is the reason for the existence of viable Higgs inflation models \cite{Bezrukov}.
None of these features can exist for the effective potential in the form of $U(\phi)V(\phi)$ as in scalar-torsion theory. For the usual quadratic coupling (we study only this form of the coupling in the present paper) $V_{\mathrm{eff}}(\phi)$ is growing for an arbitrarly growing $V(\phi)$, so the dynamics inevitably pushes the scalar field toward zero, as we have seen in our previous paper. On the contrary, for the decreasing potentials studied in the present paper the effective potential can have a minimum, resulting in de Sitter solution. For a quadratic coupling this happens if $V(\phi)$ scales less steep than $\phi^{-2}$. This condition for a de Sitter solution to exist has been obtained earlier in \cite{we} and in a more general context in \cite{Laur}, and we now have a very easy way to explain it.
Steeper potentials result in unlimited growing of the scalar field, similar to the quintessence scenario in GR. We have identified a tracker solution which is always an attractor in its range of existence. It is shown that when the coupling constant $\xi$ is low enough for the tracker to exist (see the corresponding condition in Table 2), a vacuum regime is stable, so the scalar field dominates at late time. We also list several unstable asymptotic regimes. A full description of the scalar field dynamics requires matching solutions studied in the present paper with GR solutions valid for small $\phi$. We leave this for future work.
\section*{Acknowledgements}
~~~~The work was supported by RSF Grant \textnumero 16-12-10401 and by the Russian Government
Program of Competitive Growth of Kazan Federal University. Authors are grateful to Emmanuel Saridakis for discussions.
|
\section{Introduction}
\label{sec:intro}
A \emph{directed graph} (or simply \emph{digraph}) $D$ consists of a non-empty finite set
$V (D)$ of elements called \emph{vertices}
and a finite set $A(D)$ of ordered pairs of distinct vertices called \emph{arcs} (or \emph{directed edges}).
We call $V (D)$ the vertex set and $A(D)$ the arc set of $D$.
We will often denote $D$ as $D = (V,A)$.
For an arc $(u, v)$, the vertex $u$ is called the \emph{tail} and the vertex $v$ is called the \emph{head}.
The head and tail of an arc are called the \emph{end-vertices}.
The above definition of a digraph implies that we allow a digraph to have arcs with the same end-vertices
(for example, both $(u,v)$ and $(v,u)$ may be in $A$).
In this paper we only consider simple digraphs.
That is, we do not allow \emph{parallel} (also called \emph{multiple}) arcs, i.e.,
pairs of arcs with the same tail and the same head, or \emph{loops} (i.e., arcs whose heads and tails coincide).
When parallel arcs and loops are admissible we speak of \emph{directed pseudographs};
directed pseudographs without loops are \emph{directed multigraphs} (\cite{chartrand:1996}).
For more information about graphs and digraphs see, e.g., \cite{chartrand:1996}.
For a positive integer $n$,
let $[n]=\{1,2,\dots ,n\}$,
$\mathcal{D}_n$ denote the set of all digraphs with vertex set $[n]$ and
$2^{\mathcal{D}_n}$ denote the set of all subsets of $\mathcal{D}_n$.
A \emph{random digraph} is a probability space $(\mathcal{D}_n,2^{\mathcal{D}_n}, P)$,
and we write $\mathbf{D}= ( \mathcal{D}_n, P)$
where $P$ is a probability measure.
We call a random digraph as \emph{degenerate} if all the probability mass is on one digraph.
We can also think of $\mathbf{D}$ as
the outcome of an experiment of picking a digraph from $\mathcal{D}_n$ with distribution $P$.
For every $D\in \mathcal{D}_n$, we write $P(\{D\})$ as $P(D)$ for brevity in notation.
Also, for a measure space $(\Omega, \mathcal{F}, \mu )$,
$\mathcal{F}^n$ and $ \mu^n $ denote the usual product $\sigma$-algebra and product measure, respectively.
For the set of real numbers, we consider the Borel $\sigma$-algebra, and
throughout this paper we suppress the $\sigma$-algebra notation as long as
there is no necessity nor ambiguity.
\begin{example}
\label{ex:Dnm}
(Uniform Random Digraph Model)
For positive integers $n$ and $m$ with $n\geq 2$ and $0<m< n(n-1)$, $\mathbf{D}(n,m)$ is the random digraph such that
\begin{align*}
P(D) =
\begin{cases}
\frac{1}{{n(n-1) \choose m}}, & \text{if } |A(D)|=m \\
0, & {\rm otherwise}
\end{cases}
\end{align*}
for every $D\in \mathcal{D}_n$.
In other words, $\mathbf{D}(n,m)$ picks a digraph uniformly at random among the ones with vertex set $[n]$
and having exactly $m$ arcs.
Note that there are ${ n(n-1) \choose m}$ such digraphs, and $m$ is not chosen to be 0 or $n(n-1)$ to obtain a non-degenerate random digraph.
Also notice that $\mathbf{D}(n,m)$ is the digraph version of the Erd\H{o}s-R\'{e}nyi random graph $\mathbf{G}(n,m)$ (\cite{erdos:1959}).
For some asymptotic properties of uniform random digraphs see \cite{luczak:1990} and \cite{graham:2008}.
\end{example}
A digraph $D_1$ is \emph{isomorphic} to a digraph $D_2$ (or $D_1$ and $D_2$ are \emph{isomorphic}) if there is a bijection $f:V(D_1)\rightarrow V(D_2)$ such that
$(u,v)\in A(D_1)$ if and only if $(f(u),f(v))\in A(D_2)$.
\begin{definition}
(Isomorphism Invariance)
Let $\mathbf{D}=(\mathcal{D}_n,P)$ be a random digraph.
We say that $\mathbf{D}$ is \emph{isomorphism-invariant} if $P(D_1)=P(D_2)$
whenever $D_1$ and $D_2$ are isomorphic digraphs in $\mathcal{D}_n$.
\end{definition}
Throughout the article, we only consider non-degenerate isomorphism-invariant random digraphs.
We follow the isomorphism-invariant graph classification of \cite{beer:2011}
pointing out the similarities and differences until we introduce randomness in the direction.
In Section \ref{sec:2}, we introduce the arc random digraphs (ARDs), vertex random digraphs (VRDs)
and vertex-arc random digraphs (VARDs).
In Section \ref{sec:3}, for $n\geq 4$,
we prove that there is no random digraph which is both an ARD and a VRD,
and there exist VARDs which are neither ARDs nor VRDs.
Section \ref{sec:drd} introduces the direction random digraphs (DRDs), direction-edge random digraphs (DERDs),
direction-vertex random digraphs (DVRDs) and direction-vertex-edge random digraphs (DVERDs).
Section \ref{sec:derdvard} examines the relations of DERDs with ARDs and VARDs.
In particular, we show that ARDs are the only random digraphs which are both DERD and VARD for $n\geq 4$,
and any DERD with $n\leq 3$ is a VARD.
Section \ref{sec:rnnd} presents random nearest neighbor digraphs (RNNDs)
and determines where they fit in these classifications.
Discussion and conclusions are provided in Section \ref{sec:dis}.
A list of abbreviations used in the article is provided in Table \ref{tab:abbreviations}.
\section{Preliminaries}
\label{sec:prelim}
We first summarize isomorphism-invariant random graphs introduced by \cite{beer:2011}.
A \emph{graph} $G$ is a finite non-empty set $V(G)$ of elements called \emph{vertices} together with a set $E(G)$ of
unordered pairs of vertices of $G$ called \emph{edges}.
An edge $\{u,v\}$ is denoted by $uv$ for convenience in the text.
Let $\mathcal{G}_n$ denote the set of all graphs with $V(G)=[n]$ and
$2^{\mathcal{G}_n}$ be the set of all subsets of $\mathcal{G}_n$.
A \emph{random graph} is a probability space $(\mathcal{G}_n,2^{\mathcal{G}_n}, P)$,
and we write $\mathbf{G}= ( \mathcal{G}_n, P)$ where $P$ is a probability measure.
We write $P(G)$ instead of $P(\{G\})$ for brevity in notation
\begin{table}[t]
\centering
\begin{tabular}{llll}
ARD: & Arc Random Digraph &(p. \pageref{def:ard})\\
DERD: & Direction-Edge Random Digraph &(p. \pageref{def:derd})\\
DRD: & Direction Random Digraph &(p. \pageref{def:drd})\\
DVERD: & Direction-Vertex-Edge Random Digraph &(p. \pageref{def:dvrd})\\
DVRD: & Direction-Vertex Random Digraph &(p. \pageref{def:dvrd})\\
ERG: & Edge Random Graph &(p. \pageref{def:erg})\\
GARD: & Generalized Arc Random Digraph &(p. \pageref{def:gard})\\
RNND: & Random Nearest Neighbor Digraph &(p. \pageref{def:rnnd})\\
VARD: & Vertex-Arc Random Digraph &(p. \pageref{def:vard})\\
VERG: & Vertex-Edge Random Graph &(p. \pageref{def:verg})\\
VRD: & Vertex Random Digraph &(p. \pageref{def:vrd})\\
VRG: & Vertex Random Graph &(p. \pageref{def:vrg})\\
\end{tabular}
\caption{
\label{tab:abbreviations}
A list of abbreviations used in the article together with the page numbers where they are formally defined.}
\end{table}
The random graph model was first introduced by \cite{gilbert:1959} and \cite{erdos:1959}.
The model of Gilbert corresponds to edge random graph $\mathbf{G}(n,p_e)$ in \cite{beer:2011}
in which each edge is inserted, independent of others, with probability $p_e$.
The model introduced by Erd\H{o}s and R\'{e}nyi is the uniform random graph $\mathbf{G}(n,m)$ which
picks a graph with vertex set $[n]$ uniformly at random among the ones with exactly $m$ edges.
However, in the literature,
both of these models are usually called Erd\H{o}s-R\'{e}nyi model as they developed the theory.
A graph $G_1$ is \emph{isomorphic} to a graph $G_2$ (or $G_1$ and $G_2$ are \emph{isomorphic}) if there exists a bijection $f:V(G_1)\rightarrow V(G_2)$ such that
$uv \in E(G_1)$ if and only if $f(u)f(v)\in E(G_2)$.
We say that the random graph $\mathbf{G}=(\mathcal{G}_n,P)$ is \emph{isomorphism-invariant} if $P(G_1)=P(G_2)$ whenever $G_1$ is isomorphic to $G_2$.
\begin{definition}\label{def:erg}
An \emph{edge random graph} (ERG) is a random graph $\mathbf{G}(n,p_e)=(\mathcal{G}_n, P)$ where $p_e\in [0,1]$ and
\begin{align*}
P(G)=p_e^{|E(G)|} (1-p_e)^{{n\choose 2}-|E(G)|} \text{ for every } G\in \mathcal{G}_n .
\end{align*}
\end{definition}
Let $\Omega$ be a set, $\mathbf{x}=(x_1,\dots ,x_n) \in \Omega^n$ and
$\phi: \Omega \times \Omega \rightarrow \{0,1\}$ be a symmetric function.
Then the $(\mathbf{x},\phi )$\emph{-graph}, denoted $G(\mathbf{x},\phi)$, is defined to be the graph, $G$,
with vertex set $[n]$ such that for every $i,j\in [n]$ with $i\neq j$
we have $ij\in E(G)$ if and only if $ \phi (x_i,x_j)=1$.
\begin{definition}\label{def:vrg}
Let $(\Omega, \mathcal{F}, \mu )$ be a probability space and
$\phi : \Omega \times \Omega \rightarrow \{0,1\}$ be a symmetric measurable function.
The \emph{vertex random graph} (VRG), $\mathbf{G}(n, \Omega , \mu , \phi )$,
is the random graph $(\mathcal{G}_n, P)$ satisfying
\begin{align*}
P(G)= \int \mathbf{1}_{ \{ G(\mathbf{x},\phi )=G \} } d(\mu \mathbf{x}) \text{ for every } G\in \mathcal{G}_n,
\end{align*}
where $d(\mu \mathbf{x})$ is short-hand for the product integrator $d(\mu ^n (\mathbf{x}))=d(\mu x_1)\cdots d(\mu x_n)$.
\end{definition}
Notice that in a VRG the randomness lies in the structure attached to the vertices, and
once these random structures have been assigned to the vertices, all the edges are uniquely determined.
\begin{definition}\label{def:verg}
Let $(\Omega, \mathcal{F}, \mu )$ be a probability space and
$\phi :\Omega \times \Omega \rightarrow [0,1]$ be a symmetric measurable function.
The \emph{vertex-edge random graph} (VERG), $\mathbf{G}(n, \Omega, \mu ,\phi )$, is the random graph $(\mathcal{G}_n, P)$ with
\begin{align*}
P(G)= \int P_{\mathbf{x}}(G) d(\mu \mathbf{x}), \text{ for every } G\in \mathcal{G}_n,
\end{align*}
where for given $\mathbf{x}=(x_1,\dots ,x_n)$ and $G$
\begin{align*}
P_{\mathbf{x}}(G)=\prod_{ij\in E(G)} \phi (x_i,x_j) \times \prod_{ij\notin E(G)} (1- \phi (x_i,x_j)).
\end{align*}
\end{definition}
In words, a VERG is generated as follows: a random sample of size $n$ is drawn with distribution $\mu$ from $\Omega$,
say $\mathbf{X}=(X_1,\dots , X_n)$.
Then conditional on $\mathbf{X}$, independently for each pair of distinct vertices $i$ and $j$,
the edge $ij$ is inserted with probability $\phi (X_i,X_j)$.
Observe that the same notation $\mathbf{D}(n, \Omega,\mu, \phi)$ is used for both VRGs and VERGs.
However, this causes no confusion, since $\phi$ takes values in $\{0,1\}$ for VRGs and in $[0,1]$ for VERGs.
In other words, VRGs form a special case of VERGs with $\phi$ taking values only in $\{0,1\}$.
Therefore, every VRG is a VERG.
In addition, it is easy to see that letting $\phi$ to be identically equal to $p$ gives that every ERG is a VERG.
Let $\mathbf{G_1}=(\mathcal{G}_n,P_1)$ and $\mathbf{G_2}=(\mathcal{G}_n,P_2)$ be random graphs.
The \emph{total variation distance} between $\mathbf{G_1}$ and $\mathbf{G_2}$ is defined to be
\begin{align*}
d_{ \text{TV} } (\mathbf{G_1},\mathbf{G_2})= \frac{1}{2} \sum_{G\in \mathcal{G}_n} |P_1(G)-P_2(G)|.
\end{align*}
Similarly, for any two random digraphs $\mathbf{D_1}=(\mathcal{D}_n,P_1)$ and $\mathbf{D_2}=(\mathcal{D}_n,P_2)$,
the \emph{total variation distance} between $\mathbf{D_1}$ and $\mathbf{D_2}$ is defined to be
\begin{align*}
d_{ \text{TV} } (\mathbf{D_1},\mathbf{D_2})= \frac{1}{2} \sum_{D\in \mathcal{D}_n} |P_1(D)-P_2(D)|.
\end{align*}
\section{ARDs, VRDs and VARDs}
\label{sec:2}
\subsection{Arc random digraphs}
\label{subsec:ard}
One of the most commonly studied random digraphs is the binomial (or Bernoulli) random digraph model, $\mathbf{D}(n,p_a)$, in which each of the $n(n-1)$ possible arcs is included independently with probability $p_a$.
Such random digraphs give rise to arc random digraphs.
\begin{definition}\label{def:ard}
An \emph{arc random digraph} (ARD) is a random digraph $\mathbf{D}(n,p_a)=(\mathcal{D}_n, P)$ where $0<p_a<1$ and
\begin{align*}
P(D)=p_a^{|A(D)|} (1-p_a)^{n(n-1)-|A(D)|} \ \text{ for every } D\in \mathcal{D}_n .
\end{align*}
\end{definition}
Notice that ARDs are the digraph counterparts of random graphs $\mathbf{G}(n,p_e)$ due to \cite{gilbert:1959}.
For some asymptotic properties of $\mathbf{D}(n,p_a)$ see \cite{Karp:1990}, \cite{luczak:2009}, and \cite{Krivelevich:2013}.
\begin{definition}\label{def:gard}
Let $\mathrm{p_a}:[n]\times [n]\rightarrow [0,1]$ be a function (that is not necessarily symmetric in its arguments).
The \emph{generalized arc random digraph} (GARD), $\mathbf{D}(n,\mathrm{p_a})$, is the random digraph $(\mathcal{D}_n, P)$ with
\begin{align*}
P(D)=\prod_{(i,j)\in A(D)} \mathrm{p_a}(i,j) \times \prod_{(i,j)\notin A(D)} (1-\mathrm{p_a}(i,j)) \ \text{ for every } D\in \mathcal{D}_n.
\end{align*}
\end{definition}
In other words, in a GARD each arc appears independently of others and
the arc $(i,j)$ occurs with probability $\mathrm{p_a}(i,j)$.
Note that an ARD is special case of a GARD with a constant $\mathrm{p_a}$, i.e., $\mathrm{p_a}(i,j)=p_a$ for all $i,j$.
As the classical random digraph model $\mathbf{D}(n,p_a)$ may not fit real life networks,
inhomogeneous models like GARDs are of interest for such scenarios (see, e.g., \cite{bloznelis:2012}).
Clearly, any ARD is isomorphism-invariant.
The following proposition implies that a GARD is isomorphism-invariant if and only if it is an ARD.
\begin{prop}
\label{prop:gard}
Let $\mathbf{D}$ be an isomorphism-invariant GARD.
Then $\mathbf{D}=\mathbf{D}(n,p_a)$ for some $p_a$, i.e., $\mathbf{D}$ is an ARD.
\end{prop}
\begin{proof}
We show that $\mathrm{p_a}(i,j)=\mathrm{p_a}(k,l)$ for any two ordered pairs $(i,j)$ and $(k,l)$.
First note that
\begin{align}
\label{eq:gard1}
\mathrm{p_a}(i,j)=P((i,j)\in A(\mathbf{D}))=\sum_{(i,j)\in A(D)} P(D).
\end{align}
Fix a permutation on $[n]$ which maps $i$ to $k$ and $j$ to $l$.
Observe that this permutation induces a one-to-one correspondence between the sets $\{D\in\mathcal{D}_n: (i,j)\in A(D)\}$ and $\{D'\in\mathcal{D}_n: (k,l)\in A(D')\}$ such that matched digraphs are isomorphic.
As $\mathbf{D}$ is isomorphism-invariant, this correspondence implies
\begin{align}
\label{eq:gard2}
\sum_{(i,j)\in A(D)} P(D)=\sum_{(k,l)\in A(D')} P(D').
\end{align}
Hence, the result follows by \eqref{eq:gard1} and \eqref{eq:gard2}.
\end{proof}
\subsection{Vertex random digraphs}
\label{subsec:vrd}
Let $\Omega$ be a set, $\mathbf{x}=(x_1,\dots ,x_n) \in \Omega^n$ and
$\phi: \Omega \times \Omega \rightarrow \{0,1\}$ be a function.
Then the $(\mathbf{x},\phi )$\emph{-digraph}, denoted $D(\mathbf{x},\phi)$, is defined to be the digraph, $D$,
with vertex set $[n]$ such that for all $i,j\in [n]$ with $i\neq j$ we have
\begin{align*}
(i,j)\in A(D) \text{ if and only if } \phi (x_i,x_j)=1.
\end{align*}
Clearly, every digraph $D$ with $V(D)=[n]$ is an $(\mathbf{x},\phi)$-digraph for some choice of $\Omega,\mathbf{x}$ and $\phi$.
More specifically, choose $\mathbf{x}$ to be the identity function on $\Omega=[n]$ and
define $\phi (i,j)=\mathbf{1}_{ \{ (i,j)\in A(D) \} }$ where $\mathbf{1}_{\{\cdot \}}$ is the indicator function.
\begin{definition}\label{def:vrd}
Let $(\Omega, \mathcal{F}, \mu )$ be a probability space and
$\phi : \Omega \times \Omega \rightarrow \{0,1\}$ be a measurable function.
The \emph{vertex random digraph} (VRD), $\mathbf{D}(n, \Omega , \mu , \phi )$, is the random digraph $(\mathcal{D}_n, P)$ with
\begin{align*}
P(D)= \int \mathbf{1}_{ \{ D(\mathbf{x},\phi )=D \} } d(\mu \mathbf{x}) \text{ for every } D\in \mathcal{D}_n.
\end{align*}
\end{definition}
Note that in a VRD the randomness resides in the structure attached to the vertices, as in VRGs, and
when these random structures are assigned to the vertices, all the arcs are uniquely determined.
\begin{example}
Proximity Catch Digraphs (PCDs)( \cite{ceyhan:2011}):
Let $(\Omega, \mathcal{F}, \mu)$ be a probability space.
The \emph{proximity map} $N(\cdot )$ is a function from $\Omega$ to $\mathcal{F}$.
The \emph{proximity region} associated with $x\in \Omega$, denoted $N(x)$, is the image of $x\in \Omega$ under $N(\cdot )$.
The points in $N(x)$ are thought of as being ``closer'' to $x\in \Omega$ than the points in $\Omega\backslash N(x)$.
For a given $\mathbf{x}=(x_1,x_2,\dots ,x_n)$ the \emph{proximity catch digraph} is the digraph with the vertex set $V=[n]$
and the arc set $A=\{(i,j): x_j\in N(x_i)\}$.
In other words, we insert the arc $(i,j)$ if and only if $x_j$ is in the proximity region of $x_i$.
Note that for a given $N(\cdot)$, a random PCD is a VRD, $\mathbf{D}(n,\Omega,\mu,\phi)$,
with $\phi(x_i,x_j)=\mathbf{1}_{ \{x_j\in N(x_i)\} }$.
For instance, one can take $\Omega=\mathbb{R}$, $N(x)=[x,\infty)$ and $\phi(x,y)=\mathbf{1}_{ \{x\leq y \}}$.
\end{example}
\begin{example}
Random Intersection Digraphs (\cite{bloznelis:2010}):
Let $n$ and $m$ be positive integers,
and $\mu$ be a distribution on $2^{[m]}\times 2^{[m]}$ (ordered pairs of subsets of $[m]$).
Given two collections of subsets $S_1,\dots , S_n$ and $T_1,\dots , T_n$ of the set $[m]$,
define the intersection digraph with vertex set $[n]$ such that
the arc $(i,j)$ is present in the digraph whenever $S_i \cap T_j$ is nonempty for $i\neq j$.
$\mathbf{D}(n,m,\mu)$ is the random intersection digraph generated by independent and identically distributed pairs of random subsets $(S_i,T_i)$ under $\mu$, $1\leq i\leq n$.
Note that $\mathbf{D}(n,m,\mu)$ is a VRD with $\Omega =2^{[m]}\times 2^{[m]}$ and
$\phi((S,T),(S',T'))=\mathbf{1}_{ \{ S \cap T' \neq \emptyset \}}$.
\end{example}
By letting $\Omega=[0,1]$, $\mu$ be the uniform distribution over $[0,1]$ and $\phi(x,y)=\mathbf{1}_{\{x\leq p_a\}}$,
we see that every $\mathbf{D}(2,p_a)$ is a VRD.
Recall that in a VRD, $\mathbf{D}(n,\Omega,\mu, \phi)$, $\phi$ is not required to be symmetric.
However, if $\phi$ is a symmetric function, whenever we see the arc $(i,j)$ in $A(\mathbf{D})$,
we see the arc $(j,i)$ as well.
In this case, for every $D\in \mathcal{D}_n$ in which there exists $(i,j)\in A(D)$ with $(j,i)\notin A(D)$, we have $P(D)=0$.
On the other hand, in an ARD, $\mathbf{D}(n,p_a)$, we have $P(D)>0$ for every $D\in \mathcal{D}_n$.
Therefore, whenever $\phi$ is symmetric and nonconstant $\mu^2$-a.s.,
$\mathbf{D}(n,\Omega,\mu, \phi)$ is not an ARD.
For instance, one can take $\Omega=\mathbb{R}^d$, $\mu$ to be an a.e. continuous distribution and $\phi(x,y)=\mathbf{1}_{ \{||x-y||_d \leq r \} }$,
where $||\cdot ||_d$ is the usual Euclidean norm in $\mathbb{R}^d$ and $r$ is a fixed positive real number.
Notice that these random digraphs are random PCDs in which $N(x)$ is the closed ball with radius $r$ and center $x$.
If we consider symmetric arcs as one edge only,
these type of random digraphs reduce to what is called \emph{random geometric graphs}.
For more information about random geometric graphs see \cite{penrose:2003}.
\subsection{Vertex-arc random digraphs}
\label{subsec:vard}
We now generalize the random digraphs introduced in the previous two subsections by combining the structures where the randomness lies.
\begin{definition}\label{def:vard}
Let $(\Omega, \mathcal{F}, \mu)$ be a probability space and
$\phi :\Omega \times \Omega \rightarrow [0,1]$ be a measurable function.
The \emph{vertex-arc random digraph} (VARD), $\mathbf{D}(n, \Omega, \mu ,\phi )$, is the random digraph $(\mathcal{D}_n, P)$ with
\begin{align*}
P(D)= \int P_{\mathbf{x}}(D) d(\mu \mathbf{x}), \text{ for every } D\in \mathcal{D}_n,
\end{align*}
where for given $\mathbf{x}=(x_1,\dots ,x_n)$ and $D=(V,A)$
\begin{align*}
P_{\mathbf{x}}(D)=\prod_{(i,j)\in A} \phi (x_i,x_j) \times \prod_{(i,j)\notin A} (1- \phi (x_i,x_j)).
\end{align*}
\end{definition}
The construction of a VARD is almost same with VERGs.
A random sample of size $n$ is drawn with distribution $\mu$ from $\Omega$,
say $\mathbf{X}=(X_1,\dots , X_n)$, and
then conditional on $\mathbf{X}$, independently for each pair of distinct vertices $i$ and $j$,
the arc $(i,j)$ is inserted with probability $\phi (X_i,X_j)$.
Note that we use the same notation $\mathbf{D}(n, \Omega,\mu, \phi)$ for both VRDs and VARDs.
But, since $\phi$ takes values only 0 or 1 for VRDs and in $[0,1]$ for VARDs,
this causes no confusion.
Particularly, VRDs form a special case of VARDs with $\phi$ taking values only in $\{0,1\}$.
Therefore, every VRD is a VARD.
Moreover, it is easy to verify that letting $\phi$ to be identically equal to $p_a$ gives that every ARD is a VARD.
\begin{prop}
\label{prop:isoinv}
Every VARD is isomorphism-invariant.
\end{prop}
\begin{proof}
Let $\mathbf{D}(n, \Omega, \mu ,\phi )$ be a VARD and $D,D'\in \mathcal{D}_n$ be isomorphic digraphs.
Then there exists a permutation $\sigma$ on $[n]$ such that
\begin{align*}
(i,j)\in A(D) \ \Leftrightarrow \ (\sigma(i),\sigma(j))\in A(D').
\end{align*}
Let $\sigma^{-1}$ be the inverse of $\sigma$ and $\mathbf{y}=(y_1,\dots ,y_n)$ such that
$y_i=x_{\sigma^{-1}(i)}$ for all $1\leq i \leq n$, i.e., $x_i=y_{\sigma(i)}$ for all $1\leq i \leq n$.
Then note that
\begin{align}
P_{\mathbf{x}}(D)&=\prod_{(i,j)\in A(D)} \phi (x_i,x_j) \times \prod_{(i,j)\notin A(D)} (1- \phi (x_i,x_j)) \nonumber\\
&=\prod_{(i,j)\in A(D)} \phi (y_{\sigma(i)},y_{\sigma(j)}) \times \prod_{(i,j)\notin A(D)} (1- \phi (y_{\sigma(i)},y_{\sigma(j)})) \nonumber \\
&=\prod_{(\sigma(i),\sigma(j))\in A(D')} \phi (y_{\sigma(i)},y_{\sigma(j)}) \times \prod_{(\sigma(i),\sigma(j))\notin A(D')} (1- \phi (y_{\sigma(i)},y_{\sigma(j)})) \nonumber \\
&=\prod_{(i,j)\in A(D')} \phi (y_i,y_j) \times \prod_{(i,j)\notin A(D')} (1- \phi (y_i,y_j)) \nonumber \\
&=P_{\mathbf{y}}(D'). \label{eq:iso1}
\end{align}
As $\mathbf{y}$ is a permutation of $\mathbf{x}$, Fubini's theorem and \eqref{eq:iso1} imply that
\begin{align}
\label{eq:iso2}
P(D)=\int P_{\mathbf{x}}(D) \mu(d\mathbf{x})=\int P_{\mathbf{y}}(D') \mu(d\mathbf{y}).
\end{align}
Furthermore, the change of variables that maps $y_i$ to $x_i$ in the integrant above results
\begin{align}
\label{eq:iso3}
\int P_{\mathbf{y}}(D') \mu(d\mathbf{y})=\int P_{\mathbf{x}}(D') \mu(d\mathbf{x})=P(D'),
\end{align}
since the mapping is a permutation and the Jacobian of a permutation matrix is $\pm 1$.
Thus, the results in \eqref{eq:iso2} and \eqref{eq:iso3} together imply that $P(D)=P(D')$,
and so the desired result follows.
\end{proof}
As a corollary, we easily see that any VRD is isomorphism-invariant since every VRD is a VARD.
\section{Inclusion/exclusion relations between ARDs, VRDs and VARDs}
\label{sec:3}
In the previous section we have shown that every ARD is a VARD and so is every VRD,
and every VARD is isomorphism-invariant.
In this section we prove that for $n\geq 4$ there exists no random digraph which is both ARD and VRD,
and the union of the classes ARDs and VRDs is not the entire class of VARDs.
The following theorem implies that the families ARDs and VRDs are disjoint for $n\geq 4$.
\begin{thm}
\label{thm:1}
If an ARD, $\mathbf{D}(n,p_a)$, with $n\geq 4$ is represented as a VARD, $\mathbf{D}(n,\Omega, \mu, \phi)$,
then $\phi (x,y)=p_a$ $\mu^2$-a.s.
\end{thm}
\begin{proof}
Suppose that an ARD, $\mathbf{D}(n,p_a)$, with $n\geq 4$ is represented as a VARD, $\mathbf{D}(n,\Omega, \mu, \phi)$.
For the proof of the theorem,
we borrow some tools from functional analysis which are presented in the proof of the Theorem 4.2. in \cite{beer:2011}.
Let $h: \Omega\times \Omega \rightarrow [0,1]$ be a symmetric measurable function
and $T$ be the integral operator with kernel $h$ on the space $L^2(\Omega,\mu)$ of $\mu$-square-integrable functions on $\Omega$:
\begin{align*}
(Tg)(x)=\int h(x,y)g(y)d(\mu y).
\end{align*}
Since $h$ is bounded and $\mu$ is a finite measure, the kernel $h$ is in $L^2(\mu \times \mu )$.
Integral operators with such kernels are Hilbert-Schmidt operators and are thus compact operators.
Moreover, as $h$ is symmetric, the integral operator $T$ is self-adjoint,
which implies that $L^2(\Omega,\mu)$ has an orthonormal basis $(\psi_i)_{i\geq 1}$ of eigenfunctions for $T$
such that $T\psi _i=\lambda_i \psi_i$ for not necessarily distinct real eigenvalues $\lambda_i$
with $\lambda_i\rightarrow 0$ as $i\rightarrow \infty$ (see Chapter VI in \cite{reedsimon:1980}).
We may assume that $\lambda_1$ is the largest eigenvalue.
Then we have
\begin{align*}
h(x,y)=\sum_{i\geq 1} \lambda_i \psi_i(x)\psi_i(y) ~~~ \mu^2\text{-a.s.}
\end{align*}
with the sum converging in $L^2$.
As $\psi_i$'s are orthonormal, it follows that
\begin{align}
\label{eq:3.1}
&\E(h(X_1,X_2)h(X_2,X_3)h(X_3,X_4)h(X_4,X_1)) \nonumber\\
&=\int \int \int \int h(x_1,x_2)h(x_2,x_3)h(x_3,x_4)h(x_4,x_1)d(\mu x_1)d(\mu x_2)d(\mu x_3)d(\mu x_4) \nonumber \\
&=\sum_{i\geq 1} \lambda_i^4.
\end{align}
Now let $E_1$ be the event that both $(1,2)$ and $(2,1)$ are in $A(\mathbf{D})$.
As $\mathbf{D}$ is an ARD $\mathbf{D}(n,p_a)$, it is easy to see that $P(E_1)=p_a^2$.
On the other hand, since $\mathbf{D}$ is represented as a VARD, $\mathbf{D}(n,\Omega, \mu, \phi)$, we have
\begin{align*}
P(E_1)=\E(\phi(X_1,X_2)\phi(X_2,X_1)).
\end{align*}
Thus, letting $h(x,y)=\phi(x,y)\phi(y,x)$ gives $p_a^2=\E(h(X_1,X_2))$.
As
\begin{align*}
\E(h(X_1,X_2))=\int \int h(x,y)d(\mu x)d(\mu y)=\langle T\mathbf{1}, \mathbf{1} \rangle\leq \lambda_1,
\end{align*}
we get $p_a^2\leq \lambda_1$, where $\mathbf{1}$ is the function with constant value 1.
Let $E_2$ be the event that
$(1,2),(2,1),(2,3),(3,2),(3,4),(4,3),(4,1),(1,4)\in A(\mathbf{D})$.
Since $\mathbf{D}$ is an ARD, $\mathbf{D}(n,p_a)$, it is easy to see that
\begin{align}
\label{eq:3.3}
P(E_2)=p_a^8.
\end{align}
By the representation of $\mathbf{D}$ as a VARD, $\mathbf{D}(n,\Omega, \mu, \phi)$, we also have
\begin{align}
\label{eq:3.4}
P(E_2)=\E(h(X_1,X_2)h(X_2,X_3)h(X_3,X_4)h(X_4,X_1)).
\end{align}
Now combining the results in \eqref{eq:3.1}, \eqref{eq:3.3} and \eqref{eq:3.4} gives
\begin{align}
\label{eq:3.5}
p_a^8=\sum_{i\geq 1}\lambda_i^4.
\end{align}
Since $p_a^2\leq \lambda_1$, we have $p_a^8\leq \lambda_1^4$ and thus, by \eqref{eq:3.5} we obtain that $\lambda_1=p_a^2$
and $\lambda_i=0$ for every $i\geq 2$, that is $h(x,y)=p_a^2\psi_1(x)\psi_1(y)$.
But then we have
\begin{align*}
p_a^2\int \psi_1^2(x)d(\mu x)=p_a^2=\E(h(X_1,X_2))
=p_a^2\int \int \psi_1(x)\psi_1(y) d(\mu x)d(\mu y)=p_a^2 \left ( \int \psi_1(x)d(\mu x) \right )^2,
\end{align*}
which implies that
\begin{align}
\label{eq:3.6}
\int \psi_1^2(x)d(\mu x)= \left ( \int \psi_1(x)d(\mu x) \right )^2,
\end{align}
since $p\neq 0$.
As the equality in equation \eqref{eq:3.6} is the equality in the Cauchy-Schwarz inequality for $\psi_1$ and $\mathbf{1}$,
we see that $\psi_1$ is constant $\mu$-a.s.
Since $\int \psi_1^2(x)d(\mu x)=1$, we get $\psi_1=1$ $\mu$-a.s. or $\psi_1=-1$ $\mu$-a.s.,
and therefore $h(x,y)=p_a^2$ $\mu^2$-a.s., that is
\begin{align}
\label{eq:3.7}
\phi(x,y)\phi(y,x)=p_a^2~~~ \mu^2 \text{-a.s.}
\end{align}
Next, let $E_3$ be the event that neither of the arcs $(1,2)$ and $(2,1)$ is in $A(\mathbf{D})$,
and $E_4$ be the event that none of the arcs $(1,2),(2,1),(2,3),(3,2),(3,4),(4,3),(4,1),(1,4)$ is in $A(\mathbf{D})$.
Choosing $h(x,y)$ to be $(1-\phi(x,y))(1-\phi(y,x))$ allows us to follow the same arguments above for $E_1$ and $E_2$
replaced with $E_3$ and $E_4$, respectively, and with $1-p_a$ taking place of $p_a$.
Therefore, we obtain that
\begin{align}
\label{eq:3.8}
(1-\phi(x,y))(1-\phi(y,x))=(1-p_a)^2 ~~~ \mu^2 \text{-a.s.}
\end{align}
Finally, the equations in \eqref{eq:3.7} and \eqref{eq:3.8} give the desired result.
\end{proof}
\begin{remark}
The function $h$ in the proof of Theorem \ref{thm:1} is taken to be symmetric.
Otherwise, the operator $T$ needs not to be self-adjoint and
hence the succeeding arguments in the proof are not true.
So, $h$ being symmetric is a crucial condition for the proof.
If $\phi$ is given to be symmetric,
one can take $h=\phi$ and obtain $\phi=p_a$ $\mu^2$-a.s. (as in the proof of Theorem 4.2. in \cite{beer:2011}).
However, in our case, $\phi$ is not supposed to be symmetric.
Note that we tackle this hurdle by taking $h(x,y)$ to be $\phi(x,y)\phi(y,x)$ and
$(1-\phi(x,y))(1-\phi(y,x))$, respectively, and obtain the desired result.
\end{remark}
As any VRD is a VARD with $\phi$ taking values in $\{0,1\}$, by Theorem \ref{thm:1} we have the following corollary.
\begin{cor}
\label{cor:ardnotvrd}
Any ARD, $\mathbf{D}(n,p_a)$, with $n\geq 4$ is not a VRD.
\end{cor}
We next show that union of ARD and VRD families do not constitute the entire class of VARDs when $n\geq 4$.
\begin{thm}
\label{thm:onlyvard}
There exist VARDs with $n\geq 4$ which are neither a VRD nor an ARD.
\end{thm}
\begin{proof}
Let $0<a<b<1$ be real numbers.
Consider a VARD, $\mathbf{D}(n, \Omega,\mu, \psi)$, with $n\geq 4$ such that
$\phi(x,y)\in \{a,b\}$, and $\psi(x,y)\neq \psi(y,x)$ for any $x\neq y$.
Equivalently, we have
\begin{align}
\label{eq:psiab}
\psi(x,y)\psi(y,x)=ab \text{ and } (1-\psi(x,y)) (1-\psi(y,x))=(1-a)(1-b) \text{ for } x\neq y.
\end{align}
For example, one can take $\Omega=\mathbb{R}$, $\mu$ to be a continuous distribution and $\psi(x,y)=a\mathbf{1}_{ \{x\leq y\} } + b\mathbf{1}_{ \{y< x\} }$.
Now suppose that it has another VARD representation $\mathbf{D}(n, \Omega', \nu,\phi)$.
We claim that $\phi$ satisfies the same properties of $\psi$ given in \eqref{eq:psiab} $\nu^2$-a.s.
Recall that in the proof of Theorem \ref{thm:1}, the properties of an ARD, $\mathbf{D}(n,p_a)$, that we used are
\begin{align}
\label{eq:ardprops}
P(E_2)=p_a^8=(p_a^2)^4=(P(E_1))^4 \text{ and } P(E_4)=(1-p_a)^8=((1-p_a)^2)^4=(P(E_3))^4.
\end{align}
Notice that the equations in \eqref{eq:ardprops} hold for $\mathbf{D}(n, \Omega,\mu, \psi)$ when $p_a^2$ and $(1-p_a)^2$ are replaced with $ab$ and $(1-a)(1-b)$, respectively.
That is, for $\mathbf{D}(n, \Omega,\mu, \psi)$ we have
\begin{align*}
P(E_2)=(ab)^4=(P(E_1))^4 \text{ and } P(E_4)=((1-a)(1-b))^4=(P(E_3))^4.
\end{align*}
Therefore, following the same arguments in the proof of Theorem \ref{thm:1} we obtain
\begin{align*}
\phi(x,y)\phi(y,x)=ab \text{ and } (1-\phi(x,y))(1-\phi(y,x))=(1-a)(1-b)~~~ \nu^2 \text{-a.s.}
\end{align*}
and hence the claim follows.
Then, by the choice of $a,b$ and $\psi$, we see that $\mathbf{D}(n, \Omega,\mu, \psi)$ has neither an ARD nor a VRD representation.
\end{proof}
Recall that Corollary \ref{cor:ardnotvrd} and Theorem \ref{thm:onlyvard} imply that for $n\geq 4$
$\text{ARD}\cap \text{VRD}=\emptyset$ and $\text{VARD}\backslash (\text{ARD} \cup \text{VRD} ) \neq \emptyset$, respectively, in Figure \ref{fig:vards}.
\begin{figure}\center
\scalebox{.8}{
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.5cm,y=1.5cm]
\clip(0.8537834428690544,0.8212498658964582) rectangle (9.162062220187407,6.526727615465726);
\draw (1.,6.)-- (9.,6.);
\draw (9.,6.)-- (9.,1.);
\draw (9.,1.)-- (1.,1.);
\draw (1.,6.)-- (1.,1.);
\draw [rotate around={0.:(5.,3.3)}] (5.,3.3) ellipse (5.4cm and 3.1cm);
\draw [rotate around={0.:(3.5,3.3)}] (3.5,3.3) ellipse (1.8cm and 1.3cm);
\draw [rotate around={0.:(6.5,3.3)}] (6.5,3.3) ellipse (1.8cm and 1.3cm);
\draw (5,6.2) node[anchor=center] {Isomorphism-Invariant Random Digraphs};
\draw (5.,5.57) node[anchor=center] {VARD};
\draw (3.5,4.3) node[anchor=center] {ARD};
\draw (6.5,4.3) node[anchor=center] {VRD};
\end{tikzpicture}}
\caption{Venn diagram of vertex-arc random digraphs for $n\geq 4$.
The results of this paper show that $\text{ARD}\cap \text{VRD} = \emptyset$ and all the four regions in the figure are nonempty for $n\geq 4$.
}
\label{fig:vards}
\end{figure}
\begin{remark}
{\bf Approximation to VARDs by VRDs.}
However, any VARD can be arbitrarily closely approximated by VRDs.
That is, for any VARD $\mathbf{D}$ and $\epsilon>0$,
there exists a VRD $\mathbf{D}'$ such that $d_{ \text{TV} } (\mathbf{D},\mathbf{D}') < \epsilon$.
This result is a straightforward extension of approximation of VERGs by VRGs which immediately follows by letting
$\psi(y_1,y_2)$ to be the indicator of the event $\phi(x_1,x_2)\geq f_1(a_2)$ in the proof of Theorem 3.3 in \cite{beer:2011}.
\end{remark}
\section{Direction random digraphs}
\label{sec:drd}
One can also obtain isomorphism-invariant random digraphs by first generating an isomorphism-invariant random graph
and then assigning directions randomly to each edge.
Along this line, we first generate an isomorphism-invariant random graph, $\mathbf{G}= ( \mathcal{G}_n, P_{\mathbf{G}})$,
and then for each edge $ij\in E(\mathbf{G})$, independent of other edges,
pick a one sided or two sided direction randomly between $i$ and $j$.
For a given direction probability $1/2 \leq p_d <1$,
we put only the arc $(i,j)$ with probability $1-p_d$, only the arc $(j,i)$ with probability $1-p_d$ and
both of the arcs with probability $2p_d-1$.
Observe that the arc $(i,j)$ is put with probability $p_d$.
Also, note that we omit the case $p_d=1$ because it removes randomness in the direction.
The \emph{underlying graph} of a digraph $D$, denoted $U(D)$, is the graph obtained by replacing each arc of $D$ with an edge, disallowing multiple edges between two vertices (\cite{chartrand:1996}).
\begin{definition}
The \emph{underlying random graph} of a random digraph $\mathbf{D}= ( \mathcal{D}_n, P_{\mathbf{D}})$ is the random graph
$\mathbf{G}= ( \mathcal{G}_n, P_{\mathbf{G}})$ such that
\begin{align*}
P_{\mathbf{G}}(G)=\sum_{U(D)=G} P_{\mathbf{D}}(D) \text{ for every } G\in \mathcal{G}_n.
\end{align*}
\end{definition}
For instance, the underlying random graph of an ARD, $\mathbf{D}(n,p_a)$, is an ERG, namely,
$\mathbf{G}(n,p_e)$ with $p_e=2p_a-p_a^2$.
Moreover, notice also that the underlying random graph of a VARD, $\mathbf{D}(n, \Omega,\mu, \phi)$, is the VERG,
$\mathbf{G}(n, \Omega,\mu, \phi_u)$, where $\phi_u(x,y)=\phi(x,y)+\phi(y,x)-\phi(x,y)\phi(y,x)$.
In particular, the underlying random graph of a VRD is a VRG.
For a digraph $D\in \mathcal{D}_n$,
let $n_a(D)=|A(D)|$ and $n_e(D)=|E(U(D))|$ (i.e., the number of edges of the underlying graph of $D$).
Also, let $n_s(D)$ denote the number of pairs of vertices $i$ and $j$ such that both $(i,j)$ and $(j,i)$ are in $A(D)$
(i.e., the number of symmetric arcs in $D$),
and $n_{as}(D)$ denote the number of arcs $(i,j)$ in $A(D)$ with $(j,i) \notin A(D)$.
We write $n_a,n_e,n_s$ and $n_{as}$, respectively, dropping the digraph $D$ in the notation for brevity.
Note that $n_e=n_s+n_{as}$ and $n_a=2n_s+n_{as}$.
\begin{definition}\label{def:drd}
Let $\mathbf{G}= ( \mathcal{G}_n, P_{\mathbf{G}})$ be an isomorphism-invariant random graph and $1/2\leq p_d < 1$.
A \emph{direction random digraph} (DRD) is a random digraph $\mathbf{D} =( \mathcal{D}_n, P)$ with
\begin{align*}
P(D)=P_{\mathbf{G}}(U(D)) (1-p_d)^{n_{as}} (2p_d-1)^{n_s} \text{ for every } D\in \mathcal{D}_n ,
\end{align*}
and we say that $\mathbf{D}$ is \emph{generated by} $\mathbf{G}$ with \emph{direction probability} $p_d$.
\end{definition}
A natural question is why not start with a non-random graph and insert directions randomly to the edges to obtain DRDs.
There is a simple answer to the question.
Unfortunately, if directions are randomly inserted to the edges of a (fixed) graph,
the resulting random digraph is not isomorphism-invariant unless we start with an \emph{empty graph} (the graph with no edges) or a \emph{complete graph} (the graph with all possible edges).
Notice that $ \mathbf{G}$ is the underlying random graph of a DRD generated by $\mathbf{G}$.
Observe that if the digraphs $D_1$ and $D_2$ are isomorphic, then so are the (underlying) graphs $U(D_1)$ and $U(D_2)$,
and we also have $ n_s(D_1)=n_s(D_2)$ and $n_{as}(D_1)=n_{as}(D_2)$.
Thus, a DRD is isomorphism-invariant only if it is generated by an isomorphism-invariant random graph.
moreover, notice that we may consider a (fixed) graph as a degenerate random graph.
Also, it is easy to see that the empty graph and the complete graph with vertex set $[n]$ are the only graphs in $\mathcal{G}_n$ which are isomorphic to no other graph in $\mathcal{G}_n$,
and therefore these two graphs are the only isomorphism-invariant degenerate random graphs.
\subsection{DERDs, DVRDs and DVERDs}
We next provide three classes of direction random digraphs which are generated by ERGs, VRGs or VERGs.
\begin{definition}\label{def:derd}
The direction random digraph generated by an ERG, $\mathbf{G}(n,p_e)$, with direction probability $p_d$ is called
\emph{direction-edge random digraph} (DERD) and denoted $\mathbf{D}(n,p_e, p_d)$.
\end{definition}
Notice that letting $p_d$ to be 1/2 avoids symmetric arcs, and hence in the case of $p_d=1/2$,
after generating an ERG each edge is independently oriented in one of the two directions with equal probability
(e.g., see the model in \cite{Subramanian:2003}).
For example, letting $p_e=1$ and $p_d=1/2$ gives a \emph{random tournament} in which
each edge of a complete graph is independently oriented in one direction with equal probability.
For more information about tournaments, see \cite{moon:1968}.
\begin{definition}\label{def:dvrd}
A direction random digraph generated by a VRG is called \emph{direction-vertex random digraph} (DVRD).
A \emph{direction-vertex-edge random digraph} (DVERD) is a direction random digraph generated by a VERG.
\end{definition}
Notice that the underlying random graphs of a DERD, a DVRD and a DVERD are an ERG, a VRG and a VERG, respectively.
Clearly any ERG or VRG is a VERG, and hence every DERD and DVRD has a DVERD representation.
In addition, the results in \cite{beer:2011} imply the following:
A non-degenerate DRD which is both a DERD and a DVRD is either with $n\leq 3$ or generated by an ERG, $\mathbf{G}(n,p_e)$, with $p_e=1$.
For every $n\geq 6$, there exist DVERDs which are neither DERDs nor DVRDs.
Moreover, for $n\geq 3$, there exist DRDs which are not among DVERDs, and
for $n\leq 3$, any DVERD is also a DVRD.
These results are illustrated in Figure \ref{fig:drds}.
\begin{remark}
{\bf Approximation to DVERDs by DVRDs.}
Let $\mathbf{D}= ( \mathcal{D}_n, P_{\mathbf{D}})$ be a DVERD generated by a VERG,
$\mathbf{G}= ( \mathcal{G}_n, P_{\mathbf{G}})$, with direction probability $p_d$.
By Theorem 3.3 in \cite{beer:2011}, for any $\epsilon >0$ there exists a VRG, $\mathbf{G}'=( \mathcal{G}_n, P_{\mathbf{G}'})$, satisfying $d_{ \text{TV}}(\mathbf{G}, \mathbf{G}') < \epsilon$.
Let $\mathbf{D}'= ( \mathcal{D}_n, P_{\mathbf{D}'})$ be the DVRD generated by $\mathbf{G}'$ with the same direction probability $p_d$.
Then, it is easy to see that
\begin{align*}
\sum_{U(D)=G} |P_{\mathbf{D}}(D) -P_{\mathbf{D}'}(D)| = | P_{\mathbf{G}}(G)-P_{\mathbf{G}'}(G)|
\end{align*}
for every $G\in \mathcal{G}_n$,
and therefore we get $d_{ \text{TV}}(\mathbf{D}, \mathbf{D}')=d_{ \text{TV}}(\mathbf{G}, \mathbf{G}')$
which implies that the total deviation distance between $\mathbf{D}$ and $\mathbf{D}'$ is less than $\epsilon$.
\end{remark}
\begin{figure}\center
\scalebox{.8}{
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.5cm,y=1.5cm]
\clip(0.8537834428690544,0.8212498658964582) rectangle (9.162062220187407,6.526727615465726);
\draw (1.,6.)-- (9.,6.);
\draw (9.,6.)-- (9.,1.);
\draw (9.,1.)-- (1.,1.);
\draw (1.,6.)-- (1.,1.);
\draw [rotate around={0.:(5.,3.3)}] (5.,3.3) ellipse (5.4cm and 3.1cm);
\draw [rotate around={0.:(4.3,3.3)}] (4.3,3.3) ellipse (1.5cm and 1.3cm);
\draw [rotate around={0.:(5.7,3.3)}] (5.7,3.3) ellipse (1.5cm and 1.3cm);
\draw [rotate around={0.:(5.0,3.3)}] (5.,3.3) ellipse (3.3cm and 2.3cm);
\draw (5,6.2) node[anchor=center] {Isomorphism-Invariant Random Digraphs};
\draw (5.,5.57) node[anchor=center] {DRD};
\draw (4.3,4.3) node[anchor=center] {DERD};
\draw (5.7,4.3) node[anchor=center] {DVRD};
\draw (5,5.) node[anchor=center] {DVERD};
\end{tikzpicture}}
\caption{Venn diagram of direction random digraphs.
The results of the paper imply that all the six regions in the figure are nonempty.
In particular, the region $\text{DERD} \cap \text{DVRD}$ only consists of DRDs with $n\leq 3$ and DRDs generated by $\mathbf{G}(n,p_e=1)$.
}
\label{fig:drds}
\end{figure}
\section{Inclusion/exclusion relations of DERDs with respect to VARDs}
\label{sec:derdvard}
In this section, for $n\geq 4$, we show that a random digraph is both a DERD and a VARD if and only if it is an ARD, and
any DERD with $n\leq 3$ is also a VARD.
\begin{prop}
\label{prop:derdandard}
A DERD, $\mathbf{D}(n,p_e,p_d)$, is an ARD, $\mathbf{D}(n,p_a)$, if and only if
\begin{align*}
p_d=\frac{1}{1+\sqrt{1-p_e}} \text{ and } p_a=1- \sqrt{1-p_e}.
\end{align*}
\end{prop}
\begin{proof}
Suppose that $\mathbf{D}(n,p_e,p_d)$ is an ARD $\mathbf{D}(n,p_a)$.
Then we have $p_ep_d=p_a$ since both are $P((1,2)\in A(\mathbf{D}))$.
Similarly we have $p_e(2p_d-1)=p_a^2$ as both are $P( \{ (1,2), (2,1) \} \subset A(\mathbf{D}))$.
Solving these two equations gives $p_d=(1\pm \sqrt{1-p_e})/p_e$.
If $p_e=1$, then definitely $p_d=1$.
Otherwise, $(1- \sqrt{1-p_e})/p_e <1<(1+ \sqrt{1-p_e})/p_e$ and hence $p_d=(1- \sqrt{1-p_e})/p_e$.
Note that $(1- \sqrt{1-p_e})/p_e= 1/(1+ \sqrt{1-p_e})$ and so $1/2\leq p_d \leq 1$.
Finally, $p_a=p_e p_d=1- \sqrt{1-p_e}$.
We next show that whenever $p_d=1/(1+\sqrt{1-p_e})$ and $p_a=1- \sqrt{1-p_e}$,
$\mathbf{D}(n,p_e,p_d)$ is $\mathbf{D}(n,p_a)$.
Note that in that case, $1-p_d=p_a(1-p_a)/p_e$, $2p_d-1=p_a^2/p_e$ and $1-p_e=(1-p_a)^2$.
Therefore, for a given $D\in \mathcal{D}_n$ we have
\begin{align*}
P(D)&=p_e^{n_e} (1-p_e)^{{n\choose 2}-n_e} (1-p_d)^{n_{as}} (2p_d-1)^{n_s}\\
&=p_e^{n_e} (1-p_e)^{(n(n-1)-2n_e)} \frac{p_a^{n_{as}}(1-p_a)^{n_{as}}}{p_e^{n_{as}}} \frac{p_a^{2n_s}}{p_e^{n_s}}\\
&=p_e^{n_e-n_{as}-n_s} p_a^{n_{as}+2n_s} (1-p_a)^{n(n-1)-2n_e+n_{as}}\\
&=p_a^{n_a} (1-p_a)^{n(n-1)-n_a},
\end{align*}
since $n_e=n_{as}+n_s$ and $n_a=n_{as}+2n_s$.
Thus, the desired result follows.
\end{proof}
In fact, for $n\geq4$, the family of ARDs is the intersection of the classes DERDs and VARDs.
\begin{thm}
\label{thm:derdandvard}
If a DERD, $\mathbf{D}(n,p_e,p_d)$, with $n\geq 4$ has a VARD representation $\mathbf{D}(n, \Omega,\mu, \phi)$,
then $p_d = 1/(1+ \sqrt{1-p_e})$ and $\phi(x,y)=p_ep_d\ $ $\mu^2$-a.s.
\end{thm}
\begin{proof}
Suppose $\mathbf{D}(n,p_e,p_d)$ has a VARD representation $\mathbf{D}(n, \Omega,\mu, \phi)$.
Note that, as in any ARD, the events $(i,j)\in A(\mathbf{D})$ and $(k,l)\in A(\mathbf{D})$ are independent in a DERD whenever $\{i,j\}\neq \{k,l\}$.
Therefore, one can apply the method used in the proof of Theorem \ref{thm:1} and obtain
\begin{align}
\label{eq:derd1}
\phi(x,y) \phi(y,x)=p_e(2p_d-1) ~~~ \mu^2 \text{-a.s.}
\end{align}
and
\begin{align}
\label{eq:derd2}
(1-\phi(x,y))(1-\phi(y,x))=1-p_e ~~~ \mu^2 \text{-a.s.}
\end{align}
Solving the equations in \eqref{eq:derd1} and \eqref{eq:derd2} yields
\begin{align}
\phi(x,y)+\phi(y,x)=2p_ep_d ~~~ \mu^2 \text{-a.s.} \label{eq:derdsum}
\end{align}
and
\begin{align}
\phi (x,y)=p_ep_d \pm \sqrt{(1-\sqrt{1-p_e}-p_e p_d) (1+\sqrt{1-p_e}-p_e p_d)}~~~ \mu^2 \text{-a.s.} \label{eq:derd3}
\end{align}
If $p_d > 1/(1+\sqrt{1-p_e})=(1-\sqrt{1-p_e})/p_e$,
then the numbers in the right-hand side of \eqref{eq:derd3} have imaginary parts,
and hence we get a contradiction since $\phi$ takes only real values.
Suppose $p_d \leq 1/(1+\sqrt{1-p_e})$.
In any VARD, $\mathbf{D}(n,\Omega,\mu,\phi)$, with $n\geq 3$ we have
\begin{align}
P(\{ (1,2),(1,3)\}\subset A(\mathbf{D}))&=\int \int \int \phi(x_1,x_2)\phi(x_1,x_3)d(\mu x_1)d(\mu x_2)d(\mu x_3) \nonumber \\
&=\int \left( \int \phi(x_1,x_2)d(\mu x_2) \right ) \left( \int \phi(x_1,x_3)d(\mu x_3) \right )d(\mu x_1) \nonumber \\
&=\int \left( \int \phi(x_1,x_2)d(\mu x_2) \right )^2 d(\mu x_1) \nonumber \\
&\geq \left ( \int \int \phi(x_1,x_2)d(\mu x_1)d(\mu x_2) \right )^2=\left (P((1,2)\in A(\mathbf{D})) \right )^2 \label{eq:derd4}
\end{align}
by Fubini's theorem and
the Cauchy-Schwarz inequality applied to the constant function $\mathbf{1}$ and $\int \phi(x_1,x_2)d(\mu x_2)$.
On the other hand, in any DERD, $\mathbf{D}(n,p_e,p_d)$, with $n\geq 3$ we have
\begin{align*}
P( \{ (1,2),(1,3)\} \subset A(\mathbf{D}))=(p_e p_d)^2=\left (P((1,2)\in A(\mathbf{D})) \right )^2.
\end{align*}
Therefore, we have the equality in the Cauchy-Schwarz inequality in \eqref{eq:derd4}.
Thus, $\int \phi(x_1,x_2)d(\mu x_2)=c$ $\mu$-a.s. for some constant $c$.
Since
\begin{align*}
p_ep_d=P((1,2)\in A(\mathbf{D}))=\int \int \phi(x_1,x_2)d(\mu x_2)d(\mu x_1)=\int c\ d(\mu x_1)=c,
\end{align*}
we obtain $c=p_ep_d$, that is,
\begin{align}
\int \phi(x,y)d(\mu y)=p_ep_d \ \ \mu \text{-a.s.} \label{eq:derdint1}
\end{align}
Similarly, in a VARD, $\mathbf{D}(n,\Omega,\mu,\phi)$, with $n\geq 4$ we have
\begin{align}
&P( \{ (1,2),(2,3), (1,4), (4,3)\} \subset A(\mathbf{D})) \nonumber \\
&=\int \int \int \int \phi(x_1,x_2)\phi(x_2,x_3) \phi(x_1,x_4)\phi(x_4,x_3) d(\mu x_1)d(\mu x_2)d(\mu x_3) d(\mu x_4) \nonumber \\
&=\int \int \left( \int \phi(x_1,x_2) \phi(x_2,x_3)d(\mu x_2) \right ) \left( \int \phi(x_1,x_4) \phi(x_4,x_3)d(\mu x_4) \right )d(\mu x_1) d(\mu x_3) \nonumber \\
&=\int \int \left( \int \phi(x_1,x_2)\phi(x_2,x_3)d(\mu x_2) \right )^2 d(\mu x_1) d(\mu x_3) \nonumber \\
&\geq \left ( \int \int \int \phi(x_1,x_2) \phi(x_2,x_3) d(\mu x_1)d(\mu x_2)d(\mu x_3) \right )^2
=\left (P( \{ (1,2),(2,3)\} \subset A(\mathbf{D})) \right )^2 \label{eq:derd5}
\end{align}
by Fubini's theorem and
the Cauchy-Schwarz inequality applied to the constant function $\mathbf{1}$ and $\int \phi(x_1,x_2)\phi(x_2,x_3)d(\mu x_2) $. Since in a DERD, $\mathbf{D}(n,p_e,p_d)$, with $n\geq 4$ we have
\begin{align*}
P( \{ (1,2),(2,3), (1,4), (4,3)\} \subset A(\mathbf{D}))=(p_e p_d)^4=\left (P( \{(1,2), (2,3)\} \subset A(\mathbf{D})) \right )^2,
\end{align*}
we obtain $\int \phi(x_1,x_2)\phi(x_2,x_3)d(\mu x_2)$ is constant $ \mu^2$-a.s.
by the equality in Cauchy-Schwarz inequality in \eqref{eq:derd5}.
By the equality in \eqref{eq:derd5}, one can easily verify that
\begin{align}
\int \phi(x,y)\phi(y,z)d(\mu y)=(p_ep_d)^2\ \ \mu^2\text{-a.s.} \label{eq:derdint2}
\end{align}
Let $s(x,y)=i(\phi (x,y)-p_ep_d)$.
Combining the results in \eqref{eq:derdsum}, \eqref{eq:derd3}, \eqref{eq:derdint1} and \eqref{eq:derdint2} gives
\begin{align}
s(x,y)=i(\phi (x,y)-p_ep_d)=-i(\phi (y,x)-p_ep_d)=\overline{s(y,x)} \ \ \mu^2\text{-a.s.} \label{eq:derds1}
\end{align}
and
\begin{align}
\int s(x,y)s(y,z)d(\mu y)=0\ \ \mu^2\text{-a.s.} \label{eq:derds2}
\end{align}
Let $T$ be the integral operator with kernel $s$ on the space $L^2(\Omega,\mu)$
\begin{align*}
(Tg)(x)=\int s(x,y)g(y)d(\mu y).
\end{align*}
Since $s$ is bounded and $\mu$ is a finite measure, the kernel $s$ is in $L^2(\mu \times \mu )$.
Moreover, the integral operator $T$ is compact and self-adjoint by \eqref{eq:derds1},
which implies that $L^2(\Omega,\mu)$ has an orthonormal basis $(\psi_i)_{i\geq 1}$ of eigenfunctions for $T$
such that $T\psi _i=\lambda_i \psi_i$ for not necessarily distinct eigenvalues $\lambda_i$, and
\begin{align}
s(x,y)=\sum_{i\geq 1} \lambda_i \psi_i(x)\overline{\psi_i(y)} ~~~ \mu^2\text{-a.s.} \label{eq:derds3}
\end{align}
with the sum converging in $L^2$ (\cite{reedsimon:1980}).
Since $\psi_i$'s are orthonormal, the equations \eqref{eq:derds2} and \eqref{eq:derds3} imply
\begin{align}
\sum_{i\geq 1} \lambda_i^2 \psi_i(x)\overline{\psi_i(z)}=0 ~~~ \mu^2\text{-a.s.} \label{eq:derds4}
\end{align}
Therefore, for any $m\geq 1$, by multiplying the equation in \eqref{eq:derds4} by $\psi_m(z)\overline{\psi_m(x)}$ and integrating over $x$ and $z$ we obtain $\lambda_m^2=0$, i.e., $\lambda_m=0$ for each $m$.
Thus, $s(x,y)=0\ \mu^2$-a.s. and hence $\phi(x,y)=p_ep_d \ \mu^2$-a.s. which implies $p_d = 1/(1+\sqrt{1-p_e})$.
\end{proof}
\begin{remark}
Notice that Theorem \ref{thm:derdandvard} and Proposition \ref{prop:derdandard} together imply Theorem \ref{thm:1}.
However, we provide the proof of Theorem \ref{thm:1} to keep the proof of Theorem \ref{thm:derdandvard} shorter and also to point out similarities and the differences with the techniques used in \cite{beer:2011}.
\end{remark}
\begin{figure}\center
\scalebox{.8}{
\begin{tikzpicture}[line cap=round,line join=round,>=triangle 45,x=1.5cm,y=1.5cm]
\clip(0.8537834428690544,2.4212498658964582) rectangle (9.162062220187407,7.7);
\draw (1.,6.5)-- (9.,6.5);
\draw (9.,6.5)-- (9.,2.5);
\draw (9.,2.5)-- (1.,2.5);
\draw (1.,6.5)-- (1.,2.5);
\draw [rotate around={0.:(3.6,4.3)}] (3.6,4.3) ellipse (3.5cm and 2.1cm);
\draw [rotate around={0.:(6.4,4.3)}] (6.4,4.3) ellipse (3.5cm and 2.1cm);
\draw [rotate around={0.:(7.3,4.3)}] (7.3,4.3) ellipse (1.5cm and 1.3cm);
\draw (5,6.7) node[anchor=center] {Isomorphism-Invariant Random Digraphs};
\draw (5.,4.3) node[anchor=center] {ARD};
\draw (3.6,5.9) node[anchor=center] {DERD};
\draw (6.4,5.9) node[anchor=center] {VARD};
\draw (7.3,5.3) node[anchor=center] {VRD};
\end{tikzpicture}}
\caption{Venn diagram of DERDs and VARDs for $n\geq 4$.
The results in this paper indicate that all the five regions in the figure are nonempty.
In addition, the intersection of the classes DERDs and VARDs is the family of ARDs,
i.e., $\text{DERD}\cap \text{VARD}= \text{ARD}$.
}
\label{fig:derdvard}
\end{figure}
However, for $n\leq 3$, any DERD has a VARD representation.
\begin{thm}
\label{thm:derd3isvard}
Any DERD, $\mathbf{D}(n,p_e,p_d)$, with $n\leq 3$ is also a VARD.
\end{thm}
\begin{proof}
Let $\oplus$ and $\ominus$ denote addition and subtraction modulo 1, respectively.
In other words, for real numbers $0\leq x,y <1$,
\begin{align*}
x\oplus y =
\begin{cases}
x+y, & \text{if } x+y<1 \\
x+y-1, & \text{if } x+y \geq 1
\end{cases}
\end{align*}
and
\begin{align*}
x\ominus y =
\begin{cases}
x-y, & \text{if } x-y\geq 0 \\
x-y+1, & \text{if } x-y< 0
\end{cases}
\end{align*}
If $U_1,U_2,U_3$ are independent uniform random variables over $[0,1)$,
then so are $U_1\oplus U_2, U_2\oplus U_3$ and $U_3\oplus U_1$ (see Lemma 4.5 in \cite{beer:2011}).
Therefore, $\mathbf{G}(3, p_e)$ can be represented as a vertex random graph $\mathbf{G}(3,[0,1), \nu,f)$ where
$\nu$ is the uniform distribution on $[0,1)$ and $f(x,y)=\mathbf{1}_{ \{ x\oplus y \leq p_e \}}$ (\cite{beer:2011}).
Let $g(x,y)=\mathbf{1}_{ \{x\ominus y\leq 1/2 \}}+(2p_d-1)\mathbf{1}_{ \{ x\ominus y>1/2 \} }$ for every $0\leq x,y <1$.
We claim that $\mathbf{D}(3,p_e,p_d)$ is a VARD, $\mathbf{D}(3,\Omega ,\mu, \phi)$, where
$\Omega=[0,1)\times [0,1)$, $\mu$ is product of two uniform distributions on $[0,1)$ and
$\phi( (u_1,u_1'), (u_2,u_2'))=f(u_1,u_2)g(u_1',u_2')$.
First note that
\begin{align}
g(x,y)+g(y,x)=2p_d \ \text{ and } \ g(x,y)g(y,x)=2p_d-1, \label{eq:derd3n1}
\end{align}
for every $ 0\leq x,y<1$.
As $f$ is a symmetric indicator function, the equations in \eqref{eq:derd3n1} imply
\begin{align}
\phi( (u_1,u_1'), (u_2,u_2')) \phi( (u_2,u_2'), (u_1,u_1')) &=f(u_1,u_2)(2p_d-1), \label{eq:derd3n2}\\
\phi( (u_1,u_1'), (u_2,u_2')) (1- \phi( (u_2,u_2'), (u_1,u_1'))) &=f(u_1,u_2)( g(u_1',u_2')-(2p_d-1)), \label{eq:derd3n3} \\
(1-\phi( (u_1,u_1'), (u_2,u_2'))) (1- \phi( (u_2,u_2'), (u_1,u_1'))) &=1-f(u_1,u_2). \label{eq:derd3n4}
\end{align}
We next focus on the function $g$.
It is easy to see that
\begin{align}
\int_0^1 g(x,y)~ dy= \int_0^1 g(x,y)~ dx=\frac{1}{2} 1+\frac{1}{2} (2p_d-1)=p_d, \label{eq:derd3n5}
\end{align}
for every $0\leq x,y<1$.
Consider the circle obtained by identifying the end points of the interval $[0,1]$ such that
$1/4$ is on the arc that starts from 0 and ends at $1/2$ along the clockwise direction.
Then, $x\ominus y$ is equal to the length of the arc of this circle
which starts from $x$ and ends at $y$ along the counterclockwise direction.
Notice that $g(x,y)g(y,z)$ and $g(x,y)g(y,z)g(z,x)$ depends on the ordering of $x,y,z$ along counterclockwise direction and
whether the points $x,y,z$ form an acute or obtuse triangle.
If $x,y,z$ form an acute triangle, there are basically two cases for the ordering, $x,y,z$ or $x,z,y$.
In the first case, $g(x,y)g(y,z)=(2p_d-1)^2$, and in the latter case $g(x,y)g(y,z)=1^2=1$.
If $x,y,z$ form an obtuse triangle, all six permutations of $x,y,z$ (x,y,z; x,z,y; y,x,z; y,z,x; z,x,y; z,y,x) are possible with the point at the middle corresponding to the obtuse angle.
Then, we have $g(x,y)g(y,z)= (2p_d-1)^2, (2p_d-1), (2p_d-1), (2p_d-1), (2p_d-1), 1$, respectively.
Moreover, it easy to show that three uniformly at random points on the circle form an acute triangle with probability $1/4$.
Therefore, we obtain
\begin{align}
\int_{[0,1)^3} g(x,y)g(y,z) ~ dx dy dz =\frac{1}{4}\cdot \frac{1}{2}((2p_d-1)^2+1)+
\frac{3}{4} \cdot \frac{1}{6} ((2p_d-1)^2+4(2p_d-1)+1) =p_d^2. \label{eq:derd3n6}
\end{align}
Similarly, we have
\begin{align}
\int_{[0,1)^3} g(x,y)g(y,z)g(z,x) ~ dx dy dz =\frac{1}{8}((2p_d-1)^3+1)+
\frac{1}{8} (3(2p_d-1)^2+3(2p_d-1)) =p_d^3. \label{eq:derd3n7}
\end{align}
By using the results in \eqref{eq:derd3n1}-\eqref{eq:derd3n7}, one can easily verify that
\begin{align*}
\int P_{\mathbf{x}}(D) d(\mu \mathbf{x})=p_e^{n_e} (1-p_e)^{3-n_e} (1-p_d)^{n_{as}} (2p_d-1)^{n_s},
\end{align*}
for every $D\in \mathcal{D}_3$, and hence the desired result follows.
Furthermore, the same setting works for $n=2$ as well.
\end{proof}
\begin{remark}
{\bf Is every DERD with $n=3$ a VRD?}
Note that the function $\phi$ constructed in the proof of Theorem \ref{thm:derd3isvard} is binary (only takes the values 0 or 1) if and only if $p_d=1/2$.
Hence, by Theorem \ref{thm:derd3isvard} we see that any $\mathbf{D}(3,p_e,1/2)$ has a VRD representation.
But, by Proposition \ref{prop:derdandard}, $\mathbf{D}(3,p_e,1/2)$ is an ARD only if $p_e=0$ which gives a degenerate random digraph, and hence $\mathbf{D}(3,p_e,1/2)$ does not yield a non-degenerate ARD.
Other than the degenerate ones, is there any DERD $\mathbf{D}(3,p_e,p_d)$ with $p_d>1/2$ which is also a VRD?
Furthermore, is there an ARD with $n=3$ which has a VRD representation?
For now, these questions remain to be open, but, we conjecture that any DERD with $n=3$ is also a VRD.
\end{remark}
However, for $n=2$ the families DERDs and VRDs coincide with the all isomorphism-invariant random digraphs.
Let $D_1,D_2,D_3$ and $D_4$ be the digraphs with vertex set $[2]$ which only has the arc (1,2), only the arc (2,1), both of the arcs and none of the arcs, respectively.
Note that to obtain an isomorphism-invariant random digraph necessary and sufficient condition is $P(D_1)=P(D_2)$.
Let $\mathbf{D}$ be the random digraph with $P(D_1)=P(D_2)=p_1$, $P(D_3)=p_2$ and $P(D_4)=1-2p_1-p_2$.
First observe that $\mathbf{D}$ is an ARD if and only if $\sqrt{p_2}(1-\sqrt{p_2})=p_1$.
Letting $p_e=2p_1+p_2$ and $p_d=(p_1+p_2)/(2p_1+p_2)$ gives that $\mathbf{D}$ is a DERD $\mathbf{D}(2,p_e,p_d)$.
With the same $p_e$ and $p_d$,
let $\phi( (u_1,u_1'), (u_2,u_2'))=\mathbf{1}_{ \{u_1\oplus u_1' \leq p_e\} } \mathbf{1}_{ \{ u_2\ominus u_2' \leq p_d \} }$.
Then, it is easy to see that $\mathbf{D}$ is also VRD $\mathbf{D}(2,\Omega ,\mu, \phi)$, where
$\Omega=[0,1)\times [0,1)$, $\mu$ is the uniform distribution on $[0,1)^2$.
Therefore, when $n=2$, any isomorphism-invariant random digraph is both a DERD and a VRD (hence also a VARD).
\begin{remark}
{\bf Positive Dependence:}
Recall that by the inequality in \eqref{eq:derd4}, for any {\rm VARD} $\mathbf{D}$ we have the positive dependence
\begin{align}
P( \{ (1,2),(1,3)\} \subset A(\mathbf{D}))\geq P((1,2)\in A(\mathbf{D}))P((1,3)\in A(\mathbf{D}))= P((1,2)\in A(\mathbf{D}))^2. \label{eq:posdep}
\end{align}
Furthermore, the inequality in \eqref{eq:posdep} can be generalized by H\"{o}lder's inequality as follows
\begin{align}
P( \{ (1,2),\dots ,(1,m)\} \subset A(\mathbf{D}))\geq \prod_{i=2}^m P((1,i)\in A(\mathbf{D})) =
P((1,2)\in A(\mathbf{D}))^{m-1} \label{eq:posdepm}
\end{align}
for every {\rm VARD} $\mathbf{D}$ and $2\leq m \leq n$.
Similarly, we have the same inequality in \eqref{eq:posdepm} for any {\rm DVERD} as well and
note that equality holds for every {\rm DERD}.
However, there are random digraphs other than {\rm DERD}s satisfying equality in \eqref{eq:posdepm} for each $m$.
For example, consider the {\rm VRD}, $\mathbf{D}(n,[0,1),\mu,\phi)$,
where $\mu$ is the uniform distribution over $[0,1)$ and $\phi(x,y)=\mathbf{1}_{ \{x\ominus y\geq 3/8 \}} \mathbf{1}_{ \{y\ominus x\geq 3/8 \}}$.
Clearly, in this case, we have $P( \{ (1,2), \dots ,(1,m) \} \subset A(\mathbf{D}))=(1/4)^{m-1}$ and
$P((1,i)\in A(\mathbf{D}))=1/4$ for each $i$.
Also, it is easy to verify that $\mathbf{D}(n,[0,1),\mu,\phi)$ has no {\rm DERD} representation since
$P( \{ (1,2),(1,3), (2,3) \} \subset A(\mathbf{D}))=0$.
In the same manner, one can easily obtain similar results for random graphs.
In other words, for any {\rm VERG}, $\mathbf{G}(n,\Omega,\mu,\phi)$, and $2\leq m \leq n$, we have
\begin{align}
P( \{ \{1,2\}, \{1,3\}, \dots , \{1,m\} \} \subset E(\mathbf{G}))\geq P( \{1,2\} \in E(\mathbf{G}))^{m-1}, \label{eq:vergposdepm}
\end{align}
and equality holds for every {\rm ERG}.
The underlying random graph of the {\rm VRD}, $\mathbf{D}(n,[0,1),\mu,\phi)$, described above is an example for
random graphs with no {\rm ERG} representation which attains equality in \eqref{eq:vergposdepm} for every $m$.
\end{remark}
\section{Where do random nearest neighbor digraphs reside?}
\label{sec:rnnd}
We determine the class relationship for one of the most commonly studied random digraphs, namely,
random nearest neighbor (NN) digraphs (e.g., see \cite{friedman:1983}, \cite{eppstein:1997},\cite{cuzick:1990} and \cite{penrose:2001}).
Let $n\geq 3, k \geq 1$ and $d\geq 1$ be integers with $k<n-1$.
Let $\mu$ be a probability distribution over $\mathbb{R}^d$
with density function $f$ that is assumed to be continuous almost everywhere with respect to Lebesgue measure.
Let $|\cdot|$ denote a fixed norm on $\mathbb{R}^d$ and $X=(X_1,\dots ,X_n)$ be i.i.d. vectors in $\mathbb{R}^d$ drawn from $\mu$.
For given $\mathbf{x}=(x_1,\dots ,x_n)$, the set of $k$ \emph{nearest neighbors} ($k$NNs) of $x_i$ is the closest $k$ points to $x_i$
among the points $\{x_1,\dots ,x_n\}\backslash \{x_i\}$ with respect to the given norm $|\cdot |$
and denoted as $k\text{NN}_{\mathbf{x}}(x_i)$.
As the occurrence of a tie is an event with zero probability for points from an a.e. continuous $f$,
we may assume that $k\text{NN}_{\mathbf{x}}(x_i)$ is well defined for each $i$ with probability 1.
The $k$ \emph{nearest neighbor digraph} of $\mathbf{x}$ is the digraph with vertex set $V=[n]$
and the arc set $A=\{(i,j): x_j \in k\text{NN}_{\mathbf{x}}(x_i) \}$,
(i.e., the arc $(i,j)$ is inserted if and only if
$x_j$ is one of the $k$NNs of $x_i$) and denoted as $k\text{NND}(\mathbf{x})$.
\begin{definition}\label{def:rnnd}
The \emph{random nearest neighbor digraph} (RNND) is the random digraph $\mathbf{D}(n,[k],d, \mu, |\cdot |)$ with
\begin{align*}
P(D)=\int \mathbf{1}_{ \{ k\text{NND}(\mathbf{x})=D \}} d(\mu \mathbf{x} ) \text{ for every } D\in \mathcal{D}_n.
\end{align*}
\end{definition}
Notice that we picked $k$ to be less than $n-1$, because otherwise, we obtain a degenerate random digraph.
\begin{prop}
Every RNND is isomorphism-invariant.
\end{prop}
\begin{proof}
Let $\mathbf{D}(n,[k],d,\mu, |\cdot |)$ be a RNND and $D,D'\in \mathcal{D}_n$ be isomorphic digraphs.
Then there exists a permutation $\sigma$ on $[n]$ such that
\begin{align*}
(i,j)\in A(D) \ \Leftrightarrow \ (\sigma(i),\sigma(j))\in A(D').
\end{align*}
Let $\sigma^{-1}$ be the inverse of $\sigma$ and $\mathbf{y}=(y_1,\dots ,y_n)$
such that $y_i=x_{\sigma^{-1}(i)}$ for all $1\leq i \leq n$, i.e., $y_{\sigma(i)}=x_i$.
Then it is easy to see that
\begin{align*}
k\text{NND}(\mathbf{x})=D \ \Leftrightarrow \ k\text{NND}(\mathbf{y})=D'.
\end{align*}
The rest of the proof is similar to that of Proposition \ref{prop:isoinv}.
\end{proof}
As in VRDs, in the construction of a RNND, once $\mathbf{x}$ is fixed, then the arcs are uniquely determined.
However, in a VRD, by definition, inserting the arc $(i,j)$ only depends on $x_i$ and $x_j$
whereas in a RNND it depends on all the data points.
The following proposition implies that a RNND is not a VRD.
\begin{prop}
\label{prop:onlyrnnd}
A RNND is neither a VARD nor a DRD.
\end{prop}
\begin{proof}
We first show that no RNND has a VARD representation.
Recall that by the inequality in \eqref{eq:posdep},
for any VARD with $n\geq 3$ we have
\begin{align}
P( \{ (1,2),(1,3)\} \subset A(\mathbf{D})) \geq \left (P((1,2)\in A(\mathbf{D})) \right )^2. \label{eq:rnotvd1}
\end{align}
On the other hand, in a RNND, we have
\begin{align}
P( \{ (1,2),(1,3) \} \subset A(\mathbf{D}))=\frac{k(k-1)}{(n-1)(n-2)}< \left(\frac{k}{n-1}\right)^2=(P((1,2)\in A(\mathbf{D})))^2 \label{eq:rnotvd2}
\end{align}
by symmetry, and hence the result follows by \eqref{eq:rnotvd1} and \eqref{eq:rnotvd2}.
We show that there is no RNND which is also a DRD by contradiction.
Suppose that a RNND, $\mathbf{D}=(\mathcal{D}_n, P)$, is a DRD
generated by the random graph $\mathbf{G}=(\mathcal{G}_n, P_{\mathbf{G}})$ and with direction probability $p_d$.
Let $G$ be a graph in $\mathcal{G}_n$ with $P_{\mathbf{G}}(G)>0$, and $D$ be a digraph in $\mathcal{D}_n$ such that
$U(D)=G$, $n_s(D)=0$ and containing a vertex which is the tail of no arc.
In other words, $D$ is a digraph whose underlying graph is $G$, containing no symmetric arcs and
there exists a vertex $v$ in $V(D)$ such that $v$ is the head of every arc incident to $v$.
Then, as $\mathbf{D}$ is a DRD, we have
\begin{align}
P(D)= P_{\mathbf{G}}(G)(1-p_d)^{n_{as}}>0, \label{eq:rnndanddrd}
\end{align}
since $P_{\mathbf{G}}(G)>0$ and $p_d<1$.
On the other hand, for any given $\mathbf{x}=(x_1,\dots ,x_n)$,
every vertex is the tail of exactly $k$ arcs in $k\text{NND}(\mathbf{x})$.
Therefore, we obtain $P(D)=0$ which contradicts with \eqref{eq:rnndanddrd}.
\end{proof}
\begin{remark}\label{rem:underrnnd}
For any set of points in $\mathbb{R}^d$, the number of points sharing a common $k$NN is bounded above by a constant which is independent of the number of points in the set (see, \cite{yukich:1998}).
That is, there exists a number $c$ which only depends on $d,k$ and the norm $|\cdot |$ such that
in any $k$NND a vertex is the head of at most $c$ arcs.
Therefore, a vertex of the underlying graph of a $k$NND is incident to at most $c+k$ edges.
Hence, if $\mathbf{G}$ is the underlying random graph of a RNND with $n\geq c+k+2$,
then we have $P( \{ \{1,2\}, \{1,3\}, \dots , \{1,n\} \} \subset E(\mathbf{G}))=0$ which implies that
$\mathbf{G}$ is not a VERG by the inequality in \eqref{eq:vergposdepm}.
\end{remark}
\begin{remark}
One can also generate NN type random digraphs other than RNNDs.
For instance, in the construction of a RNND insert the arc $(i,j)$ if and only if $x_j$ is the $k$-th NN of $x_i$
(i.e., insert only the one to its $k$-th NN instead of putting arcs from each point to its all $k$NNs).
We can generalize RNNDs to $\mathbf{D}(n,S_k,d,\mu, |\cdot |)$ where $S_k$ is a nonempty subset of $[k]$ and
we insert the arc $(i,j)$ if and only if $x_j$ is the $s$-th NN of $x_i$ for some $s\in S_k$.
Then the results for RNNDs in this section are also valid for any $\mathbf{D}(n,S_k,d,\mu, |\cdot |)$,
i.e., every $\mathbf{D}(n,S_k,d,\mu, |\cdot |)$ is isomorphism-invariant, has no VARD or DRD representation, and for large $n$,
has an underlying random graph which is not a VERG.
\end{remark}
Note that Proposition \ref{prop:onlyrnnd} implies that the regions $\text{VARD}^c$ in Figure \ref{fig:vards},
$\text{DRD}^c$ in Figure \ref{fig:drds} and $(\text{DERD} \cup \text{VARD})^c$ in Figure \ref{fig:derdvard} are nonempty.
For $n=3$, the only possible value of $k$ is 1.
In this case, the pair with the minimum distance are NNs of each other
and the NN of the remaining point is one of the points in this pair.
Thus, by symmetry we have $P(A(\mathbf{D})=\{(i,j),(j,i),(k,i)\})=1/6$ for every pairwise distinct $i,j,k \in \{1,2,3\}$,
and therefore any RNND $\mathbf{D}(3,1,d,\mu, |\cdot |)$ is a uniform distribution over six digraphs
independent of $d,\mu$ and $|\cdot|$.
Also, note that the underlying random graph of $\mathbf{D}(3,1,d,\mu, |\cdot |)$ is always $\mathbf{G}(3,2)$.
Observe that any RNND and $\mathbf{D}(n,nk)$ have the same number of arcs.
However, these two random digraphs are different.
Because, for the event $E=\{ \{(1,2), \dots ,(1,k+2) \} \subset A (\mathbf{D} ) \}$ we have
$P(E)=0$ in a RNND since each vertex is tail of exactly $k$ arcs,
whereas $P(E)>0$ in $\mathbf{D}(n,nk)$ since $nk \geq k+1$.
Recall that $n_e=n_a-n_s$,
and hence the number of edges in the underlying graph of a $k$NND is $nk$ minus the number of symmetric arcs.
It is easy to see that for $n>3$ there exist $k$NNDs with different number of symmetric arcs,
and therefore the underlying random graph of a RNND with $n>3$ is not a $\mathbf{G}(n,m)$.
\section{Discussion and Conclusions}
\label{sec:dis}
In this paper, we present four families, namely, ARDs, VRDs, VARDs and DRDs, of isomorphism-invariant random digraphs based on where randomness resides.
First three of these classes are extensions of the isomorphism-invariant random graph classes presented in \cite{beer:2011} to digraphs.
The family of DRDs is obtained by randomly assigning directions to edges of isomorphism-invariant random graphs,
and includes three families, DERDs, DVRDs and DVERDs.
The main results of this paper are illustrated in Figures \ref{fig:vards}-\ref{fig:derdvard}.
For $n\geq 4$, we show that there is no random digraph that is both an ARD and a VRD
(which is the digraph counterpart of the result in \cite{beer:2011},
that there is no non-degenerate random graph which is both an ERG and a VRG for $n\geq 4$).
\cite{beer:2011} also show that for every $n\geq 6$, there exist VERGs which neither belong to ERGs nor VRGs.
We reduce the lower bound for $n$ from 6 to 4 by using non-symmetric structure of the function $\phi$, and obtain the digraph counterpart of their result, i.e.,
there exist VARDs which have no ARD or VRD representation, for $n\geq 4$.
However, for DRDs we have the same lower bound 6 for $n$; that is,
for $n\geq 6$ there exist DVERDs which neither belong to DERDs or DVRDs.
We also show that for $n\geq 4$ ARDs are the only random digraphs which have both DERD and VARD representations.
The method we use for the latter result is not applicable for the intersection of the families DVRDs and VARDs,
since we lose the independence of the edges in a VRG.
Therefore, identifying all random digraphs with both DVRD and VARD representations is a challenging problem, and remains open.
For $n=3$, we show that any DERD has a VARD representation and any DERD whose edge probability is $1/2$ is also a VRD.
However, the question whether there is other DERDs with VRD representation is open,
and we conjecture that any DERD is a VRD as well for $n=3$.
Yet, when $n=3$, every DVERD is a DVRD.
However, in the case of $n=2$, any isomorphism-invariant random digraph has DERD, DVRD and VRD representations.
We also study RNNDs and determine where they lie in these classifications.
We show that no RNND has a DRD or a VARD representation,
and the underlying random graph of a RNND with large $n$ is not a VERG.
\section*{Acknowledgments}
EC was supported by the European Commission under the Marie Curie International Outgoing Fellowship Programme
via Project \# 329370 titled PRinHDD.
|
\section{#1}\label{#2}\setcounter{equation}{0}}
\textwidth 16truecm \textheight 8.4in\oddsidemargin0.33truecm\evensidemargin0.7truecm\voffset-.8truecm
\def\nnewpage{}
\f
\begin{document}
\def{\colr \underline{??????}\colb}{{\colr \underline{??????}\colb}}
\def\nto#1{{\colC \footnote{\em \colC #1}}}
\def\fractext#1#2{{#1}/{#2}}
\def\fracsm#1#2{{\textstyle{\frac{#1}{#2}}}}
\def{}
\def\colr{{}
\def\colg{{}
\def\colb{{}
\def\cole{{}
\def\colA{{}
\def\colB{{}
\def\colC{{}
\def\colD{{}
\def\colE{{}
\def\colF{{}
\ifnum\coloryes=
\definecolor{coloraaaa}{rgb}{0.1,0.2,0.8
\definecolor{colorbbbb}{rgb}{0.1,0.7,0.1
\definecolor{colorcccc}{rgb}{0.8,0.3,0.9
\definecolor{colordddd}{rgb}{0.0,.5,0.0
\definecolor{coloreeee}{rgb}{0.8,0.3,0.9
\definecolor{colorffff}{rgb}{0.8,0.3,0.9
\definecolor{colorgggg}{rgb}{0.5,0.0,0.4
\def\colg{\color{colordddd}
\def\colb{\color{black}
\def\colr{\color{red}
\def\cole{\color{colorgggg}
\def\colA{\color{coloraaaa}
\def\colB{\color{colorbbbb}
\def\colC{\color{colorcccc}
\def\colD{\color{colordddd}
\def\colE{\color{coloreeee}
\def\colF{\color{colorffff}
\def\colG{\color{colorgggg}
\f
\ifnum\isitdraft=
\chardef\coloryes=1
\baselineskip=17p
\input macros.te
\def\blackdot{{\color{red}{\hskip-.0truecm\rule[-1mm]{4mm}{4mm}\hskip.2truecm}}\hskip-.3truecm
\def\bdot{{\colC {\hskip-.0truecm\rule[-1mm]{4mm}{4mm}\hskip.2truecm}}\hskip-.3truecm
\def\purpledot{{\colA{\rule[0mm]{4mm}{4mm}}\colb}
\def{\purpledot
\els
\baselineskip=15pt
\def\blackdot{{\rule[-3mm]{8mm}{8mm}}
\def\purpledot{{\rule[-3mm]{8mm}{8mm}}
\def{}
\f
\def\fbox{\fbox{\bf\tiny I'm here; \today \ \currenttime}}{\fbox{\fbox{\bf\tiny I'm here; \today \ \currenttime}}}
\def\nts#1{{\hbox{\bf ~#1~}}}
\def\nts#1{{\colr\hbox{\bf ~#1~}}}
\def\ntsf#1{\footnote{\hbox{\bf ~#1~}}}
\def\ntsf#1{\footnote{\colr\hbox{\bf ~#1~}}}
\def\bigline#1{~\\\hskip2truecm~~~~{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}{#1}\\
\def\biglineb{\bigline{$\downarrow\,$ $\downarrow\,$}
\def\biglinem{\bigline{---}
\def\biglinee{\bigline{$\uparrow\,$ $\uparrow\,$}
\def\widetilde{\widetilde}
\newtheorem{Theorem}{Theorem}[section]
\newtheorem{Corollary}[Theorem]{Corollary}
\newtheorem{Proposition}[Theorem]{Proposition}
\newtheorem{Lemma}[Theorem]{Lemma}
\newtheorem{Remark}[Theorem]{Remark}
\newtheorem{assumption}[Theorem]{Assumptions}
\newtheorem{definition}[Theorem]{Definition}
\def\thesection.\arabic{equation}{\thesection.\arabic{equation}}
\def\hfill$\Box$\\{\hfill$\Box$\\}
\def\hfill$\Box$\\{\hfill$\Box$\\}
\def\comma{ {\rm ,\qquad{}} }
\def\commaone{ {\rm ,\qquad{}} }
\def\dist{\mathop{\rm dist}\nolimits}
\def\sgn{\mathop{\rm sgn\,}\nolimits}
\def\Tr{\mathop{\rm Tr}\nolimits}
\def\div{\mathop{\rm div}\nolimits}
\def\supp{\mathop{\rm supp}\nolimits}
\def\divtwo{\mathop{{\rm div}_2\,}\nolimits}
\def\re{\mathop{\rm {\mathbb R}e}\nolimits}
\def\mathbb C{\mathbb C}
\def\mathbb R{\mathbb R}
\def\mathbb L{\mathbb L}
\def\mathbb N{\mathbb N}
\def\mathbb P{\mathbb P}
\def\mathbb E{\mathbb E}
\def\mathbb F{\mathbb F}
\def\mathbb G{\mathbb G}
\def\mathbb Q{\mathbb Q}
\def{\tilde C}{{\widetilde C}}
\def{\cal Z}{{\cal Z}}
\def{\cal D}{{\cal D}}
\def\epsilon{\epsilon}
\def\delta{\delta}
\def{\mathbf{1}}{{\mathbf{1}}}
\def\indeq{\qquad{}}
\def\hbox{\huge\textbullet}{.}
\def\semicolon{\,;}
\title{Existence of invariant measures for the stochastic damped Schr\"odinger equation}
\author{Ibrahim Ekren, Igor Kukavica, and Mohammed Ziane}
\maketitle
\date{}
\begin{center}
\end{center}
\medskip
\indent Departement fur Mathematik, ETH Zurich, Ramistrasse 101, CH-8092, Zurich\\
\indent email: [email protected]\\
\indent Department of Mathematics, University of Southern California, Los Angeles, CA 90089\\
\indent e-mails: kukavica\char'100usc.edu, ziane\char'100usc.edu
\begin{abstract}
In this paper, we address the long time behavior of solutions of the
stochastic Schr\"odinger equation in ${\mathbb R}^{d}$.
We prove the existence of an invariant measure and establish
asymptotic compactness of solutions, implying in particular
the existence of an ergodic measure.
\end{abstract}
\noindent\thanks{\em Mathematics Subject Classification\/}:
\noindent\thanks{\em Keywords:\/}
Invariant measures,
stochastic Schr\"odinger equation,
white noise,
long time behavior,
asymptotic compactness,
tightness,
Feller property,
Aldous criterion
\startnewsection{Introduction}{sec1}
The main purpose of the paper is to study the long time behavior of
the stochastic damped Schr\"odinger equation
\begin{equation}
du + (\lambda u+i\Delta u -i|u|^{2\sigma} u )dt
= \Phi dW_t
\label{EQ01}
\end{equation}
in an unbounded domain.
Our main result
provides the existence of an invariant measure
of the Markov semigroup for the equation \eqref{EQ01}
driven by an
additive noise.
In addition, using the asymptotic compactness, we prove that the set
of invariant measures is closed and convex leading to an existence of
an ergodic measure.
The problem of
existence of an invariant measure for stochastic
partial differential equations
with dissipation
and in a bounded domain
is now relatively well-understood with
the construction of the invariant measure following the classical
Krylov-Bogolyubov procedure.
The smoothing
properties of the equation and the
boundedness of the domain guarantee the necessary compactness.
For example, the existence of invariant measures for the reaction
diffusion equations, for the Navier-Stokes equations,
complex Ginzburg-Landau, and
fractionally dissipated Euler equations was established in
\cite{CGV,F1,F2}.
Also, for the primitive equations, the invariant measure
was constructed in \cite{GKVZ}.
In the case of nondegenerate noise,
a coupling method can be used to establish
existence and uniqueness of the ergodic measure. For instance,
in the case of Schr\"odinger equation with
nondegenerate noise and when the domain is bounded,
Debussche and Odasso
established in \cite{DO}
the existence of a unique ergodic measure
(cf.~also [DV,DZ,GMR,HM,KS,MR]).
The main goal of this paper is to address the existence of an
invariant measure for the stochastic damped Schr\"odinger
equation in an unbounded domain.
The main difficulties are the
the lack of smoothing and compactness properties of the solution operator in finite time.
For instance, the coupling method is not
expected to work in this situation
since Foias-Prodi type estimates, necessary for the
approach, are not available.
In order to overcome these difficulties, we establish an asymptotic
compactness property of the solution operator
(cf.~Lemma~\ref{lemma-tightness}).
Namely, we prove that for every sequence of
solutions resulting from $H^1$-bounded initial conditions and for
every sequence of times diverging to $\infty$, there exists a
subsequence of solutions and a sequence of times such that marginals of
these solutions at these times converge in distribution in $H^1$. For
this purpose we employ the conserved quantities used classically for the
deterministic analog of the equations. We also use the energy
equation approach introduced in the deterministic setting case by
J.~Ball \cite{B}. His method was further developed
to more general deterministic situations, in particular to establish
the existence and regularity
of attractors for the damped KdV equation \cite{GR,R}
and for the damped
Schr\"odinger equation \cite{G1,G2,G3,GK,GL}.
Two byproducts of the asymptotic compactness property
established in this paper
is the existence of an invariant measure for the stochastic Schr\"odinger equation and the compactness of the set of invariant measures.
We note that the existence and uniqueness of solutions was established
by de~Bouard and Debussche in \cite{DD}.
The paper is organized as follows.
In Section~\ref{sec4}, we prove an abstract tightness result that links the evolution of some scalar quantities to the asymptotic compactness stated above.
The main feature of the $k$-th order scalar quantity
is that it is equivalent to the $H^k$ norm, while the drift of
square of its expectation is continuous in $H^{k-1}$ norm.
We also make an Aldous type continuity assumption (cf.~(iii) in Definition~\ref{energy-evolution}) which allows us
to use Aldous criterion \cite{Bi} for convergence of distributions in $L_{\rm loc}^{2}$
to pass to a limiting martingale solution \cite{D,MiR1,MiR2}.
We note that while the linear part is assumed to be a Schr\"odinger type
operator $i\Delta$, our criterion can be used for more general linear
operators as well
after suitable adjustments.
In Section~\ref{sec5}, we use this asymptotic compactness criterion
for the Schr\"odinger equation by considering the first two classical
Schr\"odinger invariants and
prove the main tightness lemma.
The paper is concluded by showing that the set of invariant measures
is closed and convex, which implies the existence of an ergodic measure.
\startnewsection{Notations}{sec2}
For functions
$u,v\in L^2(\mathbb R^{d})=L^2(\mathbb R^d;\mathbb{C})$,
denote by
$\Vert u\Vert_{L^2}$ the $L^2(\mathbb R^d)$
norm of $u$ and by $(u,v)=\int_{\mathbb R^d} u(x)\overline v(x)dx$,
the
$L^2$-inner product of $u$ and $v$. We fix a basis
$\{e_i\}_{i\geq 0}$
of $L^2(\mathbb R^d)$ that consists of smooth and compactly supported functions.
For a Banach space $B$ and with $T>0$ and $p\ge1$,
denote by $L^p([0,T];B)$ the space of functions from $[0,T]$ into $B$
with integrable $p$-th power over $[0,T]$ and by $C([0,T];B)$ the set
of continuous functions from $[0,T]$ into $B$.
Similarly to functional spaces, for $p>0$, denote by $\mathbb L^p
(\Omega,B)$ the space of random variables with values in $B$ and a
finite $p$-th moment.
Denote by $\Delta =\sum_i\partial^2_{i}$ the Laplace operator and by $H^r (\mathbb R^d)$ the Sobolev space of functions $u$ satisfying
\begin{equation}\Vert u\Vert^2_{H^r}
=\int_\mathbb R (1-\Delta)^{\fractext{r}{2}}
(u(x)\overline u(x)) dx<\infty,
\end{equation}
with the inner product denoted by $(u,v)_{H^r}$.
Write ${\cal B}
(H^1(\mathbb R^d) )$
for the set of Borel measurable subsets of $H^1(\mathbb R^d)$.
Also, denote by $L^2_{\rm loc}(\mathbb R^d)$ the space of locally square
integrable functions which with the usual metric is a
complete
metric space.
For a Hilbert space $H$, we write ${{\rm HS}(L^2,H)}$ for the space
of linear operators $\Phi\colon L^2(\mathbb R^d) \to H$
with finite Hilbert-Schmidt norm
\begin{equation}
\Vert\Phi\Vert_{{\rm HS}(L^2,H)}
=
\left(
\sum_{i=1}^\infty \Vert\Phi e_i\Vert_H^2
\right)^{1/2}
\hbox{\huge\textbullet}
\end{equation}
\startnewsection{The Schr\"odinger equation}{sec3}
We fix a probability space $(\Omega, \mathbb F,\mathbb P)$ carrying a countable
family of independent Brownian motions
$\{B^i_t\}_{i\in\mathbb N,t\geq 0}$ and define the Wiener process
\begin{equation}
W_t=\sum_{i\in\mathbb N}e_i B_t^i\hbox{\huge\textbullet}
\end{equation}
Fix $\lambda>0$. In this paper,
we investigate the long time behavior
of solutions of the stochastic damped nonlinear Schr\"odinger equation
\begin{equation}\label{equation-schro}
du + (\lambda u+i\Delta u -i|u|^{2\sigma} u )dt
= \Phi dW_t,
\end{equation}
on the space-time
domain $[0,\infty)\times \mathbb R^d$
with an additive noise, by establishing
the existence of an invariant measure and the asymptotic
tightness of
solutions of
the equation.
We emphasize that unlike in \cite[Assumption~H1]{MR}, our problem
in the whole space $\mathbb R^d$ does not allow any compact embeddings.
Recall the functionals
\begin{align}
M(v)&=|v|^2_{L^2}\\
H(v)&=\frac{1}{2} \int_{\mathbb R^d} |\nabla v(x)|^2 dx -\frac{1}{2\sigma +2}\int_{\mathbb R^d}|v(x)|^{2\sigma +2}dx
\end{align}
which are
classical invariant quantities for the Schr\"odinger equation.
The existence of solutions for the equation \eqref{equation-schro} was
proven in \cite{DD}.
In order to be able to apply the existence results in
\cite{DD}, we make the following assumptions.
\begin{assumption}\label{assumption-existence-schro}
i) $0 \leq \sigma <\fractext{2}{(d-2)}$ if $d\geq 3$ or $\sigma\ge0$
if $d=1,2$.\\
ii) $\Phi\in HS(L^2(\mathbb R^d); H^1(\mathbb R^d))$.
\end{assumption}
We now
recall the existence result from \cite[Theorem~3.4, Propositions~3.2 and~3.4]{DD}.
\cole
\begin{Theorem}
Under Assumptions~\ref{assumption-existence-schro}, for every
$\mathbb F_0$ measurable, $H^1(\mathbb R^n)$ valued random variable $u_0$,
there exists an $H^1(\mathbb R^d)$-valued
and continuous solution $\{u_t\}_{t\geq 0}$
of \eqref{equation-schro} with the initial condition $u_0$.
Additionally, the quantities $M$ and $H$ evolve as
\begin{align}\label{evolution-M}
dM(u_s)
+2\lambda M(u_s)ds =2 \sum_{i} Re \left( u_s,{\Phi e_i}\right)dB^i(s)+ \Vert\Phi \Vert_{HS(L^2; L^2)}^2ds
\end{align}
and
\begin{align}\label{evolution-H}
&dH(u_s)+2\lambda H(u_s)ds \notag\\
\indeq &=
\frac{\lambda\sigma}{\sigma+1} \int |u(\widetilde s,x)|^{2\sigma+2}dx d\widetilde s
- \sum_i \re \left( \Delta u(s) +|u(s)|^{2\sigma} u(s),\Phi e_i
\right)dB^i_s\notag
\nonumber\\&\indeq
+\left(\frac{\Vert\nabla \Phi \Vert^2_{{HS(L^2; L^2)}}}{2}
-\frac{\Vert|u(s)|^\sigma\Phi\Vert^2_{{HS(L^2;
L^2)}}}{2}\right)ds
\nonumber\\&\indeq
-{\sigma}\sum_i \left( |u(s)|^{2\sigma-2},(\re(\overline u(s) \Phi e_i))^2\right)ds,
\end{align}
where $|u(s)|^\sigma\Phi$ is the operator that to a function $v$ associates the function $|u(s)|^\sigma \Phi v$.
\end{Theorem}
\colb
Note that the results in \cite{DD} are given for $\lambda =0$ but one
can easily pass from $\lambda=0$ to any $\lambda> 0$.
\subsection{The Semigroup}
Let $u_0\in H^1(\mathbb R^d)$ be a deterministic initial condition, and let
$u$ be the corresponding solution of \eqref{equation-schro}. For all $B\in {\cal B} (H^1(\mathbb R^d) )$ we define the transition probabilities of the equation by
\begin{equation}\label{transition-prob}
P_t (u_0,B)=\mathbb P(u_t\in B)
\hbox{\huge\textbullet}
\end{equation}
For any $H^1(\mathbb R^d)$-valued measure $\nu$, we denote by
$(\nu P_t)(\cdot)=\int_{H^1(\mathbb R^d)} P_t(v,\cdot)\nu (dv)$ the distribution
at time~$t$ of the solution of \eqref{equation-schro} with the initial condition having the distribution $\nu$.
For any function $\xi\in C_b(H^1(\mathbb R^d);\mathbb R)$ and $t\geq 0$, denote
\begin{equation}\label{definition-psi}
P_t \xi(u_0) =\mathbb E\left[\xi(u_t)\right]=\int_{H^1(\mathbb R^d)} \xi(v)P_t(u_0, dv)
\hbox{\huge\textbullet}
\end{equation}
\begin{definition}
{\rm Let $\mu$ be a probability measure on $H^1(\mathbb R^d)$. We say that $\mu$
is an invariant measure for $P_t$ if we have
\begin{align}
\int_{H^1(\mathbb R^d)} \xi(v)\mu( dv)=\int_{H^1(\mathbb R^d)} P_t\xi(v') \mu( dv')
\end{align}
for all $\xi\in C_b(H^1(\mathbb R^d);\mathbb R)$ and $t\geq 0$.
}
\end{definition}
\subsection{Main results concerning the Schr\"odinger equation}
The following statement is the main result of this paper.
\cole
\begin{Theorem}
\label{T01}
Under Assumptions~\ref{assumption-existence-schro},
there exists an invariant measure for $P_t$.
\end{Theorem}
\colb
The main ingredient in the proof is the following lemma.
\cole
\begin{Lemma}
\label{lemma-tightness}
Under Assumptions~\ref{assumption-existence-schro} the
following two tightness assertions hold.\\
i) For all sequences
of times $t_n\to\infty$ and
$\mathbb F_0$-measurable initial conditions
$u^n_0\in H^1(\mathbb R^d)$ with distributions
$\nu^n$ satisfying
\begin{equation*}
\mathbb E\left[\Vert u^n_0\Vert _{H^1}^{4\vee\lceil{4d\sigma}\rceil}+\Vert u^n_0\Vert _{L^2}^{\lceil{4\sigma(2-d)+8}\rceil}\right]\leq R
\end{equation*}
for some $R>0$,
the family
of measures
\begin{equation}
\bigl\{
(\nu^n P_{t_n})(\cdot)
:
n\in\mathbb N
\bigr\}
\label{EQ02}
\end{equation}
on $H^{1}({\mathbb R^d})$
is tight.\\
ii) For all compact sets $K\subseteq H^1(\mathbb R^d)$ the family of probabilities
\begin{equation}
\left\{P_{s}(v,\cdot):s\in[0,1],\,v\in K\right\}
\label{EQ03}
\end{equation}
on $H^1(\mathbb R^d)$
is tight.
\end{Lemma}
\colb
Assuming the lemma,
we now prove the main theorem. The lemma
is then proven in Section~\ref{sec5} below.
\begin{proof}[Proof of Theorem~\ref{T01}]
An invariant measure is constructed using the classical
Krylov-Bogolyubov theorem, which requires the Feller property of the semigroup and the tightness of averaged measures
\begin{equation}
\mu_n(\cdot):=\frac{1}{n}\int_0^n P_t(0,\cdot) dt\hbox{\huge\textbullet}
\end{equation}
The Feller property is a consequence of \cite[Proposition~3.5]{DD}.
Thus in order to conclude the proof, we only need to
show tightness of the family of measures $\mu_{n}$.
Let $\epsilon>0$. Lemma~\ref{lemma-tightness} applied to the family
$\{P_{k}(0,\cdot);k\in \mathbb N\}$ gives the existence of a compact set
$K_\epsilon\subseteq H^1(\mathbb R^d)$ such that
\begin{equation}
\sup_{k} P_k(0,K_\epsilon^c)
\leq
\frac{\epsilon}{2}
\hbox{\huge\textbullet}
\end{equation}
We then consider the family of probabilities $\{P_s(v,\cdot): s\in[0,1],\, v\in K_\epsilon\}$. By the second part of Lemma~\ref{lemma-tightness}, this family is tight.
Therefore, there exists another compact set $A_\epsilon\subseteq H^1(\mathbb R^d)$ such that
\begin{equation}
\sup_{ s\in[0,1], v\in K_\epsilon} P_s(v,A_\epsilon^c)
\leq
\frac{\epsilon}{2}
\hbox{\huge\textbullet}
\end{equation}
By a direct computation
\begin{align}
\mu_n(A^c_\epsilon)&=\frac{1}{n}\int_0^n P_t(0,A_\epsilon^c) dt = \frac{1}{n}\sum_{k=0}^{n-1} \int_{k}^{k+1} P_t(0,A_\epsilon^c) dt\\
&=\frac{1}{n}\sum_{k=0}^{n-1} \int_{k}^{k+1} \int_{H^1(\mathbb R^d)}P_{k} (0,dv) P_{t-k}(v,A_\epsilon^c) dt\\
&=\frac{1}{n}\sum_{k=0}^{n-1} \int_{k}^{k+1} \left(\int_{H^1(\mathbb R^d)\cap K_\epsilon^c}P_{k} (0,dv) P_{t-k}(v,A_\epsilon^c) +\int_{H^1(\mathbb R^d)\cap K_\epsilon}P_{k} (0,dv) P_{t-k}(v,A_\epsilon^c) \right)dt
\end{align}
whence
\begin{align}
\mu_n(A^c_\epsilon)
&\leq \frac{1}{n}\sum_{k=0}^{n-1} \left(P_{k} (0,K_\epsilon^c) +P_{k} (0,K_\epsilon) \sup_{ s\in[0,1],\, v\in K_\epsilon} P_s(v,A_\epsilon^c)\right) \\
&\leq \frac{1}{n}\sum_{k=0}^{n-1} \left(P_{k} (0,K_\epsilon^c) +\sup_{ s\in[0,1],\, v\in K_\epsilon} P_s(v,A_\epsilon^c)\right) \leq \epsilon
\hbox{\huge\textbullet}
\end{align}
We have thus shown that the set of measures $\{\mu_n\}$ is
tight, concluding the proof of the theorem.
\end{proof}
The rest of the paper is devoted to the proof of
Lemma~\ref{lemma-tightness} and to establishing
the compactness of the set of invariant measures.
\startnewsection{An abstract tightness result}{sec4}
In this section, we give certain distributional convergence results that
we use below to prove Lemma~\ref{lemma-tightness}.
\cole
\begin{Lemma}\label{cv-loc-hi}
Let $k\in{\mathbb N}_0$, and
let $\xi_n$ and $\xi$ be an $H^k(\mathbb R^d)$-valued square integrable random variables such that $\xi_n \to \xi$ in distribution in $L^2_{\rm loc}(\mathbb R^d)$.
Assume that $\mathbb E[\Vert \xi_n\Vert^2_{H^k}]\to \mathbb E[\Vert \xi\Vert^2_{H^k}]$ as $n\to \infty$ and suppose
that the family
$\{\Vert \xi_n\Vert^2_{H^k}:n\in{\mathbb N}\}$
is uniformly integrable.
Then $\xi_n$ converges to $\xi$ in distribution in $H^k(\mathbb R^d)$.
\end{Lemma}
\colb
Note that when $k=0$, we have
$H^0(\mathbb R^d)=L^2(\mathbb R^d)$.
\begin{proof}[Proof Lemma~\ref{cv-loc-hi}]
Let $\{f_i\}$ be a complete orthonormal system for $H^k(\mathbb R^d)$
consisting of smooth compactly supported functions.
We first claim that
\begin{equation}\label{limit-Pro}
\lim_{N\to \infty }\sup_n \mathbb E\left[\sum_{i=N}^\infty |(\xi_n,f_i)_{H^k}|^2\right]=0
\end{equation}
which then quickly implies asserted convergence.
Let $\epsilon>0$.
By the uniform integrability assumption, there exists $R>0$ such that
\begin{equation}\sup_n \mathbb E\left[ \Vert \xi_n\Vert^2_{H^k}{\mathbf{1}}_{ \{\Vert \xi_n\Vert^2_{H^k}\geq R\}}\right] \leq \epsilon
\end{equation}
and, by possibly enlarging $R$, we may also assume that
\begin{equation}\mathbb E\left[ \Vert \xi\Vert^2_{H^k}{\mathbf{1}}_{ \{\Vert \xi\Vert^2_{H^k}\geq R\}}\right] \leq \epsilon\hbox{\huge\textbullet}
\end{equation}
For all $N\in{\mathbb N}$, the convergence in distribution in $L^2_{\rm loc}(\mathbb R^d)$ and the fact that $\{f_i\}$ have compact support imply
\begin{equation}
\mathbb E\left[\left(\sum_{i=1}^N |(\xi_n,f_i)_{H^k}|^2\right)\wedge R\right]\to\mathbb E\left[\left(\sum_{i=1}^N |(\xi,f_i)_{H^k}|^2\right)\wedge R\right]
\end{equation}
as $n\to\infty$.
Since
\begin{align}
&\left|\mathbb E\left[\sum_{i=1}^N |(\xi_n,f_i)_{H^k}|^2\right] -\mathbb E\left[\sum_{i=1}^N |(\xi,f_i)_{H^k}|^2\right]\right|
\nonumber\\&\indeq
\leq \left|\mathbb E\left[\left(\sum_{i=1}^N |(\xi_n,f_i)_{H^k}|^2\right)\wedge R\right]-\mathbb E\left[\left(\sum_{i=1}^N |(\xi,f_i)_{H^k}|^2\right)\wedge R\right]\right|
\nonumber\\&\indeq\indeq
+ \mathbb E\left[ \Vert \xi_n\Vert^2_{H^k}{\mathbf{1}}_{ \{\Vert \xi_n\Vert^2_{H^k}\geq R\}}\right] + \mathbb E\left[ \Vert \xi\Vert^2_{H^k}{\mathbf{1}}_{ \{\Vert \xi\Vert^2_{H^k}\geq R\}}\right]
\end{align}
we have that
\begin{equation}\lim_n \mathbb E\left[\sum_{i=1}^N |(\xi_n,f_i)_{H^k}|^2\right] =\mathbb E\left[\sum_{i=1}^N |(\xi,f_i)_{H^k}|^2\right]\hbox{\huge\textbullet}
\end{equation}
This convergence combined with the assumption $\mathbb E[\Vert \xi_n\Vert^2_{H^k}]\to \mathbb E[\Vert \xi\Vert^2_{H^k}]$ implies
\begin{equation}\label{convergence-remainder}
\mathbb E\left[\sum_{i=N+1}^\infty |(\xi_n,f_i)_{H^k}|^2\right]\to\mathbb E\left[\sum_{i=N+1}^\infty |(\xi,f_i)_{H^k}|^2\right]
\hbox{\huge\textbullet}
\end{equation}
Since $\xi$ is $H^k(\mathbb R^d)$-square integrable,
there is $N_0\in{\mathbb N}_0$ such that
\begin{equation}
\mathbb E\left[\sum_{i=N_0+1}^\infty |(\xi,f_i)_{H^k}|^2\right]
\leq
\frac{\epsilon}{2}
\hbox{\huge\textbullet}
\label{EQ26}
\end{equation}
Then, using \eqref{convergence-remainder}, there exists
$n_\epsilon\in{\mathbb N}$
for which
\begin{equation}
\sup_{n\geq n_\epsilon}
\mathbb E\left[\sum_{i=N_0+1}^\infty |(\xi_n,f_i)_{H^k}|^2\right]\leq \epsilon
\hbox{\huge\textbullet}
\label{EQ27}
\end{equation}
The family $\{\Vert\xi_n\Vert_{H^k}: n=1,\ldots, n_\epsilon-1\}$ is square integrable. Therefore,
\begin{equation}
\lim_{N\to \infty}
\mathbb E\left[\sum_{i=N}^\infty |(\xi_n,f_i)_{H^k}|^2\right]=0
\comma n\leq n_{\epsilon}-1
\hbox{\huge\textbullet}
\label{EQ29}
\end{equation}
By \eqref{EQ27} and \eqref{EQ29}, there exists
$N_1\geq N_0$ such that
\begin{equation}
\sup_{n\in{\mathbb N}}
\mathbb E\left[\sum_{i=N_1+1}^\infty |(\xi_n,f_i)_{H^k}|^2\right]\leq \epsilon
\hbox{\huge\textbullet}
\label{EQ33}
\end{equation}
Therefore, \eqref{limit-Pro} is established.
By \cite[Theorem~1.13]{Pr},
the convergence \eqref{limit-Pro} then
implies the tightness in distribution in $H^k(\mathbb R^d)$ of the laws of $\{\xi_n\}$. Note that any limiting measure can only be the distribution of $\xi$. Thus
\begin{equation}\label{limit-l2}
\xi_n\to \xi
\end{equation}
in distribution in $H^k(\mathbb R^d)$.
\end{proof}
We shall work on the space
${\cal Z}=C([0,T]; L^{2}_{\rm loc}(\mathbb R))$.
Denote by $z$ the canonical process on this space
and ${\cal D}$ its right continuous filtration.
We state our main theorem for an SPDE of the form
\begin{align}\label{EQ00}
&du(t) = \bigl(
-i \Delta u(t)+b(u(t))
\bigr)dt +\Phi dW_t
\end{align}
with
\begin{align}
&u(t)\in L^2_{\rm loc}(\mathbb R)\notag
\end{align}
where $b\colon \mathbb C \to \mathbb C$
is, for simplicity,
a sum of terms of the form $u^{m} |u|^{a}$ where $m\in{\mathbb N_0}$ and
$a\ge0$.
Its maximal degree ($\max\{m+a\}$) is assumed to be less than
\begin{equation*}
\frac{2 d}{d-2k}
\end{equation*}
if $d>2k$.
\begin{definition}\label{defn-weak-sol}
{\rm A measure $\nu$ on ${\cal Z}$ is a martingale solution of the
equation \eqref{EQ00} if for all $\phi$ smooth and compactly
supported functions
\begin{equation}
\int_0^T \big(|b(z_s)|,|\phi|\bigr) ds <\infty, \, \nu \mbox{-a.s.}
\end{equation}
and if
\begin{equation}
M^\phi_t
=(z_t-z_0,\phi) -\int_0^t (-i\Delta z_s +b(z_s),\phi)ds
\label{EQ38}
\end{equation}
and
\begin{equation}
(M^\phi_t)^2-\int_0^t \sum_i \left(\Phi e_i,\phi\right)^2ds
\label{EQ48}
\end{equation}
are $\nu$-local martingales.
We say that $\nu$ is a $H^k$ square integrable martingale solution if
\begin{equation}\sup_{t\in [0,T]} \mathbb E^\nu \left[\Vert z_t\Vert_{H^k}^2\right]<\infty\hbox{\huge\textbullet}
\end{equation}
}
\end{definition}
\begin{Remark}{\rm
i) Note that a martingale solution can be obtained from any strong solution of \eqref{EQ00}. Indeed, let $u$ be a solution of \eqref{EQ00} on the interval $[0,T]$.
Define the measure
\begin{equation}
\nu(dz)
=\int_\Omega \delta_{\{\{u_{s}(\omega)\}_{s\in [0,T]}\}}(dz)\mathbb P(d\omega)
\label{EQ39}
\end{equation}
meaning the measure on ${\cal Z}$ such that for all continuous bounded $F\colon{\cal Z}\to\mathbb R$
we have
\begin{equation}
\int_{\cal Z} F(z)\nu (dz)= \mathbb E[F(\{u(s)\}_{s\in[0,T]})]
\hbox{\huge\textbullet}
\end{equation}}
\end{Remark}
In order to facilitate the statement of the main result of this section,
we introduce the concept of $H^{k}$-evolution property.
\begin{definition}\label{energy-evolution}
{\rm Let $k\in{\mathbb N}$.
The equation \eqref{EQ00} has the $H^k$-norm evolution property if
for $i=0,\ldots, k$
there exist continuous functions
$F_i\colon L^2_{\rm loc}\to \mathbb R$,
$\widetilde F_i \colon\mathbb R \times L^2_{\rm loc}\to \mathbb R$,
and $ G_i \colon\mathbb R \times \mathbb R \times L^2_{\rm loc}\to \mathbb R$
satisfying the following conditions.\\
i) For all $t,r$ the functions $F_i(\cdot),\widetilde F_i(t,\cdot)$ and $G_i(t,r,\cdot)$ are continuous in $H^{i-1}$-topology on bounded sets of $H^{i}$ (for $i=0$ we require the continuity in $L^2_{\rm loc}$ on bounded sets of $L^2$).\\
ii) For all $t,r$ the functions $F_i(\cdot),\widetilde F_i(r,\cdot)$ and $G(r,\cdot)$ have at most polynomial growth in the $H^i$-norm. \\
iii) For all $H^k$ square integrable martingale solutions $\nu$ of \eqref{EQ00}, the conservation equality
\begin{equation}\label{conserved-quantities}
\mathbb E^\nu\left[\Vert z_t\Vert^2_{H^i}\right]-e^{-2\lambda (t-s)}\mathbb E^\nu\left[\Vert z_s\Vert^2_{H^i}\right]=\mathbb E^\nu\left[F_i(z_t)-\widetilde F_i(t-s,z_s)\right]
+\int_s^t \mathbb E^\nu\left[G_i(t-r,z_r)\right]dr
\end{equation}
holds for all $t\ge s$.
}
\end{definition}
We now state a theorem,
which, combined with Lemma~\ref{cv-loc-hi},
gives us a tightness result needed for
the Krylov-Bogolyubov procedure.
\cole
\begin{Theorem}\label{thm-limit-conservation}
Assume that the equation \eqref{EQ00}
has the $H^k$-norm evolution property,
and let $u^n$ be a sequence of
strong solutions of \eqref{EQ00} satisfying
the following conditions:\\
a) We have a uniform bound
\begin{equation}\label{unif-bound-theo}
\gamma=
\sup_{r\geq 0}\sup_{k\geq i\geq 0} \sup_{n,t} \mathbb E\left[\Vert b(u^n_t)\Vert_{L^1}^2+| F_i(u^n_t)|^2+|\widetilde F_i(r,u^n_t)|^2+|G_i(r,u^n_t)|^2+ \Vert u^n_t\Vert^4_{H^k}\right]<\infty
\hbox{\huge\textbullet}
\end{equation}
b) For every sequence of stopping times $T_n$ and positive numbers $\delta_n$ such that $\delta_n\to 0$ as $n\to \infty$, we have
\begin{equation}\label{aldous-criterion}
\mathbb E\left[\Vert u^n_{T_n+\delta_n}-u^n_{T_n}\Vert^2_{L^2}\right]\to 0\mbox{~~as~} n\to \infty
\hbox{\huge\textbullet}
\end{equation}
c) There exists a sequence
$t_n\to \infty $ and an $H^k$-valued random variable
$\xi$ such that
$u^n_{t_n}\to \xi$ in distribution in $L^2_{\rm loc}$. \\
Then $\mathbb E[\Vert u^n_{t_n}\Vert^2_{H^k}]\to \mathbb E[\Vert \xi\Vert^2_{H^k}]$ as $n\to \infty$.
\end{Theorem}
\colb
\begin{Remark}{\rm
The powers in \eqref{unif-bound-theo}
have been chosen so we can obtain the uniform
integrability of the family and then the De la Vallee Poussin's
theorem can be applied.}
\end{Remark}
\begin{proof}[Proof of Theorem~\ref{thm-limit-conservation}] Since
\begin{equation}\liminf_n \mathbb E[\Vert u^n_{t_n}\Vert^2_{H^k}]\geq\mathbb E[\Vert\xi\Vert^2_{H^k}]
\end{equation}
we only need to
prove
\begin{equation}
\limsup_n \mathbb E[\Vert u^n_{t_n}\Vert^2_{H^k}]\leq\mathbb E[\Vert\xi\Vert^2_{H^k}]
\hbox{\huge\textbullet}
\label{EQ06}
\end{equation}
We establish \eqref{EQ06}
by induction on $k$, reasoning by contradiction at each step.
For $k=0$ (cf.~Step~1), we use \eqref{aldous-criterion}
and the Aldous's criterion to obtain
a compactness of measures induced by the process $\{u^n\}$.
Then using \eqref{conserved-quantities}
for a limiting measure we obtain a contradiction.
{\it Step 1:} First we prove \eqref{EQ06} for $k=0$.
We assume that the convergence does not hold. This means that, passing to a subsequence,
there exists $\epsilon>0$ such that
\begin{equation}\mathbb E[ \Vert u^n_{t_n}\Vert^2_{L^2}] \geq \mathbb E[ \Vert \xi\Vert^2_{L^2}]+\epsilon
\comma n\in{\mathbb N}
\label{EQ08}
\hbox{\huge\textbullet}
\end{equation}
We now pick $T>0$ such that
$3 \gamma^{1/2} e^{-2\lambda T}\leq \epsilon$.
Note that, by \eqref{unif-bound-theo}, the sequence
$\{u^n_{t_n-T}\}$ satisfies
\begin{equation}
\sup_{n} \mathbb E\left[\Vert u^n_{t_n-T}\Vert^4_{L^2}\right]\leq \gamma
\hbox{\huge\textbullet}
\end{equation}
Therefore, passing to a further subsequence, there exists
an $L^2$-valued random variable $\xi_{-T}$
such that $u^n_{t_n-T}$ converges in distribution in
$L^2_{\rm loc}(\mathbb R^d)$ to $\xi_{-T}$.
Define a sequence of measures $\nu^n$ on ${\cal Z}$ by
\begin{equation}
\nu^n(dz)
=\int_\Omega \delta_{\{\{u^n_{t_n-T+r}(\omega)\}_{r\in [0,T]}\}}(dz)\mathbb P(d\omega)
\hbox{\huge\textbullet}
\end{equation}
The assumption \eqref{aldous-criterion} and the Aldous criterion
\cite[Theorem~16.10]{B} imply that the sequence
$\{\nu^n\}_{n=1}^{\infty}$ is tight in distribution in ${\cal Z}$.
Taking a further subsequence,
we obtain the existence of $\nu$ such that
\begin{equation}
\mathbb E^{\nu^n}\left[F(z)\right] =\mathbb E\left[F(\{u^n_{t_n-T+s}\}_{s\in[0,T]})\right] \to \mathbb E^{\nu}\left[F(z)\right]\mbox{~~as~}n\to \infty
\comma F\in C_{b}({\cal Z})
\hbox{\huge\textbullet}
\end{equation}
Identifying the marginals, we easily see that the distribution of $z_T$ under $\nu$ is the same as the distribution of $\xi$. Similarly, the distribution of $z_0$ under $\nu$ is the same as the distribution of $\xi_{-T}$.
We write the equation \eqref{conserved-quantities} at times $t_n$ and $t_n-T$ for the measure $\nu^n$
\begin{equation}\label{norm-evolution-n}
\mathbb E\left[\Vert u^n_{t_n}\Vert^2_{L^2}\right]-e^{-2\lambda T }\mathbb E\left[\Vert u^n_{t_n-T}\Vert_{L^2}^2\right]=\mathbb E\left[F_0(u^n_{t_n})-\widetilde F_0(T,u^n_{t_n-T})\right]
+\int_0^T \mathbb E\left[G_0(T-r,u^n_{t_n-T+r})\right]dr\hbox{\huge\textbullet}
\end{equation}
We claim that by the assumptions (i), (ii), and (a),
we have sufficient integrability and continuity at the right hand side of the equation to use the convergence of $\nu^n$ to $\nu$ and pass to the limit to obtain
\begin{equation}\lim_n \Bigl(
\mathbb E\left[\Vert u^n_{t_n}\Vert^2_{L^2}\right]-e^{-2\lambda T }\mathbb E\left[\Vert u^n_{t_n-T}\Vert^2_{L^2}\right]
\Bigr)
=\mathbb E^\nu\left[F_0(z_T)-\widetilde F_0(T,z_0)\right]
+\int_0^T \mathbb E^\nu\left[G_0(T-r,z_r)\right]dr\hbox{\huge\textbullet}\end{equation}
Indeed, the convergence of $\nu^n$ to $\nu$ in ${\cal Z}$ implies that for all $s\in [0,T]$ and every function $\xi\colon L^2_{\rm loc}(\mathbb R^d)\to \mathbb R$ continuous and bounded we have
\begin{equation}\mathbb E\left[\xi(u^n_{t_n-T+s})\right]\to\mathbb E^\nu \left[\xi(z_s)\right]\hbox{\huge\textbullet}\end{equation}
Note that, by the assumption (i), the mappings $F_0(\cdot)$, $\widetilde F_0(T,\cdot)$, and $G_0(T-s,\cdot)$ are continuous in $L^2_{\rm loc}(\mathbb R^d)$ on bounded sets of $L^2(\mathbb R^d)$. Additionally the assumption (ii) and the uniform bound \eqref{unif-bound-theo}
allows us to truncate $F_0(\cdot)$, $\widetilde F_0(T,\cdot)$, and $G_0(T-s,\cdot)$ when they are large in order to obtain
\begin{equation}\mathbb E\left[\psi(u^n_{t_n-T+s})\right]\to\mathbb E^\nu \left[\psi(z_s)\right]\end{equation}
for $\psi=F_0(\cdot)$, $\psi=\widetilde F(T,\cdot)$, and $\psi=G_0(T-s,\cdot)$. Thus
\begin{equation}\lim_n \Bigl(\mathbb E\left[\Vert u^n_{t_n}\Vert^2_{L^2}\right]-e^{-2\lambda T }\mathbb E\left[\Vert u^n_{t_n-T}\Vert^2_{L^2}\right]\Bigr)=\mathbb E^\nu\left[F_0(z_T)-\widetilde F_0(T,z_0)\right]
+\int_0^T \mathbb E^\nu\left[G_0(T-r,z_r)\right]dr\hbox{\huge\textbullet}\end{equation}
We shall show in Step~3 that $\nu$ is a $L^2$-square integrable martingale solution of \eqref{EQ00}. Using this result and
by the assumption \eqref{conserved-quantities} one has
\begin{equation}\mathbb E^\nu\left[F_0(z_T)-\widetilde F_0(T,z_0)\right]
+\int_0^T \mathbb E^\nu\left[G_0(T-r,z_r)\right]dr=\mathbb E^\nu\left[\Vert z_T\Vert^2_{L^2}\right]-e^{-2\lambda T}\mathbb E^\nu\left[\Vert z_0\Vert^2_{L^2}\right]
\hbox{\huge\textbullet}
\end{equation}
Noting the bound \eqref{unif-bound-theo}, we may pass to the limit and obtain
\begin{align}
&\lim_n \left(\mathbb E\left[\Vert u^n_{t_n}\Vert^2_{L^2}\right]-e^{-2\lambda T }\mathbb E\left[\Vert u^n_{t_n-T}\Vert^2_{L^2}\right]\right)
\nonumber\\&\indeq
=\mathbb E^\nu\left[\Vert z_T\Vert^2_{L^2}\right]-e^{-2\lambda T}\mathbb E^\nu\left[\Vert z_0\Vert^2_{L^2}\right]
\nonumber\\&\indeq
=\mathbb E\left[\Vert \xi\Vert^2_{L^2}\right]-e^{-2\lambda T}\mathbb E\left[\Vert \xi_{-T}\Vert^2_{L^2}\right]
\hbox{\huge\textbullet}
\label{EQ05}
\end{align}
By convergence of $u_{t_n-T}^{n}$ to $\xi_{-T}$ in distribution in $L^2_{\rm loc}$ and the Fatou's lemma, we obtain
\begin{equation}
\mathbb E\left[\Vert \xi_{-T}\Vert^2_{L^2}\right]
\leq \liminf_n \mathbb E\left[\Vert u^n_{t_n-T}\Vert^2_{L^2}\right] \leq \gamma^{1/2}
\hbox{\huge\textbullet}\end{equation}
Using \eqref{EQ08} and \eqref{EQ05}, we obtain
\begin{align}
\epsilon &\leq
\liminf_n
\mathbb E\left[\Vert u^n_{t_n}\Vert^2_{L^2}\right]- \mathbb E\left[\Vert \xi\Vert^2_{L^2}\right]
\nonumber\\&
\leq \limsup_n e^{-2\lambda T }\left(\mathbb E
\left[\Vert u^n_{t_n-T}\Vert^2_{L^2}\right]
-\mathbb E\left[\Vert \xi_{-T}\Vert^2_{L^2}\right]\right)\leq \gamma^{1/2} e^{-2\lambda T } \leq \frac{2\epsilon}{3}
\end{align}
which is a contradiction.
{\it Step 2:} Now we prove that $\nu$ is a $L^2$-square integrable martingale solution of \eqref{EQ00}.
Note that the uniform bound \eqref{unif-bound-theo}, the lower semicontinuity of the $L^2$ norm with respect to the $L^2_{\rm loc}$ topology, and the distributional convergence of $\nu^n$ to $\nu$ give that for all $t\in[0,T]$
$$
\mathbb E^\nu \left[\Vert z_t\Vert_{L^2}^2\right]\leq \gamma
\hbox{\huge\textbullet}
$$
Additionally the choice of the power for $b$ implies that the mapping $z\in{\cal Z} \to M_t^{\phi}(z)$ is continuous. Thus for all $f$ bounded continuous
$$\lim_n \mathbb E^{\nu^n}[f(M_t^\phi)]= \mathbb E^{\nu}[f(M_t^\phi)]
\hbox{\huge\textbullet}$$
For all $\phi$ smooth, there exists $K_{\phi,T}$ depending only on $\phi$ and $T$ such that
\begin{align}\label{bound-mphit}
|M_t^\phi|^2 \leq K_{\phi,T}\left(\Vert z_0\Vert^2_{L^2}+\Vert z_t\Vert^2_{L^2}+\int_0^t \Vert z_r\Vert^2_{L^2}+\Vert b(z_r)\Vert^2_{L^1} dr\right)
\comma t\in[0,T]
\hbox{\huge\textbullet}
\end{align}
We will use these points to prove that $M_t^\phi$ is a martingale under $\nu$.
We fix a family of smooth truncation functions $\Psi_R$ satisfying $|\Psi_R|\leq 2R$ and $\Psi_R(x)=x$ if $|x|\leq R$. For $0\leq s_1\leq s_2\leq\ldots\leq s_m\leq s\leq t\leq T$ smooth compactly supported functions functions $\phi_i$ and a random variable of the form $F=F(\int \phi_1 z_{s_1} dx,\ldots,\int \phi_m z_{s_m}dx)$ smooth and bounded by $1$, we have the equalities
\begin{align*}
\left | \mathbb E^\nu \left[ (M^{\phi}_t-M^\phi_s)F\right]\right|&= \left | \mathbb E^\nu \left[ (\Psi_R(M^{\phi}_t)-\Psi_R(M^\phi_s))F\right] \right| + \mathbb E^\nu \left[ |\Psi_R(M^{\phi}_t)-M^{\phi}_t|\right]+\mathbb E^\nu \left[ |\Psi_R(M^{\phi}_s)-M^{\phi}_s|\right]\\
&=\lim_n \left | \mathbb E^{\nu^n} \left[ (\Psi_R(M^{\phi}_t)-\Psi_R(M^\phi_s))F\right] \right| + \frac{1}{R}\left( \mathbb E^\nu \left[ |M^{\phi}_t|^2\right]+\mathbb E^\nu \left[ |M^{\phi}_s|^2\right]
\right) \\
&=\lim_n \left | \mathbb E^{\nu^n} \left[ (M^{\phi}_t-M^\phi_s)F\right] \right| + \frac{4T\gamma K_{\phi,T}}{R}
\hbox{\huge\textbullet}
\end{align*}
Note that by the martingale property of $M_t^\phi$ under $\nu^n$ we have $\mathbb E^{\nu^n} \left[ (M^{\phi}_t-M^\phi_s)F\right] =0$. Thus, taking $R$ to infinity we get
$ \mathbb E^\nu \left[ (M^{\phi}_t-M^\phi_s)F\right]=0$ which is sufficient to claim that $M_t^\phi$ is a $\nu$ martingale.
Due to the smoothness of $\phi$, the continuity of $z$ in $L^2_{\rm
loc}$ and \eqref{bound-mphit}, $M_t^\phi$ is a continuous and square
integrable martingale under $\nu$. We now proceed to characterize its quadratic variation.
By the definition of martingale solutions under $\nu^n$ the process $\fractext{M_t^\phi}{\sqrt{\sum_i (\Phi e_i,\phi)^2}}$ is a Brownian motion and thus has Gaussian independent increments. By the distributional convergence of $\nu^n$ and the continuity of $\fractext{M_t^\phi}{\sqrt{\sum_i (\Phi e_i,\phi)^2}}$ in the $L^2_{\rm loc}$ topology with respect to $z$ the distribution and the independence of the increments still hold under $\nu$. Thus the continuous process $\fractext{M_t^\phi}{\sqrt{\sum_i (\Phi e_i,\phi)^2}}$ is a Brownian motion under $\nu$ which implies that \eqref{EQ48} holds under $\nu$.
{\it Step 3:} For the induction step, assume that
$\mathbb E[\Vert u^n_{t_n}\Vert^2_{H^r}]\to \mathbb E[\Vert \xi\Vert^2_{H^r}]$
as $n\to \infty$ for $r=0,\ldots,k-1$.
We need to show that
\begin{equation}\mathbb E[\Vert u^n_{t_n}\Vert^2_{H^k}]\to \mathbb E[\Vert \xi\Vert^2_{H^k}]\hbox{\huge\textbullet}\end{equation}
Note that using Lemma~\ref{cv-loc-hi} at each step of
the induction one can also show that
\begin{equation}u^n_{t_n}\to \xi\end{equation} in distribution in $H^r(\mathbb R^d)$ for all $r\leq k-1$.
In order to obtain a contradiction,
assume that the convergence we are proving does not hold.
This means that, up to a subsequence, there exists $\epsilon>0$ such that
\begin{equation}\mathbb E[ \Vert u^n_{t_n}\Vert^2_{H^k}] \geq \mathbb E[ \Vert \xi\Vert^2_{H^k}]+\epsilon
\comma n\in{\mathbb N}\hbox{\huge\textbullet}\end{equation}
Similarly to the previous step, we introduce $T$ such that $3\gamma^{1/2} e^{-2\lambda T}\leq \epsilon$ and define the measures $\nu^n$ on ${\cal Z}$. We also prove similarly that there exist an $H^k$-valued random variable $\xi_{-T}$, a distribution $\nu$ on ${\cal Z}$ which is a $H^k$ square integrable solution of \eqref{EQ00} and a subsequence of $t_n$ (still denoted $t_n$) such that $\nu^n\to \nu$ on ${\cal Z}$ as $n\to\infty$ and $u^n_{t_n-T}\to \xi_{-T}$ in distribution in $L^2_{\rm loc}(\mathbb R^d)$.
Note that for all $s\in[0,T]$ the family $u^n_{t_n-T+s}$ converges in distribution in $L^2_{\rm loc}(\mathbb R^d)$ to the distribution of $z_s$ under $\nu$. Therefore, using the induction hypothesis on the family $u^n_{t_n-T+s}$ and $z_s$ we obtain
\begin{equation}
u^n_{t_n-T+s}\to z_s
\end{equation}
in distribution in $H^r(\mathbb R^d)$
and
\begin{equation}
\mathbb E[\Vert u^n_{t_n-T+s}\Vert^2_{H^r}]\to \mathbb E^\nu [\Vert z_s\Vert^2_{H^r}]
\end{equation}
for $r\leq k-1$ as $n\to \infty$.
We first use \eqref{conserved-quantities} on $\nu^n$ for $k$ to obtain
\begin{equation}
\mathbb E\left[\Vert u^n_{t_n}\Vert^2_{H^k}\right]-e^{-2\lambda T}\mathbb E\left[\Vert u^n_{t_n-T}\Vert^2_{H^k}\right]=\mathbb E\left[F_k( u^n_{t_n})-\widetilde F_k(T, u^n_{t_n})\right]
+\int_0^T \mathbb E\left[G_k(T-s, u^n_{t_n-T+s})\right]ds
\hbox{\huge\textbullet}
\end{equation}
We have proven that the distribution of $u^n_{t_n-T+s}$ converges in distribution in $H^{k-1}(\mathbb R^d)$ to the distribution of $z_s$ under $\nu$.
Similarly to the previous step, we have enough integrability and continuity on the right hand side of the equation to use this convergence and pass to the limit to obtain
\begin{equation}\lim_n \mathbb E\left[\Vert u^n_{t_n}\Vert^2_{H^k}\right]-e^{-2\lambda T}\mathbb E\left[\Vert u^n_{t_n-T}\Vert^2_{H^k}\right]=\mathbb E^\nu\left[F_k(z_T)-\widetilde F_k(T,z_0)\right]
+\int_0^T \mathbb E^\nu\left[G_k(T-s,z_s)\right]ds\hbox{\huge\textbullet}
\end{equation}
Using Fatou's lemma and \eqref{unif-bound-theo} we have that $\nu$ is a $H^k(\mathbb R^d)$ square integrable solution of \eqref{EQ00}. Thus, by assumption, Definition~\ref{energy-evolution}~(iii) gives
\begin{equation}\mathbb E^\nu\left[\Vert z_T\Vert^2_{H^k}\right]-e^{-2\lambda T}\mathbb E^\nu\left[\Vert z_0\Vert^2_{H^k}\right]=\mathbb E^\nu\left[F_k(z_T)-\widetilde F_k(T,z_0)\right]
+\int_0^T \mathbb E^\nu\left[G_k(T-s,z_s)\right]ds,
\end{equation}
which implies
\begin{equation}\lim_n \Bigl(\mathbb E\left[\Vert u^n_{t_n}\Vert^2_{H^k}\right]-e^{-2\lambda T}\mathbb E\left[\Vert u^n_{t_n-T}\Vert^2_{H^k}\right]\Bigr)=\mathbb E^\nu\left[\Vert z_T\Vert^2_{H^k}\right]-e^{-2\lambda T}\mathbb E^\nu\left[\Vert z_0\Vert^2_{H^k}\right]\hbox{\huge\textbullet}
\end{equation}
Using the same arguments as in the previous step, we obtain a contradiction.
\end{proof}
\startnewsection{Proofs of tightness for the Schr\"odinger equation}{sec5}
We now return to the Schr\"odinger equation \eqref{equation-schro}.
We fix $\lambda>0$; thus all the constants are allowed to
depend on $\lambda$.
Also, recall that we impose
Assumptions~\ref{assumption-existence-schro}
on $\sigma$ and $\Phi$.
\cole
\begin{Lemma}\label{bounds}
For every $k\in {\mathbb N}$
we have
\begin{equation}\label{l2-bounds}
\sup_{t\geq 0} \mathbb E[M(u(s))^{k}]
\leq
C_k(\mathbb E[|M(u_0)|^k]+1)
\end{equation}
and
\begin{equation}\label{h1-bounds}
\sup_{t\geq 0} \mathbb E[H(u(s))^{k}]\leq C_k(\mathbb E[|H(u_0)|^k]+1)
\end{equation}
where $C_k\ge0$ is a constant.
\end{Lemma}
\colb
\begin{proof}[Proof of Lemma~\ref{bounds}]
Using similar ideas as in \cite{EKZ}, one can show that the local martingale appearing in \eqref{evolution-M} is a martingale. Thus we have
\begin{equation}\mathbb E[M(u(t))]+2\lambda \int_0^t \mathbb E[M(u(s))]ds =\mathbb E[M(u_0)]+ t \Vert\Phi\Vert^2_{{HS(L^2;L^)}}\hbox{\huge\textbullet}
\end{equation}
Solving this ODE for $\mathbb E[M(u(t))]$, we get
\begin{equation}\mathbb E[M(u(t))]=e^{-2\lambda t}\mathbb E[M(u_0)]+\Vert\Phi\Vert^2_{{HS(L^2; L^2)}}\int_0^t e^{-2\lambda (t-s)}ds\leq \mathbb E[M(u_0)]+\frac{1}{2\lambda} \Vert\Phi\Vert^2_{{HS(L^2; L^2)}}
\end{equation}
which proves \eqref{l2-bounds} for $k=1$.
For general $k$ we proceed by induction. We assume the existence of $C_k$ for a given $k\geq 1$ and apply Ito's lemma to $M(u(t))^{k+1}$ to obtain
\begin{align}
&dM^{k+1}(u(t))+2(k+1)\lambda M^{k+1}(u(t))dt
\nonumber\\&\indeq
= (k+1) M^{k}(u(t))\Vert\Phi \Vert_{{HS(L^2; L^2)}}^2 dt
\nonumber\\&\indeq\indeq
+ \frac{k(k+1)}{2} M^{k-1}(u(t)) \sum_i Re(u(t),\Phi e_i)^2 dt+\widetilde M
\end{align}
where similarly $\widetilde M$ can be shown to be a martingale.
Thus the function $\mathbb E\left[M^{k+1}(u(t))\right]$ satisfies the ODE
\begin{align}
&\left(\mathbb E\left[M^{k+1}(u(t))\right]\right)'+2(k+1)\lambda\mathbb E\left[M^{k+1}(u(t))\right]
\nonumber\\&\indeq
= (k+1)\mathbb E\left[ M^{k}(u(t))\Vert\Phi \Vert_{{HS(L^2; L^2)}}^2\right]+ \frac{k(k+1)}{2} \mathbb E\left[M^{k-1}(u(t)) \sum_i Re(u(t),\Phi e_i)^2\right]=:g_k(t),
\end{align}
where $g_k$ is a bounded function of $t$ by the induction assumption. By solving this ODE, we see that the function $\mathbb E\left[M^{k+1}(u(t))\right]$ is bounded.
Repeating the same argument for $H^k(u(t))$ we obtain \eqref{h1-bounds}.
\end{proof}
\nnewpage
In order to obtain the tightness of the averaged measures, we use
Lemma~\ref{cv-loc-hi}
and
Theorem~\ref{thm-limit-conservation}.
The last ingredient we need is the following lemma.
\cole
\begin{Lemma}\label{schr-aldous}
Under the assumptions of Lemma~\ref{lemma-tightness},
for all stopping times $T_n$ and real numbers $\delta_n$
such that $\delta_n\to0$
as $n\to\infty$, we have
\begin{equation}\label{lemma-aldous-criterion}
\mathbb E\left[\Vert u^n_{T_n+\delta_n}-u^n_{T_n}\Vert^2_{L^2}\right]\to 0\mbox{~~as~} n\to \infty\hbox{\huge\textbullet}
\end{equation}
\end{Lemma}
\colb
\begin{proof}[Proof of Lemma~\ref{schr-aldous}]
We denote by $S^\lambda$ the semigroup associated with the linear part of the equation. With this notation, we have
\begin{align}
u^n_{T_n+\delta_n}-u^n_{T_n}
&=
S^\lambda(\delta_n) u^n_{T_n}- u_{T_n}
+ i\int_{0}^{\delta_n} S^\lambda(\delta_n-s)(|u^n(T_n+s)|^{2\sigma}u^n(s)) ds
\nonumber\\&\indeq
+\int_{0}^{\delta_n} S^\lambda(\delta_n-s)\Phi dW_{T_n+s}
\hbox{\huge\textbullet}
\end{align}
The lemma would follow from the following three convergence statements:
\begin{align}
&\mathbb E\left[\Vert S^\lambda(\delta_n) u^n_{T_n}- u^n_{T_n}\Vert^2_{L^2}\right]\to 0,
\label{EQ09}
\\
&\mathbb E\left[\Vert \int_{0}^{\delta_n} S^\lambda(\delta_n-s)(|u^n(T_n+s)|^{2\sigma}u^n(s)) ds \Vert^2_{L^2}\right]\to 0,
\label{EQ10}
\\
&\mathbb E\left[\Vert \int_{0}^{\delta_n} S^\lambda(\delta_n-s)\Phi dW_{T_n+s}\Vert^2_{L^2}\right]\to 0
\label{EQ11}
\hbox{\huge\textbullet}
\end{align}
By PDE arguments, the first convergence is obvious.
For the second convergence \eqref{EQ10}, we simply write
\begin{align}
&\mathbb E\left[\left\Vert \int_{0}^{\delta_n} S^\lambda(\delta_n-s)(|u(T_n+s)|^{2\sigma}u(s)) ds\right \Vert^2_{L^2}\right]
\nonumber\\&\indeq
\leq \delta_n \int_{0}^{\delta_n} \mathbb E\left[\Vert S^\lambda(\delta_n-s)(|u(T_n+s)|^{2\sigma}u(s)) \Vert^2_{L^2}\right]ds
\hbox{\huge\textbullet}
\end{align}
Given the uniform bounds \eqref{l2-bounds}, the integrand is uniformly bounded and
the convergence thus holds.
For the third convergence \eqref{EQ11}, we use the the Burkholder-Davis-Gundy Inequality
\cite[Lemma~5.24]{DZ} and obtain
\begin{equation}
\mathbb E\left[\left\Vert \int_{0}^{\delta_n} S^\lambda(\delta_n-s)\Phi dW_{T_n+s}\right\Vert^2_{L^2}\right]\leq \int_0^{\delta_n }\Vert S^\lambda(\delta_n-s)\Phi \Vert_{HS(L^2,L^2)}^2 ds \to 0
\end{equation}
as $n\to\infty$.
\end{proof}
\subsection{Proof of Lemma~\ref{lemma-tightness}}
{\it Proof of (i):}
We show that the assumptions of Theorem~\ref{thm-limit-conservation} with $k=1$ are satisfied for the equation~\eqref{equation-schro} and the set
$ \{(\nu^nP_{t_n})(\cdot):n\in\mathbb N\}$
is relatively weakly compact over $H^1(\mathbb R^d)$.
We define
\begin{align}
&F_0 =\widetilde F_0=0
\nonumber\\&
G_0(t,r,v):=e^{-2\lambda (t-r)}\Vert \Phi\Vert_{HS(L^2,L^2)}^2,\\
&F_1(v)=\frac{1}{2\sigma +2} \int |v(x)|^{2\sigma+2} dx\\
&\widetilde F_1(r,v):=e^{-2\lambda r}\frac{1}{2\sigma +2} \int |v(x)|^{2\sigma+2} dx,\\
&G_1(r,v):= e^{-2\lambda r}\biggl(\int_{\mathbb R^d} |v(x)|^{2\sigma+2} dx +\Vert \nabla \Phi\Vert^2_{HS(L^2,L^2)}
\nonumber\\&\indeq\indeq\indeq\indeq\indeq\indeq
- \Vert |v|^\sigma \Phi\Vert^2_{HS(L^2,L^2)}-\sigma \sum_i Re(|v|^{2\sigma -2 }v^2, (\Phi e_i)^2) \biggr)
\hbox{\huge\textbullet}
\end{align}
We apply the Gagliardo-Nirenberg interpolation inequality to obtain
\begin{equation}\Vert v\Vert_{L^{2\sigma+2}}^{2\sigma+2}\leq C\Vert v\Vert_{H^1}^{d\sigma}\Vert v\Vert^{\sigma(2-d)+2}_{L^2} \end{equation}
which shows that $F_1(\cdot)$, $\widetilde F_1(r,\cdot)$, and $G_1(r,\cdot)$ are continuous in $L^2(\mathbb R^d)$ on bounded sets of $H^1(\mathbb R^d)$. They also have at most polynomial growth in $H^1(\mathbb R^d)$ and given the bounds on $u^n_0$ and Lemma~\eqref{bounds}, with $b(u)=|u|^{2\sigma}u$, we have the bound \eqref{unif-bound-theo}. Additionally, given the assumption \ref{assumption-existence-schro} on $\sigma$, we can easily verify that the degree of $b$ satisfy for $d\geq 2$,
$$2\sigma +1 <\frac{d+2}{d-2}\leq \frac{2d}{d-2}.$$
Since $\nu$ is a $H^1$-square integrable martingale solution of \eqref{equation-schro}, by \cite[Theorem~2.4]{D}, we can extend the probability space $({\cal Z},{\cal D},\nu)$ to obtain a family of Brownian motions $\hat B^i$ such that the $H^{-1}(\mathbb R^d)$-valued continuous martingale $ M_t= z(t)-z(0) +\int_0^t (\lambda z(s)-i\Delta z(s) -i|z(s)|^{2\sigma} z(s) )ds$ can be represented as
\begin{equation}dM_t=\sum_i \Phi e_i d\hat B^i_t\hbox{\huge\textbullet}\end{equation}
Similarly to \cite[Propositions~3.2 and~3.3]{DD}, we apply Ito's lemma to $M(z(t))$ and $H(z(t))$ on this probability space to obtain that \eqref{conserved-quantities} holds for $i=0,1$ under $\nu$. This shows that the equation \eqref{equation-schro} has the $H^1$-norm evolution property.
Next, we consider the sequence of measures ${(\nu^nP_{t_n})(dv)}$ as measures on the space $L^2_{\rm loc}(\mathbb R^d)$. Denote by $B_k\subseteq \mathbb R^d$ the ball with radius $k$, centered at the origin. Given the uniform estimates \eqref{h1-bounds}, by the compact embedding of the space $H^1(B_k)$ in $L^2(B_k)$ and a successive application of Prokhorov's theorem, we obtain that there exists a subsequence of $\{t_n,u^n_0\}$, which we still denote $\{t_n,u^n_0\}$, and a distribution $\mu$ on $H^1(\mathbb R^d)$ such that
\begin{equation}{(\nu^nP_{t_n})(dv)}\to\mu\mbox{~~in distribution in~}L^2_{\rm loc}(\mathbb R^d)\hbox{\huge\textbullet}\end{equation}
We also note that the solutions $u^n$ satisfy the assumptions (iv) and (v) as consequences of \eqref{h1-bounds} and \eqref{lemma-aldous-criterion} respectively.
Thus by
Lemma~\ref{cv-loc-hi}
and
Theorem~\ref{thm-limit-conservation},
the convergence
\begin{equation}{(\nu^nP_{t_n})(dv)}\to\mu\end{equation}
is in fact in distribution in $H^1(\mathbb R^d)$ which is what we claimed.
{\it Proof of (ii):}
We choose $(s_n,v_n)\in [0,T]\times K$. By the compactness of the two sets there exist a subsequence of $(s_n,v_n)$, still denoted $(s_n,v_n)$, and $(s,v)\in[0,1]\times K$ such that $(s_n,v_n)\to (s,v)$. We claim that $P_{s_n}(v_n,\cdot)$ converges in distribution in $H^1(\mathbb R^d)$ to $P_s(v,\cdot)$.
We denote by $u^n$ and $u$ the solutions of \eqref{equation-schro} with initial data $v^n$ and $v$ respectively. In order to show this convergence we prove that we have
\begin{align}\label{conv-ii}\sup_{t\in[0,1]} \bigl(\Vert u^n_t -u_t\Vert_{H^1}+\Vert u_{s_n}-u_s\Vert_{H^1}\bigr)\to 0,\,\mathbb P\mbox{-a.s}.\end{align}
The convergence $\Vert u_{s_n}-u_s\Vert_{H^1}\to 0$ is a direct consequence of $u\in C([0,1],H^1),\,\mathbb P$-a.s.
It is shown in \cite{DD} that
\begin{equation}
\int_0^\cdot S^\lambda(\cdot-r)\Phi dW_r\in C([0,1];H^1(\mathbb R^d))\cap L^{\fractext{4(\sigma+1)}{\sigma d}}(0,1,W^{1,2\sigma +2}(\mathbb R^d))
\qquad
\mathbb P\mbox{-a.s.}
\label{EQ04}
\end{equation}
Thus, applying \cite[Proposition~3.5]{DD},
we also have $\mathbb P$-a.s. $|u^n-u|_{C([0,1];H^1(\mathbb R^d))}\to0$ as $n\to\infty$.
We now show that \eqref{conv-ii} implies the convergence $P_{s_n}(v^n,\cdot) \to P_s(v,\cdot)$.
We pick $\xi\colon H^1(\mathbb R^d)\to\mathbb R$ uniformly continuous and bounded. Then
\begin{align}
|P_s\xi (v)-P_{s_n}\xi(v^n)|&\leq \mathbb E[|\xi(u_s)-\xi(u^n_{s_n})|]
\nonumber
\\
&\leq \mathbb E[|\xi(u_s)-\xi(u_{s_n})|]+ \mathbb E[|\xi(u_{s_n})-\xi(u^n_{s_n})|]
\hbox{\huge\textbullet}
\end{align}
Note that \eqref{conv-ii} and the uniform continuity of $\xi$ imply that $\mathbb P$-a.s. $|\xi(u_s)-\xi(u_{s_n})|+|\xi(u_{s_n})-\xi(u^n_{s_n})|\to 0$ as $n\to\infty$. By the dominated convergence theorem, we obtain $|P_s\xi (v)-P_{s_n}\xi(v^n)|\to 0$ as $n\to \infty$.
\qed
\section{Compactness of the set of invariant measures}
In this section, we establish the existence
of an ergodic measure.
\cole
\begin{Theorem}
\label{T.compactness}
Under Assumptions~\ref{assumption-existence-schro},
the set of $H^1(\mathbb R^d)$-valued invariant measures is a convex and compact subset of
the space of probability measures on $H^1(\mathbb R^d)$.
\end{Theorem}
\colb
\begin{proof}
Note that the convexity is trivial,
so we only need to show compactness.
Let $\mu$ be such a measure and $(u(t))$ the solution of \eqref{equation-schro} having distribution $\mu$ at all time. For simplicity of notation, we denote $M_s=M(u(s))$ and $H_s=H(u(s))$. Our first objective is to prove the integrability of these semi-martingales.
We fix $R_0,R>0$ and define $\tau_{R}:=\inf\{s\geq 0: M_s\geq R\}$. We apply \eqref{evolution-M} on the event ${\{M_0\leq R_0\}}$ and obtain
\begin{align}
M_{t\wedge \tau_R} =&M_0 e^{-2\lambda {t\wedge \tau_R}} +\Vert\Phi \Vert_{HS(L^2; L^2)}^2 \int_0^{t\wedge \tau_R} e^{-2\lambda({t\wedge \tau_R}-s)} ds \\
&+2 \int_0^{t\wedge \tau_R} e^{-2\lambda({t\wedge \tau_R}-s)}Re(u(s),\Phi e_i) dB^i_s
\hbox{\huge\textbullet}
\end{align}
Note that by the localization the expectation of the stochastic integral vanishes. Therefore,
\begin{equation}\mathbb E\left[M_{t\wedge \tau_R}{\mathbf{1}}_{\{M_0\leq R_0\}}\right] =\mathbb E\left[M_0{\mathbf{1}}_{\{M_0\leq R_0\}} e^{-2\lambda {t\wedge \tau_R}} \right]+\Vert\Phi \Vert_{HS(L^2; L^2)}^2\mathbb E\left[ \int_0^{t\wedge \tau_R} e^{-2\lambda({t\wedge \tau_R}-s)} {\mathbf{1}}_{\{M_0\leq R_0\}} ds\right]\hbox{\huge\textbullet}\end{equation}
For fixed $R_0$, the integrands on the right hand side are uniformly bounded and the integrand on the left hand side is non-negative. We apply the dominated convergence theorem for the right side and Fatou's lemma for the left to obtain that
\begin{equation}\mathbb E\left[M_t{\mathbf{1}}_{\{M_0\leq R_0\}} \right]\leq \mathbb E\left[M_0{\mathbf{1}}_{\{M_0\leq R_0\}} \right]e^{-2\lambda t}+\frac{\Vert\Phi \Vert_{HS(L^2; L^2)}^2}{2\lambda}\hbox{\huge\textbullet} \end{equation}
Therefore, we can choose $t_{R_0}>0$ such that for all ${R_0}>0$, we have
\begin{equation}\mathbb E\left[M_{t_{R_0}}{\mathbf{1}}_{\{M_0\leq {R_0}\}} \right]\leq \frac{\Vert\Phi \Vert_{HS(L^2; L^2)}^2}{\lambda}\hbox{\huge\textbullet}\end{equation}
Noting also that the distribution of $M_{t_{R_0}}$ is $\mu$ we obtain there exists $f_{R_0}(v) \to 1$ $\mu$-a.s. as ${R_0}\to\infty$ and
\begin{equation}\mathbb E\left[M_{t_{R_0}}{\mathbf{1}}_{\{M_0\leq {R_0}\}} \right]=\int \Vert v\Vert_{L^2}^2f_{R_0}(v)\mu(dv)\leq \frac{\Vert\Phi \Vert_{HS(L^2; L^2)}^2}{\lambda}\hbox{\huge\textbullet}\end{equation}
Taking the limit ${R_0}\to \infty$, we obtain
\begin{equation}
\int \Vert v\Vert_{L^2}^2\mu(dv)\leq \frac{\Vert\Phi \Vert_{HS(L^2; L^2)}^2}{\lambda}
\hbox{\huge\textbullet}
\label{EQ07}
\end{equation}
Similarly to the proof of Lemma~\ref{bounds}, we apply Ito's lemma to $M^{k+1}(u(t))$, localize with stopping times and prove that there exists $C_k(\Phi,\lambda)$ which may a~priori depend on $\mu$ such that
\begin{equation}\int \Vert v\Vert_{L^2}^{2k}\mu (dv)\leq C_k(\Phi,\lambda)<\infty,
\comma k=1\ldots\hbox{\huge\textbullet}\end{equation}
We also apply the same procedure to $H_t$ to obtain that there exists $\widetilde C_k (\Phi,\lambda)$ that may again depend on $\mu$ such that
\begin{equation}\int \Vert v\Vert_{H^1}^{2k}\mu (dv)\leq \widetilde C_k(\Phi,\lambda)<\infty\hbox{\huge\textbullet}\end{equation}
Given this integrability, we return to \eqref{evolution-M} and\eqref{evolution-H} to prove that $C_k(\Phi,\lambda)$ and $\widetilde C_k (\Phi,\lambda)$ can be taken independent of $\mu$. Since $\mu$ is an invariant measure, we get $d \mathbb E[M_t]=d\mathbb E[H_t]=0$ and
\begin{equation}\mathbb E[M_t]= \frac{\Vert\Phi \Vert_{HS(L^2; L^2)}^2}{2\lambda}\hbox{\huge\textbullet}\end{equation}
Using the same invariance we obtain
\begin{align}
2\lambda \mathbb E[M^{k+1}_t]&= \Vert \Phi\Vert_{HS(L^2,L^2)}^2 \mathbb E[M^{k}_t] +\frac{k}{2} \mathbb E[M^{k-1}_t\sum_i Re (u(t),\Phi e_i)^2]\\
&\leq \left(\Vert \Phi\Vert_{HS(L^2,L^2)}^2 + \frac{k}{2}\right) \mathbb E[M^{k}_t],
\end{align}
which shows by induction that $C_k(\Phi,\lambda)$ may be taken independent of $\mu$.
Applying the same procedure to the equation \eqref{evolution-H}, we obtain that $\widetilde C_k(\Phi,\lambda)$ can be taken independent of $\mu$.
We now prove the sequential compactness of the set of $H^1(\mathbb R^d)$-valued invariant measures. Let $\mu^n$ be a sequence of such invariant measures of the equation \eqref{equation-schro}.
Without loss of generality, we assume that the $\sigma$-algebra $\mathbb F_0$ is rich enough so that there exists a family of $\mathbb F_0$-measurable random variables $u^n_0$ with distribution $\mu^n$.
The uniform bounds we have proven gives us
\begin{equation}\sup_n \int \Vert v\Vert_{L^2}^{2k}\mu^n (dv)\leq C_k(\Phi,\lambda)
\end{equation}
and
\begin{equation}
\sup_n \int \Vert v\Vert_{H^1}^{2}\mu^n (dv)\leq \widetilde C_k (\Phi,\lambda),\end{equation}
which a~fortiori imply
\begin{equation}\mathbb E\left[\Vert u^n_0\Vert _{H^1}^{4\vee\lceil{4d\sigma}\rceil}+\Vert u^n_0\Vert _{L^2}^{\lceil{4\sigma(2-d)+8}\rceil}\right]\leq R\hbox{\huge\textbullet}\end{equation}
Therefore, Lemma~\ref{lemma-tightness} and the fact that $\mu^n$ is an invariant measure show that the family
\begin{equation}\bigl\{(\mu^n P_{t_n})(\cdot):n\in\mathbb N\bigr\}=\{\mu^n:n\in\mathbb N\}\end{equation} is tight. Noting that the set of invariant measures is closed, we obtain the required compactness.
\end{proof}
\cole
\begin{Corollary}
\label{c.ergodic}
Under Assumptions~\ref{assumption-existence-schro}, there exists an ergodic invariant measure.
\end{Corollary}
\colb
\begin{proof}
By the Krein-Milman theorem, the compactness of the set of invariant measures implies that there exists at least one invariant measure that is an extremal point of this set. Proposition 3.2.7 of \cite{DZ} then implies that such a measure is ergodic.
\end{proof}
\section*{Acknowledgments}
I.K.~was supported in part by the NSF grant DMS-1311943,
while M.Z.~was supported in part by the NSF grant DMS-1109562.
|
\section{Introduction} \label{s1}
It has been designed a great number of iterative methods for solving various
optimization problems. The custom optimization problem consists in finding
an element in a feasible set $X \subseteq \mathbb{R}^{N}$ that yields
the minimal value of some goal function $\mu : \mathbb{R}^{N} \to \mathbb{R}$ on $X $. For brevity, we write
this problem as
\begin{equation} \label{eq:1.1}
\min \limits _{\mathbf{x} \in X} \to \mu (\mathbf{x}).
\end{equation}
It is well known that problems with the convex smooth goal function and convex feasible set
constitute one of the most investigated classes in optimization; see e.g. \cite{PD78,Pol83}.
The conditional gradient method is one of the oldest methods in this field. It was first suggested in
\cite{FW56} for the case when the goal function is quadratic and
the feasible set is polyhedral and further was developed by many authors;
see e.g. \cite{LP66,PD78,DR68,Dun80,Pol83}. We recall that the main idea of this method consists in
linearization of the goal function. That is, given the current iterate $\mathbf{x}^{k}\in X$, one finds some
solution $\mathbf{y}^{k}$ of the problem
\begin{equation} \label{eq:1.2}
\min_{\mathbf{y} \in X} \to \langle \mu'(\mathbf{x}^{k}),\mathbf{y}\rangle
\end{equation}
and defines $\mathbf{p}^{k}=\mathbf{y}^{k}-\mathbf{x}^{k}$ as a descent direction at $\mathbf{x}^{k}$.
Taking a suitable stepsize $\lambda_{k} \in (0,1]$, one sets $\mathbf{x}^{k+1}=\mathbf{x}^{k}+\lambda_{k}\mathbf{p}^{k}$
and so on.
During rather long time, this method was not considered as very efficient due to its
relatively slow convergence in comparison with Newton and projection type methods.
However, it has gained a great amount of attention very recently due to several
features significant for many applications, where huge dimensionality and inexact data
create certain drawbacks for more rapid methods. Moreover, in the case of
a polyhedral feasible set its auxiliary problem (\ref{eq:1.2}) appears simpler than those in the
other methods, and its solution yields usually so-called sparse approximations; see e.g.
\cite{Cla10,Jag13,FG16} and the references therein. It should be noted that
a great number of applications reduce to problem (\ref{eq:1.1}), where
\begin{equation} \label{eq:1.3}
\mu(\mathbf{x})=f(\mathbf{x})+h(\mathbf{x}),
\end{equation}
$f : \mathbb{R}^{N} \to \mathbb{R}$ is a smooth, but not
necessary convex function, and $h : \mathbb{R}^{N} \to \mathbb{R}$ is
not necessary smooth, but rather simple and convex function.
The appearance of the non-smooth term is caused by
regularization or exact penalty techniques; see e.g.
\cite{Pol83,FSS13}. In this case one can apply the partial linearization (PL for short) method
from \cite{MF81} (see \cite{Pat98,BLM09} for further development), where problem (\ref{eq:1.2}) is replaced with the following:
\begin{equation} \label{eq:1.4}
\min_{\mathbf{y} \in X} \to \langle \mu'(\mathbf{x}^{k}),\mathbf{y}\rangle + h(\mathbf{y}).
\end{equation}
The usefulness of this approach becomes clear if problem
(\ref{eq:1.1}), (\ref{eq:1.3}) is (partially) decomposable, which is typical for very
large dimensional problems.
For instance, let
$$
h(\mathbf{x}) = \sum \limits_{i} h_{i}(\mathbf{x}_{i}) \mbox{ and } \ X = \prod \limits_{i} X_{i}
$$
where $\mathbf{x}_{i} \in X_{i}$. Then (\ref{eq:1.4}) becomes equivalent to several
independent problems of the form
\begin{equation} \label{eq:1.5}
\min\limits _{\mathbf{y}_{i} \in X_{i}} \to \left\{
\left\langle \mathbf{y}_{i}, \frac{\partial f(\mathbf{x}^{k})}{\partial
\mathbf{x}_{i}}\right\rangle +h_{i}(\mathbf{y}_{i})\right\}.
\end{equation}
In case $h \equiv 0$, this decomposition method was considered in \cite{MESS67}.
However, even solution of all the partial problems of form (\ref{eq:1.5})
may appear too expensive. A randomized block-coordinate variant of the
conditional gradient method was rather recently proposed in \cite{LJSP13}.
A general scheme of block-descent methods for such problems was given in \cite{Pat99}.
We recall for instance that
various engineering problems based on the so-called group LASSO regression method
have this format (see \cite{YL06,MGB08}), as well as many problems of
network resource allocation in wireless multi-user interfering systems
(see \cite{SFSPP14}). We give several additional examples of such decomposable
applied problems in Section \ref{s6}.
The main goal of this paper is to suggest a modification of PL
methods for decomposable composite optimization problems
of form (\ref{eq:1.1}), (\ref{eq:1.3}), which maintains the basic
convergence properties, but enables one to reduce the computational
expenses per iteration. We follow the approach suggested in
\cite{Kon15d} for regularized splitting methods. The main difference of this method
consists in utilizing PL technique without any regularization in order to simplify the auxiliary problem,
but this implies the dis-continuity of the descent mapping and requires new
substantiation schemes. We take the inexact Armijo type linesearch rule,
which makes our method different from those in \cite{MF81,BLM09}
even in the non-decomposable case.
In what follows, we denote by
$\mathbb{R}^{s}$ the real $s$-dimensional Euclidean space, all
elements of such spaces being column vectors represented by a lower
case Roman alphabet in boldface, e.g. $\mathbf{x}$. We use
superscripts to denote different vectors, and subscripts to denote
different scalars or components of vectors. For any vectors
$\mathbf{x}$ and $\mathbf{y}$ of $\mathbb{R}^{s}$, we denote by $
\langle \mathbf{x}, \mathbf{y} \rangle $ their scalar product, i.e.,
$$
\langle \mathbf{x}, \mathbf{y} \rangle =\mathbf{x}^{\top} \mathbf{y} =\sum_{i=1}^{s} x_{i}y_{i},
$$
and by $\| \mathbf{x} \|$ the Euclidean norm of $\mathbf{x}$, i.e.,
$\| \mathbf{x}\|=\sqrt{ \langle \mathbf{x}, \mathbf{x} \rangle }$.
We denote by $\mathbb{R}^{s}_{+}$ the non-negative orthant in $\mathbb{R}^{s}$, i.e.
$\mathbb{R}^{s}_{+} = \{ \mathbf{u} \in \mathbb{R}^{s} \ | \ u_{i} \geq 0 \ i=1,\ldots,s \}$.
We also set $\mathcal{R}=\mathbb{R}\bigcup \{-\infty, +\infty\}$.
Given a function $f:\mathbb{R}^{s} \to \mathcal{R}$, we can define its domain
$$
{\rm dom} f=\{ \mathbf{x} \in \mathbb{R}^{s} \ | \ f(\mathbf{x}) > -\infty \}.
$$
For any set $X$, $\Pi (X)$ denotes the family of all nonempty subsets
of $X$.
\section{Problem formulation and preliminary properties}\label{s2}
We first formulate a partitionable optimization problem of form
(\ref{eq:1.1}), (\ref{eq:1.3}). We set
$\mathcal{N}=\{1,\ldots,N\}$ and suppose that there exists a
partition
$$
\mathcal{N}=\bigcup\limits_{i=1}^{n} \mathcal{N}_{i}
$$
with $|\mathcal{N}_{i}|=N_{i}$, $N=\sum \limits_{i=1}^{n} N_{i}$,
and $\mathcal{N}_{i}\bigcap \mathcal{N}_{j}=\varnothing$ if $i\neq
j$ such that
\begin{equation} \label{eq:2.1}
X=X_{1}\times \dots \times X_{n}=\prod \limits_{i=1}^{n} X_{i},
\end{equation}
where $X_{i}$ is a non-empty, convex, and compact set in
$\mathbb{R}^{N_{i}}$ for $i=1, \dots, n$. Then, any point
$\mathbf{x}=(x_{1}, \dots,x_{N})^{\top} \in \mathbb{R}^{N}$ is
represented by $\mathbf{x}=(\mathbf{x}_{1}, \dots,
\mathbf{x}_{n})^{\top}$ where $\mathbf{x}_{i}=(x_{j})_{j \in
\mathcal{N}_{i}} \in \mathbb{R}^{N_{i}}$ for $i=1, \dots, n$.
Also, we suppose that
\begin{equation} \label{eq:2.2}
h(\mathbf{x}) = \sum \limits^{n} _{i=1} h_{i}(\mathbf{x}_{i}),
\end{equation}
where $h_{i}: \mathbb{R}^{N_{i}} \rightarrow \mathcal{R}$ is convex, proper,
lower semi-continuous, and ${\rm dom} h_{i} \supseteq X_{i}$ for $i=1, \dots, n$. Then
the function
$h$ is also convex, proper, and lower semi-continuous and we can define its
subdifferential
$$
\partial h(\mathbf{x})=\partial h_{1}(\mathbf{x}_{1})\times \dots \times \partial
h_{n}(\mathbf{x}_{n}), \quad \forall \mathbf{x}\in X.
$$
So, our problem (\ref{eq:1.1}), (\ref{eq:1.3}),
(\ref{eq:2.1})--(\ref{eq:2.2}) is rewritten as
\begin{equation} \label{eq:2.3}
\min \limits _{\mathbf{x} \in X_{1}\times \dots \times X_{n}} \to
\mu(\mathbf{x})=\left\{f(\mathbf{x})+\sum \limits^{n} _{i=1} h_{i}(\mathbf{x}_{i})\right\}.
\end{equation}
Its solution set will be denoted by $X^{*}$ and the optimal value of the
function by $\mu^{*}$, i.e.
$$
\mu^{*} = \inf \limits _{ \mathbf{x} \in X} \mu(\mathbf{x}).
$$
We suppose that the function $f : \mathbb{R}^{N} \to
\mathbb{R}$ is smooth, but not necessary convex. Set
$\mathbf{g}(\mathbf{x})=f'(\mathbf{x})$, then
$$
\mathbf{g}(\mathbf{x})=(\mathbf{g}_{1}(\mathbf{x}), \dots,
\mathbf{g}_{n}(\mathbf{x}))^{\top}, \ \mbox{where} \
\mathbf{g}_{i}(\mathbf{x})=\left(\frac{\partial
f(\mathbf{x})}{\partial x_{j}}\right)_{j \in \mathcal{N}_{i}} \in
\mathbb{R}^{N_{i}}, \ i=1, \dots, n.
$$
From the assumptions above it follows that the function $\mu$ is
directionally differentiable at each point $\mathbf{x} \in X$, that
is, its directional derivative with respect to any vector
$\mathbf{d}$ is defined by the formula:
\begin{equation}\label{eq:2.4}
\mu '(\mathbf{x}; \mathbf{d})= \langle
{\mathbf{g} (\mathbf{x}), \mathbf{d}} \rangle + h'(\mathbf{x};
\mathbf{d}), \ \mbox{with} \ h'(\mathbf{x}; \mathbf{d})=\sum
\limits^{n} _{i=1} \max_{\mathbf{b}_{i} \in
\partial h_{i}(\mathbf{x}_{i})} \langle {\mathbf{b}_{i},
\mathbf{d}_{i}} \rangle ;
\end{equation}
see e.g. \cite{Cla83}.
We need the optimality condition for problem (\ref{eq:2.3}).
\begin{proposition} \label{pro:2.1} \cite[Proposition 2.1]{Kon15d}
(a) Each solution of problem (\ref{eq:2.3}) is a solution of the
mixed variational inequality (MVI for short): Find a point $\mathbf{x}^{*}
\in X=X_{1}\times \dots \times X_{n}$ such that
\begin{equation} \label{eq:2.5}
\begin{array}{c}
\displaystyle \sum \limits_{i=1}^{n} \left[ \langle {\mathbf{g}_{i}
(\mathbf{x}^{*}), \mathbf{y}_{i}-\mathbf{x}^{*}_{i}} \rangle +
h_{i}(\mathbf{y}_{i})-h_{i}(\mathbf{x}^{*}_{i}) \right] \geq 0 \\
\displaystyle \quad \forall \mathbf{y}_{i}\in X_{i}, \quad
\mbox{for} \ i=1, \dots, n.
\end{array}
\end{equation}
(b) If $f$ is convex, then each solution of MVI (\ref{eq:2.5})
solves problem (\ref{eq:2.3}).
\end{proposition}
In what follows, we denote by $X^{0}$ the solution set of MVI
(\ref{eq:2.5}) and call it the set of {\em stationary points} of
problem (\ref{eq:2.3}).
For each point $\mathbf{x}\in X$ we can define a point
$\mathbf{y}(\mathbf{x})=(\mathbf{y}_{1}(\mathbf{x}), \dots,
\mathbf{y}_{n}(\mathbf{x}))^{\top}\in X$ such that
\begin{equation} \label{eq:2.6}
\begin{array}{c}
\displaystyle \sum \limits_{i=1}^{n} \left[ \langle \mathbf{g}_{i}
(\mathbf{x}),
\mathbf{y}_{i}-\mathbf{y}_{i}(\mathbf{x}) \rangle +
h_{i}(\mathbf{y}_{i})-h_{i}(\mathbf{y}_{i}(\mathbf{x})) \right] \geq 0 \\
\displaystyle \quad \forall \mathbf{y}_{i}\in X_{i}, \quad
\mbox{for} \ i=1, \dots, n.
\end{array}
\end{equation}
This MVI gives a necessary and sufficient optimality condition for
the optimization problem:
\begin{equation} \label{eq:2.7}
\min \limits_{\mathbf{ y}\in X_{1}\times \dots \times X_{n}} \to \sum \limits_{i=1}^{n} \Phi _{i}
(\mathbf{x}, \mathbf{y}_{i}),
\end{equation}
where
\begin{equation} \label{eq:2.8}
\Phi _{i} (\mathbf{x}, \mathbf{y}_{i})= \langle \mathbf{g}_{i}
(\mathbf{x}),\mathbf{y}_{i}\rangle + h_{i}(\mathbf{y}_{i})
\end{equation}
for $i=1, \dots, n$; cf. (\ref{eq:1.4}). Under the above assumptions
the point $\mathbf{y}(\mathbf{x})$ exists, but is not defined uniquely in general,
hence we can define the set $Y(\mathbf{x})$ of these points at $\mathbf{x}$,
thus defining the set-valued mapping $\mathbf{ x}\mapsto Y(\mathbf{ x})$.
Observe that all the components of $\mathbf{y}(\mathbf{ x})$ can be
found independently, i.e. (\ref{eq:2.7})--(\ref{eq:2.8}) is
equivalent to $n$ independent optimization problems of the form
\begin{equation} \label{eq:2.9}
\min \limits_{\mathbf{ y}_{i}\in X_{i}} \to \Phi _{i} (\mathbf{x}, \mathbf{y}_{i}),
\end{equation}
for $i=1, \dots, n$ and $\mathbf{y}_{i}(\mathbf{x})$ just solves
(\ref{eq:2.9}). Therefore,
$$
Y(\mathbf{x}) = Y_{1}(\mathbf{x})\times \dots \times Y_{n}(\mathbf{x}),
$$
where each set $Y_{i}(\mathbf{x})$ is non-empty, convex, and compact.
Moreover, if we set
$$
\sigma _{i} (\mathbf{x}, \mathbf{y}_{i})=\Phi _{i} (\mathbf{x}, \mathbf{x}_{i})-\Phi _{i} (\mathbf{x}, \mathbf{y}_{i})= \langle \mathbf{g}_{i}
(\mathbf{x}),\mathbf{x}_{i}-\mathbf{y}_{i}\rangle + h_{i}(\mathbf{x}_{i})-h_{i}(\mathbf{y}_{i})
$$
and
$$
\varphi (\mathbf{x})= \sum \limits_{i=1}^{n} \varphi _{i} (\mathbf{x}), \
\varphi _{i} (\mathbf{x})= \max\limits_{\mathbf{ y}_{i}\in X_{i}}\sigma _{i} (\mathbf{x}, \mathbf{y}_{i}) \quad \mbox{for} \ i=1, \dots, n;
$$
then
$$
\varphi _{i} (\mathbf{x})= \sigma _{i} (\mathbf{x}, \mathbf{y}_{i}(\mathbf{x})), \ i=1, \dots, n;
$$
for any $\mathbf{y}(\mathbf{x})=(\mathbf{y}_{1}(\mathbf{x}), \dots,
\mathbf{y}_{n}(\mathbf{x}))^{\top} \in Y(\mathbf{x})$. We can choose the most suitable format for
the definition of a point of $Y(\mathbf{x})$.
We recall that given a set $V \subseteq \mathbb{R}^{s}$,
a set-valued mapping $Q: V \to \Pi (\mathbb{R}^{s})$ is said to be {\em closed} on a set
$W \subseteq V$, if for each pair of sequences $\{ \mathbf{u}^{k} \} \to \mathbf{u}$, $\{ \mathbf{q}^{k} \} \to \mathbf{q}$
such that $\mathbf{u}^{k} \in W$ and $\mathbf{q}^{k} \in Q(\mathbf{u}^{k})$,
we have $\mathbf{q} \in Q(\mathbf{u})$.
We also need continuity type properties of the marginal functions.
\begin{lemma} \label{lm:2.1}
(a) The function $\varphi : \mathbb{R}^{N} \to
\mathcal{R}$ is lower semi-continuous on $X$;
(b) The mapping $\mathbf{x}\mapsto Y(\mathbf{x})$ is
closed on $X$.
\end{lemma}
{\bf Proof.}
Assertion (a) has been proved in \cite[Lemma 4]{BLM09}. To obtain (b), take sequences
$\{\mathbf{x}^{k}\} \to \bar \mathbf{x}$, $\{\mathbf{y}^{k}\} \to \bar \mathbf{y}$ with
$\mathbf{y}^{k} \in Y(\mathbf{x}^{k})$. Then from (\ref{eq:2.3}) we have
$$
\begin{array}{c}
\displaystyle \sum \limits_{i=1}^{n} \left[ \langle \mathbf{g}_{i}
(\mathbf{x}^{k}), \mathbf{u}_{i}-\mathbf{y}^{k}_{i} \rangle +
h_{i}(\mathbf{u}_{i})-h_{i}(\mathbf{y}^{k}_{i}) \right] \geq 0 \\
\displaystyle \quad \forall \mathbf{u}_{i}\in X_{i}, \quad
\mbox{for} \ i=1, \dots, n.
\end{array}
$$
Since $\mathbf{g}$ is continuous and $h$ is lower semi-continuous,
taking the limit $k \to \infty$ gives
$$
\begin{array}{c}
\displaystyle \sum \limits_{i=1}^{n} \left[ \langle \mathbf{g}_{i}
(\bar \mathbf{x}), \mathbf{u}_{i}-\bar \mathbf{y} \rangle +
h_{i}(\mathbf{u}_{i})-h_{i}(\bar \mathbf{y}) \right] \geq 0 \\
\displaystyle \quad \forall \mathbf{u}_{i}\in X_{i}, \quad
\mbox{for} \ i=1, \dots, n;
\end{array}
$$
hence $\bar \mathbf{y} \in Y(\bar \mathbf{x})$ and $\mathbf{x}\mapsto Y(\mathbf{x})$
is closed.
\QED
We now show that $\varphi$ can serve as a gap function for problem (\ref{eq:2.3}).
\begin{proposition} \label{pro:2.2}
(a) For any point $\mathbf{x}\in X$ it holds that $\varphi (\mathbf{x})\geq 0$, or, equivalently,
$\varphi_{i} (\mathbf{x})\geq 0$ for $i=1, \dots, n$;
(b) $ \mathbf{x}\in X^{0} \Longleftrightarrow \mathbf{x}\in Y(\mathbf{ x})\Longleftrightarrow
\varphi (\mathbf{x})= 0 \Longleftrightarrow \varphi_{i} (\mathbf{x})= 0, i=1, \dots, n$;
\end{proposition}
{\bf Proof.} Since $\sigma _{i} (\mathbf{x}, \mathbf{x}_{i})=0$, assertion (a) is true.
Next, if $ \mathbf{x}=\mathbf{y}(\mathbf{ x}) \in Y(\mathbf{ x})$, then
(\ref{eq:2.6}) implies $\mathbf{x}\in X^{0}$, $\varphi (\mathbf{x}) \leq 0$ and $\varphi_{i} (\mathbf{x})\leq 0$
for $ i=1, \dots, n$, hence, by (a), $\varphi (\mathbf{x})= 0$ and $\varphi_{i} (\mathbf{x})= 0$
for $ i=1, \dots, n$. Conversely, let
$\mathbf{x}$ solve MVI (\ref{eq:2.5}), but $ \mathbf{x}\notin Y(\mathbf{x})$ or $\varphi (\mathbf{x})> 0$.
Then there exists an index $l$ and a point $\mathbf{x}'_{l} \in X_{l}$ such that
$\sigma _{l} (\mathbf{x}, \mathbf{x}'_{l})>0$. It follows that
$$
-\sum \limits_{i\neq l} \sigma _{i} (\mathbf{x}, \mathbf{x}_{i})- \sigma _{l} (\mathbf{x}, \mathbf{x}'_{l})<0,
$$
i.e. $\mathbf{x} \notin X^{0}$, which is a contradiction. This means that
assertion (b) is true. \QED
We see that the value $\varphi(\mathbf{x})$ can serve as accuracy measure at a point
$\mathbf{x}$. We establish now a useful descent property. Define for brevity
$I=\{1, \dots, n\}$.
\begin{lemma} \label{lm:2.2} Take
any points $\mathbf{x} \in X$, $\mathbf{y}(\mathbf{x}) \in Y(\mathbf{x})$ and an index $s \in I$. If
$$
\mathbf{d}_{i}= \left\{ {
\begin{array}{ll}
\displaystyle
\mathbf{y}_{s}(\mathbf{x})-\mathbf{x}_{s} \quad & \mbox{if} \ i=s, \\
\mathbf{0} \quad & \mbox{if} \ i\neq s; \\
\end{array}
} \right.
$$
then
\begin{equation} \label{eq:2.11}
\mu' (\mathbf{x}; \mathbf{d}) \leq -\varphi_{s}(\mathbf{x}).
\end{equation}
\end{lemma}
{\bf Proof.} Due to the definition of $\mathbf{d}$ and (\ref{eq:2.4}), we have
$$
\mu '(\mathbf{x}; \mathbf{d})= \langle {\mathbf{g} (\mathbf{x}),
\mathbf{d}} \rangle + h'(\mathbf{x}; \mathbf{d})=\langle
\mathbf{g}_{s} (\mathbf{x}), \mathbf{d}_{s}\rangle +
\max_{\mathbf{b}_{s} \in
\partial h_{s}(\mathbf{x}_{s})} \langle {\mathbf{b}_{s},
\mathbf{d}_{s}} \rangle.
$$
By convexity, we have
$$
\langle {\mathbf{b}_{s},
\mathbf{d}_{s}} \rangle \leq
h_{s}(\mathbf{y}_{s}(\mathbf{x}))-h_{s}(\mathbf{x}_{s})
$$
for any $\mathbf{b}_{s} \in\partial h_{s}(\mathbf{x}_{s})$. It follows that
$$
\mu '(\mathbf{x}; \mathbf{d}) \leq \langle
\mathbf{g}_{s} (\mathbf{x}), \mathbf{y}_{s}(\mathbf{x})-\mathbf{x}_{s} \rangle +
h_{s}(\mathbf{y}_{s}(\mathbf{x}))-h_{s}(\mathbf{x}_{s})=-\varphi_{s}(\mathbf{x}),
$$
hence (\ref{eq:2.11}) holds true. \QED
\section{The descent method with inexact line-search}\label{s3}
Denote by $\mathbb{Z}_{+}$ the set of non-negative integers. The
basic cycle of the descent PL method with inexact line-search
for MVI (\ref{eq:2.5}) is described as follows.
\medskip
\noindent {\bf Basic cycle (PL).} Choose a point $\mathbf{x}^{0}\in X$
and numbers $\delta > 0$, $\beta \in (0,1)$, $\theta
\in (0,1)$.
At the $k$-th iteration, $k=0,1,\ldots$, we have a point
$\mathbf{x}^{k}\in X$.
{\em Step 1:} Choose an index $s \in I$ such that
$\varphi_{s}(\mathbf{x}^{k}) \geq \delta$, set $s_{k}=s$,
$$
\mathbf{d}^{k}_{i}= \left\{ {
\begin{array}{ll}
\displaystyle
\mathbf{y}_{s}-\mathbf{x}^{k}_{s} \quad & \mbox{if} \ i=s_{k}, \\
\mathbf{0} \quad & \mbox{if} \ i\neq s_{k}; \\
\end{array}
} \right.
$$
where $\mathbf{y}_{s}=\mathbf{ y}_{s}(\mathbf{x}^{k})\in Y_{s}(\mathbf{x}^{k})$ and go to Step 3.
Otherwise (i.e. when $\varphi_{i}(\mathbf{x}^{k}) <
\delta$ for all $i \in I$) go to Step 2.
{\em Step 2:} Set $\mathbf{z}=\mathbf{x}^{k}$ and stop.
{\em Step 3:} Determine $m$ as the smallest number in
$\mathbb{Z}_{+}$ such that
\begin{equation} \label{eq:3.1}
\mu (\mathbf{x}^{k}+\theta ^{m} \mathbf{d}^{k})
\leq \mu (\mathbf{x}^{k})-\beta \theta ^{m}\varphi_{s}(\mathbf{x}^{k}),
\end{equation}
set $\lambda_{k}=\theta ^{m}$,
$\mathbf{x}^{k+1}=\mathbf{x}^{k}+\lambda_{k}\mathbf{d}^{k}$, and
$k=k+1$. The iteration is complete.
\medskip
\begin{lemma} \label{lm:3.1} The line-search procedure in Step 3 is always finite.
\end{lemma}
{\bf Proof.}
If we suppose that the line-search procedure is infinite, then
$$
\theta^{-m}(\mu (\mathbf{x}^{k}+\theta ^{m} \mathbf{d}^{k}) - \mu
(\mathbf{x}^{k}))>-\beta \varphi_{s}(\mathbf{x}^{k}),
$$
for $m \to \infty$, hence, by taking the limit we have $ \mu'
(\mathbf{x}^{k};\mathbf{d}^{k}) \geq -\beta \varphi_{s}(\mathbf{x}^{k})$,
but Lemma \ref{lm:2.2} gives $ \mu'
(\mathbf{x}^{k};\mathbf{d}^{k}) \leq -\varphi_{s}(\mathbf{x}^{k})$, hence $ (1-\beta
)\varphi_{s}(\mathbf{x}^{k}) \leq 0$, a contradiction. \QED
We recall that a single-valued mapping $\mathbf{p} : \mathbb{R}^{s} \to \mathbb{R}^{s}$
is said to be {\em uniformly continuous} on a set
$V \subset \mathbb{R}^{s}$, if for any number $\varepsilon >0$ there exists a number $\tau >0$
such that $\| \mathbf{p}(\mathbf{x})-\mathbf{p}(\mathbf{y})\| < \varepsilon$ for
each pair of points $\mathbf{x}, \mathbf{y} \in V$ with $\| \mathbf{x}-\mathbf{y}\| < \tau$.
Our convergence analysis will be based on the following property.
\begin{proposition} \label{pro:3.1}
Suppose in addition that the gradient map $\mathbf{g} : \mathbb{R}^{N} \to \mathbb{R}^{N}$
is uniformly continuous on $X$. Then the number of iterations in Basic cycle (PL) is finite.
\end{proposition}
{\bf Proof.}
By construction, we have $- \infty < \mu ^{*}\leq \mu
(\mathbf{x}^{k})$ and $\mu (\mathbf{x}^{k+1})\leq \mu
(\mathbf{x}^{k})-\beta \delta \lambda_{k}$, hence
\begin{equation} \label{eq:3.1a}
\lim \limits_{k\rightarrow \infty }\lambda_{k}=0.
\end{equation}
Besides, the sequence $\{\mathbf{x}^{k}\}$ is bounded and must have limit points,
Suppose that the sequence $\{\mathbf{x}^{k}\}$ is infinite. Since
the set $I$ is finite, there is an index $s_{k}=s$, which is
repeated infinitely. Take the corresponding subsequence $\{k_{l}\}$.
We intend to evaluate the
difference $\mu (\mathbf{x}^{k_{l}}+\lambda_{k_{l}}\mathbf{d}^{k_{l}}) -
\mu (\mathbf{x}^{k_{l}})$, but we temporarily remove these indices for more convenience.
Then, using the mean value theorem and convexity of $h_{i}$, we have
\begin{eqnarray*}
\displaystyle && \mu (\mathbf{x}+\lambda \mathbf{d})
- \mu (\mathbf{x}) =f (\mathbf{x}+\lambda \mathbf{d})
- f(\mathbf{x}) + h_{s}(\mathbf{x}_{s}+\lambda \mathbf{d}_{s})-h_{s}(\mathbf{x}_{s}) \\
&& \leq \lambda \left\{\langle \mathbf{g}_{s}(\mathbf{x}),\mathbf{y}_{s}-\mathbf{x}_{s} \rangle
+ h_{s}(\mathbf{y}_{s})-h_{s}(\mathbf{x}_{s})\right\}
+ \lambda \langle \mathbf{g}_{s}(\mathbf{x}+\xi\lambda \mathbf{d})-\mathbf{g}_{s}(\mathbf{x}),\mathbf{y}_{s}-\mathbf{x}_{s} \rangle \\
&& \leq -\lambda \varphi_{s}(\mathbf{x})+
\lambda \|\mathbf{g}_{s}(\mathbf{x}+\xi\lambda \mathbf{d})-\mathbf{g}_{s}(\mathbf{x}) \| \|\mathbf{d}_{s} \|,
\end{eqnarray*}
where $\xi=\xi_{k_{l}} \in (0,1)$. Since $X_{s}$ is bounded, $ \|\mathbf{d}_{s} \|\leq C_{s}< \infty$.
Due to the uniform continuity of $\mathbf{g}$, there exists a number $\lambda'>0$
such that
$$
\|\mathbf{g}_{s}(\mathbf{x}+\xi\lambda \mathbf{d})-\mathbf{g}_{s}(\mathbf{x}) \| \leq (1-\beta) \delta/C_{s}
$$
if $\lambda \leq \lambda'$, besides, $\varphi_{s}(\mathbf{x}) \geq \delta$.
It follows that
$$
\mu (\mathbf{x}+\lambda \mathbf{d})
- \mu (\mathbf{x}) \leq -\lambda \varphi_{s}(\mathbf{x})+ \lambda (1-\beta) \delta
\leq -\beta\lambda \varphi_{s}(\mathbf{x})
$$
if $\lambda \leq \lambda'$, hence $\lambda_{k_{l}} \geq \bar \lambda >0$ by the stepsize rule in Basic cycle (PL),
which contradicts (\ref{eq:3.1a}). \QED
The whole method involves the upper level whose iterations (stages)
contain Basic cycle (LP) with decreasing values of $\delta$.
\medskip
\noindent {\bf Method (Upper level).} Choose a point
$\mathbf{z}^{0}\in X$ and a sequence $\{\delta _{l}\} \searrow 0$.
At the $l$-th stage, $l=1,2,\ldots$, we have a point
$\mathbf{z}^{l-1}\in X$ and a number $\delta _{l}$. Apply Basic
cycle (LP) with $\mathbf{x}^{0}=\mathbf{z}^{l-1}$, $\delta=\delta _{l}$
and obtain a point $\mathbf{z}^{l}= \mathbf{z}$ as its output.
\medskip
\begin{theorem} \label{thm:3.1} Suppose in addition that the gradient
map $\mathbf{g} : \mathbb{R}^{N} \to \mathbb{R}^{N}$
is uniformly continuous on $X$. Then the
sequence $\{\mathbf{z}^{l}\}$ generated by the method with Basic
cycle (LP) has limit points, all these limit points are solutions of
MVI (\ref{eq:2.5}). Besides, if $f$ is convex, then
\begin{equation} \label{eq:3.2}
\lim \limits_{l\rightarrow \infty} \mu (\mathbf{z}^{l})=\mu^{*};
\end{equation}
and all the limit points of $\{\mathbf{z}^{l}\}$ belong to $X^{*}$.
\end{theorem}
{\bf Proof.}
Following the proof of Proposition \ref{pro:3.1}, we see that
$\mu (\mathbf{z}^{l+1})\leq \mu (\mathbf{z}^{l})$, hence
$$
\lim \limits_{l\rightarrow \infty }\mu (\mathbf{z}^{l})=\tilde \mu.
$$
Besides, the
sequence $\{\mathbf{z}^{l}\}$ is bounded and must have limit points.
Take an arbitrary limit point $\mathbf{\bar z}$ of
$\{\mathbf{z}^{l}\}$, then
$$
\lim \limits_{t\rightarrow \infty }\mathbf{z}^{l_{t}}=\mathbf{\bar
z}.
$$
For $l>0$ we have
$$
\varphi_{i}(\mathbf{z}^{l}) \leq \delta_{l} \ \mbox{for all} \ i \in I,
$$
hence $\varphi(\mathbf{z}^{l}) \leq n \delta_{l}$. Due to
Lemma \ref{lm:2.1}, taking the limit $l=l_{t}\rightarrow
\infty$, we obtain $\varphi(\mathbf{\bar z}) \leq0$ and
$\mathbf{\bar z} \in X$. Due to Proposition
\ref{pro:2.2}, this means that $\varphi(\mathbf{\bar z}) =0$
and that the point $\mathbf{\bar z}$
solves MVI (\ref{eq:2.5}). Next, if $f$ is convex, then by
Proposition \ref{pro:2.1} (b), each limit point of
$\{\mathbf{z}^{l}\}$ solves problem (\ref{eq:2.3}). It follows that
$\tilde \mu=\mu^{*}$ and (\ref{eq:3.2}) holds. \QED
In case $h \equiv 0$, the method is a new decomposable version of the
conditional gradient method.
Although the dimensions $N_{i}$ can be arbitrary, we think that the proposed PL method
may have preferences, in particular, over the method from \cite{Kon15d}, in case when $N_{i}>1$
and all the sets $X_{i}$ are polyhedrons. Also, it may have preferences
over the usual conditional gradient and partial linearization methods
if the number of subsets $n$ is rather large.
\begin{remark} \label{rm:3.1} The initial boundedness requirement for the feasible set
$X$ was made in Section \ref{s2} only for more simplicity of exposition and
can be replaced with proper coercivity assumptions. In fact, instead of compactness of each set
$X_{i}$ we can require their closedness and add e.g. the following conditions.
{\bf (C1)} {\em For each $i\in I $ and for each
sequence $\{\mathbf{u}^{l}_{i}\}$ such that $\mathbf{u}^{l}_{i}\in X_{i}$ and
$\{\|\mathbf{u}^{l}_{i}\|\}\to \infty$ as $l \to \infty$, we have
$\{h_{i}(\mathbf{u}^{l}_{i})/\|\mathbf{u}^{l}_{i}\|\}\to +\infty$.}
{\bf (C2)} {\em For each
sequence $\{\mathbf{u}^{l}\}$ such that $\mathbf{u}^{l}\in X$
and $\{\|\mathbf{u}^{l}\|\}\to \infty$ as $l \to \infty$, we have
$\{\mu(\mathbf{u}^{l})\}\to +\infty$.}
Then {\bf (C1)} provides existence of a solution of auxiliary problem (\ref{eq:2.7})--(\ref{eq:2.8}),
moreover, the sequence $\{\mathbf{d}^{k}\}$ is bounded if so is $\{\mathbf{x}^{k}\}$.
From {\bf (C2)} it follows that $ \mu ^{*} > - \infty$, problems
(\ref{eq:2.3}) and (\ref{eq:2.5}) have solutions, and that
the sequence $\{\mathbf{x}^{k}\}$ is bounded. Therefore,
all the assertions of Section \ref{s3} remain true.
Also, we supposed that ${\rm dom} h_{i} \supseteq X_{i}$ for $i=1, \dots, n$ only for
more simplicity of exposition. Set
$$
D=\prod \limits_{i=1}^{n} ({\rm dom} h_{i} \bigcap X_{i}).
$$
It suffices to assume $D \neq \varnothing$. Then we should only take the initial point
$\mathbf{z}^{0}\in D$.
\end{remark}
\section{Modifications of the linesearch procedure}\label{s4}
Due to Lemma \ref{lm:3.1} the current Armijo rule in (\ref{eq:3.1})
provides its finite implementation e.g. in comparison with
the one-dimensional minimization rule. This version can also be substantiated
under the same assumptions, but we are interested in developing
line-search procedures that are concordant to the partition of the space given
in Section \ref{s2} and do not require calculation of all the components of the
gradient and new point at each iteration. In fact, rule (\ref{eq:3.1})
involves some shift in one component $\mathbf{x}_{s}$,
but utilizes the value of the cost function
at the trial point. That is, we have to calculate the value of $f$ together
with only one component $h_{s}$.
Let us first consider the {\em convex case} where
the function $f$ is convex. Then, we can replace (\ref{eq:3.1}) with
the following:
\begin{equation} \label{eq:4.1}
\langle \mathbf{g}_{s} (\mathbf{x}^{k}+\theta^{m}\mathbf{d}^{k}),\mathbf{d}^{k}_{s}\rangle
+ \theta ^{-m}\left\{ h_{s}(\mathbf{x}^{k}_{s}+\theta ^{m}
\mathbf{d}^{k}_{s})-h_{s}(\mathbf{x}^{k}_{s}) \right\}\leq -\beta
\varphi_{s}(\mathbf{x}^{k}).
\end{equation}
Since the
trial point $\mathbf{x}^{k}+\theta ^{m}\mathbf{d}^{k}$ has the shift
from $\mathbf{x}^{k}$ only in $\mathbf{d}^{k}_{s}$, it can be
implemented independently of other variables. From
(\ref{eq:4.1}) it now follows that
\begin{eqnarray*}
\displaystyle && \mu (\mathbf{x}^{k}+\theta ^{m} \mathbf{d}^{k})
- \mu (\mathbf{x}^{k}) =f (\mathbf{x}^{k}+\theta ^{m} \mathbf{d}^{k})
- f(\mathbf{x}^{k}) +
h_{s}(\mathbf{x}^{k}_{s}+\theta ^{m} \mathbf{d}^{k}_{s})-h_{s}(\mathbf{x}^{k}_{s}) \\
&& \leq \theta ^{m} \langle \mathbf{g}_{s} (\mathbf{x}^{k}+\theta
^{m}\mathbf{d}^{k}),\mathbf{d}^{k}_{i}\rangle +
h_{s}(\mathbf{x}^{k}_{s}+\theta ^{m}
\mathbf{d}^{k}_{i})-h_{s}(\mathbf{x}^{k}_{s}) \leq -\beta
\theta ^{m} \varphi_{s}(\mathbf{x}^{k}),
\end{eqnarray*}
and (\ref{eq:3.1}) holds true. It easy to see that all the assertions of
Section \ref{s3} remain true for this version.
Moreover, we can utilize even a pre-defined stepsize in the {\em Lipschitz gradient case}.
Let us suppose that partial gradients of the function $f$ are
Lipschitz continuous, i.e.,
$$
\| \mathbf{g}_{i}
(\mathbf{x}+\mathbf{d}^{(i)})-\mathbf{g}_{i}(\mathbf{x})\| \leq
L_{i}\|\mathbf{d}^{(i)}\|=L_{i}\|\mathbf{d}_{i}\|
$$
for any vector $\mathbf{x}$, where
$$
\mathbf{d}^{(i)}_{j}= \left\{ {
\begin{array}{ll}
\displaystyle
\mathbf{d}_{i} \quad & \mbox{if} \ j=i, \\
\mathbf{0} \quad & \mbox{if} \ j\neq i; \\
\end{array}
} \right.
$$
for $i \in I$ and any vector $\mathbf{d}=(\mathbf{d}_{1},
\dots,\mathbf{d}_{n})^{\top} \in \mathbb{R}^{N}$.
Clearly, this property holds if the gradient of $f$ is Lipschitz continuous with
some constant $L >0$, then $L_{i} \leq L$ for each $i \in I$.
It is known that any function $\phi$ having the Lipschitz continuous gradient
satisfies the inequality
$$
\phi (\mathbf{y}) \leq \phi(\mathbf{x})+\langle
\phi'(\mathbf{x}),\mathbf{y}-\mathbf{x} \rangle +0.5L_{\phi}
\|\mathbf{y}-\mathbf{x} \|^{2};
$$
see \cite[Lemma 1.2]{DR68}. Similarly, for any vectors $\mathbf{x}$
and $\mathbf{d}$, we have
$$
f (\mathbf{x}+\mathbf{d}^{(i)}) \leq f(\mathbf{x})+\langle
\mathbf{g}_{i}(\mathbf{x}),\mathbf{d}_{i} \rangle +0.5L_{i}
\|\mathbf{d}_{i} \|^{2} \quad \forall i \in I.
$$
If $\mathbf{d}_{i}= \mathbf{y}_{i}(\mathbf{x})-\mathbf{x}_{i}$, then
we have
\begin{eqnarray*}
\displaystyle && \mu (\mathbf{x}+\lambda \mathbf{d}^{(i)})
- \mu (\mathbf{x}) =f (\mathbf{x}+\lambda \mathbf{d}^{(i)})
- f(\mathbf{x}) + h_{i}(\mathbf{x}_{i}+\lambda \mathbf{d}_{i})-h_{i}(\mathbf{x}_{i}) \\
&& \leq \lambda \left\{\langle \mathbf{g}_{i}(\mathbf{x}),\mathbf{d}_{i} \rangle
+ h_{i}(\mathbf{y}_{i}(\mathbf{x}))-h_{i}(\mathbf{x}_{i})\right\}
+0.5L_{i} \lambda^{2}\|\mathbf{d}_{i} \|^{2} \\
&& \leq -\lambda \varphi_{i}(\mathbf{x})+0.5L_{i} \lambda^{2}\|\mathbf{d}_{i} \|^{2}
\leq -\beta \lambda \varphi_{i}(\mathbf{x}),
\end{eqnarray*}
if
\begin{equation} \label{eq:4.2}
\lambda \leq \bar \lambda_{(i)}(\mathbf{x})=2(1-\beta) \varphi_{i}(\mathbf{x})/(\|\mathbf{d}_{i} \|^{2}L_{i}).
\end{equation}
It follows that (\ref{eq:3.1}) holds with $\lambda_{k} \geq \min\{1, \theta
\bar \lambda_{(s)}(\mathbf{x}^{k})\}>0$.
Moreover, we can simply set $\lambda_{k} = \lambda_{(s)}(\mathbf{x}^{k})>0$,
and all the assertions of Proposition \ref{pro:3.1} and Theorem \ref{thm:3.1}
remain true for this version. This modification reduces the computational expenses
essentially since calculations of the goal function values are not necessary and we can calculate
values of the partial gradients $\mathbf{g}_{i}$ and functions $h_{i}$ only for
necessary separate components. Clearly, the adaptive
PL method admits other modifications and extensions, e.g.
selection of a group of indices in $I$ instead of only one
component.
These opportunities make the method very flexible and suitable for
parallel and distributed computations applicable for very
high-dimensional optimization problems; see e.g.
\cite{BT89,Pat99,Jag13,FSS13}.
\section{Convergence rates}\label{s5}
In this section, we give some convergence rates for the adaptive PL method.
We suppose that all the basic assumptions of Section \ref{s2}
hold, but will also utilize some additional conditions.
We first establish the finite termination property
under the following {\em sharp solution condition}, which modifies those
in \cite[Chapter 7, \S 1, Section 3]{Pol83} and
\cite[Section 2.2]{Kon01}.
There exist a number $\tau > 0$ and a point $\bar \mathbf{x} \in X$ such that
$$
\langle \mathbf{g}(\bar \mathbf{x}), \mathbf{x} - \bar \mathbf{x}
\rangle + h (\mathbf{x}) - h (\bar \mathbf{x}) \geq \tau \| \mathbf{x} - \bar \mathbf{x} \| \quad \forall \mathbf{x} \in X.
$$
\begin{theorem} \label{thm:5.1}
Let a sequence $\{\mathbf{z}^{l}\}$ be generated by the method with Basic
cycle (LP). Suppose that the function $f$ is convex,
its gradient is Lipschitz continuous with constant $L < \infty$, and that
the sharp solution condition holds. Then
there exists a stage number $t$ such that $X^{*}= Y(\mathbf{z}^{t})$.
\end{theorem}
{\bf Proof.} First we note that the sharp solution condition implies
$\bar \mathbf{x} \in X^{0}$, and, by convexity, $X^{0}=X^{*}$; see Proposition \ref{pro:2.1}.
Next, if there exists some other point $\tilde \mathbf{x} \in X$,
which provides the sharp solution condition, then, again by convexity, we must have
\begin{eqnarray*}
&& \langle \mathbf{g} (\tilde \mathbf{x}),\bar \mathbf{x}-\tilde \mathbf{x} \rangle +
h(\bar \mathbf{x})-h(\tilde \mathbf{x})
\leq \langle \mathbf{g} (\bar \mathbf{x}),\bar \mathbf{x}-\tilde \mathbf{x} \rangle +
h(\bar \mathbf{x})-h(\tilde \mathbf{x}) \\
&& \leq -\tau \| \bar \mathbf{x} - \tilde \mathbf{x} \|<0,
\end{eqnarray*}
which is a contradiction. Hence, $X^{*}=\{\bar \mathbf{x} \}$.
From the sharp solution condition for any point $ \mathbf{x} \in X$ we have
\begin{eqnarray*}
&& \langle \mathbf{g} (\mathbf{z}^{l}),\bar \mathbf{x}-\mathbf{x} \rangle +
h(\bar \mathbf{x})-h(\mathbf{x}) \\
&& =\langle \mathbf{g} (\bar \mathbf{x}),\bar \mathbf{x}-\mathbf{x} \rangle +
h(\bar \mathbf{x})-h(\mathbf{x})
+ \langle \mathbf{g} (\mathbf{z}^{l})-\mathbf{g} (\bar \mathbf{x}),\bar \mathbf{x}-\mathbf{x} \rangle \\
&& \leq -\tau \| \bar \mathbf{x} - \mathbf{x} \| + L \| \mathbf{z}^{l}-\bar \mathbf{x}\| \| \bar \mathbf{x} - \mathbf{x} \| \\
&& = -\tau \| \bar \mathbf{x} - \mathbf{x} \| (1 - L \| \mathbf{z}^{l}-\bar \mathbf{x}\| ).
\end{eqnarray*}
From Theorem \ref{thm:3.1} we now have $ \{\| \mathbf{z}^{l}-\bar \mathbf{x}\|\}\to 0$
as $l\to +\infty$. Hence
$$
\langle \mathbf{g} (\mathbf{z}^{l}),\bar \mathbf{x}-\mathbf{x} \rangle +
h(\bar \mathbf{x})-h(\mathbf{x})<0 \quad \forall \mathbf{x} \in X, \mathbf{x}\neq \bar \mathbf{x},
$$
for $l$ large enough. It follows that there exists a number $t$
such that $ Y(\mathbf{z}^{t})=\{\bar \mathbf{x} \}$. \QED
In the method, each stage contains a finite number of iterations of
the basic cycle. Therefore, it seems suitable
to derive its complexity estimate, which gives the total amount of
work of the method. We now suppose in addition
that the function $f$ is convex and its partial gradients satisfy Lipschitz continuity
conditions with constants $L_{i}$ for each $i \in I$. Then it was
shown in Section \ref{s4} that we can take the stepsize
\begin{equation} \label{eq:5.1a}
\lambda_{k} = \lambda_{(s)}(\mathbf{x}^{k})
=2(1-\beta) \varphi_{s}(\mathbf{x}^{k})/(\|\mathbf{d}^{k}_{s} \|^{2}L_{s}) \geq
2(1-\beta) \varphi_{s}(\mathbf{x}^{k})/(\rho^{2}L),
\end{equation}
where
$$
L= \max_{s \in I}L_{s}, \ \rho= \max_{s \in I}\rho_{s}, \ \rho_{s}= {\rm diam} X_{s};
$$
see (\ref{eq:4.2}).
We take the value $\Phi(\mathbf{x})=\mu (\mathbf{x})- \mu ^{*}$ as an
accuracy measure for our method. In other words, given a starting
point $ \mathbf{z}^{0}$ and a number $\varepsilon > 0$, we define
the complexity of the method, denoted by $V(\varepsilon )$, as the
total number of iterations at $l(\varepsilon )$ stages such that
$l(\varepsilon )$ is the maximal number $l$ with
$\Phi(\mathbf{z}^{l}) \geq \varepsilon$, hence,
\begin{equation} \label{eq:5.1}
V (\varepsilon ) \leq \sum ^{l(\varepsilon )} _{l=1}
V_{l},
\end{equation}
where $V_{l}$ denotes the total number of iterations at stage
$l$. We proceed to estimate the right-hand side of (\ref{eq:5.1}).
To change $\delta _{l}$, we apply the geometric rate:
\begin{equation} \label{eq:5.2}
\delta _{l} = \nu ^{l}\delta_{0}, l=0,1,\ldots;
\quad \nu \in (0,1), \delta_{0}>0.
\end{equation}
By (\ref{eq:3.1}), we have
$$
\mu (\mathbf{x}^{k+1})
\leq \mu (\mathbf{x}^{k})-\beta \lambda_{k} \delta _{l},
$$
hence, in view of (\ref{eq:5.1a}), we obtain
\begin{equation} \label{eq:5.3}
V_{l} \leq \rho^{2}L \Phi(\mathbf{z}^{l-1})/(2\beta(1-\beta)\delta^{2} _{l}).
\end{equation}
Under the above assumptions, for some $ \mathbf{ x}^{*} \in X^{*}$ it holds that
\begin{eqnarray*}
&& \mu (\mathbf{z}^{l}) - \mu(\mathbf{x}^{*}) = f (\mathbf{z}^{l}) - f(\mathbf{x}^{*})
+ h (\mathbf{z}^{l}) - h(\mathbf{x}^{*}) \\
&& \leq \langle \mathbf{g}(\mathbf{z}^{l}),\mathbf{z}^{l}-\mathbf{x}^{*}
\rangle + h(\mathbf{z}^{l})-h(\mathbf{x}^{*}) \\
&& \leq \max\limits_{\mathbf{ y}\in X} \left\{ \langle \mathbf{g}(\mathbf{z}^{l}),
\mathbf{z}^{l}-\mathbf{y} \rangle + h(\mathbf{z}^{l})-h(\mathbf{y}) \right\} \\
&& = \varphi (\mathbf{z}^{l}) \leq n \delta_{l}.
\end{eqnarray*}
Using this estimate in (\ref{eq:5.3}) gives
$$
V_{l} \leq \rho^{2}Ln \delta_{l-1}/(2\beta(1-\beta)\delta^{2} _{l}).
$$
From (\ref{eq:5.2}) it follows that
$$
V_{l} \leq \rho^{2}Ln \nu ^{-l}/(2\beta(1-\beta)\delta_{0}\nu)=(C_{1}/\nu)\nu ^{-l}.
$$
Besides, since $\varepsilon \leq \Phi(\mathbf{z}^{l}) \leq n \delta_{l}=n \delta_{0}\nu ^{l}$, we have
$$
\nu^{-l(\varepsilon )} \leq n \delta_{0}/\varepsilon.
$$
Combining both the inequalities in (\ref{eq:5.1}), we obtain
\begin{eqnarray*}
V (\varepsilon ) && \leq C_{1}\nu^{-1}\sum ^{l(\varepsilon )}
_{l=1} \nu^{-l} = C_{1} (\nu^{-l(\varepsilon )}-1)/(1-\nu) \\
&& \leq C_{1} ((n \delta_{0}/\varepsilon)-1)/(1-\nu).
\end{eqnarray*}
We have obtained the complexity estimate.
\begin{theorem} \label{thm:5.2}
Let a sequence $\{\mathbf{z}^{l}\}$ be generated by the method with Basic
cycle (LP). Suppose that the function $f$ is convex and its partial
gradients satisfy Lipschitz continuity conditions with constants
$L_{i}$ for each $i \in I$. Then the method has the complexity
estimate
$$
V (\varepsilon ) \leq C_{1} ((n \delta_{0}/\varepsilon)-1)/(1-\nu),
$$
where $C_{1}=\rho^{2}Ln/(2\beta(1-\beta)\delta_{0})$.
\end{theorem}
We observe that the order of the estimates is similar to that in the
usual conditional gradient methods under the same assumptions; see
e.g. \cite{LP66,PD78,Pol83}.
\section{Some examples of applications}\label{s6}
We intend now to give some examples of applied problems
which reduce to decomposable composite optimization problems
of form (\ref{eq:2.3}), where utilization of
the proposed adaptive PL method may give certain preferences.
\subsection{Selective classification problems}
One of the most popular approaches to data classification is
support vector machine techniques; see e.g. \cite{Bur98,Aga15}.
The simplest linear support vector machine problem for data classification
consists in creating an optimal hyperplane separating two convex hulls
of a collection of known
points $\mathbf{x}^{i} \in \mathbb{R}^{m}$, $i=1, \ldots,l$ attributed to previous data observations
with different labels $ y_{i} \in \{-1,+1\}$, $i=1, \ldots,l$,
where $m$ is the number of features. That is, the distance between the hyperplane and each
convex hull should be as long as possible. This separation of the
feature space enables us to classify new data points.
However, this requirement appears too strong for real problems
where the so-called soft margin approach, which minimizes the penalties for
mis-classification, is utilized. This problem can be formulated
as the optimization problem
$$
\min \limits _{\mathbf{w} \in \mathbb{R}^{n}} \to (1/p)\|\mathbf{w}\|^{p}_{p}
+ C \sum \limits^{l} _{i=1} L( \langle \mathbf{w}, \mathbf{x}^{i} \rangle; y_{i}),
$$
where $L$ is a loss function and $C > 0$ is a penalty parameter. The usual choice is $L(z; y) = \max\{0; 1-yz\}$
and $p$ is either 1 or 2. Taking $p=2$, we can rewrite this problem as
$$
\min \limits _{\mathbf{w}, \mathbf{\xi}} \to 0.5\|\mathbf{w}\|^{2}
+ C \sum \limits^{l} _{i=1} \xi_{i},
$$
subject to
$$
1-y_{i}\langle \mathbf{w}, \mathbf{x}^{i} \rangle \leq \xi_{i}, \ \xi_{i} \geq 0, \ i=1, \ldots,l.
$$
In this formulation, each observation $i$ is attributed to some data point $\mathbf{x}^{i}$, however,
it seems worthwhile to use sets here since any object may be often represented by some
set of features, this is also the case for noisy observations;
see \cite{SKP08}. So, let object $i$ be represented by a set $X_{i} \in \mathbb{R}^{m}$.
Suppose it is the convex hull of the points $\mathbf{x}^{ik}$, $k=1, \ldots,t$, which thus have the same
label $ y_{i} \in \{-1,+1\}$.
Then we write the soft margin classification problem as follows:
$$
\min \limits _{\mathbf{w}, \mathbf{\xi}} \to 0.5\|\mathbf{w}\|^{2}
+ C \sum \limits^{l} _{i=1} \xi_{i},
$$
subject to
\begin{eqnarray*}
1-y_{i}\langle \mathbf{w}, \mathbf{x}^{ik} \rangle \leq \xi_{i}, && \ k=1, \ldots,t; \ i=1, \ldots,l; \\
\xi_{i} \geq 0, && \ i=1, \ldots,l;
\end{eqnarray*}
which somewhat differs from those in \cite{SKP08}.
Here $\xi_{i}$ is the set slack variable
and we impose the penalty for the sum of set slacks. By using the
convex optimization theory, we can write its dual that
has the quadratic programming format:
$$
\max\limits _{\mathbf{\alpha}} \to \sum \limits^{l} _{i=1} \sum \limits^{t} _{k=1} \alpha_{ik} -
0.5 \left\| \sum \limits^{l}_{i=1}\sum \limits^{t} _{k=1} (\alpha_{ik}y_{i})\mathbf{x}^{ik}\right\|^{2}
$$
subject to
\begin{eqnarray*}
&& \sum \limits^{t} _{k=1} \alpha_{ik} \leq C, \ i=1, \ldots,l; \\
&& \alpha_{ik} \geq 0, \ k=1, \ldots,t; \ i=1, \ldots,l;
\end{eqnarray*}
The basic solution of the primal problem is given by the formula:
$$
\mathbf{w}= \sum \limits^{l}_{i=1}\sum \limits^{t} _{k=1} (\alpha_{ik}y_{i})\mathbf{x}^{ik}.
$$
At the same time, we observe that the dual problem falls into
format (\ref{eq:2.3}) and its feasible set is the corresponding
Cartesian product. Hence, our method can be suitable in the high-dimensional case,
where the number of sets is also very large.
\subsection{Network equilibrium problems}
Various network equilibrium problems represent one of the main tools for
evaluation of flows distribution in traffic and communication networks.
We now describe for instance the path flow formulation
of the network equilibrium problem with elastic demands; see e.g.
\cite{Mag84} and references therein.
The model is determined on an oriented graph, each of its arc being associated with some flow
and some cost (for instance, time of delay), which depends on the values of arc flows.
Usually, the number of nodes and arcs is very large for applied problems.
Let us be given a graph with a finite set of nodes $\mathcal{V}$ and a
set of oriented arcs $\mathcal{A}$ which join the nodes so that any
arc $a=(i, j)$ has the origin $i$ and the destination $j$. Next,
among all the pairs of nodes of the graph we extract a subset of
origin-destination (O/D) pairs $\mathcal{M}$ of the form $m=(i \to
j)$. Besides, each pair $m \in \mathcal{M}$ is associated with a
variable flow demand $v_{m}$ and with the set of paths
$\mathcal{P}_{m}$ which connect the origin and destination for this
pair. We suppose that each $v_{m}$ is non-negative with some upper bound $\gamma_{m} \leq +\infty$ for $m
\in \mathcal{M}$. Denote by $\tau_{m}$ the minimal path cost for the pair $m$
and suppose that it depends on the flow demand, i.e.
$\tau_{m}= \tau_{m}(v_{m})$. Also, denote by $u_{p}$ the path flow for the path $p$.
Then the feasible set of flows/demands $W$ can be
defined as follows:
$$
W = \prod_{m \in \mathcal{M}} W_{m},
$$
where
$$
W_{m}= \left\{ \mathbf{w}_{m}=(\mathbf{u}_{m},v_{m})
\ \vrule \ \begin{array}{c}
\sum _{p \in \mathcal{P}_{m}} u_{p} =v_{m}, u_{p} \geq 0, \
p \in \mathcal{P}_{m}, \\ 0 \leq v_{m} \leq \gamma_{m}, \
\end{array}
\right\}, \ \forall m \in \mathcal{M},
$$
where $\mathbf{u}_{m}=(u_{p})_{ p \in \mathcal{P}_{m}}$.
Given a flow vector $\mathbf{u} = (\mathbf{u}_{m})_{ m \in \mathcal{M}}$,
one can determine the arc flow
$$
f_{a}= \sum _{m \in \mathcal{M}} \sum _{p \in \mathcal{P}_{m}}
\alpha _{p a} u_{p}
$$
for each arc $a \in \mathcal{A}$, where
$$
\alpha _{p a} =\cases {
1 &\textrm{ if arc $a$ belongs to path $p$,} \cr
0 &\textrm{ otherwise.} \cr
}
$$
If the vector $ \mathbf{f} = (f_{a})_{a \in \mathcal{A}}$ of arc flows is
known, one can determine the arc cost $c_{a}(f_{a})$. We suppose for simplicity that
it depends on the arc flow of just this arc. Usually, arc costs are monotone
increasing functions of arc flows. Then one can compute costs for each path
$p$:
$$
g_{p}(\mathbf{u}) = \sum _{a \in \mathcal{A}} \alpha _{p a} c_{a}(f_{a}).
$$
We say that
a feasible flow / demand pair $ (\mathbf{u}^{*}, \mathbf{v}^{*} ) \in W$ is an {\em
equilibrium point} if it satisfies the following conditions:
\begin{equation} \label{eq:6.1}
\forall m \in \mathcal{M}, \ \exists \lambda_{m} \ \mbox{such that}
\ g_{p}(\mathbf{u}^{*}) \left\{
\begin{array}{ll}
\geq \lambda_{m} \quad & \mbox{if} \quad u^{*}_{p} = 0, \\
=\lambda_{m} \quad & \mbox{if} \quad u^{*}_{p} >0,
\end{array}
\right. \quad \forall p \in \mathcal{P}_{m};
\end{equation}
and
\begin{equation} \label{eq:6.2}
\tau_{m}(v_{m}^{*}) \left\{
\begin{array}{ll}
\leq \lambda_{m} \quad & \mbox{if} \quad v^{*}_{m} = 0, \\
=\lambda_{m} \quad & \mbox{if} \quad v^{*}_{m} \in (0,\gamma_{m}); \\
\geq \lambda_{m}
\quad & \mbox{if} \quad v^{*}_{m} =\gamma_{m};
\end{array}
\right. \quad \forall p \in \mathcal{P}_{m}.
\end{equation}
However, conditions (\ref{eq:6.1})--(\ref{eq:6.2}) determine
equivalently the following VI: Find a pair $
(\mathbf{u}^{*}, \mathbf{v}^{*}) \in W$ such that
\begin{equation} \label{eq:6.3}
\sum _{m \in \mathcal{M}} \sum _{p \in \mathcal{P}_{m}}
g_{p}(\mathbf{u}^{*}) (u_{p}- u^{*}_{p})-\sum _{m \in \mathcal{M}}
\tau_{m}(v_{m}^{*}) (v_{m}- v^{*}_{m}) \geq 0 \quad \forall (\mathbf{u}, \mathbf{v})
\in W.
\end{equation}
Furthermore, due to the separability of the functions $c_{a}$ and $\tau_m$, their continuity
implies integrability, i.e., then there exist functions
$$
\eta_a(f_a) = \int \limits_{0}^{f_a} c_a(t) dt \ \forall a \in \mathcal{A}, \
\sigma_m(v_m) = \int \limits_{0}^{v_m} \tau_m(t) dt \ \forall m \in \mathcal{M}.
$$
It follows that VI (\ref{eq:6.3}) also gives an optimality condition of the following optimization problem:
\begin{equation}\label{eq:6.4}
\min \limits_{(\mathbf{u}, \mathbf{v}) \in W} \rightarrow
\left\{ \sum\limits_{a \in \mathcal{A}} \eta_a(f_a) -
\sum\limits_{m \in \mathcal{M}} \sigma_m(v_m)\right\}.
\end{equation}
Hence, each solution of (\ref{eq:6.4}) is a solution to VI (\ref{eq:6.3}),
the inverse assertion is true if the functions $\eta_a$ and $-\sigma_m$ are convex,
this seems rather natural. However, this problem falls into the basic format
(\ref{eq:2.3}) and the suggested PL method can be applied to this problem.
The basic auxiliary problem consists in finding an element
$(\bar\mathbf{u}_{s}, \bar v_{s})=\mathbf{y}_{s}(\mathbf{x}^{k})\in Y_{s}(\mathbf{x}^{k})$
with $\mathbf{x}^{k}=(\mathbf{u}^{k}, \mathbf{v}^{k})$,
which is now corresponds to a solution of the problem
\begin{equation}\label{eq:6.5}
\min \limits_{(\mathbf{u}_{s}, v_{s}) \in W_{s}} \rightarrow
\left\{ \sum _{p \in \mathcal{P}_{s}}
g_{p}(\mathbf{u}^{k})u_{p} -
\sigma_s(v_{s})\right\}
\end{equation}
for some selected pair $s \in \mathcal{M}$. This solution can be found with the simple procedure below,
which is based on optimality conditions (\ref{eq:6.1})--(\ref{eq:6.2}).
First we calculate a shortest path $t \in \mathcal{P}_{s}$ for the pair $s$ with the minimal
cost $\tilde \lambda=g_t(\mathbf{u}^k)$.
Case 1. If $\tau_{s}(0) \le \tilde \lambda$, then set $\bar v_{s}=0$ and
$\bar u_{p}=0$ for all $p \in \mathcal{P}_{s}$, $\lambda_{s} = \tilde \lambda$.
Otherwise go to Case 2.
Case 2. If $\tau_{s}(\gamma_{s}) \ge \tilde \lambda$,
set $\lambda_{s} = \tilde \lambda$,
$\bar v_{s}=\gamma_{s}$, $\bar u_{t}=\gamma_{s}$, and
$\bar u_{p}=0$ for all $p \in \mathcal{P}_{s}$, $p \neq t$.
Otherwise go to Case 3.
Case 3. We have $\tau_{s}(\gamma_{s}) < \tilde \lambda < \tau_{s}(0)$. By continuity of
$\tau_{s}$, we find the value $\bar v_{s} \in [0, \gamma_{s}]$ such that
$\tau_{s}(\bar v_{s}) =\tilde \lambda$, set $\lambda_{s} = \tilde \lambda$,
$\bar u_{t}=\bar v_{s}$, and $\bar u_{p}=0$ for all $p \in \mathcal{P}_{s}$, $p \neq t$.
We supposed above that each function $\tau_m$ is continuous, i.e.
that each function $\sigma_m$ is smooth. However, the described procedure for
problem (\ref{eq:6.5})
is extended easily to the case where $-\sigma_m$ is convex
and continuous, then $\tau_m$ can be set-valued.
At the same time, we note that the network equilibrium problem with fixed demands
differs only in somewhat simplified formulation
of problem (\ref{eq:6.4}).
Clearly, the described method remains convergent in these cases and seems in general
simpler and more flexible in comparison with the usual conditional gradient
and projection type methods.
\subsection{Penalty method for decomposable optimization problems}
A great number of optimization problems related to large scale systems
are written as follows:
\begin{equation}\label{eq:6.6}
\max \ \to \ \sum \limits_{i=1}^{n} \langle \mathbf{c}_{i}, \mathbf{x}_{i} \rangle
\end{equation}
subject to
\begin{eqnarray}
&& \sum \limits_{i=1}^{n} A_{i}\mathbf{x}_{i} = \mathbf{b}_{0}, \label{eq:6.7} \\
&& \mathbf{x}_{i} \in X_{i} = \{ \mathbf{y} \in \mathbb{R}^{l_{i} }_{+} \ | \ B_{i}\mathbf{y} \leq \mathbf{b}_{i}
\}, \ i=1,\ldots,n; \label{eq:6.8}
\end{eqnarray}
for instance, it can be attributed to the total income maximization
in a system containing $n$ subsystems (producers),
who utilize common and particular factors. That is,
producer $i$ chooses an output vector $\mathbf{x}_{i}\in \mathbb{R}^{l_{i}}$,
his/her consumption rates are described by an $m_{0} \times l_{i}$
matrix $A_{i}$ of common factors and by an $m_{i} \times l_{i}$
matrix $B_{i}$ of particular factors, whereas the vector
$\mathbf{c}_{i}$ denotes prices of his/her outputs, the vector $\mathbf{b}_{i} \in
\mathbb{R}^{m_{i}}$ (respectively, $\mathbf{b}_{0}\in \mathbb{R}^{m_{0}}$) denotes inventories of
particular (respectively, common) factors; see e.g. \cite{Las70,Pol83}.
Due to its very large dimensionality, a suitable decomposition approach
can be utilized to reduce the computer memory and calculation expenses.
For instance, the price (Dantzig-Wolfe) decomposition principle
replaces problem (\ref{eq:6.6})--(\ref{eq:6.8}) with its dual defined with the help of
the Lagrangian including only the term associated with the common constraints in (\ref{eq:6.7}).
However, we can also utilize the penalty approach and replace problem (\ref{eq:6.6})--(\ref{eq:6.8}) with
the sequence of auxiliary problems of the form
\begin{equation}\label{eq:6.9}
\min \ \to \ 0.5 \tau \left\| \sum \limits_{i=1}^{n} A_{i}\mathbf{x}_{i} - \mathbf{b}_{0} \right\|^{2}
-\sum \limits_{i=1}^{n} \langle \mathbf{c}_{i}, \mathbf{x}_{i} \rangle
\end{equation}
subject to
\begin{equation}\label{eq:6.10}
\mathbf{x}_{i} \in X_{i}, \ i=1,\ldots,n;
\end{equation}
where $\tau>0$ is a penalty parameter. Clearly, problem (\ref{eq:6.9})--(\ref{eq:6.10}) also
falls into the basic format (\ref{eq:2.3}) and
application of the suggested PL (conditional gradient) method leads
to some other decomposition method for the initial problem (\ref{eq:6.6})--(\ref{eq:6.8}).
In fact, the partial gradient of the cost function at $\mathbf{x}$ is written as follows
$$
\mathbf{g}_{i}(\mathbf{x})=\tau A^{\top}_{i}\left[ \sum \limits_{j=1}^{n} A_{j}\mathbf{x}_{j} - \mathbf{b}_{0} \right]
-\mathbf{c}_{i},
$$
hence
$$
\mathbf{g}_{i}(\mathbf{x}^{k}+\theta \mathbf{d}^{k})=\mathbf{g}_{i}(\mathbf{x}^{k})+ \theta \tau A^{\top}_{i}A_{s}\mathbf{d}^{k}_{s},
$$
and we can make shifts only in the selected component $\mathbf{d}^{k}_{s}$
at each iteration. Besides, in order to find
$\mathbf{y}_{s}^{k}=\mathbf{y}_{s}(\mathbf{x}^{k})\in Y_{s}(\mathbf{x}^{k})$, we have to solve the separate problem
$$
\min \limits_{\mathbf{y}_{s} \in X_{s}} \rightarrow
\langle \mathbf{g}_{s}(\mathbf{x}^{k}), \mathbf{y}_{s} \rangle.
$$
Combining this method with proper regulation of $\tau$, we obtain a sequence
convergent to a solution of (\ref{eq:6.6})--(\ref{eq:6.8}).
\section{Computational experiments}\label{s7}
In order to compare the performance of the presented method
with the usual non-decomposable version we carried
out preliminary series of computational experiments. For simplicity, we took only
the smooth problems, i.e. set $h \equiv 0$.
Hence, we compared the usual conditional gradient method (CGM) from \cite{PD78} and our method which is treated as
its adaptive version (ACGM). We took the even partition of $\mathbb{R}^{N}$,
i.e., set $N_{i}=t=N/n$ for $i=1, \dots, n$. Next, each set $X_{i}$ was chosen to be the standard
simplex in $\mathbb{R}^{t}$, i.e.,
$$
X_{i}=\left\{ \mathbf{u}\in \mathbb{R}^{t}_{+} \ \vrule \ \sum \limits_{i=1}^{t}
u_{i}=1 \right\}.
$$
We took $\Delta_{k}=\varphi(\mathbf{x}^{k})$
as accuracy measure and chose the accuracy $0.1$.
We chose the same starting point $(1/t)e$,
where $\mathbf{e}$ denote the vector of units in $\mathbb{R}^{N}$,
and the rule $\delta_{l+1}=\nu \delta_{l}$ with $\nu = 0.5$ for (ACGM). The methods were
implemented in Delphi with double precision arithmetic.
We report the number of iterations (it) and the total number of calculations (cl) of the
partial gradients $\mathbf{g}_{i}$ for attaining the desired accuracy.
In the first series, we took the convex quadratic cost function.
We chose $\mu(\mathbf{x})=f_{1}(\mathbf{x})$ where
$$
f_{1} (\mathbf{x})= 0.5 \langle P\mathbf{x},\mathbf{x} \rangle -\langle \mathbf{q},\mathbf{x} \rangle,
$$
the elements of the matrix $P$ are defined by
$$
p_{ij}= \left\{ {
\begin{array}{rl}
\displaystyle
\sin (i) \cos (j) \quad & \mbox{if} \ i<j, \\
\sin (j) \cos (i) \quad & \mbox{if} \ i>j, \\
\sum \limits_{s\neq i} | p_{is}| +1 \quad & \mbox{if} \ i=j;
\end{array}
} \right.
$$
and elements of the vector $\mathbf{q}$ are defined by $q_{j}=\sin (j)/j$ for all $i,j$.
The results are given in Table \ref{tbl:1}.
\begin{table}
\caption{The numbers of iterations (it) and partial gradients calculations
(cl)} \label{tbl:1}
\begin{center}
\begin{tabular}{|rr|rr|rr|}
\hline
&{} & (CGM) &{} & (ACGM) &{} \\
\hline
$N$ &{} $n$ & it &{} cl & it &{} cl \\
\hline
10 & 5 & 15 & 75 & 9 & 28 \\
\hline
20 & 5 & 50 & 250 & 108 & 189 \\
\hline
50 & 5 & 143 & 715 & 452 & 676 \\
\hline
100 & 5 & 257 & 1285 & 775 & 1161 \\
\hline
50 & 10 & 228 & 2280 & 632 & 1048 \\
\hline
100 & 10 & $\Delta_{500}=0.11$ 500 & 5000 & $\Delta_{1500}=0.127$ 1500 & 2515 \\
\hline
80 & 20 & $\Delta_{500}=0.3$ 500 & 10000 & 766 & 1646 \\
\hline
100 & 20 & $\Delta_{500}=0.367$ 500 & 10000 & 1328 & 2820 \\
\hline
100 & 25 & $\Delta_{500}=0.4$ 500 & 12500 & 980 & 2346 \\
\hline
100 & 50 & $\Delta_{500}=0.76$ 500 & 25000 & 236 & 1036 \\
\hline
\end{tabular}
\end{center}
\end{table}
In the second series, we took the composite convex cost function
$$
\mu(\mathbf{x})=f_{1}(\mathbf{x})+f_{2}(\mathbf{x}),
$$
where $f_{1}$ was defined as above
and
$$
f_{2}(\mathbf{x}) =1/(\langle \mathbf{c},\mathbf{x}\rangle+\tau),
$$
where $c_{i}= 2+\sin(i)$ for $ i=1,\ldots,N$
and $\tau=5$. The results are
given in Table \ref{tbl:2}.
\begin{table}
\caption{The numbers of iterations (it) and partial gradients calculations
(cl)} \label{tbl:2}
\begin{center}
\begin{tabular}{|rr|rr|rr|}
\hline
&{} & (CGM) &{} & (ACGM) &{} \\
\hline
$N$ &{} $n$ & it &{} cl & it &{} cl \\
\hline
10 & 5 & 15 & 75 & 10 & 32 \\
\hline
20 & 5 & 49 & 245 & 113 & 189 \\
\hline
50 & 5 & 139 & 695 & 475 & 666 \\
\hline
100 & 5 & 240 & 1200 & 779 & 1161 \\
\hline
50 & 10 & 231 & 2310 & 620 & 1003 \\
\hline
100 & 10 & $\Delta_{500}=0.11$ 500 & 5000 & $\Delta_{1500}=0.125$ 1500 & 2515 \\
\hline
80 & 20 & $\Delta_{500}=0.3$ 500 & 10000 & 766 & 1674 \\
\hline
100 & 20 & $\Delta_{500}=0.36$ 500 & 10000 & 1329 & 2920 \\
\hline
100 & 25 & $\Delta_{500}=0.39$ 500 & 12500 & 1011 & 2350 \\
\hline
100 & 50 & $\Delta_{500}=0.775$ 500 & 25000 & 236 & 1040 \\
\hline
\end{tabular}
\end{center}
\end{table}
In almost all the cases, (ACGM) showed some preference
over (CGM) in the number of partial
gradients calculations. At the same time, tuning
parameters of (ACGM) needs further
investigations.
\section{Conclusions}
We described a new adaptive component-wise
method for decomposable composite optimization problems involving
non-smooth functions, where the feasible set is the
Cartesian product. The method consists in selective component-wise steps
together with a special control of tolerance sequences. We showed
that this keeps the convergence properties of the usual PL
one together with reduction of the computational
expenses. We describe several classes of significant applications
for the new method. The preliminary results of computational tests showed rather
satisfactory convergence.
\section*{Acknowledgement}
This work was supported by the RFBR grant, project No. 16-01-00109a
and by grant No. 297689 from Academy of Finland.
|
\section*{Introduction}
Let $\Omega$ be a bounded smooth strictly pseudoconvex domain of $\mathbb{C}^n$, i.e., there exists a smooth
strictly plurisubharmonic function $\rho$ defined on a bounded neighbourhood of $\bar{\Omega}$ such that
$$
\Omega=\{\rho<0\}.
$$
Let $A\geq 0,T>0$. We consider the equation
\begin{equation}\label{KRF}
\begin{cases}
\begin{array}{ll}
\dot{u}=\log\det (u_{\alpha\bar{\beta}})-Au+f(z,t)\;\;\;&\mbox{on}\;\Omega\times (0,T),\\
u=\varphi&\mbox{on}\;\partial\Omega\times [0,T),\\
u=u_0&\mbox{on}\;\bar{\Omega}\times\{ 0\},\\
\end{array}
\end{cases}
\end{equation}
where $\dot{u}=\frac{\partial u}{\partial t}$,
$u_{\alpha\bar{\beta}}=\frac{\partial^2 u}{\partial z_{\alpha}\partial\bar{z}_{\beta}}$,
$u_0$ is a plurisubharmonic function in a neighbourhood of $\bar{\Omega}$
and $\varphi, f$ are smooth in $\bar{\Omega}\times [0,T]$.
If $u_0$ is a smooth strictly
plurisubharmonic function in $\bar{\Omega}$ and some compatibility conditions are satisfied on
$\partial\Omega\times\{0\}$, then \eqref{KRF} admits a unique solution
$u\in C^{2;1}(\bar{\Omega}\times [0,T))$, and then $u$ is smooth outside $\partial\Omega\times \{0\}$
(\cite{HL10}, see also Section \ref{sec HL}). In more general cases, we define
the weak solution
\begin{Def}
The function $u\in USC(\bar{\Omega}\times [0,T))$ (upper semicontinuous function) is called a
\emph{weak solution} of
\eqref{KRF} if there exist $u_m\in C^{\infty}(\bar{\Omega}\times [0,T))$ satisfying
\begin{equation}\label{KRF_weak}
\begin{cases}
\begin{array}{ll}
u_m(.,t)\in SPSH(\Omega),\\
\dot{u}_m=\log\det (u_m)_{\alpha\bar{\beta}}-Au_m+f(z,t)\;\;\;&\mbox{on}\;\Omega\times (0,T),\\
u_m\searrow\varphi&\mbox{on}\;\partial\Omega\times [0,T),\\
u_m\searrow u_0&\mbox{on}\;\bar{\Omega}\times\{ 0\},\\
u=\lim\limits_{m\rightarrow \infty} u_m,
\end{array}
\end{cases}
\end{equation}
where $SPSH(\Omega)=\{\mbox{strictly plurisubharmonic functions on } \Omega\}$.
\end{Def}
The equation \eqref{KRF} always admits a unique weak solution, and the weak solution has been described
in case $u_0$ has zero Lelong numbers or $u\geq N\sum\log |z-a_j|+O(1)$ (\cite{Do15a,Do15b},
see also Section \ref{sec weak}). In this article, we will consider the general case where
$u_0$ has positive Lelong numbers.
The corresponding problem in compact K\"ahler manifolds was considered and solved by Di Nezza anh Lu
\cite{DL14}. The maximal solution (as weak solution in case of domains) is smooth
outside $D_t\times \{t\}$, where $D_t$ is an analytic set. In case of domains on $\mathbb{C}^n$, by condition
$u_0= \varphi (.,0)$ on $\partial\Omega$ (and $\varphi$ is smooth), $(u_0-\sup_{\Omega}u_0-1)$
can be approximated as in
Theorem \ref{Dem appr} by functions contained in Cegrell's class $\mathcal{E} (\Omega)$ and then for
any $\epsilon>0$, the set $\{z\in\Omega: \nu(u_0,z)\geq \epsilon\}$ contains a finite number of
points \cite{Ceg04}. Hence we can describe more precisely the singular set of the weak solution:
$\Omega\times\{t\}$ contains only finitely many singular points of the weak solution, for any $0<t<T$.
In the domain $\Omega\times [0,T)$,
the set of singular points of the weak solution is $\cup \{a_j\}\times (0,\epsilon_j)$,
where $\nu(u_0, a_j)>0$
and $\epsilon_j$ is a positive number
which is bounded by constants depending on $A$ and $\nu(u_0, a_j)$.
For the convenience, we denote by $\epsilon_A$ the functions
\begin{equation}\label{epsilonA.eq}
\begin{cases}
\epsilon_A(x)=\frac{x}{2n} \mbox{ if } A=0,\\
\epsilon_A(x)=\frac{1}{A}(\log (Ax+2n)-\log (2n)) \mbox{ if } A>0.
\end{cases}
\end{equation}
where $x>0$ and $0<t<T$.
Our main result is following
\begin{The}\label{main}
Let $A\geq 0, T>0$ and $\Omega$ be a bounded smooth strictly pseudoconvex domain of $\mathbb{C}^n$.
Let $\varphi, f$ be smooth functions in $\bar{\Omega}\times [0,T]$
and $u_0$ be a plurisubharmonic function in a neighbourhood of $\bar{\Omega}$
such that $u_0(z)=\varphi(z,0)$ for any $z\in\partial\Omega$.
Then the weak solution u of \eqref{KRF} satisfies
\begin{itemize}
\item[(i)] If $m>\frac{1}{T}$ and
$\{z\in\Omega: \nu_{u_0}(z)\geq\frac{2}{m}\}=\{a_{m,1},...,a_{m,N(m)}\}$ then there exist nonegetive numbers
$\epsilon_{m,1},...,\epsilon_{m,N(m)}\in (0,T]$ such that
\begin{itemize}
\item[(a)] $\epsilon_A(\nu (u_0,a_{m,j}))\leq \epsilon_{m,j}\leq \epsilon_A(n\nu (u_0,a_{m,j}))$ for
any $j=1,...,N(m)$.\\
\item[(b)]$\nu (u(.,t),a_{m,j})>0$ for any $j=1,...,N(m)$ and $t<\epsilon_{m,j}$. Moreover, if $\epsilon_{m,j}<T$ then
$\nu (u(.,t),a_{m,j})\searrow 0$ as $t\nearrow \epsilon_{m,j}$.\\
\item[(c)]$u\in C^{\infty}(Q_m)$, where $Q_m= \bar{\Omega}\times (\frac{1}{m},T)\setminus
\cup_{j\leq N(m)}\{a_{m,j}\}\times (\frac{1}{m},\epsilon_{m,j}] .$\\
\item[(d)] $\dot{u}=\log\det (u_{\alpha\bar{\beta}})-Au+f(z,t)\;\;\mbox{on}\;
Q_m . $\\
\end{itemize}
\item[(ii)] $u=\varphi$ on $\partial\Omega\times [0,T)$ and $\lim\limits_{t\to 0}u(z,t)=u_0(z)$ for any
$z\in\Omega$. In particular, $u(z,t)\stackrel{L^1}{\longrightarrow}u_0(z)$
as $t\to 0$.\\
\end{itemize}
\end{The}
Let us first recall some preliminaries.
\section{Coherent analytic sheaves on pseudoconvex domains}
In this section, we recall some properties of coherent analytic sheaves on pseudoconvex domains in
$\mathbb{C}^n$. The readers can find more details in \cite{Hor90} (chap VI and chap VII). Corollary \ref{Noether U}
will be used to prove Proposition \ref{Prop fini sum app}.
If $\Omega$ is a domain in $\mathbb{C}^n$, we let $\mathcal{O} (\Omega)$ denote the set of holomorphic functions on $\Omega$.
If $z\in\mathbb{C}^n$, we let $\mathcal{O}_z(\mathbb{C}^n)$ or $\mathcal{O}_z$ for short denote the set of equivalence classes of functions $f$ which are holomorphic in some neighbourhood of $z$, under the equivalence relation $f\sim g$ if $f=g$ in some
neighbourhood of $z$. If $f$ is holomorphic in a neighbourhood of $z$, we write $f_z$ for the residue class of $f$
in $\mathcal{O}_z$, which is called the germ of $f$ at $z$.
It follows from Theorem 6.3.3 and Theorem 6.3.5 in \cite{Hor90} that:
\begin{The}
Let $a\in\mathbb{C}^n$. Then $\mathcal{O}_a$ is a Noetherian ring. If $\mathcal{I}$ is a ideal of $\mathcal{O}_a$, and
$\mathcal{I}\ni f_j\stackrel{j\to\infty}{\longrightarrow} f\in\mathcal{O}_a$ in sense of simple convergence, then
$f\in\mathcal{I}$. Here $f_j\longrightarrow f$ means that the coefficient of $(z-a)^{\alpha}$ in $f_j$ converges
to the coefficient of $(z-a)^{\alpha}$ in $f$ for every $\alpha$.
\end{The}
By the same argument, we also have:
\begin{The}\label{the germ}
Let $\Omega$ be a neighbourhood of $0\in\mathbb{C}^n$. Then for any $\{f_j\}_{j=1}^{\infty}\subset\mathcal{O}_(\Omega)$, there exist
$M>0$ and $0<r<d(0,\partial\Omega)$ such that the ideal $\mathcal{I}$ of $\mathcal{O} (\Delta_r^n)$ which is generated by
$\{f_j\}_{j=1}^{\infty}$ is also generated by $\{f_j\}_{j=1}^M$. Moreover, $\mathcal{I}$ is closed in the topology
of uniform convergence on compact subsets of $\Delta_r^n$.
\end{The}
\begin{Def}
Let $X$ and $\mathcal{F}$ be two topological spaces and $\pi$ be a mapping $\mathcal{F}\to X$ such that
\begin{itemize}
\item[(i)] $\pi$ maps $\mathcal{F}$ onto $X$.
\item[(ii)] $\pi$ is a local homeomorphism, that is, every point in $\mathcal{F}$ has an open neighbourhood which
is mapped homeomorphically by $\pi$ on an open set in $X$.
\end{itemize}
Then $\mathcal{F}$ is called a sheaf on $X$ and $\pi$ is called the projection on $X$. If $U$ is a subset of $X$, a
section of $\mathcal{F}$ over $U$ is a continuous map $\varphi: U\to\mathcal{F}$ such that $\pi\varphi=Id$ on $U$.
The set of all sections of $\mathcal{F}$ over $U$ is denoted by $\Gamma (U,\mathcal{F})$. If $x\in X$,
then $\mathcal{F}_x=\pi^{-1}\{x\}$ is called the stalk of $\mathcal{F}$ at $x$.
\end{Def}
\begin{Ex} Let $\Omega$ be a open subset of $\mathbb{C}^n$.
The sheaf $\mathcal{O}_{\Omega}$ (or $\mathcal{O}$ for short) of germs of holomorphic functions (analytic functions)
on $\Omega\subset \mathbb{C}^n$ is the topological space defined by
\begin{itemize}
\item[(i)] $\mathcal{O}=\cup_{z\in\Omega}\mathcal{O}_z$.\\
\item[(ii)] The topology in $\mathcal{O}$ is the strongest topology such that for every open subset $U\subset\Omega$
and for every $f\in\mathcal{O} (U)$, the map
$U\ni z\mapsto f_z\in\mathcal{O}$
is continuous.
\end{itemize}
The projection on $\Omega$ is the map $\pi : \mathcal{O}\to \Omega$ defined by $\pi (\mathcal{O}_z)=\{z\}$.
The set of sections of $\mathcal{O}$ over $U$ is $\Gamma (U, \mathcal{O})=\mathcal{O} (U)$ and the stalk of $\mathcal{O}$ at $z$ is the set
of germs at $z$ of holomorphic functions in some neighbourhood of $z$.
\end{Ex}
\begin{Def}
Let $\Omega$ be an open subset of $\mathbb{C}^n$. A sheaf $\mathcal{F}$ on $\Omega$ is called an analytic sheaf if
it is a sheaf of $\mathcal{O}$-modules, i.e., $\mathcal{F}_z$ is an $\mathcal{O}_z$ module for every $z\in\Omega$
and the product of a section of $\mathcal{O}$ and a section of $\mathcal{F}$ is a section of $\mathcal{F}$.
\end{Def}
\begin{Def}
An analytic sheaf $\mathcal{F}$ on $\Omega$ is called coherent if
\begin{itemize}
\item[(i)]$\mathcal{F}$ is locally finitely generated, i.e., for every $z\in\Omega$ there exists a neighbourhood
$U\subset\Omega$ and a finite number of sections $f_1,...,f_q\in\Gamma (U,\mathcal{F})$ so that $\mathcal{F}_z$
is generated by $(f_1)_z,...,(f_q)_z$ as $\mathcal{O}_z$ module for every $z\in\Omega$.
\item[(ii)] If $U$ is an open subset of $\Omega$ and $f_1,..., f_q\in\Gamma (U,\mathcal{F})$, then the sheaf of
relations $\mathcal{R} (f_1,...,f_q)$ is locally finitely generated.
\end{itemize}
\end{Def}
\begin{The}
Every locally finitely generated subsheaf of $\mathcal{O}^p$ is coherent.
\end{The}
\begin{Cor}
Let $\{f_j\}_{j=1}^{\infty}\subset\mathcal{O}_(\Omega)$. Assume that $\mathcal{F}$ is the analytic sheaf generated by
$\{f_j\}_{j=1}^{\infty}$ over $\Omega$, i.e., the germs $(f_1)_z,...,(f_q)_z...$ generate $\mathcal{F}_z$
for every $z\in\Omega$. Then $\mathcal{F}$ is coherent.
\end{Cor}
\begin{The}
Let $\Omega$ be a Stein manifold, $K$ an $\Omega$-holomorphically convex compact subset of $\Omega$, and
$\mathcal{F}$ a coherent analytic sheaf on a neighbourhood of $K$. Then
\begin{itemize}
\item[(i)] There exist finitely many sections $f_1, ..., f_q$ of $\mathcal{F}$ over a neighbourhood of $K$
which generate $\mathcal{F}$ there.
\item[(ii)] If $f_1,..,f_q$ are sections of $\mathcal{F}$ over a neighbourhood of $K$ which generate
$\mathcal{F}$ there and if $f$ is an arbitraly section of $\mathcal{F}$ over a neighbourhood of $K$, then
one can find $c_1,...,c_q$ analytic in a neighbourhood of $K$ so that $f=\sum\limits_{j=1}^qc_jf_j$ there.
\end{itemize}
\end{The}
\begin{Cor}\label{Noether U}
Let $W$ be a pseudoconvex domains in $\mathbb{C}^n$ and $\Omega$ be a open subset of $W$, $\bar{\Omega}\subset W$.
Let $\{f_j\}_{j=1}^{\infty}\subset\mathcal{O}(W)$. Assume that $\mathcal{I}$ is the ideal of $\mathcal{O}(\Omega)$
generated by $f_1|_{\Omega},...,f_j|_{\Omega}...$ Then there exist finitely many functions
$f_1,...,f_q$ such that $\mathcal{I}$ is generated by $f_1|_{\Omega},...,f_q|_{\Omega}$.
\end{Cor}
\section{Demailly's approximation}
We recall Demailly's approximation theorem, which allows to approximate plurisubharmonic functions by multiples
of logarithms of holomorphic functions. The readers can find the proof of this theorem in \cite{Dem92} (or
\cite{Dem14}).
\begin{The}\label{Dem appr}
Let $\varphi$ be a plurisubharmonic function on a bounded pseudoconvex open set $\Omega\subset\mathbb{C}^n$. For every $m>0$, let $\mathcal{H}_{\Omega} (m\varphi)$ be the Hilbert space of holomorphic functions $f$ on $\Omega$ such that
$\int_{\Omega}|f|^2e^{-2m\varphi}dV_{2n}<+\infty$ and let
$\varphi_m=\frac{1}{2m}\log\sum |g_{m,l}|^2$ where $(g_{m,l})$ is an orthonormal basis of $\mathcal{H}_{\Omega}(m\varphi)$.
Then there are constants $C_1, C_2>0$ independent of $m$ such that\\
\begin{itemize}
\item[(a)] $\varphi(z)-\frac{C_1}{m}\leq \varphi_m (z)\leq \sup\limits_{|\zeta-z|<r}\varphi (\zeta)
+\frac{1}{m}\log\frac{C_2}{r^n}\;$ for every $z\in\Omega$ and $r<d(z,\partial\Omega)$. In particular, $\varphi_m$
converges to $\varphi$ pointwise and in $L^1_{loc}$ topology on $\Omega$ when $m\to\infty$.
\item[(b)] $\nu (\varphi, z)-\frac{n}{m} \leq \nu (\varphi_m, z)\leq \nu (\varphi, z)$ for every $z\in\Omega$.
\end{itemize}
\end{The}
Using Theorem \ref{Dem appr}, we will prove the following proposition, which will be use to prove the main theorem.
\begin{Prop}\label{Prop fini sum app}
Under the assumptions of Theorem \ref{Dem appr}, for any open subset $U\Subset\Omega$,
the following properties hold
\begin{itemize}
\item[(a)]$\int\limits_{U}e^{2m(\varphi_m-\varphi)}<\infty$.\\
\item[(b)] If $z\in U$ and $\nu (\varphi, z)< \frac{1}{m}$ then $\varphi_m$
is smooth in a neighbourhood of $z$.\\
\item[(c)] If there exist only finitely many points $a_1,...,a_l$ in
$\{z\in \bar{U}:\nu (\varphi, z) \geq \frac{1}{m} \}$,
then there exist $C,N>0$ such that
\begin{center}
$\varphi_m\geq N\sum\limits_{j=1}^l\log |z-a_j|-C$ on $U$.
\end{center}
\end{itemize}
\end{Prop}
\begin{proof}
(a) There is a corresponding result in the case of compact K\"ahler manifolds, which was proved in
\cite[p.164]{Dem10}, \cite[p.10]{Dem14}. The same arguments can be applied for the case of domains in $\mathbb{C}^n$.
Set $f_j:\Omega\times\Omega\to \mathbb{C}$, $f_j(z,w)=g_{m,j}(z)\overline{g_{m,j}(\bar{w})}$.
We have, for any $z,w\in V$ and $l>0$,
\begin{center}
$\sum\limits_{j\leq l}|f_j(z,w)|\leq (\sum\limits_{j\leq l}|g_{m,j}(z)|^2)^{1/2}
(\sum\limits_{j\leq l}|g_{m,j}(\bar{w})|^2)^{1/2}\leq e^{m(\varphi_m(z)+\varphi(\bar{w}))}.$
\end{center}
Then the series $\sum\limits_{j=1}^{\infty}f_j(z,w)$ converges uniformly on compact subsets of $\Omega\times\Omega$.
Let $U\Subset V\Subset \Omega$.
We denote by $\mathcal{J}$ the ideal of $\mathcal{O} (V\times V)$ generated by
$\{f_j\}_{j=1}^{\infty}$. It follows from Theorem \ref{the germ}
that $f=\sum\limits_{j=1}^{\infty}f_j\in\mathcal{J}$ .
Moreover, by Corollary
\ref{Noether U} we can choose $M\gg 1$ such that $\mathcal{J}$ is generated by $f_1,...,f_M$. Hence,
\begin{center}
$f=\sum\limits_{j=1}^Ma_jf_j,$
\end{center}
where $a_j\in\mathcal{O}(V\times V)$. Then there exists $C>0$ such that
\begin{center}
$e^{2m\varphi_m}=f(z,\bar{z})\leq\sum\limits_{j=1}^M|a_j(z,\bar{z})||g_{m,j}(z)|^2\leq
C\sum\limits_{j=1}^M|g_{m,j}(z)|^2,$
\end{center}
for any $z\in U$. Thus
\begin{center}
$\int\limits_Ue^{2m(\varphi_m-\varphi)}\leq CM<\infty.$
\end{center}
(b) If $\in U$ and $\nu (\varphi, z)< \frac{1}{m}$, then $2m\varphi$ is intergrable in a neighborhood of
$z$ (see \cite{Sko72}). Hence,
there exists $g\in\mathcal{H}_{\Omega}(m\varphi)$ such that $g(z)\neq 0$. Hence $\varphi_m(z)\neq -\infty$.
Thus $\varphi_m$ is smooth in a neighbourhood of $z$.\\
(c) This is a corollary of Lemma \ref{lemgz}.
\end{proof}
\begin{Lem}\label{lemgz}
Let $g_1,...,g_M\in\mathcal{O} (\Delta^n)$ such that
$$\{g_1=...=g_M=0\}=\{0\}.$$
Then there exist $C,N>0$ such that, on $\Delta^n_{1/2}$,
$$|g_1|^2+...+|g_M|^2\geq C(|z_1|^2+...+|z_n|^2)^N.$$
\end{Lem}
\begin{proof}
It follows from Hilbert's Nullstellensatz theorem (see \cite[p.19]{Huy05}) that there exist $N>0$, $1>r>0$
and holomorphic functions $f_{jk}\in\mathcal{O} (\Delta^n_r)$ satisfying, on $\Delta^n_r$,
\begin{center}
$z_k^N=\sum\limits_{j=1}^m g_j.f_{ij}$,
\end{center}
for $k=1,...,n$. Then there exists $C_1>0$ such that, on $\Delta^n_{r/2}$,
$$|g_1|^2+...+|g_M|^2\geq C_1(|z_1|^2+...+|z_n|^2)^N.$$
In the other hand,
$\inf\limits_{\Delta^n_{1/2}\setminus\Delta^n_{r/2}}(|g_1|^2+...+|g_M|^2)>0$. Then there exists $C_2>0$ such
that, on $\Delta^n_{1/2}\setminus\Delta^n_{r/2}$,
$$|g_1|^2+...+|g_M|^2\geq C_2(|z_1|^2+...+|z_n|^2)^N.$$
Denote $C=\max\{C_1,C_2\}$, we have, on $\Delta^n_{1/2}$,
$$|g_1|^2+...+|g_M|^2\geq C(|z_1|^2+...+|z_n|^2)^N.$$
\end{proof}
Lemma \ref{lemgz} is also an immediate corollary of Lojasiewicz inequality:
\begin{The}
Let $f: U\rightarrow \mathbb{R}$ be a real-analytic function on an open set $U$ in $\mathbb{R}^n$,
and let $Z$ be the zero locus of $f$. Assume that $Z$ is not empty.
Then, for any compact set $K$ in $U$, there are positive constants $C$ and $\alpha$ such that,
for every $x\in K$
\begin{center}
$dist (x,Z)^{\alpha}\leq C|f(x)|.$
\end{center}
\end{The}
We refer the reader to \cite{Loj59, Mal66, JKS92} for more details.
Now we recall some previous results on Parabolic complex Monge-Amp\`ere equation.
\section{Parabolic complex Monge-Amp\`ere equation}
\subsection{Hou-Li theorem}\label{sec HL}
\begin{flushleft}
The Hou-Li theorem \cite{HL10} states
that equation \eqref{KRF} has a unique smooth solution when the conditions are good
enough. We state the precise problem to be studied:
\end{flushleft}
\begin{equation}
\label{HLKRF}
\begin{cases}
\begin{array}{ll}
\dot{u}=\log\det (u_{\alpha\bar{\beta}})+f(t,z,u)\;\;\;&\mbox{on}\;\Omega\times (0,T),\\
u=\varphi&\mbox{on}\;\partial\Omega\times [0,T),\\
u=u_0&\mbox{on}\;\bar{\Omega}\times\{ 0\}.\\
\end{array}
\end{cases}
\end{equation}
We first need the notion of a subsolution to \eqref{HLKRF}.
\begin{Def}
\label{HLsubsol}
A function $\underline{u}\in C^{\infty}(\bar{\Omega}\times [0,T))$
is called a \emph{subsolution} of the equation \eqref{HLKRF} if and only if
\begin{equation}
\label{subsolu.houli}
\begin{cases}
\underline{u}(.,t) \mbox{is a strictly plurisubharmonic function,}\\
\dot{\underline{u}}\leq \log\det (\underline{u})_{\alpha\bar{\beta}}+f(t,z,\underline{u}),\\
\underline{u}|_{\partial\Omega\times (0,T)}=\varphi|_{\partial\Omega\times (0,T)},\\
\underline{u}(.,0)\leq u_0 .
\end{cases}
\end{equation}
\end{Def}
\begin{The}\label{houli}
Let $\Omega\subset\mathbb{C}^n$ be a bounded domain with smooth boundary. Let $T\in (0,\infty]$. Assume
that
\begin{itemize}
\item $\varphi$ is a smooth function in $\bar{\Omega}\times [0,T)$.
\item $f$ is a smooth function in $[0,T)\times\bar{\Omega}\times\mathbb{R}$ non increasing in the
lastest variable.
\item $u_0$ is a smooth strictly plurisubharmonic funtion in a neighbourhood of $\Omega$.
\item $u_0(z)=\varphi (z,0),\;\forall z\in\partial\Omega$.
\item The compatibility condition is satisfied, i.e.
$$
\dot{\varphi}=\log\det (u_0)_{\alpha\bar{\beta}}+f(t,z,u_0),
\;\;\forall (z,t)\in \partial\Omega\times \{ 0 \}.
$$
\item There exists a subsolution to the equation \eqref{HLKRF}.
\end{itemize}
Then there exists a unique solution
$u\in C^{\infty}(\Omega\times (0,T))\cap C^{2;1}(\bar{\Omega}\times [0,T))$ of the equation
\eqref{HLKRF}.
\end{The}
If $\Omega$ is a bounded smooth strictly pseudoconvex domain of $\mathbb{C}^n$ then
a subsolution always exists on $\bar{\Omega}\times [0,T')$
(for example, $\underline{u}= M\rho+\varphi$, where $M\gg 1$ and
$\rho$ is a smooth strictly plurisubhamonic function such that $\rho|_{\partial\Omega}=0$ ), for any
$0<T'<T$, and so Theorem \ref{houli} does not need the additional assumption of existence of a subsolution.
Using regularity theories (see, for example, \cite{GT83, CC95, Lieb96, Do15a}),
we can conclude that the solution $u$
of \eqref{HLKRF} is smooth outside $\partial\Omega\times\{0\}$
if the assumption of Theorem \ref{houli} holds. If $\Omega$ is a bounded smooth strictly pseudoconvex domain of $\mathbb{C}^n$,
$\varphi$ is smooth in $\bar{\Omega}\times [0,T]$ and
$f$ is smooth in $[0,T]\times\bar{\Omega}\times\mathbb{R}$, then $u$ is the solution of \eqref{HLKRF}
in $\bar{\Omega}\times [0, T+\delta)$, where $0<\delta\ll 1$. Hence $u$ can be approximated in
$\bar{\Omega}\times [0,T]$ by the smooth functions $u(z,t+\frac{1}{m})$. Thus, for approximations, $u$
is as good as smooth in $\bar{\Omega}\times [0,T]$.
\subsection{Maximum principle}
\begin{flushleft}
The following maximum principle is a basic tool to
establish upper and lower bounds in the sequel (see \cite{BG13} and \cite{IS13} for the proof).
\end{flushleft}
\begin{The}\label{max prin}
Let $\Omega$ be a bounded domain of $C^n$ and $T>0$.
Let $\{\omega_t\}_{0<t<T}$ be a continuous family of continuous
positive definite Hermitian forms on $\Omega$. Denote by $\Delta_t$ the Laplacian with
respect to $\omega_t$:
$$
\Delta_t f=\dfrac{n\omega_t^{n-1}\wedge dd^cf}{\omega_t^n},\;\forall f\in C^{\infty}(\Omega).
$$
Suppose that
$H \in C^{\infty}(\Omega \times (0,T)) \cap C(\bar{\Omega}\times [0,T))$ and satisfies\\
\begin{center}
$(\frac{\partial}{\partial t}-\Delta_t)H \leq 0 \:$
or
$\: \dot{H}_t \leq \log\dfrac{(\omega_t+dd^c H_t)^n}{\omega_t^n}$.
\end{center}
Then $\sup\limits_{\bar{\Omega}\times [0,T)} H = \sup\limits_{\partial_P(\Omega \times
[0,T))}H$. Here we denote $\partial_P(\Omega\times (0,T))=\left(\partial\Omega\times (0,T) \right)
\cup \left( \bar{\Omega}\times\{ 0\}\right)$.
\end{The}
\begin{Cor}(Comparison principle)\label{compa log}
Let $\Omega$ be a bounded domain of $\mathbb{C}^n$ and $A\geq 0, T>0$. Let
$u,v\in C^{\infty}(\Omega\times (0,T))\cap C(\bar{\Omega}\times [0,T))$
satisfy:
\begin{itemize}
\item $u(.,t)$ and $v(.,t)$ are strictly plurisubharmonic functions for any $t\in [0,T)$,
\item $\dot{u}\leq \log\det (u_{\alpha\bar{\beta}})-Au +f(z,t),$
\item $\dot{v}\geq \log\det (v_{\alpha\bar{\beta}})-Av+f(z,t),$
\end{itemize}
where $f\in C^{\infty}(\bar{\Omega}\times [0,T))$.\\
Then $\sup\limits_{\Omega\times (0,T)}(u-v)\leq max\{ 0,
\sup\limits_{\partial_P(\Omega\times (0,T))}(u-v)\}$.
\end{Cor}
\begin{Cor}\label{compa lap}
Let $\Omega$ be a bounded domain of $\mathbb{C}^n$ and $T>0$. We denote
by $L$ the operator on $C^{\infty}(\Omega\times (0,T))$ given by
$$
L(f)=\dfrac{\partial f}{\partial t}-\sum a_{\alpha\bar{\beta}}
\dfrac{\partial^2f}{\partial z_{\alpha}\partial \bar{z}_{\beta}}-b.f,
$$
where $a_{\alpha\bar{\beta}}, b\in C(\Omega\times (0,T))$, $(a_{\alpha\bar{\beta}}(z,t))$
are positive definite Hermitian matrices and $b(z,t)<0$.\\
Assume that $\phi\in C^{\infty}(\Omega\times (0,T))\cap C(\bar{\Omega}\times [0,T))$
satisfies\\
$$L(\phi)\leq 0.$$
Then $\phi\leq max(0,\sup\limits_{\partial_P(\Omega\times (0,T))}\phi).$
\end{Cor}
\subsection{Weak solution of Parabolic Monge-Amp\`ere equation}\label{sec weak}
\begin{flushleft}
We recall some properties of weak solution of \eqref{KRF}, which were proved in \cite{Do15a} and \cite{Do15b}.
\end{flushleft}
\begin{Prop}\label{first results}
Let $A\geq 0,T>0$ and let $\Omega$ be a bounded smooth strictly pseudoconvex domain of $\mathbb{C}^n$. Assume that $u_0$
is a plurisubharmonic function in a neighbourhood of $\bar{\Omega}$
and $\varphi, f$ are smooth in $\bar{\Omega}\times [0,T]$.
Then there exists a unique weak solution $u$ of \eqref{KRF}. Moreover,
\begin{itemize}
\item[(i)] $u|_{\partial\Omega\times [0,T)}=\varphi$ ;
$u(z,t)\stackrel{L^1}{\longrightarrow}u_0$ as $t\searrow 0$.
\item[(ii)] If Lelong numbers $\nu (u,a)=0$ for every $a\in\Omega$
then $u\in C^{\infty}(\bar{\Omega}\times (0,T))$ and $u$ satisfies \eqref{KRF} in $\bar{\Omega}\times (0,T)$
in the classical sense.
\item[(ii)] If there exist $l\in\mathbb{N}, a_j\in\Omega, N_j\geq n_j\geq 0$ such that
$$\sum\limits_{j=1}^l n_j\log |z-a_j|+C_0\geq u_0\geq \sum\limits_{j=1}^l N_j\log |z-a_j|-C_0,$$
then $u$ satisfies
\begin{itemize}
\item[(a)] $u\in C^{\infty}(Q)$, where $Q=(\bar{\Omega}\times (0,T))\setminus
(\cup(\{a_j\}\times (0,\epsilon_A(N_j)])$.
\item[(b)] $u=-\infty$ on $\cup(\{a_j\}\times [0,\min\{T,\epsilon_A(n_j)\}))$.
\item[(c)] $\dot{u}=\log\det u_{\alpha\bar{\beta}}-Au+f(z,t)$ in $Q$.
\end{itemize}
\end{itemize}
\end{Prop}
\begin{Prop} \label{weaker condition.lem.sec weak}
Assume that there exists $u_m\in C^{\infty}(\bar{\Omega}\times [0,T))$ satisfying
\begin{equation}\label{KRF_weak.sec weak}
\begin{cases}
\begin{array}{ll}
u_m(.,t)\in SPSH(\Omega)&\forall t\in [0,T),\\
u_m(z,t)+2^{-m}\geq u_{m+1}(z,t)&\forall (z,t)\in \bar{\Omega}\times [0,T),\\
\dot{u}_m=\log\det (u_m)_{\alpha\bar{\beta}}-Au_m+f(z,t)\;\;&\forall (z,t)\in\Omega\times (0,T),\\
u_m(z,t)\longrightarrow\varphi(z,t)&\forall (z,t)\in\partial\Omega\times [0,T),\\
u_m(z,0)\longrightarrow u_0(z)&\forall z\in\bar{\Omega}.
\end{array}
\end{cases}
\end{equation}
Then $u=\lim u_m$ is a weak solution of \eqref{KRF}.
\end{Prop}
\begin{Prop}\label{singular.prop.weak sec}
If there is $a\in\Omega$ such that $\nu_{u_0}(a)>0$, then the weak solution $u$ of \eqref{KRF} satisfies $u(a,t)=-\infty$ for $t\in [0,\epsilon_A(\nu_{u_0}(a)))$.
\end{Prop}
\section{Proof of Theorem \ref{main}}
\subsection{Some technique lemmas}
We present some lemmas, which will be used to prove the main theorem. The following two lemmas were proved in
\cite{Do15b}.
\begin{Lem}\label{utminusu0.lem.weak sec}
Suppose that $\psi,g\in C^{\infty}(\bar{\Omega}\times [0,T])$.
Assume that $v\in C^{\infty}(\bar{\Omega}\times [0,T))$ satisfies
\begin{equation}
\begin{cases}
\begin{array}{ll}
v(.,t)\in SPSH(\bar{\Omega}),\\
\dot{v}=\log\det (v_{\alpha\bar{\beta}})-Av+g(z,t)\;\;\;&\mbox{on}\;\Omega\times (0,T),\\
v=\psi&\mbox{on}\;\partial\Omega\times [0,T).\\
\end{array}
\end{cases}
\end{equation}
Then $$v(z,t)-v(z,0)\geq -C(t),$$ for any $(z,t)\in \bar{\Omega}\times [0,T)$. Here $C(t)$ is defined by
$$C(t)=\inf\limits_{1>\epsilon>0}((-n\log\epsilon+A\sup |\psi|+
\sup |g|)t-\epsilon\inf\rho)+
\sup\limits_{t'\in [0,t]}\sup\limits_{z\in\partial\Omega}|\psi (z,t')-\psi (z,0)|,$$
where $\rho\in C^{\infty}(\bar{\Omega})$ such that $dd^c\rho\geq dd^c|z|^2$ and $\rho|_{\partial\Omega}=0$.
\end{Lem}
\begin{Lem}\label{dotu.lem}
Assume that $u\in C^{\infty}(\bar{\Omega}\times [0,T))$ satisfies
\begin{equation}
\begin{cases}
\dot{u}=\log\det (u_{\alpha\bar{\beta}})+f(z,t)\;\;\mbox{on}\;\; \Omega\times (0,T),\\
u=\varphi\;\;\;\mbox{on}\;\; \partial\Omega\times [0,T).
\end{cases}
\end{equation}
Then
$$\dfrac{u(z,t)-\sup u_0}{t}-B\leq \dot{u}(z,t)\leq \dfrac{u(z,t)-u_0(z)}{t}+B, \;\;\forall
(z,t)\in\bar{\Omega}\times [0,T),$$
where $B= 2\sup |\dot{\varphi}|+T\sup |\dot{f}|+n$ and $u_0=u(., 0)$.
\end{Lem}
We will also need the following elementary observation
\begin{Lem}\label{de Giorgi}
Let $g:\mathbb{R}^+\rightarrow \mathbb{R}^+$ be a decreasing right-continuous function. Assume that there exist $\alpha, B>1$
such that $g$ satisfies
\begin{center}
$tg(t+s)\leq B (g(s))^{\alpha}\;\;\forall t,s>0$.
\end{center}
Then $g(s)>0$ for all $s\geq s_{\infty}$, where
$$s_{\infty}=\dfrac{2Bg(0)^{\alpha-1}}{1-2^{-\alpha+1}}.$$
\end{Lem}
We refer the reader to \cite{EGZ09} for the proof of Lemma \ref{de Giorgi}. Using this lemma, we prove Lemma
\ref{lemestiu}.
\begin{Lem}\label{lemestiu}
Let $\Omega$ be a bounded open subset of $\mathbb{C}^n$. Let $u, u_0, \psi$ be plurisubharmonic functions in $\Omega$
such that $u\in C^{\infty}(\bar{\Omega})$ and $\psi$ is bounded near $\partial\Omega$. Assume also that
\begin{center}
$(dd^cu)^n\leq Be^{a_1(u-a_2u_0)}dV$,\\
$\int_{\Omega}e^{b(\psi-u_0)}dV\leq B,$
\end{center}
where $a_1,a_2,b,B>0, b>a_1a_2$. Then there exists $C>0$ depending
only on $a_1,a_2,b,B,n, \Omega$ and $\liminf\limits_{z\to\partial\Omega} (u-a_2\psi)$ such that
\begin{center}
$u\geq a_2\psi -C$ on $\Omega.$
\end{center}
\end{Lem}
There is a corresponding result in the case of compact K\"ahler manifolds, which was proved in
\cite{DL14}. The same arguments can be applied for the case of domains in $\mathbb{C}^n$.
For the reader's convenience, we recall the arguments here.
\begin{proof}
Without loss of generality, we can assume that $\liminf\limits_{z\to\partial\Omega} (u-a_2\psi)=0$.
For $t\in\mathbb{R}$, we denote
$$U_t=\{z\in\Omega: u(z)< a_2\psi (z)-t\}.$$
Let $M>\sup |u|/a_2$ and $\psi_M=\max\{\psi, -M\}$.
Then, for any $t,s>0$ and $v\in PSH(\Omega)$ such that $-1\leq v\leq 0$, we have
\begin{flushleft}
$\begin{array}{ll}
t^n\int\limits_{U_{t+s}}(dd^cv)^n
=\int\limits_{\{u<a_2\psi_M -t-s\}}(tdd^cv)^n
&\leq \int\limits_{\{u< a_2\psi_M +tv-s\}}(dd^c(a_2\psi_M+tv-s))^n\\[14pt]
& \leq \int\limits_{\{u<a_2\psi_M+tv-s\}}(dd^cu)^n\\[14pt]
&\leq \int \limits_{U_s}Be^{a_1(u-a_2u_0)}dV\\[14pt]
&\leq \int \limits_{U_s}Be^{-s}e^{a_1a_2(\psi- u_0)}dV\\[14pt]
&\leq B\int\limits_{U_s}e^{a_1a_2(\psi- u_0)}dV\\[14pt]
&\leq B\left(\int\limits_{U_s}e^{b(\psi-u_0)}\right)^{\frac{a_1a_2}{b}}
\left(\int\limits_{U_s}dV\right)^{1-\frac{a_1a_2}{b}}\\[14pt]
&\leq B^{1+\frac{a_1a_2}{b}}\lambda (U_s)^{1-\frac{a_1a_2}{b}},
\end{array}$
\end{flushleft}
where $\lambda$ is Lebesgue measure. Hence, we have
\begin{equation}\label{eq estiu capvol}
t^nCap(U_{t+s}, \Omega)\leq B^{1+\frac{a_1a_2}{b}}\lambda (U_s)^{1-\frac{a_1a_2}{b}}.
\end{equation}
Moreover, it follows from \cite{AT84,Ze01} that for any $p>0$, there exists $C_p>0$ depending only on $p,n$ and
$\Omega$ such that
\begin{equation}\label{eq estiu volcap}
\lambda(U_t)\leq C_p Cap (U_t,\Omega)^p.
\end{equation}
Combining \eqref{eq estiu capvol} and \eqref{eq estiu volcap}, we obtain
\begin{equation}
t^n\lambda(U_{t+s})^{1/p}\leq B^{1+\frac{a_1a_2}{b}}C_p^{1/p}\lambda (U_s)^{1-\frac{a_1a_2}{b}}.
\end{equation}
Let $p=\frac{2b}{b-a_1a_2}$. Applying Lemma \ref{de Giorgi} for $g(t)=\lambda(U_{t})^{1/(pn)}$ and
$\alpha=\frac{p(b-a_1a_2)}{b}$, there exists $s_{\infty}>0$ depending only on
$a_1,a_2,b,B,n, \Omega$ such that $\lambda (U_{s_{\infty}})=0$. By the plurisubharmonicity of
$u$ and $\psi$, we conclude that $u\geq a_2\psi-s_{\infty}$.
\end{proof}
\subsection{Proof of Theorem \ref{main}} We consider the case $A=0$ and then we use it to prove the result in the case $A>0$.
\begin{flushleft}
\it{Step 1: Construct an approximation.}
\end{flushleft}
Let us first contruct a sequence of solutions as in Proposition \ref{weaker condition.lem.sec weak} such that
\begin{itemize}
\item[(i)]$\varphi_k=\varphi$ on $\partial\Omega\times (\epsilon, T)$
for any $\epsilon>0$ and $k\gg 1$;
\item[(ii)]
$\sup\limits_{k\geq 0}\sup\limits_{z\in\partial\Omega}|\varphi_k (z,t)-\varphi_k (z,0)|\rightarrow 0$
as $t\searrow 0$.
\end{itemize}
Using the convolution of $u_0+\frac{|z|^2}{k}$ with mollifiers,
we can take $u_{0,k}\in C^{\infty}
(\bar{\Omega})\cap SPSH(\bar{\Omega})$ such that
\begin{equation}\label{u0m.eq.proof exist.sec weak}
u_{0,k}\searrow u_0.
\end{equation}
Note that $u_0|_{\partial\Omega}$ is continuous. Then
\begin{equation}\label{deltam.eq.proof exist.sec weak}
\delta_k=\sup\limits_{z\in\partial\Omega}(u_{0,k}(z)-u_0(z))
\stackrel{k\rightarrow\infty}{\longrightarrow} 0.
\end{equation}
We define $g_k\in C^{\infty}(\bar{\Omega})$ and
$\varphi_k\in C^{\infty}(\bar{\Omega}\times [0,T))$ by\\
$$g_k=\log\det (u_{0,k})_{\alpha\bar{\beta}}+f(z,0),$$
$$\varphi_k=\zeta(\frac{t}{\epsilon_k}) (tg_k+u_{0,k})+(1-\zeta(\frac{t}
{\epsilon_k}))\varphi,$$
where $\zeta$ is a smooth function on $\mathbb{R}$ such that $\zeta$ is decreasing,
$\zeta|_{(-\infty,1]}=1$ and $\zeta|_{[2,\infty)}=0$. $\epsilon_k>0$ are
chosen such that the sequences $\{\epsilon_k\}$,
$\{\epsilon_k\sup|g_k|\}$ are decreasing to $0$.
$u_{0,k}$ and $\varphi_k$ satisfy the compatibility condition.
By Theorem \ref{houli}, there exists $u_k\in C^{\infty}(\bar{\Omega}\times [0,T))$ satisfying
\begin{equation}\label{KRF_m.sec weak}
\begin{cases}
\begin{array}{ll}
\dot{u}_k=\log\det (u_k)_{\alpha\bar{\beta}}+f(z,t)\;\;\;&\mbox{on}\;\Omega\times (0,T),\\
u_k=\varphi_k&\mbox{on}\;\partial\Omega\times [0,T),\\
u_k=u_{0,k}&\mbox{on}\;\bar{\Omega}\times\{ 0\}.\\
\end{array}
\end{cases}
\end{equation}
It is easy to verify that $u_k$ satisfies the conditions in Proposition \ref{weaker condition.lem.sec weak}. Then
$u(z,t)=\lim\limits_{k\to\infty} u_k(z,t)$ is the weak solution of \eqref{KRF}.
Moreover, $\varphi_k=\varphi$ on $\partial\Omega\times (\epsilon, T)$ for any $\epsilon>0$ and $k\gg 1$.
\begin{flushleft}
\it {Step 2: Smoothness of weak solution in
$(\bar{\Omega}\setminus\{\nu (u_0,z)\geq \frac{2}{m}\})\times (\frac{1}{m},T)$.}
\end{flushleft}
Let $m>\frac{1}{T}$ and $\epsilon <\frac{1}{m}$.
Applying Lemma \ref{utminusu0.lem.weak sec} and Lemma \ref{dotu.lem} for $u_k (z,t+\epsilon)$, $k\gg 1$,
we have,
for any $(z,t)\in\bar{\Omega}\times (\epsilon, T)$,
\begin{flushleft}
$\begin{array}{ll}
(dd^cu_k)^n =e^{\dot{u}_k-f(z,t)}dV
&\leq C_1e^{\frac{u_k-u_k(z,\epsilon)}{t-\epsilon}}dV\\
&\leq C_2e^{\frac{u_k-u_{0,k}}{t-\epsilon}}dV\\
&\leq C_2e^{\frac{u_k-u_0}{t-\epsilon}}dV,
\end{array}$
\end{flushleft}
where $C_1, C_2>0$ are independent of $k$.
Assume that $W$ is an open neighbourhood of $\bar{\Omega}$ such that $W$ is bounded pseudoconvex and $u_0\in PSH(W)$.
Let $l=\frac{1+\epsilon}{2(1/m-\epsilon)}$ and let $\{g_{l,j}\}$ be an orthonormal basis of
$\mathcal{H}_W (l u_0)$.
Denote $v_l=\frac{1}{2l}\log\sum\limits_{j=1}^{\infty}|g_{l,j}|^2$. Applying Lemma \ref{lemestiu} for $a_1=\frac{1}{1/m-\epsilon}$,
$a_2=1$ and $b=2l$, we have
\begin{equation}
u_k(z,\frac{1}{m})\geq v_l(z)-C_3,
\end{equation}
where $C_3>0$ is independent of $k$. Then
\begin{equation}
u(z,\frac{1}{m})\geq v_l(z)-C_3.
\end{equation}
Hence, applying Proposition \ref{Prop fini sum app}, we have
\begin{equation}\label{main s2 eq3}
u(z,\frac{1}{m})\geq N\sum\limits_{\nu(u_0,a)\geq \frac{1}{l}}\log |z-a|+O(1),
\end{equation}
where $N>0$.
It follows from Proposition \ref{first results} that $u$ is smooth and satisfies \eqref{KRF} in the
classical sense in
$(\bar{\Omega}\setminus\{z:\nu (u,z)\geq \frac{1}{l}\})\times (\frac{1}{m},T)$.
When $\epsilon\searrow 0$, we conclude that
$u$ is smooth and satisfies \eqref{KRF} in the
classical sense in
$(\bar{\Omega}\setminus\{z:\nu (u,z)\geq \frac{2}{m}\})\times (\frac{1}{m},T)$ and
$u|_{\partial\Omega\times [0,T)}=\varphi|_{\partial\Omega\times [0,T)}$.
\begin{flushleft}
\it{Step 3: The set of singular points of $u$}.
\end{flushleft}
Assume that $\{z:\nu (u_0,z)\geq \frac{2}{m}\}=\{a_{m,1},...,a_{m,N(m)}\}$. For any $j=1,...,N(m)$, we denote
\begin{equation}
\epsilon_{m,j}=\sup\{T>t\geq 0: \nu(u(.,t'),a_{m,j})>0, \forall 0\leq t\leq t'\}.
\end{equation}
We need to show that
\begin{itemize}
\item $u$ is smooth and satisfies \eqref{KRF} in the classical sense in $Q_m:=\bar{\Omega}\times (\frac{1}{m},T)\setminus
\cup_{j\leq N(m)}\{a_{m,j}\}\times (\frac{1}{m},\epsilon_{m,j}].$
\item $\frac{\nu (u_0,a_{m,j})}{2n}\leq \epsilon_{m,j}\leq \frac{\nu (u_0,a_{m,j})}{2}.$
\item $\nu (u(.,t),a_{m,j})>0$ for any $j=1,...,N(m)$ and $t<\epsilon_{m,j}.$
\item If $\epsilon_{m,j}<T$ then $\nu (u(.,t),a_{m,j})\searrow 0$ as $t\nearrow \epsilon_{m,j}$. \\
\end{itemize}
By Step 2, for any $T>\epsilon>0$, there is $r>0$ such that $u$ is smooth in $(B(a_{m,j},r)\setminus\{a_{m,j}\})\times (\epsilon,T)$
for any $j=1,..., N(m)$.
If there is $\epsilon\in (0,T)$ satisfying $\nu(u(.,\epsilon),a_{m,j})=0$ then
it follows from Proposition \ref{first results}
that $u\in C^{\infty}(B(a_{m,j},r)\times (\epsilon,T))$. By the definition of $\epsilon_{m,j}$, we conclude that $u$ is smooth in
$Q_m$. Clearly, $u$ satisfies \eqref{KRF} in the classical sense in $Q_m$. \\
By Propostion \ref{singular.prop.weak sec} and by the smoothness of $u$ in $Q_m$, we have
$\epsilon_{m,j}\geq \frac{\nu (u_0,a_{m,j})}{2n}$.
If $\nu (u_0,a_{m,j})<2T$, we have, as in the step 2,
\begin{equation}\label{main s3 eq1}
u(a_{m,j},c)\geq v_{\frac{1+\epsilon}{2(c-\epsilon)}}+O(1) =: v_l+O(1),
\end{equation}
for any $0<c<T$ and $c>\epsilon>0$.
If $2c>\nu (u_0, a_{m,j})$ then for any $\epsilon\ll 1$ we have
$$\nu (u_0, a_{m,j}) <\dfrac{1}{l}.$$
It follows from Proposition
\ref{Prop fini sum app} that $v_l$ is smooth in a neighbourhood of $a_{m,j}$. By \eqref{main s3 eq1}, we have
$$c\geq\epsilon_{m,j}.$$ Letting $c\rightarrow\frac{\nu (u_0, a_{m,j})}{2}$,
we obtain $\epsilon_{m,j}\leq \frac{\nu (u_0, a_{m,j})}{2}$. Thus $\epsilon_{m,j}$ satisfies
\begin{center}
$\frac{\nu (u_0,a_{m,j})}{2n}\leq \epsilon_{m,j}\leq \frac{\nu (u_0,a_{m,j})}{2}.$
\end{center}
By definition of $\epsilon_{m,j}$, we have $\nu (u(.,t),a_{m,j})>0$ for any $j=1,...,N(m)$ and $t<\epsilon_{m,j}$. Applying Lemma
\ref{utminusu0.lem.weak sec} for $u_k$ and letting $k\rightarrow\infty$, we conclude that $\nu (u(.,t),a_{m,j})$ is
non increasing in $t$. If $\epsilon_{m,j}<T$ then
by Proposition \ref{singular.prop.weak sec} and by the smoothness of $u$ in $Q_m$, we have, for any
$0<\epsilon<\epsilon_{m,j}$,
\begin{center}
$\dfrac{\nu (u(.,\epsilon),a_{m,j})}{3n}+\epsilon\leq\epsilon_{m,j}.$
\end{center}
Hence, $\nu (u(.,t),a_{m,j})\searrow 0$ as $t\nearrow \epsilon_{m,j}$.
\begin{flushleft}
\it{Step 4: Continuity at zero.}
\end{flushleft}
Applying Lemma \ref{utminusu0.lem.weak sec}, we have
\begin{equation}\label{liminf.conti0.proofmain2}
\liminf\limits_{t\to 0}u(z,t)\geq u_0(z),
\end{equation}
for any $z\in\bar{\Omega}$.
Note that $u$ is the limit of a decreasing sequence of smooth functions, then $u\in USC(\bar{\Omega}\times [0,T))$. We have
\begin{equation}\label{limsup.conti0.proofmain2}
\limsup\limits_{t\to 0}u(z,t)\leq u_0(z),
\end{equation}
for any $z\in\bar{\Omega}$.
Combining \eqref{liminf.conti0.proofmain2} and \eqref{limsup.conti0.proofmain2}, we obtain
$$\lim\limits_{t\to 0}u(z,t)=u_0(z).$$
Hence, by the dominated convergence theorem, $u(., t)\rightarrow u_0$ in $L^1$, as $t\to 0$.
\begin{flushleft}
\it{Step 5: The case $A>0$.}
\end{flushleft}
Assume that $u$ is the weak solution of \eqref{KRF} with $A>0$.
We set $$v(z,t)=(At+1)u(z,\frac{\log (At+1)}{A}).$$ Then we can verify that $v$ is the weak solution of
\begin{equation}
\begin{cases}
\begin{array}{ll}
\dot{v}=\log\det (v_{\alpha\bar{\beta}})-n\log (At+1)+f(z,\frac{\log (At+1)}{A})\;\;
&\mbox{on}\; \Omega\times (0,\frac{e^{AT}-1}{A}),\\
v(z,0)=u_0(z)&\mbox{on}\;\Omega,\\
v(z,t)=\varphi (z,\frac{\log (At+1)}{A})&\mbox{on}\; \partial\Omega\times [0,T).
\end{array}
\end{cases}
\end{equation}
Using the case $A=0$, we conclude the similar result for the case $A>0$.
|
\section{Introduction}
In \cite{Sullivan1,Sullivan2}, Sullivan introduced multiple notions of {\sl tautness} of a foliation, and proved the equivalence of these notions for $C^2$ foliations. The conceptually simplest of these was highlighted further by Gabai (cf, Definition~2.11 of \cite{G1}). When $T\mathcal F$ is only $C^0$, Definition~2.11 of \cite{G1} admits two natural interpretations. Moreover, the notions of tautness introduced by Sullivan are distinct when the criterion that $\mathcal F$ be $C^2$ is dropped.
In this paper we describe several notions of tautness that arise in the study of $C^0$ foliations, $C^{1,0}$ or smoother foliations, and in geometry. We examine how these notions are related, and give examples in \S~\ref{Examples} to show that these notions are different.
The implications of the differences between versions of tautness are clearest in geometry and contact topology. We show in Proposition~\ref{no vp flow} that some smoothly taut foliations are not transverse to a volume preserving flow, whereas by Theorem~\ref{vp flow}, $C^{1,0}$ everywhere taut foliations are always transverse to such flows.
In Theorem~\ref{limit of both} we construct smoothly taut $C^{\infty,0}$ foliations that are $C^0$ approximated both by overtwisted contact structures and by everywhere taut foliations. This contrasts sharply with Theorem~\ref{everwheretautimplies} which shows that $C^{1,0}$ everywhere taut foliations are $C^0$ approximated by, and only by, weakly symplectically fillable, universally tight contact structures.
Since topologically taut $C^0$ foliations are isotopic to everywhere taut $C^{\infty,0}$ foliations (Corollary~\ref{top approx}), the differences between the versions of tautness are unimportant when working with foliations up to topological conjugacy. In particular, for example, an $L$-space does not admit a transversely orientable, topologically taut, $C^0$ foliation (Corollary~\ref{L space}).
This study was inspired by an observation of Vincent Colin, who pointed out to us that if $T\mathcal F$ is only continuous, then $T\mathcal F$ is not uniquely integrable and this impacts the existence of transversals. We thank Vincent for helpful conversations, and in particular for explaining his work, \cite{ColinFirmo}, where these issues arise in the search for a foliation approximated by a pair of contact structures. We also thank the Banff International Research Station for their hospitality, which made these conversations possible.
\section{Definitions}
We begin by recalling the definition of foliation, paying careful attention to smoothness.
\begin{definition}\label{folndefn1} Let $M$ be a smooth 3-manifold with empty boundary. Let $k$ be a non-negative integer or infinity. A {\sl codimension one foliation}, $\mathcal F$, is a decomposition of $M$ into a disjoint union of connected surfaces, called the {\sl leaves} of $\mathcal F$, together with a collection of charts $U_i$ covering $M$, with $\phi_i:\mathbb R^2 \times \mathbb R \to U_i$ a homeomorphism, such that the preimage of each component of a leaf intersected with $U_i$ is a horizontal plane.
The foliation $\mathcal F$ is $C^k$ if the charts $(U_i,\phi_i)$ can be chosen so that each $\phi_i$ is a $C^k$ diffeomorphism.
The foliation $\mathcal F$ is $C^{k,0}$ if the charts $(U_i,\phi_i)$ can be chosen so that the restriction of each $\phi_i$ to a horizontal plane is a $C^k$ immersion and so that the tangent planes of the leaves vary continuously.
\end{definition}
Notice that $T\mathcal F$ exists and is continuous if and only if $\mathcal F$ is $C^{1,0}$. Different amounts of transverse smoothness can also be specified by defining $C^{k,l}$ foliations, $k\ge l$. For the purposes of this paper it is enough to know that $C^{k,1}$ foliations are necessarily $C^1$. See Definition~2.1 of \cite{KR3} for the general definition.
\begin{definition} \cite{KR2} \label{flowboxdefn}
Let $\mathcal F$ be either a $C^k$ or $C^{k,0}$ foliation, and let $\Phi$ be a smooth transverse flow. A {\sl flow box}, $F$, is an $(\mathcal F,\Phi)$ compatible closed chart, possibly with corners. That is, it is a submanifold diffeomorphic to $D\times I$, where $D$ is either a closed $C^k$ disk or polygon (a closed disk with at least three corners), $\Phi$ intersects $F$ in the arcs $\{(x,y)\}\times I$, and each component of $D\times \partial I$ is embedded in a leaf of $\mathcal F$. The components of $\mathcal F\cap F$ give a family of $C^k$ graphs over $D$.
\end{definition}
A {\sl flow box decomposition} is a finite cover of $M$ by flow boxes so that their interiors are pairwise disjoint and intersections along boundaries satisfy certain conditions. The constraints imposed on boundary intersections are not important for what follows, but can be found in Definition~3.1 of \cite{KR4}. A flow box compatible isotopy is an isotopy which fixes setwise the flow boxes, and the cells of flow boxes, in a given flow box decomposition. Given $(M,\mathcal F,\Phi)$, there is a flow box decomposition of $M$, and a flow box decomposition of a submanifold of $M$ always extends to a flow box decomposition of $M$ (Proposition~3.2 of \cite{KR4}).
There are two natural definitions of a {\sl transversal to $\mathcal F$.}
\begin{definition} Let $\mathcal F$ be a $C^0$ foliation of a closed manifold. A {\sl topological transversal} $\gamma$ is a curve which is {\sl topologically} transverse to $\mathcal F$; namely, no nondegenerate sub-arc of $\gamma$ isotopes relative to its endpoints into a leaf of $\mathcal F$.
\end{definition}
\begin{definition} Let $\mathcal F$ be a $C^{1,0}$ foliation of a closed manifold. A curve $c$ is {\sl smoothly transverse} to $\mathcal F$ if it is smooth and $T\gamma$ and $T\mathcal F$ span $TM$ at each point of $c$. For brevity, such a curve is called a {\sl transversal} to $\mathcal F$.
\end{definition}
Notice that a smoothly embedded topological transversal is not necessarily a transversal.
\begin{lemma}\label{isotopetosmooth}
Let $\mathcal F$ be a $C^1$ foliation of $M$, and let $\gamma$ be a topological transversal to $\mathcal F$ passing through a point $p\in M$. There is an isotopy of $M$ relative to $p$ taking $\gamma$ to a transversal through topological transversals to $\mathcal F$.
\end{lemma}
\begin{proof}
Pick a $C^1$ diffeomorphism taking a flow box to a horizontally foliated subset of $\mathbb R^3$. Then, in coordinates, a straight line can connect any point of the lower boundary to any point of the upper boundary. Since the straight line can be modified near the end points to be tangent to any non-horizontal vector, the segments can be glued to give a $C^1$ transverse curve in the manifold. This curve can then be $C^1$ approximated with a smooth transverse curve.
\end{proof}
Therefore, when $\mathcal F$ is $C^1$, there is a transversal through $p$ if and only if there is a topological transversal through $p$. Moreover, when $\mathcal F$ is $C^1$, there is a transversal through every point if there is a transversal through every leaf. However, even for $C^{\infty,0}$ foliations, this is not true (Proposition~\ref{no vp flow}), and hence Definition~2.11 of \cite{G1} gives rise to three distinct notions of tautness.
\begin{definition} A $C^0$ foliation $\mathcal F$ of a closed 3-manifold is {\sl topologically taut} if for every leaf $L$ of $\mathcal F$ there is a simple closed topological transversal to $\mathcal F$ that has nonempty intersection with $L$.
\end{definition}
\begin{definition} A $C^{1,0}$ foliation $\mathcal F$ is {\sl smoothly taut} if for every leaf $L$ of $\mathcal F$ there is a simple closed transversal to $\mathcal F$ that has nonempty intersection with $L$.
\end{definition}
\begin{definition} Let $\mathcal F$ be a $C^{1,0}$ foliation of a closed manifold. The foliation $\mathcal F$ is {\sl everywhere taut}, or simply {\sl taut}, if for every point $p$ of $M$ there is a simple closed transversal to $\mathcal F$ that contains $p$.
\end{definition}
In the absence of sufficient smoothness, these three notions of tautness differ, and they are frequently confused in the literature. Compare Theorem~2.1 of our paper \cite{KR2} with Theorem~\ref{vp flow} for one of many instances of this.
In practice, everywhere taut is the most useful form of tautness. Referring to everywhere taut foliations as taut allows most theorems to be stated without additional hypotheses.
In Section~\ref{Examples}, we give examples showing that the inclusions of foliations $$\{\mbox{everywhere taut}\}\subset \{\mbox{smoothly taut}\}\subset\{ \mbox{topologically taut}\}$$ are proper, even when restricting to $C^{\infty,0}$ foliations.
Restricting to $C^1$ foliations, these proper inclusions become equalities.
\begin{lemma}\label{C1} For $C^1$ foliations, topologically taut implies smoothly taut, and smoothly taut implies everywhere taut.
\end{lemma}
\begin{proof} Suppose $\mathcal F$ is topologically taut. By Lemma~\ref{etautlemma}, there is a topological transversal through every point $p$ of $M$. By Lemma~\ref{isotopetosmooth}, there is therefore a smoothly transverse transversal through every point.
\end{proof}
We include a fourth notion of tautness, introduced by Sullivan in \cite{Sullivan2}, since it motivates the usage of ``taut''.
\begin{definition} Let $\mathcal F$ be a $C^{k,0}$ foliation of a closed manifold with $k \ge 2$. The foliation is {\sl geometrically taut} if there exists a metric on $M$ such that every leaf of $\mathcal F$ is a minimal surface. An excellent reference for this material is the Appendix of \cite{Hass}.
\end{definition}
\section{Standard foliation lemmas} \label{classicallemmas}
We begin by translating three classical results into the context of $C^0$ foliations. In each case, the original proof translates immediately to yield the claimed result. The proofs of each of the following three lemmas can be found, for example, in the proof of Proposition~4, Chapter VII, of \cite{CN}, and are similar. For completeness, we illustrate the key ideas involved by including a proof of the first lemma.
\begin{lemma}\label{join} If $\mathcal F$ is a topologically taut $C^0$ foliation of $M$, then there is a connected closed topological transversal that intersects every leaf of $\mathcal F$.
\end{lemma}
\begin{proof} Let $\gamma$ be a collection of topological transversals that intersect every leaf of $\mathcal F$. Suppose $\gamma$ is not connected. Since $M$ is connected and the set of points on leaves that intersect a single closed transversal is an open set, there are distinct components of $\gamma$ that intersect a common leaf $L$ at points $p$ and $q$ respectively. Let $\alpha$ be an arc in $L$ connecting $p$ and $q$, and let $D\subset L$ be a disk neighborhood of $\alpha$. The foliation restricts to a product homeomorphic to $D \times I$ near $D$. The two transversals can be replaced by a connected transversal by inserting a half twist in $D\times I$.
\end{proof}
\begin{lemma} \label{top everywhere taut}
Suppose $\mathcal F$ is $C^0$ and topologically taut. For every point $p$ in $M$, there is a topological transversal through $p$.\qed
\end{lemma}
\begin{lemma} \label{etautlemma} If $L$ is a noncompact leaf of a $C^0$ foliation, then there is a topological transversal that has nonempty intersection with $L$. \qed
\end{lemma}
A {\sl Reeb component} is a foliation of a solid torus whose boundary is a leaf and such that all other leaves are homeomorphic to planes. See Example~\ref{phantomReeb}. A foliation is {\sl Reebless} if it contains no Reeb component.
For $\mathcal F$ a transversely oriented foliation, a {\sl dead end component} \cite{Thurstonthesis} $C$ of $\mathcal F$ is a connected submanifold of $M$ that is cobounded by a finite collection of torus leaves $T_1,...T_n$ of $\mathcal F$ so that, for one of the two choices of transverse orientation of $\mathcal F$, $C$ lies on the positive side of each $T_i$.
The next result is well known and can be proved using ideas found in \cite{Goodman}.
\begin{prop} [Theorem~1, \cite{solodov}] \label{suegoodman}
A transversely oriented $C^0$ foliation $\mathcal F$ is topologically taut if and only if it contains no dead end components.
\end{prop}
These results are not true if topological transversal is replaced by smooth transversal. One issue that arises is the existence of {\sl phantom Reeb components} and, more generally, {\sl phantom dead end components.}
\section{Phantom leaves}\label{Examples}
In this section, we give examples that highlight some differences between $C^1$ foliations and foliations with only continuous tangent plane field. We begin by producing examples of Reebless $C^{1,0}$ foliations, and even Reebless $C^{\infty,0}$ foliations, that contain the tangent plane field of a compressible torus.
\begin{definition}
Let $\mathcal F$ be a $C^{1,0}$ foliation. A {\sl phantom} surface of $\mathcal F$ is a $C^1$ embedded surface in $M$ that is not contained in a leaf of $\mathcal F$, but has tangent plane field contained in $T\mathcal F$.
\end{definition}
\begin{figure}[htbp]
\centering
\includegraphics[width=3.75in]{Tangent_torus}
\caption{A Reeb-like annulus and a Reeb annulus with a collar on the compact leaves.}
\label{tangent torus}
\end{figure}
\begin{example}[Phantom Reeb component] \label{phantomReeb} Let $A = [-1,1] \times S^1$ be the annulus shown in Figure~\ref{tangent torus}, and let $\theta=0 \in S^1$. Let $a_0$ be a smooth embedded arc in $A$ with boundary $\{\pm1\} \times \{0\}$ that is symmetric when the first coordinate is negated, is tangent to $ \{\pm 1/2\} \times S^1$ and is transverse to $\{x\} \times S^1$ for $x \neq \pm 1/2$. As $\theta$ varies over $S^1$, let $a_\theta$ be the result of translating $a_0$ in the $S^1$ coordinate by $\theta$.
The union of the $a_\theta$ is a $C^{\infty,0}$ transversely orientable foliation of $A$. The curves $\{\pm 1/2\} \times S^1$ are integral curves for the foliation, and the transverse orientation may be chosen so that it points out of $S^1 \times[-1/2, 1/2]$. Call the foliated annulus $[-1/2,1/2] \times S^1$ a {\sl phantom Reeb annulus}.
Rotating $A$ about $\{0\} \times S^1$ produces a $C^{\infty, 0}$ Reeb-like foliation, called a {\sl phantom Reeb component}, $\mathcal R$ of $D^2 \times S^1$. The torus, $T$, of radius $1/2$ about $\{0\} \times S^1$, is an example of a phantom torus.
Figure~\ref{tangent torus} also shows a foliation that when rotated about $\{0\} \times S^1$ is both $C^0$ close to $\mathcal R$ and has an actual Reeb component. This foliation also has a product foliation by tori in a small neighborhood of the boundary of the Reeb component.
\end{example}
\begin{example} [$C^{\infty,0}$ and topologically taut, but not smoothly taut] \label{top not smooth}
Produce a foliation on $A = [-1,1] \times S^1$ first by letting $\{\pm 1\} \times S^1$ be leaves. Let $\theta=0 \in S^1$, and let $b_0$ be a non-compact smooth arc that agrees with $a_0$ on $ [-1/2,1/2] \times S^1$, is transverse to $\{x\} \times S^1$ for $x \in (-1,-1/2) \cup (1/2,1)$, limits on $\{1\} \times S^1$ in the direction of increasing $\theta$, and limits on $\{-1\} \times S^1$ in the direction of decreasing $\theta$. For other $\theta \in S^1$ let $b_\theta$ be given by translating $b_0$ through an angle of $\theta$ in the $S^1$ coordinate. See Figure~\ref{topologically taut}.
Identifying each pair of points $( \pm 1, \theta)$ gives a $C^{\infty,0}$ transversely orientable foliation $\mathcal T$ of a torus.
\end{example}
\begin{figure}[htbp]
\centering
\includegraphics[width=1.75in]{Topologically_taut}
\caption{Topologically, but not smoothly, taut}
\label{topologically taut}
\end{figure}
\begin{prop} The foliation $\mathcal T$ is topologically taut, but not smooth\-ly taut.
\end{prop}
\begin{proof} It is straightforward to show that $\mathcal T$ is topologically taut. Since the single compact leaf, $\{\pm 1\} \times S^1$ of $\mathcal T$ is parallel to the Reeb-like annulus, no smooth closed transversal can intersect it.
\end{proof}
Next we note that a transverse torus can often be used to create a phantom torus.
\begin{prop} \label{phantomtori}
Let $\mathcal F$ be a $C^{1,0}$ foliation of $M$. Suppose there is a smoothly embedded torus $T$ in $M$ that is smoothly transverse to each leaf of $\mathcal F$, with $\mathcal F\cap T$ everywhere taut.
Then there is an isotopy of $M$, supported on a small neighborhood of $T$, that takes $\mathcal F$ to a $C^{1,0}$ foliation with phantom torus $T$.
\end{prop}
\begin{proof}
Set $\lambda=\mathcal F\cap T$. Since $\lambda$ is everywhere taut (as a foliation of $T$) it can be isotoped to be transverse to the smooth product foliation $\{t\}\times S^1$, for some choice of smooth product structure $ S^1\times S^1$ on $T$. Hence, there is an isotopy of $M$, supported on a small neighborhood of $T$, and
a smooth parametrization $(x,y,z)\in S^1\times [-1,1] \times S^1$ of a smaller neighborhood of $T$, so that $T$ is given by $y=0$, $\partial/\partial y$ is tangent to $\mathcal F$, $\mathcal F$ meets $S^1\times [-1,1] \times S^1$ in the product foliation $\lambda\times [-1,1]$, where $[-1,1]$ describes the range of the $y$-coordinate, and $\partial/\partial z$ is smoothly transverse to $\mathcal F$.
Let $b(y)$ be a damped version of $y^{1/3}$. That is, it is $0$ away from $0$ and has a vertical tangency at $y=0$. Now consider the homeomorphism $(x, y, z) \to (x, y, z + b(y))$. It takes $\mathcal F$ to a new foliation tangent to $T$.
\end{proof}
Let $c$ be a transversal to a $C^{\infty,0}$ foliation $\mathcal F$. Applying Proposition~\ref{phantomtori} to $\mathcal F$ and $T=\partial N(c)$, where $N(c)$ is a sufficiently small smooth regular neighborhood of $c$, yields examples of $C^{\infty,0}$ foliations of $M$ that are smoothly taut, but not everywhere taut.
\begin{example} [$C^{\infty,0}$ and smoothly taut, but not everywhere taut] \label{phantomReebinsert} Let $\mathcal F$ be a smoothly taut, transversely oriented, $C^{\infty,0}$ foliation of $M$ and let $c$ be a smooth closed transverse curve. Identify $N(c)$ and $\mathcal F|N(c)$ with $D^2 \times S^1$ foliated by disks. Produce a $C^{\infty, 0}$ foliation $\mathcal H$ by replacing $\mathcal F$ on $D^2 \times S^1$ by $\mathcal R$.
\end{example}
\begin{prop}\label{no vp flow} The foliations of Example~\ref{phantomReebinsert} are smoothly taut, but they are not everywhere taut. They admit no closed dominating 2-form, and they are not transverse to any volume preserving flow for any choice of Riemmanian metric.
\end{prop}
\begin{proof} Since $T$ is a separating integral surface, no smooth closed trans\-versal can intersect a point of $T$. The existence of a dominating 2-form would contradict Stokes theorem applied to $T$. Since $T$ separates, a volume preserving flow transverse to $\mathcal H$ can not exist.
\end{proof}
The construction of Example~\ref{phantomReebinsert} will be used in Corollary~\ref{limits of Reeb} to show that all foliations are, up to isotopy, limits of foliations with Reeb components.
Each of the preceding examples took advantage of a phantom Reeb component. More generally, one can construct examples with one or more phantom dead end components.
\begin{definition}
For $\mathcal F$ a transversely oriented foliation, a {\sl phantom dead end component} $C$ of $\mathcal F$ is a connected submanifold of $M$ that is cobounded by a finite collection of tori $T_1,...T_n$, where $T_1$ is a phantom leaf and each $T_j,j>2,$ is either a leaf or a phantom leaf for $\mathcal F$, so that, for one of the two choices of transverse orientation of $\mathcal F$, $C$ lies on the positive side of each $T_i$.
\end{definition}
Notice that if $C$ is a phantom dead end component of $\mathcal F$, cobounded by tori $T_1,...,T_n$, and $T_j$ is a leaf of $\mathcal F$, then there is no transversal through $T_j$.
\begin{example} [Phantom dead ends obstruct smooth tautness] \label{examplelala}
Let $\mathcal F$ be a topologically taut $C^{\infty,0}$ foliation containing a torus leaf $T_1$. Suppose further that there is a torus $T_2$ isotopic to $T_1$ that meets $\mathcal F$ transversely in a $C^1$ foliation without Reeb annuli. Proposition~\ref{phantomtori} can be applied to $\mathcal F$ and $T_2$ to obtain a $C^{\infty,0}$ foliation $\mathcal G$ so that $T_1$ is a leaf of $\mathcal G$, and $T_1$ and $T_2$ cobound a dead end component $T\times I$ of $\mathcal G$. Since $\mathcal F$ is topologically taut, so is $\mathcal G$. However, since there is no transversal through $T_1$, the foliation $\mathcal G$ is not smoothly taut.
There are many examples of foliated manifolds $(M,\mathcal F)$ satisfying these conditions. For example, begin with a manifold $X$ with torus boundary and a topologically taut foliation $\mathcal F_0$ transverse to $\partial X$. Let $\mathcal F_1$ be the foliation of $X$ obtained from $\mathcal F_0$ by adding as leaf $T_1=\partial X$, and modifying the leaves of $\mathcal F_0$ in a collar of $\partial X$ so that they spiral about $\partial X$. Let $(M,\mathcal F)$ be the double of $(X,\mathcal F_1)$.
\end{example}
In fact, even with the hypotheses of smooth tautness, the usual argument (see Lemma~\ref{join}) for combining several smooth transversals produces a single curve that may be only topologically transverse. This is illustrated by the following family of examples.
\begin{example} [$C^{\infty,0}$ and smoothly taut, but there is no transversal that has nonempty intersection with every leaf] Begin with a 3-manifold $X$ with torus boundary and smoothly taut foliation $\mathcal F_0$ transverse to $\partial X$. Choose $(X,\mathcal F_0)$ so that $\mathcal F_0$ has minimal set disjoint from $\partial X$. Now let $(M,\mathcal F)$ be the double of $(X,\mathcal F_1)$. There are disjoint tori $T_1$ and $T_2$ in $M$ parallel to $\partial X$, and these can be chosen to lie transverse to $\mathcal F$. Applying Proposition~\ref{phantomtori} to $\mathcal F$ and the tori $T_1$ and $T_2$ yields a $C^{\infty,0}$ foliation $\mathcal G$ that is smoothly taut and has phantom dead end component. However, there is no connected transversal that has nonempty intersection with each leaf of $\mathcal G$.
\end{example}
\begin{example} The previous examples are built around the separating properties of certain phantom tori. The same effects can be achieved without tangent tori. Figure~\ref{crease} shows an alternate version of the first foliation described in Example~\ref{phantomReeb}.
\end{example}
\begin{figure}[htbp]
\centering
\includegraphics[width=1.75in]{Crease}
\caption{}
\label{crease}
\end{figure}
\section{Approximating foliations by foliations}
Next we show that topologically taut foliations can be $C^0$ approximated by isotopic smoothly taut foliations which can in turn be approximated by isotopic everywhere taut foliations. Thus in seeking geometric applications of foliations, it is enough to prove the existence of a topologically taut foliation.
\begin{definition} A {\sl plaque} of a flow box $F$ is a connected component of a leaf of $\mathcal F$ intersected with $F$. A {\sl plaque neighborhood in $F$} is a connected closed set with nonempty interior that is a union of plaques.
\end{definition}
The next lemma is immediate by Proposition~3.10 of \cite{KR4}.
\begin{lemma}\label{create U} Suppose $\mathcal F$ is a $C^{\infty,0}$ foliation, and $U$ is a small open regular neighborhood of a finite collection of pairwise disjoint plaque neighborhoods. There is a $C^0$ small isotopy of $M$ taking $\mathcal F$ to a $C^0$ close $C^{\infty,0}$ foliation that is smooth when restricted to $U$. \qed
\end{lemma}
\begin{lemma}\label{smoothly connect} Let $F=D^2 \times I$ be a flow box for $(\mathcal F, \Phi)$. Suppose $\mathcal F$ is smooth on a plaque neighborhood $U$ of $F$. If $p \in D^2 \times \{0\}$ and $q \in D^2 \times \{1\}$, then there exists a transversely smooth arc $\alpha$ agreeing with $\Phi$ in a neighborhood of $p$ and $q$.
\end{lemma}
\begin{proof} Choose a segment of $\Phi$ that starts at $p$ and ends at point $u_1 \in U$. Pick another segment that starts at $q$ and ends at $u_2 \in U$ where $u_2$ lies in a plaque just above the plaque containing $u_1$. These segments can be combined with a transverse arc connecting $u_1$ and $u_2$ to produce the desired smooth arc $\alpha$.
\end{proof}
\begin{thm}\label{top approx by smooth} Given a topologically taut $C^{1,0}$ foliation $\mathcal F$, there is a $C^0$ small isotopy of $M$ taking $\mathcal F$ to a $C^0$ close, smoothly taut, $C^{\infty, 0}$ foliation.
\end{thm}
\begin{proof} Let $\gamma$ be a collection of topological transversals that intersect every leaf of $\mathcal F$. Choose a collection of flow boxes that cover $\gamma$, that intersect only along their horizontal boundaries, and that have vertical boundaries disjoint from $\gamma$. Extend these to a flow box decomposition $\mathcal B$ of $M$.
Choose a plaque neighborhood in each flow box of $\mathcal B$, and let $U$ be a small regular neighborhood of their union. Apply Lemma~\ref{create U} to obtain an isotopy taking $\mathcal F$ to a foliation smooth on $U$, and denote the image of $\gamma$ under this isotopy by $\gamma'$. Lemma~\ref{smoothly connect} can be used to replace $\gamma'$ with a transversely smooth curve $c$ one flow box at a time.
\end{proof}
\begin{lemma}\label{everywhere} Let $\mathcal B$ be a flow box decomposition for an topologically taut $C^{1,0}$ foliation $\mathcal F$. There is a $\mathcal B$ compatible isotopy taking $\mathcal F$ to a $C^0$ close $C^{\infty,0}$ foliation $\mathcal G$ for which there exists a connected closed transversely smooth curve $c$ such that $c$ intersects every plaque of every flow box.
\end{lemma}
\begin{figure}[h]
\centering
\includegraphics[width=3in]{transversals}
\caption{}
\label{transversals}
\end{figure}
\begin{proof} By Theorem~4.1 and Proposition~3.10 of \cite{KR4} there is a $\mathcal B$ compatible isotopy taking $\mathcal F$ to $\mathcal G$, where $\mathcal G$ is $C^{\infty,0}$ and smooth on a set $U$ that contains a plaque neighborhood of every flow box of $\mathcal B$. By Lemma~\ref{top everywhere taut}, there is a closed topological transversal $\gamma_i$ intersecting $U_i$, for each component of $U_i$ of $U$. Using the method of Lemma~\ref{join}, construct a connected topological transversal $\gamma$ from the $\gamma_i$ so that $\gamma$ intersects every component of $U$.
Now replace $\gamma$ with a smooth closed transversal $c$ by performing a small isotopy that preserves $\mathcal B$ and smoothness of the foliation on $U$. To do this, first choose small flow boxes containing $\gamma$ and subordinate to $\mathcal B$, and then apply the smoothing operation of Lemma~\ref{top approx by smooth}.
Let $F$ be a flow box and choose $p\in F$. As shown in Figure~\ref{transversals}, consider the segment $\Phi_p$ of $\Phi$ that contains $p$, flows up and down, crosses both horizontal boundary components of $F$, and ends flowing up at $u_i \in U_i$ and flowing down at $u_j \in U_j$. Let $\tau$ be a small push off of a segment of $c$ that starts in $U_i$ flowing up and ends in $U_j$. Smoothly joining $\tau$ and $\Phi_p$ through $U_i \cup U_j$ produces a smooth transversal $c_p$ that passes through every plaque of $F$.
The original transversal $c$ can be smoothly connected with $c_p$ using a half twist. Repeat this construction for every flow box of $\mathcal B$.
\end{proof}
\begin{cor}\label{top approx} Given a topologically taut $C^{1,0}$ foliation $\mathcal F$, there is a $C^0$ small isotopy of $M$ taking $\mathcal F$ to a $C^0$ close, everywhere taut, $C^{\infty, 0}$ foliation.
\end{cor}
\begin{proof} Since there was no restriction on the choice of $p$ in the proof of Lemma~\ref{everywhere}, this follows immediately.
\end{proof}
The next corollary explores the limiting interaction between foliations with and without Reeb components.
\begin{cor}\label{limits of Reeb} A $C^{0}$ Reebless foliation is isotopic to a $C^0$ limit of $C^{\infty,0}$ foliations with Reeb components. \qed
\end{cor}
\begin{proof}
Let $\mathcal F$ be a $C^0$ Reebless foliation, and let $\gamma$ be a closed topological transversal to $\mathcal F$. To see that such a curve exists, pick a flow transverse to $\mathcal F$, and consider an arc of the flow that starts and ends in a single flow box. The method of Lemma~\ref{join} can be used to replace the arc with a topological transversal.
Theorem~\ref{top approx by smooth} is stated for taut foliations, but the method of proof allows a single topological transversal, $\gamma$, to be isotoped to a transversal $c$. Thus applying Theorem~\ref{top approx by smooth} produces an isotopic $C^{\infty,0}$ foliation and a transversal $c$.
The method of Example~\ref{phantomReebinsert} produces an isotopic $C^{\infty,0}$ foliation $\mathcal H$ that is the $C^0$ limit of foliations with Reeb components.
\end{proof}
\section{Dominating 2-forms and volume preserving flows}
For completeness we include a $C^{1,0}$ version of Sullivan's theorem \cite{Sullivan2}.
\begin{thm}\label{vp flow} Let $\mathcal G$ be an everywhere taut, transversely oriented, $C^{1,0}$ foliation. There exists a smooth, closed 2-form $\omega$ on $M$ such that $\omega$ is positive on $T\mathcal G$. Moreover, fixing any Riemannian metric on $M$, there is a smooth volume preserving flow $\Phi$ transverse to $\mathcal G$.
\end{thm}
\begin{proof} Let $\beta$ be a smooth 2-form on a disk $D$ such that $\beta$ is $0$ near $\partial D$ and is otherwise positive on $TD$. Let $\pi:S^1 \times D^2 \to D^2$ be projection so that $\pi^*\beta$ is a closed form on $S^1 \times D^2$.
Given a transversal $\gamma$ to $\mathcal G$, let $N(\gamma)$ be a small solid torus neighborhood foliated by disks of $\mathcal G|N(\gamma)$. Choose a diffeomorphism $h:N(\gamma) \to S^1 \times D^2$ that maps leaves of $\mathcal G|N(\gamma)$ to disks that are transverse to the first coordinate. Then $h^*\pi^*\beta$ is positive on $T\mathcal G$ in a neighborhood of $\gamma$ and non-negative at all points of $N(\gamma)$.
Since $M$ is compact and $\mathcal G$ is everywhere taut, there is a finite collection of smooth transversals $\gamma_i$ such that at every point of $M$, at least one of $h_i^*\pi^*\beta$ is positive on $T\mathcal G$. If follows that $\omega = \Sigma_i h_i^*\pi^*\beta$ has the desired properties.
Let $\Omega$ be a volume form on $M$. The equation $\omega = X \lrcorner \Omega$ uniquely determines a vector field $X$ that is transverse to $\mathcal G$. Let $\Phi$ be the associated flow for $X$. By Cartan's formula,
$$\mathcal{L}_X \Omega = X \lrcorner d\Omega + d(X\lrcorner\Omega) = d(X\lrcorner \Omega) = d\omega =0,$$ and it follows that $\Phi$ preserves volume.
\end{proof}
\begin{cor}\label{topdomform}
Let $\mathcal F$ be a topologically taut, transversely oriented, $C^{1,0}$ foliation. There is a $C^0$ small isotopy of $M$ taking $\mathcal F$ to a $C^0$ close, everywhere taut, $C^{\infty, 0}$ foliation $\mathcal G$ satisfying
\begin{enumerate}
\item there exists a smooth, closed 2-form $\omega$ on $M$ such $\omega$ is positive on $T\mathcal G$, and,
\item fixing any Riemannian metric on $M$, there is a smooth volume preserving flow $\Phi$ transverse to $\mathcal G$.
\end{enumerate}
\end{cor}
\begin{proof}
This follows immediately from Corollary~\ref{top approx} and Theorem~\ref{vp flow}.
\end{proof}
\section{Approximating foliations by contact structures}
In this section we contrast the main result of \cite{bowden} and \cite{KR3}, which gives properties of any contact approximation of an everywhere taut foliation with the corresponding result, Theorem~\ref{limit of both}, for topologically taut foliations.
\begin{thm} [Theorem~1.2 of \cite{bowden} and Theorem~1.2, \cite{KR3}, approximation without tautness] \label{approx} Let $M$ be a closed, connected, oriented 3-manifold, and let $\mathcal F$ be a transversely oriented $C^{1,0}$ foliation on $M$. Then $\mathcal F$ can be $C^0$ approximated by a positive (respectively, negative) contact structure if and only if $\mathcal F$ is not a foliation of $S^1\times S^2$ by spheres.
\end{thm}
\begin{thm}[\cite{bowden,KR2, KR3}]\label{everwheretautimplies} Let $\mathcal G$ be an everywhere taut, $C^{1,0}$ foliation on a manifold other than $S^2 \times S^1$ and let $\Phi$ be a transverse volume preserving flow. Then $\mathcal G$ can be $C^0$ approximated by a positive (respectively, negative) contact structure. Moreover, any contact structure that is transverse to $\Phi$ is weakly symplectically fillable and universally tight.
\end{thm}
\begin{cor} \label{topologicallytautimplies}
Let $\mathcal F$ be a topologically taut, $C^{1,0}$ foliation on a manifold other than $S^2 \times S^1$. Then $\mathcal F$ can be approximated by a a pair of contact structures $\xi_{\pm}$, $\xi_+$ positive and $\xi_-$ negative, such that $(M,\xi_+)$ and $(-M,\xi_-)$ are weakly symplectically fillable and universally tight.
\end{cor}
\begin{proof}
This follows immediately from Corollary~\ref{topdomform} and Theorem~\ref{everwheretautimplies}.
\end{proof}
When the condition on a foliation is weakened from everywhere taut to smoothly taut,
it no longer follows that any positive contact structure sufficiently close to $\mathcal F$ must be weakly symplectically fillable and universally tight.
\begin{thm}\label{limit of both}
There exist $C^{\infty,0}$ smoothly taut, transversely oriented foliations which can be $C^0$ approximated both by weakly symplectically fillable, universally tight contact structures and by overtwisted contact structures.
\end{thm}
\begin{proof} Let $\mathcal F$ be any transversely oriented, everywhere taut $C^{\infty,0}$ foliation, and let $c$ be a transversal to $\mathcal F$ that has nonempty intersection with every leaf of $\mathcal F$. As described in Example~\ref{phantomReebinsert}, replace a regular neighborhood of $c$ foliated by disks by a phantom Reeb component to create a new $C^{\infty,0}$ foliation $\mathcal H$. The foliation $\mathcal H$ is smoothly taut, but not everywhere taut.
By Corollary~\ref{topologicallytautimplies}, $\mathcal H$ can be approximated by a weakly symplectically fillable, universally tight contact structure.
The construction of $\mathcal H$ involved the creation of a phantom Reeb component near a smooth closed transversal. Replacing a neighborhood of this phantom Reeb component with the $C^0$ close Reeb foliation described in Example~\ref{phantomReeb} gives a $C^0$ close $C^{\infty,0}$ foliation $\mathcal H_1$. Thus $M$ can be written as the union of three codimension 0 pieces: $S$, a solid torus with the Reeb foliation, $T \times I$, the product of a torus and interval foliated as smooth product, and $C$, the closure of the complement of $(T \times I) \cup S$, foliated by the restriction of $\mathcal H_1$.
Set $I=[a,b]$, and consider the contact structures $\xi_{\epsilon,n}= \mbox{ker } \alpha_{\epsilon,n}$ on $T\times I$ given by
$$\alpha_{\epsilon,n} = dz+\epsilon (cos \,nz \,dx + sin\, nz \,dy).$$
Such contact structures have constant slope characteristic foliations on each $T \times \{z\}$. By constraining $\epsilon>0, n, a,$ and $b$ appropriately, we obtain a contact structure $\xi=\xi_{\epsilon,n}$ on $T \times I$ such that
\begin{enumerate}
\item $\xi$ has Giroux torsion greater than 1,
\item $\xi$ strictly dominates $\mathcal F$ along $\partial(T \times I)$, and
\item $\xi$ is $C^0$ close to the product foliation by vertical tori.
\end{enumerate}
The first condition means that all slopes occur, at least once, as $z$ moves across $I$. The second condition is that the slopes at $a$ and $b$ must be close to, and strictly dominate, in the sense of Definition~5.3 of \cite{KR2}, i.e., the slope of $\mathcal F \cap \partial(T \times I)$ when viewed from outside of $C \cup S$. The third condition follows by choosing $\epsilon>0$ sufficiently small (see Proposition 2.3.1 \cite{ET}).
The next step is to extend $\xi$ over all of $S \cup C$. Let $J=[a - \delta, b + \delta]$ and regard $\xi$ as defined on $T \times J$, a neighborhood of $T \times I$ in $M$ for which $\mathcal H_1 \cap \partial (T \times J) = \mathcal F \cap \partial (T \times J)$. If $\delta$ is small enough, $\xi$ will dominate $\mathcal F \cap \partial(T \times J)$, or equivalently, $\xi$ dominates $\mathcal H_1 \cap \partial (T \times J)$.
Since $c$ has nonempty intersection with every leaf of $\mathcal F$, condition (2) is what is required to apply the techniques of \cite{KR2} to extend $\xi$ from $T \times J$ to the remaining portion $C\cup S$ while continuing to approximate $\mathcal H_1$.
Choose $z_0\in I$ so that the slope of the characteristic foliation of $\xi$ in $T \times\{z_0\}$ matches the slope of a compressing disk. Then any closed curve of the characteristic foliation on $T\times\{z_0\}$ bounds an overtwisted disk.
\end{proof}
Since Theorem~{3.4} of \cite{calegari} shows that any $C^0$ foliation can be isotoped to a $C^{\infty,0}$ foliation, Corollary~\ref{topologicallytautimplies} implies the following.
\begin{cor} [Corollary~1.7, \cite{KR3}] \label{L space}
An $L$-space does not admit a transversely orientable, topologically taut, $C^0$ foliation. \qed
\end{cor}
|
\section{Introduction}
\label{intro}
Magnetic reconnection refers to topological reconfiguration of magnetic field, which is widely accepted as the mechanism of energy release in solar flares \citep{prie02}. In the standard two-ribbon flare model \citep{carm64,stur68,hira74,kopp76}, oppositely-directed magnetic field lines reconnect at a current sheet, creating closed field lines that cross over a polarity inversion line (PIL) below the current sheet and a plasmoid structure above. More sets of anti-parallel field lines flow into the current sheet and reconnect at a growing height, producing two flare ribbons moving outward away from the PIL. In this classical two-dimensional (2D) picture, magnetic reconnection takes place at a null-line (or X-line) between anti-parallel coplanar field lines, forming a post-flare arcade and ribbons outlining the feet of the arcade.
In reality, magnetic reconnection in solar flares must be three-dimensional (3D). In the 3D framework, magnetic reconnection takes place preferentially at a separator (intersection of two separatrices; \citealt{baum80,lauy90,long96,long05}) or at its generalization, a hyperbolic flux tube (HFT; intersection of two quasi-separatrix layers, or QSLs; \citealt{demo96,demo97,tito02}). The 3D magnetic reconnection is manifest in many two-ribbon flare observations \citep{gorb88,mand91,demo93,demo07}, such as bi-directional spread of post-flare arcade \citep{aula07}, anti-parallel motions of hard X-ray sources \citep{aula06,demo07}, elongation motion of flare ribbons \citep{qiuj02,qiuj10,qiuj09,flet04}, and $J$-shaped flare ribbons \citep{aula12}. Besides two-ribbon flares, circular-ribbon flares also exhibit a special configuration of 3D magnetic reconnection with a null point and a fan-spine structure \citep{mass09,sunx13,yang15}.
In this Letter, we present observational evidence for a different scenario from those reported previously: an X-shaped ribbon flare with the ribbons intersecting at the center, revealed by the high-resolution ultraviolet images from the {\em Interface Region Imaging Spectrograph} ({\em IRIS}; \citealt{depo14}). Simultaneously, the extreme-ultraviolet images from the Atmospheric Imaging Assembly (AIA; \citealt{leme12}) on board the {\em Solar Dynamics Observatory} ({\em SDO}) show converging non-coplanar loops. These elements suggest a 3D reconnection scenario with a separator, and for the first time, that separator and the reconnection occurring there have been observed.
\section{Observations}
The M2.3 flare presented here was observed by {\em IRIS} in slit-jaw 1330 \AA~images (SJI 1330 \AA) from 15:17 UT to 16:05 UT on 2014 November 9. Figure \ref{fig-obs} gives an overview of the flare. The flare 1--8 \AA~soft X-ray emission starts at $\sim$15:22 UT and peaks at 15:32 UT, as shown in the top panel. Also shown are light curves of different parts of the flare ribbons in SJI 1330 \AA, which observe the upper chromosphere and lower transition region.
The lower panels of the figure and the animation show the flare evolution in SJI 1330 \AA, and AIA 1600, 171, and 131 \AA~passbands, demonstrating formation of an X-shaped flare ribbon. Brightenings along four branches of the ribbon in the northeast (NE), southeast (SE), northwest (NW), and southwest (SW) spread toward each other and finally cross at one point around the flare peak, when the ribbons exhibit an X-shape. Furthermore, we see evident post-flare loops in AIA 131 and 171 \AA~images in the decay phase, which appear to converge toward the center of the X-shape.
Here, we analyze flare ribbon evolution observed in SJI 1330 \AA~which have a pixel size of $0.\!\!^{\prime\prime}166$ and a cadence of 37~s. We also use the AIA and HMI (Helioseismic and Magnetic Imager; \citealt{scho12}) data from {\em SDO} to construct magnetic topology of the flare region. The AIA images have a pixel scale of $0.\!\!^{\prime\prime}6$. The HMI magnetic field data used here are taken from pre-flare. All the images from {\em IRIS} and {\em SDO} are co-aligned by comparing the sunspot features shown in both SJI 2832 \AA~and AIA 1700 \AA~images.
\section{Analysis and Results}
\subsection{Flare Ribbon Motion}
This flare shows an unusual ribbon pattern as revealed by high spatial-resolution SJI 1330 \AA. The ribbon's apparent motion reveals three stages of flare evolution: the initial (15:22--15:28 UT), main (15:28--15:32 UT), and last (15:32--15:38 UT) stages, denoted by the red vertical dotted lines in the top panel of Figure \ref{fig-obs}. The top panels of Figure \ref{fig-ribbon} present the three-stage motion pattern of footpoint brightenings in 1330 \AA. The SJI 1330 \AA~observe brightening in the flare ribbons as well as some low-temperature loops. To distinguish brightening of footpoints from loops or other transient non-ribbon features, we select flaring pixels whose intensity is enhanced to be more than 25 times the intensity of the quiescent-Sun for more than two minutes \citep{long07}, and identify these pixels as footpoints of flare loops.
Seen from Figure \ref{fig-ribbon}, in the initial stage, two ribbons on the right of the flare region (NW ribbon in positive magnetic fields and SW ribbon in negative fields) brighten first and generally show an elongation motion toward the left. These ribbons are on the two sides of the PIL and are nearly parallel with the PIL along the east-west direction, and their morphology resembles the standard two-ribbon flare configuration. In the subsequent main stage, the other two ribbons on the left (NE ribbon in negative fields and SE ribbon in positive fields) are brightened from the outer-most ends, and the brightenings spread to the right as well as approach each other. In the meanwhile, the previously brightened NW and SW ribbons spread further, converging toward each other. The four ribbons, when fully formed, intersect at one point marked in the top panels of Figure \ref{fig-ribbon}, and form an unusual X-shape extending from this point, which we refer as the X-point in the foregoing text. Since the X-point is located at the center of a quadrupolar magnetic structure, it is plausible that a separator may be anchored there. In the last stage, all the ribbons display an outward motion; in particular, the NW and SW ribbons move outward away from the local PIL like in the 2D picture.
We track the ribbon brightening to measure the apparent motion, which is then shown in the two middle panels of Figure \ref{fig-ribbon}. For the inward motion occurring in the initial and main stages, we measure the distance of the ribbon fronts from the X-point. In the initial stage, the fronts of the NW and SW ribbons both move toward the X-point but in the direction nearly parallel to the PIL separating the two ribbons. At the end of this stage, the NW ribbon is $\sim$3 Mm from the X-point. In the following main stage, the fronts of the NW and SW ribbons have changed the motion direction to also approach each other until they converge at the X-point; the apparent speed of the converging motion is $\sim$17 and $\sim$85 km s$^{-1}$ for the NW and SW ribbons, respectively. Simultaneously, the NE and SE ribbons are brightened and their fronts converge to the X-point at the speeds of $\sim$47 and $\sim$66 km s$^{-1}$, respectively. The four ribbons intersect at the X-point at 15:31 UT.
In the main stage, whereas the leftmost fronts of the NW and SW ribbons are moving toward the X-point, the previously-formed portions of these ribbons also start to spread outward away from the PIL. This separation motion continues into, and dominates, the last stage. To measure the separation motion of NW and SW ribbons, we track the distance of the ribbon fronts along the direction perpendicular to the PIL at a point marked by a triangular symbol in the top left panel of Figure \ref{fig-ribbon}. The speed of this separation motion is much smaller, of mostly $<$ 9 km s$^{-1}$.
Apparent motion of flare ribbons is a consequence of energy release along flare loops successively formed by magnetic reconnection. Therefore, magnetic reconnection flux can be measured by summing up magnetic flux swept up by flare ribbons \citep{forb84,pole86}. The bottom panel of Figure \ref{fig-ribbon} shows reconnection flux and its time derivative, the reconnection rate, measured in the positive and negative fields respectively. It is seen that the reconnection flux grows rapidly in the main stage, which is dominated by the ribbon's inward motion. The total reconnection flux approaches 10$^{21}$ Mx with the positive and negative fluxes reasonably balanced. The reconnection rate reaches its maximum of 8$\times$10$^{18}$ Mx s$^{-1}$ early in the main stage and diminishes toward the end of this stage. All of these suggest that most of reconnection and subsequent energy release take place during the main stage of the flare, when ribbons form the peculiar X-shape.
\subsection{Magnetic Topology by 3D Field Extrapolation}
According to a common topological picture\footnote{Here we refer only to the interpretation of this flare, in which we find evidence for separator reconnection. We recognize that some other events show evidence for reconnection associated with more general topological features like QSLs or HFTs.}, flare ribbons form at the bases of separatrices of the coronal field. A positive ribbon separates the positive polarity connecting to one negative source, say $N1$, from that connecting to another, $N2$. Similarly, the negative ribbon separates field lines originating at $P1$ from those originating at $P2$. The reconnection responsible for the flare occurs as field lines from two connectivities, $P1$--$N1$ and $P2$--$N2$, are eliminated to form field lines of the other two: $P1$--$N2$ and $P2$--$N1$. The increase of the latter two, and decrease in the former two, cause the prototypical spreading of the ribbons. The reconnection is supposed to occur at a separator formed by the intersection of the two separatrices. This is the one location where field lines from all four distinct connectivities are in close proximity. In the vast majority of cases, the separator\footnote{Note that numerous separators could be found in magnetic topological models \citep{long07a,parn10}.} is indirectly inferred, and believed to lie high in the corona, where the flare reconnection is actually occurring.
The present flare has the unusual property that its positive and negative flare ribbons cross one another at a point on the surface, namely the X-point. This suggests that the separatrices themselves intersect along a separator which extends all the way down to the solar surface. In this case, at least a portion of the reconnection is taking place at the lower boundary, where it is more readily observed.
To explore this hypothesis, we produced a topological model of the magnetic field using the Magnetic Charge Topology method \citep[MCT;][]{long05}. A section of the 14:34:13 UT line-of-sight HMI magnetogram, shown in the upper left panel of Figure~\ref{fig-model}, was extracted, and its flux partitioned into over 150 distinct polarity regions. These flux regions were replaced by magnetic point sources containing the same net flux and situated at their centroids. This approximation has been found to change the actual connectivity of the extrapolated field by no more than 15\%, but to render the field's topological features clear and easily identified \citep{long09}. From these point sources, we extrapolated a linear force-free field (LFFF) with $\alpha=+3\times10^{-10}\,{\rm cm}^{-1}$, in order to identify the connectivities, and the skeleton defining it. This value of $\alpha$ was chosen to produce field lines with the best overall resemblance to those observed in AIA 171 \AA~images.
The lower panels of Figure \ref{fig-model} show representative field lines of different connectivities from the LFFF extrapolation. Cyan lines show field lines from each of the pre-reconnection domains, while yellow lines show those field lines resulting from reconnection around the X-point. Magenta field lines show connections in other domains not immediately adjacent to the X-point.
Red and blue curves in the lower panels show the surface manifestation of the separatrices whose intersection forms the separator. These same structures are over-plotted, with the same colors, in the upper panels as well. Green and violet curves in those panels show how these separatrices extend along the positive and negative polarities, respectively. The topological model of the separatrices can therefore be seen to follow the flare ribbons, conforming to our basic understanding of flare reconnection. In addition, the HMI magnetogram in Figure \ref{fig-model} shows that, near the X-point, the negative regions are clearly divided to be distinct sources, and the positive regions can be considered to be distinct sources as well, though the division is less prominent. The magnetic field with such a feature is inclined to the formation of a null point (and thus the X-point) and a genuine separatrix instead of a less distinct QSL. The separatrices cross at a photospheric null point located west and south of the observed X-point. Such a geometrical discrepancy is expected in a topological model of this kind, since a MCT model does not accurately model the actual field. We have then performed a LFFF extrapolation from the full-resolution, line-of-sight HMI magnetogram, rather than a point-source approximation of it (i.e.\ MCT), and found a null point just above the lower boundary, situated slightly closer to the X-point. A non-linear force-free field (NLFFF) extrapolation using pre-processed HMI vector data has shown a similar, low-lying null point, still closer to the actual location (X. Sun, private communications). In all cases the skeleton from the null point outlines the same topological structure. We expect the same will be true of the actual magnetic field, of which the MCT, LFFF and NLFFF, are decreasingly crude models.
The actual field, like the LFFF extrapolation, has distinct-connectivity families separated by separatrices whose intersection with the surface appears along the flare ribbons. In the present field, these separatrices intersect at a separator that extends down to the surface, therefore allowing reconnection to occur very near the solar surface. According to this picture, field lines are forged through the process of reconnection --- the post-reconnection flux --- and should be energized in the process. These field lines, shown in yellow and magenta in the lower panels of Figure \ref{fig-model}, match those coronal loops most evident in the hot AIA channels (e.g., 94 \AA) shown in the lower right panel of that figure. These images show little or no evidence of the other family of field lines (cyan lines), presumably because they are not energized by the flare.
In the LFFF model employed here, the separator connecting to the X-point, together with another separator also lying within the fan of the X-point-related null, are found and shown as cyan curves in the top panels of Figure 3. Each of the separators is a single field line along which the separatrix surfaces intersect transversally. In the actual field, pre-reconnection stressing will deform that line into a current-carrying ribbon \citep{long01,long02,long04}. This deformation resembles that observed in 2D models of X-points, which stretch into current sheets under stress \citep{syro71}. In either case, the structure elongates to create a surface across which pre-reconnection domains meet to reconnect. Reconnected flux is then expelled through the sheet, emerging at its tips where post-reconnection flux lies. Following this logic we expect the surface intersection of our separator to be stretched horizontally (i.e.\ east-west). This is, in fact, what SJI 1330 \AA~show during the most intense periods of reconnection.
\section{Conclusion \& Discussions}
\label{discussion}
In this Letter, we report an atypical X-shaped flare ribbon, its evolution, and magnetic topology. These observations present the first evidence, as far as we know, for a special scenario of separator reconnection. Reconstruction of magnetic topology in the active region suggests the presence of separators connecting to the X-point outlined by the ribbons. The inward motion of four flare ribbons as well as converging post-flare loops illustrate sequential 3D reconnection between two sets of non-coplanar loops that approach laterally. It is also evident that initiation of this sequence of 3D reconnection at the X-point is related to the flare development nearby.
The reconnection configuration and evolution in this flare can be sketched in Figure \ref{fig-cartoon}, demonstrating a current sheet that extends from the right of the X-shaped ribbon first horizontally and then vertically into the lower atmosphere at the X-point. Prior to the formation of the X-shaped ribbon, reconnection starts and spreads leftward along the horizontal current, perhaps triggered by a disturbed overlying flux rope\footnote{The flux rope structure is visible in AIA 131 \AA~images (see the animation of Figure \ref{fig-obs}).}, forming the NW and SW ribbons nearly parallel to each other. Reconnection then proceeds downward along the vertical current with two sets of non-coplanar (inflow) loops reconnecting laterally (left panel). The unusual inward motion of the ribbons map the footpoints of outflow loops formed at progressively lower heights. As the flux rope expands (right panel), the reconnection site moves upward, now almost in a direction perpendicular to the horizontal current, producing progressively high-lying post-flare (outflow) loops as well as the apparent outward motion of the ribbons (similar to the standard 2D reconnection scenario). Such a two-stage reconnection evolution is reported for the first time for an X-ribbon flare occurring along a curved separator line in a complex magnetic field.
\acknowledgments
The authors thank the anonymous referee for constructive comments to improve the manuscript. We thank Xudong Sun for helpful discussions including results of his NLFFF modeling. This project was supported by NSFC under grants 10933003, 11373023, and 11403011, and by NKBRSF under grants 2011CB811402 and 2014CB744203. Y.L. is also supported by the Postdoctoral Science Foundations from Jiangsu Province and China Postdoctoral Office, and by the Fundamental Research Funds for the Central Universities. The work at MSU is also supported by the NSF grant 1460059. {\em IRIS} is a NASA small explorer mission. \textit{SDO} is a mission of NASA's Living With a Star Program.
\bibliographystyle{apj}
|
\section{The example}\label{sec:gb}
Seeking a converse to Theorem~\ref{thm:zerotc}, let $J \subseteq I$ be ideals with the same tight closure. In \cite[Proposition 3.1, Theorem 3.4, and Theorem 3.5]{nmeYao-HK}, we gave several criteria under which a converse to Theorem~\ref{thm:zerotc} holds. An analysis of the ideas surrounding
\cite[Proposition 3.1]{nmeYao-HK} yields the following observation: \emph{The critical situation occurs when there exist prime ideals $\ideal{p} \subsetneq \ideal{m}$ such that $\ideal{p}, \ideal{m} \in \Ass_R(I^{[q]}/J^{[q]})$ for infinitely many values of $q$.} One may ask whether this can happen. For instance, in
\cite[Example 2.2]{nmeYao-HK}, the critical situation does not occur for the ideals $J \subseteq I$ in $R$ unless it already was an issue for the ideals $\ideal{b} \subseteq \ideal{a}$ in $A$. Indeed, for each $q$, there is a bijective correspondence between the sets $\Ass_A(\ideal{a}^{[q]}/\ideal{b}^{[q]})$ and $\Ass_R(I^{[q]}/J^{[q]})$, given by $\ideal{p} \mapsto \ideal{p} R$.
However, the situation outlined above can happen, as shown below. Moreover, the expected converse to Theorem~\ref{thm:zerotc} holds, at least in the given example. Note that
the example below does not appear to arise as one of the special cases delineated in \cite[\S 3]{nmeYao-HK}. Therefore, we had to use computational methods.
Before we get to the specific characteristic $p$ situation, we give a somewhat more general construction, which works over any field, and may be of independent interest. As we will be using Gr\"obner basis techniques, we set some notations and recall some facts:
\begin{defn}
Let $A$ be a polynomial ring, $>$ a monomial order, and $f\in A \setminus \{0\}$. The expressions $\lt(f)$ and $\lm(f)$ denote, respectively, the leading term and the leading monomial of $f$ with respect to the given order.
Given two elements $f,g \in A \setminus \{0\}$, the \emph{S-polynomial} of $f$ and $g$ is given by \[
S(f,g) := \frac{\lcm(\lt(f), \lt(g))}{\lt(f)} \cdot f - \frac{\lcm(\lt(f), \lt(g))}{\lt(g)} \cdot g,
\]
where $\lcm$ means the least common multiple.
\end{defn}
The following theorem is a slightly nonstandard (albeit well-established) form of the \emph{Buchberger criterion}:
\begin{thm}\label{thm:Bucrit}\cite[Theorem 2.9.3]{CLO-book1e3}
Let $A$ be a polynomial ring over a field, let $>$ be a monomial order, and let $G = \{g_1, \dotsc, g_n\}$ be a finite subset of $A$. Then $G$ is a Gr\"obner basis if and only if there exist elements $a_{ijk} \in A$ such that for each pair $(j,k)$ with $1\leq j <k \leq n$, we have \[
S(g_j,g_k) = \sum_{i=1}^n a_{ijk} g_i,
\]
in such a way that for each nonzero $a_{ijk}$, we have $\lm(S(g_j, g_k)) \geq \lm (a_{ijk} g_i)$ with respect to the given monomial order.
\end{thm}
\begin{thm}\label{thm:elim} \cite[Theorem 4.3.11 and the discussion which follows]{CLO-book1e3}
Let $A$ be a polynomial ring over a field $k$, let $I$ be an ideal of $A$ and $0 \neq u\in A$. Let $r$ be an indeterminate over $A$, and let $B = A[r]$ be a polynomial ring, ordered with lexicographic order in such a way that $r>x$ for all variables $x$ of $A$. Let $\ideal{a} := rIB + (1-r)uB \subseteq B$. Then $\ideal{a} \cap A = I \cap (u)$, and if $F$ is a Gr\"obner basis of $\ideal{a}$ in $B$, then $\ideal{a} \cap A$ is generated by the set of elements of $F$ whose leading terms are not multiples of $r$.
\end{thm}
\begin{construction}\label{cons}
Let $k$ be an arbitrary field, let $m \in \field{N}$ such that $m\geq 4$, and let $n=2m+1$. We also impose the condition that if $p$ is the characteristic of $k$, then $p \nmid m$, which is automatically satisfied if $p=0$. Let $A := k[s,x,y]$, $\ideal{m} := (s,x,y) \subseteq A$, $g=xy(x-y)(x+y-sy)$, and $\ideal{e} := (x^n, y^n, g) \subseteq A$. Let $f := \sum_{j=2}^{n-1} (-1)^j x^{n+1-j} y^j$. Let $\ideal{h} := \ideal{e} + (f)$. Let $\ideal{b} := (x,y)^{n+2}$. Then we will show the following: \begin{enumerate}
\item $\ideal{b} \subseteq \ideal{e}$,
\item $sf \in \ideal{e}$ (hence, $\ideal{h} \subseteq (\ideal{e} :s)$),
\item $xf, yf \in \ideal{e}$ (hence, $\ideal{m} \subseteq (\ideal{e} :f)$),
\item $f \notin \ideal{e}$ (hence, $(\ideal{e}:f) \neq A$, so that $\ideal{m} = (\ideal{e} :f)$),
\item $\ideal{h}$ is $s$-saturated (that is, $(\ideal{h} : s) = \ideal{h}$),
\item $\ideal{e} : \ideal{m}^\infty = \ideal{e} : s^\infty = \ideal{h}$, and
\item $H^0_\ideal{m}(A/\ideal{e}) \cong A/\ideal{m}$.
\end{enumerate}
To see (1), take a typical monomial generator $x^i y^j$ of $\ideal{b}$. That is, $i+j = n+2$. Since $x^n, y^n \in \ideal{e}$, we may assume that $1\leq j \leq n-1$, so that $i\geq 3$. Note that modulo $g$, we have \[
x^3y \equiv s x^2 y^2 - (s-1) x y^3.
\]
Multiplying this by $x^{i-3} y^{j-1}$, we have $x^i y^j \in (x^{i-1} y^{j+1}, x^{i-2} y^{j+2}, g)$. Then apply induction to obtain $x^i y^j \in (x^2 y^n, x y^{n+1}, g) \subseteq \ideal{e}$.
To see (2), note that modulo $g$, we have \[
s x y^2 (x-y) \equiv xy (x^2 - y^2).
\]
Using this congruence, we have:
\begin{align*}
s(f - xy^n) &= s x y^2(x-y)\left(\sum_{j=0}^{m-1} x^{n-3-2j} y^{2j}\right) \\
&\equiv xy(x^2 -y^2)\left(\sum_{j=0}^{m-1} x^{n-3-2j} y^{2j}\right) \\
&= xy(x^{2m} - y^{2m}) = x^{n} y - x y^{n}.
\end{align*}
Thus, $sf \in (x^n, y^n, g) = \ideal{e}$, as required.
To see (3), let $t=s-1$. Modulo $g$, we have the equivalence \[
tx y^2 (x-y) \equiv x^2 y (x-y).
\]
It follows by induction (on $i$) that for all integers $i\geq 1$, $a\geq 1$, and $b\geq i+1$, we have $t^{i} x^a y^b (x-y) \equiv x^{a+i} y^{b-i} (x-y)$ (modulo $g$). In particular (letting $a=1$ and $b=n$), for all $1\leq i\leq n-1$, we have \[
t^i x y^n(x - y)\equiv x^{i+1}y^{n-i} (x-y).
\]
modulo $g$. Note also that $-x f+ x^{n} y^2 =yf - x^2 y^n = \sum_{j=1}^{m-1} x^{2j+1} y^{n-2j}(x-y)$. But by the above (since $2(m-1) = n-1$), this latter sum is congruent (modulo $g$) to $y^n \cdot \left(\sum_{j=1}^{m-1} t^{2j} x (x-y) \right)$. Thus, $-xf, yf \in (x^n, y^n, g) = \ideal{e}$, as required.
In order to demonstrate (4), we require the introduction of Gr\"obner bases into the discussion. From now on, we will use \emph{lexicographic}\footnote{We emphasize here that we are \emph{not} using degree-lexicographic order. So for instance, in this ordering, we have $s>x^2$. Indeed, $s>x^{200}$.} order, with $s>x>y$. We claim that \[
G := \{g, x^n, x^{n-1} y^3, x^{n-2} y^4, \cdots, x^3 y^{n-1}, y^{n} \}
\] is a Gr\"obner basis of $\ideal{e}$ with respect to lex order. First, since the elements of $G$ consists of the generating set $\{g, x^n, y^n\}$ of $\ideal{e}$ along with some elements of $\ideal{b}$ (an ideal which by (1) is contained in $\ideal{e}$), it follows that $G$ is indeed a generating set for $\ideal{e}$. To show that it is a Gr\"obner basis, we shall find $a_{ijk}$ as in Theorem~\ref{thm:Bucrit}. But since the S-polynomial of a pair of monomials is always 0, we only need to look at the S-polynomials $S(m,g)$ for monomials $m$ of $G$. In the following list, we represent each S-polynomial in two ways. First, we write it in lexicographic order, and then we write it in the form given by Theorem~\ref{thm:Bucrit}: \begin{itemize}
\item $S(y^{n}, g) = -sxy^{n+1} - x^3y^{n-1} + xy^{n+1} = (-sxy)y^{n} - 1(x^3 y^{n-1}) + (xy)y^{n}$.
\item $S(x^3 y^{n-1}, g) = -s x^2 y^{n} - x^4 y^{n-2} + x^2 y^{n} = (-s x^2) y^{n} - 1(x^4 y^{n-2}) + (x^2) y^{n}$.
\item For any $i$ with $4 \leq i \leq n-2$, we have $x^{i-1} y^{n+3-i}, x^{i+1} y^{n+1-i} \in G$. And \begin{align*}
S(x^i y^{n+2-i}, g) &= - s x^{i-1} y^{n+3-i} - x^{i+1} y^{n+1-i} + x^{i-1} y^{n+3-i} \\
&= (-s+1) x^{i-1} y^{n+3-i} + (-1)x^{i+1}y^{n+1-i}.
\end{align*}
\item $S(x^{n-1} y^3, g) = -s x^{n-2}y^4 - x^n y^2 + x^{n-2} y^4 = (-s+1) x^{n-2} y^4 - (y^2) x^{n}$.
\item $S(x^{n}, g) = -s x^{n-1} y^3 - x^{n+1}y + x^{n-1} y^3 = (-s+1) x^{n-1} y^3 + (xy) x^{n}$.
\end{itemize}
Thus, $G$ is a Gr\"obner basis of $\ideal{e}$. The leading term $x^{n-1} y^2$ of $f$ is manifestly not divisible by any of the leading terms of $G$, which means that the output of the division algorithm of $f$ by $G$ is $f$. Thus, $f \notin \ideal{e}$, as required.
To demonstrate (5), we will use Theorem~\ref{thm:elim}. Accordingly, let $B := k[r,s,x,y]$, ordered lexicographically with $r>s>x>y$, and consider the ideal $\ideal{a} := r\ideal{h} B+ (1-r) sB$ of $B$. We claim that the entries of the following vector comprise a Gr\"obner basis of $\ideal{a}$. Note that it ends with all the elements of $s G \cup \{sf\}$ except for $s x^{n-1}y^3$.
\[
\left[ \begin{matrix}
rs-s\\
r x^{n}\\
c := r x^3y - r x y^3 - s x^2 y^2 + s x y^3 \\
d := m r x^2 y^{n-1} + \sum_{j=1}^{n-3} (-1)^{j-1} j s x^{n-1-j} y^{j+2} \\
r y^{n} \\
-sg = s^2 x^2 y^2 - s^2 xy^3 - s x^3y + s x y^3\\
s x^{n} \\
s f = \sum_{j=2}^{n-1} (-1)^j s x^{n+1-j}y^j \\
s x^{n-2} y^4 \\
s x^{n-3} y^5 \\
\vdots \\
s x^3 y^{n-1} \\
s y^{n}
\end{matrix} \right]
\]
(This is a vector of length $n+5$, and we label the elements $F_0$ through $F_{n+4}$.) First we have to show that the ideal generated by the entries of $F$ is exactly $\ideal{a}$. To see that $\ideal{a} \subseteq (F)$, \begin{itemize}
\item $rg = (-x^2 y^2 + x y^3) (rs-s) + 1\cdot c$, and
\item $rf = (x-y) \left(\displaystyle \sum_{j=1}^{m-1} j x^{n-3-2j} y^{2j-1}\right)c + d + mx (-r y^{n} + s y^{n})$.
\end{itemize}
To see that $F \subseteq \ideal{a}$, \begin{itemize}
\item $c = 1\cdot (rg) + (-x^2 y^2 + x y^3) (-rs+s)$,
\item $d =(m-1)(-sx+x)(ry^{n}) + (-x+y)\left(\displaystyle \sum_{j=1}^{m-1} j x^{n-3-2j} y^{2j-1}\right)(rg) + 1\cdot (rf)+ \left(\displaystyle \sum_{j=1}^{n-3} (-1)^{j-1} jx^{n-j-1}y^{j+2}\right)(-rs+s)$,
\end{itemize}
and for each element $u \in G \cup \{f\}$, we have $ru \in \ideal{a}$, so that \begin{itemize}
\item $su = s\cdot (ru) + u \cdot(-rs+s)$.
\end{itemize}
Thus, $\ideal{a} = (F)$.
Taking all the S-polynomials $S_{jk} = S(F_j, F_k)$ such that $j<k$ and $F_j$, $F_k$ are not both monomials (and note that the only non-monomials are $F_i$ for $i=0,2,3,5,7$), we may obtain the following list. For these choices of $a_{ijk}$, the diligent reader may easily verify the conditions of Theorem~\ref{thm:Bucrit}:
\begin{itemize}
\item $S_{01} = -F_6$
\item $S_{02} = x y^3 F_0 + F_5$
\item $S_{03} = ((-x+y)\sum_{j=1}^{m-1} j x^{n-3-2j} y^{2j-1})F_5 - F_7 + (m-1)(sx -x) F_{n+4}$
\item $S_{04} = -F_{n+4}$
\item $S_{05} = (s x y^3 + x^3y-xy^3)F_0 - F_5$
\item $S_{06} = -F_6$
\item $S_{07} = (\sum_{j=3}^{n-1} (-1)^{j-1} x^{n-j+1}y^j) F_0 - F_7$
\item $S_{0i} = -F_i$, for $8\leq i \leq n+4$
\item $S_{12} = (\sum_{j=1}^{m-1} x^{n-3-2j} y^{2j}) F_2 + xF_4 + F_7 - x F_{n+4}$
\item $S_{13} = (\sum_{j=1}^{n-3} (-1)^j j x^{n-3-j} y^{j+2})F_6$
\item $S_{15} = (\sum_{j=1}^{n-2} r x^{n-2-j} y^j)F_5 + (rxy+ry^2)F_6 + (rsxy-rx^2 - rxy)F_{n+4}$
\item $S_{17} = - sx^2 F_4+ryF_7$
\item $S_{23} = -mxyF_4 - y F_7 + (\sum_{j=1}^{n-4} (-1)^{j-1} jF_{j+7}) + ((1-m)x^2 + mxy)F_{n+4}$
\item $S_{24} = -x y^2 F_4 + (-x^2 y + x y^2) F_{n+4}$
\item $S_{25} = (s x^2 y^3 - s x y^4 + x^4 y - x^2 y^3) F_0 + (-sy-x)F_5$
\item $S_{26} = -s(\sum_{j=1}^{m-1} x^{n-3-2j}y^{2j})F_2 - sxF_4 - sF_7 + sxF_{n+4}$
\item $S_{27} = (s x^{n-5} y^2 + (-x+y)\sum_{j=2}^{m-1} jsx^{n-3-2j}y^{2j-1})F_2 - s F_3 + (m-1)(rx-sx)F_{n+4}$
\item $S_{2i} = -sF_{i+1} + (-r+s) F_{i+2}$, for $8\leq i \leq n+1$
\item $S_{2,n+2} = -sF_{n+3} + (-r+s)x^2 F_{n+4}$
\item $S_{2,n+3} = (-rxy-sx^2+sxy)F_{n+4}$
\item $S_{2,n+4} = (-rxy^2 - sx^2y + s xy^2)F_{n+4}$
\item $S_{34} = (\sum_{j=1}^{n-4}(-1)^{j-1} j F_{j+7}) - (n-3) x^2 F_{n+4}$
\item $S_{35} = m(sxy^n + x^3 y^{n-2} - xy^n)F_0 + [s(x-y)(\sum_{j=1}^{m-1} j x^{n-3-2j} y^{2j-1}) - m y^8] F_5 +sF_7 + (m-1) (-s^2 x + s x) F_{n+4}$
\item $S_{36} = (\sum_{j=1}^{n-3} (-1)^{j-1} j x^{n-j-3}y^{j+2}) F_6$
\item $S_{37} = ms(\sum_{j=0}^{n-4} (-1)^j x^{n-2-j}y^j) F_4 + (\sum_{j=1}^{n-4} (-1)^{j-1} jsx^{n-4-j}y^{j+2})F_6 - (n-3)sxy^{n-5} F_8$
\item $S_{3i} = x^{n+3-i} (sy F_7 + s(\sum_{j=1}^{n-4} (-1)^j jF_{j+7}) - sx^2 F_{n+4})$ for $8\leq i\leq n+3$
\item $S_{3,n+4} = (\sum_{j=1}^{n-4} (-1)^{j-1}js F_{j+7}) - (n-3) sx^2 F_{n+4}$
\item $S_{45} = (s^2 xy-sxy) F_4 + r F_{n+3}$
\item $S_{47} = (\sum_{j=1}^{n-3} (-1)^{j-1} s x^{n-1-j} y^j) F_4$
\item $S_{56} = -(\sum_{j=1}^{n-2} x^{n-2-j} y^j)F_5 -(xy+y^2)F_6 + (-sxy+x^2+xy)F_{n+4}$
\item $S_{57} = -(\sum_{j=1}^{m-1} x^{n-3-2j} y^{2j}) F_5 -yF_6+(-sx+x)F_{n+4}$
\item $S_{58} = -yF_7 - F_8 - (s+2)F_9 + (\sum_{j=10}^{n+3} (-1)^{j-1}F_j)+ x^2 F_{n+4}$
\item $S_{5i} = - F_{i-1} + (-s+1)F_{i+1}$ for $9\leq i\leq n+2$
\item $S_{5,n+3} = -F_{n+2}+ (-s+1)x^2 F_{n+4}$
\item $S_{5,n+4} = - F_{n+3} + (-s+1)xyF_{n+4}$
\item $S_{67} = y F_7 - x^2 F_{n+4}$
\item $S_{7i} = xy^{i-8} ((\sum_{j=9}^{n+3} (-1)^j F_j) + (-x^2+xy)F_{n+4})$ for $8\leq i\leq n+3$
\item $S_{7,n+4} = (\sum_{j=1}^{n-3} (-1)^j x^{n-1-j} y^j)F_{n+4}$
\end{itemize}
Hence, the entries of $F$ give a Gr\"obner basis of $\ideal{a}$. By Theorem~\ref{thm:elim}, it follows that the elements of $F$ whose leading term does not involve $r$ forms a generating set for the ideal $\ideal{h} \cap (s)$ of $A$. That is, $\ideal{h} \cap (s) = (s y^{n}, s x^3 y^4 (x,y)^{n-5}, s f, s x^{n}, sg)$. Dividing by $s$, we get $(\ideal{h} : s) = (y^{n}, x^3 y^4(x,y)^{n-5}, f, x^{n}, g) = \ideal{h}$ (since $x^3 y^4 (x,y)^{n-5} \subseteq (x,y)^{n+2} = \ideal{b} \subseteq \ideal{h}$), as required.
To see (6), first note that $\ideal{e} : \ideal{m}^\infty = \ideal{e} : s^\infty$, since $\ideal{e}$ contains powers of both $x$ and $y$. But $\ideal{e} \subseteq \ideal{h}$, so from (2) and (5), we have $\ideal{h} \subseteq (\ideal{e}:s) \subseteq (\ideal{h}:s) = \ideal{h}$, whence all are equalities. Thus, $(\ideal{e} : s^\infty) = (\ideal{h} : s^\infty) = \ideal{h}$, as required.
Finally, to see (7), it follows from (6) and (4) that \[
H^0_\ideal{m}(A/\ideal{e}) = \frac{\ideal{e} : \ideal{m}^\infty}{\ideal{e}} = \frac{\ideal{h}}{\ideal{e}} = \frac{\ideal{e} + (f)}{\ideal{e}} \cong \frac{A}{(\ideal{e} :f)} = A/\ideal{m}.
\]
\end{construction}
\begin{example}
Let $p$ be an odd prime number. Let $k$ be a field of characteristic $p$, and $R := k[s,x,y] / (xy (x-y)(x+y-sy))$. This is the ring used by Katzman in \cite{KaUAss}, with variable change given by $s=t+1$.
Consider the ideals $J := (x^p, y^p)$ and $I=(x,y)^p$ of $R$. As shown in Katzman's paper, $J^* = I$. Now fix a power $q=p^e$ of $p$, $e\geq 1$, and let $n=pq$ in Construction~\ref{cons}. Let $\ideal{b}$, $\ideal{e}$, $A$, $\ideal{m}$, $g$, $\ideal{h}$, and $f$ be as in that construction. The conditions of the construction are satisfied, since $p\geq 3$, whence $n= pq \geq 9$, and $p$ can never divide $(pq-1)/2$. Then $R=A/(g)$ and $J^{[q]} = \ideal{e}/(g) \subseteq R$. In particular, letting $z$ be the image of $f$ in $R$, we have $z\notin J^{[q]}$.
However, we claim that $z\in I^{[q]}$. To see this, it is enough to show (in the ring $R$ -- that is, modulo $g$) that for all $j=2, 3, \dotsc, pq-1$, we have $x^j y^{pq+1-j} \in (x^q, y^q)^p$. For $j=q, q+1$ this is clear, and for $j\geq q+2$, the assertion follows from the equation $x^j y^{pq+1-j} = (t+1) x^{j-1} y^{pq-j+2} - t x^{j-2} y^{pq-j+3}$, along with induction, showing that in these cases, $x^j y^{pq+1-j} \in (x^q y^{pq-q})$. For $2\leq j \leq q-1$, we have \begin{align*}
x^j y^{pq+1-j}
&= x^{q+1} y^{pq-q} - \sum_{i=1}^{j-2} x^{q+i} y^{pq-q-i}(y-x) - x^j y^{pq-q+2-j}(x^{q-1}-y^{q-1}) \\
&= x^{q+1} y^{pq-q} - \sum_{i=1}^{j-2} t^i x^{q} y^{pq-q}(y-x) - t^{j-1}x y^{pq-q+1}(x^{q-1}-y^{q-1}) \\
&= x^{q+1} y^{pq-q} - \sum_{i=1}^{j-2} t^i x^{q} y^{pq-q}(y-x) - t^{j-1}x^q y^{pq-q+1}
+ t^{j-1}x y^{pq} \\
& \in (x^q, y^q)^p = I^{[q]}.
\end{align*}
Let $\ideal{p} := (x,y)$. We claim that $\ideal{p} \in \Ass_R(I^{[q]}/ J^{[q]})$. Since $\ideal{p}$ is minimal over $J^{[q]}$, it suffices to show that $I^{[q]}_\ideal{p}/ J^{[q]}_\ideal{p} = (I^{[q]} / J^{[q]})_\ideal{p} \neq 0$. To do this, it suffices to show that $((x^{pq},y^{pq}) +(g))A_P$ is properly contained in $((x^q, y^q)^p + (g))A_P$, where $P := (x,y) \subseteq A$. But in the ring $C := L[x,y]$ (where $L:=k(s)$, the fraction field of $k[s]$), the ideal $(x^{pq},y^{pq}, g)C$ is primary to $P' = (x,y)C$, which is a maximal ideal of $C$. So to show that $((x^q, y^q)^p + (g))C_{P'} / ((x^{pq},y^{pq}) +(g))C_{P'} = ((x^q, y^q)^p + (g) / (x^{pq}, y^{pq},g))_{P'}$ is nonzero over $A_P = C_{P'}$, it suffices to show that the $C$-module $((x^q, y^q)^p + (g))C / (x^{pq},y^{pq}, g)C \neq 0$. For this, it is enough to show that $x^q y^{(p-1)q} \notin \ideal{c} := (x^{pq},y^{pq}, xy(x-y))C$, since $xy(x-y)$ is a factor of $g$. Suppose that $x^q y^{(p-1)q} \in \ideal{c}$. Then there exist polynomials $a,b,c \in C$ such that \[
x^q y^{(p-1)q} = a x^{pq} + b y^{pq} + cxy(x-y).
\]
From degree considerations (taking the homogeneous degree $pq$-part of the above equation), we may assume that $a,b \in L$. Then making the substitution ($x=1$, $y=0$) in the displayed equation yields $a=0$, whereas the substitution ($x=0$, $y=1$) yields $b=0$. So $x^q y^{(p-1)q} = cxy(x-y)$. But then the substitution $x=y=1$ leads to the conclusion that $1=0$, a manifest contradiction. Hence $x^q y^{(p-1)q} \notin \ideal{c}$, so that $\ideal{p} \in \Ass_R(I^{[q]}/ J^{[q]})$, as required.
We also know from Construction~\ref{cons} that $\ideal{m} = (J^{[q]} : z)$, so that since $z\in I^{[q]}$, we have $\ideal{m} \in \Ass_R(I^{[q]}/J^{[q]})$ as well. So we are in the ``critical situation'' described at the beginning of this note.
Moreover, \[
\len_R(H^0_\ideal{m}(I^{[q]}/J^{[q]})) \leq \len_A(H^0_\ideal{m}(A/\ideal{e})) = \len_A(A/\ideal{m}) = 1,
\]
a constant, which shows that $\rjj{R_\ideal{m}}{J_\ideal{m}}{I_\ideal{m}} = 0$, since $\dim R/J = 1>0$. Hence, the expected converse to Theorem~\ref{thm:zerotc} holds for this specific example.
\end{example}
\section*{Acknowlegements}
We wish to thank Ezra Miller and Kirsten Schmitz for discussions regarding
this note. We used Macaulay 2 \cite{M2hyper} for some of the computations.
\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{#1}}
\newcommand\frH{\mathfrak{H}}
\begin{document}
\title[A Sobolev space theory for time fractional SPDE]{A Sobolev Space theory for stochastic partial differential equations
with time-fractional derivatives}
\author{Ildoo Kim}
\address{Center for Mathematical Challenges, Korea Institute for Advanced Study, 85 Hoegiro Dongdaemun-gu,
Seoul 130-722, Republic of Korea} \email{<EMAIL>}
\thanks{The research of the first author was supported by the TJ Park Science
Fellowship of POSCO TJ Park Foundation}
\author{Kyeong-Hun Kim}
\address{Department of Mathematics, Korea University, 1 Anam-dong,
Sungbuk-gu, Seoul, 136-701, Republic of Korea}
\email{<EMAIL>}
\thanks{This work was supported by Samsung Science and Technology Foundation under Project Number SSTF-BA1401-02}
\author{Sungbin Lim}
\address{Department of Mathematics, Korea University, 1 Anam-dong, Sungbuk-gu, Seoul,
136-701, Republic of Korea} \email{<EMAIL>}
\subjclass[2010]{60H15, 45K05, 35R11}
\keywords{Stochastic partial differential equations, time-fractional evolution equation, Sobolev space theory, space-time white noise}
\maketitle
\begin{abstract}
In this article we present an $L_p$-theory ($p\geq 2$) for the time-fractional quasi-linear
stochastic partial differential equations (SPDEs) of type
$$
\partial^{\alpha}_tu=L(\omega,t,x)u+f(u)+\partial^{\beta}_t \sum_{k=1}^{\infty}\int^t_0 ( \Lambda^k(\omega,t,x)u+g^k(u))dw^k_t,
$$
where $\alpha\in (0,2)$, $\beta <\alpha+\frac{1}{2}$, and $\partial^{\alpha}_t$ and $\partial^{\beta}_t$ denote the Caputo derivative of order $\alpha$ and $\beta$ respectively. The processes $w^k_t$, $k\in \bN=\{1,2,\cdots\}$, are independent one-dimensional Wiener processes defined on a probability space $\Omega$, $L$ is a second order operator of either divergence or non-divergence type, and $\Lambda^k$ are linear operators of order up to two. The coefficients of the equations depend
on $\omega (\in \Omega), t,x$ and are allowed to be discontinuous.
This class of
SPDEs can be used to describe random effects on transport of
particles in medium with thermal memory or particles subject to
sticking and trapping.
\vspace{1mm}
We prove uniqueness and existence results of strong solutions in appropriate Sobolev spaces, and obtain maximal $L_p$-regularity of the solutions.
Our result certainly
includes the $L_p$-theory for SPDEs driven by space-time white noise if the space dimension $d< 4-2(2\beta-1)\alpha^{-1}$. In particular, if $\beta<1/2+\alpha/4$ then we can cover space-time white noise driven SPDEs with space dimension $d=1,2,3$. This is a quite interesting result since in case of the classical SPDE, i.e. $\alpha=\beta=1$, $d$ must be one.
\end{abstract}
\mysection{Introduction}
Fractional calculus and related equations have become important
topics in many science and engineering fields. For instance, they
appear in mathematical modeling \cite{MS, V}, control engineering
\cite{caponetto2010fractional, podlubny1999fractional}, biophysics
\cite{glockle1995fractional, langlands2009fractional},
electromagnetism \cite{engheia1997role, tarasov2006electromagnetic},
polymer science \cite{bagley1983theoretical, metzler1995relaxation},
hydrology \cite{benson2000application, SBMW}, and even finance
\cite{raberto2002waiting, scalas2000fractional}.
While the classical heat equation $\partial_t u=\Delta u$ describes
the heat propagation in homogeneous mediums, the time-fractional
diffusion equation $\partial^\alpha_t u=\Delta u$, $\alpha\in
(0,1)$, can be used to model the anomalous diffusion exhibiting
subdiffusive behavior, due to particle sticking and trapping
phenomena \cite{metzler1999anomalous, metzler2004restaurant}. The
fractional wave equation $\partial^{\alpha}_t u=\Delta u$,
$\alpha\in (1,2)$ governs the propagation of mechanical diffusive
waves in viscoelastic media \cite{mainardi1995fractional}.
The
fractional differential equations have an another important issue in
the probability theory related to non-Markovian diffusion processes
with a memory \cite{metzler2000boundary, MK}.
However, so far, the
study of time-fractional partial differential equations is mainly
restricted to deterministic equations.
For the results on deterministic equations, we refer the reader e.g.
to \cite{SY,Za} ($L_2$-theory),
\cite{zacher2005maximal} ($L_p$-theory),
and \cite{Pr1991,clement1992global,KKL2014} ($L_q(L_p)$-theory).
Also see
\cite{clement2004quasilinear} for $BUC_{1-\beta}([0,T];X)$-type estimates,
\cite{clement2000schauder} for Schauder estimates, \cite{zacher2013giorgi} for DeGirogi-Nash type estimate, and \cite{zacher2010weak} for Harnack inequality.
\vspace{1mm}
In this article we prove existence and uniqueness results in Sobolev spaces for the time-fractional SPDEs of non-divergence type
\begin{align}
\partial_{t}^{\alpha}u & =[a^{ij}u_{x^{i}x^{j}}+b^{i}u_{x^{i}}+cu+f(u)]\nonumber \\
& \qquad+\partial_{t}^{\beta}\int_{0}^{t}[\sigma^{ijk}u_{x^{i}x^{j}}+\mu^{ik}u_{x^{i}}+\nu^{k}u+g^{k}(u)]dw_{s}^{k}
\label{eq:target SPDE non-div}
\end{align}
as well as of divergence type
\begin{align}
\partial_{t}^{\alpha}u & =\left[D_{x^{i}}\left(a^{ij}u_{x^{j}}+b^{i}u+f^{i}(u)\right)+cu+h(u)\right]\nonumber \\
& \qquad+\partial_{t}^{\beta}\int_{0}^{t}[\sigma^{ijk}u_{x^{i}x^{j}}+\mu^{ik}u_{x^{i}}+\nu^{k}u+g^{k}(u)]dw_{s}^{k}.
\label{eq:target SPDE div}
\end{align}
Here, $\alpha$ and $\beta$ are arbitrary constants satisfying
\begin{equation}
\label{beta}
\alpha\in(0,2), \quad \quad \beta\in
(-\infty,\alpha+\frac{1}{2}),
\end{equation}
and the equations are interpreted by their integral forms (see
Definition \ref{defn 1}). The notation $\partial^{\gamma}_t$ denotes the
Caputo derivative and the Riemann-Liouville integral of order
$\gamma$ if $\gamma\geq 0$ and if $\gamma\leq 0$, respectively (see
Section \ref{section2}).
The coefficients
$a^{ij},b^{i},c,\sigma^{ijk},\mu^{ik}$, and $\nu^{k}$ are functions
depending on
$(\omega,t,x)\in\Omega\times[0,\infty]\times\mathbb{R}^{d}$, and the
nonlinear terms $f$, $f^{i}$, $h$, and $g^{k}$ depend on $(\omega,t,x)$ and the
unknown $u$.
The indices $i$ and $j$ go from 1 to $d$ and
$k$ runs through $\{1,2,3,\cdots\}$. Einstein's summation
convention on $i$, $j$, and $k$ is assumed throughout the article. By having infinitely many Wiener processes in the equations, we
cover SPDEs for measure valued processes, for instance, driven by
space-time white noise (see Section \ref{space-time}). Concerning the leading coefficients $a^{ij}$, while in the divergence case we assume they are measurable in $\omega$, piecewise continuous in $t$, and uniformly continuous in $x$, in the non-divergence case we further assume $\sup_{\omega,t}|a^{ij}(t,\cdot)|_{C^{|\gamma|}(\bR^d)}<\infty$ to get $H^{\gamma+2}_p$-valued solutions. Here $\gamma\in \bR$ and $p\geq 2$.
\vspace{1mm}
Our motivation of studying time-fractional SPDEs of type (\ref{eq:target SPDE non-div}) and (\ref{eq:target SPDE div}) naturally arises from the consideration
of the heat equation in a material with thermal memory and random external forces.
For a detailed derivation of these equations, we refer the reader to \cite{CKK}.
If $\alpha=\beta=1$ then (\ref{eq:target SPDE non-div}) and (\ref{eq:target SPDE div}) are classical second-order SPDEs of non-divergence and divergence types. Hence our equations are far-reaching generalizations. The condition $\beta<\alpha+1/2$ in (\ref{beta}) is necessary to make sense of the equations. We refer again to \cite{CKK} for the reason.
\vspace{1mm}
To the best of our knowledge, \cite{desch2008stochastic,desch2013maximal, desch2011p} made first attempt to study the mild solutions to time-fractional SPDEs.
The authors in \cite{desch2008stochastic,desch2013maximal, desch2011p} applied $H^{\infty}$-functional calculus technique to obtain $L_{p}(L_q)$-estimates ($p,q\geq 2$) for the mild solution to
the integral equation of the type
\begin{equation}
\label{eqn h}
U(t)+A \int^t_0k_1(t-s)U(s)ds=\int^t_0k_2(t-s)G(s)dW_s.
\end{equation}
Here, $k_1(t)=\frac{1}{\Gamma(\alpha)}t^{\alpha-1}$, $k_2=\frac{1}{\Gamma(\beta')}t^{\beta'-1}$, and $A$ is the generator of a bounded analytic semigroup on $L_q$ and assumed to admit a bounded
$H^{\infty}$-calculus on $L_q$. Actually, due to Lemma \ref{lem:s int}(iii), (\ref{eqn h}) is equivalent to
\begin{equation}
\label{eqn h2}
\partial^{\alpha}_t U=-AU+\partial^{\beta}_t\int^t_0 G(s)dW_s,
\end{equation}
where $\beta:=1+\alpha-\beta'$. Thus equation (\ref{eqn h2}) is similar to our equations, but it is quite simple compared to ours. For instance, the operator $A$ in (\ref{eqn h}) is independent of $(\omega,t)$, and equation (\ref{eqn h}) contains only a additive noise. The extra condition $\beta>\alpha-1$ is also assumed there.
We also refer to a recent article \cite{CKK}, where an $L_2$-theory for time-fractional SPDEs is presented under the extra condition $\alpha,\beta\in (0,1)$.
The main tool used in \cite{CKK} for the $L_2$-estimate is the Parseval's identity.
\vspace{1mm}
In this article we exploit Krylov's analytic approach to study the strong solutions of much general equations. To obtain $L_p$-estimates of solutions, we control the sharp functions of the solutions in terms of the maximal functions
of free terms $f$, $h$, and $g$, and then apply Hardy-Littlewood theorem and Fefferman-Stein theorem. The lack of integrability of derivatives of kernels related to the representation
of solutions to some model equations causes main difficulties in carrying out this procedure.
We
prove that for any $\gamma\in \bR$ and $p\geq 2$, under a minimal
regularity assumption (depending on $\gamma$) on the coefficients and the nonlinear terms, equation (\ref{eq:target SPDE non-div}) with
zero initial condition has a unique $H^{\gamma+2}_p$-valued
solution, and for this solution the following estimate holds:
\begin{equation}
\label{es}
\|u\|_{\bH^{\gamma+2}_p(T)}\leq N\left(\|f(0)\|_{\bH^{\gamma+2}_p(T)}
+\|g(0)\|_{\bH^{\gamma+c'_0}_p(T, l_2))}\right),
\end{equation}
where $\bH^{\nu}_p(T)=L_p(\Omega\times [0,T]; H^{\nu}_p)$,
$\bH^{\nu}_p(T,l_2)=L_p(\Omega\times [0,T]; H^{\nu}_p(l_2))$ and
$c'_0>\frac{(2\beta-1)_+}{\alpha}=:c_0$ if $\beta = 1/2$, and
$c'_0=c_0$ if $\beta\neq 1/2$. The result for $\gamma\leq 0$ is
needed to handle SPDEs driven by space-time white noise with the space dimension $d< 4-2(2\beta-1)\alpha^{-1}$. For
divergence type equation (\ref{eq:target SPDE div}), we prove uniqueness, existence, and a version of
(\ref{es}) for $\gamma=-1$.
\vspace{1mm}
Our main results, Theorem
\ref{thm:main results non-div} and Theorem \ref{thm:main results
div}, substantially improve the results of
\cite{desch2008stochastic,desch2013maximal, desch2011p} in the
sense that (i) we study the strong solutions, (ii) our coefficients
are much general and allowed to be discontinuous and depend on
$(\omega,t,x)$, (iii) the second and lower order derivatives of
solutions appear in the stochastic part of our equations, (iv)
non-linear terms are also considered, (v) we do not impose the
lower bound of $\beta$ and there is no restriction on $\gamma$, and (vi) we also cover SPDEs driven by space-time white noise with space dimension
$d< 4-2(2\beta-1)\alpha^{-1}$.
\vspace{1mm}
This article is organized as follows. In Section 2 we present some preliminaries
on the fractional calculus and introduce our main results. We prove
a parabolic Littlewood-Paley
inequality for a model time-fractional SPDE in Section 3. The unique solvability and a priori estimate for the model equation are
obtained in Section 4. We prove Theorems \ref{thm:main results non-div}
and \ref{thm:main results div} in Section 5 and 6, respectively. In Section \ref{space-time} we give an application to SPDE driven by space-time white noise.
\vspace{1mm}
Finally we introduce some notation used in this article. We use ``:='' to denote
a definition. As usual, $\mathbb{R}^{d}$ stands for the $d$-dimensional
Euclidean space of points $x=(x_{1},\ldots,x_{d})$, $B_{r}(x):=\{y\in\mathbb{R}^{d}:|x-y|<r\}$,
and $B_{r}:=B_{r}(0)$.
$\mathbb{N}$ denotes the natural number system
and $\mathbb{C}$ indicates the complex number system. For $i=1,\ldots,d$,
multi-indices $\mathfrak{a}=(\mathfrak{a}_{1},\ldots,\mathfrak{a}_{d})$,
$\mathfrak{a}_{i}\in\{0,1,2,\ldots\}$, and functions $u(x)$ we set
\[
u_{x^{i}}=\frac{\partial u}{\partial x^{i}}=D_{^{i}}u,\quad D_{x}^{\mathfrak{a}}u=D_{1}^{\mathfrak{a}_{1}}\cdots D_{d}^{\mathfrak{a}_{d}}u,\quad|\mathfrak{a}|=\mathfrak{a}_{1}+\cdots+\mathfrak{a}_{d}.
\]
We also use the notation $D_{x}^{m}$ for a partial derivative of
order $m$ with respect to $x$.
By $C_{c}^{\infty}(\mathbb{R}^{d};H)$,
we denote the collection of $H$-valued smooth functions having compact support
in $\mathbb{R}^{d}$, where $H$ is a Hilbert space.
In particular, $C_{c}^{\infty}:=C_{c}^{\infty}(\mathbb{R}^{d}; \bR)$. $\cS(\bR^d)$ denotes the Schwartz class on $\bR^d$.
For $p\geq 1$ and a normed space $F$ by $L_{p}(\mathcal{O};F)$ we denote the set
of $F$-valued Lebesgue measurable function $u$ on $\mathcal{O}$ satisfying
\[
\left\Vert u\right\Vert _{L_{p}(\mathcal{O};F)}:=\left(\int_{\mathcal{O}}\|u(x)\|_{F}^{p}dx\right)^{1/p}<\infty.
\]
We write $L_{p}(\mathcal{O})=L_{p}(\mathcal{O};\mathbb{R})$ and $L_{p}=L_{p}(\mathbb{R}^{d})$.
Generally, for a given measure space $(X,\mathcal{M},\mu)$, $L_{p}(X,\cM,\mu;F)$
denotes the space of all $F$-valued $\mathcal{M}^{\mu}$-measurable functions
$u$ so that
\[
\left\Vert u\right\Vert _{L_{p}(X,\cM,\mu;F)}:=\left(\int_{X}\left\Vert u(x)\right\Vert _{F}^{p}\mu(dx)\right)^{1/p}<\infty,
\]
where $\mathcal{M}^{\mu}$ denotes the completion of $\cM$ with respect to the measure $\mu$.
If there is no confusion for the given measure and $\sigma$-algebra, we usually omit the measure and the $\sigma$-algebra. We denote by
\[
\mathcal{F}\left( f\right) (\xi)=\frac{1}{(2\pi)^{d/2}}\int_{\mathbb{R}^{d}}e^{-i\xi\cdot x}f(x)dx,\quad\mathcal{F}^{-1}( g) (x):=\frac{1}{(2\pi)^{d/2}}\int_{\mathbb{R}^{d}}e^{i\xi\cdot x}g(\xi)d\xi,
\]
the Fourier and the inverse Fourier transforms of $f$ in $\mathbb{R}^{d}$ respectively. $\lfloor a\rfloor$
is the greatest integer which is less than or equal to $a$, whereas
$\lceil a\rceil$ denotes the smallest integer which is greater than
or equal to $a$.
$a \wedge b := \min \{a,b\}$, $a \vee b := \max \{a,b\}$, $a_+:=a \vee 0$, and $a_-:= -(a \wedge 0)$.
If we write $N=N(a,b,\cdots)$, this means
that the constant $N$ depends only on $a,b,\cdots$. Throughout
the article, for functions depending on $(\omega,t,x)$, the argument
$\omega \in \Omega$ will be usually omitted.
\mysection{Main Results}
\label{section2}
First we introduce some elementary facts related to the fractional calculus.
We refer the reader to \cite{Po, SKM, baleanu2012fractional,herrmann2011fractional} for more details.
For $\varphi\in L_{1}((0,T))$ and $n=1,2,\cdots$, define $n$-th order integral
$$I^n_t \varphi (t):=\int^t_0 (I^{n-1} \varphi)(s)ds), \quad \quad (I^0_t\varphi:=\varphi).
$$
In general, the Riemann-Liouville fractional integral
of the order $\alpha\geq 0$ is defined as
$$
I_{t}^{\alpha}\varphi:=\frac{1}{\Gamma(\alpha)}\int_{0}^{t}(t-s)^{\alpha-1}\varphi(s)ds, \quad 0\leq t\leq T.
$$
By Jensen's inequality, for $p\in[1,\infty]$,
\begin{equation}
\label{eq:Lp continuity of I}
\left\Vert I_{t}^{\alpha}\varphi\right\Vert _{L_{p}(0,T)}\leq
N(T,\alpha)\left\Vert \varphi\right\Vert _{L_{p}(0,T)}.
\end{equation}
Thus $I^{\alpha}_t\varphi(t)$ is well-defined and finite for almost all $t\leq T$. This inequality shows that if $1\leq p<\infty$ and $\varphi_n \to \varphi$ in $L_p([0,T])$, then $I^{\alpha}\varphi_n$ also converges to $I^{\alpha}_t\varphi$ in $L_p([0,T])$. The inequality for $p=\infty$ implies that if $f_n(\omega,t)$ converges in probability uniformly in
$[0, T ]$ then so does $I^{\alpha}_t f_n$.
Using Fubini's theorem one can easily show for any $\alpha,\beta\geq 0$
\begin{equation}
\label{eqn 4.15.3}
I^{\alpha}_tI^{\beta}_t \varphi=I^{\alpha+\beta}_t \varphi, \quad
\text{$(a.e.)$} \,\, t\leq T.
\end{equation}
It is known that if $p>\frac{1}{\alpha}$ and $\alpha-\frac{1}{p}\notin\mathbb{N}$ then (see \cite[Theorem 3.6]{SKM})
\begin{align}
\label{eq:sobolev imbedding}
\|I_{t}^{\alpha}\varphi\|_{C^{\alpha-\frac{1}{p}}([0,T])}\leq N(p,T,\alpha) \|\varphi\|_{L_{p}(0,T)}.
\end{align}
Let $\alpha\geq 0$, $\varphi \in C^{\alpha}([0,T])$, and $m$ be the maximal integer such that $m<\alpha$.
It is also known that, for any $\beta\geq 0$ (see \cite[Theorem 3.2]{SKM})
\begin{equation}
\label{eqn 8.30-1}
\left\|I_{t}^{\beta}\left(\varphi-\sum_{k=0}^m \frac{\varphi^{(k)}(0)}{k!}t^k\right)
\right\|_{\mathcal{C}^{\alpha+\beta}([0,T])}
\leq
N(\beta)\left\|\varphi-\sum_{k=0}^m \frac{\varphi^{(k)}(0)}{k!}t^k\right\|_{\mathcal{C}^{\alpha}([0,T])}
\end{equation}
if either $\alpha+\beta\notin\mathbb{N}$ or
$\alpha,\beta\in\mathbb{N}\cup\{0\}$.
Next we introduce the fractional derivative $D^{\alpha}_t$, which is (at least formally) the inverse operator of $I^{\alpha}_t$.
Let $\alpha\geq 0$ and $\lfloor \alpha \rfloor =n-1$ for some $n\in \mathbb{N}$. Then obviously
$$n-1\leq \alpha <n, \quad \quad n-\alpha\in (0,1].
$$
For a function $\varphi(t)$ which is $(n-1)$-times differentiable and $\left(\frac{d}{dt}\right)^{n-1} I_t^{n-\alpha} \varphi$ is absolutely continuous on $[0,T]$,
the Riemann-Liouville fractional derivative $D_{t}^{\alpha}$ and the Caputo fractional derivative $\partial_{t}^{\alpha}$ are defined as
\begin{equation}
\label{eqn 4.15}
D_{t}^{\alpha}\varphi:=\left(\frac{d}{dt}\right)^{n}\left(I_{t}^{n-\alpha}\varphi\right),
\end{equation}
and
$$
\partial^{\alpha}_t\varphi:=D^{\alpha-(n-1)}_t\left(\varphi^{(n-1)}(t)-\varphi^{(n-1)}(0)\right).
$$
By \eqref{eqn 4.15.3} and \eqref{eqn 4.15}, for any $\alpha,\beta\geq 0$,
\begin{align}
\label{eqn 4.20.1}
D^{\alpha}_t I_{t}^{\beta} \varphi=\begin{cases}
D_{t}^{\alpha-\beta}\varphi & :\alpha>\beta\\
I_{t}^{\beta-\alpha}\varphi& :\alpha\leq\beta,
\end{cases}
\end{align}
and $D^{\alpha}_tD^{\beta}_t=D^{\alpha+\beta}_t$.
Using \eqref{eqn 4.15.3}-\eqref{eqn 4.20.1}, one can check
\begin{align}
\label{eqn 8.30}
\partial_{t}^{\alpha}\varphi= D_{t}^{\alpha} \left(\varphi(t)-\sum_{k=0}^{n-1}\frac{t^{k}}{k!}\varphi^{(k)}(0)\right).
\end{align}
Thus if
$\varphi(0)=\varphi^{(1)}(0)=\cdots = \varphi^{(n-1)}(0)=0$ then $D^{\alpha}_t\varphi=\partial^{\alpha}_t \varphi$
and by (\ref{eqn 8.30}) and (\ref{eqn 8.30-1}),
\begin{equation}
\label{extra con}
\|\partial^{\beta}_t \varphi\|_{C^{\alpha-\beta}([0,T])}
\leq \|I^{\lfloor \beta\rfloor+1-\beta}_t \varphi\|_{C^{\lfloor \beta\rfloor+1-\beta+\alpha}([0,T]}
\leq N \|\varphi\|_{C^{\alpha }([0,T])} \qquad \forall \beta \leq \alpha,
\end{equation}
where either $\alpha-\beta\notin\mathbb{N}$ or
$\alpha,\beta\in\mathbb{N}\cup\{0\}$.
\begin{remark}
Banach space valued fractional calculus can be defined as above on the basis of Bochner's integral and Pettis's integral.
See e.g. \cite{agarwal2015fractional} and references therein.
\end{remark}
Let $(\Omega,\mathscr{F},P)$ be a complete probability
space and $\{\mathscr{F}_{t},t\geq0\}$ be an increasing filtration of
$\sigma$-fields $\mathscr{F}_{t}\subset\mathscr{F}$, each of which
contains all $(\mathscr{F},P)$-null sets. We assume that
an independent family of one-dimensional
Wiener processes $\{w_{t}^{k}\}_{k\in\mathbb{N}}$ relative to the
filtration $\{\mathscr{F}_{t},t\geq0\}$ is given on $\Omega$.
By $\cP$, we denote the predictable $\sigma$-field generated by $\mathscr{F}_{t}$, i.e.
$\cP$ is the smallest $\sigma$-field containing every set $A \times (s,t]$, where $s<t$ and $A \in \mathscr{F}_s$.
For $p\geq 2$ and $\gamma \in \bR$, let
$H_{p}^{\gamma}=H_{p}^{\gamma}(\mathbb{R}^{d})$ denote the class of all
tempered distributions $u$ on $\mathbb{R}^{d}$ such that
\begin{equation}
\label{eqn norm}
\| u\| _{H_{p}^{\gamma}}:=\|(1-\Delta)^{\gamma/2}u\|_{L_{p}}<\infty,
\end{equation}
where
$$
(1-\Delta)^{\gamma/2} u = \cF^{-1} \left((1+|\xi|^2)^{\gamma/2}\cF (u) \right).
$$
It is well-known that if $\gamma=1,2,\cdots$, then
$$
H^{\gamma}_p=W^{\gamma}_p:=\{u: D^{\mathfrak{a}}_x u\in L_p(\bR^d), \, \,\,|\mathfrak{a}|\leq \gamma\}, \quad \quad H^{-\gamma}_p=(H^{\gamma}_{p/{(p-1}})^*.
$$
For a tempered distribution $u\in H^{\gamma}_p$ and $\phi\in
\cS(\bR^d)$, the action of $u$ on $\phi$ (or the image of $\phi$
under $u$) is defined as
$$(u,\phi)=\left((1-\Delta)^{\gamma/2}u ,
(1-\Delta)^{-\gamma/2}\phi \right)=\int_{\bR^d}
(1-\Delta)^{\gamma/2}u \cdot (1-\Delta)^{-\gamma/2}\phi \,dx.
$$
Let $l_2$ denote the set of all sequences $a=(a^1,a^2,\cdots)$ such that
$$|a|_{l_{2}}:=\left(\sum_{k=1}^{\infty}|a^{k}|^{2}\right)^{1/2}<\infty.
$$
By $H_{p}^{\gamma}(l_{2})=H_{p}^{\gamma}(\bR^d,l_2)$ we denote the class of all $l_2$-valued
tempered distributions $v=(v^1,v^2,\cdots)$ on $\mathbb{R}^{d}$ such that
$$
\|v\|_{H_{p}^{\gamma}(l_{2})}:=\||(1-\Delta)^{\gamma/2}v|_{l_2}\|_{L_{p}}<\infty.
$$
We introduce stochastic Banach spaces:
$$
\mathbb{H}_{p}^{\gamma}(T):=L_{p}\left(\Omega\times[0,T],\mathcal{P};H_{p}^{\gamma}\right),\quad\mathbb{L}_{p}(T)=\mathbb{H}_{p}^{0}(T)
$$
$$
\mathbb{H}_{p}^{\gamma}(T,l_{2}):=L_{p}\left(\Omega\times[0,T],\mathcal{P};H_{p}^{\gamma}(l_{2})\right),\quad\mathbb{L}_{p}(T,l_{2})=\mathbb{H}_{p}^{0}(T,l_{2}).
$$
For instance, $u\in \bH^{\gamma}_p(T)$ if and only if $u$ is an $H^{\gamma}_p$-valued $\cP^{dP\times dt}$ -measurable process defined on $\Omega\times [0,T]$ such that
$$
\|u\|_{\bH^{\gamma}_p(T)}:=\left(\bE \int^T_0 \|u\|^p_{H^{\gamma}_p}dt \right)^{1/p}<\infty.
$$
Here $\cP^{dP \times dt}$ is the completion of $\cP$ w.r.t $dP\times dt$. We write $g\in\mathbb{H}_{0}^{\infty}(T,l_{2})$ if $g^{k}=0$ for
all sufficiently large $k$, and each $g^{k}$ is of the type
$$
g^{k}(t,x)=\sum_{i=1}^{n}1_{(\tau_{i-1},\tau_{i}]}(t)g^{ik}(x),
$$
where $\tau_{i}\leq T$ are stopping times with repect to $\mathscr{F}_{t}$
and $g^{ik}\in C_{c}^{\infty}(\mathbb{R}^{d})$. It is known \cite[Theorem 3.10]{Krylov1999}
that $\mathbb{H}_{0}^{\infty}(T,l_{2})$ is dense in $\mathbb{H}_{p}^{\gamma}(T,l_{2})$
for any $\gamma$.
We use $U_{p}^{\alpha,\gamma}$ to denote the family of $H_{p}^{\gamma+(2 -2/(\alpha p))_+}$-valued
$\mathscr{F}_{0}$-measurable random variables $u_{0}$ such that
$$
\|u_{0}\|_{U_{p}^{\alpha, \gamma}}:=\left(\mathbb{E}\|u_{0}\|_{H_{p}^{\gamma+(2 -2/(\alpha p))_+}}^{p}\right)^{1/p}<\infty,
$$
where $(2 -2/(\alpha p))_+ = \frac{|2 -2/(\alpha p)|+2 -2/(\alpha p)}{2}$.
\vspace{2mm}
(i) and (iii) of Lemma \ref{lem:s int} below are used e.g. when we apply $I^{\alpha}_t$ and $D^{\alpha}_t$ to the time-fractional SPDEs, and (ii) can be used in the approximation arguments.
\begin{lem}
\label{lem:s int}
\noindent
(i) Let $\alpha\geq 0$ and $h\in L_{2}(\Omega\times[0,T],\mathcal{P};l_{2})$.
Then the equality
\begin{equation}
I^{\alpha}\left(\sum_{k=1}^{\infty}\int_{0}^{\cdot}h^{k}(s)dw_{s}^{k}\right)(t)=\sum_{k=1}^{\infty}\left(I^{\alpha}\int_{0}^{\cdot}h^{k}(s)dw_{s}^{k}\right)(t)\label{eq:equality s int}
\end{equation}
holds for all $t\leq T$ $(a.s.)$ and also in $L_{2}(\Omega\times[0,T])$, where the convergence of the series in both sides is understood in probability sense.
\noindent
(ii) Suppose $\alpha\geq 0$ and $h_{n} \to h$ in $L_{2}(\Omega\times[0,T],\mathcal{P} ; l_{2})$
as $n\rightarrow\infty$. Then
$$
\sum_{k=1}^{\infty}\left(I^{\alpha}\int_{0}^{\cdot}h_{n}^{k}dw_{s}^{k}\right)(t) \longrightarrow\sum_{k=1}^{\infty}\left(I^{\alpha}\int_{0}^{\cdot}h^{k}dw_{s}^{k}\right)(t)
$$
in probability uniformly on $[0,T]$.
\noindent
(iii) If $\alpha>1/2$ and $L_{2}(\Omega\times[0,T],\mathcal{P} ; l_{2})$, then
\begin{align*}
\frac{\partial}{\partial t}\left(I^{\alpha}\sum_{k=1}^{\infty}\int_{0}^{\cdot}h^{k}(s)dw_{s}^{k}\right)(t)
&=\frac{1}{\Gamma(\alpha)}\sum_{k=1}^{\infty}\int_{0}^{t}(t-s)^{\alpha-1}h^{k}(s)dw_{s}^{k}
\end{align*}
$(a.e.)$ on $\Omega \times [0,T]$.
\end{lem}
\begin{proof}
See Lemmas 3.1 and 3.3 of \cite{CKK}.
\end{proof}
\begin{rem}
\label{rem:s int}
By \cite[Remark 3.2]{Krylov1999}, for any $g\in \mathbb{H}^{\gamma}_p(T,l_2)$ and $\phi\in C^{\infty}_c(\fR^d)$
\begin{equation}
\label{eqn 5.21.1}
\bE \left[\sum_k \int^T_0 (g^k,\phi)^2 ds \right]\leq N(p,\phi) \|g\|^{2}_{\bH^{\gamma}_p(T,l_2)}.
\end{equation}
Thus if $g_{n}\rightarrow g$
in $\mathbb{H}_{p}^{\gamma}(T,l_{2})$, then $(g_n,\phi)\to (g,\phi)$ in $L_{2}(\Omega\times[0,T],\mathcal{P};l_{2})$.
Therefore, one can apply Lemma \ref{lem:s int} $(ii)$ with $h_{n}(t)=\left(g_{n}(t,\cdot),\phi\right)$
and $h(t)=\left(g(t,\cdot),\phi\right)$.
\end{rem}
\vspace{3mm}
Let $\alpha\in(0,2)$, $\beta< \alpha+\frac{1}{2} $ and set
$$
\Lambda:=\max(\lceil \alpha \rceil , \lceil \beta \rceil).
$$
\begin{defn}
\label{def:solution space}
Define
$$
\cH^{\gamma}_p(T):=\bH^{\gamma+2}_p(T) \cap \left\{u: I^{\Lambda-\alpha}u \in L_p(\Omega;C([0,T]; H^{\gamma}_p))\right\},
$$
that is,
$u \in \mathcal{H}_{p}^{\gamma+2}(T)$ iff $u\in\mathbb{H}_{p}^{\gamma+2}(T)$ and $I^{\Lambda-\alpha}u$ has a $H^\gamma_p$-valued continuous version $\bI^{\Lambda-\alpha}_tu$.
The norm in $\cH^{\gamma+2}_p(T)$ is defined as
\begin{align*}
\|u\|_{\mathcal{H}_{p}^{\gamma+2}(T)}
&:=\|u\|_{\mathbb{H}_{p}^{\gamma+2}(T)}+\left(\mathbb{E}\sup_{t\leq T} \|\bI^{\Lambda-\alpha}u(t,\cdot)\|_{H_{p}^{\gamma}}^{p}\right)^{1/p}.
\end{align*}
\end{defn}
\begin{defn}
\label{defn 1}
Let $u \in \mathcal{H}_{p}^{\gamma_1+2}(T)$, $f\in\mathbb{H}_{p}^{\gamma_2}(T)$,
$g\in\mathbb{H}_{p}^{\gamma_3}(T,l_{2})$, $u_0\in U_{p}^{\alpha,
\gamma_4}$, and
$v_0\in U_{p}^{\alpha-1,\gamma_4}$
for some $\gamma_i\in \bR$ ($i=1,2,3,4$).
We say that $u$ satisfies
\begin{equation}
\label{eqn 7.15}
\partial_{t}^{\alpha}u(t,x)=f(t,x)+\partial_{t}^{\beta}\int_{0}^{t}g^{k}(s,x)dw_{s}^{k}, \quad t\in (0,T],
\end{equation}
$$
u(0,\cdot)=u_0, \quad \quad \partial_tu(0,\cdot)=v_0~(\text{if}~ \alpha >1)
$$
if
for any $\phi\in \cS(\bR^d)$ the equality
\begin{eqnarray}
\nonumber
&&(\bI^{\Lambda-\alpha}_tu(t)-I^{\Lambda-\alpha}_t( u_0+tv_01_{\alpha>1}), \phi) \\
&&=I_{t}^{\Lambda}\left(f(t,\cdot),\phi\right)+\sum_{k=1}^{\infty}I_{t}^{\Lambda-\beta}\int_{0}^{t}\left(g^{k}(s,\cdot),\phi\right)dw_{s}^{k}
\label{eq:solution space 2}
\end{eqnarray}
holds for all $t\in[0,T]$ $(a.s.)$ (see Remark \ref{rem 7.15} for an equivalent version of (\ref{eq:solution space 2})).
In this case we say (\ref{eqn 7.15}) holds in the sense of distributions.
We say $u$ (or (\ref{eqn 7.15})) has zero initial condition
if (\ref{eq:solution space 2}) holds with $u_0=v_0=0$.
\end{defn}
Below we discuss how the space $U^{\alpha,\gamma}_p$ is chosen and show why (\ref{eq:solution space 2}) is an appropriate interpretation of (\ref{eqn 7.15}).
\begin{rem}
\label{remark 8.17}
In this article we always assume $u(0)=1_{\alpha>1}\partial_t u(0)=0$. The space $U^{\alpha,\gamma}_p$ is defined for later use. It turns out that for the solution to the equation
$$
\partial^{\alpha}_tu=\Delta u, \quad t>0 \, ; \quad u(0,\cdot)=u_0, \quad 1_{\alpha>1}\partial_t u(0,\cdot)=1_{\alpha >1} v_0,
$$
we have, for any $\gamma \in \bR$ and $\kappa>0$,
$$
\|u\|_{L_p((0,T),H^{\gamma+2}_p)}\leq N \left(\|u_0\|_{U^{\alpha,\gamma'}_p}+1_{\alpha >1}\|v_0\|_{U^{\alpha-1,\gamma'}_p}\right),
$$
where $\gamma'=\gamma+\kappa 1_{\beta=1/2}$.
\end{rem}
\begin{rem}
If $\alpha=\beta=1$ then $\Lambda=1$ and (\ref{eq:solution space 2}) coincides with classical definition of the weak solution \cite[Definition 3.1]{Krylov1999}.
\end{rem}
\begin{rem}
\label{rem 7.15}
(i) Let $u$, $f,g,u_0$, and $v_0$ be given as in Definition \ref{defn 1}. We claim that (\ref{eq:solution space 2}) holds for all $t\leq T$ (a.s.) if and only if the equality
\begin{equation}
\label{eqn 7.15.2}
\left(u(t)-u_0-t\partial_tv_01_{\alpha >1},\phi\right) =I_{t}^{\alpha}\left(f(t),\phi\right)+\sum_{k=1}^{\infty}I_{t}^{\alpha-\beta}\int_{0}^{t}\left(g^{k}(s),\phi\right)dw_{s}^{k}
\end{equation}
holds for almost all $t\leq T$ (a.s.). Indeed, applying $D^{\Lambda-\alpha}_t$ to (\ref{eq:solution space 2}) and using (\ref{eqn 4.20.1}),
we get equality (\ref{eqn 7.15.2}) for almost all $t\leq T$ $(a.s.)$. Here $I^{\alpha-\beta}_t:=D^{\beta-\alpha}_t$ if $\alpha\leq \beta$.
Note that if $\alpha\leq \beta$, the last term of (\ref{eqn 7.15.2}) makes sense due to Lemma \ref{lem:s int}(iii) and the assumption $\beta-\alpha<1/2$.
For the other direction, we apply $I_t^{\Lambda-\alpha}$ to (\ref{eqn 7.15.2}) and get
(\ref{eq:solution space 2}) for all $t\leq T$ $(a.s.)$. This is because $(\bI^{\Lambda-\alpha}_tu,\phi)$ is continuous in $t$ by the assumption $u\in \cH^{\gamma_1+2}_p(T)$.
Also, taking $D^{\alpha}_t$ to (\ref{eqn 7.15.2}), we formally get a distributional version of (\ref{eqn 7.15}):
$$
(\partial^{\alpha}_t u,\phi)=(f(t),\phi)+\partial^{\beta}_t\int^t_0 (g^k,\phi)dw^k_t, \quad \text{$(a.e.)$}\, t\leq T.
$$
(ii) Let $\beta<1/2$ and $u(0)=1_{\alpha>1}u'(0)=0$. Denote
$$
\bar{f}(t)=\frac{1}{\Gamma(1-\beta)}\sum_k \int^t_0(t-s)^{-\beta}g^k(s)dw^k_s.
$$
Then from (\ref{eqn 7.15.2}) and Lemma \ref{lem:s int}(iii) it follows that the equality
$$
\left(u(t),\phi\right) =I_{t}^{\alpha}\left(f(t)+\bar{f}(t),\phi\right)
$$
holds for almost all $t\leq T$ (a.s.). Therefore (\ref{eq:solution space 2}) holds for all $t\leq T$ (a.s.) with $f+\bar{f}$ and $0$ in place of $f$ and $g$, respectively.
\end{rem}
To use some deterministic results later in this article we show our intepretation of (\ref{eqn 7.15}) coincides with the one in \cite{KKL2014, zacher2005maximal, Za}.
In the following remark $u$ is not random and $\gamma_1=\gamma_2=\gamma$.
\begin{rem}
\label{rem deterministic}
Denote $\fH^{\gamma+2}_p(T)=L_p([0,T] ; H^{\gamma+2}_p)$ and $\fL_p(T)=\fH^{0}_p(T)$.
We dnoete by $\fH^{\alpha,\gamma+2}_{p,0}(T)$ the completion of
$ C^{\infty}_c((0,\infty)\times \bR^d) $
with the norm
$$\|\cdot\|_{\fH^{\alpha,\gamma+2}_p(T)}:=\|\cdot\|_{\fH^{\gamma+2}_p(T)}+\|\partial^{\alpha}_t \cdot\|_{\fH^{\gamma}_p(T)}.
$$
That is, $u\in \fH^{\alpha,\gamma+2}_{p,0}(T)$ if and only if there exists a sequence $u_n \in C^{\infty}_c((0,\infty)\times \bR^d)$ such that
$\|u_n -u\|_{\fH^{\gamma+2}_p(T)}\to 0$ and
$f_n:=\partial^{\alpha}_tu_n$ is a Cauchy sequence in $\fH^{\gamma}_p(T)$, whose limit is defined as $\partial^{\alpha}_t u$.
\vspace{2mm}
The following two statements are equivalent:
\vspace{1mm}
$\bullet$ $u\in \fH^{\alpha,\gamma+2}_{p,0}(T)$ and $\partial^{\alpha}_tu=f$ in $\fH^{\gamma}_p(T)$.
\vspace{3mm}
$\bullet$ $u\in \cH^{\gamma+2}_{p}(T)$, $f\in \bH^{\gamma}_p(T)$, and $u$ satisfies $\partial^{\alpha}_tu=f$ with zero initial condition
in the sense of Definition \ref{def:solution space}.
\vspace{2mm}
First, let $u\in \fH^{\alpha,\gamma+2}_{p,0}(T)$ and $\partial^{\alpha}_tu=f$ in $\fH^{\gamma}_p(T)$. Take $u_n$ and $f_n$ as above.
Then since $u_n, f_n \in C([0,T];H^{\gamma}_p)$, we have
\begin{equation*}
u_n(t)=I_{t}^{\alpha}f_n(t), \quad \forall \,t\leq T,
\end{equation*}
and letting $n\to \infty$ we conclude
\begin{equation}
\label{det sol rel}
u(t)=I_{t}^{\alpha}f(t), \quad (a.e.)\, \, t\leq T.
\end{equation}
Taking $I^{\Lambda -\alpha}$ to both sides of (\ref{det sol rel}) and recalling $\Lambda \geq 1$, one easily finds that $I^{\Lambda - \alpha}u$ has an $H^{\gamma}_p$-valued continuous version.
Therefore, by Remark \ref{rem 7.15}, $u\in \cH^{\gamma+2}_{p}(T)$ and it satisfies $\partial^{\alpha}_tu=f$ with the zero initial condition in the sense of Definition of \ref{def:solution space}.
Next, let $u\in \cH^{\gamma+2}_{p}(T)$ satisfy $\partial^{\alpha}_tu=f$ in the sense of Definition of \ref{def:solution space}
with zero initial condition.
Then by (\ref{eqn 7.15.2}),
$$
u(t)=I^{\alpha}_t f(t) \quad \text{in}\quad H^{\gamma}_p(\bR^d), \quad \text{$(a.e.)$}~\,t\in [0,T].
$$
Extend $u$ so that $u(t)=0$ for $t<0$. Take $\eta\in C^{\infty}_c((1,2))$ with the unit integral, and denote $\eta_{\varepsilon}(t)=\varepsilon^{-1}\eta(t/{\varepsilon})$,
$$
u^{\varepsilon}(t):=u \star \eta_{\varepsilon}(t) := \int_{\bR} u(s)\eta_\varepsilon(t-s)ds=\int^t_0 u(s)\eta_{\varepsilon}(t-s)ds,
$$
and $f^{\varepsilon}:=f \star \eta_{\varepsilon}$.
Note $u^{\varepsilon}(t)=0$ for $t<\varepsilon$, and thus $u^{\varepsilon}\in C^n([0,T];H^{\gamma}_p)$ for any $n$.
Multiplying by a smooth function which equals one for $t\leq T$ and vanishes for $t>T+1$, we may assume $u^{\varepsilon}\in C^{\infty}_c((0,\infty);H^{\gamma}_p)$. Obviously
$\partial^{\alpha}_tu^{\varepsilon}=f^{\varepsilon}$ in $\fH^{\gamma}_p(T)$,
$\|u^{\varepsilon}-u\|_{\fH^{\gamma+2}_p(T)}\to 0$ and $\|f^{\varepsilon}-f\|_{\fH^{\gamma}_p(T)}\to 0$ as $\varepsilon \downarrow 0$.
Next choose a smooth function $\zeta(x)\in C^{\infty}_c(B_1(0))$ with unit integral, and denote $u^{\varepsilon,\delta}(t,x)=u^{\varepsilon}* \delta^{-d}\zeta(\cdot/{\delta})=\delta^{-d}\int_{\bR^d}u^{\varepsilon}(t,y)\zeta((x-y)/{\delta})dy$ and define $f^{\varepsilon,\delta}$ similarly. Then we still have $\partial^{\alpha}_tu^{\varepsilon,\delta}=f^{\varepsilon,\delta}$.
For any $\varepsilon'>0$, choose $\varepsilon$ and $\delta$ so that
$\|u^{\varepsilon,\delta}-u^{\varepsilon}\|_{\fH^{\gamma+2}_p(T)}+
\|\partial^{\alpha}_t(u^{\varepsilon,\delta}-u^{\varepsilon})\|_{\fH^{\gamma}_p(T)}\leq \varepsilon'$.
After this, multiplying by appropriate smooth cut-off functions of $x$, we can approximate $u^{\varepsilon,\delta}$ and $f^{\varepsilon,\delta}$
with functions in
$C^{\infty}_c((0,\infty)\times \bR^d)$, and therefore we may assume
$u^{\varepsilon,\delta}, f^{\varepsilon,\delta} \in C^{\infty}_c((0,\infty)\times \bR^d)$. Thus it follows that
$u\in \fH^{\alpha,\gamma+2}_{p,0}(T)$ and it satisfies $\partial^{\alpha}_tu=f$ as the limit in $\fH^{\gamma}_p(T)$.
\end{rem}
\begin{thm}
\label{lem:solution space}
(i) For any $\gamma, \nu\in \bR$, the map $(1-\Delta)^{\nu/2}:\cH^{\gamma+2}_p \to \cH^{\gamma-\nu+2}_p(T)$ is an isometry.
\vspace{1mm}
\noindent
(ii) Let $u \in \cH_p^{\gamma+2}(T)$ satisfy (\ref{eqn 7.15}).
Then
\begin{align}
\notag
\mathbb{E}\sup_{t\leq T} \|\bI^{\Lambda-\alpha}u(t,\cdot)\|_{H_{p}^{\gamma}}^{p}
&\leq N\big(\bE \|u(0)\|^p_{H^{\gamma}_{p}}+1_{\alpha >1}\bE \|\partial_t u(0)\|^p_{H^{\gamma}_p}\\
\label{eq:7.2.1}
&\quad\quad+\|f\|_{\mathbb{H}_{p}^{\gamma}(T)} +\|g\|_{\mathbb{H}_{p}^{\gamma}(T,l_{2})} \big),
\end{align}
where $N=N(d,p,T)$.
\vspace{1mm}
\noindent
(iii)
$\mathcal{H}_{p}^{\gamma+2}(T)$ is a Banach space,
\noindent
(iv) Let $\theta:=\min\{1,\alpha,2(\alpha-\beta)+1\}$. Then
there exists a constant $N=N(d,\alpha,\beta,p,T)$ so that for any
$t\leq T$ and $u\in\mathcal{H}_{p}^{\gamma+2}(T)$ satisfying (\ref{eqn 7.15}) with the zero initial condition,
\begin{equation}
\left\Vert u\right\Vert _{\mathbb{H}_{p}^{\gamma}(t)}^{p}
\leq N\int_{0}^{t}(t-s)^{\theta-1} \left(\|f\|^p_{\mathbb{H}_{p}^{\gamma}(s)}
+\|g\|^p_{\mathbb{H}_{p}^{\gamma}(s,l_{2})} \right)ds.
\label{eq:solution space estimate 1}
\end{equation}
\end{thm}
\begin{proof}
(i) For any $u \in \cH_p^{\gamma+2}(T)$,
$(1-\Delta)^{\nu/2}\bI^{\Lambda - \alpha}_tu$
is an $H_p^{\gamma-\nu+2}$-valued continuos version of $(1-\Delta)^{\nu/2}I_t^{\Lambda -\alpha}u$.
Thus it is obvious.
(ii) Due to (i), we may assume that $\gamma=0$. Take a nonnegative
function $\zeta\in C_{c}^{\infty}(\bR^d)$ with unit integral. For
$\varepsilon>0$ define
$\zeta_{\varepsilon}(x)=\varepsilon^{-d}\zeta(x/\varepsilon)$, and
for tempered distributions $v$ on $\bR^d$ put
$v^{(\varepsilon)}(x):=v\ast\zeta_{\varepsilon}(x)$.
Note that for each $t \in (0,T)$, $u^{(\varepsilon)}(t,x)$ is an
infinitely differentiable function of $x$.
By plugging
$\zeta_{\varepsilon}(\cdot-x)$ in (\ref{eq:solution space 2}) in place of $\phi$, for any $x$
\begin{align}
\label{eq:7.3.1}
(\bI^{\Lambda-\alpha}u)^{(\varepsilon)} (t,x)
= I_{t}^{\Lambda}f^{(\varepsilon)}(t,x)
+I_{t}^{\Lambda-\beta}\int_{0}^{t}g^{(\varepsilon)k}(s,x)dw_{s}^{k},
\quad \forall \,t\leq T\, \text{$(a.s.)$}.
\end{align}
Observe that
\begin{align}
\label{b eqn 1}
\mathbb{E}\sup_{t\leq T}
\left\| I_{t}^{\Lambda}f^{(\varepsilon)} (t,\cdot) \right\|_{p}^{p}
\leq N\mathbb{E}\int_{0}^{T}\|f^{(\varepsilon)}(s,\cdot)\|_{p}^{p}\,ds.
\end{align}
Also,
by (\ref{eq:Lp continuity of I}), the Burkholder-Davis-Gundy inequality, and the H\"older inequality,
\begin{align}
\notag
\mathbb{E}\sup_{t\leq T}\left\| I_{t}^{\Lambda-\beta} \sum_k \int_{0}^{t}g^{(\varepsilon)k}(s,\cdot)dw_{s}^{k} \right\|_{p}^{p}
& \leq N \int_{\bR^d} \mathbb{E} \sup_{t\leq T} \left| \sum_k \int_{0}^{t}g^{(\varepsilon)k}(s,x)dw_{s}^{k}\right|^p dx \\
\label{b eqn 2}
& \leq N \mathbb{E} \int_{0}^{T} \|g^{(\varepsilon)}(s,\cdot)\|_{L_{p}(l_{2})}^{p}ds.
\end{align}
Thus from \eqref{eq:7.3.1},
\begin{align}
\notag
\mathbb{E}\sup_{t\leq T} \left\|(\bI_{t}^{\Lambda-\alpha}u)^{(\varepsilon)}(t,\cdot) \right\|_{p}^{p} &\leq N(\|f^{(\varepsilon)}\|^p_{\bL_p(T)}+
\|g^{(\varepsilon)}\|^p_{\bL_p(T,l_2)} ) \notag \\
&\leq N (\|f\|^p_{\bL_p(T)}+
\|g\|^p_{\bL_p(T,l_2)} ).
\label{eq:7.3.2}
\end{align}
By considering $(\bI_{t}^{\Lambda-\alpha}u)^{(\varepsilon)}-(\bI_{t}^{\Lambda-\alpha}u)^{(\varepsilon')} $ instead of $(\bI_{t}^{\Lambda-\alpha}u)^{(\varepsilon)}$,
we easily see that $(\bI_{t}^{\Lambda-\alpha}u)^{(\varepsilon)}$ is a Cauchy sequence in $L_{p}(\Omega ; C([0,T];L_{p}))$.
Let $\bar{u}$ be the limit in this space.
Then since $(\bI^{\Lambda-\alpha}_tu)^{(\varepsilon)}$ converges to $\bI^{\Lambda-\alpha}u$ in $\bL_p(T)$, we conclude $\bar{u}=\bI^{\Lambda-\alpha}u$, and
get (\ref{eq:7.2.1}) by considering the limit of (\ref{eq:7.3.2}) as $\varepsilon \to 0$ in the space $L_{p}(\Omega ; C([0,T];L_{p}))$.
(iii) By (\ref{eq:Lp continuity of I}),
$I_t^{\Lambda - \alpha}u_n$ converges to $I_t^{\Lambda - \alpha}u$ in $\bH_p^{\gamma+2}(T)$
if $u_n$ converges to $u$ in $\bH_p^{\gamma+2}(T)$.
Moreover, both $\bH_p^{\gamma+2}(T)$ and $L_{p}(\Omega ; C([0,T];H^\gamma_{p}))$ are Banach spaces.
Therefore, $\cH_p^{\gamma+2}(T)$ is a Banach space.
(iv) As in the proof of (ii), we only consider the case $\gamma=0$. By (\ref{eqn 7.15.2}), for each $x \in \bR^d$, $(a.s.)$
$$
u^{(\varepsilon)}(t,x)=I^{\alpha}_t f^{(\varepsilon)}(t,x)+I^{\alpha-\beta}_t\int^t_0 g^{(\varepsilon)k}(s,x) dw^k_s \quad (a.e.)~ t \in [0,T].
$$
Note
\begin{align*}
\|I^{\alpha}_t f^{(\varepsilon)}\|_{\mathbb{L}_{p}(t)}^{p}
\leq N I_{t}^{\alpha}\|f^{(\varepsilon)}\|_{\mathbb{L}_{p}(\cdot)}^{p} (t)
\leq N I_{t}^{\alpha}\|f\|_{\mathbb{L}_{p}(\cdot)}^{p}(t) \quad \forall \,t\in[0,T].
\end{align*}
By Lemma \ref{lem:s int} and the stochastic Fubini theorem (note if $\alpha < \beta$ then we define $I^{\alpha-\beta}_t=\frac{\partial}{\partial t}I_t^{\alpha+1-\beta}$), for each $x$ $(a.s.)$
$$
v^{\varepsilon}(t,x):=I^{\alpha-\beta}_t\int^t_0 g^{(\varepsilon)k} (s,x) dw^k_s
=c(\alpha,\beta)\int^t_0(t-s)^{\alpha-\beta}g^{(\varepsilon)k}(s.x) dw^k_s
$$
for almost all $t \in [0,T]$.
Thus by the Burkholder-Davis-Gundy inequality and
the H\"older inequality, for any $t\leq T$,
\begin{align*}
\|v^\varepsilon\|_{\mathbb{L}_{p}(t)}^{p}
&\leq N\mathbb{E}\int_{0}^{t}\int_{\mathbb{R}^{d}}\left(I_{s}^{2(\alpha-\beta)+1}\left(|g^{(\varepsilon)}|_{l_{2}}^{2}(\cdot,x)\right)(s) \right)^{p/2}dxds\\
& \leq NI_{t}^{2(\alpha-\beta)+1}\left(\|g\|_{\mathbb{L}_{p}(\cdot,l_{2})}^{p}\right)(t).
\end{align*}
Observe that for $s\leq t\leq T$,
$$
(t-s)^{\alpha-1}+(t-s)^{2(\alpha-\beta)}\leq N(t-s)^{\theta-1}
$$
where $N$ depends on $\alpha,\beta$ and $T$. Thus, for any $t\leq T$
\begin{align*}
\|u^{(\varepsilon)}\|_{\mathbb{L}_{p}(t)}^{p}
& \leq NI_{t}^{\alpha}\left(\|f\|_{\mathbb{L}_{p}(\cdot)}^{p}\right)(t)+NI_{t}^{2(\alpha-\beta)+1}\left(\|g\|_{\mathbb{L}_{p}(\cdot,l_{2})}^{p}\right)(t)\\
& \leq NI_{t}^{\theta}\left(\|f\|_{\mathbb{L}_{p}(\cdot)}^{p}+\|g\|_{\mathbb{L}_{p}(\cdot,l_{2})}^{p}\right)(t).
\end{align*}
The claim of (iv) follows from Fatou's lemma.
\end{proof}
Assumption \ref{assu:common} below will be used for both divergence type and non-divergence type equations. As mentioned before, the argument $\omega$ is omitted for functions depending on $(\omega,t,x)$.
\begin{assumption}
\label{assu:common}
$(i)$ The coefficients $a^{ij}$, $b^{i}$, $c$, $\sigma^{ijk}$, $\mu^{ik}$, $\nu^{k}$ are
$\mathcal{P}\otimes\mathcal{B}(\mathbb{R}^{d})$-measurable.
\noindent $(ii)$ The leading coefficients $a^{ij}$ are continuous in
$x$ and piecewise continuous in $t$ in the following sense: there
exist stopping times $0=\tau_{0}<\tau_{1}<\tau_{2}<\cdots<\tau_{M_0}
= T$ such that
\begin{equation}
\label{cond 8.26}
a^{ij}(t,x)=\sum_{n=1}^{M_0}
a_{n}^{ij}(t,x)1_{(\tau_{n-1},\tau_n]}(t),
\end{equation}
where each $a_{n}^{ij}$ are uniformly continuous in $(t,x)$, that is for any $\varepsilon>0$,
there exists a $\delta >0$ such that
$$
|a^{ij}_n(t,x)-a^{ij}_n(s,y)|\leq\varepsilon, \quad \forall \, \omega \in \Omega
$$
whenever $|(t,x)-(s,y)|\leq \delta$.
\noindent
$(iii)$ There exists a constant $\delta_0\in(0,1]$ so that for any $n,\omega,t,x$
\begin{equation}
\label{elliptic}
\delta_0 |\xi|^{2}\leq a_{n}^{ij}(t,x)\xi^{i}\xi^{j}\leq \delta_0^{-1}|\xi|^{2},\quad \forall \xi\in\mathbb{R}^{d},
\end{equation}
$$
|b^{i}(t,x)|+|c(t,x)|+|\sigma^{ij}(t,x)|_{l^{2}}+|\mu^{i}(t,x)|_{l_{2}}+|\nu(t,x)|_{l^{2}}\leq \delta^{-1}_0.
$$
\noindent
$(iv)$ $\sigma^{ijk}=0$ if $\beta\geq1/2$, and $\mu^{ik}=0$
if $\beta\geq1/2+\alpha/2$ for every $i,j,k,\omega,t,x$.
\end{assumption}
Recall for $a\in \bR$, $a_+:=a \vee 0$.
For $\kappa\in (0,1)$, denote
\begin{equation}
\label{c_0}
c_0=c_0(\alpha,\beta)=\frac{(2\beta-1)_+}{\alpha}, \quad c'_0=c'_0(\kappa)=c_0+\kappa 1_{\beta=1/2}.
\end{equation}
Note that $c'_0\in [0,2)$ because $\beta<\alpha+\frac{1}{2}$, and $c_0=c'_0=0$ if $\beta<1/2$.
\begin{rem}
(i) Assumption \ref{assu:common}(iv) is made on the basis of the model equation
$$
\partial^{\alpha}_tu=(\Delta u + \tilde{f})dt +\partial^{\beta}_t \int^t_0 g^k dw^k_s, \quad \quad u(0)=1_{\alpha>1}u'(0)=0,
$$
for which the following sharp estimate holds (see Lemma \ref{lem:L2 result} and Theorem \ref{thm:model eqn}):
for any $\gamma \in \bR$ and $\kappa>0$,
\begin{equation}
\label{model esti}
\|u\|_{\bH^{\gamma+2}_p(T)}\leq c\left(\|\tilde{f}\|_{\bH^{\gamma}_p(T)}+\|g\|_{\bH^{\gamma+c'_0}_p(T,l_2)}\right).
\end{equation}
Thus to have $H^{\gamma+2}_p$-valued solutions we need
$\tilde{f}\in \bH^{\gamma}_p(T)$ and $g\in \bH^{\gamma+c'_0}_p(T,l_2)$. In particular if $\beta<1/2$ then the solution is twice more differentiable than $g$. This enables us to have the second derivatives of solutions in the stochastic parts of equations
(\ref{eq:target SPDE non-div}) and (\ref {eq:target SPDE div}).
(ii) For the solution of stochastic heat equation $du=\Delta u dt + g(u)dW_t$ (this is the case when $\alpha=\beta=1$), the solution is once more differentiable than $g$ (i.e. $\|\nabla u\|_{L_p} \approx \|g\|_{L_p}$), and if $g$ contains any second-order derivatives of $u$ then one cannot control $\nabla u$ and any other derivatives of $u$.
\end{rem}
\begin{rem}
Due to (\ref{model esti}) we need $c'_0>c_0$ if $\beta=1/2$. This is why in Assumption
\ref{assu:non-div} below we impose extra smoothness on the coefficients and free terms of the stochastic parts when $\beta=1/2$.
\end{rem}
To describe the regularity of the coefficients we introduce the following space introduced e.g. in \cite{Krylov1999}. Fix $\delta_1>0$, and for each $r\geq 0$, let
$$
B^{r}:=
\begin{cases}
L_{\infty}(\mathbb{R}^{d})&: r=0\\
C^{r-1,1}(\mathbb{R}^{d})&: r=1,2,3,\cdots\\
C^{r+\delta_1}(\mathbb{R}^{d})&: \text{otherwise},
\end{cases}
$$
where $C^{r+\delta_1}(\mathbb{R}^{d})$ and
$C^{r-1,1}(\mathbb{R}^{d})$ are the H\"{o}lder space and the
Zygmund space respectively. We also define the space
$B^{r}(l_{2})$ for $l_2$-valued functions
using $|\cdot|_{l_2}$ in place of $|\cdot|$.
It is well-known (e.g. \cite[Lemma 5.2]{Krylov1999}) that for any $\gamma\in \bR$, $u\in H^{\gamma}_p$ and $a\in B^{|\gamma|}$,
\begin{equation}
\label{multi}
\|au\|_{H^{\gamma}_p}\leq N(d,p,\delta_1,\gamma) |a|_{B^{|\gamma|}} \|u\|_{H^{\gamma}_p},
\end{equation}
and similarly for any $b\in B^{|\gamma|}(l_2)$,
\begin{align}
\label{multi2}
\|bu\|_{H^{\gamma}_p(l_2)}\leq N(d,p,\delta_1,\gamma) |b|_{B^{|\gamma|}(l_2)} \|u\|_{H^{\gamma}_p}.
\end{align}
The following assumption is only for the divergence type equation.
We use the notation $f^i(u)$, $h(u)$, and $g(u)$ to denote $f^i(t,x,u)$, $h(t,x,u)$, and $g(t,x,u)$, respectively.
Take $c'_0$ from (\ref{c_0}) and note $ c'_0-1<1$.
\begin{assumption}
\label{assu:div}
$(i)$ There exists a $\kappa\in (0,1)$ so that for any $u\in \bH^1_p(T)$,
$$f^i(u) \in \bL_p(T), \quad h(u) \in \bH^{-1}_p(T), \quad g(u) \in \bH_p^{c'_0-1}(T,l_2).
$$
$(ii)$ For any $\varepsilon>0$ there exists $K_{1}=K_{1}(\varepsilon)$
so that
\begin{align} \nonumber
&\left\Vert f^{i}(t,\cdot,u)-f^{i}(t,\cdot,v)\right\Vert _{L_{p}}
+\left\Vert h(t,\cdot,u)-h(t,\cdot,v)\right\Vert _{H_{p}^{-1}(l_2)}\\
&+\left\Vert g(t,\cdot,u)-g(t,\cdot,v)\right\Vert _{H_{p}^{c'_{0}-1}(l_2)}
\leq\varepsilon\|u-v\|_{H_{p}^{1}}+K_{1}\|u-v\|_{L_{p}}
\label{eq:assumption div}
\end{align}
for all $u,v\in H_{p}^{1}$ and $\omega,t$.
\noindent
$(iii)$ There exists a constant $K_{2}>0$ such that
$$
\left|\sigma^{ij}(t,\cdot)\right|_{B^{1}(l_{2})}+\left|\mu^{i}(t,\cdot)\right|_{B^{|c'_{0}-1|}(l_{2})}+\left|\nu(t,\cdot)\right|_{B^{|c'_{0}-1|}(l_{2})}\leq K_{2},\quad\forall i,j,\omega,t.
$$
\end{assumption}
\vspace{3mm}
Note that (\ref{eq:assumption div}) is certainly satisfied if $f^i(v)$, $h(v)$, and $g(v)$ are Lipschitz continuous with repsect to $v$ in their corresponding spaces uniformly on $\omega$ and $t$ . Indeed, if $g(v)$ is Lipschitz continuous then using $c'_0-1<1$ and an interpolation inequality (see e.g. \cite[Section 2.4.7]{T}), we get for any $\varepsilon>0$,
$$
\|g(u)-g(v)\|_{H^{c'_0-1}_p(l_2)}\leq N \|u-v\|_{H^{c'_0-1}_p}\leq \varepsilon \|u-v\|_{H^1_p}+K(\varepsilon)\|u-v\|_{L_p}.
$$
Finally we give our main result for divergence equation (\ref{eq:target SPDE div}).
\begin{thm}
\label{thm:main results div}
Let $p\geq2$.
Suppose that Assumptions \ref{assu:common} and \ref{assu:div} hold.
Then divergence type equation \eqref{eq:target SPDE div}
with the zero initial condition has a unique solution
$u\in\mathcal{H}_{p}^{1}(T)$ in the sense of Definition \ref{def:solution space}, and for this solution we have
\begin{equation}
\|u\|_{\mathbb{H}_{p}^{1}(T)}\leq N\left(\|f^i(0)\|_{\mathbb{L}_{p}(T)}+\|h(0)\|_{\mathbb{H}_{p}^{-1}(T)}+\|g(0)\|_{\mathbb{H}_{p}^{c'_{0}-1}(T)}\right),\label{eq:a priori estimate div}
\end{equation}
where the constant $N$ depends only on $d$, $p$, $\alpha$, $\beta$, $\kappa$, $\delta_0$, $\delta_1$, $K_1$, $K_2$, and $T$.
\end{thm}
Next we introduce our result for non-divergence equation.
To have $H^{\gamma+2}_p$-valued solution we assume the following conditions.
\begin{assumption}
\label{assu:non-div}
$(i)$ There exists a $\kappa\in (0,1)$ so that for any $u\in \bH^{\gamma+2}_p(T)$,
$$f(u) \in \bH^{\gamma}_p(T), \quad g(u) \in \bH_p^{\gamma+c'_0}(T,l_2).
$$
\noindent
$(ii)$ There exists a constant $K_3$ so that for any $\omega$, $t$, $i$, $j$,
\begin{equation}
\label{multiplier}
|a^{ij}(t,\cdot)|_{B^{|\gamma|} }+ |b^i(t,\cdot)|_{B^{|\gamma|} }+ |c(t,\cdot)|_{B^{|\gamma|} }
\leq K_3,
\end{equation}
and
$$
|\sigma^{ij}(t,\cdot)|_{B^{|\gamma+c'_{0}|}(l_{2})}+|\mu^{i}(t,\cdot)|_{B^{|\gamma+c'_{0}|}(l_{2})}+|\nu(t,\cdot)|_{B^{|\gamma+c'_{0}|}(l_{2})}
\leq K_3.
$$
\noindent
$(iii)$ For any $\varepsilon>0$, there exists a constant $K_{4}=K_4(\varepsilon)>0$
such that
\begin{align}
\notag
\left\Vert f(t,u)-f(t,v)\right\Vert _{H^{\gamma}_{p}}+ \left\Vert g(t,u)-g(t,v)\right\Vert _{H_{p}^{\gamma+c'_{0}}(l_{2})}\\
\leq\varepsilon\|u-v\|_{H_{p}^{\gamma+2}}+K_{4}\|u-v\|_{H^{\gamma}_{p}}\label{eq:assumption non-div},
\end{align}
for any $u,v\in H_{p}^{\gamma+2}$ and $\omega,t$.
\end{assumption}
See \cite{Krylov1999} for some examples of (\ref{eq:assumption non-div}). Here we introduce only one nontrivial example.
Let $\gamma+2-d/p>n$ for some $n\in \{0,1,2,\cdots\}$ and $f_0=f_0(x)\in H^{\gamma}_p$. Take
$$
f(u)=f_0(x) \sup_x |D^n_xu|.
$$
Take a $\delta>0$ so that $\gamma+2-d/p-n>\delta$. Using a Sobolev embedding $H^{\gamma+2-\delta}_p \subset C^{\gamma+2-\delta-d/p}\subset C^n$,
we get for any $u, v \in H^{\gamma+2}_p$ and $\varepsilon>0$,
\begin{eqnarray*}
&& \|f(u)-f(v)\|_{H^{\gamma}_p}\leq \|f_0\|_{H^{\gamma}_p} \sup_x |D^n_x(u-v)|\leq N|u-v|_{C^n}\\
&\leq& N\|u-v\|_{H^{\gamma+2-\delta}_p}\leq \varepsilon \|u-v\|_{H^{\gamma+2}_p}+K(\varepsilon)\|u-v\|_{H^{\gamma}_p}.
\end{eqnarray*}
\vspace{3mm}
Here is our main result for non-divergence equation \eqref{eq:target SPDE non-div}.
\begin{thm}
\label{thm:main results non-div}
Let $\gamma\in \bR$ and $p\geq2$.
Suppose that Assumptions \ref{assu:common} and \ref{assu:non-div} hold.
Then non-divergence type equation
\eqref{eq:target SPDE non-div} with zero initial condition
has a unique solution $u\in\mathcal{H}_{p}^{\gamma+2}(T)$ in the sense of Definition \ref{def:solution space}, and for this solution
\begin{equation} \label{eq: a priori estimate non-div}\|u\|_{\mathbb{H}_{p}^{\gamma+2}(T)}\leq
N\left(\|f(0)\|_{\mathbb{H}_{p}^{\gamma}(T)}+\|g(0)\|_{\mathbb{H}_{p}^{\gamma+c'_{0}}(T,l_{2})}\right),
\end{equation}
where the constant $N$ depends only on $d$, $p$, $\alpha$, $\beta$, $\kappa$, $\delta_0$, $\delta_1$, $K_3$, $K_4$, and $T$.
\end{thm}
\mysection{Parabolic Littlewood-Paley inequality}
In this section we obtain a sharp $L_p$-estimate for solutions to the model equation
\begin{equation}
\label{plp}
\partial^{\alpha}_tu=\Delta u +\partial^{\beta}_t \int^t_0 g^k dw^k_s.
\end{equation}
For this, we prove the parabolic Littlewood-Paley inequality related to the equation. For the classical case $\alpha=\beta=1$ we refer to \cite{IPK,Kr01, kr94}.
Consider the fractional diffusion-wave equation
\begin{equation}
\label{deter}
\partial_{t}^{\alpha}u(t,x)=\Delta u(t,x), \quad u(0)=u_0, \quad 1_{\alpha>1}u'(0)=0.
\end{equation}
By taking the Fourier transform and the inverse Fourier transform with respect to $x$, we formally find that $u(t)=p(t)*u_0$ is a solution to this problem if $p(t,x)$ satisfies
\begin{equation}
\label{eqn 8.17.2}
\partial^{\alpha}_t \cF(p)=-|\xi|^2 \cF(p),\quad \cF(p)(0,\xi)=1, \quad 1_{\alpha>1}\cF\left(\frac{\partial p}{\partial t}\right)(0,\xi)=0.
\end{equation}
It turns out that (see \cite{eik2004, KS2015} or Lemma \ref {prop:kernel esti. of q} below) there exists a function $p(t,x)$, called the fundamental solution, such that it satisfies (\ref{eqn 8.17.2}). It is also true that $p$ is infinitely differentiable in $(0,\infty)\times\mathbb{R}^{d}\setminus \{0\}$ and $\lim_{t\rightarrow 0} \frac{\partial^np(t,x)}{\partial t^n}=0$ if $x\neq 0$.
Define
\begin{equation}
q_{\alpha,\beta}(t,x):=\begin{cases}
I_{t}^{\alpha-\beta}p(t,x) & :\alpha\geq\beta\\
D_{t}^{\beta-\alpha}p(t,x) & :\alpha<\beta,
\end{cases}
\label{eq:definition of q_=00007Ba,b=00007D}
\end{equation}
and
$$
q(t,x):=q_{\alpha,1}(t,x).
$$
Note that $q_{\alpha,\beta}$ is well defined due to above mentioned properties of $p$.
Moreover $D_{t}^{\beta-\alpha}p(t,x)=\partial_{t}^{\beta-\alpha}p(t,x)$ since $p(0,x)=0$ if $x\neq 0$.
\vspace{1mm}
In the following lemma we collect some important properties of $p(t,x)$, $q(t,x)$, and $q_{\alpha,\beta}(t,x)$ taken from
\cite{eik2004} and \cite{KS2015}.
\begin{lem}
\label{prop:kernel esti. of q}
Let $d\in \bN$, $\alpha\in(0,2)$,
$\beta<\alpha+\frac{1}{2}$, and $\gamma\in[0,2)$.
\vspace{1mm}
\noindent
$(i)$ There exists a fundamental solution $p(t,x)$ satisfying above mentioned properties. It also holds that
for any $t\neq0$ and $x\neq0$,
\begin{equation}
\label{eqn 8.17.1}
\partial_{t}^{\alpha}p(t,x)=\Delta p(t,x), \quad
\frac{\partial p(t,x)}{\partial t}=\Delta q(t,x),
\end{equation}
and for each $x\neq 0$, $\frac{\partial}{\partial t}p(t,x)\to 0$ as $t\downarrow 0$.
Moreover, $\frac{\partial}{\partial t}p(t,\cdot)$ is integrable in $\mathbb{R}^{d}$ uniformly on $t \in [\varepsilon,T]$ for any $\varepsilon>0$.
\vspace{1mm}
\noindent
$(ii)$ If $n\leq3$, $D_{x}^{n}q(t,\cdot)$ is integrable in $\mathbb{R}^{d}$ uniformly on $t \in [\varepsilon,T]$ for any $\varepsilon>0$.
\vspace{1mm}
\noindent
$(iii)$ There exist constants $c=c(d,\alpha)$ and $N=N(d,\alpha)$ such that if $|x|^2 \geq t^\alpha$,
\begin{align}
\label{p ker est}
\left|p(t,x)\right|
&\leq N |x|^{-d} \exp\left\{ -c |x|^{\frac{2}{2-\alpha}} t^{-\frac{\alpha}{2-\alpha}}\right\}.
\end{align}
\vspace{1mm}
\noindent
$(iv)$
It holds that
\begin{align}
\label{mit rela}
\mathcal{F}\{D_t^\sigma q_{\alpha,\beta}(t,\cdot)\}(\xi)=t^{\alpha-\beta-\sigma}E_{\alpha,1+\alpha-\beta-\sigma}(-|\xi|^{2}t^{\alpha}),
\end{align}
where $E_{a,b}(z)$, $a>0$, is the Mittag-Leffler function defined as
$$
E_{a,b}(z):=\sum_{k=0}^{\infty}\frac{z^{k}}{\Gamma(ak+b)}, \quad z\in \bC.
$$
\vspace{1mm}
\noindent
$(v)$
There exists a constant $N=N(d,\gamma,\alpha,\beta)$ such that
$$
\left|D_t^\sigma (-\Delta)^{\gamma/2}q_{\alpha,\beta}(1,x)\right|
+\left|D_t^\sigma(-\Delta)^{\gamma/2}\partial_t q_{\alpha,\beta}(1,x)\right|
\leq N\left(|x|^{-d+2-\gamma}\wedge|x|^{-d-\gamma}\right)
$$
if $d \geq 2$,
and
\begin{align*}
&\left|D_t^\sigma(-\Delta)^{\gamma/2}q_{\alpha,\beta}(1,x)\right|
+\left|D_t^\sigma(-\Delta)^{\gamma/2}\partial_t q_{\alpha,\beta}(1,x)\right| \\
&\qquad \leq N\left(\{|x|^{1-\gamma}(1+\ln|x|1_{\gamma=1})\}\wedge|x|^{-1-\gamma}\right)
\end{align*}
if $d=1$.
Furthermore, for each $n\in\mathbb{N}$
\begin{align}
\notag
&\left|D_t^\sigma D_{x}^{n}(-\Delta)^{\gamma/2}q_{\alpha,\beta}(1,x)\right|
+\left|D_t^\sigma D_{x}^{n}(-\Delta)^{\gamma/2}\partial_t q_{\alpha,\beta}(1,x)\right| \\
\label{de ker est}
&\qquad \leq N(d,\gamma,\alpha,\beta,n)\left(|x|^{-d+2-\gamma-n}\wedge|x|^{-d-\gamma-n}\right).
\end{align}
\vspace{1mm}
\noindent
(vi)
The scaling properties hold:
\begin{equation}
\label{eq:q scaling prop}
q_{\alpha,\beta}(t,x)=t^{-\frac{\alpha d}{2}+\alpha-\beta}q_{\alpha,\beta}(1,xt^{-\frac{\alpha}{2}}),
\end{equation}
\begin{equation}
\label{eq:q gamma scaling prop}
D_t^\sigma (-\Delta)^{\gamma/2}q_{\alpha,\beta}(t,x)
=t^{-\sigma-\frac{\alpha(d+\gamma)}{2}+\alpha-\beta}
D_t^\sigma (-\Delta)^{\gamma/2}q_{\alpha,\beta}(1,xt^{-\frac{\alpha}{2}}).
\end{equation}
\end{lem}
\begin{proof}
(i), (ii), (iii), and (v) are easily obtained from Theorem 2.1 and Theorem 2.3 of \cite{KS2015}.
The proof of (iv) can be found in Seciotn 6 of \cite{KS2015}.
For the scaling property (vi), see \cite[(5.2)]{KS2015}.
\end{proof}
The following result is well-known, for instance if $\alpha \in (0,1]$.
For the completeness of the article, we give a proof.
\begin{corollary}
\label{zero converge}
Let $f \in C_0^2(\bR^d)$. Then
$$
\int_{\bR^d} p(t,x-y)f(y)dy
$$
converges to $f(x)$ uniformly as $t \downarrow 0 $.
\end{corollary}
\begin{proof}
By (\ref{mit rela}), for any $t>0$,
$$
\int_{\bR^d} p(t,y)dy =\cF p (0) =E_{\alpha,1}(0)=1.
$$
Also (\ref{eq:q scaling prop}) shows that $\|p(t,\cdot)\|_{L_1(\bR^d)}$ is a constant function of $t$. For any $\delta >0$,
\begin{align*}
&\left|\int_{\bR^d} p(t,x-y)f(y)dy -f(x) \right|\\
&=\left|\int_{\bR^d} p(t,y)(f(x-y)-f(x))dy \right| \\
&\leq \int_{|y| < \delta} \left|p(t,y)(f(x-y)-f(x))\right|dy +\int_{|y| > \delta} \left|p(t,y)(f(x-y)-f(x))\right|dy \\
&=: \cI(\delta) + \cJ(\delta).
\end{align*}
Since $f \in C_0^2(\bR^d)$, for any $\varepsilon >0$, one can take a small $\delta$ so that $\cI(\delta) < \varepsilon$.
Moreover due to (\ref{p ker est}), for fixed $\delta>0$, $\cJ(\delta) \to 0$ as $t \downarrow 0$.
The corollary is proved.
\end{proof}
In the remainder of this section, we restrict the range of $\beta$ so that
\begin{equation}
\label{eq:assum KLP ineq}
\frac{1}{2}<\beta< \alpha+\frac{1}{2} .
\end{equation}
Thus by definition \eqref{c_0}, we have
$${c_1}:= 2 - c'_0=2-\frac{2\beta-1}{\alpha} \in(0,2).
$$
In the following section (i.e. Section 4) we prove that if $g\in \bH^{\infty}_0(T,l_2)$ then the unique solution (in the sense of Definition \ref{def:solution space}) to equation
(\ref{plp}) with the zero initial condition
is given by the formula
\begin{equation}
\label{eqn 9.3}
u=\int^t_0\int_{\bR^d}q_{\alpha,\beta}(t-s,x-y)g^k(s,y)dy dw^k_s.
\end{equation}
By Burkerholder-Davis-Gundy inequality
\begin{eqnarray}
\label{bdc}
&&\|(-\Delta)^{c_1/2}u\|^p_{\bL_p(T)}\\
&&\leq N \bE \int_{\bR^d} \int^T_0 \left[\int^t_0 \left(\int_{\bR^d} (-\Delta)^{c_1/2}q_{\alpha,\beta}(t-s,x-y)g(s,y) dy\right)^2_{l_2} ds\right]^{p/2} dt dx. \nonumber
\end{eqnarray}
Our goal is to control the right hand side of (\ref{bdc}) in terms of $\|g\|_{\bL_p(T,l_2)}$. For this, we introduce some definitions as follows. Let $H$ be a Hilbert space. For $g\in C_{c}^{\infty}(\mathbb{R}^{d+1}; H)$, define
$$
T_{t-s}^{\alpha,\beta}g(s,\cdot)(x):=\int_{\mathbb{R}^{d}}q_{\alpha,\beta}(t-s,x-y)g(s,y)dy.
$$
Note that, due to Lemma \ref{prop:kernel esti. of q}(v), $(-\Delta)^{{c_1}/2}q_{\alpha,\beta}(t,\cdot)\in L_1 (\mathbb{R}^{d})$ for all $t>0$.
Therefore, for any $t>s$
$$
(-\Delta)^{{c_1}/2}T_{t-s}^{\alpha,\beta}g(s,\cdot)\in L_1 (\mathbb{R}^{d})
$$
and
$$
(-\Delta)^{{c_1}/2}T_{t-s}^{\alpha,\beta}g(s,\cdot)(x)=\int_{\mathbb{R}^{d}}(-\Delta)^{{c_1}/2}q_{\alpha,\beta}(t-s,x-y)g(s,y)dy.
$$
We also define the sublinear operator $\cT$ as
\begin{align*}
\mathcal{T}g(t,x)
&:=\left[\int_{-\infty}^{t}\left| (-\Delta)^{{c_1}/2}T_{t-s}^{\alpha,\beta}g(s,\cdot)(x)\right| _{H}^{2}ds\right]^{1/2},
\end{align*}
where $|\cdot|_H$ denotes the given norm in the Hilbert space $H$. $\cT$ is sublinear due to the Minkowski inequality
\begin{equation}
\label{sublinear}
\|f+g\|_{L_2((-\infty,t);H)}\leq \|f\|_{L_2((-\infty,t);H)}+\|g\|_{L_2((-\infty,t);H)}.
\end{equation}
Now we introduce a parabolic version of Littlewood-Paley inequality. The proof is given at the end of this section.
\begin{thm}
\label{thm:L-P}
Let $H$ be a separable Hilbert space, $p\in[2,\infty)$,
$T\in(-\infty,\infty]$, and $\alpha\in(0,2)$. Assume that \eqref{eq:assum KLP ineq} holds.
Then for any $g\in C_{c}^{\infty}(\mathbb{R}^{d+1};H)$,
\begin{equation}
\int_{\mathbb{R}^{d}}\int_{-\infty}^{T}\left|\mathcal{T}g(t,x)\right|^{p}dtdx\leq N\int_{\mathbb{R}^{d}}\int_{-\infty}^{T}|g(t,x)|_{H}^{p}dtdx,
\label{eq:L-P ineq}
\end{equation}
where $N=N(d,p,\alpha,\beta)$.
\end{thm}
\begin{rem}
\label{rem:extension of cT}
By Theorem \ref{thm:L-P}, the operator
$\mathcal{T}$ can be continuously extended onto $L_{p}(\mathbb{R}^{d+1} ;H)$.
We denote this extension by the same notation $\mathcal{T}$.
\end{rem}
\begin{rem}
\label{rem 9.3}
Take $u$ and $g$ from (\ref{eqn 9.3}). Extend $g(t)=0$ for $t\leq
0$. Note that the right hand side of (\ref{bdc}) is $\bE
\int_{\mathbb{R}^{d}}\int_{-\infty}^{T}\left|\mathcal{T}g(t,x)\right|^{p}dtdx$.
Thus, using (\ref{eq:L-P ineq}) (actually Remark \ref{rem:extension
of cT}) for each $\omega$ and taking the expectation, we get
$$
\|(-\Delta)^{c_1/2}u\|^p_{\bL_p(T)}\leq N
\|g\|^p_{\bL_p(T,l_2)}.
$$
\end{rem}
First we prove Theorem \ref{thm:L-P} for $p=2$. The following lemma is a slight extension of \cite[Lemma 3.8]{CKK}, which is proved only for $\alpha\in (0,1)$ with constant $N$ depending also on $T$. For the proof we use the following well-known property of the Mittag-Leffler function: if $\alpha\in (0,2)$ and $b\in \bC$, then there exist positive constants $\varepsilon=\varepsilon(\alpha)$
and $C=C(\alpha,b)$ such that
\begin{equation}
\label{eqn 8.19.2}
|E_{\alpha,b}(z)|\leq C(1 \wedge |z|^{-1}), \quad \quad \pi-\varepsilon \leq |\arg (z)|\leq \pi.
\end{equation}
See \cite[Lemma 3.1]{SY} for the proof of (\ref{eqn 8.19.2}).
\begin{lem}
\label{lem:L2 result}
Suppose that the assumptions in Theorem \ref{thm:L-P} hold.
Then for any $T\in (-\infty, \infty]$ and $g\in C_{c}^{\infty}(\mathbb{R}^{d+1};H)$,
\begin{equation}
\label{eqn 8.19.1}
\int_{\mathbb{R}^{d}}\int_{-\infty}^{T}|\mathcal{T}g(t,x)|^{2}dtdx\leq N\int_{\mathbb{R}^{d}}\int_{-\infty}^{T}\left| g(t,x)\right| _{H}^{2}dtdx,
\end{equation}
where $N=N(d,p,\alpha,\beta)$ is independent of $T$.
\end{lem}
\begin{proof}
{\bf{Step 1}}. First, assume $g(t,x)=0$ for $t \leq 0$. In this case we may assume $T>0$ because the left hand side of (\ref{eqn 8.19.1}) is zero if $T\leq 0$.
We prove (\ref{eq:L-P ineq}) for $T=1$. Since $g(t,x)=\cT g(t,x)=0$ for $t \leq 0$,
by Parseval's identity and (\ref{mit rela}),
\begin{align*}
& \int_{\mathbb{R}^{d}}\int_{-\infty}^{1}|\mathcal{T}g(t,x)|^{2}dtdx\\
& =\int_{0}^{1}\int_{0}^{t}\int_{\bR^d}|\xi|^{2{c_1}}\left|\cF\left\{ q_{\alpha,\beta}(t-s,\cdot)\right\} (\xi)\right|^{2}\left| \mathcal{F}\{g\}(s,\xi)\right| _{H}^{2}d\xi dsdt\\
& \leq \int_{|\xi|\leq 1}\int_{0}^{1}\left| \mathcal{F}\{g\}(s,\xi)\right| _{H}^{2}\left(\int_{s}^{1}|\xi|^{2{c_1}}\left|t^{\alpha-\beta}E_{\alpha,1-\beta+\alpha}(-|\xi|^{2}t^{\alpha})\right|^{2}dt\right)dsd\xi\\
& \quad+\int_{|\xi|\geq1}\int_{0}^{1}\left| \mathcal{F}\{g\}(s,\xi)\right| _{H}^{2}\left(\int_{s}^{1}|\xi|^{2{c_1}}
\left|t^{\alpha-\beta}E_{\alpha,1-\beta+\alpha}(-|\xi|^{2}t^{\alpha})\right|^{2}dt\right)dsd\xi \\
& \leq N\int_{0}^{1}\int_{\mathbb{R}^{d}}\left| g(t,x)\right| _{H}^{2}dxdt\\
& \quad+N\int_{|\xi|\geq1}\int_{0}^{1}\left| \mathcal{F}\{g\}(s,\xi)\right| _{H}^{2}\left(\int_{s}^{1}|\xi|^{2{c_1}}
\left|t^{\alpha-\beta}E_{\alpha,1-\beta+\alpha}(-|\xi|^{2}t^{\alpha})\right|^{2}dt\right)dsd\xi,
\end{align*}
where the last inequality is due to (\ref{eqn 8.19.2}) and the condition $\alpha-\beta>-1/2$. Thus to prove our assertion for $T=1$ we only need to prove
$$
\sup_{\xi} \left(1_{|\xi|\geq 1}
|\xi|^{2c_1}\int^{1}_0\left|t^{\alpha-\beta}E_{\alpha,1-\beta+\alpha}(-|\xi|^{2}t^{\alpha})\right|^{2}dt\right)
<\infty.
$$
By (\ref{eqn 8.19.2}), if $|\xi|\geq 1$ (recall we assumed
$\beta>1/2$ in this section),
\begin{align*}
& |\xi|^{2c_1}\int^{1}_0\left|t^{\alpha-\beta}E_{\alpha,1-\beta+\alpha}(-|\xi|^{2}t^{\alpha})\right|^{2}dt\\
& \leq N |\xi|^{2{c_1}}\int_{0}^{|\xi|^{-2/\alpha}} t^{2(\alpha-\beta)} dt
+N|\xi|^{2c_1}\int_{|\xi|^{-2/\alpha}}^{1} \left|\frac{t^{\alpha-\beta}}{|\xi|^{2}t^{\alpha}}\right|^{2} dt\\
&\leq N |\xi|^{2(c_1-2+\frac{2\beta-1}{\alpha})}+N|\xi|^{2c_1-4}\left(|\xi|^{2(\frac{2\beta-1}{\alpha})}-1\right)\leq 3N.
\end{align*}
Therefore, the case $T=1$ is proved.
For arbitrary $T>0$,
we use (\ref{eq:q gamma scaling prop}), which implies
\begin{equation}
\label{8.20.3}
(-\Delta)^{{c_1}/2}q_{\alpha,\beta}(T(t-s),x)
=T^{-\frac{\alpha(d+{c_1})}{2}+\alpha-\beta}(-\Delta)^{{c_1}/2}q_{\alpha,\beta}(t-s,T^{-\frac{\alpha}{2}}x),
\end{equation}
and consequently
\begin{equation}
\label{8.20.4}
\mathcal{T}g(Tt,x)=\mathcal{T}\tilde{g}(t,T^{-\frac{\alpha}{2}}x),
\end{equation}
where $\tilde{g}(t,x)=g(Tt,T^{\frac{\alpha}{2}}x)$.
By using the result proved for $T=1$,
\begin{align*}
\int_{\mathbb{R}^{d}}\int_{-\infty}^{T}|\mathcal{T}g(t,x)|^{2}dtdx & =T^{1+\frac{\alpha d}{2}}\int_{\mathbb{R}^{d}}\int_{-\infty}^{1}|\mathcal{T}\tilde{g}(t,x)|^{2}dtdx\\
& \leq NT^{1+\frac{\alpha d}{2}}\int_{\mathbb{R}^{d}}\int_{-\infty}^{1}|\tilde{g}(t,x)|^{2}dtdx\\
& =N\int_{\mathbb{R}^{d}}\int_{-\infty}^{T}|g(t,x)|^{2}dtdx.
\end{align*}
Thus \eqref{eq:L-P ineq} holds for all $T>0$ with a constant independent of $T$. It follows that
\eqref{eq:L-P ineq} also holds for $T= \infty$.
{\bf{Step 2}}. General case. Take $a\in \bR$ so that $g(t,x)=0$ for $t\leq a$. Then obviously, for $\bar{g}(t,x):=g(t+a,x)$ we have $\bar{g}(t)=0$ for $t\leq 0$. Thus it is enough to apply the result for Step 1 with $\bar{g}$ and $T-a$ in place of $g$ and $T$ respectively.
\end{proof}
For a real-valued measurable function $h$ on $\mathbb{R}^{d}$, define
the maximal function
$$
\bM_xh(x):=\sup_{r>0}\frac{1}{|B_{r}(x)|}\int_{B_{r}(x)}|h(y)|dy=\sup_{r>0}\aint_{B_{r}(x)}|h(y)|dy.
$$
The Hardy-Littlewood maximal theorem says
\begin{equation}
\label{hl}
\|\bM_x h\|_{L_p(\bR^d)}\leq N(d,p) \|h\|_{L_p(\bR^d)}, \quad \forall p>1.
\end{equation}
For a function $h(t,x)$, set
$$
\bM_xh(t,x)=\bM_x\left(h(t,\cdot)\right)(x),\quad\bM_{t}h(t,x)=\bM_{t}\left(h(\cdot,x)\right)(t),
$$
and
$$
\bM_t \bM_x h(t,x)=\bM_t\left(\bM_xh(\cdot,x)\right)(t).
$$
To evaluate $\bM_t \bM_x h(t,x)$, we first fix $t$ and estimate $(\bM_x h(t,\cdot))(x)$. After this, we fix $x$ and regard $(\bM_x h(t,\cdot))(x)$
as a function of $t$ only to estimate the maximal function with respect to $t$.
Denote
\begin{equation}
Q_{0}:=[-2^{\frac{2}{\alpha}},0]\times[-1,1]^{d}.\label{eq:Q zero}
\end{equation}
\begin{lem}
\label{lem:first lemma}
Let $g\in C_{c}^{\infty}(\mathbb{R}^{d+1};H)$
and assume that $g=0$ outside of $[-4^{\frac{2}{\alpha}},4^{\frac{2}{\alpha}}]\times B_{3d}$.
Then for $(t,x)\in Q_{0}$,
$$
\int_{Q_{0}}\left|\mathcal{T}g(s,y)\right|^{2}dsdy\leq N\bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x),
$$
where $N=N(d,\alpha,\beta)$.
\end{lem}
\begin{proof}
By Lemma \ref{lem:L2 result},
$$
\int_{Q_{0}}\left|\mathcal{T}g(s,y)\right|^{2}dsdy\leq\int_{-4^{\frac{2}{\alpha}}}^{0}\int_{B_{3d}}\left| g(s,y)\right| _{H}^{2}dyds.
$$
For any $(t,x)\in Q_{0}$ and $y\in B_{3d}$, since $|x-y|\leq|x|+|y|\leq\sqrt{d}+3d\leq4d$,
we obtain
\begin{align*}
\int_{-4^{\frac{2}{\alpha}}}^{0}\int_{B_{3d}}\left| g(s,y)\right| _{H}^{2}dyds
& \leq\int_{-4^{\frac{2}{\alpha}}}^{0}\int_{|x-y|\leq4d}\left| g(s,y)\right| _{H}^{2}dyds\\
& \leq N\int_{-4^{\frac{2}{\alpha}}}^{0}\bM_x\left| g(s,x)\right| _{H}^{2}ds\\
& \leq N\bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x).
\end{align*}
The lemma is proved.
\end{proof}
Here is a generalization of Lemma \ref{lem:first lemma}.
\begin{lem}
\label{lem:second lemma}
Let $g\in C_{c}^{\infty}(\mathbb{R}^{d+1};H)$
and assume that $g(t,x)=0$ for $|t|\geq4^{\frac{2}{\alpha}}$. Then for any $(t,x)\in Q_{0}$,
$$
\int_{Q_{0}}\left|\mathcal{T}g(s,y)\right|^{2}dsdy\leq N(d,\alpha,\beta)\bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x).
$$
\end{lem}
\begin{proof}
Take $\zeta\in C_{c}^{\infty}(\mathbb{R}^{d})$ such that $\zeta=1$
in $B_{2d}$ and $\zeta=0$ outside $B_{3d}$.
Recall that $\cT$ is a sublinear operator, and therefore
$$
\mathcal{T}g\leq\mathcal{T}(\zeta g)+\mathcal{T}( (1-\zeta)g).
$$
Since $\mathcal{T}(\zeta g)$ can be estimated by Lemma \ref{lem:first lemma},
we may assume that $g(t,x)=0$ for $x\in B_{2d}$. Let $0>s>r>-4^{\frac{2}{\alpha}}$.
Then by \eqref{eq:q gamma scaling prop},
\begin{align}
& \left| (-\Delta)^{{c_1}/2}T_{s-r}^{\alpha,\beta}g(r,\cdot)(y)\right| _{H}\nonumber \\
& \leq(s-r)^{-\frac{\alpha d}{2}+\alpha-\beta-\frac{\alpha{c_1}}{2}}\int_{\mathbb{R}^{d}}\left|(-\Delta)^{{c_1}/2}q_{\alpha,\beta}(1,(s-r)^{-\frac{\alpha}{2}}y)\right|\left| g(r,y-z)\right| _{H}dz\nonumber \\
\label{eq:5-12-1}
& =(s-r)^{-\frac{\alpha d}{2}-\frac{1}{2}}\int_{\mathbb{R}^{d}}\left|(-\Delta)^{{c_1}/2}q_{\alpha,\beta}(1,(s-r)^{-\frac{\alpha}{2}}y)\right|\left| g(r,y-z)\right| _{H}dz.
\end{align}
To proceed further we use the following integration by parts formula : if $F$ and $G$ are smooth enough then for any $0<\varepsilon <R<\infty$,
\begin{align}
\int_{\epsilon\leq|\eta|\leq R}F(\eta)G(|\eta|)d\eta & =-\int_{\epsilon}^{R}G'(\rho)\left[\int_{|\eta|\leq\rho}F(\eta)d\eta\right]d\rho\nonumber \\
& \qquad+G(R)\int_{|\eta|\leq R}F(\eta)d\eta-G(\epsilon)\int_{|\eta|\leq\epsilon}F(\eta)d\eta.
\label{eq:integration by parts}
\end{align}
Indeed, (\ref{eq:integration by parts}) is obtained by applying integration by parts to
\begin{align*}
\int_\varepsilon^R G(\rho)\frac{d}{d\rho} \left(\int_{B_\rho(0)}F(z)\, dz \right)d\rho
&=\int_\varepsilon^R G(\rho)\left(\int_{\partial B_\rho(0)} F(s) \, dS_{\rho} \right) d\rho \\
&=\int_{R \geq |z| \geq \varepsilon} F(z) G(|z|)\,dz.
\end{align*}
Observe
that if $(s,y)\in Q_{0}$ and $\rho>1$, then
\begin{equation}
|x-y|\leq2d,\quad B_{\rho}(y)\subset B_{2d+\rho}(x)\subset B_{(2d+1)\rho}(x),\label{eq:2}
\end{equation}
whereas if $\rho\leq1$ then for $z\in B_{\rho}(0)$,
$|y-z|\leq\sqrt{d}+1\leq2d$ and thus $g(r,y-z)=0$.
Therefore by (\ref{eq:integration by parts}) and(\ref{de ker est}),
\begin{align*}
& (s-r)^{-\frac{\alpha d}{2}-\frac{1}{2}}\int_{\mathbb{R}^{d}}\left|(-\Delta)^{{c_1}/2}q_{\alpha,\beta}\left(1,(s-r)^{-\frac{\alpha}{2}}y\right)\right|\left| g(r,y-z)\right| _{H}dz\\
& \leq N(s-r)^{-\frac{\alpha d}{2}-\frac{1}{2}-\frac{\alpha}{2}}\int_{1}^{\infty} \left((s-r)^{-\frac{\alpha}{2}}\rho\right)^{-d-1-{c_1}}
\left[\int_{|z|\leq\rho}\left| g(r,y-z)\right|_{H}dz\right]d\rho\\
& \leq N(s-r)^{\alpha-\beta}\int_{1}^{\infty}\rho^{-d-1-{c_1}}\left[\int_{|z|\leq\rho}| g(r,y-z)| _{H}dz\right]d\rho\\
& \leq N(s-r)^{\alpha-\beta}\int_{1}^{\infty}\rho^{-1-{c_1}}\left[\aint_{B_{3\rho}(x)}\left| g(r,z)\right| _{H}dz\right]d\rho\\
& \leq N(s-r)^{\alpha-\beta}\bM_x\left| g\right| _{H}(r,x).
\end{align*}
Then due to the fact that $(\bM_x\left| g\right| _{H})^{2}\leq\bM_x\left| g\right| _{H}^{2}$,
\begin{align*}
\int_{Q_{0}}
&\left|\mathcal{T}g(s,y)\right|^{2}dsdy=\int_{Q_{0}}\int_{-\infty}^{s}\left| (-\Delta)^{{c_1}/2}T_{s-r}^{\alpha,\beta}g(r,\cdot)(y)\right| _{H}^{2}drdsdy\\
&\leq N\int_{Q_{0}}\int_{-4^{\frac{2}{\alpha}}}^{s}\left[\bM_x\left| g\right| _{H}^{2}(r,x)(s-r)^{2(\alpha-\beta)}\right]drds\\
&\leq N\int_{-4^{\frac{2}{\alpha}}}^{0}\left(\int_{r}^{0}(s-r)^{2(\alpha-\beta)}ds\right)\bM_x\left| g\right| _{H}^{2}(r,x)dr\leq N\bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x).
\end{align*}
The lemma is proved.
\end{proof}
\begin{lem}
\label{lem:third lemma}
Let $g\in C_{c}^{\infty}(\mathbb{R}^{d+1};H)$ and assume $g(t,x)=0$
outside of $(-\infty,-3^{\frac{2}{\alpha}})\times B_{3d}$. Then for
any $(t,x)\in Q_{0}$,
$$
\int_{Q_{0}}\left|\mathcal{T}g(s,y)\right|^{2}dsdy\leq N\bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x),
$$
where $N=N(d,\alpha,\beta)$.
\end{lem}
\begin{proof}
Note that $g(s,\cdot)=0$ for $s\geq-3^{\frac{2}{\alpha}}$. Recalling
\eqref{eq:q gamma scaling prop}, we have
\begin{align*}
& \left|\mathcal{T}g(s,y)\right|^{2} \leq\int_{-\infty}^{s}\left| (-\Delta)^{{c_1}/2}T_{s-r}^{\alpha,\beta}g(r,\cdot)(y)\right| _{H}^{2}dr\\
& =\int_{-\infty}^{-3^{\frac{2}{\alpha}}}\left| (s-r)^{-\frac{\alpha d}{2}-\frac{1}{2}}\int_{\mathbb{R}^{d}}(-\Delta)^{{c_1}/2}q_{\alpha,\beta}\left(1,(s-r)^{-\frac{\alpha}{2}}z\right)g(r,y-z)dz\right| _{H}^{2}dr\\
& \leq\int_{-\infty}^{-3^{\frac{2}{\alpha}}}(s-r)^{-\alpha d-1}\left[\int_{\mathbb{R}^{d}}\left|(-\Delta)^{{c_1}/2}q_{\alpha,\beta}\left(1,(s-r)^{-\frac{\alpha}{2}}z\right)\right|\left| g(r,y-z)\right| _{H}dz\right]^{2}dr.
\end{align*}
If $|z|\geq 4d$, then $g(r,y-z)=0$ since $y\in Q_{0}$ and $|y-z|\geq|z|-|y|\geq3d$.
Therefore, by Minkowski's inequality and Lemma \ref{prop:kernel esti. of q},
\begin{align*}
&\int_{[-1,1]^{d}} \left|\int_{\mathbb{R}^{d}}\left|(-\Delta)^{{c_1}/2}q_{\alpha,\beta}\left(1,(s-r)^{-\frac{\alpha}{2}}z\right)\right|\left| g(r,y-z)\right| _{H}dz\right|^{2}dy\\
& \leq\int_{[-1,1]^{d}}\left|\int_{|z|\leq4d}\left|(-\Delta)^{{c_1}/2}q_{\alpha,\beta}\left(1,(s-r)^{-\frac{\alpha}{2}}z\right)\right|\left| g(r,y-z)\right| _{H}dz\right|^{2}dy\\
& \leq\left(\int_{|z|\leq4d}\left[\int_{[-1,1]^{d}}\left| g(r,y-z)\right| _{H}^{2}dy\right]^{1/2}\left|(-\Delta)^{{c_1}/2}q_{\alpha,\beta}\left(1,(s-r)^{-\frac{\alpha}{2}}z\right)\right|dz\right)^{2}\\
& \leq\left(\int_{|z|\leq4d}\left[\int_{B_{5d}(0)}\left| g(r,y)\right| _{H}^{2}dy\right]^{1/2}\left|(-\Delta)^{{c_1}/2}q_{\alpha,\beta}\left(1,(s-r)^{-\frac{\alpha}{2}}z\right)\right|dz\right)^{2}\\
& \leq N\bM_x\left| g\right| _{H}^{2}(r,x)\left(\int_{|z|\leq4d}\left|(-\Delta)^{{c_1}/2}q_{\alpha,\beta}\left(1,(s-r)^{-\frac{\alpha}{2}}z\right)\right|dz\right)^{2}\\
& \leq N(s-r)^{\alpha(d+\hat{c}-2)}\bM_x\left| g\right| _{H}^{2}(r,x),
\end{align*}
where $\hat{c}\in(1,2)$ if ${c_1}=1$ and $d=1$, and otherwise
$\hat{c}={c_1}$. Since $|s-r|\sim|r|$ for $r<-3^{\frac{2}{\alpha}}$
and $-2^{\frac{2}{\alpha}}<s<0$, we have
\begin{align*}
\int_{Q_{0}} & \left|\mathcal{T}g(s,y)\right|^{2}dsdy
=\int_{-2^{\frac{2}{\alpha}}}^{0}\int_{[-1,1]^{d}}\left|\mathcal{T}g(s,y)\right|^{2}dsdy\\
& \leq N\int_{-2^{\frac{2}{\alpha}}}^{0}\int_{-\infty}^{-3^{\frac{2}{\alpha}}}(s-r)^{\alpha(\hat{c}-2)-1}\bM_x\left| g\right| _{H}^{2}(r,x)dr ds\\
& \leq N\int_{-\infty}^{-3^{\frac{2}{\alpha}}}\bM_x\left| g\right| _{H}^{2}(r,x)\frac{dr}{|r|^{\alpha(2-\hat{c})+1}}\\
& \leq N\int_{-\infty}^{-3^{\frac{2}{\alpha}}}\left(\int_{-r}^{0}\bM_x\left| g\right| _{H}^{2}(s,x)ds\right)\frac{dr}{|r|^{\alpha(2-\hat{c})+2}}\\
& \leq N\bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x)\int_{3^{\frac{2}{\alpha}}}^{\infty}\frac{dr}{r^{\alpha(2-\hat{c})+1}}\leq N\bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x).
\end{align*}
The lemma is proved.
\end{proof}
\begin{lem}
\label{lem:forth lemma}
Let $g\in C_{c}^{\infty}(\mathbb{R}^{d+1};H)$
and assume that $g(t,x)=0$ outside of $(-\infty,-3^{\frac{2}{\alpha}})\times B_{2d}^{c}$.
Then for any $(t,x)\in Q_{0}$,
$$
\int_{Q_{0}}\int_{Q_{0}}\left|\mathcal{T}g(s,y)-\mathcal{T}g(r,z)\right|^{2}dsdydrdz\leq N\bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x),
$$
where $N=N(d,\alpha,\beta)$.
\end{lem}
\begin{proof}
Due to Poincar\'e's inequality, it is enough to show
\begin{equation}
\int_{Q_{0}}\left(\left|\frac{\partial}{\partial s}\mathcal{T}g\right|^{2}+\left|D_{y}\mathcal{T}g\right|^{2}\right)dsdy\leq N\bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x).
\label{eq:3-1}
\end{equation}
Because of the similarity, we only prove
\begin{equation}
\label{eq:3-1}
\int_{Q_{0}} \left|D_{y}\mathcal{T}g\right|^{2} dsdy\leq N\bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x).
\end{equation}
Note that since $g(s,\cdot)=0$ for $s\geq-3^{\frac{2}{\alpha}}$,
\begin{align*}
D_{x}\mathcal{T}g(t,x) & =D_{x}\left[\int_{-\infty}^{-3^{\frac{2}{\alpha}}}\left| (-\Delta)^{{c_1}/2}T_{t-s}^{\alpha,\beta}g(s,\cdot)(x)\right| _{H}^{2}ds\right]^{1/2}\\
& \leq\left[\int_{-\infty}^{-3^{\frac{2}{\alpha}}}\left| D_{x}(-\Delta)^{{c_1}/2}T_{t-s}^{\alpha,\beta}g(s,\cdot)(x)\right| _{H}^{2}ds\right]^{1/2},
\end{align*}
where the above inequality is from Minkowski's inequality.
Recall \eqref{eq:q gamma scaling prop}. Thus for any $(s,y) \in Q_0$,
\begin{align*}
& \left|D_{y}\mathcal{T}g(s,y)\right|^{2}\\
& \leq\int_{-\infty}^{-3^{\frac{2}{\alpha}}}\left| D_{y}(-\Delta)^{{c_1}/2}T_{s-r}^{\alpha,\beta}g(r,\cdot)(y)\right| _{H}^{2}dr\\
& =\int_{-\infty}^{-3^{\frac{2}{\alpha}}}\left| (s-r)^{-\frac{\alpha d}{2}-\frac{1}{2}-\frac{\alpha}{2}}\int_{\mathbb{R}^{d}}D_{x}(-\Delta)^{{c_1}/2}q_{\alpha,\beta}\left(1,(s-r)^{-\frac{\alpha}{2}}z\right)g(r,y-z)dz\right| _{H}^{2}dr\\
& \leq\int_{-\infty}^{-3^{\frac{2}{\alpha}}}(s-r)^{-\alpha d-1-\alpha} \\
&\qquad \qquad \qquad \qquad \times \left[\int_{\mathbb{R}^{d}}\left|D_{x}(-\Delta)^{{c_1}/2}q_{\alpha,\beta}\left(1,(s-r)^{-\frac{\alpha}{2}}z\right)\right|\left| g(r,y-z)\right| _{H}dz\right]^{2}dr.
\end{align*}
Since $g(r,y-z)=0$ if $|z|\leq d$ and $y\in[-1,1]^{d}$,
\begin{align*}
& \int_{Q_{0}}|D_{y}\mathcal{T}g(s,y)|^{2}dsdy\\
& \leq\int_{Q_{0}}\int_{-\infty}^{-4^{\frac{2}{\alpha}}}(s-r)^{-\alpha(d+1)-1}\\
&\qquad \times\left[\int_{|z|\geq d}\left|D_{x}(-\Delta)^{{c_1}/2}q_{\alpha,\beta}\left(1,(s-r)^{-\frac{\alpha}{2}}z\right)\right|\left| g(r,y-z)\right| _{H}dz\right]^{2}drdsdy.
\end{align*}
Let $(t,x) \in Q_0$.
By using (\ref{eq:integration by parts}) and Lemma \ref{prop:kernel esti. of q}(v),
\begin{align*}
& \int_{|z|\geq d}\left|D_{x}(-\Delta)^{{c_1}/2}q_{\alpha,\beta}\left(1,(s-r)^{-\frac{\alpha}{2}}z\right)\right|\left| g(r,y-z)\right| _{H}dz\\
& \leq N(s-r)^{-\frac{\alpha}{2}}\int_{d}^{\infty}\left((s-r)^{-\frac{\alpha}{2}}\rho\right)^{-d-{c_1} - \varepsilon} \left(\int_{B_{\rho}(y)}\left| g(r,z)\right| _{H}dz\right)d\rho\\
& \leq N(s-r)^{\frac{\alpha }{2}(d+{c_1} + \varepsilon -1)}\bM_x\left| g(r,x)\right| _{H},
\end{align*}
where $\varepsilon \in [0,2]$ is taken so that ${c_1} + \varepsilon \in (1,2)$.
Therefore,
\begin{align*}
\int_{Q_{0}}\left|D_{y}\mathcal{T}g(s,y)\right|^{2}dsdy & \leq N\int_{-2^{\frac{2}{\alpha}}}^{0}\left[\int_{-\infty}^{-3^{\frac{2}{\alpha}}}(s-r)^{\alpha(c_1+\varepsilon -2)-1}\bM_x\left| g(r,x)\right| _{H}^{2}dr\right]ds\\
& \leq N\int_{-\infty}^{-3^{\frac{2}{\alpha}}}\left(\int_{-r}^{0}\bM_x\left| g(r,x)\right| _{H}^{2}ds\right)\frac{dr}{|r|^{\alpha(c_1+\varepsilon -2)-2}}\\
& \leq N\bM_{t}\bM_x\left| g(t,x)\right| _{H}^{2}\int_{3^{\frac{2}{\alpha}}}^{\infty}\frac{dr}{r{}^{\alpha(c_1+\varepsilon -2)-1}}\\
& \leq\bM_{t}\bM_x\left| g(t,x)\right| _{H}^{2}.
\end{align*}
Thus (\ref{eq:3-1}) and the lemma are proved.
\end{proof}
For a measurable function $h(t,x)$ on $\mathbb{R}^{d+1},$ we define
the sharp function
$$
h^{\#}(t,x)=\sup_{Q}\aint_{Q}\left|h(r,z)-h_{Q}\right|drdz,
$$
where
$$
h_{Q}=\aint_{Q}h(s,y)dyds
$$
and the supremum is taken over all $Q\subset\mathbb{R}^{d+1}$ containing
$(t,x)$ of the form
\begin{align*}
Q
&=Q_{R}(s,y), \quad R>0 \\
&=(s-R^{\frac{2}{\alpha}}/2,s+R^{\frac{2}{\alpha}}/2)\times(y^{1}-R/2,y^{1}+R/2)\times\cdots\times(y^{d}-R/2,y^{d}+R/2).
\end{align*}
By Fefferman-Stein theorem,
\begin{equation}
\label{fs}
\|h^{\#}\|_{L_p(\bR^{d+1})}\leq N \|h\|_{L_p(\bR^{d+1})},
\quad \quad p>1.
\end{equation}
Also note that for any $c\in \bR$,
\begin{eqnarray}
\nonumber
\aint_{Q}\left|h(r,z)-h_{Q}\right|^2drdz&=&\aint_{Q}\left|\aint_Q(h(r,z)-h(s,y))dsdy\right|^2drdz\\
&\leq&4 \aint_Q |h(r,z)-c|^2 drdz. \label{4times}
\end{eqnarray}
\begin{proof}[{\bf{Proof of Theorem \ref{thm:L-P}}}]
If $p=2$, (\ref{eq:L-P ineq}) follows from Lemma \ref{lem:L2 result}.
Hence we assume $p>2$.
First we prove for each $Q=Q_{R}(s,y)$ and $(t,x)\in Q$,
\begin{equation}
\aint_{Q}|\mathcal{T}g-\left(\mathcal{T}g\right)_{Q}|^{2}dsdy\leq N\bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x).\label{eq:target in KLP}
\end{equation}
Note that for any $h_{0}\in\mathbb{R}$ and $h\in\mathbb{R}^{d}$,
\begin{align*}
&\mathcal{T}g(t-h_{0},x-h) \\
& =\left[\int_{-\infty}^{t-h_{0}}\left| (-\Delta)^{{c_1}/2}T_{t-h_{0}-s}^{\alpha,\beta}g(s,\cdot)(x-h)\right| _{H}^{2}ds\right]^{1/2}\\
& =\left[\int_{-\infty}^{t-h_{0}}\left| (-\Delta)^{{c_1}/2}\int_{\mathbb{R}^{d}}q_{\alpha,\beta}(t-h_{0}-s,x-h-y)g(s,y)dy\right| _{H}^{2}ds\right]^{1/2}\\
& =\left[\int_{-\infty}^{t}\left| (-\Delta)^{{c_1}/2}\int_{\mathbb{R}^{d}}q_{\alpha,\beta}(t-s,x-y)\bar{g}(s,y)dy\right| _{H}^{2}ds\right]^{1/2}\\
& =\mathcal{T}\bar{g}(t,x)
\end{align*}
where $\bar{g}(s,y):=g(s-h_{0},y-h)$. This shows that to prove (\ref{eq:target in KLP}) we may assume $(s+R^{\frac{2}{\alpha}},y)=(0,0)$.
Also, due to (\ref{8.20.3}) (or (\ref{8.20.4})),
$$
\mathcal{T}g(c^{\frac{2}{\alpha}}\cdot,c\cdot)(t,x)=\mathcal{T}g(c^{\frac{2}{\alpha}}t,cx).
$$
Since dilations do not affect averages, it suffices to prove \eqref{eq:target in KLP} with $R=2$, i.e.
$$Q=Q_{0}=[-2^{\frac{2}{\alpha}},0]\times[-1,1]^{d}.
$$
Now we take a function $\zeta\in C_{c}^{\infty}$
such that $\zeta=1$ on $[-3^{\frac{2}{\alpha}},3^{\frac{2}{\alpha}}]$,
$\zeta=0$ outside of $[-4^{\frac{2}{\alpha}},4^{\frac{2}{\alpha}}]$,
and $0\leq\zeta\leq1$. We also choose a function $\eta\in C_{c}^{\infty}(\mathbb{R}^{d})$
such that $\eta=1$ on $B_{2d}$, $\eta=0$ outside of $B_{3d}$,
and $0\leq\eta\leq1$. Set
$$
g_{1}(t,x)=g\zeta,\quad g_{2}=g(1-\zeta)\eta,\quad g_{3}=g(1-\zeta)(1-\eta).
$$
Observe that $g=g_{1}+g_{2}+g_{3}$ and
$$
(-\Delta)^{{c_1}/2}T_{t-s}^{\alpha,\beta}g_{1}(s,y)=\zeta(s)(-\Delta)^{{c_1}/2}T_{t-s}^{\alpha,\beta}g(s,y),
$$
\begin{equation}
\label{eqn 8.20.5}
\mathcal{T}g\leq\mathcal{T}g_{1}+\mathcal{T}(g_{2}+g_{3}),
\end{equation}
\begin{equation}
\label{eqn 8.20.7}
\cT g_3\leq \mathcal{T}(g_{2}+g_{3})\leq\mathcal{T}g.
\end{equation}
(\ref{eqn 8.20.5}) is because $\cT$ is sublinear (see (\ref{sublinear})), and (\ref{eqn 8.20.7}) comes
from the facts $g_3=(1-\eta)(g_2+g_3)$, $g_2+g_3=(1-\zeta)g$, $|1-\eta(s)|\leq 1$, and $|1-\zeta(s)|\leq 1$.
Hence for any constant
$c$,
\begin{equation}
\label{eqn indeed1}
|\mathcal{T}g-c|\leq|\mathcal{T}g_{1}|+|\mathcal{T}(g_{2}+g_{3})-c|
\end{equation}
and
\begin{equation}
\label{eqn indeed2}
\left|\mathcal{T}(g_{2}+g_{3})-c\right|\leq\left|\mathcal{T}g_{2}\right|+\left|\mathcal{T}g_{3}-c\right|.
\end{equation}
Indeed, (\ref{eqn indeed1}) is from (\ref{eqn 8.20.5}) if $c\leq \cT g$, and if $c>\cT g$ then it follows from $\cT(g_2+g_3)\leq \cT g$. Similarly, (\ref{eqn indeed2}) is obvious if $c\leq \cT(g_2+g_3)$, and $c>\cT(g_2+g_3)$ we use $\cT g_3 \leq \cT(g_2+g_3)$.
Therefore, for any $c\in \bR$,
\begin{align*}
\left|\mathcal{T}g(s,y)-c\right| & \leq\left|\mathcal{T}g_{1}(s,y)\right|+\left|\mathcal{T}(g_{2}+g_{3})(s,y)-c\right|\\
& \leq\left|\mathcal{T}g_{1}(s,y)\right|+\left|\mathcal{T}g_{2}(s,y)\right|+\left|\mathcal{T}g_{3}(s,y)-c\right|,
\end{align*}
and by (\ref{4times})
\begin{align*}
\aint_{Q_0} & \left|\mathcal{T}g(s,y)-\left(\mathcal{T}g\right)_{Q_0}\right|^{2}dsdy\leq4\aint_{Q_0}\left|\mathcal{T}g(s,y)-c\right|^{2}dsdy\\
\leq & 16\aint_{Q_0}\left|\mathcal{T}g_{1}(s,y)\right|^{2}dsdy+16\aint_{Q_0}\left|\mathcal{T}g_{2}(s,y)\right|^{2}dsdy+16\aint_{Q_0}\left|\mathcal{T}g_{3}(s,y)-c\right|^{2}dyds.
\end{align*}
Note $g_1$ and $g_2$ satisfy the conditions of Lemma \ref{lem:second lemma} and \ref{lem:third lemma}, respectively, and thus
\begin{eqnarray*}
&&\aint_{Q_0}\left|\mathcal{T}g_{1}(s,y)\right|^{2}dsdy+\aint_{Q_0}\left|\mathcal{T}g_{2}(s,y)\right|^{2}dsdy\\
&&\leq N \left( \bM_{t}\bM_x\left| g_1\right| _{H}^{2}(t,x)+ \bM_{t}\bM_x\left| g_2\right| _{H}^{2}(t,x)\right)\leq N \bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x).
\end{eqnarray*}
The second inequality above is due to $|g_i|\leq |g|$ ($i=1,2,3$). Take $c=\left(\mathcal{T}g_{3}\right)_{Q_0}$ and note that
\begin{equation}
\label{eqn 8.20.8}
\aint_{Q_0}\left|\mathcal{T}g_{3}(s,y)-\left(\mathcal{\mathcal{T}}g_{3}\right)_{Q_0}\right|^{2}dsdy\leq \aint_{Q_0}\aint_{Q_0}\left|\mathcal{T}g_{3}(s,y)-\mathcal{T}g_{3}(r,z)\right|^{2}drdzdsdy.
\end{equation}
Note also, on $Q_0$, $\cT g_3$ does not depend on the values of
$g_3(t,x)$ for $t >0$. Hence the above two integrals do not change
if we replace $g_3$ by $g_3 \xi$, where $\xi\in C^{\infty}(\bR)$ so
that $0\leq \xi \leq 1$, $\xi=1$ for $t\leq 1$, and $\xi=0$ for
$t\geq 2^{2/{\alpha}}$. Now it is easy to check that $g_3\xi$
satisfies the assumptions of Lemma \ref{lem:forth lemma}, and
therefore the right hand side of (\ref{eqn 8.20.8}) is controlled by
$$
\bM_{t}\bM_x\left| g_3\xi\right| _{H}^{2}(t,x)\leq \bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x).
$$
Hence (\ref{eq:target in KLP}) is finally proved.
We continue the proof of the theorem.
By (\ref{eq:target in KLP}) and Jensen's inequality
$$
(\cT g)^{\#}(t,x)\leq N \left(\bM_{t}\bM_x\left| g\right| _{H}^{2}(t,x) \right)^{1/2}.
$$
Therefore by Fefferman-Stein theorem (\cite[Theorem 4.2.2]{Stein1993}) and
Hardy-Littlewood maximal theorem (\cite[Theorem 1.3.1]{Stein1993}),
\begin{eqnarray*}
\|\cT g\|_{L_p(\bR^{d+1})}&\leq& N \|(\cT g)^{\#}\|_{L_p(\bR^{d+1})}\\
&\leq& N \|\bM_{t}\bM_x\left| g\right| _{H}^{2}\|^{1/2}_{L_{p/2}(\bR^{d+1})}\\
&\leq& N \|\bM_x|g|^2_H\|^{1/2}_{L_{p/2}(\bR^{d+1})}\leq N \||g|_H\|_{L_p(\bR^{d+1})}.
\end{eqnarray*}
This proves the theorem if $T=\infty$. Note that if $T<\infty$ the left hand side of (\ref{thm:L-P}) does not depend on the value of $g$ for $t\geq T$. Take $\tilde \xi \in C^{\infty}(\bR)$ such that $0\leq \tilde \xi \leq 1$, $\tilde \xi=1$ for $t\leq T$ and $\tilde \xi=0$ for $t\geq T+\varepsilon$, $\varepsilon>0$. Then it is enough to apply the result for $T=\infty$ with $g\tilde \xi$. Since $\varepsilon>0$ is arbitrary the theorem is proved.
\end{proof}
\mysection{Model Equation}
\label{section 4}
Let $\alpha\in (0,2)$ and $\beta \in \left(-\infty, \alpha+\frac{1}{2} \right)$. In this section we obtain the uniqueness, existence, and sharp estimate of strong solutions to the model equation
\begin{equation}
\label{eq:model equation}
\partial_{t}^{\alpha}u(t,x)=\Delta u (t,x)+\partial_{t}^{\beta}\int_{0}^{t}g^{k}(s,x)dw_{s}^{k}, \quad t>0
\end{equation}
with the zero initial condition $u(0,x)=0$ (additionally $\partial_{t}u(0,x)=0$ if $\alpha>1$).
The following lemma is used to estimate solutions to the equation when $\beta <1/2$.
\begin{lem}
\label{lem:pde approach}
Let $\gamma\in\mathbb{R}$, $p>2$, $\beta<\frac{1}{2}$, and $g\in\mathbb{H}_{p}^{\gamma}(T,l_{2})$. Then
for any $t \in [0,T]$,
$$
\mathbb{E}\int_{0}^{t}\left\Vert\sum_{k=1}^{\infty} \partial_{t}^{\beta}\int_{0}^{r}g^{k}(s,\cdot)dw_{s}^{k}\right\Vert _{H_{p}^{\gamma}}^{p}dr
\leq N(d,p,\beta,T) I_{t}^{1-2\beta}\|g\|_{\mathbb{H}_{p}^{\gamma}(\cdot,l_{2})}^{p}(t).
$$
In particular,
$$
\mathbb{E} \int_{0}^{t}\left\Vert\sum_{k=1}^{\infty} \partial_{t}^{\beta}\int_{0}^{r}g^{k}(s,\cdot)dw_{s}^{k}\right\Vert _{H_{p}^{\gamma}}^{p}dr
\leq N \|g\|_{\mathbb{H}_{p}^{\gamma}(t,l_{2})}^{p}.
$$
\end{lem}
\begin{proof}
Due to the isometry $(I-\Delta)^{\gamma/2}:H^{\gamma}_p \to L_p$, we only need to prove the case $\gamma=0$.
By Lemma \ref{lem:s int} (iii),
\begin{align*}
\partial_{t}^{\beta}\left(\sum_{k=1}^{\infty}\int_{0}^{r}g^{k}(s,x)dw_{s}^{k}\right) =\frac{1}{\Gamma(\alpha)}\sum_{k=1}^{\infty}\int_{0}^{t}(t-s)^{-\beta}g^{k}(s,x)dw_{s}^{k},
\end{align*}
for almost all $t\leq T$ (a.s.). By the Burkholder-Davis-Gundy
inequality and the H\"{o}lder inequality, for all $t\leq T$,
\begin{align*}
\mathbb{E}\int_{0}^{t}&\left\Vert \frac{1}{\Gamma(\alpha)}\sum_{k=1}^{\infty}\int_{0}^{r}(r-s)^{-\beta}g^{k}(s,\cdot)dw_{s}^{k}\right\Vert _{L_{p}}^{p}dr\\
& \leq N\mathbb{E}\int_{\mathbb{R}^{d}}\int_{0}^{t}\left(\int_{0}^{r}(r-s)^{-2\beta}\left|g(s,x)\right|_{l_{2}}^{2}ds\right)^{p/2}drdx\\
& \leq N\mathbb{E}\int_{\mathbb{R}^{d}}\int_{0}^{t}\left(\int_{0}^{r}(r-s)^{-2\beta(\frac{2}{p} + \frac{p-2}{p})}\left|g(s,x)\right|_{l_{2}}^{2}ds\right)^{p/2}drdx\\
& \leq N\mathbb{E}\int_{\mathbb{R}^{d}}\int_{0}^{t}\int_{0}^{r}(r-s)^{-2\beta}\left|g(s,x)\right|_{l_{2}}^{p}dsdrdx \\
& = N \int_{0}^{t}(t-s)^{-2\beta}\|g\|_{\mathbb{L}_{p}(s,l_{2})}^{p}ds = N I_{t}^{1-2\beta}\|g\|_{\mathbb{L}_{p}(\cdot,l_{2})}^{p}(t).
\end{align*}
The lemma is proved.
\end{proof}
A version of Lemma \ref{lem:solution representation} can be found in \cite{CKK} for $p=2$ and $\alpha, \beta\in (0,1)$. But solutions spaces are slightly different, and our proof is more rigorous.
\begin{lem}
\label{lem:solution representation}
Let $g\in\mathbb{H}_{0}^{\infty}(T,l_{2})$ and
define
\begin{align}
\label{sto sol re}
u(t,x) :=\sum_{k=1}^{\infty}\int_{0}^{t}\int_{\mathbb{R}^{d}}q_{\alpha,\beta}(t-s,x-y)g^{k}(s,y)dydw_{s}^{k}.
\end{align}
Then $u\in \cH^2_{p}(T)$ and satisfies \eqref{eq:model equation} with the zero initial conditon in the sense of distributions
(see Definition \ref{def:solution space}).
\end{lem}
\begin{proof}
Let $(t,x) \in [0,T] \times \bR^d$. Set
$$
v(t,x):=\sum_{k=1}^{\infty}\int_{0}^{t}g^{k}(s,x)dw_{s}^{k},\quad w(t,x):=I_t^{\alpha-\beta}v(t,x),
$$
where $I_t^{\alpha-\beta}v=D_{t}^{\beta-\alpha}v$ if $\alpha <
\beta$. Note that since $g\in \bH^{\infty}_0(T,l_2)$, by the
Kolmogorov continuity theorem
$$
v\in C^{1/2-\varepsilon}([0,T],H^m_p)
$$ for any $\varepsilon>0$ and $m$.
Thus $w\in C^{\delta}([0,T],H_p^m)$ for some $\delta>0$ (see (\ref{extra con})).
By Fubini's theorem if $\alpha\geq \beta$ and fractional integration by parts (e.g. \cite[Lemma 2.3]{CKK}) if $\alpha < \beta$,
$$
\int^t_0 I_s^{\alpha-\beta} p(s,x-y)(\int_0^{t-s} g^k(r,y)dw^k_r) ds=\int_{0}^{t} p(t-s,x-y)I_s^{\alpha-\beta} \int_0^s g^{k}(r,y)dw_{r}^{k}ds.
$$
Here $I^{\alpha}_s p(s,x-y)$ and $I^{\alpha-\beta}_s \int^s_0 g^k(r,y)dw^k_r$ are used to denote $(I^{\alpha-\beta}_t p(\cdot,x-y))(s)$ and
$(I^{\alpha-\beta}_t \int^{\cdot}_0 g^k(r,y)dw^k_r)(s)$, respectively. Thus, using the stochastic Fubini theorem (see \cite[Lemma 2.7]{krylov2011ito}) we get, for each $(t,x)$, $(a.s.)$
\begin{align*}
\int_0^t u(s,x)ds
&= \sum_{k=1}^\infty \int_{\mathbb{R}^{d}} \int_{0}^{t}I_s^{\alpha-\beta} p(s,x-y)\int_0^{t-s} g^k(r,y)dw^k_r ds dy \\
&=\sum_{k=1}^\infty \int_{\mathbb{R}^{d}} \int_{0}^{t} p(t-s,x-y)I_s^{\alpha-\beta} \int_0^s g^{k}(r,y)dw_{r}^{k}dy ds \\
&= \int_{0}^{t}\int_{\mathbb{R}^{d}} p(t-s,x-y)w(s,y)dy ds.
\end{align*}
Due to
the continuity with respect to $t$, for each $x$ we get
$$
\int_0^t u(s,x)ds= \int_{0}^{t}\int_{\mathbb{R}^{d}} p(t-s,x-y)w(s,y)dy ds, \quad \forall t \leq T~\text{$(a.s.)$}
$$
and therefore $(a.s.)$
\begin{align}
\label{u partial tp}
u(t,x) &=\frac{\partial}{\partial t}\int_{0}^{t}
\int_{\mathbb{R}^{d}} p(t-s,x-y) w(s,y)dy ds,~\quad (a.e.) \,\,
t\leq T.
\end{align}
In other words, the above equality holds $(a.e.)$ on $\Omega \times [0,T] \times \bR^d$.
Next we claim that
\begin{align}
\label{rep claim}
u(t,x)-w(t,x)= \int_{0}^{t}\int_{\mathbb{R}^{d}}q(t-s,x-y) \Delta
w(s,y)dy ds
\end{align}
$(a.e.)$ on $\Omega \times [0,T] \times \bR^d$.
By the definition of the differentiation, for each $(\omega,t,x)$,
\begin{eqnarray*}
&& \frac{\partial}{\partial t} \int_{0}^{t} \int_{\mathbb{R}^{d}}p(t-s,x-y)w(s,y)dy ds \\
&=&\lim_{h \downarrow 0}\frac{1}{h}\int_{t}^{t+h} \int_{\mathbb{R}^{d}}\left(p(t+h-s,x-y) \right)w(s,y)dy ds \\
&&+\lim_{h\downarrow 0}\int_{0}^{t} \int_{\mathbb{R}^{d}}\left[\frac{p(t+h-s,x-y)-p(t-s,x-y)}{h}\right]w(s,y)dy ds.
\end{eqnarray*}
By the mean value theorem, the integration by parts, and Lemma \ref{prop:kernel esti. of q}(i) and (ii),
\begin{eqnarray*}
&&\lim_{h\downarrow 0}\int_{0}^{t} \int_{\mathbb{R}^{d}}\left[\frac{p(t+h-s,x-y)-p(t-s,x-y)}{h}\right]w(s,y)dy ds\\
&=&\lim_{h\downarrow 0}\int_{0}^{t} \int_{\mathbb{R}^{d}}\frac{\partial p}{\partial t}(t+\theta h-s, x-y)w(s,y)dy ds, \quad \theta\in (0,1)\\
&=&\lim_{h\downarrow 0}\int_{0}^{t} \int_{\mathbb{R}^{d}} q(t+\theta h-s,x-y) \Delta w(s,y)dy ds\\
&=&\int^t_0 \int_{\mathbb{R}^{d}} q(t-s) \Delta w(s,y)dy ds.
\end{eqnarray*}
For the last equality above we used the $L_1$-continuity of the integrable function \cite[Theorem 9.5]{Ru}, which implies that
for any $f \in L_1([0,t+\varepsilon])$, where $\varepsilon>0$, it holds that
$\lim_{h \to 0} \int_0^t | f(s+h) -f(s)|dt =0$.
On the other hand, due to Corollary \ref{zero converge},
\begin{eqnarray*}
\lim_{h \downarrow 0}\frac{1}{h}\int_{t}^{t+h}
\int_{\mathbb{R}^{d}}\left(p(t+h-s,x-y) \right)w(s,y)dy ds =w(t,x).
\end{eqnarray*}
Thus (\ref{rep claim}) is proved due to (\ref{u partial tp}), and
from (\ref{rep claim}) it easily follows that $u$ has a
$H^2_p$-valued continuous version since $g \in \bH_0^\infty(T,l_2)$.
It only remains to show that $u$ satisfies (\ref{eq:model
equation}). By representation formula (\ref{rep claim}), it follows
that $u-w\in \fH^{\alpha,2}_{p,0}(T)$ $(a.s.)$, and
\begin{eqnarray}
\label{det sol rep}
\partial_t^\alpha (u-w)
&=& \Delta (u-w) + \Delta w(t,x) \\
& =&\Delta u \nonumber
\end{eqnarray}
in $\fL_p(T)$. See Remark \ref{rem deterministic} for spaces
$\fH^{\alpha,2}_{p,0}(T)$ and $\fL_p(T)$. Actually in (\cite[Lemma
3.5]{KKL2014}), it is proved that (\ref{rep claim}) gives the unique
solution to (\ref{det sol rep}) in the space
$\fH^{\alpha,2}_{p,0}(T)$ if $\Delta w$ is sufficiently smooth.
However one can easily check that this representation holds even if
$\Delta w \in L_p([0,T] \times\bR^{d})$ by using an approximation argument.
It follows from (\ref{eqn 7.15.2}) and Remark \ref{rem
deterministic} that for any $\phi\in C^{\infty}_c(\bR^d)$, $(a.s.)$
$$
(u(t)-w(t),\phi)=I^{\alpha}(\Delta u,\phi), \quad
\text{$(a.e.)$}\,\, t\leq T.
$$
Taking $(w(t),\phi)$ to the right hand side of the equality and
using the continuity of $u$ with respect to $t$,
we get
$$
(u(t),\phi)=I^{\alpha}_t(\Delta u, \phi) +I^{\alpha-\beta}_t\int^t_0 (g^k,\phi)dw^k_s, \quad \, \forall t\leq T~ \text{$(a.s.)$}.
$$
Therefore $u$ is a solution to (\ref{eq:model equation}) in the
sense of distributions
because $u$ itself is an $H^2_p$-valued continuous process.
The lemma is proved.
\end{proof}
Recall, for $\kappa\in (0,1)$,
$$
c'_0=c'_0(\kappa)=\frac{(2\beta-1)_+}{\alpha} +\kappa 1_{\beta=1/2} \quad \in [0,2).
$$
\begin{thm}
\label{thm:model eqn}
Let $\gamma\in\mathbb{R}$ and $p\geq 2$. Suppose $g\in\mathbb{H}_{p}^{\gamma+c'_{0}}(T,l_{2})$ for some $\kappa>0$. Then,
equation \eqref{eq:model equation} with zero initial condition
has a unique solution $u\in\mathcal{H}_{p}^{\gamma+2}(T)$ in the sense
of distributions, and for this solution
we have
\begin{equation}
\label{eq:model a priori}
\|u\|_{\cH_{p}^{\gamma+2}(T)}\leq N\|g\|_{\mathbb{H}_{p}^{\gamma+c'_{0}}(T,l_{2})},
\end{equation}
where $N=N(d,p,\alpha,\beta ,\kappa, T)$.
Furthermore, if $\beta > 1/2$ then
\begin{equation}
\label{model t ind}
\|u_{xx}\|_{\mathbb{H}_{p}^{\gamma}(T)}\leq N\| \Delta^{c'_0/2}g\|_{\mathbb{H}_{p}^{\gamma}(T,l_{2})},
\end{equation}
where $N=N(d,p,\alpha,\beta)$ is independent of $T$.
\end{thm}
\begin{proof}
Due to the isometry $(I-\Delta)^{\gamma /2}:H^{\gamma}_p \to L_p$, we only need to prove the case $\gamma=0$.
Recall that as discussed in Remark \ref{rem deterministic} for the
deterministic case, our sense of solutions introduced in Definition
\ref{def:solution space} coincides with the one in \cite{KKL2014}.
Therefore the uniqueness result easily follows from the
deterministic result (\cite[Theorem 2.9]{KKL2014}, cf.
\cite{zacher2005maximal}). Therefore it is sufficient to prove the
existence of the solution and estimates (\ref{eq:model a priori})
and (\ref{model t ind}).
{\bf{Step 1}}. First, assume $g\in\mathbb{H}_{0}^{\infty}(T,l_{2})$.
Define
$$
u(t,x)=\sum_{k=1}^{\infty}\int_{0}^{t}\int_{\mathbb{R}^{d}}q_{\alpha,\beta}(t-s,y)g^{k}(s,x-y)dydw_{s}^{k}.
$$
Then by Lemma \ref{lem:solution representation},
$u\in \cH^2_{p}(T)$ is a solution to equation \eqref{eq:model equation} with the zero initial condition.
Thus we only need to prove the estimates. We divide the proof acoording to the range of $\beta$.
\textbf{Case 1:} $\beta > \frac{1}{2}$.
Due to the inequality (e.g. p.41 of \cite{Krylov2008}).
$$
\|u_{xx}\|_{\bL_{p}(T)} \leq N \|\Delta u\|_{\bL_{p}(T)},
$$
to get \eqref{model t ind}, it suffices to show
\begin{align}
\label{delta u}
\|\Delta u\|_{\bL_{p}(T)}\leq N\| \Delta^{c'_0/2}g\|_{\bL_p(T,l_{2})}.
\end{align}
Denote
$$
v=(-\Delta)^{c'_0 /2}u, \quad \bar{g}=(-\Delta)^{c'_0 /2}g .
$$
By the Burkholder-Davis-Gundy
inequality and Remark \ref{rem:extension of cT},
\begin{align*}
\|\Delta u\|^p_{\bL_p(T)}=\left\Vert (-\Delta)^{(2-c'_0)/2}v\right\Vert _{\mathbb{L}_{p}(T)}^{p}
& \leq N\mathbb{E}\int_{0}^{T}\int_{\mathbb{R}^{d}}\left|\mathcal{T}\bar{g}(t,x)\right|^{p}dxdt\\
& \leq N\mathbb{E}\int_{0}^{T}\int_{\mathbb{R}^{d}}|\bar{g}(t,x)|_{l_{2}}^{p}dxdt,
\end{align*}
where $N=N(d,p,\alpha,\beta)$.
Next we prove (\ref{eq:model a priori}).
By Lemma \ref{lem:solution space}(iv) and (\ref{delta u}),
\begin{align}
\notag
\|u\|^p _{\bL_{p}(T)}
&\leq N\int_{0}^{T}(T-s)^{\theta-1} \left(\|\Delta u\|^p_{\bL_p(s)}
+\|g\|^p_{\bL_p(s,l_2)} \right)ds \\
\notag
&\leq N\int_{0}^{T}(T-s)^{\theta-1} \|g\|^p_{\bH^{c'_0}_p(s,l_2)} ds\\
\label{mo u est}
&\leq N\|g\|^p_{\bH_p^{c_0'}(T,l_2)}\int^T_0(T-s)^{\theta-1}ds \leq N\|g\|^p_{\bH_p^{c_0'}(T,l_2)}.
\end{align}
Combining (\ref{model t ind}), (\ref{mo u est}), and (\ref {eq:7.2.1}), we get
(\ref{eq:model a priori}).
\textbf{Case 2:} $\beta<\frac{1}{2}$.
In this case, ${c_0'}=0$ and we apply the result of the deterministic
equation from \cite{KKL2014}. By Remarks \ref {rem 7.15}(ii) and \ref{rem deterministic}, $u$ satisfies
$$
\partial_{t}^{\alpha}u=\Delta u+\bar{f}
$$
in the sense of \cite[Definition 2.4]{KKL2014}, where
$$
\bar{f}(t)=\frac{1}{\Gamma(1-\beta)}\sum_k \int^t_0(t-s)^{-\beta}g^k(s)dw^k_s.
$$
Due to \cite[Theorem 2.9]{KKL2014} and Lemma \ref{lem:pde approach},
$$
\|u\|^p_{\bH^2_p(T)}\leq N\|\bar{f}\|^p_{\bL_p(T)}\leq N\|g\|^p_{\bL_p(T,l_2)},
$$
which togegher with (\ref {eq:7.2.1}) yields \eqref{eq:model a priori}.
\textbf{Case 3:} $\beta=\frac{1}{2}$.
Put $\delta=\frac{\kappa \alpha}{2}$. Write $\tilde{\beta}=\frac{1}{2}+\delta$ and define
$$
v(t,x)=\sum_{k=1}^{\infty}\int_{0}^{t}\int_{\mathbb{R}^{d}}q_{\alpha,\tilde{\beta}}(t-s,x-y)g^{k}(s,y)dydw_{s}^{k}.
$$
Since $0<\delta<\alpha$
and $\frac{1}{2}<\tilde{\beta}<2$, the result from Case 1 with ${c_0'}=(2 \tilde \beta -1)/\alpha =\kappa $
implies that $v\in \cH^2_{p}$ satisfies
\begin{align*}
\partial_{t}^{\alpha}v(t,x)=\Delta v(t,x)+\sum_{k=1}^{\infty}\partial_{t}^{\tilde{\beta}}\int_{0}^{t}g^{k}(s,x)dw_{s}^{k},
\end{align*}
with the zero initial condition and
$$
\left\Vert v\right\Vert _{\cH_{p}^{2}(T)}\leq N\left\Vert g\right\Vert _{\mathbb{H}_{p}^{c'_{0}}(T,l_{2})}.
$$
Since $I^{\delta}_tv$ satisfies \eqref{eq:model equation}, by the uniqueness of solutions, we conclude that $I_{t}^{\delta}v(t,x)=u(t,x)$.
Therefore,
\begin{align*}
\left\Vert u\right\Vert _{\cH_{p}^{2}(T)} & =\|I_{t}^{\delta}v\|_{\cH_{p}^{2}(T)} \leq N\left\Vert v\right\Vert _{\cH_{p}^{2}(T)}\leq N\left\Vert g\right\Vert _{\mathbb{H}_{p}^{c'_{0}}(T,l_{2}).}
\end{align*}
Thus the theorem is proved if $g\in\mathbb{H}_{0}^{\infty}(T,l_{2})$.
{\bf{Step 2}}.
For general $g\in\mathbb{H}_{p}^{{c_0'}}(T,l_{2})$, take a
sequence $g_{n}\in\mathbb{H}_{0}^{\infty}(T,l_{2})$ so that
$g_{n}\rightarrow g$ in $\mathbb{H}_{p}^{{c_0'}}(T,l_{2})$.
Define $u_{n}$ as the solution of equation \eqref{eq:model equation}
with $g_{n}$ in place of $g$.
Then
\begin{equation}
\|u_{n}\|_{\cH_{p}^{2}(T)}\leq N\|g_{n}\|_{\mathbb{H}_{2}^{{c_0'}}(T,l_{2})},\label{eq:model thm proof 1}
\end{equation}
\begin{equation}
\label{eq:model thm proof 2}
\|u_{n}-u_{m}\|_{\cH_{p}^{2}(T)}\leq N\|g_{n}-g_{m}\|_{\mathbb{H}_{p}^{{c_0'}}(T,l_{2})}.
\end{equation}
Thus, $u_{n}$ converges to $u$ in $\cH_{p}^{2}(T)$ and $u$ becomes a solution to equation \eqref{eq:model equation}. Indeed, to check $u$ is a solution, let $\phi \in \cS$ and then we have
$$
(\bI^{\Lambda-\alpha}_tu_n(t), \phi)
=I_{t}^{\Lambda}\left(\Delta u_n(t,\cdot),\phi\right)+\sum_{k=1}^{\infty}I_{t}^{\Lambda-\beta}\int_{0}^{t}\left(g^{k}_n(s,\cdot),\phi\right)dw_{s}^{k}, \quad \forall t\leq T.
$$
Taking the limit and using (\ref{eq:model thm proof 2}) we conclude that $I^{\Lambda-\alpha}u$ has a continuous version and therefore the above equality holds for all $t\leq T$ (a.s.) with $u$ and $g$ in place of $u_n$ and $g_n$ respectively.
The theorem is proved.
\end{proof}
\mysection{Proof of Theorem \ref{thm:main results non-div}}
\label{pf non-div thm}
First we introduce a version of method of continuity used in this article. Later we will take $L_0=\Delta$ and $\Lambda_0=0$.
\begin{lem}[Method of continuity]
Let $L_0$, $L_1$ be continuous operators from $\cH_p^{\gamma+2}(T)$ to $\bH_p^\gamma(T)$ and
$\Lambda_0$, $\Lambda_1$ be continuous operators from $\cH_p^{\gamma+2}(T)$ to $\bH_p^{\gamma+c_0'}(T,l_2)$.
For $\lambda \in [0,1]$ and $u \in \cH_p^{\gamma+2}(T)$, denote $L_\lambda u= \lambda L_1 u + (1-\lambda)L_0u$
and $\Lambda_\lambda u= \lambda \Lambda_1 u + (1-\lambda)\Lambda_0 u$.
Suppose that for any $f \in \bH_p^\gamma(T)$ and $g \in \bH_p^{\gamma+c_0'}(T,l_2)$ the equation
\begin{align*}
\partial_{t}^{\alpha}u
=L_0u+f+\partial_{t}^{\beta}\int_{0}^{t} (\Lambda^k_0 u+g^{k})dw_{s}^{k}
\end{align*}
with zero initial codition has a solution $u$ in $\cH_{p}^{\gamma+2}(T)$.
Also assume that if $u \in \cH_{p}^{\gamma+2}(T)$ has zero initial condition and satisfies (in the sense of distributions) the equation
\begin{equation}
\label{eqn 8.25.7}
\partial_{t}^{\alpha}u
=L_\lambda u+f+\partial_{t}^{\beta}\int_{0}^{t}(\Lambda^k_\lambda u +g^{k})dw_{s}^{k},
\end{equation}
then the following ``a priori estimate" holds:
\begin{align}
\label{824 a pr}
\|u\|_{\cH_p^{\gamma+2}(T)} \leq N_0 \left(\|f\|_{\bH_p^\gamma(T)} + \|g\|_{\bH_p^{\gamma+c_0'}(T,l_2)}\right),
\end{align}
where $N_0$ is independent of $\lambda$, $u$, $f$, and $g$.
Then for any $\lambda\in [0,1]$, $f \in \bH_p^{\gamma}(T)$, and
$g \in \bH_p^{\gamma+c_0'}(T,l_2)$ the equation
\begin{align}
\label{eqn 8.25.1}
\partial_{t}^{\alpha}u
=L_{\lambda} u+f+\partial_{t}^{\beta}\int_{0}^{t}(\Lambda^k_{\lambda} u +g^{k})dw_{s}^{k}
\end{align}
with zero initial condition has a unique solution $u$ in $\cH_{p}^{\gamma+2}(T)$.
\end{lem}
\begin{proof}
The uniqueness easily follows from (\ref{824 a pr}).
Let $J$ be the set of all $\lambda\in [0,1]$ for which equation (\ref{eqn 8.25.1}) has a solution in $\cH^{\gamma+2}_{p}(T)$ for any $f \in \bH_p^{\gamma}(T)$ and $g \in \bH_p^{\gamma+c_0'}(T,l_2)$. By the assumption $0\in J$. Thus to prove the lemma it suffices to show that there exists $\varepsilon>0$ depending only on $N_0$ and the boundedness of the operators $L_i$ and $\Lambda_i$ ($i=0,1$) such that $\lambda \in J$ whenever $\lambda_0\in J$ and $|\lambda-\lambda_0|<\varepsilon$.
Let $\lambda_0 \in [0,1]$ and $\lambda\in [0,1]$. Fix $u^0\in \cH^{\gamma+2}_{p}$.
By the assumption, we can inductively define $u^{n+1} \in \cH^{\gamma+2}_{p}(T)$ as the solution to
\begin{eqnarray}
\nonumber
\partial_t^\alpha u^{n+1}
&=&L_{\lambda_0}u^{n+1} +(-L_{\lambda_0}u^n +L_{\lambda}u^n+f)\\
&&+\partial_{t}^{\beta}\int_{0}^{t} (\Lambda_{\lambda_0} u^{n+1}+(-\Lambda_{\lambda_0} u^n+\Lambda_{\lambda} u^n+g^{k}) )dw_{s}^{k}.
\label{824 n eqn}
\end{eqnarray}
Note that for $u^{n+1}-u^n \in \cH^{\gamma+2}_{p}(T)$ satisfies
\begin{align*}
\partial_t^\alpha (u^{n+1} -u^n)
&=L_{\lambda_0}(u^{n+1}-u^n)+(\lambda-\lambda_0)(L_1-L_0)(u^n-u^{n-1}) \\
&~~~ +\partial_{t}^{\beta}\int_{0}^{t} \Lambda^k_{\lambda_0} (u^{n+1}-u^n)
+(\lambda-\lambda_0)(\Lambda_{1}-\Lambda_0)(u^n-u^{n-1}) dw_{s}^{k}.
\end{align*}
By a priori estmate (\ref{824 a pr}), we have
\begin{align*}
&\|u^{n+1}-u^n\|_{\cH_p^{\gamma+2}(T)} \\
&\leq N_0|\lambda-\lambda_0| (\|(L_1-L_0)(u^n-u^{n-1}))\|_{\bH^{\gamma}_p(T)}
+\|(\Lambda_1-\Lambda_0)(u^n-u^{n-1})\|_{\bH^{\gamma+c'_0}_p(T,l_2)})
\\
&\leq N|\lambda-\lambda_0| \|u^n-u^{n-1}\|_{\cH_p^{\gamma+2}(T)},
\end{align*}
where the second inequality is due to the continuity of operators
$L_0$, $L_1$, $\Lambda_0$, and $\Lambda_1$. Note that the constant
$N$ above does not depend on $\lambda$ and $\lambda_0$ as well. Thus
if $\varepsilon N<1/2$ and $|\lambda-\lambda_0|\leq \varepsilon$
then $u_n$ becomes a Cauchy sequence in $\cH^{\gamma+2}_{p,0}(T)$
and therefore the limit $u$ of $u^n$ becomes a solution to equation
(\ref{eqn 8.25.1}), which is easily checked by taking the limit in
(\ref{824 n eqn}). The lemma is proved.
\end{proof}
Next we present an estimate for a deterministic equation of non-divergence type. We use the space $\fH^{\alpha,\gamma+2}_{p,0}(T)$ introduced in Remark \ref{rem deterministic}.
\begin{lem}
\label{det11}
Let $a^{ij}$ be given as in (\ref{cond 8.26}), that is
\begin{equation}
\label{a}
a^{ij}(t,x)=\sum_{n=1}^{M_0}
a_{n}^{ij}(t,x)1_{(\tau_{n-1},\tau_n]}(t)
\end{equation}
where $\tau_n$ and $a^{ij}_n$ are non-random, and $a^{ij}$ satisfy (\ref{elliptic}) and (\ref{multiplier}) with the constants $\delta_0$ and $K_3$ given there. Then for any solution $u\in \fH^{\alpha,\gamma+2}_{p,0}(T)$ to the deterministic equation
\begin{equation}
\label{eqn 8.26.2}
\partial^{\alpha}_t u=a^{ij}u_{x^ix^j}+f
\end{equation}
in $\fH^{\gamma}_p(T)$,
it holds that
\begin{equation}
\label{apriori det}
\|u\|_{\fH^{\gamma+2}_p(T)}\leq N \|f\|_{\fH^{\gamma}_p(T)},
\end{equation}
where $N$ depends only on $\alpha$, $p$, $\gamma$, $\delta_0$, $K_3$, $T$, $M_0$, and the modulus of continuity of $a^{ij}_n$. In particular, $N$ depends on $M_0$ but independent of the choice of $\tau_1,\cdots, \tau_{M_0-1}$.
\end{lem}
\begin{proof}
If $\gamma=0$ then this lemma is proved in \cite[Theorem 2.9]{KKL2014} under the condition that $a^{ij}_n$ are uniformly continuous in $(t,x)$, but without the condition $|a^{ij}|_{B^{|\gamma|}}\leq K_3$. The proof for the case $\gamma \neq 0$ depends on the one for $\gamma=0$.
\vspace{3mm}
We divide the proof into several steps.
(Step 1). Assume that $a^{ij}$ are independent of $(t,x)$. In
this case (\ref{apriori det}) holds due to \cite[Theorem
2.9]{KKL2014} (or see \cite{zacher2005maximal,Za}) if $\gamma=0$.
For the case $\gamma \neq 0$ it is enough to apply the operator
$(1-\Delta)^{\gamma/2}$ to the equation.
We show that (\ref{apriori det}) leads to
\begin{equation}
\label{indT}
\|u_{xx}\|_{\fH^{\gamma}_p(T)}\leq N_0
\|f\|_{\fH^{\gamma}_p(T)},
\end{equation}
where $N_0=N_0(\alpha,p,\gamma,\delta_0)$, and thus independent of $T$. Obviously, to prove the independency of $T$ we only need to consider the case $\gamma=0$, and for this case, it is enough to notice that $v(t,x):=u(Tt,T^{\alpha/2}x)$ satisfies $\partial^{\alpha}_t v=a^{ij}v_{x^ix^j}+T^{\alpha}f(Tt,T^{\alpha/2}x)$ in
$\fL_p(1)$ and use the result for $T=1$.
(Step 2). (perturbation in $x$). Assume that $a^{ij}$ depends only
on $x$. Recall we are assuming
\begin{equation}
\label{K_3}
\sup_{i,j,\omega}|a^{ij}(\cdot)|_{B^{|\gamma|}} \leq K_{3}.
\end{equation}
In this step we prove that there exists a positive constant
$\varepsilon_{1}=\varepsilon_1(N_0)$, thus independent of $T$ and
$K_{3}$, so that \eqref{indT} holds with new constant
$N=N(N_0,K_3)$ if
\begin{align}
\label{eps_1:perturbation}
\sup_{i,j, t,x,y}|a^{ij}(x)-a^{ij}(y)| \leq \varepsilon_{1}.
\end{align}
Set
$a_{0}^{ij}:=a^{ij}(0)$, and rewrite (\ref{eqn 8.26.2}) as
$$
\partial_{t}^{\alpha}u=a^{ij}_{0}u_{x^{i}x^{j}}+f+(a^{ij}-a^{ij}_{0})u_{x^{i}x^{j}}.
$$
By the result of Step 1, for each $t\leq T$
\begin{align}
\label{eq:6.25.1}
\|u_{xx}\|_{\fH^{\gamma}_{p}(t)} \leq N_0 \left(\|f\|_{\fH^{\gamma}_{p}(t)}
+\|(a^{ij}-a^{ij}_0)u_{x^ix^j}\|_{\fH^{\gamma}_{p}(t)}\right).
\end{align}
By \eqref{multi},
$$
\|(a^{ij}-a_{0}^{ij})u_{x^{i}x^{j}}\|_{H_{p}^{\gamma}}
\leq N(d,\gamma) |a^{ij}-a_{0}^{ij}|_{B^{|\gamma|}} \|u_{xx}\|_{H_{p}^{\gamma}}.
$$
It follows from (\ref{eq:6.25.1}) that
\begin{equation}
\label{leadto}
\|u_{xx}\|_{\fH^{\gamma}_{p}(t)} \leq N_0 \|f\|_{\fH^{\gamma}_{p}(T)}
+N_0N(d,\gamma) |a^{ij}(t,\cdot)-a_{0}^{ij}|_{B^{|\gamma|}} \|u_{xx}\|_{\fH^{\gamma}_{p}(t)}.
\end{equation}
Hence we get (\ref{indT}) with $2N_0$ in place of $N_0$ if
\begin{align}
\label{claim:coefficient}
|a^{ij}-a_{0}^{ij}|_{B^{|\gamma|}} \leq \frac{1}{2N(d,\gamma)N_{0}}=:\varepsilon_2.
\end{align}
Now we take $\varepsilon_1=\varepsilon_2/2$ and assume (\ref{eps_1:perturbation}) holds. Fix a small constant $\rho>0$ so that $\rho^{(|\gamma|)\wedge 1}K_3\leq \varepsilon_2/2$, and set
$$
a^{ij}_{\rho}(t,x):= a^{ij}(\rho x),\quad u_{\rho}(t,x):=u(\rho^{\frac{2}{\alpha}}t,\rho x), \quad
f_{\rho}(t,x):=\rho^{2}f(\rho^{\frac{2}{\alpha}}t,\rho x).
$$
Note that $u_{\rho}(t,x)$ satisfies
\begin{align*}
\partial_{t}^{\alpha}u_{\rho} = a_{\rho}^{ij}(u_{\rho})_{x^{i}x^{j}}+f_{\rho}, \quad \quad t\leq \rho^{-2/{\alpha}}T.
\end{align*}
By the definition of $B^{|\gamma|}$, (\ref{K_3}), and the choice of $\rho$,
\begin{align*}
|a^{ij}_{\rho}(\cdot)-a^{ij}_{\rho}(0)|_{B^{|\gamma|}} \leq \sup_x |a^{ij}-a^{ij}_0| + 1_{\gamma \neq 0}\rho^{(|\gamma|)\wedge 1}K_{3}\leq \varepsilon_2.
\end{align*}
Thus by the above arguments which lead to (\ref{leadto}) and (\ref{claim:coefficient}), we get for each $t\leq \rho^{-2/{\alpha}}T$,
$$
\|(u_{\rho})_{xx}\|_{\fH^{\gamma}_{p}(t)} \leq 2N_0 \|f_{\rho}\|_{\fH^{\gamma}_p(t)}.
$$
Consequently, for each $t\leq T$,
$$
\|u_{xx}\|_{\fH^{\gamma}_p(t)}\leq N(K_3, N_0) \|f\|_{\fH^{\gamma}_p(t)}.
$$
As before, this and (\ref{eq:solution space estimate 1}) yield (\ref{apriori det}). Before moving to next step we emphasize that we take $\varepsilon_1=(4N(d,\gamma)N_0)^{-1}$ and therefore it does not depend on $T$ and $K_3$.
(Step 3). (Partition of unity). We still assume $a^{ij}$ is independent of $t$. Choose a $\delta_1$ so that
\begin{align}
\label{eq:freezing}
|a^{ij}(x)-a^{ij}(y)| \leq \frac{\varepsilon_{1}}{2}
\end{align}
whenever $|x-y|\leq 4\delta_{1}$. For this $\delta_{1}$, take a sequence of functions $\zeta_{n}\in C_{c}^{\infty}(\mathbb{R}^{d})$, $n\in\mathbb{N}$, so that $0\leq \zeta_n \leq 1$, the support of $\zeta_{n}$ lies in $B_{\delta_{1}}(x_{n})$ for some $x_{n}\in\mathbb{R}^{d}$,
$$
\sup_{x\in\mathbb{R}^{d}} \sum_{n\in\mathbb{N}} \left| D_{x}^{\mathbf{n}}\zeta_{n}(x) \right| \leq M(\delta_{1},\mathbf{\mathbf{n}})<\infty
$$
for any multi-index $\mathbf{n}\in\mathbb{Z}^{d}$ and,
$$
\inf_{x\in\mathbb{R}^{d}}\sum_{n\in\mathbb{N}}\left| \zeta_{n}(x) \right| \geq \vartheta >0.
$$
It is well-known (\cite[Lemma 6.7]{Krylov1999}) that for any $\gamma \in\mathbb{R}$ and $\mathbf{n}\in\mathbb{N}$,
\begin{align}
\label{eq:partition of unity}
\|h\|_{H_{p}^{\gamma}}^{p}\leq N \sum_{n\in\mathbb{N}}\|h \zeta_{n}\|_{H_{p}^{\gamma}}^{p}
\leq N \|h\|_{H_{p}^{\gamma}}^{p},\quad \sum_{n\in\mathbb{N}} \|u D_{x}^{\mathbf{n}}\zeta_{n}\|_{H_{p}^{\gamma}}^{p} \leq N \|u\|_{H_{p}^{\gamma}}^{p}
\end{align}
where $N$ depend only on $d$,$\gamma$, $M(\delta_{1}$, $\mathbf{n})$, and $\vartheta$.
Take a nonnegative $\eta\in C_{c}^{\infty}(\mathbb{R}^{d})$ so that $0\leq \eta\leq 1$, $\eta=1$ on $B_{1}$, and $\eta =0$ outside $B_{2}$. Write
$$
u_{n}=u\zeta_{n},\quad \eta_{n}(x)=\eta\left(\frac{x-x_{n}}{\delta_{1}}\right)
$$
and define
\begin{eqnarray}
\label{defa}
a_{n}^{ij}(x):=\eta_{n}(x)a^{ij}(x)+(1-\eta_{n}(x))a^{ij}(x_{n})
\end{eqnarray}
Then, because $\eta_n=1$ on the support of $\zeta_n$, $u_{n}(t,x)$ satisfies
$$
\partial_{t}^{\alpha}u_{n}(t,x) = a_{n}^{ij}(u_{n})_{x^{i}x^{j}}+\bar{f}_{n},
$$
where
$$
\bar{f}_{n}(t,x)
:=f(t,x)\zeta_{n}+(a^{ij}_nu_{x^{i}x^{j}}\zeta_{n}-a_{n}^{ij}(u_{n})_{x^{i}x^{j}}).
$$
Note that
$$
a^{ij}_nu_{x^{i}x^{j}}\zeta_{n}-a^{ij}_n(u_{n})_{x^{i}x^{j}} = a^{ij}(2 u_{x^{i}}(\zeta_{n})_{x^{i}}+u(\zeta_{n})_{x^{i}x^{j}}).
$$
Due to (\ref{eq:freezing}), for each $x,y\in \bR^d$,
\begin{align*}
|a_{n}^{ij}(t,x)-a_{n}^{ij}(t,y)|&= |\eta_n(x)(a^{ij}(x)-a^{ij}(x_n))-\eta_n(y)(a^{ij}(y)-a^{ij}(x_n))|\\
&\leq |\eta_n(x)(a^{ij}(x)-a^{ij}(x_n))|+|\eta_n(y)(a^{ij}(y)-a^{ij}(x_n))| \leq \varepsilon_{1}.
\end{align*}
Also note that $(a^{ij}_n)$ satisfies the uniform ellipticity condition with the same constant $\delta_0$.
Therefore, by the result from Step 2 and (\ref{eq:partition of unity}), for each $t\leq T$,
\begin{align}
\nonumber
&\|u\|_{\fH_{p}^{\gamma+2}(t)}^{p} \leq N\sum_{n\in\mathbb{N}}\|u_{n}\|_{\fH_{p}^{\gamma+2}(t)}^{p}
\leq N\sum_{n\in\mathbb{N}} \|\bar{f}_{n}\|_{\fH_{p}^{\gamma}(t)}^{p}\\
&\leq N\|u\|^p_{\fH^{\gamma+1}_p(t)}+N\|f\|^p_{\fH_p(t)} \nonumber\\
&\leq \varepsilon \|u\|_{\fH_{p}^{\gamma+2}(t)}^{p} +N(\varepsilon) \|u\|_{\fH_{p}^{\gamma}(t)}^{p} +N \|f\|^p_{\fH_{p}^{\gamma}(t)}. \label{eqn 8.26.8}
\end{align}
We take $\varepsilon = 1/2$, and to drop the term $\|u\|_{\fH^{\gamma}_p(t)}$ above we use (\ref{eq:solution space estimate 1}), which implies
\begin{eqnarray*}
\|u\|^p_{\fH^{\gamma}_p(t)}
&\leq& N \int_{0}^{t}(t-s)^{\theta-1} \|a^{ij} u_{x^ix^j} + f\|^p_{\fH^{\gamma}_p(s)}ds \\
&\leq& N \int_{0}^{t}(t-s)^{\theta-1} (\|u\|^p_{\fH^{\gamma+2}_p(s)}+ \|f\|^p_{\fH^{\gamma}_p(s)}) ds \\
&\leq& N \int_{0}^{t}(t-s)^{\theta-1} (\|u\|^p_{\fH^{\gamma}_p(s)}+ \|f\|^p_{\fH^{\gamma}_p(s)}) ds
\end{eqnarray*}
where the last inequality is due to (\ref{eqn 8.26.8}). Therefore by applying fractional Gronwall's lemma (\cite[Corollary 2]{YGD}),
we obtain (\ref{apriori det}).
We remark that up to this step,
the constant $N$ of (\ref{apriori det}) depends only on $\delta_0$, $p$, $K_3$, $\alpha$, $\gamma$, $T$, and the modulus of continuity of $a^{ij}$.
(Step 4) (general case). Recall that in Step 3 we proved the lemma
when $a^{ij}$ are independent of $t$. For the general case, it is
enough to repeat Steps 5 and 6 of the proof of \cite[Theorem
2.9]{KKL2014}. Indeed, in \cite{KKL2014} the lemma is proved when
$\gamma=0$, and the proof is first given for time-independent
$a^{ij}$, and then this result is extended for the general case.
This method of generalization works exactly same for any $\gamma\in
\bR$.
\end{proof}
We continue the proof of Theorem \ref{thm:main results non-div}.
\vspace{2mm}
{\bf{Case A. Linear case}}. Suppose $f$ and $g$ are independent of
$u$, and $b^i=c=\mu^{ik}=\nu^k=0$. To apply the method of
continuity, for each $\lambda\in [0,1]$ denote
$$
(a^{ij}_{\lambda})=\lambda (a^{ij})+(1-\lambda)I_{d\times d}, \quad \sigma^{ijk}_{\lambda}=\lambda \sigma^{ijk},
$$
where $I_{d\times d}$ is the $d\times d$-identity matrix.
Then
$$
L_{\lambda}u:=\lambda a^{ij}u_{x^ix^j}+(1-\lambda)\Delta u=a^{ij}_{\lambda}u_{x^ix^j}
$$
and
$$
\Lambda^k_{\lambda}u:=\lambda \sigma^{ijk}u_{x^ix^j}=\sigma^{ijk}_{\lambda}u_{x^ix^j}.
$$
Due to the method of continuity and Theorem \ref{thm:model eqn}, we only need
to prove a priori estimate (\ref{824 a pr}) holds given that a solution $u\in \cH^{\gamma+2}_{p,0}(T)$ to equation (\ref{eqn 8.25.7}) already exists.
Note that for any $\lambda \in [0,1]$ the coefficients
$a^{ij}_{\lambda}$ and $\sigma_{\lambda}$ satisfy the
same conditions assumed for $a^{ij}$ and $\sigma^{ijk}$, that is,
conditions specified in Assumptions \ref{assu:common} and \ref{assu:div} with the same constants used there.
This shows that by considering $a^{ij}_{\lambda}$ and $\sigma_{\lambda}$
in place of $a^{ij}$ and $\sigma$, we only need to prove (\ref{824 a pr})
for $\lambda=1$.
By Theorem \ref{thm:model eqn}, the equation
\begin{equation}
\label{v}
\partial_{t}^{\alpha}v(t,x)= \Delta v (t,x)+\sum_{k=1}^{\infty}\partial_{t}^{\beta}\int_{0}^{t}(\sigma^{ijk}u_{x^ix^j}+g^{k})dw_{s}^{k}
\end{equation}
has a unique solution $v\in \cH^{\gamma+2}_{p,0}(T)$, and moreover
\begin{equation}
\label{1}
\|v\|_{\cH^{\gamma+2}_p(T)} \leq
N\|g\|_{\bH^{\gamma+c_0'}_p(T,l_{2})}.
\end{equation}
Indeed, (\ref{1}) is obvious if $\beta \geq 1/2$ because $\sigma^{ijk}=0$ in this case.
If $\beta<1/2$, then by Theorem \ref{thm:model eqn} and Lemma \ref{lem:pde approach}, for each $t\leq T$,
\begin{eqnarray*}
\|v\|^p_{\cH^{\gamma+2}_p(t)}&\leq& N I^{1-2\beta}_t \|\sigma^{ij}u_{x^ix^j}+g\|^p_{\bH^{\gamma}_p(\cdot,l_2)}(t)\\
&\leq& N I^{1-2\beta}_t \|u\|^p_{\cH^{\gamma+2}_p(\cdot)}(t)+N
\|g\|^p_{\bH^{\gamma}_p(T,l_2)}.
\end{eqnarray*}
Therefore (\ref{1}) follows from the fractional Gronwall's lemma.
Note that $\bar{u}=u-v$ satisfies the equation
\begin{align*}
\partial_{t}^{\alpha}\bar{u}(t,x)= a^{ij}\bar{u}_{x^{i}x^{j}}(t,x)
+a^{ij}(t)v_{x^{i}x^{j}}(t,x)-\Delta v(t,x) +f(t,x).
\end{align*}
By Lemma \ref{det11},
\begin{eqnarray}
\|\bar{u}\|_{\cH^{\gamma+2}_p(T)}&\leq& N
\|a^{ij}(t)v_{x^{i}x^{j}}(t,x)-\Delta v(t,x)
+f(t,x)\|_{\bH^{\gamma}_p(T)} \nonumber\\
&\leq& N \|v\|_{\bH^{\gamma+2}_p(T)}+N\|f\|_{\bH^{\gamma}_p(T)}.
\label{2}
\end{eqnarray}
Since $u=\bar{u}+v$, the desired estimate follows from (\ref{1}) and
(\ref{2}).
\subsection*{B: General case.}
Write
$$
\bar{f}:=b^{i}u_{x^{i}}+cu+f(u) ,\quad \bar{g}^{k}:=\mu^{ik} u_{x^{i}}+\nu^{k}u+g^{k}(u).
$$
Note that $\mu^{ik}= 0 $ if ${c_0'}\geq 1$.
Then by (\ref{multi}), (\ref{multi2}), and Assumption \ref{assu:non-div} (iii),
\begin{align*}
&\|\bar{f}(u)-\bar{f}(v)\|_{H_{p}^{\gamma}} + \|\bar{g}(u)-\bar{g}(v)\|_{H_{p}^{\gamma+{c_0'}}(l_{2})} \\
&\leq N \left(\|u-v\|_{H_{p}^{\gamma+1}}+\|\mu^{i}(u-v)_{x^{i}}\|_{H_{p}^{\gamma+{c_0'}}(l_{2})} +\|u-v\|_{H_{p}^{\gamma+{c_0'}}}\right) \\ &\quad\qquad\qquad\qquad\qquad+ \|f(u)-f(v)\|_{H_{p}^{\gamma}}+\|g(u)-g(v)\|_{H_{p}^{\gamma+{c_0'}}(l_{2})} \\
& \leq \varepsilon \|u-v\|_{H_{p}^{\gamma+2}}+N\|u-v\|_{H_{p}^{\gamma}}
\end{align*}
where $N$ depends on $d,p,m,\kappa,K_{3},K_{4}$, and $\varepsilon$.
Hence considering $\bar{f}$ and $\bar{g}^{k}$ in place of $f$ and
$g^{k}$, we may assume $b^{i}=c=\mu^{ik}=\nu^{k}=0$.
For each $u\in\mathcal{H}_{p}^{\gamma+2}(T)$, consider the equation
\begin{align*}
\partial_{t}^{\alpha}v = a^{ij}v_{x^{i}x^{j}}+f(u)+\sum_{k=1}^{\infty}\int_{0}^{t}[\sigma^{ijk}v_{x^{i}x^{j}}+g^{k}(u)]dw_{s}^{k}
\end{align*}
with zero initial condition. By the result of Case A, this equation
admit a unique solution $v\in\mathcal{H}_{p}^{\gamma+2}(T)$. By
denoting $v=\mathcal{R}u$, we can define an operator
$$
\mathcal{R}:\mathcal{H}_{p}^{\gamma+2}(T)\rightarrow\mathcal{H}_{p}^{\gamma+2}(T).
$$
By Lemma \ref{lem:solution space}(ii), \eqref{eq:assumption non-div}, and the result of Case A, for each $t\leq T$,
\begin{align*}
\|\mathcal{R}u-\mathcal{R}v\|_{\mathcal{H}_{p}^{\gamma+2}(t)}^{p} & \leq N_{0}\left(\|f(u)-f(v)\|_{\mathbb{H}_{p}^{\gamma}(t)}^{p}+\|g(u)-g(v)\|_{\mathbb{H}_{p}^{\gamma+{c_0'}}(t,l_{2})}^{p}\right)\\
& \leq N_{0}\varepsilon^{p}\|u-v\|_{\mathcal{H}_{p}^{\gamma+2}(t)}^{p}+N_{1}\|u-v\|_{\mathbb{H}_{p}^{\gamma}(t)}^{p}\\
& \leq N_{0}\varepsilon^{p}\|u-v\|_{\mathcal{H}_{p}^{\gamma+2}(t)}^{p}+N_{1}\int_{0}^{t}(t-s)^{\theta-1}\|u-v\|_{\mathcal{H}_{p}^{\gamma+2}(s)}^{p}ds,
\end{align*}
where $N_{1}$ depends also on $\varepsilon$. Next, we fix $\varepsilon$
so that $\Theta:=N_{0}\varepsilon^{p}<2^{-2}$. Then repeating the
above inequality and using the identity
\begin{align*}
\int_{0}^{t} & (t-s_{1})^{\theta-1}\int_{0}^{s_{1}}(s_{1}-s_{2})^{\theta-1}\cdots\int_{0}^{s_{n-1}}(s_{n-1}-s_{n})^{\theta-1}ds_{n}\cdots ds_{1}\\
&=\frac{\left\{ \Gamma(\theta)\right\} ^{n}}{\Gamma(n\theta+1)}t^{n\theta},
\end{align*}
we get
\begin{align*}
\|\mathcal{R}^{n}u-\mathcal{R}^{n}v\|_{\mathcal{H}_{p}^{\gamma+2}(t)}^{p} & \leq\sum_{k=0}^{n}\binom{n}{k}\Theta^{n-k}\left(T^{\theta}N_{1}\right)^{k}\frac{\left\{ \Gamma(\theta)\right\} ^{k}}{\Gamma(k\theta+1)}\|u-v\|_{\mathcal{H}_{p}^{\gamma+2}(t)}^{p}\\
& \leq2^{n}\Theta^{n}\max_{k}\left[\frac{\left\{ \Theta^{-1}T^{\theta}N_{1}\Gamma(\theta)\right\} ^{k}}{\Gamma(k\theta+1)}\right]\|u-v\|_{\mathcal{H}_{p}^{\gamma+2}(t)}^{p}\\
& \leq\frac{1}{2^{n}}N_{2}\|u-v\|_{\mathcal{H}_{p}^{\gamma+2}(t)}^{p}.
\end{align*}
For the second inequality above we use $\sum_{k=0}^{n}\binom{n}{k}=2^{n}$.
It follows that if $n$ is sufficiently large then $\mathcal{R}^{n}$
is a contraction in $\mathcal{H}_{p}^{\gamma+2}(T)$, and this yields all
the claims. The theorem is proved.
\mysection{Proof of Theorem \ref{thm:main results div}}
\label{pf div thm}
We first prove a result for a deterministic equation of divergence type.
\begin{lem}
\label{det112}
Let $a^{ij}$ be given as in (\ref{a}) with non-random $\tau_n$ and $a^{ij}_n$. Suppose $a^{ij}$ satisfy the uniform ellipticity (\ref{elliptic}) and $a^{ij}_n$ are uniformly continuous in $(t,x)$. Then for any solution $u\in \fH^{\alpha,1}_{p,0}(T)$ to the deterministic equation
\begin{equation}
\label{eqn 8.27.2}
\partial^{\alpha}_t u=D_{x^i}(a^{ij}u_{x^ix^j}+f^i)+h
\end{equation}
in $\fH^{-1}_p(T)$,
it holds that
\begin{equation}
\label{apriori det div}
\|u\|_{\fH^{1}_p(T)}\leq N (\|f^i\|_{\fL_p(T)}+\|h\|_{\fH^{-1}_p(T)}),
\end{equation}
where $N$ depends only on $\alpha$, $p$, $\gamma$, $\delta_0$, $T$, $M_0$, and the modulus of continuity of $a^{ij}_n$.
\end{lem}
\begin{proof} We divide the proof into three steps.
(Step 1). Let $a^{ij}$ depend only on $t$. In this case, (\ref{apriori det div}) is a consequence of (\ref{apriori det}) with $\gamma=-1$,
which is because $\|D_{x^i}f^i\|_{H^{-1}_p}\leq N\|f^i\|_{L_p}$.
(Step 2). We prove there exists $\varepsilon_2>0$, which depends also on $T$, such that (\ref{apriori det div}) holds if
\begin{equation}
\label{varepsilon2}
\sup_{t,x}|a^{ij}(t,x)-a^{ij}(t,y)|\leq \varepsilon_2.
\end{equation}
Denote
$a_{0}^{ij}(t):=a^{ij}(t,0)$, and rewrite the equation as
\begin{align*}
\partial_{t}^{\alpha}u = D_{x^{i}}(a_{0}^{ij}u_{x^{i}}+ \bar{f}^{i})+h
\end{align*}
where
\begin{align*}
\bar{f} ^{i}:=f^{i}+\sum_{j=1}^{d}\left(a^{ij}-a_{0}^{ij}\right)u_{x^{j}}.
\end{align*}
Note that $a_{0}^{ij}$ is independent of $x$. By the result of Step 1, for each $t\leq T$,
\begin{align*}
\|u\|_{\fH^1_{p}(t)} & \leq N_3(\|f\|_{\fL_{p}(t)}+\|(a^{ij}-a^{ij}_0)u_{x^j}\|_{\fL_{p}(t)}+N\|h\|_{\fH^{-1}_p(t)}).
\end{align*}
Observe that
$$
\|(a^{ij}(t,\cdot)-a_{0}^{ij}(t))u_{x^{i}}(t,\cdot)\|_{L_{p}} \leq N(d,p) \sup_{t,x}|a^{ij}(t,x)-a_{0}^{ij}(t)| \|u(t,\cdot)\|_{H_{p}^{1}}.
$$
Therefore, our claim follows if (\ref{varepsilon2}) holds with $\varepsilon_2=(2N(d,p) N_3)^{-1}$.
(Step 3). We introduce a partition of unity $\zeta^n$ as in the proof of Lemma \ref{det11}, and define $\eta$ and $a^{ij}_n$ as in (\ref{defa}) so that each $(a^{ij}_n)$ satisfies (\ref{varepsilon2}). Note $u^n(t,x)=u\zeta^n$ satisfies
$$
\partial^{\alpha}_tu^n=D_{x^i}(a^{ij}_n u^n_{x^j} +\bar{f}^{n,i})+\bar{h}^n
$$
where
$$
\bar{f}^{n,i}=f^i\zeta^n-a^{ij}u\zeta^n_{x^j}, \quad \quad \bar{h}^n =h\zeta^n-a^{ij}u_{x^j}\zeta^n_{x^i}.
$$
Therefore, using Step 2 and $\|\cdot\|_{H^{-1}_p}\leq N\|\cdot\|_{L_p}$, we get
\begin{eqnarray*}
\|u\|^p_{\fH^1_p(t)}&\leq& N\sum_{n\in \bN}\|u^n\|^p_{\fH^1_p(t)}\leq N\sum_{n\in \bN}(\|\bar{f}^{n}\|^p_{\fL_p(t)}+\|\bar{h}^n\|^p_{\fH^{-1}_p(t)})\\
&\leq& N(\|f^i\|^p_{\fL_p(t)}+\|h\|^p_{\fH^{-1}_p(t)})+N\|u\|^p_{\fL_p(t)}+N\|a^{ij}u_{x^j}\|^p_{\fH^{-1}_p(t)}.
\end{eqnarray*}
Here we claim that for any $\varepsilon>0$,
\begin{equation}
\label{eqn 8.27.5}
\|a^{ij}u_{x^j}\|_{H^{-1}_p}\leq \varepsilon \|u\|_{H^1_p}+N(\varepsilon)\|u\|_{L_p}.
\end{equation}
Indeed, since $a^{ij}$ are uniformly continuous in $x$ uniformly in $t$, considering appropriate convolution we can take a sequence of $C^1$-functions $a^{ij}_n$ which uniformly converges to $a^{ij}$ with respect to $x$ uniformly in $t$. Thus
\begin{eqnarray*}
\|a^{ij}u_{x^j}\|_{H^{-1}_p}&\leq& \|(a^{ij}_n-a^{ij})u_{x^j}\|_{H^{-1}_p}+ \|a^{ij}_nu_{x^j}\|_{H^{-1}_p}\\
&\leq& \sup_{t,x} |a^{ij}_n-a^{ij}|\|u_x\|_{L_p}+|a^{ij}_n|_{B^1}\|u_x\|_{H^{-1}_p}.
\end{eqnarray*}
This certainly proves (\ref{eqn 8.27.5}). Taking small $\varepsilon$ and using the interpolation
$\|u\|_{L_p}\leq \varepsilon' \|u\|_{H^1_p}+N(\varepsilon')\|u\|_{H^{-1}_p}$, we get for each $t\leq T$,
$$
\|u\|^p_{\fH^1_p(t)}\leq N\|f^i\|^p_{\fL_p(t)}+N\|h\|^p_{\fH^{-1}_p(t)}+N\|u\|^p_{\fH^{-1}_p(t)}
$$
The last term $\|u\|_{\fH^{-1}_p(t)}$ can be easily dropped as before using (\ref{eq:solution space estimate 1}) and Gronwall's lemma. The lemma is proved.
\end{proof}
Now we prove Theorem \ref{thm:main results div}.
\vspace{2mm}
(Step 1). Suppose $f^i,h$ and $g$ are independent
of $u$ and $b^i=c=\nu^{ik}=0$. In this case, by the method of
continuity and Theorem \ref{thm:model eqn} we only need to show a
priori estimate \eqref{eq:a priori estimate div} holds given that a
solution $u\in \cH^1_{p}(T)$ already exists. See the proof of
Theorem \ref{thm:main results non-div} for details.
In this case estimate \eqref{eq:a priori estimate div} follows from Lemma \ref{det112} and the arguments in Case 1 of the proof of Theorem
\ref{thm:main results non-div}. Indeed, take the function $v\in \cH^1_{p}(T)$ from (\ref{v}), which is a solution to
$$
\partial_{t}^{\alpha}v(t,x)= \Delta v (t,x)+\sum_{k=1}^{\infty}\partial_{t}^{\beta}\int_{0}^{t}(\sigma^{ijk}u_{x^ix^j}+g^{k})dw_{s}^{k}.
$$
By (\ref{1}),
$$
\|v\|_{\cH^{1}_p(T)} \leq
N\|g\|_{\bH^{c_0'-1}_p(T,l_{2})}.
$$
Note that $\bar{u}:=u-v$ satisfies
$$
\partial^{\alpha}_t\bar{u}=D_{x^i}(a^{ij}\bar{u}_{x^j}+\bar{f}^i)+h, \quad \bar{f}^i:=(a^{ij}-\delta^{ij})v_{x^j}.
$$
Thus one can estimate $\|\bar{u}\|_{\cH^1_p(T)}$ using Lemma \ref{det112}, and this leads to \eqref{eq:a priori estimate div} since $u=\bar{u}+v$.
(Step 2). General case. The proof is almost identical to that of Case B of the proof of Theorem \ref{thm:main results non-div}. We put
$$
\bar{f}^i=b^iu+f(u), \quad \bar{h}(u)=cu+h(u), \quad \bar{g}^k=\nu^{ik}u+\nu^k u+g^k(u).
$$
Then, as before, one can check these functions satisfy condition (\ref{eq:assumption div}), and therefore we may assume $b^i=c=\nu^{ik}=\nu^k=0$.
Then, using Step 1, we define the operator $\cR: \cH^1_{p}(T)\to \cH^1_{p}(T)$ so that $v=\cR u$ is the solution to the problem
$$
\partial^{\alpha}_t v=D_{x^i}(a^{ij}v_{x^i}+f^i(u))+h(u) +\partial^{\beta}_t\int^t_0 (\sigma^{ijk}v_{x^ix^j}+g^k(u))dw^k_t
$$
with zero initial condition. After this, using the arguments used in
the proof of Theorem \ref{thm:main results non-div}, one easily
finds that $\cR^n$ is a contraction in $\cH^1_{p}(T)$ if $n$ is
large enough. This proves the theorem.
\qed
\mysection{SPDE driven by space-time white noise}
\label{space-time}
In this section we assume
\begin{equation}
\label{eqn st}
\beta<\frac{3}{4}\alpha+\frac{1}{2},
\end{equation}
and the space dimension $d$ satisfies
\begin{equation}
\label{dimension}
d<4-\frac{2(2\beta-1)_+}{\alpha}=:d_0.
\end{equation}
Note
$d_0\in (1,4]$ due to (\ref{eqn st}). If $\beta< \frac{\alpha}{4}+1/2$ then one can take $d=1,2,3$. Also, $\alpha=\beta=1$ then $d$ must be $1$.
\vspace{3mm}
In this section we study the SPDE
\begin{equation}
\label{space-time}
\partial^{\alpha}_t u=\left(a^{ij}u_{x^ix^j}+b^iu_{x^i}+cu+f(u)\right)
+\partial^{\beta}_t \int^t_0 h(u) \, dB_t
\end{equation}
where the coefficients $a^{ij}$, $b^i$, $c$ are functions depending on
$(\omega,t,x)$, the functions $f$ and $h$ depend on $(\omega,t,x)$ and the unknown $u$, and
$B_t$ is a cylindrical Wiener process on $L_2(\bR^d)$.
Let $\{\eta^k:k=1,2,\cdots\}$ be an orthogonal basis of $L_2(\bR^d)$. Then (see \cite[Section 8.3]{Krylov1999})
$$
dB_t=\sum_{k=1}^{\infty} \eta^k dw^k_t
$$
where
$ w^k_t:=(B_t,\eta^k)_{L_2}$ are independent one dimensional Wiener processes. Hence one can rewrite (\ref{space-time}) as
$$
\partial^{\alpha}_t u= \left(a^{ij}u_{x^ix^j}+b^iu_{x^i}+cu+f(u)\right)
+ \sum_{k=1}^\infty\partial^{\beta}_t \int^t_0 g^k(u)\, dw^k_t,
$$
where
$$
g^k(t,x,u)=h(t,x,u)\eta^k(x).
$$
\begin{lem}
\label{lemma 1.26}
Assume
\begin{equation}
\label{algebraic}
{\kappa_0} \in \left(\frac{d}{2},d\right], \quad 2\leq 2r\leq p, \quad 2r<\frac{d}{d-{\kappa_0}},
\end{equation}
and $h(x,u)$, $\xi(x)$ are functions of $(x,u)$ and $x$ respectively such that $|h(x,u)-h(x,v)|\leq \xi(x) |u-v|$.
For $u\in L_p(\bR^d)$, set $g^k(u)=h(x,u(x)) \eta_k(x)$. Then
$$
\|g(u)-g(v)\|_{H^{-{\kappa_0}}_p(l_2)}\leq N \|\xi\|_{L_{2s}}\|u-v\|_{L_p},
$$
where
$s=r/{(r-1)}$
is the conjugate of $r$ and $N=N(r)<\infty$. In particular, if $r=1$ and $\xi$ is bounded, then
$$
\|g(u)-g(v)\|_{H^{-{\kappa_0}}_p(l_2)}\leq N \|u-v\|_{L_p}.
$$
\end{lem}
\begin{proof}
It is well-known (e.g. \cite[p.132]{Stein1970}, \cite[Exercise 12.9.19]{Krylov2008}) that there exists a Green function $G(x)$, which decays exponentially fast at infinity and behaves like $|x|^{{\kappa_0}-d}$ so that the equality holds:
$$
\|g(u)-g(v)\|_{H^{-{\kappa_0}}_p(l_2)}=\|\bar{h}\|_{L_p},
$$
where
\begin{eqnarray*}
\bar{h}(x)
&:=&\left(\int_{\bR}|G(x-y)|^2|h(y,u(y))-h(y,v(y))|^2 dy\right)^{1/2}\\
&\leq& \left(\int_{\bR}|G(x-y)|^2\xi^2(y)|u(y)-v(y)|^2dy\right)^{1/2}=:\tilde{h}(x).
\end{eqnarray*}
By H\"older's inequality,
\begin{align*}
|\tilde{h}(x)|
\leq \|\xi\|_{L_{2s}} \cdot \left(\int_{\bR}|G(x-y)|^{2r}|u(y)-v(y)|^{2r}dy\right)^{1/(2r)}.
\end{align*}
Note that $\|G\|_{L_{2r}} <\infty$ since $2r<\frac{d}{d-{\kappa_0}}$.
Therefore applying Minkowski's inequality, we have
$$
\|\tilde{h}\|_{L_p}\leq N\|\xi\|_{L_{2s}} \|G\|_{2r}\|u-v\|_{L_p} \leq N\|\xi\|_{L_{2s}} \|u-v\|_{L_p}.
$$
The lemma is proved.
\end{proof}
\begin{remark}
\label{hu rmk}
By following the proof of Lemma \ref{lemma 1.26}, one can easily check that
$$
\|g(u)\|_{H^{-{\kappa_0}}_p(l_2)}\leq N \|h(u)\|_{L_p}.
$$
\end{remark}
\begin{assumption}
\label{ass 3}
\noindent
(i) The coefficients $a^{ij},b^i$, and $c$ are $\mathcal{P}\otimes\mathcal{B}(\mathbb{R})$-measurable.
\noindent
(ii) The functions $f(t,x,u)$ and $g(t,x,u)$ are $\mathcal{P}\otimes\mathcal{B}(\mathbb{R}^d\times \bR)$-measurable.
\noindent
(iii) For each $\omega,t,x,u$ and $v$,
$$
|f(t,x,u)-f(t,x,v)|\leq K|u-v|, \quad |h(t,x,u)-h(t,x,v)|\leq \xi(t,x)|u-v|,
$$
where $\xi$ depends on $(\omega,t,x)$.
\end{assumption}
Denote
$$
f_0=f(t,x,0), \quad h_0=h(t,x,0).
$$
\begin{thm}
\label{thm space-time}
Suppose Assumption \ref{ass 3} holds and
$$
\|f_0\|_{\bH^{-\kappa_0-c'_0}_p(T)}+\|h_0\|_{\bL_p(T)} + \sup_{\omega,t}\|\xi\|_{2s} \leq K<\infty,
$$
where $\kappa_0$ and $s$ satisfy
\begin{equation}
\label{eqn 9.10}
\frac{d}{2}<\kappa_0< \left(2-\frac{(2\beta-1)_+}{\alpha}\right)\wedge d, \quad \quad \frac{d}{2\kappa_0-d} <s
\end{equation}
and $c'_0$ from (\ref{c_0}).
Also assume that the coefficients $a^{ij}$, $b^i$, and $c$ satisfy Assumption \ref{assu:common} and (\ref{multiplier}) with $\gamma:=-\kappa_0-c'_0$. Then
equation (\ref{space-time}) with zero initial condition has a unique solution $u\in \cH^{2-\kappa_0-c'_0}_{p}(T)$, and for this solution we have
$$
\|u\|_{\cH^{2-\kappa_0-c'_0}_p(T)}\leq N \|f_0\|_{\bH^{-\kappa_0-c'_0}_p(T)}+N\|h_0\|_{\bL_p(T)}.
$$
\end{thm}
\begin{proof}
We only need to check if the conditions for Theorem \ref{thm:main results non-div} are satisfied with $\gamma:=-\kappa_0-c'_0$. Since $f(u)$ is Lipschitz continuous, we only check the conditions for $g^k(u):=h(u)\eta_k$.
Let $r$ be the conjugate of $s$ and then
$2r<\frac{d}{d-{\kappa_0}}$ due to the assumption $\frac{d}{2\kappa_0-d} <s$.
Recall $\gamma$ is chosen such that $\gamma+c'_0=-{\kappa_0}$. Thus, By Lemma \ref{lemma 1.26}, for any $\varepsilon>0$,
$$
\|g(u)-g(v)\|_{H^{\gamma+c'_0}_p(l_2)}\leq N\|\xi\|_{L_{2s}} \|u-v\|_{L_p}\leq \varepsilon \|u-v\|_{H^{\gamma+2}_p}+N(\varepsilon)\|u-v\|_{H^{\gamma}_p},
$$
where the second inequality is due to $\gamma+2>0$, which is equivalent to ${\kappa_0} +c'_0<2$.
Therefore all the conditions for Theorem \ref{thm:main results non-div} are checked.
The theorem is proved.
\end{proof}
\begin{remark}
(i) By (\ref {dimension}), there always exists $\kappa_0$ satisfying (\ref{eqn 9.10}).
(ii) The constant $2-\kappa_0-c'_0$ gives
the regularity of the solution $u$.
To see how smooth the solution is,
recall $c'_0=(2\beta-1)_+/{\alpha} +\kappa 1_{\beta=1/2}$.
It follows
$$
0< 2-\kappa_0-c'_0 < \begin{cases}
2-\frac{d}{2}-\frac{2\beta-1}{\alpha} \quad &\hbox{if } \beta >1/2\\
2-\frac{d}{2} &\hbox{if } \beta \leq 1/2 .
\end{cases}
$$
If $\xi$ is bounded one can take $r=1$ and ${\kappa_0} \approx \frac{d}{2}$, thus $2-\kappa_0-c'_0$ can be as close as one wishes to the above upper bounds.
\end{remark}
\begin{remark}
Take $\alpha=1$ and $\beta\leq 1$ so that the integral form of (\ref{space-time}) becomes
\begin{eqnarray*}
u(t,x)=\int^t_0(a^{ij}u_{x^ix^j}+b^iu_{x^i}+cu+f(u))dt+ I^{1-\beta}_t \int^t_0 h(u)dB_t.
\end{eqnarray*}
By the stochastic Fubini theorem, at least formally
$$
I^{1-\beta}_t \int^t_0 h(u)dB_t= \frac{1}{\Gamma(2-\beta)} \int^t_0 h(u(s))(t-s)^{1-\beta}dB_s.
$$
If $\beta=1$ then the classical theory (see e.g. \cite[Section 8]{Krylov1999}) requires $d=1$ to have meaningful solutions, that is locally integrable solutions. By Theorem \ref{thm space-time}, if $\beta<3/4$ then it is possible to take $d=1,2,3$. This might be because the operator $I^{1-\beta}_t$ gives certain smoothing effect to $B_t$ in the time direction.
\end{remark}
|
\section{Introduction} \label{intro}
The purpose of this survey is to give readers interested in operator algebraic quantum group theory a light introduction to the study of approximation properties for locally compact quantum groups. This survey is not a complete one on this subject by any means. Instead, we choose in these notes to focus our attention on four important concepts: amenability, co-amenability, the Haagerup property, and weak amenability.
The study of approximation properties for locally compact groups has a long history, and has many interactions with various branches of mathematics, including probability theory, ergodic theory and dynamics, and (of course) operator algebra theory. From the perspective of operator algebras, the study of group approximation properties gained central importance with the seminal work of Haagerup \cite{Ha78}, who showed that the reduced free group C$^\ast$-algebras $C^*_r(\mathbb{F}_k)$ have the metric approximation property, despite being non-nuclear when $k \ge 2$. Haagerup obtained this result by first showing that the free groups possess a certain approximation property, now called {\it the Haagerup property}, which can be thought of as an amenability-like condition. Building on this work, Cowling and Haagerup \cite{CoHa89} introduced the concept of {\it weak amenability} for locally compact groups and showed that this weak form of amenability is also satisfied by the free groups and real rank one semisimple Lie groups. The connection to operator algebras here is that a discrete group is weakly amenable if and only if its group von Neumann algebra $\mathcal L(G)$ has the weak$^*$ completely bounded approximation property \cite{Ha86}. In recent years, the Haagerup property and weak amenability for groups has also had spectacular application's in Popa's deformation/rigidity program for von Neumman algebras. See \cite{OzPo10,ChSi13} and the references therein.
The theory of locally compact quantum groups generalizes the theory of locally compact groups, and contains many interesting examples arising from various free probabilistic, combinatorial, and non-commutative geometric constructions (see \cite{BaSp,RaWe16, wo1,wo2,wo3}). In particular, locally compact quantum groups give rise to interesting and highly non-trivial quantum analogues of group C$^\ast$-/von Neumann algebras, and it is natural to develop a parallel theory of approximation properties for quantum groups and study its interaction with the structure of these corresponding operator algebras. The main goal of this survey is to develop some of this theory, focusing on both the parallels and the distinctions between the classical and quantum worlds. At this stage in the game, the theory of approximation properties for general locally compact quantum groups is still in its infancy, with may unresolved open problems and issues. It is our hope that this survey will inspire readers to pursue questions on this interesting research topic.
\subsection{Organization of the paper} The rest of this paper is organized as follows: Section \ref{prelim} contains some basic facts on locally compact quantum groups that we will use throughout the paper. In section \ref{amen}, we begin our study of approximation properties with a discussion of amenability and co-amenability for locally compact quantum groups. Here, we study the dualities between amenability and co-amenability and their connections with nuclearity/injectivity for quantum group operator algebras. We also give a proof of the quantum Kesten amenability criterion for discrete quantum groups, and use this to characterize the amenability of free orthogonal quantum groups and Woronowicz' $SU_q(2)$ quantum group. In section \ref{fourieralg}, we start to pave the way for studying more general approximation properties for quantum groups by introducing the Fourier algebra and describing the completely bounded Fourier multipliers on this algebra. In Section \ref{hap/wa}, we use these structures to define the Haagerup property and weak amenability for general locally compact quantum groups, and discuss some basic features of these properties. In Section \ref{discr}, we restrict our attention to discrete quantum groups for the remainder of the paper. Here we study the interaction between (weak) amenability and the Haagerup property with various operator algebra approximation properties, like nuclearity, injectivity, the Haagerup approximation property, and the completely bounded approximation property. In the unimodular case, we prove several equivalences between quantum group approximation properties and operator algebra approximation properties. In Section \ref{CAP} we introduce {\it central} analogues of (weak) amenability and the Haagerup property, and show that these properties are equivalent to their non-central versions in the unimodular case. In Section \ref{CAP-monoidal}, we show how central approximation properties for discrete quantum groups are preserved by monoidal equivalences between their compact duals. Finally, in Section \ref{app} we use the theory developed in the previous sections to study approximation properties for free quantum groups, focusing mainly on the case of free orthogonal quantum groups.
\subsection{Some further reading}
As stated before, we only focus in this paper on a few aspects of the theory of quantum group approximation properties. In particular, we don't say anything about Kazhdan's property (T) for locally compact quantum groups. For the very recent and remarkable developments on this topic, we refer the reader to \cite{ChNg15, Ar15, FiMuPa15}. We also don't consider any of the applications of these approximation properties to structural properties of the operator algebras associated to these quantum groups (e.g. strong solidity, prime factorization results, ...). See \cite{Is15a,Is15b, dCFY}. Another aspect of the theory that we do not touch here are {\it external approximation properties} for quantum groups. For example, there have been very recent and striking developments on questions related to residual finiteness, (inner)linearity, hyperlinearity, and the Kirchberg factorization property for discrete quantum groups. See \cite{BaBi10, BaFrSk12, BrCoVe15, Ch, BaWa16}.
\subsection*{Acknowledgements}
First and foremost, I would like to thank Uwe Franz, Adam Skalski, and Piotr Soltan for organizing such a wonderful graduate school, for inviting me to speak there, and for encouraging me to write this survey. I would also like to thank \'Etienne Blanchard and Stefaan Vaes for kindly sharing with me their unpublished notes \cite{BlVa02}.
\section{Preliminaries} \label{prelim}
In the following, we write $\otimes$ for the minimal tensor product of C$^\ast$-algebras or tensor product of Hilbert spaces, $\overline{\otimes}$ for the spatial tensor product of von Neumann algebras and $\odot$ for the algebraic tensor product. All inner products are taken to be conjugate-linear in the second variable. Given vectors $\xi,\eta$ in a Hilbert space $H$, we denote by $\omega_{\xi,\eta} \in \mathcal B(H)_*$ the vector state $x \mapsto \langle x \xi |\eta\rangle$. If $\xi = \eta$, we simply write $\omega_\xi = \omega_{\xi,\xi}$.
\subsection{Locally compact quantum groups}
Let us first recall from \cite{KuVa00} and \cite{KuVa03} that
a (von Neumann algebraic) \textit{locally compact quantum group (lcqg)} is
a quadruple $\mathbb{G} = (M,\Delta, \varphi, \psi)$, where
$M$ is a von Neumann algebra, $\Delta:M \to M\overline{\otimes}M$ is a co-associative coproduct, i.e. a
unital normal $\ast$-homomorphism such that
\[
(\iota \otimes \Delta) \circ \Delta = (\Delta \otimes \iota) \circ \Delta,
\]
and $\varphi$ and $\psi$ are normal faithful semifinite weights on $M$ such that
\[
(\iota \otimes \varphi)\Delta(x) = \varphi(x) 1 \quad \mbox{and} \quad
(\psi \otimes \iota)\Delta(x) = \psi(x) {1}
\qquad (x\in M^+).
\]
We call $\varphi$ and $\psi$ the
\textit{left Haar weight} and the \textit{right Haar weight} of $\mathbb{G}$, respectively, and
we write $L^\infty(\mathbb{G}) $ for the quantum group von Neumann algebra $M$.
Associated with each locally compact quantum group $\mathbb{G}$, there is a \emph{reduced} quantum group
C*-algebra
$C_0(\mathbb{G})\subseteq L^\infty(\mathbb{G})$ with coproduct
\[
\Delta: x\in C_0(\mathbb{G}) \mapsto \Delta(x) \in M(C_0(\mathbb{G})\otimes C_0(\mathbb{G})) \subseteq L^\infty(\mathbb{G}) \overline {\otimes} L^\infty(\mathbb{G}).
\]
Here we let $M(B)$ denote the {\it multiplier algebra } of a C*-algebra $B$.
Therefore, $(C_0(\mathbb{G}), \Delta)$ together with $\varphi$ and $\psi$ restricted to $C_0(\mathbb{G})$ is a
C*-algebraic locally compact quantum group in the sense of \cite{KuVa00}.
Using the left Haar weight $\varphi$, we can apply the GNS construction to obtain an inner product
\[
\langle {\Lambda_\varphi(x) | \Lambda_\varphi(y)} \rangle = \varphi(y^* x)
\]
on $ \mathfrak N_{\varphi}= \{x \in L^\infty(\mathbb{G}) : \varphi(x^*x) < \infty \}$,
and thus obtain a Hilbert space $L^2(\mathbb{G})= L^2(\mathbb{G}, \varphi)$. Here, $\Lambda_\varphi: \mathfrak N_{\varphi} \to L^2(\mathbb{G})$ is the canonical injection.
The quantum group von Neumann algebra $L^\infty(\mathbb{G})$ is standardly represented on $L^2(\mathbb{G})$ via the unital normal
$\ast$-homomorphism $\pi : L^\infty(\mathbb{G}) \to \mathcal B(L^2(\mathbb{G}))$ satisfying $\pi(x)\Lambda_\varphi(y) = \Lambda_\varphi(xy)$.
There exists a (left) \textit{fundamental unitary operator} $W$ on $L^2(\mathbb{G})\otimes L^2(\mathbb{G})$,
which satisfies the pentagonal relation
\[
W_{12} W_{13} W_{23} = W_{23} W_{12}.
\]
Here, and later,
we use the standard leg number notation: $W_{12} = W\otimes I, W_{13}
= \Sigma_{23} W_{12} \Sigma_{23}$ and $W_{23} = 1 \otimes W$, where
$\Sigma: L^2(\mathbb{G}) \otimes L^2(\mathbb{G}) \to L^2(\mathbb{G}) \otimes L^2(\mathbb{G}) $ is the flip map.
In this case, the coproduct $\Delta$ on $L^\infty(\mathbb{G})$ can be expressed as $\Delta(x) = W^*(1\otimes x)W$.
Let $L^1(\mathbb{G}) = L^\infty(\mathbb{G})_*$ be the predual of $L^\infty(\mathbb{G})$.
Then the pre-adjoint of $\Delta$ induces on $L_1(\mathbb{G})$ a completely contractive Banach algebra product
\[
\star = \Delta_*: \omega \otimes \omega' \in L^1(\mathbb{G}) \widehat \otimes L^1(\mathbb{G}) \mapsto \omega \star \omega' = (\omega \otimes \omega' ) \Delta \in L^1(\mathbb{G}),
\]
where we let $\widehat \otimes$ denote the operator space projective tensor product. We will also use the symbol $\star$ to denote the induced left/right module actions of $L^1(\mathbb{G})$ on $L^\infty(\mathbb{G})$:
\[\omega \star x = (\iota \otimes \omega)\Delta(x) \quad \&\quad x \star \omega = (\omega \otimes \iota) \Delta(x) \qquad (x \in L^\infty(\mathbb{G}), \ \omega \in L^1(\mathbb{G})).\]
Using the Haar weights, one can construct an antipode $S$ on $L^\infty(\mathbb{G})$ satisfying $(S \otimes \iota)W = W^*$. Since, in general, $S$ is
unbounded on $L^\infty(\mathbb{G})$, we can not use $S$ to define an involution on $L^1(\mathbb{G})$.
However, we can consider a dense subalgebra $L^1_\sharp (\mathbb{G})$ of $L^1(\mathbb{G})$, which is defined to be
the collection of all $\omega\in L^1(\mathbb{G})$ such that there exists $\omega^\sharp \in L^1(\mathbb{G})$ with
$\langle \omega^\sharp, x\rangle = \overline{ \langle \omega, S(x)^* \rangle}$ for each $x\in \mathcal{D}(S)$.
It is known from \cite{Ku01} and \cite[Section~2]{KuVa03} that
$L^1_\sharp(\mathbb{G})$ is an involutive Banach algebra with involution $\omega\mapsto\omega^\sharp$
and norm $\|\omega\|_\sharp = \mbox{max}\{\|\omega\|, \|\omega^\sharp\|\}$.
\subsection{Representations of locally compact quantum groups}
A \emph{representation} $(\pi, H)$ of a locally compact quantum group $\mathbb{G}$ is a non-degenerate completely contractive homomorphism $\pi: L^1(\mathbb{G}) \to \mathcal B(H)$ whose restriction to $L^1_\sharp(\mathbb{G})$ is a $\ast$-homomorphism.
It is shown in \cite[Corollary 2.13]{Ku01} that each representation $(\pi, H)$ of $\mathbb{G}$ corresponds uniquely to a unitary operator $U_\pi$
in $M(C_0(\mathbb{G}) \otimes \mathcal K(H))\subseteq L^\infty (\mathbb{G}) \overline{\otimes} \mathcal B(H)$ such that
\[
(\Delta \otimes \iota) U_\pi = U_{\pi, 13} U_{\pi, 23}.
\]
The correspondence is given by
\[
\pi(\omega) = (\omega \otimes \iota) U_\pi \in \mathcal B(H) \qquad (\omega \in L^1(\mathbb{G})).
\]
We call $U_\pi $ the \emph{unitary representation} of $\mathbb{G}$ associated with $\pi$.
Given two representations $(\pi,H_\pi) $ and $ (\sigma ,H_\sigma)$, we can obtain new representations by forming their tensor product and direct sum. The unitary operator
\[
U_{\pi \otimes \sigma} := U_{\pi,12} U_{\sigma,13} \in M(C_0(\mathbb{G}) \otimes\mathcal K(H_\pi \otimes H_\sigma))
\]
determines a representation
\[
\pi \otimes \sigma: \omega \in L^1(\mathbb{G}) \mapsto \mathcal (\omega \otimes \iota)U_{\pi \otimes \sigma} \in \mathcal B(H_\pi \otimes H_\sigma).
\]
We call the representation $({\pi \otimes \sigma}, H_\pi \otimes H_\sigma)$ the \textit{tensor product} of $\pi$ and $\sigma$.
The \emph{direct sum} $(\pi\oplus \sigma, H_\pi \oplus H_\sigma)$ of $\pi$ and $\sigma$ is given by the representation
\[
\pi\oplus \sigma : \omega\in L^1(\mathbb{G}) \to \pi(\omega) \oplus \sigma(\omega)\in \mathcal B(H_\pi \oplus H_\sigma).
\]
For representations $(\pi,H)$ of $\mathbb{G}$, one can consider all of the usual notions of (ir)reducibility, intertwiner spaces, unitary equivalence, and so on, that can be considered for representations of general Banach algebras. In particular, given $\pi,\sigma$ as above, we write
\begin{align*}\text{Mor}(\pi, \sigma) &= \{T \in \mathcal B(H_\pi,H_\sigma): T\pi(\omega) = \sigma(\omega)T \qquad (\omega \in L^1(\mathbb{G})) \}\\
&= \{T \in \mathcal B(H_\pi,H_\sigma): (1 \otimes T)U_\pi = U_\sigma(1 \otimes T)\},
\end{align*}
for the space of intertwiners between $\pi$ and $\sigma$.
The \emph{left regular representation} of $\mathbb{G}$ is defined by
\[
\lambda : \omega \in L^1(\mathbb{G}) \mapsto (\omega \otimes \iota )W\in \mathcal B(L^2(\mathbb{G})).
\]
This is an injective completely contractive homomorphism from $L^1(\mathbb{G})$ into $\mathcal B(L^2(\mathbb{G}))$ such that
$\lambda(\omega^\sharp) = \lambda(\omega)^*$ for all $\omega\in L^1_\sharp(\mathbb{G})$.
Therefore, $\lambda$ is a $*$-homomorphism when restricted to $L^1_\sharp(\mathbb{G})$.
The norm closure of $\lambda(L^1(\mathbb{G}))$ yields C$^\ast$-algebra $C_0(\widehat{\mathbb{G}})$ and the $\sigma$-weak closure of
$\lambda(L^1(\mathbb{G}))$ yields a von Neumann algebra $L^\infty(\widehat{\mathbb{G}})$.
There is a coproduct on $L^\infty(\widehat{\mathbb{G}})$ given by
\[
\hat{\Delta}: \hat x \in L^\infty(\widehat{\mathbb{G}}) \to \hat\Delta(\hat x) = \hat W^*(1\otimes \hat x) \hat W \in
L^\infty(\widehat{\mathbb{G}})\overline{\otimes} L^\infty(\widehat{\mathbb{G}}),
\]
where $\hat W = \Sigma W^*\Sigma$.
We can find
Haar weights $\hat \varphi$ and $ \hat \psi$ to turn $\widehat{\mathbb{G}} = (L^\infty(\widehat{\mathbb{G}}) , \hat \Delta, \hat \varphi, \hat \psi)$
into a locally compact quantum group -- the {\it Pontryagin dual} of $\mathbb{G}$.
Repeating this argument for the dual quantum group $\widehat{\mathbb{G}}$ , we get the left regular representation
\[
\hat \lambda : \hat\omega \in L^1(\widehat{\mathbb{G}}) \mapsto (\hat \omega \otimes \iota )\hat W= (\iota \otimes \hat \omega )W^*
\in \mathcal B(L^2(\mathbb{G})).
\]
It turns out that $C_0(\mathbb{G})$ and $L^\infty(\mathbb{G})$ are just
the norm and $\sigma$-weak closure of $\hat \lambda(L^1(\widehat{\mathbb{G}}))$ in $\mathcal B(L^2(\mathbb{G}))$, respectively.
This gives us the quantum group version of Pontryagin duality $\widehat{\widehat{\mathbb{G}}}= \mathbb{G}$.
There is a \emph{universal representation}
$\pi_u: L^1(\mathbb{G}) \to \mathcal B(H_u)$ and
we obtain the {\it universal quantum group C$^\ast$-algebra} $C^u_0(\widehat{\mathbb{G}}) = \overline{\pi_u(L^1(\mathbb{G}))}^{\|\cdot\|}$.
By the universal property, every representation $\pi: L^1(\mathbb{G}) \to \mathcal B(H)$ uniquely corresponds to a surjective
$\ast$-homomorphism $\hat \pi$ from $C^u_0(\widehat{\mathbb{G}})$
onto the C*-algebra $C_\pi(\widehat{\mathbb{G}}) = \overline{\pi(L^1(\mathbb{G}))}^{\|\cdot\|}$ such that $\pi = \hat \pi \circ \pi_u .$
In particular, the left regular representation $(\lambda, L^2(\mathbb{G}))$ uniquely determines a surjective $\ast$-homomorphism $\hat \pi_\lambda$ from $C^u_0(\widehat{\mathbb{G}})$ onto $ C_0(\widehat{\mathbb{G}})$. Considering the duality, we can obtain the universal quantum group C$^\ast$-algebra $C^u_0(\mathbb{G}) $ and we denote by $ \pi_{\hat \lambda} : C^u_0(\mathbb{G}) \to C_0(\mathbb{G})$ the canonical surjective $\ast$-homomorphism. As shown in \cite{Ku01}, $C^u_0(\mathbb{G})$ admits a coproduct $\Delta_u:C^u_0(\mathbb{G}) \to M(C^u_0(\mathbb{G}) \otimes C^u_0(\mathbb{G}))$ which turns $(C^u_0(\mathbb{G}), \Delta_u)$ into a C$^\ast$-algebraic locally compact quantum group with left and right Haar weights given by $\varphi \circ \pi_{\hat \lambda}$ and $\psi \circ \pi_{\hat \lambda}$, respectively.
It is also shown in \cite{Ku01} that the fundamental unitary $W$ of $\mathbb{G}$ admits a ``semi-universal'' version $\text{\reflectbox{$\mathbb W$}}\in M(C_0(\mathbb{G}) \otimes C^u_0(\widehat{\mathbb{G}}))$ which has the property that for each representation $\pi: L^1(\mathbb{G}) \to \mathcal B(H)$, the corresponding unitary operator $U_\pi$ can be expressed as
\[(\iota \otimes \hat \pi)\text{\reflectbox{$\mathbb W$}} = U_\pi.
\]
Moreover, $\text{\reflectbox{$\mathbb W$}}$ satisfies the following relations
\begin{align} \label{bichar_relation}
(\Delta \otimes \iota) \text{\reflectbox{$\mathbb W$}} = \text{\reflectbox{$\mathbb W$}}_{13}\text{\reflectbox{$\mathbb W$}}_{23} ~\mbox{and} ~
(\iota \otimes \hat \Delta_u) \text{\reflectbox{$\mathbb W$}} = \text{\reflectbox{$\mathbb W$}}_{13} \text{\reflectbox{$\mathbb W$}}_{12}.
\end{align}
\subsection{Compact and discrete quantum groups} A locally compact quantum group is called
\textit{compact} if $C_0(\mathbb{G})$ is unital, or equivalently $\varphi = \psi$ is a $\Delta-$bi-invariant state (up to a scale factor). In this case, we will write $C(\mathbb{G})$ for $C_0(\mathbb{G})$.
We say $\mathbb{G}$ is \textit{discrete} if $\widehat{\mathbb{G}}$ is a compact quantum group, or equivalently, if $L^1(\mathbb{G})$ is unital. For a self-contained treatment of compact/discrete quantum groups and their duality, we refer to \cite{PoWo90, EfRu94, Ti08}.
Let $\mathbb{G}$ be a discrete quantum group. Denote by $\text{Irr}(\widehat{\mathbb{G}})$ the collection of equivalence classes of finite dimensional irreducible representations of $\widehat{\mathbb{G}}$. For each $\pi \in \text{Irr}(\hat \mathbb{G})$, select a representative unitary representation $( U^\pi,H_\pi)$. Then $U^\pi$ can be identified with a unitary matrix
\[
U^\pi = [ u_{ij}^\pi] \in M_{n(\pi)}( C(\widehat{\mathbb{G}})),
\]
where $n(\pi) = \dim H_\pi$. The linear subspace $\mathcal O(\widehat{\mathbb{G}}) \subseteq C(\widehat{\mathbb{G}})$ spanned by $\{ u_{ij}^\pi : \pi \in \text{Irr}(\widehat{\mathbb{G}}), \ 1 \le i,j \le n(\pi)\}$ is a dense Hopf-$\ast$-subalgebra of $C(\widehat{\mathbb{G}})$. The coproduct on $\mathcal O(\widehat{\mathbb{G}})$ is given by restricting $\hat \Delta $ to $\mathcal O(\widehat{\mathbb{G}})$, and we have
\[
\hat \Delta (u^\pi_{ij} ) = \sum_{k=1}^{n(\pi)}u^\pi_{ik} \otimes \hat u^\pi_{kj}.
\]
We call $\mathcal O(\widehat{\mathbb{G}})$ the \textit{algebra of polynomial functions on $\widehat{\mathbb{G}}$}.
The Haar state $\hat \varphi = \hat \psi$ is always faithful when restricted
to $\mathcal O(\widehat{\mathbb{G}})$, and $C^u_0(\widehat{\mathbb{G}})$ can be identified with the universal enveloping C$^\ast$-algebra of $\mathcal O(\widehat{\mathbb{G}})$. This allows us to regard $\mathcal O(\widehat{\mathbb{G}})$ as a dense $\ast$-subalgebra of $C^u_0(\widehat{\mathbb{G}})$, and the semi-universal fundamental unitary of
the discrete quantum group $\mathbb{G}$ is then given by
\begin{align} \label{bichar_compact} \text{\reflectbox{$\mathbb W$}} = \bigoplus_{\pi \in \text{Irr}(\widehat{\mathbb{G}})} ( U^\pi)^* = \bigoplus_{\pi \in \text{Irr}(\widehat{\mathbb{G}})} \sum_{1 \le i,j \le d_\alpha} e_{ij}^{\pi} \otimes u_{ji}^{\pi \ast} \in M(C_0( \mathbb{G}) \otimes C^u_0(\widehat{\mathbb{G}}).
\end{align}
In general, for any discrete quantum group $\mathbb{G}$, we have
\[C_0(\mathbb{G}) =\bigoplus_{\pi \in \text{Irr}(\widehat{\mathbb{G}})}^{c_0} \mathcal B(H_\pi) \quad \mbox{and} \quad L^\infty(\mathbb{G}) = \prod_{\pi \in \text{Irr}(\widehat{\mathbb{G}})} \mathcal B(H_\pi).
\]
Denote by $C_c(\mathbb{G})$ the (algebraic) direct sum $\bigoplus_{\pi \in \text{Irr}(\widehat{\mathbb{G}})} \mathcal B(H_\pi)$. Then
$C_c(\mathbb{G})$ forms a common core for the Haar weights $\varphi$ and $\psi$ on $\mathbb{G}$.
For each $\pi \in \text{Irr}(\widehat{\mathbb{G}})$, denote by $p_\pi \in L^\infty(\mathbb{G})$ the minimal central projection whose support is $\mathcal B(H_\pi)$.
Let $\text{Tr}_{\pi}$
be the canonical trace on $\mathcal B(H_\pi)$ (with $\text{Tr}_{\pi}(1) = n(\pi)$).
One can then find positive invertible matrices $Q_\pi \in \mathcal B(H_\pi)$ which satisfy
\[
d(\pi) =\text{Tr}_\pi(Q_\pi) = \text{Tr}_\alpha(Q_\pi^{-1}) \quad \mbox{and} \quad
\Delta(Q_\pi) = Q_\pi \otimes Q_\pi.
\] The left and right Haar weights on $\mathbb{G}$ are given by the formulas
\begin{align} \label{eqn:weights}
\varphi(p_\pi x)= d(\pi) \text{Tr}_\alpha(Q_\pi^{-1} p_\pi x) \quad \mbox{and}
\qquad \psi(p_\pi x) = d(\pi) \text{Tr}_\pi(Q_\pi p_\pi x).
\end{align} See \cite[Equations (2.12)-(2.13)]{PoWo90}.
The number $d(\pi)$ appearing above is called the {\it quantum dimension} of $\pi$. Note that $d(\pi) \ge n(\pi)$, and we have equality for all $\pi$ if and only if $\varphi = \psi$. In this case, we say that $\mathbb{G}$ is a {\it unimodular discrete quantum group}.
The matrices $Q_{\pi} \in \mathcal B(H_\pi)$ defined above can be used to describe many other structures associated to $\mathbb{G}$ and $\widehat{\mathbb{G}}$. For $\mathbb{G}$, the modular automorphism group $\sigma_t$ turns out to be equal to the scaling group $\tau_t$, and \[\sigma_t(x) = \tau_t(x) = Q_\pi^{-it}xQ_\pi^{it} \qquad (x \in \mathcal B(H_\pi), \ t \in \mathbb{R}). \]
On the dual side, the {\it Schur orthogonality relations} for the Haar state $\hat \varphi$ are given by
\[\hat \varphi(u^\pi_{ij}(u^{\pi'}_{kl})^*) = \frac{\delta_{\pi\pi'}\delta_{ik}Q_{\pi, lj}}{d(\pi)} \quad \& \quad \hat \varphi((u^\pi_{ij})^*u^{\pi'}_{kl}) = \frac{\delta_{\pi\pi'}\delta_{jl}(Q_{\pi}^{-1})_{ki}}{d(\pi)}. \]
The {\it Woronowicz characters} $(f_z)_{z \in \mathbb{C}}$ of $\widehat{\mathbb{G}}$ is the family of characters $f_z: \mathcal O(\widehat{\mathbb{G}}) \to \mathbb{C}$ given by
\[f_z(u_{ij}^\pi) = (Q_\pi^z)_{ij} \qquad (\pi \in \text{Irr}(\widehat{\mathbb{G}}), \ 1 \le i,j \le n(\pi)).\] Note that \[f_z\star f_w = f_{z+w} \quad \& \quad f_{z}(x^*) = \overline{f_{-\bar z}(x)} \qquad (z,w \in \mathbb{C}, \ x \in \mathcal O(\widehat{\mathbb{G}})).\]
So that $f_z$ is a $\ast$-character if $z \in i\mathbb{R}$. The Woronowicz characters can also be used to describe the antipode $\hat S$, the (analytic continuation of the) scaling group $\hat \tau_z$ and modular group $\hat \sigma_z$, and the unitary antipode $\hat R$:
\[\hat \tau_z(x) = f_{-iz}\star x\star f_{iz}, \quad \hat S^2(x) = \hat\tau_{-i}(x), \quad \hat\sigma_z(x) = f_{iz}\star x\star f_{iz}, \quad \hat S \circ \hat \tau_{i/2}. \]
\subsection{A few basic examples} We mention here a few examples of compact quantum groups that we will come to at several points in this paper. These are the free unitary quantum groups, the free orthogonal quantum groups, and Woronowicz' $SU_q(2)$ quantum group. We leave it to the reader to check that the following examples do indeed satisfy Woronowicz' axioms for a compact (matrix) quantum group \cite{wo2}.
\begin{notat}
Given a complex $\ast$-algebra $A$ and a matrix $X = [x_{ij}] \in M_N(A)$, we denote by $\bar X$ the matrix $[x_{ij}^*] \in M_N(A)$.
\end{notat}
\begin{defn}[\cite{VaWa96}] Let $N \ge 2$ be an integer and let $F \in \text{GL}_N(\mathbb{C})$.
\begin{enumerate}
\item The \textit{free unitary quantum group} $U_F^+$ (with parameter matrix $F$) is the compact quantum group given by the universal C$^\ast$-algebra
\begin{align} \label{eqn:defining} C^u(U^+_F) = C^*\big(\{v_{ij}\}_{1 \le i,j \le N} \ | \ V = [v_{ij}] \text{ is unitary } \& \ F \bar V F^{-1} \text{ is unitary } \big),\end{align} together with coproduct $\Delta:C^u(U^+_F) \to C^u(U^+_F) \otimes C^u(U_F^+)$ given by
\[\Delta(v_{ij}) = \sum_{k=1}^N v_{ik} \otimes v_{kj} \qquad (1 \le i,j \le N).\]
\item Let $\lambda \in \mathbb{R}\backslash \{0\}$ and assume that $F \bar F = \lambda 1$. The \textit{free orthogonal quantum group} $O_F^+$ (with parameter matrix $F$) is the compact quantum group given by the universal C$^\ast$-algebra
\begin{align} \label{eqn:defining} C^u(O^+_F) = C^*\big(\{u_{ij}\}_{1 \le i,j \le N} \ | \ U = [u_{ij}] \text{ is unitary } \& \ U = F \bar U F^{-1}\big),\end{align} together with coproduct $\Delta:C^u(O^+_F) \to C^u(O^+_F) \otimes C^u(O_F^+)$ given by
\[\Delta(u_{ij}) = \sum_{k=1}^N u_{ik} \otimes u_{kj} \qquad (1 \le i,j \le N).\]
\end{enumerate}
\end{defn}
\begin{rem}
In the above definition, the coproduct $\Delta$ is defined so that the matrix of generators $V = [v_{ij}]$ (resp. $U = [u_{ij}]$) is always a (fundamental) representation of the compact matrix quantum group $U^+_F$ (resp. $O_F^+$).
\end{rem}
\begin{rem}
Note that the above definition for $O^+_F$ makes sense for any $F\in \text{GL}_N(\mathbb{C})$. The additional condition $F \bar F \in \mathbb{R} 1$ is equivalent to the requirement that $U$ is always an {\it irreducible} representation of $O^+_F$. Indeed, Banica \cite{Ba96} showed that $U$ is irreducible if and only if $F\bar F = \lambda 1$ ($\lambda \in \mathbb{R}$).
\end{rem}
We now define Woronowicz' $SU_q(2)$ quantum group.
\begin{defn}[\cite{wo2}]
Let $q \in [-1,1]\backslash \{0\}$ and let $C(SU_q(2)) = C^*(\alpha, \gamma)$ be the universal unital C$^\ast$-algebra generated by elements $\alpha, \gamma$ subject to the relations which make the $2 \times 2$ matrix
\[W = [w_{ij}] = \left(\begin{matrix} \alpha & -q\gamma^*\\
\gamma & \alpha^* \end{matrix}\right) \quad \text{unitary}.
\]
Defining $\Delta(w_{ij}) = \sum_{k=1}^2 w_{ik} \otimes w_{kj}$, $\Delta$ extends to a co-product on $C(SU_q(2))$, yielding Woronowicz' $SU_q(2)$-quantum group.
\end{defn}
\begin{rem}
Actually, the quantum groups $SU_q(2)$ are just spacial cases of free orthogonal quantum groups. Indeed, for $q \in [-1,1]\backslash \{0\}$, we have \[SU_q(2) = O^+_{F_q} \quad \text{where} \quad F_q = \left( \begin{matrix} 0 &1\\
-q^{-1} & 0 \end{matrix}\right) \in \text{GL}_{2}(\mathbb{C}).\]
That is, there is a $\ast$-isomorphism $C(O^+_{F_q}) \to C(SU_q(2))$ given by $u_{ij} \mapsto w_{ij}$. This fact can be readily checked by comparing generators and relations.
\end{rem}
\section{Amenability and co-amenability} \label{amen}
We begin our discussion of approximation properties of locally compact quantum groups with arguably the most fundamental and natural approximation properties: {\it amenability}. Recall that a locally compact group $G$ is said to be amenable if there exists a finitely additive, left translation-invariant Borel probability measure on $G$. In functional analytic language, this means the existence of a state $m \in L^\infty(G)^*$ such that
\[
m(L_xf) = m(f) \qquad (f \in L^\infty(G), \ g \in G),
\]
where $L_gf: t \mapsto f(g^{-1}t)$ is the left translate of $f$ by $g$. Such a state $m$ is called a {\it left invariant mean} on $L^\infty(G)$. The class of amenable groups is quite broad and includes, for example, all compact, abelian, solvable, and nilpotent locally compact groups. For a very nice introduction to the topic of amenability for locally compact groups, we refer the reader to the monographs \cite{Pa, Ru}.
Our goal now is to extend he concept of amenability to general locally compact quantum groups. To do this, we first observe that if $m \in L^\infty(G)^*$ is left-invariant, then $m$ is also invariant under the natural dual right action of $L^1(G)$ on $L^\infty(G)$:
\[
(f, \omega) \mapsto f \star \omega = (\omega \otimes \iota)\Delta(f) \qquad (\omega \in L^1(G), \ f \in L^\infty(G)),
\]
where $\Delta f (s,t) = f(st)$ is the usual corproduct on $L^\infty(G)$.
Thus, if $m$ is a left-invariant mean on $L^\infty(G)$, then
\begin{align} \label{top-inv}
m(\omega \otimes \iota)\Delta(f)= m(f) \int_{G}\omega(x)dx = \omega(1)m(f) \qquad (\omega \in L^1(G), \ f \in L^\infty(G)).
\end{align}
In \cite{Ru}, the above condition \eqref{top-inv} is called {\it topological left invariance} for $m$. Although topological left-invariance is an a priori weaker notion than left-invariance, it is known \cite[Chapter 1]{Ru} that the existence of a topologically invariant mean on $L^\infty(G)$ is equivalent to amenability of $G$. Since the notion of topological left-invariance translates directly to the quantum setting, we make the following definition.
\begin{defn}
A lcqg is $\mathbb{G}$ is called {\it amenable} if there exists a state $m \in L^\infty(\mathbb{G})^*$ such that $m(\omega \otimes \iota)\Delta = \omega(1)m$ for all $\omega \in L_1(\mathbb{G})$. We call such a state $m$ a left-invariant mean on $L^\infty(\mathbb{G})$.
\end{defn}
\begin{rem}
There is the obvious notion of a right-invariant mean $m \in L^\infty(\mathbb{G})^*$, which satisfies $m(\iota \otimes \omega)\Delta = \omega(1)m$ for all $\omega \in L^1(\mathbb{G})$. Hence we could define amenability in terms of right-invariant means. It turns out that our preference for left over right is irrelevant, since an easy calculation shows that a state $m \in L^\infty(\mathbb{G})^*$ is a left-invariant mean if and only if $m\circ R$ is a right-invariant mean, where $R = S \circ \tau_{i/2}:L^\infty(\mathbb{G}) \to L^\infty(\mathbb{G})$ is the unitary antipode. (Recall that $R$ is a $\ast$-anti-automorphism, $R^2 = \iota$, and $\Delta = \sigma (R \otimes R)\Delta R$, where $\sigma$ is the tensor flip map.)
\end{rem}
\subsection{Some examples of amenable lcqgs}
At this point we should mention at least a few non-trivial examples: Every compact quantum group is clearly amenable (it's Haar state is an invariant mean), among the discrete quantum groups we will see later on that Woronowicz' $\widehat{SU_q(2)}$ is amenable and $\widehat{O^+_F}$ is amenable if and only if $F \in \text{GL}_2(\mathbb{C})$. In fact, it is more generally known that the dual $\widehat{G_q}$ of the $q$-deformation of a simply connected semi-simple compact Lie group $G$ is amenable (see, for example, \cite{NeTu13} and the references therein). In the locally compact quantum case, we just content ourselves to mention the work of Caspers \cite{Ca14}, showing that that dual of $SU_q(1,1)$ is amenable.
\subsection{Amenability, nuclearity, and injectivity}
In this section we will show how amenability for a lcqg is intimately related to two of the most fundamental approximation properties in operator algebra theory: injectivity and nuclearity. Let is begin by recalling these concepts.
\begin{defn}
Let $H$ be a Hilbert space and $M \subset \mathcal B(H)$ a von Neumann algebra. $M$ is called {\it injective} if there exists a conditional expectation $E:B(H) \to M$. I.e., a unital completely positive map $E$ from $\mathcal B(H)$ onto $M$ such that $E^2 = E$.
\end{defn}
Although the above definition of injectivity may, at the face of it, appear to not have anything to do with ``approximations'', it turns out to be equivalent to the weak* completely positive approximation property, thanks to some deep work of Connes \cite{co1}. For future reference, let us recall this concept.
\begin{defn} \label{wcpap}
A von Neumann algebra $M$ has the {\it weak* completely positive approximation property (w*CPAP)} if there exists a net of finite rank, $\sigma$-weakly continuous, completely positive contractions $T_\alpha:M \to M$ such that
\[
T_\alpha \to \text{id}_M \qquad \text{pointwise $\sigma$-weakly}.
\]
\end{defn}
In the C$^\ast$-algebra world, the corresponding analogue of injectivity is nuclearity. Just like injectivity, nuclearity has many equivalent characterizations (see \cite[Chapter 2]{BrOz08} for example). In the following, we take the so-called {\it completely positive approximation property} as our definition of nuclearity.
\begin{defn} \label{cpap}
A C$^\ast$-algebra is {\it nuclear} if there exists a net of finite rank, completely positive contractions $T_\alpha:A \to A$ such that
\[
T_\alpha \to \text{id}_A \qquad \text{pointwise in norm}.
\]
\end{defn}
We now come to a result that connects amenability to injectivity/nuclearity.
\begin{thm}[\cite{BeTu03}] \label{injectivity}
Let $\mathbb{G}$ be an amenable lcqg with Pontryagin dual $\widehat{\mathbb{G}}$, and let $\hat\pi:C_0^u(\widehat{\mathbb{G}}) \to \mathcal B(H)$ be a non-degenerate $\ast$-representation. Then $\hat\pi(C_0^u(\widehat{\mathbb{G}}))$ is nuclear, and $\hat\pi(C_0^u(\widehat{\mathbb{G}}))''$ is injective. In particular, $C_0^u(\widehat{\mathbb{G}})$ and $C_0(\widehat{\mathbb{G}})$ are nuclear, and $L^\infty(\widehat{\mathbb{G}})$ is injective.
\end{thm}
\begin{proof}
Let $\pi: L^1(\mathbb{G}) \to \mathcal B(H)$ be the associated non-degenerate $\ast$-representation and let $U = U_\pi \in M(C_0(\mathbb{G})\otimes \mathcal K(H))$ be the corresponding unitary representation of $\mathbb{G}$. First note that the injectivity of $\pi(C_0^u(\widehat{\mathbb{G}}))''$ is equivalent to the nuclearity of $\pi(C_0^u(\widehat{\mathbb{G}}))$ by some very deep work of Connes \cite{co1} and Choi-Effros \cite{ChEf77}. Moreover, by a result of Tomiyama \cite{To}, injectivity of $\pi(C_0^u(\widehat{\mathbb{G}}))''$ is equivalent to that of $\pi(C_0^u(\widehat{\mathbb{G}}))'$. It is this latter fact that we will now prove.
To this end, let $m$ be a right-invariant mean on $L^\infty(\mathbb{G})$, let $\alpha:\mathcal B(H) \to L^\infty(\mathbb{G}) \overline{\otimes }\mathcal B(H)$ be given by
\[
\alpha(x) = U(1 \otimes x)U^* \qquad (x \in \mathcal B(H)),
\] and define $E:\mathcal B(H) \to \mathcal B(H)$ by \[E(x) = (m \otimes \iota)(\alpha(x)) \qquad (x \in \mathcal B(H)). \]
It is clear from the definitions that $E$ is a unital and completely positive (ucp) map. We now show that $E(\mathcal B(H)) = \pi(C_0^u(\widehat{\mathbb{G}}))'$ and $E^2 = E$. Since $(\Delta \otimes \iota)U = U_{13}U_{23}$, a simple calculation shows that
\[
(\iota \otimes \alpha) \alpha = (\Delta \otimes \iota) \alpha.
\]
From this equality, we get
\begin{align*}
\alpha(E(x)) &= (m \otimes \alpha)\alpha(x) = (m \otimes \iota) (\Delta \otimes \iota)(\alpha(x)) \\
&=1 \otimes (m \otimes \iota)\alpha(x) = 1 \otimes E(x) \qquad (x \in \mathcal B(H)),
\end{align*}
where the last equality follows from the right-invariance of $m$. As a consequence, \[E^2(x) = (m \otimes \iota)(\alpha(E(x))) = (m \otimes \iota) (1 \otimes E(x)) = E(x) \qquad (x \in \mathcal B(H)).\]
Next, observe that $E(\mathcal B(H)) = \{y \in \mathcal B(H): \alpha(y) = 1 \otimes y\}$. Indeed, we have already shown the inclusion ``$\subseteq$'', and the inclusion ``$\supseteq$'' follows because the equality $\alpha(y) = 1 \otimes y$ implies that $E(y) = (m \otimes \iota)(1 \otimes y) = y$.
To finish, we note that
\begin{align*}
\{y \in \mathcal B(H): \alpha(y) = 1 \otimes y\} &= \{y \in \mathcal B(H): (1 \otimes y)U = U(1 \otimes y)\} \\
&=\{y \in \mathcal B(H): \pi(\omega)y = y\pi(\omega) \ \forall \omega\in L^1(\mathbb{G})\} \\
&= \pi(L^1(\mathbb{G}))' = \hat \pi(C^u_0(\widehat{\mathbb{G}}))'.
\end{align*}
\end{proof}
The above theorem raises the following natural question about the connection between injectivity/nuclearity and amenability.
\begin{quest}
For a locally compact quantum group $\mathbb{G}$, does the injectivity of $L^\infty(\widehat{\mathbb{G}})$ or the nuclearity of $C_0(\widehat{\mathbb{G}})$ (respectively $C^u_0(\widehat{\mathbb{G}})$) characterize the amenability of $\mathbb{G}$?
\end{quest}
The answer to the above question (as stated) turns out to be no, even in the classical case. Let $G$ be any connected locally compact group. Connes \cite{co1} showed for instance that the group von Neumann algebra $\mathcal L(G) = L^\infty(\hat G)$ is always injective and that the reduced group C$^\ast$-algebra $C^*_\lambda(G) = C_0(\hat G)$ is always nuclear. On the other hand, we will see later in Theorem \ref{unimodular} that the answer is yes, provided $\mathbb{G}$ is a unimodular discrete quantum group. In the general case, one has to impose some additional hypotheses on the conditional expectations $E:B(L^2(\mathbb{G})) \to L^\infty(\widehat{\mathbb{G}})$ in order to yield a characterization of amenability.
One such additional hypothesis is given by the following definition.
\begin{defn}[Soltan-Viselter \cite{SoVi14}]
A lcqg $\mathbb{G}$ is called {\it quantum injective} if there exists a conditional expectation $E:\mathcal B(L^2(\mathbb{G})) \to L^\infty(\mathbb{G})$ such that $E(L^\infty(\widehat{\mathbb{G}})) \subset ZL^\infty(\mathbb{G})$ (where $ZL^\infty(\mathbb{G})$ denotes the center of $L^\infty(\mathbb{G})$).
\end{defn}
It turns out that if one considers quantum injectivity instead of ordinary injectivity, then one obtains a characterization of amenability. The theorem presented below is from \cite{SoVi14}, but similar results were obtained independently by Crann-Neufang in \cite{CrNe16}.
\begin{thm}[Soltan-Viselter \cite{SoVi14}, Crann-Neufang \cite{CrNe16}]
The following conditions are equivalent for a lcqg $\mathbb{G}$:
\begin{enumerate}
\item $\mathbb{G}$ is amenable.
\item There exists a conditional expectation $E:\mathcal B(L^2(\mathbb{G})) \to L^\infty(\widehat{\mathbb{G}})$ with $E(L^\infty(\mathbb{G})) = \C1$.
\item $\widehat{\mathbb{G}}$ is quantum injective.
\end{enumerate}
\end{thm}
\begin{proof}[Sketch]
$(1) \implies (2)$. Fix a left-invariant mean $m$ and consider the right-regular representation $V \in L^\infty(\widehat{\mathbb{G}})'\overline{\otimes }L^\infty(\mathbb{G})$, which is unitary operator that implements the coproduct on $L^\infty(\mathbb{G})$ via $\Delta(x) = V(x \otimes 1)V^*$. Define $E:\mathcal B(L^2(\mathbb{G})) \to \mathcal B(L^2(\mathbb{G}))$ by \[\omega(E(x)) = m(\omega \otimes \iota)(V(x \otimes 1)V^*) \qquad (x \in \mathcal B(L^2(\mathbb{G})), \omega \in \mathcal B(L^2(\mathbb{G}))_*).\]
Then, as in the proof of Theorem \ref{injectivity}, one can show $E:\mathcal B(L^2(\mathbb{G})) \to (L^\infty(\widehat{\mathbb{G}})')' = L^\infty(\widehat{\mathbb{G}})$ is a conditional expectation. Moreover \[E(x) = (\iota\otimes m)(V(x \otimes 1)V^*) = (\iota\otimes m)\Delta(x) = m(x)1 \qquad (x \in L^\infty(\mathbb{G})). \]
$(2) \implies (3)$ is trivial.
$(3) \implies (1)$.
Let $E$ be a conditional expectation satisfying $(3)$. Fix a state $\rho \in L^\infty(\widehat{\mathbb{G}})^*$ and define a state $m \in L^\infty(\mathbb{G})^*$ by
\[
m = \rho \circ E|_{L^\infty(\mathbb{G})}.
\]
We claim that $m$ is a left-invariant mean on $L^\infty(\mathbb{G})$. To see this, denote by \[\bar \Delta: \mathcal B(L^2(\mathbb{G})) \to \mathcal B(L^2(\mathbb{G}))\overline{\otimes} \mathcal B(L^2(\mathbb{G})); \qquad \bar \Delta(x) = W^*(1 \otimes x)W\] the canonical extension to $\mathcal B(L^2(\mathbb{G}))$ of the coproduct $\Delta$. Since any conditional expectation $E:\mathcal B(L^2(\mathbb{G})) \to L^\infty(\widehat{\mathbb{G}})$ is automatically an $L^\infty(\widehat{\mathbb{G}})$-bimodule map \cite{To} (i.e., $E \in \mathcal {CB}_{L^\infty(\widehat{\mathbb{G}})}(\mathcal B(L^2(\mathbb{G})))$), we can use a result of Effros-Kishimoto \cite{EfKi87}, which says that $E$ is a pointwise $\sigma$-weak limit of so-called {\it elementary operators}:
\[
E = \lim_\alpha E_\alpha \quad \text{pointwise $\sigma$-weakly, where} \quad E_\alpha(x) = \sum_{i=1}^{n_\alpha} a_{i,\alpha} xb_{i,\alpha} \qquad (a_{i,\alpha}, b_{i,\alpha} \in L^\infty(\widehat{\mathbb{G}})').
\]
Now fix $\omega \in L^1(\mathbb{G})$ and $x \in L^\infty(\mathbb{G})$. Then using the fact that $W \in L^\infty(\mathbb{G}) \overline{\otimes} L^\infty(\widehat{\mathbb{G}})$, we get the following $\sigma$-weak limit.
\begin{align*}
E(\omega \otimes \iota)\bar\Delta(x) &=E(\omega \otimes \iota)(W^*(1 \otimes x)W) \\
&=\lim_\alpha \sum_{i=1}^{n_\alpha} (\omega \otimes \iota) (W^*(1 \otimes a_{i,\alpha} xb_{i,\alpha})W) \\
&= (\omega \otimes \iota)W^*(1 \otimes E(x))W \\
&= (\omega \otimes \iota) (1 \otimes E(x)),
\end{align*}
where the last equality follows from the fact that $E(L^\infty(\mathbb{G})) \subseteq ZL^\infty(\widehat{\mathbb{G}})$. Applying $\rho$ to both sides of the above equality, we get
\[
m(\omega\otimes \iota)\Delta(x) = m(x)\omega(1) \qquad (x \in L^\infty (\mathbb{G})).
\]
\end{proof}
\subsection{Co-amenability}
Up to this point, we have essentially two ways to detect the amenability of a lcqg $\mathbb{G}$. Namely, finding a left or right-invariant mean on $L^\infty(\mathbb{G})$, or constructing a suitable conditional expectation $E:\mathcal B(L^2(\mathbb{G})) \to L^\infty(\widehat{\mathbb{G}})$. In order to develop further (possibly simpler) means of checking amenability, we now turn to a notion dual to amenability, called co-amenability.
\begin{defn}
A locally compact quantum group $\mathbb{G}$ is called {\it co-amenable} if there exists a state $\epsilon:C_0(\mathbb{G}) \to \mathbb{C}$ such that $(\iota \otimes \epsilon)\Delta = \iota$. Such a state $\epsilon$ is called a {\it co-unit} for $C_0(\mathbb{G})$. \end{defn}
\begin{lem} If the co-unit $\epsilon$ exists, then it is unique.
\end{lem}
\begin{proof}
Let $R$ denote the unitary antipode for $L^\infty(\mathbb{G})$. Since $\sigma(R\otimes R)\Delta = \Delta R$, we have $(\epsilon R \otimes \iota)\Delta = \iota$. But then
\[
\epsilon R = \epsilon R(\iota \otimes \epsilon)\Delta =\epsilon(\epsilon R \otimes \iota)\Delta = \epsilon \implies (\iota \otimes \epsilon )\Delta = (\epsilon \otimes \iota)\Delta = \iota.
\]
Thus, if $\epsilon'$ is any other co-unit, we have $\epsilon' = (\epsilon' \otimes \epsilon)\Delta = \epsilon$.
\end{proof}
In contrast to the notion of amenability, co-amenability turns out to have many equivalent characterizations, several of which can be thought of as approximation properties for the quantum group (or its dual).
\begin{thm} \label{co-amen-characterization}
The following conditions are equivalent for a lcqg $\mathbb{G}$:
\begin{enumerate}
\item $\mathbb{G}$ is co-amenable.
\item $C_0(\mathbb{G})^*$ is unital (with respect to the convolution algebra structure induced by $\Delta$).
\item The canonical quotient map $\pi_u:C_0^u(\mathbb{G}) \to C_0(\mathbb{G})$ is injective.
\item There is a state $\epsilon \in C_0(\mathbb{G})^*$ such that $( \epsilon \otimes \iota) W = 1$.
\item There is a net $(\xi_i)_{i} \subset L^2(\mathbb{G})$ of unit vectors such that \[\lim_i\|W(\xi_i \otimes \eta) - (\xi_i \otimes \eta)\| = 0 \qquad (\eta \in L^2(\mathbb{G})).\]
\item $L^1(\mathbb{G})$ has a bounded (left/right/two-sided) approximate identity.
\end{enumerate}
\end{thm}
\begin{proof}
$(1) \iff (2) \iff (3)$: Clearly $\epsilon$ plays the role of the unit in the Banach algebra $C_0(\mathbb{G})^*$, so (1) and (2) are equivalent. Note that since \[L^1(\mathbb{G}) \subseteq C_0(\mathbb{G})^* \subseteq C^u_0(\mathbb{G})^*\] are identified as closed two-sided ideals \cite{Ku01}, it follows that $C_0(\mathbb{G})^*$ is unital if and only if $C_0(\mathbb{G})^* = C_0^u(\mathbb{G})^*$, which happens if and only if $\pi_u$ is injective.
$(1) \implies (4)$: From (1), we have
\[W = (\iota \otimes \epsilon \otimes \iota)(\Delta \otimes \iota)W =(\iota \otimes \epsilon \otimes \iota)W_{13}W_{23} = W(1 \otimes (\epsilon \otimes \iota)W).\]
Which implies that $(\epsilon \otimes \iota)W = 1$, since $W$ is unitary.
$(4) \implies (5)$: Let $\epsilon \in C_0(\mathbb{G})^*$ satisfy $(4)$, and extend it to a state $\tilde \epsilon$ on $\mathcal B(L^2(\mathbb{G}))$. Since $L^\infty(\mathbb{G})$ is in standard form on $L^2(\mathbb{G})$, we can find a net of
unit vectors $(\xi_i)_i\subset L^2(\mathbb{G})$ such that \[\tilde \epsilon(x) = \lim_i \omega_{\xi_i}(x) \qquad (x \in L^\infty(\mathbb{G})).\] But then for all $\eta \in L^2(\mathbb{G})$, we have $(\iota \otimes \omega_\eta)W \in C_0(\mathbb{G}) \subset L^\infty(\mathbb{G})$, and therefore
\begin{align*}
\|W(\xi_i \otimes \eta) - (\xi_i \otimes \eta)\|^2 &= 2\|\eta\|^2 - 2 \text{Re}(W(\xi_i \otimes \eta)|\xi_i \otimes \eta) \\
&= 2\|\eta\|^2- 2 \text{Re} (\omega_{\xi_i} \otimes \omega_\eta)W\\
&\to 2\|\eta\|^2- 2 \text{Re} (\epsilon \otimes \omega_\eta)W \\
&= 2\|\eta\|^2- 2 \omega_\eta(1) =0.
\end{align*}
$(5) \implies (6)$: Consider the net $(\omega_{i})_i \subset L^1(\mathbb{G})$ with $\omega_{i} = \omega_{\xi_i}$. Then for any $\omega = \omega_{u,v} \in L^1(\mathbb{G})$ and $x \in L^\infty(\mathbb{G})$,
\begin{align*}
&|\omega_i\star\omega(x) - \omega(x)|\\
&= |\langle(1 \otimes x)W(\xi_i \otimes u)|W\xi_i \otimes v\rangle - \langle(1 \otimes x)(\xi_i \otimes u)|\xi_i \otimes v\rangle| \\
&= |\langle(1 \otimes x)(W(\xi_i \otimes u)- (\xi_i \otimes u))|W\xi_i \otimes v\rangle + \langle(1 \otimes x)(\xi_i \otimes u)|W(\xi_i \otimes v)-\xi_i \otimes v\rangle| \\
&\le \|x\|\|(W(\xi_i \otimes u)- (\xi_i \otimes u))\| \|v\|+ \|x\|\|(W(\xi_i \otimes v)- (\xi_i \otimes v))\| \|u\| \to 0.
\end{align*}
This shows that $(\omega_i)_i$ is a contractive left BAI for $L^1(\mathbb{G})$. From this we get that $(\omega_iR)_i$ is a right BAI. Then it is a standard Banach algebra fact that $(\omega_i + \omega_iR - \omega_iR\omega_i)_i$ is a bounded two-sided approximate identity.
$(6) \implies (1)$: Without loss of generality, we can assume that our (left) bounded approximate identity $(\omega_i)_i$ converges weak$^\ast$ to $\tilde\epsilon \in L^\infty(\mathbb{G})^*$. Put $\epsilon = \tilde\epsilon|_{C_0(\mathbb{G})}$. Then for any $x \in C_0(\mathbb{G})$, $\omega \in L^1(\mathbb{G})$, we have
\[
\omega( (\epsilon \otimes \iota)\Delta x) = \epsilon((\iota \otimes \omega)\Delta x) = \lim_i \omega_i((\iota \otimes \omega)\Delta x) = \lim_i \omega_i\star \omega(x) = \omega(x).
\]
This shows that $(\epsilon \otimes \iota) \Delta = \iota$. From this formula, it is easy to see that $\epsilon$ is a character on $C_0(\mathbb{G})$, in particular, $\epsilon$ is a state.
\end{proof}
As an almost immediate consequence of the above theorem, we obtain the following important relationship between co-amenability and amenability with respect to Pontryagin duality.
\begin{cor} \label{co-amen-to-amen}
If $\widehat{\mathbb{G}}$ is co-amenable, then $\mathbb{G}$ is amenable.
\end{cor}
\begin{proof}
If $\widehat{\mathbb{G}}$ is co-amenable, then there exists a net of unit vectors $(\xi_i)_i\subset L^2(\mathbb{G})$ such that
\[
\|\hat W(\xi_i \otimes \eta) - \xi_i \otimes \eta\| \to 0 \qquad (\eta \in L^2(\mathbb{G})).\] But since $W = \sigma \hat W \sigma^*$ is unitary, we get
\[
\|W(\eta \otimes \xi_i) - \eta \otimes \xi_i\| \to 0 \qquad (\eta \in L^2(\mathbb{G})).\]
Now let $m \in L^\infty(\mathbb{G})^*$ be any state which is a weak$^\ast$-limit point of the net $(\omega_{\xi_i})_i$. Then we have, for all $\xi,\eta \in L^2(\mathbb{G})$,
\begin{align*}
\omega_{\xi,\eta}((\iota \otimes m)\Delta(x)) &= \lim_i \langle W^*(1 \otimes x)W(\xi \otimes \xi_i)|(\eta \otimes \xi_i) \rangle \\
&= \lim_i \langle (1 \otimes x)(\xi \otimes \xi_i)|(\eta \otimes \xi_i) \rangle \\
&= \omega_{\xi,\eta}(1) \lim_{i}\omega_{\xi_i} (x) = \omega_{\xi,\eta}(1) m(x).
\end{align*}
Therefore $m$ is a left-invariant mean on $L^\infty(\mathbb{G})$.
\end{proof}
A natural question that arises from the above result is whether the converse of Corollary \ref{co-amen-to-amen} is true for all locally compact quantum groups.
\begin{quest}
Does the amenability of $\mathbb{G}$ imply the co-amenability of $\widehat{\mathbb{G}}$?
\end{quest}
In its full generality, the above question is still open, however there are some special cases that the answer is known to be yes: If $\mathbb{G}$ is a classical locally compact group, a theorem of Leptin \cite{Le68} says that $\mathbb{G}$ is amenable if and only if the Fourier algebra $A(\mathbb{G}) \cong L^1(\widehat{\mathbb{G}})$ has a bounded approximate identity, which by Theorem \ref{co-amen-characterization} implies that $\widehat{\mathbb{G}}$ is co-amenable. If $\mathbb{G}$ is compact, the answer is obviously yes as well, since $L^1(\widehat{\mathbb{G}})$ is unital. If $\mathbb{G}$ is discrete, then the answer is again yes, but this is a highly non-trivial theorem of
Tomatsu \cite{To06} (also proved independently by Blanchard-Vaes \cite{BlVa02}).
\begin{thm}[Tomatsu \cite{To06}, Blanchard-Vaes \cite{BlVa02}] \label{disc-amen}
Let $\mathbb{G}$ be an amenable discrete quantum group. Then $\widehat{\mathbb{G}}$ is co-amenable.
\end{thm}
In what follows we give a very rough outline of the ideas behind this theorem. Our arguments follow closely along the lines of the approach of Blanchard-Vaes \cite{BlVa02}. In the proofs in \cite{BlVa02, To06}, the modular theory of the Haar weights on $L^\infty(\mathbb{G})$ plays a major role. For another, very recent proof of Theorem \ref{disc-amen} that avoids the use of modular theory, we refer the reader to Crann \cite{Cr15}.
\begin{proof}[Rough Sketch of Theorem \ref{disc-amen}]
Let $m$ be a left-invariant mean on $L^\infty(\mathbb{G})$. By a standard Hahn-Banach plus convexity argument, we can write $m$ as a weak$^\ast$ limit of a net of states $(\omega_j)_j \subset L^1(\mathbb{G})$ such that \begin{align} \label{l1-inv}\|\omega \star \omega_j - \omega(1)\omega_j\|_{L^1(\mathbb{G})} \to 0 \qquad (\omega \in L^1(\mathbb{G})).\end{align}
Since $L^\infty(\mathbb{G})$ is in standard position on $L^2(\mathbb{G})$, there exists a unique unit vector $\xi_j$ in the positive cone $L^2(\mathbb{G})^+$ such that $\omega_j = \omega_{\xi_j}$. Our claim is that co-amenability of $\widehat{\mathbb{G}}$ will follow if we can show that
\begin{align}\label{as-inv}
\lim_i \|\lambda(\omega)\xi_i - \omega(1)\xi_i\|_{L^2(\mathbb{G})} = 0 \qquad (\omega \in L^1(\mathbb{G})).
\end{align}
Indeed, if \eqref{as-inv} holds, then it follows that the unital linear functional
\[\omega \mapsto \omega(1) \qquad (\omega \in L^1(\mathbb{G})) \quad \text{satisfies} \quad |\omega(1)| \le \|\lambda(\omega)\|,
\]
and therefore extends uniquely to a state $\hat \epsilon:C_0(\widehat{\mathbb{G}}) \to \mathbb{C}$. But then one readily checks that for $\omega$ belonging to the dense subspace $\{\omega \in L^1(\mathbb{G}): \omega \circ S \in L^1(\mathbb{G})\}$,
\[
\omega((\hat \epsilon \otimes \iota )\hat W) = \hat \epsilon(\omega \otimes \iota)W^* = \hat \epsilon(\lambda(\omega \circ S)) = \omega(S(1)) = \omega(1).
\]
Thus, $(\hat \epsilon \otimes \iota)\hat W = 1$ and $\widehat{\mathbb{G}}$ is therefore co-amenable.
To prove \eqref{as-inv}, by linearity and density, it suffices to fix $\pi \in \text{Irr}(\widehat{\mathbb{G}})$ and prove \eqref{as-inv} for $\omega^\pi \in \mathcal B(H_\pi)_{\ast} = M_{n(\pi)}(\mathbb{C})_{\ast} \subset L^1(\mathbb{G})$. To this end, define positive normal linear functionals $\eta_j, \mu_j \in M_{n(\pi)}(L^\infty(\mathbb{G}))_*$ by
\[
\eta_j ([x_{mn}]) = \text{Tr} \otimes \omega_j([x_{mn}]) \quad \& \quad \mu_j(([x_{mn}]) = \sum_{m,n} \langle \omega_{m,n}^\pi, \omega_j \star x_{mn} \rangle \qquad ([x_{mn}] \in M_{n(\pi)}(L^\infty(\mathbb{G}))),
\]
where $\omega_{ij}^\pi(e_{kl}^\pi) = \delta_{ik}\delta_{jl}$ is the dual basis for the fixed system of matrix units $(e^\pi_{ij})_{1 \le i,j \le n_\pi} \subseteq \mathcal B(H_\pi)$. The positivity of $\eta_j$ can readily be verified based on the equation
\[ \sum_{m,n} \langle \omega_{m,n}^\pi, \omega_j \star x_{mn} \rangle = \sum_{m,n,k} \langle x_{mn}\lambda(\omega_{kn}^\pi)\xi_j|\lambda(\omega_{km}^\pi)\xi_j \rangle.\]
\begin{comment}
Consider the matrices
\[
T(j) = [(\omega_{sr}^\pi \otimes \iota)W\xi_j]_{rs} = [\lambda(\omega_{sr}^\pi)\xi_j]_{rs} \in M_{n_\pi}(L^2(\mathbb{G}))
\]
and define $\mu_j \in M_{n_\pi}(L^\infty(\mathbb{G}))_*$ by
\begin{align*}
\mu_j((x_{mn})) &= \sum_{m,n,k} \langle x_{mn}T(j)_{nk}|T(j)_{mk} \rangle \\
&= \sum_{m,n,k} \langle x_{mn} (\omega_{kn}^\pi \otimes \iota)W\xi_j| (\omega_{km}^\pi \otimes \iota)W\xi_j\rangle \\
&=... \sum_{m,n} \langle (x_{mn} \star \omega_{mn}^\pi)\xi_j |\xi_j \rangle = \sum_{m,n}
\langle \omega_j, x_{mn} \star \omega_{mn}^\pi \rangle \\
&= \sum_{m,n} \langle \omega_{m,n}^\pi, \omega_j \star x_{mn} \rangle.
\end{align*}
Similarly, define $\eta_j \in M_{n_\pi}(L^\infty(\mathbb{G}))_*$ by
\[
\eta_j ((x_{mn})) = \text{Tr} \otimes \omega_j((x_{mn})), \quad \mu_j = .
\]
\end{comment}
Now, by \eqref{l1-inv}, we have $\|\mu_j - \eta_j\|_{M_{n_\pi}(L^\infty(\mathbb{G}))_*} \to 0$, and consequently we also have \[\|(Q_\pi \otimes 1)\mu_j(Q_\pi \otimes 1) - (Q_\pi \otimes 1)\eta_j(Q_\pi \otimes 1)\|_{M_{n(\pi)}(L^\infty(\mathbb{G}))_*} \to 0,\]
where $Q_\pi$ is the modular matrix associated to $\pi$.
At this point we'd like to apply the Powers-St{\o}rmer inequality \cite{Ha75} to the above equation, which requires us to identify the unique elements in the tensor product positive cone $(S_2(H_\pi) \otimes L^2(\mathbb{G}))^+$, which implement $(Q_\pi \otimes 1)\mu_j(Q_\pi \otimes 1)$ and $(Q_\pi \otimes 1)\eta_j(Q_\pi \otimes 1)$, respectively. After some, analysis, the two families of elements in this cone turn out to be
\[(Q_\pi \otimes 1)[\lambda(\omega_{sr}^\pi)\xi_j]_{rs} \quad \& \quad (Q_\pi \otimes \xi_j) \qquad (\text{respectively}). \] From this, an application of Powers-St{\o}rmer gives
\begin{align*}
& \|(Q_\pi \otimes 1)\mu_j(Q_\pi \otimes 1) - (Q_\pi \otimes 1)\eta_j(Q_\pi \otimes 1)\|_{M_{n_\pi}(L^\infty(\mathbb{G}))_*} \to 0\\
\implies&\|(Q_\pi \otimes 1)[\lambda(\omega_{sr}^\pi)\xi_j]_{rs} -(Q_\pi \otimes \xi_j) \|_{(S_2(H_\pi) \otimes L^2(\mathbb{G}))} \to 0 \\
&\iff \|[\lambda(\omega_{sr}^\pi)\xi_j]_{rs} -1 \otimes \xi_j \|_{(S_2(H_\pi) \otimes L^2(\mathbb{G}))} \to 0
\end{align*}
Applying $\omega^\pi \otimes \iota$ to the last line, where $\omega^\pi \in B(H_\pi)^*$, we get
\[
\|\lambda (\omega^\pi)\xi_j - \omega^\pi(1)\xi_j\| \to 0.
\]
\end{proof}
\subsection{The Kesten amenability criterion}
We shall close our discussion of amenability and co-amenability by presenting a very handy criterion for establishing the (non-)amenability of a discrete quantum group, called the {\it Kesten amenability criterion}. We then use this result to determine precisely when the duals of free orthogonal quantum groups $O^+_F$ are amenable
Let $\mathbb{G}$ be a discrete quantum group with compact dual $\widehat{\mathbb{G}}$, and let $U = [u_{ij}] \in M_n(\mathcal O(\widehat{\mathbb{G}}))$ be a finite-dimensional representation of $\widehat{\mathbb{G}}$. Recall that the {\it character of $U$} is the element $\chi = \chi_U \in \mathcal O(\widehat{\mathbb{G}})$ given by
\[
\chi = \sum_{i=1}^n u_{ii}\in \mathcal O(\widehat{\mathbb{G}}).
\]
Note that that $\chi$ always satisfies the following norm inequalities \[\|\chi\|_{C(\widehat{\mathbb{G}})} \le \|\chi\|_{C^u(\widehat{\mathbb{G}})} = \hat \epsilon(\chi) = n, \] where $\hat \epsilon: C^u(\widehat{\mathbb{G}}) \to \mathbb{C}$ is the universal co-unit (which is a character satisfying $\hat \epsilon(u_{ij}) = \delta_{ij}$). Moreover, the same inequalities hold for the real and imaginary parts $\Re \chi$ and $\Im \chi$.
Recall that $\mathbb{G}$ is called {\it finitely generated} if $\widehat{\mathbb{G}}$ is a {\it compact matrix quantum group}. I.e., if there is a finite dimensional unitary representation $U$ whose matrix elements generate $\mathcal O(\widehat{\mathbb{G}})$ as a $\ast$-algebra. We call such a representation $U$ a {\it fundamental representation} of $\widehat{\mathbb{G}}$. The following theorem is the Kesten amenability criterion, which we state in the finitely generated case for simplicity. (We leave the obvious adaptation to the general case to the reader.)
\begin{thm}[Banica \cite{Ba99}]
Let $\mathbb{G}$ be a finitely generated discrete quantum group, let $U = [u_{ij}]_{ij = 1}^n$ be a fundamental (unitary) representation of $\widehat{\mathbb{G}}$, and let $\chi = \chi_U \in \mathcal O(\widehat{\mathbb{G}})$ be the character of $U$. Let $\sigma \subset [-n,n]$ be the spectrum of $\Re\chi \subset C(\widehat{\mathbb{G}})$.
Then $\mathbb{G}$ is amenable if and only if $n \in \sigma$.
\end{thm}
\begin{rem}
When $\mathbb{G} = \Gamma$ is a classical finitely generated discrete group with finite generating set $S$. The above theorem reduces to the classical Kesten amenability criterion \cite{Ke59}, which says that $\Gamma$ is amenable if and only if the Markov operator
\[
T = \frac{1}{2}\sum_{\gamma \in S} \lambda(\gamma) + \lambda(\gamma)^* \in C^*_r(\Gamma) \quad \text{satisfies} \quad |S| \in \sigma(T).
\]
\end{rem}
\begin{proof}
If $\mathbb{G}$ is amenable, then $\widehat{\mathbb{G}}$ is co-amenable and therefore there is a bounded co-unit $\hat \epsilon:C(\widehat{\mathbb{G}}) \to \mathbb{C}$ which sends $\Re\chi$ to $n$.
Thus $n \in \sigma$.
Conversely, suppose that $n \in \sigma$. Put $a_i = 1 - \Re u_{ii} \in C(\widehat{\mathbb{G}})$, which is positive. Then the positive operator $\sum_i a_i$ is not invertible. This means that there is a sequence $(\xi_k)_k \subset L^2(\mathbb{G})$ of unit vectors such that $\sum_i \langle a_i \xi_k| \xi_k\rangle \to 0$.
As each term in the above sum is non-negative, we get
\begin{align*} \langle a_i \xi_k| \xi_k\rangle \to 0 &\qquad (1 \le i \le n) \\
\iff \|u_{ii}\xi_k - \xi_k\| \to 0 &\qquad (1 \le i \le n),
\end{align*} where the second line follows from the equality
\[
\|(1-u_{ii})\xi_k \|^2 = 2 \langle (1-\Re u_{ii})\xi_k|\xi_k\rangle + \|u_{ii}\xi\|^2 -1.
\]
Now let $M_n(C(\widehat{\mathbb{G}}))$ act on $\mathbb{C}^n\otimes L^2(\mathbb{G})$ in the canonical way. As $U$ is unitary,
\[1 = \| U(e_j \otimes \xi_k)\|^2 = \|\sum_i e_i \otimes u_{ij}\xi_k\|^2 = \sum_i \|u_{ij}\xi_k\|^2 \qquad (1 \le j \le n),\]
and as $\|u_{ii}\xi_k -\xi_k \| \to 0$ we get $\|u_{ij}\xi_k\| \to 0$ if $i \ne j$.
In summary, this means that \[ u_{ij}\xi_k-\delta_{ij}\xi_k \to 0 \qquad (1 \le i, j \le n).\] Multiplying the above limit by $u_{il}^*$ and summing over $i$, we also obtain
\[
\sum_{i}( u_{il}^*u_{ij}\xi_k-\delta_{ij}u_{il}^*\xi_k) = \delta_{lj}\xi_k - u_{jl}^*\xi_k \to 0 \qquad (1 \le l,j \le n).
\]
Extending linearly and multiplicatively (using the triangle inequality), we get from the above two limits that
\[
\|x \xi_k - \hat\epsilon(x)\xi_k\| \to 0 \qquad (x \in \mathcal O(\widehat{\mathbb{G}})).
\]
In particular $|\hat \epsilon(x)| = \lim_k\|x\xi_k\|_{L^2(\mathbb{G})} \le \|x\|_{C_0(\widehat{\mathbb{G}})}$. I.e., $C_0(\widehat{\mathbb{G}})$ has a bounded co-unit, and therefore $\mathbb{G}$ is amenable.
\end{proof}
As an application of the Kesten amenability criterion, we determine when the free orthogonal quantum groups $O^+_F$ are co-amenable.
\begin{thm}[Banica \cite{Ba96}]
Let $n \ge 2$ and $F \in \text{GL}_n(\mathbb{C})$ be such that $F\bar F = \pm 1$. Then $O^+_F$ is co-amenable if and only if $n = 2$. In particular, Woronowicz' $SU_q(2)$ is co-amenable for all $q \in [-1,1] \backslash \{ 0\}$.
\end{thm}
\begin{proof}[Sketch] Consider the character $\chi = \chi_U$ of the fundamental representation $U$. Since by definition, we have $\bar U = F^{-1}UF$, it follows from the invariance of characters under equivalence of representations that
\[
\chi^* = \chi_{\bar U} = \chi_{F^{-1}UF} = \chi.
\]
Thus, to determine when $O^+_F$ is co-amenable, we can try to determine the spectrum $\sigma \subset [-n,n]$ of $\chi \in C_0(O^+_F)$. To do this, we will identify the spectral measure $\mu$ of $\chi$ with respect to the Haar state $\varphi$ (whose support is $\sigma$), and this can be done by computing the moments of $\mu$. But in this case, we have
\[
\int_\sigma t^k d\mu = \varphi(\chi^k) = (\varphi \otimes \text{Tr})(U^{\otimes k}) =
\text{Tr}((\varphi \otimes \iota)U^{\otimes k}) = \dim(\text{Mor}(1, U^{\otimes k})),
\]
since $(\varphi \otimes \iota)U^{\otimes k}$ is the orthogonal projection onto the space of fixed vectors $\text{Hom}(1, U^{\otimes k})$ of the tensor product representation $U^{\otimes k}$.
In \cite{Ba96}, Banica showed that for all $n \ge 2$, $\text{Mor}(1, U^{\otimes k}) = \{0\}$ if $k$ is odd, and
\[
\dim(\text{Mor}(1, U^{\otimes 2m})) = C_{m} = \frac{1}{m +1} {2m\choose m} \qquad (\text{the $m$th Catalan number, } m \in \mathbb{N}).
\]
The Catalan numbers are the well-known moment sequence for the semicircular measure $\mu$ on the interval $[-2,2]$ (see for example \cite{NiSp06}):
\[
d\mu(t) = \frac{1}{2\pi}\sqrt{4-t^2}1_{[-2,2]}dt.
\]
Thus $\sigma = [-2,2]$ and therefore $n \in \sigma \iff n = 2$.
\end{proof}
\section{Weakening the notion of amenability} \label{fourieralg}
We now consider various relaxations of the notion of amenability and co-amenability for locally compact quantum groups. In fact, in this survey, we will primarily focus on two such notions, namely {\it the Haagerup property} and {\it weak amenability} for locally compact quantum groups. The primary goal here will be to investigate the interplay between these approximation properties and the structure of certain operator algebras associated to a locally compact quantum group.
\subsection{Motivation - Amenability and approximate units in the Fourier algebra}
Let $G$ be a locally compact group and let $\lambda_G:G \to \mathcal U(L^2(G))$ denote the left regular representation. The {\it Fourier algebra} is the subspace of $A(G) \subset C_0(G)$ consisting of coefficient functions of $\lambda_G$. I.e.,
\[A(G) = \Big\{t \mapsto \phi_{\xi,\eta}(t) = \langle \lambda_G(t^{-1}) \xi|\eta\rangle : \xi, \eta \in L^2(G) \Big\}.\] The Fourier algebra $A(G)$ turns out to be a Banach algebra under the pointwise operations inherited from $C_0(G)$, when it is equipped with the norm
\[
\|\phi\|_{A(G)} = \inf\{\|\xi\|_2\|\eta\|_2: \phi = \phi_{\xi,\eta}\}.
\]
See \cite{Ey64} for details.
Since $A(G)$ is a nice commutative Banach algebra that can be associated to any locally compact group $G$, it is natural to ask what types of conditions on $G$ ensure that $A(G)$ has a bounded approximate identity (BAI)? It turns out that the existence of such a BAI is equivalent to the amenability of $G$.
\begin{thm}[Leptin \cite{Le68}]
A locally compact group $G$ is amenable if and only if the Banach algebra $A(G)$ has a bounded approximate identity.
Moreover, if $G$ is amenable, we can always take the BAI $(b_i)_i\subset A(G)$ to consist of normalized {\it positive definite functions} associated to $\lambda_G$. (I.e., for each $i$, $b_i = \phi_{\xi_i,\xi_i}$ for some unit vector $\xi_i \in L^2(G)$.)
\end{thm}
In other words, if $G$ is not amenable, then the Banach algebra $A(G)$ has no BAI. However, it could still happen that $A(G)$ has an {\it unbounded} approximate identity $(b_i)_i$ that is uniformly bounded in some other (smaller) norm on $A(G)$. For example, we could ask for the uniform boundedness of $(b_i)_i$ in the multiplier norm on $A(G)$:
\[
\|b_i\|_{MA(G)} = \sup_{\|a\|_{A(G)} = 1}\|b_ia\|_{A(G)}.
\]
Another scenario one could consider is a non-amenable group $G$ for which there exists a net $(b_i)_i$ of normalized positive definite functions contained in the slightly larger space $C_0(G)$ (instead of $A(G)$) for which a BAI-type condition still holds for $A(G)$. Namely,
\[
\|b_ia-a\|_{A(G)} \to 0 \qquad (a \in A(G)).
\]
It was discovered by Haagerup \cite{Ha78} that the above two scenarios actually hold for the non-amenable free groups $\mathbb{F}_k$ on $k \ge 2$ generators. With this seminal work of Haagerup, the notion of {\it weak amenability} and {\it the Haagerup property} for groups was born. Our goal now is to study these approximation properties in some detail in the general context of locally compact quantum groups.
\subsection{The Fourier algebra of a locally compact quantum group}
We begin by defining the quantum group analogue of the Fourier algebra. To do this, we consider once again the classical situation: Let $G$ be a locally compact group and let
$W \in \mathcal B(L^2(G) \otimes L^2(G))$ be the associated multiplicative unitary which describes the lcqg structure associated to $G$. It turns out that $W$ is given by
\[
W\xi(s,t) = \xi (s,s^{-1}t) \qquad (\xi \in L^2(G \times G) = L^2(G) \otimes L^2(G)),
\]
and from this it follows that the multiplicative unitary $\hat W =\Sigma W^*\Sigma$ associated to the Pontryagin dual $\hat G$ satisfies
\[
\hat W \xi(s,t) = \xi(ts,t) \qquad (\xi \in L^2(G \times G)).
\]
In particular, the dual left regular representation $\hat \lambda:L^1(\hat G) \to C_0(G)$ is given by $\hat \lambda(\omega) = (\omega \otimes \iota)\hat W$. The following proposition shows that $A(G)$ is nothing but the image of $L^1(\hat G)$ under $\hat \lambda$.
\begin{prop}
For a locally compact group $G$, we have that
\[
\hat \lambda: L^1(\hat G) \to A(G)
\]
is an isometric isomorphism of Banach algebras.
\end{prop}
\begin{proof}
Fix $\omega \in L^1(\hat G)$. Since $L^\infty(\hat G)$ is in standard position on $L^2(G)$, we can find a pair of vectors $\xi,\eta \in L^2(G)$ such that $\omega = \omega_{\xi,\eta}$ on $L^\infty(\hat G)$ and $\|\omega\|_{L^1(\hat G)} = \|\xi\|\|\eta\|$. Now, for $\alpha,\beta \in L^2(G)$, we compute
\begin{align*}
\langle \hat \lambda (\omega)\alpha|\beta \rangle &= \langle \hat W (\xi \otimes \alpha)|(\eta \otimes \beta)\rangle \\
&= \int_G \int_G \xi(ts)\alpha(t) \overline{\eta(s)\beta(t)} dsdt \\
&=\int_G \langle \lambda(t^{-1})\xi |\eta\rangle \alpha(t) \overline{\beta(t)}dt.
\end{align*}
Hence, $\hat \lambda(\omega_{\xi,\eta}) = \phi_{\xi,\eta} \in A(G)$ and $\|\phi_{\xi,\eta}\|_{A(G)} \le \|\xi\|\|\eta\|$. Taking the infimum over all $\xi, \eta$ above (noting that the homomorphism $\hat \lambda$ is injective), we get $\|\hat \lambda(\omega)\|_{A(G)} = \|\omega\|_{L^1(\hat G)}$.
\end{proof}
We now turn to the quantum setting and define the Fourier algebra and Fourier multipliers.
\begin{defn}
For a locally compact quantum group $\mathbb{G}$, the {\it Fourier algebra} is the subalgebra \[A(\mathbb{G}) := \hat\lambda(L^1(\widehat{\mathbb{G}})) \subseteq C_0(\mathbb{G}).\] For $\omega \in L^1(\widehat{\mathbb{G}})$, we write $\|\hat\lambda(\hat\omega)\|_{A(\mathbb{G})} = \|\hat \omega\|_{L^1(\widehat{\mathbb{G}})}$, thus making $A(\mathbb{G})$ a Banach algebra with respect to the norm $\|\cdot\|_{A(\mathbb{G})}$.
\end{defn}
\begin{defn}An element $a \in L^\infty(\mathbb{G})$ is called a {\it (left) Fourier multiplier} if $a A(\mathbb{G}) \subseteq A(\mathbb{G})$.
\end{defn}
Note that by the closed graph theorem, any Fourier multiplier $a \in L^\infty(\mathbb{G})$ induces a bounded linear map \[m_a:A(\mathbb{G}) \to A(\mathbb{G}); \qquad m_a(b) = ab \qquad (b\in A(\mathbb{G})). \] Equivalently, $a$ induces a bounded linear map $L_{*}:L^1(\widehat{\mathbb{G}}) \to L^1(\widehat{\mathbb{G}})$ such that
\begin{align} \label{eqn:left-mult}
L_*(\omega_1\star \omega_2) = L_*(\omega_1)\star\omega_2 \qquad (\omega_i \in L^1(\widehat{\mathbb{G}})),
\end{align}
where $L_*(\omega) = \hat\lambda^{-1}(m_a(\hat \lambda (\omega))).$
In what follows, we will frequently identify a Fourier multiplier $a$ with the corresponding bounded map $m_a$. We write $MA(\mathbb{G})$ for the Banach algebra of all Fourier multipliers, and let $M_{cb}A(\mathbb{G}) := MA(\mathbb{G}) \cap \mathcal {CB}(A(\mathbb{G}))$ be the Banach space of {\it completely bounded (cb) Fourier multipliers}. Note that $M_{cb}A(\mathbb{G})$ consists of precisely those $a \in MA(\mathbb{G})$ for which the corresponding $\sigma$-weakly continuous map \[L := (L_*)^*:L^\infty(\widehat{\mathbb{G}}) \to L^\infty(\widehat{\mathbb{G}})\] is completely bounded on $L^\infty(\widehat{\mathbb{G}})$. Finally, we call $a \in M_{cb}A(\mathbb{G})$ a {\it completely positive multiplier} (or a continuous {\it completely positive definite function} on $\widehat{\mathbb{G}}$) if the corresponding map $L$ is completely positive.
The following proposition gives some useful equivalent characterizations of those $L \in \mathcal {CB}_\sigma(L^\infty(\mathbb{G}))$ that come from cb multipliers.
\begin{prop} \label{M_cb-characterization}
Let $L \in \mathcal {CB}_\sigma(L^\infty(\widehat{\mathbb{G}}))$ with pre-adjoint $L_* \in \mathcal {CB}(L^1(\widehat{\mathbb{G}}))$. Then the following conditions are equivalent:
\begin{enumerate}
\item $L_*$ satisfies \eqref{eqn:left-mult}.
\item $\hat \Delta \circ L = (L \otimes \iota) \circ \hat \Delta$.
\item There exists a unique $a \in L^\infty(\mathbb{G})$ such that $(L \otimes \iota) (\hat W) = (1 \otimes a)(\hat W)$.
\item There exists a unique $a \in M_{cb} A(\mathbb{G})$ such that $a \hat \lambda(\hat\omega) = \hat\lambda(L_*\hat \omega)$ for all $\hat \omega \in L^1(\widehat{\mathbb{G}})$.
\end{enumerate}
Moreover, we have $M_{cb}A(\mathbb{G}) \subseteq M(C_0(\mathbb{G}))$,
\end{prop}
\begin{proof}[Sketch]
The equivalence $(1) \iff (2)$ is clear, and so is $(3) \iff (4) \implies (1)$ (recall that $\hat \lambda(\hat \omega) = (\hat \omega \otimes \iota)\hat W$). The hard part is $(1) \implies (3)$ and the containment $M_{cb}A(\mathbb{G}) \subseteq M(C_0(\mathbb{G}))$. See \cite{JuNeRu09} and \cite{Da11}, respectively, for details.
\end{proof}
\begin{rem}
Note that Condition (3) in Theorem \ref{M_cb-characterization} implies that for any $a \in M_{cb}A(\mathbb{G})$, the corresponding map $L = L^{(a)} \in \mathcal {CB}_\sigma(L^\infty(\widehat{\mathbb{G}}))$ leaves $C_0(\widehat{\mathbb{G}})$ invariant. Indeed, if $\omega \in \mathcal K(L^2(\mathbb{G}))^*$, then $(\iota \otimes \omega)\hat W \in C_0(\widehat{\mathbb{G}})$ and
\[
L((\iota \otimes \omega)\hat W) = (\iota \otimes \omega \cdot a)\hat W \in C_0(\widehat{\mathbb{G}}).
\] The result now follows from the density of $\{(\iota \otimes \omega)\hat W: \omega \in \mathcal K(L^2(\mathbb{G}))^*\}$ in $C_0(\widehat{\mathbb{G}})$.
\end{rem}
\begin{rem}
When $G$ is a locally compact group, a classical unpublished result of Gilbert (see also \cite{Jo92}), gives a useful intrinsic characterization of $M_{cb}A(G)$. Namely, a function $a \in M(C_0(G)) = C_b(G)$ belongs to $M_{cb}A(G)$ if and only if there exists a Hilbert space $H$ and $\alpha,\beta \in C_b(G,H)$ such that
\begin{align}\label{Gilbert}a(s) = \langle \alpha (st)|\beta(t) \rangle \qquad (s,t \in G).
\end{align} Moreover,
\[\|a\|_{M_{cb}A(G)} = \inf \|\alpha\|_{\infty}\|\beta\|_{\infty},\]
where the infimum runs over all $\alpha, \beta, H$ for which \eqref{Gilbert} holds.
In particular, this implies that $B(G)$, the Fourier-Stieltjes algebra of all coefficients of unitary (or even uniformly bounded) representations of $G$ on Hilbert space belong to $M_{cb}A(G)$, with $\|a\|_{M_{cb}A(G)} \le \|a\|_{B(G)}$.
It turns out that there is a quantum analogue of Gilbert's result which was obtained by Daws \cite{Da11}. We shall not need this result here and simply refer the interested reader to the above reference. What will be essential for us is the quantum analogue of the containment $B(G) \subseteq M_{cb}A(G)$, which we now state.
\end{rem}
\begin{prop} \label{B(G)-inclusion}
Let $H$ be a Hilbert space and $U \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$ a unitary representation of a lcqg $\mathbb{G}$. Let $\omega \in \mathcal K(H)^*$ and let \[a = (\iota \otimes \omega)(U^*) \in M(C_0(\mathbb{G})).\] Then $a \in M_{cb}A(\mathbb{G})$ and $\|a\|_{M_{cb}A(\mathbb{G})} \le \|\omega\|$. Moreover, if $\omega \in \mathcal K(H)^*$ is positive, then $a$ is a completely positive definite function.
\end{prop}
\begin{proof}
We start off by defining our candidate for the map $L \in \mathcal {CB}_{\sigma}(L^\infty(\widehat{\mathbb{G}}))$ induced by $a$ (as in Proposition \ref{M_cb-characterization}). Define
\[
L \in \mathcal {CB}_{\sigma}(L^\infty(\widehat{\mathbb{G}}), \mathcal B(L^2(\mathbb{G}))); \qquad L(x) = (\iota \otimes \omega) (U(x \otimes 1)U^*).
\]
It is clear from the structure of $L$ that $\|L\|_{cb} \le \|\omega\|$ and that $L$ is completely positive if $\omega$ is positive. Next, we show that $L\in \mathcal {CB}_\sigma(L^\infty(\widehat{\mathbb{G}}))$. To do this, observe that we can write
\begin{align*}
(L \otimes \iota) \hat W&= (\iota \otimes \iota \otimes \omega)(U_{13}\hat W_{12}U^*_{13}).
\end{align*}
But since $\hat W = \Sigma W^*\Sigma$, we also have
\begin{align*}
&U_{13}\hat W_{12}U^*_{13} = U_{13}\Sigma_{12} W^*_{12}\Sigma_{12} U^*_{13} = \Sigma_{12} U_{23}W^*_{12}U^*_{23}\Sigma_{12}\\
&= \Sigma_{12} U_{23}W^*_{12}U^*_{23}W_{12}W_{12}^*\Sigma_{12} \\
&= \Sigma_{12}U_{23}((\Delta \otimes \iota)U^*)W_{12}^*\Sigma_{12} \\
&=\Sigma_{12}U_{23}U_{23}^*U_{13}^*W_{12}^* \Sigma_{12}\\
&=\Sigma_{12}U_{13}^*W_{12}^*\Sigma_{12} = U_{23}^*\Sigma_{12}W_{12}^*\Sigma_{12} = U_{23}^*\hat W_{12},
\end{align*}
and therefore
\[
(L \otimes \iota) \hat W = (\iota \otimes \iota \otimes\omega)(U_{23}^*\hat W_{12}.) = (1 \otimes a)\hat W.
\]
From this last equation, it follows that $L$ maps into $L^\infty(\widehat{\mathbb{G}})$ (Indeed, just recall that the set $\{(\iota \otimes \mu)\hat{W}: \mu \in \mathcal B(L^2(\mathbb{G}))_* \}$ is $\sigma$-weakly dense in $L^\infty(\widehat{\mathbb{G}})$). Moreover, we have that $a \in M_{cb}A(\mathbb{G})$ implements $L$ by Proposition \ref{M_cb-characterization}.
\end{proof}
\begin{rem} \label{Daws-Bochner}
In Proposition \ref{B(G)-inclusion} it is shown that whenever we have a unitary representation $U \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$ and a state $\omega \in \mathcal K(H)^*$, the
adjoint map $L^{(a)} \in \mathcal {CB}_\sigma(L^\infty(\widehat{\mathbb{G}}))$ corresponding to the multiplier $a = (\iota \otimes \omega)U^* \in M_{cb}A(\mathbb{G})$ is unital and completely positive. In fact, the converse statement is also true: if $L^{(a)} \in \mathcal {CB}_\sigma(L^\infty(\widehat{\mathbb{G}}))$ is unital and completely positive, then $a \in M_{cb}A(\mathbb{G})$ is of the form $a = (\iota \otimes \omega)U^*$ for some unitary representation $U \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$ and some state $\omega \in \mathcal K(H)^*$. See \cite{Da12} for details. The proof of this result for classical locally compact groups is an easy exercise for the reader (just observe that complete positivity of $a \in M_{cb}A(G)$ implies that $a$ is a continuous positive definite function on $G$).
\end{rem}
With the above remark in mind we will call any $a \in M_{cb}A(\mathbb{G})$ of the above form
a {\it (normalized) completely positive definite function} on $\mathbb{G}$. The general theory of (completely) positive definite functions on lcqgs was studied in detail by Daws and Salmi \cite{DaSa13}. We refer the reader to this paper for some of the more subtle aspects of this theory in the quantum context.
\section{The Haagerup property and weak amenability} \label{hap/wa}
We now begin the discussion of our two approximation properties for locally compact quantum groups--the Haagerup property and weak amenability--which can be thought of as weak notions of amenability. (Or more appropriately, weak notions of co-amenability for their duals, as we shall see below).
\subsection{The Haagerup property}
The Haagerup property for general locally compact quantum groups was developed by Daws, Fima, Skalski and White \cite{DaFiSkWh16}. In what follows, we give a very brief and rather incomplete account.
\begin{defn}
We say that a lcqg $\mathbb{G}$ has the \textbf{Haagerup property} if there exists a bounded approximate identity $(a_i)_i$ for $C_0(\mathbb{G})$ consisting of normalized completely positive definite functions.
\end{defn}
In the classical setting, the Haagerup property for a locally compact group is often characterized in representation theoretic terms involving mixing unitary representations and almost invariant vectors. We shall see now that this characterization carries through in the quantum setting.
\begin{defn}
A unitary representation $U \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$ of a lcqg $\mathbb{G}$ is called a {\it mixing representation} if all of its matrix elements belong to $C_0(\mathbb{G})$. That is,
\[
(\iota \otimes \omega_{\xi,\eta})U \in C_0(\mathbb{G}) \qquad (\xi, \eta \in H).
\]
\end{defn}
\begin{ex}
The prototypical example of a mixing representation is the left-regular representation $W \in M(C_0(\mathbb{G}) \otimes C_0(\widehat{\mathbb{G}}))$. Indeed, by construction we have $(\iota \otimes \omega)W \in C_0(\mathbb{G})$ for all $\omega \in \mathcal K(L^2(\mathbb{G}))^*$.
\end{ex}
\begin{rem}
Let $\mu$ be a state on the universal C$^\ast$-algebra $C_0^u(\widehat{\mathbb{G}})$, associated to the dual quantum group $\widehat{\mathbb{G}}$. Then, following \cite{DaFiSkWh16}, we can find a Hilbert space $H$, a unitary representation $U \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$ and a cyclic unit vector $\xi \in H$ such that
\[
\langle \pi^u(\omega), \mu \rangle = \langle \omega, a \rangle \qquad (\omega \in L^1(\mathbb{G}))
\]
where $\pi^u:L^1(\mathbb{G}) \to C^u_0(\widehat{\mathbb{G}})$ is the universal representation and $a = (\iota \otimes \omega_{\xi})U \in M(C_0(\mathbb{G}))$ is the corresponding matrix element of $U$.
Clearly, if the representation $U$ is mixing, then $a \in C_0(\mathbb{G})$. But in fact the following lemma shows that the converse is true as well.
\end{rem}
\begin{lem}\label{mixing}
Let $\mu \in C_0^u(\widehat{\mathbb{G}})^*$ be a state and let $a = (\iota \otimes \omega_{\xi})U \in M(C_0(\mathbb{G}))$ be the corresponding matrix element as above. If $a \in C_0(\mathbb{G})$, then $U$ is a mixing representation.
\end{lem}
\begin{proof}
Let $\pi:L^1_\sharp(\mathbb{G}) \to \mathcal B(H)$ be the involutive $\ast$-representation corresponding to $U$. Then $\xi$ is cyclic for $\pi$, and therefore it suffices to show that
\[
(\iota \otimes \omega_{\pi(\omega_1)\xi,\pi(\omega_2)\xi})U \in C_0(\mathbb{G}) \qquad (\omega \in L^1_\sharp(\mathbb{G})).
\]
But for $\omega \in L^1_\sharp(\mathbb{G})$,
\begin{align*}
\langle\omega, (\iota \otimes \omega_{\pi(\omega_1)\xi,\pi(\omega_2)\xi})U \rangle &= \langle \pi(\omega)\pi(\omega_1)\xi| \pi(\omega_2)\eta \rangle \\
&= \langle\pi(\omega_2^\sharp \star \omega \star \omega_1)\xi|\eta\rangle \\
&= \langle \omega_2^\sharp \star \omega \star \omega_1, a\rangle = \langle \omega, \omega_1 \star a \star \omega_2^\sharp \rangle,
\end{align*}
so
\[
(\iota \otimes \omega_{\pi(\omega_1)\xi,\pi(\omega_2)\xi})U = \omega_1\star a \star \omega_2^\sharp.
\]
Thus our problem reduces to showing that the natural left/right module actions of $L^1_\sharp(\mathbb{G})$ on $L^\infty(\mathbb{G})$ leave $C_0(\mathbb{G})$ invariant. To do this, we can assume by the Cohen factorization theorem that $\omega_1 = c\cdot \omega_1' $ and $\omega_2^\sharp = d\cdot\omega_2'$ for some $c, d \in C_0(\mathbb{G})$ and $\omega_1',\omega_2' \in L^1(\mathbb{G})$. But then,
\begin{align*}
\omega_1 \star a &= (\iota \otimes c \cdot \omega_1') \Delta(a) = (\iota \otimes \omega_1')(\Delta(a) (1 \otimes c)),\\
a\star \omega_2^\sharp &= (d \cdot \omega_2' \otimes \iota)\Delta(a) = (d \cdot\omega_2' \otimes \iota)(\Delta(a)(d \otimes 1)).
\end{align*}
Finally, observing that $\Delta(C_0(\mathbb{G}))(1 \otimes C_0(\mathbb{G})) \subseteq C_0(\mathbb{G}) \otimes C_0(\mathbb{G})$ and $\Delta(C_0(\mathbb{G}))(C_0(\mathbb{G}) \otimes 1) \subseteq C_0(\mathbb{G}) \otimes C_0(\mathbb{G})$ (see \cite[Section 3]{KuVa00}), the result follows.
\end{proof}
Next, we define the notion of almost invariant vectors for a unitary representation and study its connection to the Haagerup property.
\begin{defn}
A unitary representation $U \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$ of a lcqg $\mathbb{G}$ is said to have {\it almost invariant vectors} if there is a net of unit vectors $(\xi_i)_i \subset H$ such that
\[
\|U(\eta \otimes \xi_i) - (\eta \otimes \xi_i)\| \to 0 \qquad (\eta \in L^2(\mathbb{G})).
\]
\end{defn}
\begin{ex}
The left regular representation $W$ of lcqg $\mathbb{G}$ has almost invariant vectors if and only if $\widehat{\mathbb{G}}$ is co-amenable. See Theorem \ref{co-amen-characterization}.
\end{ex}
\begin{ex}
Of course, any representation $U \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$ of a lcqg $\mathbb{G}$ with a {\it fixed vector} (i.e., $ \exists 0 \ne \xi \in H$ such that $U(\eta \otimes \xi) = \eta \otimes \xi$ for all $\eta \in H$) trivially has almost invariant vectors.
\end{ex}
\begin{thm}
The following conditions are equivalent for a lcqg $\mathbb{G}$:
\begin{enumerate}
\item $\mathbb{G}$ has the Haagerup property.
\item There is a mixing representation $U \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$ with almost invariant vectors.
\end{enumerate}
\end{thm}
\begin{rem}
Before starting the proof, we make one observation. Recall that the antipode $S:L^\infty(\mathbb{G}) \to L^\infty(\mathbb{G})$ is a densely defined invertible operator which satisfies the formal identity
\[
(S \otimes \iota)U = U^*
\]
for any unitary representation $U \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$ of $\mathbb{G}$. In particular, this identity implies that any coefficient of a unitary representation $b = (\iota \otimes \omega_{\xi,\eta})U$ belongs to the domain $\mathcal D(S)$ of $S$, and $Sb = (\iota \otimes \omega_{\xi,\eta})U^* \in M_{cb}A(\mathbb{G})$.
\end{rem}
\begin{proof}
$(1) \implies (2)$. Let $(a_i)_i \subseteq M_{cb}A(\mathbb{G}) \cap C_0(\mathbb{G})$ be a bounded approximate identity for $C_0(\mathbb{G})$ consisting of normalized completely positive definite functions. For each $i$, Remark \ref{Daws-Bochner} supplies us with a Hilbert space $H_i$, a state $\omega_i \in \mathcal K(H_i)^*$ and a unitary representation $U_i \in M(C_0(\mathbb{G}) \otimes \mathcal K(H_i))$
such that
\[
a_i = (\iota \otimes \omega_i)U_i^* \in C_0(\mathbb{G}).
\]
Moreover we may assume that $\omega_i = \omega_{\xi_i}$ for some unit vector $\xi_i \in H_i$ and that $\xi_i$ is cyclic for $U_i$. Indeed, we can amplify $(U_i,H_i)$ if necessary to obtain $\xi_i$, and then we can restrict $U_i$ to the cyclic representation generated by $\xi_i$.
Set $b_i = (\iota \otimes \omega_{\xi_i})U_i = S^{-1}a_i.$ Since $S^{-1}:\mathcal D(S^{-1}) \cap C_0(\mathbb{G}) \to C_0(\mathbb{G})$ (see \cite[Section 5]{KuVa00}), we have that $b_i \in C_0(\mathbb{G})$ for each $i$, and therefore each $U_i$ is a mixing representation by Lemma \ref{mixing}.
Put $H = \oplus_i H_i$ and $U = \bigoplus_i{U_i} \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$, where $H = \bigoplus H_i$. Then $U$ is again a mixing representation, since direct sums of mixing representations are obviously still mixing. Now, since $(a_i)_i$ is a bounded approximate identity for $C_0(\mathbb{G})$, it follows that $a_i \to 1$ $\sigma$-weakly in $L^\infty(\mathbb{G})$, and the same is true for $(b_i)_i$. Indeed, since $(b_i)_i$ is a bounded net, it suffices to show that
\[
\langle \omega, b_i\rangle \to \langle \omega, 1 \rangle \qquad (\omega \in L^1_\sharp(\mathbb{G})).
\] But for this dense subspace we have
\[
\langle \omega, b_i\rangle = \overline{\langle (\omega^\sharp)^* \circ S, S^{-1}a_i \rangle} = \overline{\langle (\omega^\sharp)^* , a_i \rangle} \to \overline{\langle (\omega^\sharp)^* ,1 \rangle} = \langle \omega , 1 \rangle.
\] In particular, this $\sigma$-weak convergence implies that for any $\eta \in L^2(\mathbb{G})$
\begin{align}\label{al-in}
\|U(\eta \otimes \xi_i) - (\eta \otimes \xi_i)\|^2 = 2\|\eta\|^2 - 2\Re\langle \omega_\eta, b_i\rangle \to 2\|\eta\|^2 - 2\Re\langle \omega_\eta, 1 \rangle = 0.
\end{align}
Therefore the vectors $(\xi_i)_i \subseteq H$ are almost invariant for the representation $U$.
$(2) \implies (1)$. We only outline this direction and refer the reader to the proof of \cite[Theorem 5.5]{DaFiSkWh16} for the full details. Let $U \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$ be a mixing representation, let $(\xi_i)_i \subset H$ be a family of almost invariant vectors for $U$, and let $b_i = (\iota \otimes \omega_{\xi_i})U \in C_0(\mathbb{G})$. It then follows from equation \eqref{al-in} that $b_i\to 1 \ \sigma\text{-weakly in }L^\infty(\mathbb{G})$. Working a little harder, one can actually show that $b_ix \to x$ and $xb_i \to x$ weakly in $C_0(\mathbb{G})$ ($x \in C_0(\mathbb{G})$), and thus the the convex hull of $(b_i)_i$ contains a bounded approximate
identity (BAI) for $C_0(\mathbb{G})$. In particular, this yields a BAI $(c_j)_j$ for $C_0(\mathbb{G})$ consisting of elements of the form $c_j = (\iota \otimes \omega_j)U$, where $\omega_j \in \mathcal K(H)^*$ is a state. Setting $a_j = S(c_j) = (\iota \otimes \omega_j)U^*$, we obtain a net of normalized completely positive definite functions on $\mathbb{G}$ which is also a BAI for $C_0(\mathbb{G})$.
\end{proof}
\begin{cor}
A locally compact quantum group has the Haagerup property if $\widehat{\mathbb{G}}$ is co-amenable.
\end{cor}
\begin{proof}
If $\widehat{\mathbb{G}}$ is co-amenable, then the left-regular representation has almost invariant vectors.
\end{proof}
\subsection{Weak amenability}
We now define another approximation property for locally compact quantum groups, called amenability. For locally compact groups, this concept is due to Cowling and Haagerup \cite{CoHa89}.
\begin{defn} \label{weakamen}
A locally compact quantum group $\mathbb{G}$ is called {\it weakly amenable} if there exists a net $(a_i)_i \subseteq A(\mathbb{G})$ such that \[\|a_ib-b\|_{A(\mathbb{G})} \to 0 \quad (b \in A(\mathbb{G})) \quad \text{and}\quad \Lambda((a_i)_i):= \limsup_i\|a_i\|_{M_{cb}A(\mathbb{G})} < \infty.\]
\end{defn}
In short, $\mathbb{G}$ is weakly amenable precisely when $A(\mathbb{G})$ has a left approximate identity that is uniformly bounded in the $\|\cdot\|_{M_{cb}A(\mathbb{G})}$-norm. We define the {\it Cowling-Haagerup constant for $\mathbb{G}$} to be the number \[\Lambda_{cb}(\mathbb{G}):= \inf \{\Lambda((a_i)_i): (a_i)_i \subset A(\mathbb{G}) \text{ satisfies Definition \ref{weakamen}}\}.\]
The next proposition shows that weak amenability is really a weak form of co-amenability for $\widehat{\mathbb{G}}$.
\begin{prop}
If $\widehat{\mathbb{G}}$ is co-amenable, then $\mathbb{G}$ is weakly amenable and $\Lambda_{cb}(\mathbb{G}) = 1$.
\end{prop}
\begin{proof}
When $\widehat{\mathbb{G}}$ is co-amenable, $A(\mathbb{G})$ has a left bounded approximate identity consisting of normalized completely positive definite functions.
\end{proof}
When $\mathbb{G}$ is a classical group or a discrete quantum group, we know that $\widehat{\mathbb{G}}$ is co-amenable if and only if $\mathbb{G}$ is amenable. This equivalence is unknown for general locally compact quantum groups. A positive solution to this problem would provide an affirmative solution to the following open problem.
\begin{quest}
Let $\mathbb{G}$ be an amenable locally compact quantum group. Does $\mathbb{G}$ have the Haagerup property and is $\Lambda_{cb}(\mathbb{G}) < \infty$?
\end{quest}
\subsection{Some examples}
\begin{ex}[See \cite{Ha78, CoHa89}]
If $G = \mathbb{F}_n, SL_2(\mathbb{Z}), SL_2(\mathbb{R}), SU(n,1), SO(n,1)$, then $G$ has the Haagerup property and $\Lambda_{cb}(G) = 1$.
\end{ex}
\begin{ex}[See \cite{Ha86}]
If $G$ is any connected simple Lie group of real rank greater than $1$ with finite center, then $G$ doe not have the Haagerup property and $\Lambda_{cb}(G) = \infty$.
\end{ex}
\begin{ex}[See \cite{Oz08}]
Let $G$ be a finitely generated word hyperbolic group, then $\Lambda_{cb}(G) < \infty$.
\end{ex}
In the quantum setting, we have the following examples.
\begin{ex}[See \cite{DaFiSkWh16,Fr12}]
Let $(\mathbb{G}_i)_{i \in I}$ be a family of discrete quantum groups. If each $\mathbb{G}_i$ has the Haagerup property, then their free product $\mathbb{G} = \ast_{i \in I} \mathbb{G}$ has the Haagerup property. Similarly, if \[\Lambda_{cb}(\mathbb{G}_i) = 1 \quad \text{for each $i$, then} \quad
\Lambda_{cb}(\mathbb{G}) = 1.
\]
\end{ex}
\begin{ex}[See \cite{Br12, Br13, Fr13, dCFY, Le14}]
Let $\mathbb{G}$ be the discrete quantum group such that $(\widehat{\mathbb{G}}$ is one of the following compact quantum groups: $O^+_F, U^+_F, S_N^+, H_N^{+(s)}$, where $F \in GL_N(\mathbb{C})$. Then $\mathbb{G}$ has the Haagerup property and $\Lambda_{cb}(\mathbb{G}) = 1$. We will discuss some of these examples in more detail in Section \ref{app}.
\end{ex}
In the non-discrete quantum setting, there turns out to be a scarcity of examples. We highlight the following remarkable result of Caspers \cite{Ca14}.
\begin{ex}[See \cite{Ca14}]
The (non-discrete) locally compact quantum group $\mathbb{G} = SU_q(1,1)$ is weakly amenable with $\Lambda_{cb}(\mathbb{G}) = 1$, co-amenable, and has the Haagerup property.
\end{ex}
\begin{comment}
\begin{thm}
If $\mathbb{G}$ is discrete and $\widehat{\mathbb{G}}$ is of Kac type, then $\mathbb{G}$ has the Haagerup property iff $(L^\infty(\widehat{\mathbb{G}}), \hat \varphi)$ has the Haagerup approximation property.
\end{thm}
\section{Kazhdan's Property (T)}
\begin{defn}
Let $U \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$ be a unitary representation. \begin{enumerate}
\item We say that $U$ has {\it almost invariant vectors} if there is a net of unit vectors $(\xi_j)_{j} \subset H$ such that
\[\|U(\eta \otimes \xi_j) - \eta \otimes \xi_j\| \to 0 \qquad (\eta\in L^2(\mathbb{G})).\]
\item We say that $U$ has an {\it invariant vector} if there exists $0 \ne \xi \in H$ such that
\[U(\eta \otimes \xi) = \eta\otimes \xi \qquad (\eta \in L^2(\mathbb{G})).\]
\item We say that $\mathbb{G}$ has property (T) if every unitary representation $U$ with almost invariant vectors has an invariant vector.
\end{enumerate}
\end{defn}
\begin{lem} TFAE for a unitary representation $U \in M(C_0(\mathbb{G}) \otimes \mathcal K(H))$:
\begin{enumerate}
\item $U$ has almost invariant vectors $(\xi_j)_{j}$.
\item The net $(\iota \otimes \omega_{\xi_j})U \to 1$ $\sigma$-weakly in $L^\infty(\mathbb{G})$.
\item The net of states $\omega_{\xi_j}\circ \phi_U \in C_0^{u}(\widehat{\mathbb{G}})$ converges to $\hat \epsilon_u$ in the weak-$\ast$ topology.
\item $\|\phi_U(a) \xi_j- \hat \epsilon_u(a)\xi_j\| \to 0$.
\end{enumerate}
If instead $\xi$ is a fixed unit vector, then the above equivalent limits can be replaced by equalities in the obvious way.
\end{lem}
\begin{proof}
This is standard.
\end{proof}
\begin{exs}
Non-classical examples have been surprisingly hard to come by. We will discuss some of these later on.
\end{exs}
\end{comment}
\section{Discrete quantum groups and operator algebra approximation properties} \label{discr}
In this section we restrict our attention to discrete quantum groups and explore some connections between amenability, weak amenability and the Haagerup property for $\mathbb{G}$, and various approximation properties for C$^\ast$-algebras and von Neumann algebras associated to $\mathbb{G}$.
\subsection{Some operator algebra approximation properties}
Let us start by defining the operator algebra approximation properties that will be relevant to us.
\begin{defn}
Let $M$ be a von Neumann algebra. We say that $M$ has the {\it weak$^\ast$ completely bounded approximation property (w$^\ast$CBAP)} if there is a net of normal, finite rank cb maps $T_i:M \to M$ such that $T_i \to \iota_M$ pointwise $\sigma$-weakly and $\limsup_i\|T_i\|_{cb} :=C < \infty$. We call the infimum of all these $C$'s $\Lambda_{cb}(M)$, the {\it Cowling-Haagerup constant of $M$}.
\end{defn}
\begin{defn}
Let $A$ be a C$^\ast$-algebra. We say that $A$ has the {\it completely bounded approximation property (CBAP)} if there is a net of finite rank cb maps $T_i:A \to A$ such that $T_i \to \iota_A$ pointwise in norm and $\limsup_i\|T_i\|_{cb} :=C < \infty$. We call the infimum of all these $C$'s $\Lambda_{cb}(A)$, the {\it Cowling-Haagerup constant of $A$}.
\end{defn}
\begin{defn}
Let $M$ be a von Neumann algebra equipped with a faithful normal semi-finite weight $\varphi$. We say that $(M,\varphi)$ has the {\it Haagerup approximation property} if there exists a net $(\Phi_i)_i$ unital completely positive maps $\Phi_i:M \to M$ such that $\varphi(\Phi_i(x)) \le \varphi(x)$ for all $x \in M_+$, and such that the induced maps
\[T_i \in \mathcal B(L^2(M,\varphi); \qquad T_i\Lambda_\varphi(x) = \Lambda_\varphi(\Phi_i(x)) \qquad (x \in \mathfrak N_\varphi)\] are compact and converge strongly to $\iota_{L^2(M,\varphi)}$.
\end{defn}
\begin{rem}
The above definition of the Haagerup property for a von Neumann algebra is taken from \cite{CaSk15}, where this notion is studied in detail. In particular, it is shown there that this property is independent of the choice of weight $\varphi$ on $M$, and therefore one can simply say that {\it $M$ has the Haagerup property}, without reference to a choice of weight. It should also be mentioned that another approach to the Haagerup property for von Neumann algebras was developed at the same time in \cite{OkTo15}, but this time using the standard form of a von Neumann algebra. It turns out that both of these approaches to the Haagerup property are equivalent \cite{CaOkSkTo14}.
\end{rem}
\begin{rem}
The above three definitions should be regarded as weaker forms of nuclearity (in the C$^\ast$-context) and injectivity (in the von Neumann context). Indeed, all three of the above properties are implied by nuclearity/injectivity -- compare with Defintions \ref{wcpap}-\ref{cpap}.
\end{rem}
We now come to the main theorem of this section, which connects quantum group approximation properties to operator algebra approximation properties in the discrete case.
\begin{thm} \label{discrete-AP}
Let $\mathbb{G}$ be a discrete quantum group with compact dual $\widehat{\mathbb{G}}$. Then we have the following implications.
\begin{enumerate}
\item $\mathbb{G}$ is amenable $\implies$ $L^\infty(\widehat{\mathbb{G}})$ is injective and $C(\widehat{\mathbb{G}})$ is nuclear.
\item $\mathbb{G}$ has the Haagerup property $\implies$ $L^\infty(\widehat{\mathbb{G}})$ has the Haagerup approximation property.
\item $\mathbb{G}$ is weakly amenable $\implies$ $L^\infty(\widehat{\mathbb{G}})$ has the w$^*$CBAP and $C(\widehat{\mathbb{G}})$ has the CBAP. In this case we also have
\[
\Lambda_{cb}(L^\infty(\widehat{\mathbb{G}})), \Lambda_{cb}(C(\widehat{\mathbb{G}})) \le \Lambda_{cb}(\mathbb{G}).
\]
\end{enumerate}
If $\mathbb{G}$ is, in addition, a {\it unimodular} discrete quantum group, then the reverse implications in $(1)-(3)$ hold, and moreover
\[
\Lambda_{cb}(L^\infty(\widehat{\mathbb{G}})) = \Lambda_{cb}(C(\widehat{\mathbb{G}})) = \Lambda_{cb}(\mathbb{G}).
\]
\end{thm}
\subsection{Some preparations}
Before beginning the proof of Theorem \ref{discrete-AP}, let us first recall that for a discrete quantum group, we have
\[
L^\infty(\mathbb{G}) = \prod_{\pi \in \text{Irr}(\widehat{\mathbb{G}})} \mathcal B(H_\pi), \qquad C_0(\mathbb{G}) =\bigoplus_{\pi \in \text{Irr}(\widehat{\mathbb{G}})}^{c_0} \mathcal B(H_\pi),\]
\[
L^2(\mathbb{G}) = L^2(\widehat{\mathbb{G}}) = \bigoplus_{\pi \in \text{Irr}(\widehat{\mathbb{G}})}^{\ell^2} L^2(\mathcal B(H_\pi), d(\pi)\text{Tr}(Q_\pi\cdot)),
\]
and the left regular representation of $\widehat{\mathbb{G}}$ is given by
\[\hat W= \prod_{\pi \in \text{Irr}(\widehat{\mathbb{G}})} \sum_{i,j} u_{ij}^\pi\otimes e_{ij}^\pi,\]
where $(e_{ij}^{\pi})_{1 \le i,j \le n(\pi)}$ denotes a fixed system of matrix units for $\mathcal B(H_\pi)$ and $U^\pi = [u_{ij}^{\pi}] \in M_{n(\pi)}(C(\mathbb{G})) = C(\mathbb{G}) \otimes \mathcal B(H_\pi)$ is a corresponding representative of $\pi$. Also recall that
\[
C_c(\mathbb{G}) = \bigoplus_{\pi \in \text{Irr}(\widehat{\mathbb{G}})}\mathcal B(H_\pi) \subset L^\infty(\mathbb{G}).
\]
denotes the $\sigma$-weakly dense subspace of finitely supported elements. A standard but crucial observation regarding $C_c(\mathbb{G})$ is that we have the norm-dense inclusions
\[
\Lambda_\varphi(C_c(\mathbb{G})) \subseteq L^2(\mathbb{G}), \quad \& \quad C_c(\mathbb{G}) \subseteq A(\mathbb{G})
\]
for any discrete quantum group $\mathbb{G}$. In particular, if $a = \hat \lambda(\omega_{\xi,\eta}) \in A(\mathbb{G})$ with $\xi,\eta \in L^2(\mathbb{G})$ and $\|a\|_{A(\mathbb{G})} = \|\xi\|_2 \|\eta\|_2$, then for any $\epsilon >0$ we can find $x,y \in C_c(\mathbb{G})$ such that $\xi' = \Lambda_\varphi(x)$ and $\eta' = \Lambda_\varphi(y)$ satisfy
\[
\|\xi - \xi'\|_2 < \frac{\epsilon}{2\|\eta\|_2} \quad \&\quad \|\eta - \eta'\|_2 < \frac{\epsilon}{2\|\xi\|_2}.
\]
Setting $b= \hat \lambda(\omega_{\xi',\eta'})$, we then have $b \in A(\mathbb{G})\cap C_c(\mathbb{G})$ and
\[
\|a-b\|_{A(\mathbb{G})} \le \|\xi-\xi'\|_2\|\eta\|_2 + \|\xi\|_2\|\eta-\eta'\|_2 < \epsilon.
\]
Note that the above type of estimate easily implies that any normalized completely positive definite function in $A(\mathbb{G})$ can be $A(\mathbb{G})$-norm approximated by normalized completely positive definite function in $C_c(\mathbb{G}) \cap A(\mathbb{G})$.
Now suppose $L = L^{(a)} \in \mathcal {CB}_\sigma(L^\infty(\widehat{\mathbb{G}}))$ is the adjoint of a cb multiplier $a = (a^\pi)_{\pi \in \text{Irr}(\widehat{\mathbb{G}})} \in M_{cb}A(\mathbb{G})\subset L^\infty(\mathbb{G})$, then the defining formula $(L \otimes \iota)\hat W = (1 \otimes a)\hat W$ tells us that
\[
(L \otimes \iota)\sum_{i,j} u^\pi_{ij} \otimes e^\pi_{ij} = \sum_{i,j} u_{ij}^\pi \otimes a^\pi e^\pi_{ij} = \sum_{i,j,k}u^\pi_{ij} \otimes a_{ki}^\pi e^{\pi}_{kj} = \sum_{kj} \sum_{i}a^{\pi}_{ki}u^\pi_{ij} \otimes e^\pi_{kj} \qquad (\pi \in \text{Irr}(\widehat{\mathbb{G}})).
\]
In other words, $L$ acts by the formula
\begin{align} \label{mult-formula}
L(u_{ij}^\pi) = \sum_{k=1}^{n(\pi)}a^\pi_{ik}u^\pi_{kj} \qquad (1 \le i,j \le n(\pi), \ \pi \in \text{Irr}(\widehat{\mathbb{G}})).
\end{align}
\subsection{Proof of Theorem \ref{discrete-AP}}
\subsubsection{For general discrete $\mathbb{G}$}
We give here only a general outline of the proof of the theorem. For further details, we refer the reader to \cite{Br12,Fr12,DaFiSkWh16,KrRu99}.
{\it Implication} (3): Suppose we have a net $(a_i)_i \subset A(\mathbb{G})$ such that
\[
\|a_ib - b\|_{A(\mathbb{G})} \to 0 \qquad (b \in A(\mathbb{G})) \quad \& \quad \limsup_i\|a_i\|_{M_{cb}A(\mathbb{G})}=C < \infty.
\]
Since $\|a_i\|_{A(\mathbb{G})} \le \|a_i\|_{M_{cb}A(\mathbb{G})}$ for each $i$, we can, without loss of generality, assume that $(a_i)_i \subset C_c(\mathbb{G})\cap A(\mathbb{G})$.
Consider now the adjoint maps $(L_i)_i \subset \mathcal{CB}_\sigma(L^\infty(\widehat{\mathbb{G}}))$, where $L_i = L^{(a_i)}$. Then $\limsup_i \|L_i\|_{cb} = C < \infty$, and the condition $\|a_ib - b\|_{A(\mathbb{G})} \to 0 \quad (b \in A(\mathbb{G})) $ implies by duality that $L_i \to \iota_{L^\infty(\widehat{\mathbb{G}})}$ pointwise $\sigma$-weakly. Finally formula \eqref{mult-formula} combined with the fact that $a_i \in C_c(\mathbb{G})$ implies that each $L_i$ is finite rank. Thus $\Lambda_{cb}(L^\infty(\widehat{\mathbb{G}})) < \Lambda_{cb}(\mathbb{G})$. For the C$^\ast$-variant, we just need to make the additional observation that $L_i \to \iota_{L^\infty(\widehat{\mathbb{G}})}$ pointwise $\sigma$-weakly implies in particular that $a_i \to 1$ pointwise in $L^\infty(\mathbb{G})$, which then implies that $\|L_ix-x\|_{C(\widehat{\mathbb{G}})} \to 0$ for all $x \in \mathcal O(\widehat{\mathbb{G}})$. Since $\mathcal O(\widehat{\mathbb{G}})$ is norm dense in $C(\widehat{\mathbb{G}})$, we see that $L_i|_{C(\widehat{\mathbb{G}})} \to \iota_{C(\widehat{\mathbb{G}})}$ pointwise in norm. In particular, $\Lambda_{cb}(C(\widehat{\mathbb{G}})) \le \Lambda_{cb}(\mathbb{G})$.
{\it Implication} (1): Thus is just a variant of the above argument, except now we have the additional fact that our net $(a_i)_i \subseteq A(\mathbb{G})\cap C_c(\mathbb{G})$ consists of normalized completely positive definite functions. This additional condition forces $L_i$ to be completely positive and unital, and we are done.
{\it Implication} (2): This is just another minor modification of ideas in the previous two implications. Suppose $(a_i)_i \subseteq C_0(\mathbb{G})$ is a bounded approximate identity consisting of normalized completely positive definite functions. Then $L_i$ is a unital completely positive map and it is $\hat \varphi$-preserving, since
\[
\hat \varphi(L_iu_{kl}^\pi) = \sum_r a_{i,kr}^\pi \hat \varphi(u^\pi_{rl}) = \delta_{\pi,\pi_0} = \hat \varphi(u_{kl}^\pi) \qquad (1 \le k,l \le n(\pi), \ \pi \in \text{Irr}(\widehat{\mathbb{G}})),
\]
where $\pi_0$ is the (equivalence class of the) trivial representation of $\widehat{\mathbb{G}}$. Finally, we observe that the $L^2$-extension $\Phi_i$ of $L_i$ is precisely $a_i \in C_0(\mathbb{G}) \subset \mathcal K(L^2(\mathbb{G})) \subset \mathcal B(L^2(\mathbb{G}))$. In particular, $\Phi_i$ is compact, and $\Phi_i \to \iota_{L^2(\mathbb{G})}$ strongly, since $(a_i)_i$ is a bounded approximate identity for $C_0(\mathbb{G})$.
\begin{rem}
Before going to the unimodular case, we would like to point out that recently a remarkable extension of the above implication (2) has been obtained by Okayasu, Ozawa and Tomatsu \cite{OkOzTo15}. Namely, for any locally compact quantum group $\mathbb{G}$ the Haagerup property implies that $L^\infty(\widehat{\mathbb{G}})$ has the Haagerup approximation property. This result is even new for classical locally compact groups.
\end{rem}
\subsubsection{The unimodular case} Now suppose that $\mathbb{G}$ is unimodular. This is equivalent to saying that the dual Haar state $\hat \varphi:L^\infty(\widehat{\mathbb{G}}) \to \mathbb{C}$ is tracial. To get the converse of the implications $(1)-(3)$ above, we need a way of manufacturing a completely bounded multiplier on $\mathbb{G}$ from an arbitrary completely bounded map $T \in \mathcal {CB}(L^\infty(\widehat{\mathbb{G}}))$. When $\mathbb{G}$ is unimodular, this is possible thanks to an ``averaging trick''
introduced by Haagerup \cite{Ha86} (for discrete groups) and later extended to unimodular discrete quantum groups by Kraus and Ruan \cite{KrRu99}.
The idea of this averaging trick is as follows: Given a completely bounded map $T:C(\widehat{\mathbb{G}}) \to L^\infty(\widehat{\mathbb{G}})$\footnote{We choose these domains/ranges to be as general as possible}, define $a = a_T \in L^\infty(\mathbb{G})$ by
\[
a = (\hat \varphi \otimes \iota)([(T \otimes \iota)\hat W]\hat W^*).
\]
In other words, if we write $a = (a^\pi)_{\pi \in \text{Irr}(\widehat{\mathbb{G}})}$ with $a^\pi \in \mathcal B(H_\pi)$, then
\begin{align}\label{averaging-formula}a^\pi_{ij} = \sum_{k=1}^{n(\pi)}\hat\varphi((Tu_{ik}^\pi)(u_{jk}^\pi)^*)= \sum_{k=1}^{n(\pi)}\langle\Lambda_{\hat\varphi}(Tu_{ik}^\pi)|\Lambda_{\hat\varphi}u_{jk}^\pi)\rangle \qquad (1 \le i,j \le n(\pi), \ \pi \in \text{Irr}(\widehat{\mathbb{G}})).
\end{align}
Our claim is that in fact $a \in M_{cb}A(\mathbb{G})$ with $\|a\|_{M_{cb}A(\mathbb{G})} \le \|T\|_{cb}$. To verify this, recall that since $\hat\varphi$ is a faithful tracial state, there exists a unique normal faithful $(\hat \varphi \otimes \hat \varphi)$-preserving conditional expectation
\[
E:L^\infty(\widehat{\mathbb{G}}) \overline{\otimes} L^\infty(\widehat{\mathbb{G}}) \to \hat\Delta(L^\infty(\widehat{\mathbb{G}})),
\]
which at the $L^2$-level just corresponds to the orthogonal projection $P:L^2(\widehat{\mathbb{G}}) \otimes L^2(\widehat{\mathbb{G}}) \to L^2(\hat\Delta(L^\infty(\widehat{\mathbb{G}})))$. In particular, it follows from the Schur orthogonality relations for matrix elements of irreducible unitary representations of $\widehat{\mathbb{G}}$ that the following formula for $E$ holds:
\begin{align}\label{E-formula}
E(u_{ij}^\pi \otimes u^\sigma_{kl}) = \frac{\delta_{\pi,\sigma}\delta_{j,k}}{n(\pi)}\Big(\sum_{1 \le r \le n(\pi)} u_{ir}^\pi \otimes u_{rl}^\pi\Big) = \frac{\delta_{\pi,\sigma}\delta_{j,k}}{n(\pi)}\hat \Delta(u^\pi_{il}) \qquad (\pi, \sigma \in\text{Irr}(\widehat{\mathbb{G}})).
\end{align}
Now define a linear map
\[
L = L_T:C(\widehat{\mathbb{G}}) \to L^\infty(\widehat{\mathbb{G}}); \qquad L = \hat \Delta^{-1}\circ E \circ (T \otimes \iota)\circ \hat \Delta.
\]
Evidently $L$ is completely bounded with $\|L\|_{cb} \le \|T\|_{cb}$, and $L$ is completely positive/unital/trace-preserving whenever $T$ has these properties. Let us now evaluate $Lu_{ij}^\pi$, for some $\pi \in \text{Irr}(\widehat{\mathbb{G}})$ and $1 \le i,j \le n(\pi)$. First note that
\begin{align*}
(T\otimes \iota) \hat \Delta(u_{ij}^\pi) &=\sum_{k=1}^{n(\pi)}Tu_{ik}^\pi \otimes u_{kj}^\pi \\
&=\sum_{k=1}^{n(\pi)}\sum_{\sigma \in \text{Irr}(\widehat{\mathbb{G}})} \sum_{1 \le r,s \le n(\sigma)} n(\sigma) \langle \Lambda_{\hat \varphi}(Tu_{ik}^\pi ) | \Lambda_{\hat \varphi}(u^\sigma_{rs}) \rangle u_{rs}^\sigma \otimes u_{kj}^\pi
\end{align*}
Where the last equality follows from the fact that $(\sqrt{n(\sigma)}\Lambda_{\hat \varphi}(u_{rs}^\sigma))_{\sigma, r,s}$ is an orthonormal basis for $L^2(\widehat{\mathbb{G}})$ (since $\hat \varphi$ is tracial), and thus
\[
Tu_{ik}^\pi = \sum_{\sigma \in \text{Irr}(\widehat{\mathbb{G}})} \sum_{1 \le r,s \le n(\sigma)} n(\sigma) \langle \Lambda_{\hat \varphi}(Tu_{ik}^\pi ) | \Lambda_{\hat \varphi}(u^\sigma_{rs}) \rangle u_{rs}^\sigma \qquad (\text{with respect to $L^2$-convergence}).
\]
From these equations we then get
\begin{align*}
Lu_{ij}^\pi& = \hat \Delta^{-1} \circ E\Big(\sum_{k=1}^{n(\pi)}\sum_{\sigma \in \text{Irr}(\widehat{\mathbb{G}})} \sum_{1 \le r,s \le n(\sigma)} n(\sigma) \langle \Lambda_{\hat \varphi}(Tu_{ik}^\pi ) | \Lambda_{\hat \varphi}(u^\sigma_{rs}) \rangle u_{rs}^\sigma \otimes u_{kj}^\pi\Big) \\
&= \hat \Delta^{-1} \circ \sum_{k=1}^{n(\pi)}\sum_{\sigma \in \text{Irr}(\widehat{\mathbb{G}})} \sum_{1 \le r,s \le n(\sigma)} n(\sigma) \langle \Lambda_{\hat \varphi}(Tu_{ik}^\pi ) | \Lambda_{\hat \varphi}(u^\sigma_{rs}) \rangle \frac{\delta_{\pi,\sigma}\delta_{s,k}}{n(\pi)}\hat \Delta(u^\pi_{rj}) \\
&= \sum_{1 \le r,k \le n(\pi)} \langle \Lambda_{\hat \varphi}(Tu_{ik}^\pi ) | \Lambda_{\hat \varphi}(u^\sigma_{rk}) \rangle u^\pi_{rj} \\
&= \sum_{1 \le r \le n(\pi)}a^\pi_{ir}u^\pi_{rj}
\end{align*}
Comparing this calculation with formula \eqref{mult-formula}, we conclude that $a \in M_{cb}A(\mathbb{G})$ and $L = L^{(a)}$ is the adjoint of the multiplier $a$, and in particular $\|a\|_{M_{cb}A(\mathbb{G})} \le \|T\|_{cb}$.
\begin{rem}\label{finite-rank}
In the case that our map $T \in \mathcal {CB}(C(\widehat{\mathbb{G}}), L^\infty(\widehat{\mathbb{G}}))$ is finite rank, we can say even more about the resulting multiplier $a$. Namely, that $a \in A(\mathbb{G})$. Indeed, if $T$ is finite rank, then there exists $x_1, \ldots, x_m \in L^\infty(\widehat{\mathbb{G}})$ and $\tau_1, \ldots , \tau_m \in C(\widehat{\mathbb{G}})^*$ such that
\[T(x) = \sum_{1 \le r \le m} \tau_r(x) x_r \qquad (x \in C(\widehat{\mathbb{G}})),\]
which yields
\[a^\pi_{ij} = \sum_{1 \le r\le m} \sum_{k = 1}^{n(\pi)}\tau_r(u_{ik}) \langle\Lambda_{\hat \varphi} (x_r)|\Lambda_{\hat\varphi}(u_{jk}) \rangle.\]
This suggests we define $b_r, c_r \in L^\infty(\mathbb{G})$ by
\[b_r^\pi = [\tau_r(u_{ij}^\pi)]_{ij}, \quad \& \quad c_r^\pi = [\langle\Lambda_{\hat\varphi} (x_r)|\Lambda_{\hat\varphi}(u_{ji}^\pi)\rangle]_{ij}, \]
so that $a = \sum_rb_rc_r$. Moreover, an easy calculation using the Schur orthogonality relations shows that
\begin{align*}
\|\Lambda_{\varphi}(c_r)\|_{L^2(\mathbb{G})}^2 &=\sum_{\pi} n(\pi)\text{Tr}((c_r^\pi)^*c_r^\pi)= \sum_\pi n(\pi) \sum_{1 \le i,j \le n(\pi)} |\langle\Lambda_{\hat\varphi} (x_r)|\Lambda_{\hat\varphi}(u_{ji}^\pi)\rangle|^2 \\
&=\|\Lambda_{\hat\varphi} x_r\|_{L^2((\widehat{\mathbb{G}})}^2 < \infty \qquad (1 \le r \le m).
\end{align*}
In particular,
\[\|\Lambda_\varphi(a)\|_{L^2(\mathbb{G})} \le \sum_{1 \le r\le m}\|b_r\|_\infty \|\Lambda_\varphi(c_r)\|_{L^2(\mathbb{G})}< \infty \implies a \in \mathfrak N_\varphi.\]
But in the unimodular case, $\mathfrak N_\varphi = \mathfrak N_\varphi^* \subseteq A(\mathbb{G})$,
more explicitly, if $p_0 \in L^\infty(\mathbb{G})$ denotes the rank one projection corresponding to the trivial representation of $\widehat{\mathbb{G}}$, and $S(=R)$ denotes the unitary antipode for $L^\infty(\mathbb{G})$, then also $Sa \in \mathfrak N_\varphi = \mathfrak N_\varphi^*$ and
\[
a = S^2(a) = S((\iota \otimes \omega_{\Lambda_\varphi(p_0), \Lambda_\varphi((Sa)^*)})W) = (\iota \otimes \omega_{\Lambda_\varphi(p_0), \Lambda_\varphi((Sa)^*)})W^* = \hat \lambda(\omega_{\Lambda_\varphi(p_0), \Lambda_\varphi((Sa)^*)}) \in A(\mathbb{G}).
\] For details, see for example \cite[Proposition 3.9]{BrRu14}.
\end{rem}
Let us now use our averaging trick to sketch the the following implications:
\begin{enumerate}
\item[(a)] $L^\infty(\widehat{\mathbb{G}})$ injective $\implies$ $\mathbb{G}$ is amenable.
\item[(b)] $\Lambda_{cb}(L^\infty(\widehat{\mathbb{G}})) \ge \Lambda_{cb}(\mathbb{G})$.
\item[(c)] $L^\infty(\widehat{\mathbb{G}})$ has the Haagerup property $\implies$ $\mathbb{G}$ has the Haagerup property.
\end{enumerate}
(Note: The C$^\ast$-variants of (a),(b) require only minor modifications which are left to the reader).
(a). Since $L^\infty(\widehat{\mathbb{G}})$ is injective, there exists a net of normal finite rank UCP maps $(T_i) \subset \mathcal {CB}_\sigma(L^\infty(\widehat{\mathbb{G}}))$ such that
$T_i \to \iota_{L^\infty(\widehat{\mathbb{G}})}$ pointwise $\sigma$-weakly. Let $(a_i)_i \subset A(\mathbb{G})$
be the corresponding net of normalized completely positive definite functions obtained by the averaging trick. Since $T_i \to \iota_{L^\infty(\widehat{\mathbb{G}})}$ pointwise $\sigma$-weakly, a comparison with formula \eqref{averaging-formula} shows that $a_i^\pi \to 1 \in \mathcal B(H_\pi)$ in norm for all $\pi \in \text{Irr}(\widehat{\mathbb{G}})$. In particular,
\[
\|a_ib - b\|_{A(\mathbb{G})} \to 0 \qquad (b \in C_c(\mathbb{G})).
\] But since $C_c(\mathbb{G})$ is dense in $A(\mathbb{G})$, this implies that $(a_i)_i$ is a bounded approximate identity for $A(\mathbb{G})$.
(b). The argument is almost identical to (a): If there exists a net of normal finite rank completely bounded maps $(T_i) \subset \mathcal {CB}_\sigma(L^\infty(\widehat{\mathbb{G}}))$ such that
$T_i \to \iota_{L^\infty(\widehat{\mathbb{G}})}$ pointwise $\sigma$-weakly, the corresponding net $(a_i)_i \subseteq A(\mathbb{G})$ will be a left approximate identity for $A(\mathbb{G})$ with
\[
\limsup_i\|a_i\|_{M_{cb}A(\mathbb{G})} \le \limsup_{i}\|T_i\|_{cb}.
\]
Thus $\Lambda_{cb}(\mathbb{G}) \le \Lambda_{cb}(L^\infty(\widehat{\mathbb{G}}))$.
(c). This is yet another variant of (a): Since $L^\infty(\widehat{\mathbb{G}})$ has the Haagerup approximation property, then we can find a net of UCP $\hat \varphi$-preserving maps $(T_i) \subset \mathcal {CB}_\sigma(L^\infty(\widehat{\mathbb{G}}))$ such that the corresponding $L^2$-extensions $\Phi_i$ are compact, and $\Phi_i \to \iota_{L^2(\widehat{\mathbb{G}})}$ pointwise in norm. Again, the point-norm convergence of $\Phi$, when compared with formula \eqref{averaging-formula}, tells us that $a^\pi_i \to 1$
for all $\pi \in \text{Irr}(\widehat{\mathbb{G}})$, and therefore
\[
\|a_ib - b\|_{C_0(\mathbb{G})} \to 0 \qquad (b \in C_0(\mathbb{G})).
\]
Moroever, the $L^2$-compactness of $\Phi_i$ turns out to imply each $a_i \in C_0(\mathbb{G})$. See \cite[Theorem 6.7]{DaFiSkWh16} for details.
\begin{rem} From Theorem \ref{discrete-AP}, we see that for unimodular discrete quantum
groups, there is a tight connection between approximation properties of $\mathbb{G}$ in terms of completely bounded Fourier multipliers and various approximation properties for $C(\widehat{\mathbb{G}})$ and $L^\infty(\widehat{\mathbb{G}})$. In the non-unimodular case, this connection only goes one way and it is an important problem in quantum group theory to understand to what extent the unimodular results persist. More precisely, we conclude this section with the following open questions.
\end{rem}
\begin{quest}
Let $\mathbb{G}$ be a non-unimodular discrete quantum group. Do we then have:
\begin{enumerate}
\item $C(\widehat{\mathbb{G}})$ is nuclear $\implies$ $\mathbb{G}$ is amenable? $L^\infty(\widehat{\mathbb{G}})$ is injective $\implies$ $\mathbb{G}$ is amenable?
\item $C(\widehat{\mathbb{G}})$ has the CBAP $\implies$ $\mathbb{G}$ is weakly amenable? $L^\infty(\widehat{\mathbb{G}})$ has the w$^\ast$CBAP $\implies$ $\mathbb{G}$ is weakly amenable?
\item $L^\infty(\widehat{\mathbb{G}})$ has the Haagerup approximation property $\implies \mathbb{G}$ has the Haagerup property?
\end{enumerate}
\end{quest}
\section{Central approximation properties for discrete quantum groups} \label{CAP}
Let $\mathbb{G}$ be a (discrete) quantum group. A multiplier $a \in MA(\mathbb{G})$ is called \textit{central} if $a \in ZL^\infty(\mathbb{G})$, the center of $L^\infty(\mathbb{G})$. I.e., if $L^\infty(\mathbb{G}) = \prod_{\pi \in \text{Irr}(\widehat{\mathbb{G}})} \mathcal B(H_\pi)$, and $p_\pi \in \mathcal B(H_\pi)$ is the central projection corresponding to the $\pi$th component of $L^\infty(\mathbb{G})$. Then $a \in ZMA(\mathbb{G}) = MA(\mathbb{G}) \cap ZL^\infty(\mathbb{G})$ if and only if
\[
a = (c_\pi p_\pi)_{\pi \in \text{Irr}(\widehat{\mathbb{G}})} \qquad (\text{with }c_\pi \in \mathbb{C}).
\]
In a similar fashion, we define $ZM_{cb}A(\mathbb{G}):=M_{cb}A(\mathbb{G})\cap ZL^\infty(\mathbb{G})$.
When studying approximation properties for $\mathbb{G}$ such as amenability, weak amenability, or the Haagerup property, we can impose an additional requirement on the Fourier multipliers that implement these properties. Namely that they be central. This idea naturally leads to the notion of {\it central approximation properties} for discrete quantum groups. More precisely, we have the following definition.
\begin{defn}
A discrete quantum group $G$ is {\it centrally amenable/centrally weakly amenable/has the central Haagerup property} if:
\begin{enumerate}
\item $\mathbb{G}$ is amenable/weakly amenable/has the Haagerup property, and
\item there exists a net $(a_i)_i \subset ZM_{cb}A(\mathbb{G})$ that implements the corresponding approximation property from (1).
\end{enumerate}
\end{defn}
\begin{rem}
For a discrete quantum group $\mathbb{G}$
, we evidently have that the central version of an approximation property for $\mathbb{G}$ implies the corresponding non-central one. In particular, in the case of weak amenability, we can define a {\it central Cowling-Haagerup constant}
\[
Z\Lambda_{cb}(\mathbb{G}): = \inf\{\limsup_i \|a_i\|_{M_{cb}A(\mathbb{G})}: (a_i)_i \subset A(\mathbb{G}) \cap ZL^\infty(\mathbb{G}) \text{ is a left approx. unit for }A(\mathbb{G})\},
\]
and we have
\[Z\Lambda_{cb}(\mathbb{G}) \ge \Lambda_{cb}(\mathbb{G}).\]
\end{rem}
In Section \ref{CAP-monoidal} we will see that there are amenable discrete quantum groups $\mathbb{G}$ that fail to be centrally amenable.
However, in the unimodular case, it turns out that $\mathbb{G}$ has a given approximation property if and only if it has the corresponding central version.
\begin{thm}[\cite{Fr13}] \label{unimodular}
Let $\mathbb{G}$ be a unimodular discrete quantum group. Then:
\begin{enumerate}
\item $\mathbb{G}$ is centrally amenable $\iff$ $\mathbb{G}$ is amenable.
\item $\mathbb{G}$ is centrally weakly amenable $\iff$ $\mathbb{G}$ is weakly amenable (and $Z\Lambda_{cb}(\mathbb{G}) = \Lambda_{cb}(\mathbb{G})$).
\item $\mathbb{G}$ has the central Haagerup property $\iff$ $\mathbb{G}$ has the Haagerup property.
\end{enumerate}
\end{thm}
\begin{proof}
The key idea is to start with an arbitrary element $a \in M_{cb}A(\mathbb{G})$ and construct from $a$ a new multiplier $\tilde{a} \in ZM_{cb}A(\mathbb{G})$ that retains the ``good'' properties of the original multiplier (i.e., positive definiteness, cb norm bounds, belonging to $C_0(\mathbb{G})$ or $A(\mathbb{G})$, ...). With this in mind, fix $a \in M_{cb}A(\mathbb{G})$ and consider the element $\tilde a = (\tilde{a}^\pi)_{\pi \in \text{Irr}(\widehat{\mathbb{G}})}\in ZL^\infty(\mathbb{G})$ given by
\[
\tilde{a}^\pi = \frac{\text{Tr}_\pi(a^{\overline{\pi}})}{n(\pi)}p_\pi \qquad (\pi \in \text{Irr}(\widehat{\mathbb{G}})).
\]
Let $\hat R$ denote the unitary antipode of $L^\infty(\widehat{\mathbb{G}})$. If we fix our representatives $(U^\pi)_{\pi \in \text{Irr}(\widehat{\mathbb{G}})}$ of $\text{Irr}(\widehat{\mathbb{G}})$ so that $U^{\bar \pi} = \overline{U^\pi}$. Then $\hat R$ is a $\sigma$-weakly continuous $\ast$-antiautomorphism of $L^\infty(\widehat{\mathbb{G}})$ satisfying
\[\hat R(u_{ij}^\pi)= (u_{ji}^\pi)^* = u_{ji}^{\bar \pi} \qquad (\pi \in \text{Irr}(\widehat{\mathbb{G}}), \ 1 \le i,j \le n(\pi)),\]
Define a $\sigma$-weakly continuous map
\begin{align} \label{centralization}\tilde{L} \in \mathcal {CB}_\sigma(L^\infty(\widehat{\mathbb{G}})); \qquad \tilde{L} = \hat \Delta^{-1} \circ E \circ ((\hat R \circ L^{(a)} \circ \hat R) \otimes \iota) \hat \Delta,
\end{align}
where $E: L^\infty(\widehat{\mathbb{G}}) \overline{\otimes}L^\infty(\widehat{\mathbb{G}}) \to \hat \Delta( L^\infty(\widehat{\mathbb{G}}))$ is the conditional expectation from \eqref{E-formula} and $L^{(a)}$ is the adjoint of the left multiplier $a$.
Now, since $\hat R$ is a $\ast$-antiautomorphism, \[\|\hat R \circ L^{(a)} \circ \hat R \|_{cb} = \|m_a\|_{cb},\] and therefore $\|\tilde{L}\|\le \|L^{(a)}\|_{cb}$.
We now claim that $\tilde{L} = L^{\tilde{a}}$, so $\tilde a \in ZM_{cb}A(\mathbb{G})$ and $\|\tilde{a}\|_{M_{cb}A(\mathbb{G})} \le \|a\|_{M_{cb}A(\mathbb{G})}$. To check this, we fix a matrix element $u_{ij}^\pi \in \mathcal O(\widehat{\mathbb{G}})$ and calculate
\begin{align*}
\tilde{L}u_{ij}^\pi &= \hat \Delta^{-1} \circ E \circ (\hat R \circ m_a \circ \hat R \otimes \iota) \hat \Delta(u_{ij}^\pi)\\
&=\hat \Delta^{-1} \circ E\Big(\sum_{k,l}a^{\bar \pi}_{kl}R(u_{li}^{\bar\pi})\otimes u_{kj}^\pi\Big) \\
&= (\hat \Delta)^{-1} \circ E\Big(\sum_{k,l}a^{\bar \pi}_{kl}u_{il}^{\pi}\otimes u_{kj}^\pi\Big) \\
&=\sum_{k,l}\delta_{k,l}a^{\bar \pi}_{kl}\frac{u_{ij}^{\pi}}{n(\pi)}=\frac{ \text{Tr}(a^{\bar{\pi}})}{n(\pi)}u_{ij}^\pi\\
&=L^{(\tilde a)}u_{ij}^\pi.
\end{align*}
Now that we have a natural method for constructing elements of $ZM_{cb}A(\mathbb{G})$ from elements of $M_{cb}A(\mathbb{G})$, the rest of the proof is relatively straightforward: One simply has to check that if a net $(a_i)_i \subseteq M_{cb}A(\mathbb{G})$ implements an approximation property for $\mathbb{G}$ such as (weak) amenability/the Haagerup property, then the corresponding net
$(\tilde{a}_i)_i$ will implement the central version of this approximation property. We leave the details to the reader
\end{proof}
\subsection{Going beyond the unimodular case} It is natural to wonder whether it is possible to generalize the centralization procedure for cb multipliers used in the proof of Theorem \ref{unimodular} to general discrete quantum groups. Looking at the above proof, one can see that a crucial tool available in the unimodular case was the existence of a $\hat \varphi \otimes \hat \varphi$-preserving conditional expectation $E: L^\infty(\widehat{\mathbb{G}}) \overline{\otimes} L^\infty(\widehat{\mathbb{G}}) \to \hat \Delta(L^\infty(\widehat{\mathbb{G}}))$, from which we could build a completely positive right inverse $\hat \Delta^\sharp := \hat \Delta^{-1} \circ E$ for the coproduct $\hat \Delta$.
When $\mathbb{G}$ is not unimodular, it turns out that there can never exist such a $\hat \varphi \otimes \hat \varphi$-preserving conditional expectation $E$, since by \cite{Ta72}, this can happen if and only if $\hat \Delta$ intertwines the modular groups $\hat \sigma_t$ and $\hat \sigma_t \otimes \hat \sigma_t$ of $L^\infty(\widehat{\mathbb{G}})$ and $L^\infty(\widehat{\mathbb{G}}) \overline {\otimes} L^\infty(\widehat{\mathbb{G}})$, respectively. But for every discrete $\mathbb{G}$ , it turns out that
\[
\hat \Delta \circ \hat \sigma_t = (\hat \tau_{t} \otimes \hat \sigma_{t})\hat \Delta,
\]
where $\hat \tau_t$ is the scaling group for $L^\infty(\widehat{\mathbb{G}})$, and from this formula it follows that $\hat \Delta$ intertwines $\hat \sigma_t$ and $\hat \sigma_t \otimes \hat \sigma_t$ if and only if $\hat \sigma_t = \hat \tau_t = \iota$ (i.e., $\mathbb{G}$ is unimodular!).
However, not all is lost in the non-unimodular case. Following \cite[Equation (2.2)]{An06}, we can define a normal Haar state-preserving unital completely positive map $\hat \Delta^\sharp: L^\infty(\widehat{\mathbb{G}}) \overline{\otimes} L^\infty(\widehat{\mathbb{G}}) \to \hat L^\infty(\widehat{\mathbb{G}})$ by the pairing
\[
(\hat \varphi \otimes \hat \varphi)([\hat\sigma_{i/2} \otimes \hat \sigma_{i/2}](x)\hat \Delta(a)) = \hat \varphi(\hat\Delta^\sharp(x)\hat\sigma_{-i/2}(a)) \qquad (a \in \mathcal O(\widehat{\mathbb{G}}) \odot \mathcal O(\widehat{\mathbb{G}}), \ x \in \mathcal O(\widehat{\mathbb{G}})).
\]
Unraveling this definition, one obtains the following concrete formula
\[
\hat \Delta^\sharp(u_{ij}^\pi \otimes u_{k,l}^\sigma) = \delta_{j,k}\delta_{\pi,\sigma}\frac{u_{il}^\pi}{d(\pi)} \qquad (\pi,\sigma \in \text{Irr}(\widehat{\mathbb{G}}), \ 1 \le i,j \le n(\pi), \ 1 \le k,l \le n(\sigma)).
\]
Now, if we use the map $\hat \Delta^\sharp$ in place of $\hat \Delta^{-1} \circ E$ in formula \eqref{centralization}, we obtain from any $a \in M_{cb}A(\mathbb{G})$ a new element $\tilde a \in ZM_{cb}A(\mathbb{G})$ with
\[
\tilde{a}^\pi = \frac{\text{Tr}(a^{\bar \pi})}{d(\pi)}p_\pi \qquad (\pi \in\text{Irr}(\widehat{\mathbb{G}})) \quad \& \quad \|\tilde a\|_{M_{cb}A(\mathbb{G})} \le \|a\|_{M_{cb}A(\mathbb{G})}.
\]
Observe that the only difference between this formula and the one in the unimodular case is that the classical dimension $n(\pi)$ is replaced by the quantum dimension $d(\pi)$. This difference has the unfortunate effect of producing only central multipliers which are ``uniformly far'' from the identity multiplier $1 \in M_{cb}A(\mathbb{G})$. Indeed, if $\|a\|_{M_{cb}(\mathbb{G})} \le C$, then $\|\tilde{a}^\pi\|_\infty \le C\frac{n(\pi)}{d(\pi)}$. In conclusion, this approach to constructing central multipliers in the non-unimodular setting for the purpose of proving central approximation properties is destined to fail.
In the next section we will see an alternate approach to establishing central approximation properties for non-unimodular discrete quantum groups using tools from monoidal equivalence for compact quantum groups.
\section{Central approximation properties and monoidal equivalence} \label{CAP-monoidal}
We continue our discussion of central approximation properties for discrete quantum groups by connecting these properties with the powerful notion of {\it monoidal equivalence} for compact quantum groups. This remarkable connection between these seemingly unrelated concepts has had a profound impact on the operator algebraic theory of compact quantum groups, and has led (for instance) to the proof of central weak amenability and central Haagerup property for all free quantum groups and the duals of quantum automorphism groups of finite dimensional C$^\ast$-algebras. See \cite{dCFY} and Section \ref{app}.
We begin with the definition of monoidal equivalence.
\begin{defn}[Bichon-De Rijdt-Vaes \cite{BiDeVa06}]
Let $\mathbb{G}_1, \mathbb{G}_2$ be two compact quantum groups. We say that $\mathbb{G}_1$ and $\mathbb{G}_2$ are {\it monoidally equivalent}, and write $\mathbb{G}_1 \sim^{mon} \mathbb{G}_2$, if there exists a bijection \[\Phi:\text{Irr}(\mathbb{G}_1) \to \text{Irr}(\mathbb{G}_2)\] together with linear isomorphisms
\[\Phi: \text{Mor}(\pi_1 \otimes \ldots \otimes \pi_n, \sigma_1 \otimes \ldots\otimes \sigma_m) \to \text{Mor}(\Phi(\pi_1)\otimes\ldots \otimes \Phi(\pi_n), \Phi(\sigma_1) \otimes \ldots \otimes \Phi(\sigma_m)) \]
such that $\Phi(1_{\mathbb{G}_1}) = 1_{\mathbb{G}_2}$ ($1_{\mathbb{G}_i}$ being the trivial representation of $\mathbb{G}_i$), and such that for any morphisms $S,T$,
\begin{align*}
\Phi(S \circ T) &=\Phi(S)\circ \Phi(T) \quad (\text{whenever $S \circ T$ is well-defined})\\
\Phi(S^*)&=\Phi(S)^* \\
\Phi(S \otimes T) &=\Phi(S) \otimes \Phi(T).
\end{align*}
\end{defn}
A monoidal equivalence between $\mathbb{G}_1$ and $\mathbb{G}_2$ means that the abstract monoidal C$^\ast$-tensor categories $\text{Rep}(\mathbb{G}_i)$ (consisting of all finite dimensional Hilbert space representations of $\mathbb{G}_i$ equipped with their morphism spaces $ \text{Mor}(\cdot, \cdot)$ and the tensor product operation $\otimes$) are isomorphic.
As a first (and very important) example, we consider monoidal equivalences between free orthogonal quantum groups $O^+_F$, where $F \in \text{GL}_N(\mathbb{C})$ with $N \ge 2$ and $F \bar F \in \mathbb{R} 1$.
\begin{thm}[\cite{BiDeVa06}] For $i=1,2$, fix $F_i \in \text{GL}_{N_i}(\mathbb{C})$ with $F_i\bar F_i = c_i1$ and $c_i \in \mathbb{R}$. Then the corresponding free orthogonal quantum groups $O^+_{F_1}$ and $O^+_{F_2}$ are monoidally equivalent if and only if \[\frac{c_1}{\text{Tr}(F_1^*F_1)} = \frac{c_2}{\text{Tr}(F_2^*F_2)}. \]
Moreover, any compact quantum group $\mathbb{G}$ such that $O^+_{F_1} \sim^{mon}\mathbb{G}$ is isomorphic to an $O^+_{F_2}$ of the above form.
\end{thm}
\begin{proof}[Sketch]
The basic idea (going back to Banica's seminal work \cite{Ba96}) is that if $U_i$ is the fundamental representation
of $O^+_{F_i}$, then the defining relations for this quantum group are equivalent to the requirement that $U_i$ be irreducible and self-conjugate ($U_i = F_i\bar{U_i}F_i^{-1})$, and that the vector \[t_{F_i} = \sum_{k=1}^{N _i}e_k \otimes F_ie_k \in \mathbb{C}^{N_i} \otimes \mathbb{C}^{N_i}\] belongs to $\text{Mor}(1, U_i \otimes U_i)$.
Now, since $O^+_{F_i}$ is the universal compact quantum group with the above properties, the Tannaka-Krein-Woronowicz reconstruction theorem \cite{Wo88} implies that $\text{Rep}(O^+_{F_i})$ must then be generated as a monoidal C$^\ast$-tensor category by the object $U_i$ together with the basic morphisms $\iota_{\mathbb{C}^{N_i}}$ and $t_{F_i}$. The monoidal equivalence $\Phi: O^+_{F_1} \sim^{mon} O^+_{F_2}$ can then be defined on these generators by setting
\[
\Phi(U_1) = U_2, \quad \Phi(\iota_{\mathbb{C}^{N_1}}) = \iota_{\mathbb{C}^{N_2}}, \quad \Phi\Big(\frac{1}{\text{Tr}(F_1^*F_1)^{1/2}}t_{F_1}\Big) = \frac{1}{\text{Tr}(F_2^*F_2)^{1/2}}t_{F_2}.
\]
and extending to all of $\text{Rep}(O^+_{F_1})$ via the categorical operations. Of course, one has to verify some details, but they can be found in \cite{BiDeVa06}.
\end{proof}
As a particular instance of the above theorem, recall that for $q \in [-1,1]\backslash \{0\}$, we have that \[SU_q(2) = O^+_{F_q} \quad \text{where} \quad F_q = \left( \begin{matrix} 0 &1\\
-q^{-1} & 0 \end{matrix}\right) \in \text{GL}_{2}(\mathbb{C}).\]
Noting that $F_q \overline{F_q} = -q1$ and $Tr(F_q^*F_q) = 1+q^{-2}$, we immediately obtain the following corollary, which says that every free orthogonal quantum group $O^+_F$
is monoidally equivalent to a suitable $SU_q(2)$. The result is quite remarkable, since $SU_q(2)$ is always co-amenable, while $O^+_F$ is generally not.
\begin{cor} \label{mon-equiv-SU}
Let $F \in \text{GL}_N(\mathbb{C})$ and $F\bar F = c1$ for $c \in \mathbb{R}$. Then $O^+_F$ is monoidally equivalent to $SU_q(2)$ for the unique value of $q \in [-1,1]\backslash \{0\}$ satisfying $\text{Tr}(F^*F) = -c(q+q^{-1})$.
\end{cor}
\subsection{Central multipliers and monoidal equivalence}
We now make the connection between monoidal equivalence and central approximation properties. The key to everything is the following theorem of Freslon \cite{Fr13} on transferring central completely bounded multipliers between duals of monoidally equivalent compact quantum groups.
\begin{thm}[Freslon \cite{Fr13}] \label{thm:mon-equiv-ap}
Let $\mathbb{G}_1,\mathbb{G}_2$ be discrete quantum groups and assume that $\Phi: \widehat{\mathbb{G}}_1 \sim^{mon}\widehat{\mathbb{G}}_2$ is a monoidal equivalence. Given any central function $a_1 = (a^\pi p_\pi)_{\pi \in \text{Irr}(\widehat{\mathbb{G}}_1)} \in ZL^\infty(\mathbb{G}_1)$, define $a_2 =(a^\pi p_{\Phi(\pi)})_{\pi \in \text{Irr}(\widehat{\mathbb{G}}_1)} \in ZL^\infty(\mathbb{G}_2)$.
Then we have that $a_1 \in ZM_{cb}A(\mathbb{G}_1)$ if and only if $a_2 \in ZM_{cb}A(\mathbb{G}_2)$, and \[\|a_1\|_{M_{cb}A(\mathbb{G}_1)} = \|a_2\|_{M_{cb}A(\mathbb{G}_2)}.\]
Moreover, $a_1$ is a normalized completely positive definite function if and only if $a_2$ is.
\end{thm}
In order to prove this theorem, the key idea is to compare the cb multiplier norms for $a_1$ and $a_2$ using the algebraic/analytic structure of the {\it linking algebra} associated to the given monoidal equivalence.
\begin{thm}[Bichon-De Rijdt-Vaes \cite{BiDeVa06}] \label{link-algebra}
Let $\mathbb{G}_1 \sim^{mon} \mathbb{G}_2$ be monoidally equivalent compact quantum groups with monoidal equivalence $\Phi$. Then
\begin{enumerate}
\item There exists a unique unital $\ast$-algebra $\mathcal B:= \mathcal O(\mathbb{G}_1,\mathbb{G}_2)$ (the linking algebra) equipped with a faithful state $\omega$ and unitary elements $X^\pi \in \mathcal B(H_\pi, H_{\Phi(\pi ))}) \otimes \mathcal B$ for each $\pi \in \text{Irr}(\mathbb{G})$, satisfying
\begin{enumerate}
\item The matrix elements of $(X^\pi)_\pi$ form a linear basis of $\mathcal B$.
\item $X^\sigma_{13}X^\gamma_{23}(S \otimes 1) = (\Phi(S) \otimes 1) X^\pi$ for each $S \in \text{Mor}(\pi, \sigma \otimes \gamma)$.
\item $(\iota \otimes \omega)(X^\pi) = 0$ if $\pi \ne 1$.
\end{enumerate}
\item There exists a unique pair of commuting actions $\mathbb{G}_1 \curvearrowright^{\delta_1} \mathcal B \curvearrowleft^{\delta_2} \mathbb{G}_2 $ (i.e., unital $\ast$-homomorphisms $\delta_1: \mathcal B \to \mathcal B \otimes \mathcal O(\mathbb{G}_1)$ and $\delta_2: \mathcal B \to \mathcal O(\mathbb{G}_2) \otimes \mathcal B$ such that \[(\delta_1 \otimes \iota)\delta_1 = (\iota \otimes \Delta_1)\delta_1, \quad (\iota \otimes \delta_2)\delta_2 = (\Delta_2 \otimes \iota) \delta_2 \] and
\[ \text{span}(\delta_1(\mathcal B) (1 \otimes \mathcal O(\mathbb{G}_1))) = \mathcal B \otimes \mathcal O(\mathbb{G}_1), \text{span}(\delta_2(\mathcal B) (\mathcal O(\mathbb{G}_2)\otimes 1)= \mathcal O(\mathbb{G}_2) \otimes \mathcal B.\] These actions are given by the formulas
\begin{align}\label{action-formula}(\iota \otimes \delta_1)(X^\pi) = X^\pi_{12}U^\pi_{13} \qquad (\iota \otimes \delta_2)(X^\pi) = U^{\Phi(\pi)}_{12}X^{\pi}_{13} \qquad (\pi \in \text{Irr}(\mathbb{G}_1)).
\end{align}
\item The state $\omega:\mathcal B \to \mathbb{C}$ is invariant under the actions $\delta_1, \delta_2$. I.e.,
\[(\iota \otimes \omega)\delta_2 = (\omega \otimes \iota)\delta_1 = \omega(\cdot)1.\]
\end{enumerate}
Denote by $B_r$ the C$^\ast$-algebra generated in the GNS-representation with respect to $\omega$ and by $B_u = C^*_u(\mathcal B)$ the universal enveloping C$^\ast$-algebra of $\mathcal B$. Then $\delta_i$ extends uniquely to a coaction of C$^\ast$-algebraic CQGs on $B_r$ and $B_u$, respectively.
\end{thm}
In order to motivate what follows, it is helpful to keep in mind the trivial example of a monoidal equivalence $\mathbb{G}_1 \sim^{mon} \mathbb{G}_1$ given by the identity map. In this case, $\mathcal O(\mathbb{G}_1,\mathbb{G}_1) = \mathcal O(\mathbb{G}_1)$, $\omega = \varphi_{\mathbb{G}_1}$ is the Haar state, and $\delta_i$ is just the left/right action of $\mathbb{G}_1$ on $\mathcal O(\mathbb{G}_1)$ induced by the coproduct $\Delta_{\mathbb{G}_1}$. Thinking about the linking algebra $\mathcal B$ in terms of this special case will help make the formulas in the above proposition appear more natural.
\begin{proof}[Rough outline of Theorem \ref{link-algebra}]
As a vector space, we define \[\mathcal B = \bigoplus_{\pi \in \text{Irr}(\mathbb{G})} \mathcal B(H_\pi,H_{\Phi(\pi)})^*\] and we define natural elements $X^\pi \in \mathcal B(H_\pi, \otimes H_{\Phi(\pi)}) \otimes \mathcal B$ by $(\omega_\sigma \otimes \iota)X^\pi = \delta_{\pi, \sigma}\omega_\sigma$ for all $\omega_\sigma \in \mathcal B(H_\sigma, H_{\Phi(\sigma)})^*$. We can then define a unique bilinear map $\mathcal B \times \mathcal B \to \mathcal B$ using the relation (b) from the statement of the theorem. The associativity
of this map follows from the calculation
\begin{align*}
&(X_{14}^aX_{24}^b)X_{34}^c((S \otimes 1)T\otimes 1) \\
&= (\Phi((S \otimes 1)T)\otimes 1)X^x \\
&=X_{14}^a(X_{24}^bX_{34}^c)((S \otimes 1)T\otimes 1) \qquad (S \in \text{Mor}(y,a \otimes b), \ T \in \text{Mor}(x,y \otimes c)),
\end{align*} and the fact that maps of the form $(S \otimes 1)T$ linearly span $\text{Mor}(x,a\otimes b \otimes c)$.
The involution $b \mapsto b^*$ on $\mathcal B$ is defined by
\[(X^\pi)^*_{13}(\Phi(t) \otimes 1) = X^{\bar \pi}_{23}(t \otimes 1) \qquad (t \in \text{Mor}(1, \pi \otimes \bar \pi)). \]
Equivalently, we have \[((\omega_{\xi, \eta} \otimes \iota)(X^\pi))^* = (\omega_{(\xi^*\otimes 1)t, (1 \otimes \eta^*)\Phi(\tilde{t})}\otimes \iota)X^{\bar \pi},\] where $\tilde t \in \text{Mor}(1, \bar \pi \otimes \pi)$ is such that $(t^* \otimes 1)(1\otimes \tilde{t}) =1$.
One can then check that the above constructed elements $X^x$ are unitaries with respect to this $\ast$-algebra structure. One can also check that if $a = \sum_{\pi, i} (\omega_{\xi_i^\pi, f_i^\pi} \otimes \iota)X^\pi \in \mathcal B$, where $(f_i^\pi)_i$ is an orthonormal basis for $H_{\Phi(\pi)}$, then \[\omega(aa^*) = \sum_{\pi,i} \frac{1}{d(\pi)}\langle Q_\pi \xi_i^\pi |\xi_i^\pi\rangle \ge 0.\]
Hence $\omega$ is a state.
Given the above description of $\mathcal B$, the construction of the actions $\delta_1, \delta_2$ according to \eqref{action-formula} is well-defined, and they automatically satisfy the claimed $\omega$-invariance condition. This $\omega$-invariance in turn implies the existence of the extension of $\delta_i$ to a C$^\ast$-algebraic action of $\mathbb{G}_i$ on $B_r$. The case of $\delta_i$ extending to an action on $B_u$ follows from standard considerations. See for example \cite[Lemma 2.13]{dC16}.
\end{proof}
By symmetry, for every monoidal equivalence of compact quantum groups $\Phi:\mathbb{G}_1 \sim^{mon}\mathbb{G}_2$, we have the inverse monoidal equivalence \[\Phi^{-1}:\mathbb{G}_2 \sim^{mon} \mathbb{G}_1,\] and we can repeat the above construction for $\Phi^{-1}$ which yields a corresponding $\ast$-algebra $(\tilde {\mathcal B}, \tilde \omega)$ with faithful state $\tilde \omega$. Let $(Y^\pi)_{\pi \in \text{Irr}(\mathbb{G}_1)}$ be the unitary matrices corresponding to the algebra $\tilde {\mathcal B} = \mathcal O(\mathbb{G}_2,\mathbb{G}_1)$. Then one has a canonical $\ast$-anti-isomorphism \[\rho: B_r \to \tilde B_r; \qquad (\iota \otimes \rho)(X^\pi) = (Y^\pi)^* \qquad (\pi \in \text{Irr}(\mathbb{G}_1)).\]
Thus $\tilde B_r \cong B_r^{op}$. See \cite{VaVe07} for details.
Before getting to the proof of Theorem \ref{thm:mon-equiv-ap}, we need one more important lemma due to Vaes and Vergnioux. See \cite[Theorem 6.1]{VaVe07}.
\begin{lem} \label{VaVe-lem}
Let $\Phi:\mathbb{G}_1 \sim^{mon}\mathbb{G}_2$ be a monoidal equivalence with reduced linking algebras $B_r$ and $\tilde{B}_r$ as above. Then there exists an injective $\ast$-homomorphism \[\theta:C(\mathbb{G}_1) \to \tilde B_r \otimes B_r; \quad \text{given by } (\iota \otimes \theta)U^\pi = Y^\pi_{12}X^\pi_{13} \in \mathcal B(H_\pi) \otimes \tilde{B_r} \otimes B_r \qquad (\pi \in \text{Irr}(\mathbb{G}_1)).\]
\end{lem}
\begin{proof}
A priori, $\theta$ is defined as a map from $\mathcal O(\mathbb{G}_1)$ to $\tilde{\mathcal B} \odot \mathcal B$, and the fact that it is a $\ast$-homomorphism is a routine calculation based on the definitions of $\mathcal B$ and $\tilde {\mathcal B}$.
To obtain the extension to a $\ast$-homomorphism $\theta:C(\mathbb{G}_1) \to \tilde B_r \otimes B_r$, we just observe that $(\tilde \omega \otimes \omega) \theta = \varphi_{\mathbb{G}_1}$. Therefore $\theta$ extends to the GNS completions.
\end{proof}
\subsection{Proof of Theorem \ref{thm:mon-equiv-ap}}
Fix a monoidal equivalence $\Phi:\widehat{\mathbb{G}}_1 \sim^{mon} \widehat{\mathbb{G}}_2$ and let $B_r$ be the corresponding reduced linking algebra. For each $\pi \in \text{Irr}(\widehat{\mathbb{G}}_1)$, let $B_\pi \subset B_r$ denote the linear span of the coefficients of the unitary $X^\pi$ (the so-called {\it spectral subspace of $\pi$}), and observe that with respect to the $L^2(B_r,\omega)$-Hilbert space structure, $B_r$ is the orthogonal direct sum $\bigoplus^{\ell^2}_{\pi \in \text{Irr}(\mathbb{G}_1)}B_\pi$. Finally, denote by $P_\pi:B_r \to B_r$ the orthogonal projection onto the spectral subspace $B_\pi$, but viewed as a finite rank idempotent map on the C$^\ast$-algebra $B_r$.
Now, observe that for each $\pi \in \text{Irr}(\widehat{\mathbb{G}}_1)$, our action $\delta_1:B_r \to B_r \otimes C(\widehat{\mathbb{G}}_1)$ from Theorem \ref{link-algebra} satisfies the equation \[(\iota \otimes L^{(p_\pi)}) \circ \delta_1 = \delta_1 \circ P_\pi,\]
where $L^{(p_\pi)} \in \mathcal CB(C(\widehat{\mathbb{G}}_1))$ is the adjoint of the multiplier $p_\pi \in ZM_{cb}A(\mathbb{G}_1)$. By continuous linear extension, this yields the equality of (possibly unbounded) linear maps
\[
(\iota \otimes L^{(a_1)}) \circ \delta_1 = \delta_1 \circ \Big( \sum_{\pi \in \text{Irr}(\widehat{\mathbb{G}}_1)}a^\pi P_\pi \Big),
\] for any $a_1 = (a^\pi p_\pi)_{\pi \in \text{Irr}(\widehat{\mathbb{G}}_1)} \in ZL^\infty(\mathbb{G}_1)$. But $\delta_1$ is an injective (hence completely isometric) $\ast$-homomorphism, hence we obtain we get
\begin{align}\label{one}\Big\|\sum_{\pi \in \text{Irr}(\widehat{\mathbb{G}}_1)}a^\pi P_\pi
\Big\|_{\mathcal {CB}(B_r)} =\Big\|\delta_1 \circ \Big(\sum_{\pi \in \text{Irr}(\widehat{\mathbb{G}}_1)}a^\pi P_\pi
\Big)\Big\|_{\mathcal {CB}(B_r)}= \|(\iota \otimes L^{(a_1)}) \circ \delta_1\|_{cb} \le \|a_1\|_{M_{cb}A(\mathbb{G}_1)}.
\end{align}
On the other hand, Lemma \ref{VaVe-lem} provides an injective $\ast$-homomorphism $\tilde\theta:C_r(\mathbb{G}_2) \to B_r \otimes \tilde B_r$, which, by construction, satisfies
\[\tilde \theta \circ L^{(p_{\Phi(\pi)})} = (P_\pi \otimes \iota) \tilde \theta \qquad (\pi \in \text{Irr}(\widehat{\mathbb{G}}_1)) \implies \tilde \theta \circ L^{(a_2)} = \Big( \Big( \sum_\pi a^\pi P_\pi \Big) \otimes \iota\Big) \tilde \theta, \]
where $a_2 = (a^\pi p_{\Phi(\pi)})_\pi \in ZL^\infty(\mathbb{G}_2)$.
Therefore
\begin{align} \label{two}\|a_2\|_{M_{cb}A(\mathbb{G}_2)} = \|\tilde \theta \circ L^{(a_2)}\|_{cb} = \Big\| \Big( \Big( \sum_\pi a^\pi P_\pi \Big) \otimes \iota\Big) \tilde \theta\Big\|_{cb} \le \Big\|\sum_{\pi \in \text{Irr}(\widehat{\mathbb{G}}_1)}a^\pi P_\pi
\Big\|_{\mathcal {CB}(B_r)}
\end{align}
From \eqref{one} and \eqref{two}, we obtain
\[
\|a_1\|_{M_{cb}A(\mathbb{G}_1)} \ge \|a_2\|_{M_{cb}A(\mathbb{G}_2)},
\]
and by switching the roles of $\mathbb{G}_1$ and $\mathbb{G}_2$, the reverse inequality must hold as well. Finally, we observe that complete positive definiteness is preserved by the bijection $ZM_{cb}A(\mathbb{G}_1) \owns a_1 \longleftrightarrow a_2 \in ZM_{cb}A(\mathbb{G}_2)$, since $L^{(a_1)}$ is unital complete contraction if and only $L^{(a_2)}$ is.
\section{Application: Approximation properties for free quantum groups} \label{app}
In this final section, our goal is to apply some of the techniques of the previous sections to study the Haagerup property and weak amenability for some examples of discrete free quantum groups. The main family of examples that we will consider are the free orthogonal quantum groups. This class of examples turns out to be the key to understanding all discrete free quantum groups. The main result we plan to discuss is the following:
\begin{thm}[\cite{dCFY}] \label{approx-prop-fqg}
Let $N \ge 2$, $F \in \text{GL}_N(\mathbb{C})$ with $F\bar F \in \R1$, and put $\mathbb{F} O_F = \widehat{O^+_F}$. Then the discrete quantum group $\mathbb{F} O_F$ has the central Haagerup property and is centrally weakly amenable with $Z\Lambda_{cb}(\mathbb{F} O_F) = 1$
\end{thm}
\begin{rem}
The quantum group $\mathbb{F} O_F$ is an example of a (discrete) free quantum group in the sense of \cite{VeVo13}. In fact, as a consequence of a structure theorem of Wang \cite{Wa02} a general free quantum group turns out to be a free product of the form
\[
\mathbb{G} = \mathbb{F} O_{F_1} \ast \mathbb{F} O_{F_2}\ast \ldots \ast \mathbb{F} O_{F_k} \ast \mathbb{F} U_{Q_1} \ast \ldots \ast \mathbb{F} U_{F_l},
\]
where $\mathbb{F} U_{Q_i}:= \widehat{U^+_{Q_i}}$ for some $Q_i \in \text{GL}_{N_i}(\mathbb{C})$ and $F_j \in \text{GL}_{N_j}(\mathbb{C})$ $F_j \bar F_j \in \mathbb{R} 1$.
In fact, Theorem \ref{approx-prop-fqg} is true for any free quantum group $\mathbb{G}$ of the above form. The key idea to proving the more general form of Theorem \ref{approx-prop-fqg} is to first prove it for the special case as stated above, then extend to the $\mathbb{F} U_Q$ case using a free product argument, and then get the general case with yet another free product argument. See \cite[Section 5]{dCFY} for details. In summary, the study of $\mathbb{F} O_F$ with $F \bar F \in \mathbb{R} 1$ is the crucial one.
\end{rem}
\begin{rem}
It is also worth mentioning that Theorem \ref{approx-prop-fqg} has also been used together with monoidal equivalence and free product results from \cite{dRVa10} and \cite{dCFY} to obtain analogous results for the discrete duals of quantum automorphism groups of finite dimensional C$^\ast$-algebras.
These results for quantum automorphism groups have, in turn, been used to prove central approximation properties for the discrete duals of free wreath products of compact quantum groups by quantum automorphism groups. See for example \cite{Lemeux, LeTa14, FiPi15}.
\end{rem}
\subsection{Theorem \ref{approx-prop-fqg} in the unimodular case}
Let us first sketch the proof of Theorem \ref{approx-prop-fqg} in the simplest case where $\mathbb{F} O_F$ is unimodular. The results obtained in this case will give us some intuition for what to expect when we consider the general case.
Now, if $\mathbb{F} O_F$ is unimodular, then (up to isomorphism of compact duals) this corresponds to only two cases: the first one being where $F = 1 \in \text{GL}_N(\mathbb{C})$, or $N = 2l$ and $F = \text{diag}(\underbrace{F_0, \ldots, F_0}_{l \text{ times }})$, where $F_0 = \left[ \begin{matrix}
0 & 1 \\
-1 &0
\end{matrix}\right]$. See \cite[Remark 4.3]{FiVe15} and \cite[Remark 5.7]{BiDeVa06}.
We begin with a useful lemma.
\begin{lem} \label{characterize-cppd}
Let $\mathbb{G}$ be a unimodular discrete quantum group and let $\mathcal ZC^u(\widehat{\mathbb{G}}) \subseteq C^u(\widehat{\mathbb{G}})$ be the C$^\ast$-subalgebra generated by the irreducible characters $\{\chi_\pi\}_{\pi \in \text{Irr} (\widehat{\mathbb{G}})} \subset C^u(\widehat{\mathbb{G}})$. Then a central element $a = (a^\pi p_\pi)_{\pi \in \text{Irr}(\widehat{\mathbb{G}})} \in ZL^\infty(\mathbb{G})$ is a normalized completely positive definite function if and only if there exists a state $\psi \in \mathcal ZC^u(\mathbb{G})^*$ such that
\[
a^\pi = \frac{\psi(\chi_{\pi})}{n(\pi)} \qquad (\pi \in \text{Irr}(\widehat{\mathbb{G}})).
\]
\end{lem}
\begin{proof}
If $a= (a^\pi p_\pi)_{\pi \in \text{Irr}(\widehat{\mathbb{G}})} \in ZM_{cb}A(\mathbb{G})$ is a normalized completely positive definite function, then the corresponding map
\[
L^{(a)}: \mathcal O(\widehat{\mathbb{G}}) \to \mathcal O(\widehat{\mathbb{G}}); \qquad L^{(a)}u_{ij}^\pi = a^\pi u_{ij}^\pi \qquad (1 \le i,j \le n(\pi)), \ \pi \in \text{Irr}(\widehat{\mathbb{G}}))
\] is unital and completely positive. It then follows that the linear functional
\[
\tilde \psi:\mathcal O(\widehat{\mathbb{G}}) \to \mathbb{C}; \qquad \tilde \psi = \hat \epsilon \circ L^{(a)}.
\]
is a state, and hence extends to a state $\tilde \psi \in C^u(\widehat{\mathbb{G}})^*$.
Setting $\psi = \tilde\psi|_{\mathcal ZC^u(\widehat{\mathbb{G}})}$ (which is a state on $\mathcal ZC^u(\widehat{\mathbb{G}})$), we obtain
\[
\psi(\chi_\pi) = \sum_{i=1}^{n(\pi)} \tilde{\psi}(u_{ii}^\pi) = \sum_{i=1}^{n(\pi)} \hat \epsilon (a^\pi u_{ii}^\pi) = n(\pi) a^\pi \qquad (\pi \in \text{Irr}(\widehat{\mathbb{G}})).
\]
Conversely, let $\psi \in \mathcal ZC^u(\widehat{\mathbb{G}})^*$ be a state. Then we can extend $\psi$ to a state
$\tilde\psi \in C^u(\widehat{\mathbb{G}})^*$ (by the Hahn-Banach theorem). Let $\tilde{a} = (\tilde{a}^\pi)_{\pi \in \text{Irr}(\widehat{\mathbb{G}})} \in L^\infty(\mathbb{G})$ be given by
\[
\tilde{a}^\pi_{ij} = \tilde{\psi}((u^\pi_{ji})^*) \qquad (1 \le i,j \le n(\pi)), \ \pi \in \text{Irr}(\widehat{\mathbb{G}})).
\]
Then $\tilde a$ is a normalized completely positive definite function on $\mathbb{G}$ (since $\psi \circ \hat S$ is a state and $L^{(\tilde a)} = ((\tilde\psi \circ \hat S) \otimes \iota)\hat \Delta$ is unital and completely positive on $\mathcal O(\widehat{\mathbb{G}})$, and this map is readily seen to factor through the quotient $C^u(\widehat{\mathbb{G}}) \to C(\widehat{\mathbb{G}})$).
Applying the averaging argument in the proof of Theorem \ref{unimodular} to the multiplier $\tilde{a}$, we obtain a new central normalized completely positive definite function $a = (a^\pi)_{\pi \in \text{Irr}(\widehat{\mathbb{G}})}$
given by
\[
a^\pi = \frac{\text{Tr}(\tilde a^\pi)}{n(\pi)} = \frac{\psi(\chi_\pi^*)}{n(\pi)} = \frac{\psi(\chi_{\bar \pi})}{n(\pi)} \qquad (\pi \in \text{Irr}(\widehat{\mathbb{G}})).
\]
\end{proof}
So, in order to prove the central Haagerup property for $\mathbb{F} O_F$ in the unimodular case, we need to get our hands on $\mathcal Z C^u(O^+_F)$. For this, we recall a few facts from Banica's seminal work \cite{Ba96}:
\begin{thm}[\cite{Ba96}] \label{banica-result}Let $F \in \text{GL}_N(\mathbb{C})$ with $F\bar F = \pm 1$. There is a maximal family $(U^k)_{k \in \mathbb{N}_0}$ of pairwise inequivalent irreducible unitary representations of $O^+_F$ such that
\[
U^0 = 1, \quad U^1 = [u_{ij}]_{1 \le i,j \le N} \quad \text{ (the fundamental representation)},
\]
and for $k \ge 1$, we have
\begin{align} \label{fusion-rules}\overline{U^k} \cong U^k \quad \& \quad
U^1 \otimes U^k \cong U^{k+1} \oplus U^{k-1} \cong U^k \otimes U^1.
\end{align}
Moreover, if $(\mu_k)_{k \in \mathbb{N}_0}$ denote the (dilated) type 2 Chebychev polynomials determined by the initial conditions and recursion
\begin{eqnarray} \label{eqn_chebyshev}
\mu_0(x) = 1, \quad \mu_1(x) = x, \quad x\mu_k(x) = \mu_{k+1}(x) + \mu_{k-1}(x) && (k \ge 1),
\end{eqnarray} and $q \le q_0 \in (0,1]$ are defined so that
\[
N = q_0 + q_0^{-1} \quad \& \quad \text{Tr}(F^*F) = q+q^{-1},
\]
then we have the following (quantum) dimension formulas for $U^k$:
\[
n(k) = \mu_k(q_0 + q_o^{-1}) \quad \& \quad d(k) = \mu_k(q+q^{-1}) \qquad (k \in \mathbb{N}_0).
\]
\end{thm}
Using Theorem \ref{banica-result} we can obtain the following proposition.
\begin{prop} \label{char-algebra}
Let $N \ge 2$ and let $F \in \text{GL}_N(\mathbb{C})$ with $F\bar F = \pm 1$ with $\mathbb{F} O_F$ unimodular. Then there is a $\ast$-isomorphism
\[
\mathcal ZC^u(O^+_F) \cong C([-N,N]) \quad \text{such that} \quad \chi_k \mapsto \mu_k|_{[-N,N]} \qquad (k \in \mathbb{N}_0),
\]
where $\chi_k$ is the character of the representation $U^k$ and $\mu_k$ is the $k$th Chebychev polynomial defined above.
\end{prop}
\begin{proof}
We have $\mathcal Z C^u(O^+_F) = C^*(\chi_k: k \in \mathbb{N}_0) = C^*(1 ,\chi_1)$, where the last equality follows from the self-conjugacy and fusion rules for the irreducible representations $(U^k)_{k \in \mathbb{N}_0}$ given by \eqref{fusion-rules}. Moreover, $\chi_1$ is self-adjoint, so the Gelfand map yields a $\ast$-isomorphism \[\mathcal Z C^u(O^+_F) \to C(\sigma(\chi_1)); \quad 1 \mapsto 1, \ \chi_1 \mapsto \mu_1|_{\sigma(\chi_1)},\]
where $\sigma(\chi_1)$ is the spectrum of $\chi_1 \in C^u(O^+_F)$. Since we also have also $\chi_k = \mu_k(\chi_1)$ for each $k \ge 0$, it remains to show that $\sigma(\chi_1) = [-N,N]$.
To this end, we first note that $\sigma(\chi_1) \subseteq [-N,N]$, since $\chi_1 = \sum_{i=1}^N u_{ii}$ is a sum of $N$ contractions. For the reverse inclusion, consider first the case where $F=1$. Then there is a quotient map $C^u(O_F^+) \to C(O_N)$ (= the C$^\ast$-algebra of continuous functions on the orthogonal group $O_N$). This map sends $\chi_1$ to the character $\chi_1^{O_N}$ of the fundamental representation of $O_N$, hence $\sigma(\chi_1) \supseteq \sigma(\chi_1^{O_N})$. But this latter set is well known to be $[-N,N]$ -- since the trace of an $N\times N$ orthogonal matrix can be real number in $[-N,N]$.
For the other case, we have $N = 2l$ and $F = \text{diag}(\underbrace{F_0, \ldots, F_0}_{l \text{ times }})$, where $F_0 = \left[ \begin{matrix}
0 & 1 \\
-1 &0
\end{matrix}\right]$. In this case, one readily sees that we have a quotient map from $C^u(O^+_F)$ onto $C(\mathbb{T})$ given by
\[U^1 = U \mapsto V = \Big\{\theta \mapsto \text{diag}\Big(\underbrace{\left[\begin{matrix} \cos \theta & \sin \theta \\
-\sin(\theta) & \cos(\theta) \end{matrix}\right], \ldots, \left[\begin{matrix} \cos \theta & \sin \theta \\
-\sin(\theta) & \cos(\theta) \end{matrix}\right] }_{l \text{ times }}\Big) \Big\} \in M_{2l}(O_{2l}), \qquad (\theta \in \mathbb{R})\]
In particular, $\chi_1$ gets sent to the function $\theta \mapsto f(\theta) = \text{Tr}(V(\theta)) = 2l\cos(\theta)$, whose spectrum is $[-2l,2l] = [-N,N]$, completing the proof.
\end{proof}
Combining Lemma \ref{characterize-cppd} and Proposition \ref{char-algebra}, we immediately obtain the following characterization of completely positive definite functions on $\mathbb{F} O_F$ in the unimodular case. This result was implicitly obtained in the $F=1$ case in \cite{Br12}. See also \cite{CiFrKu14} for a nice treatment.
\begin{thm} \label{Br12thm} \label{Khintchine}
Let $\mathbb{F} O_F$ be unimodular as above. Then there is a one-to-one correspondence between
central normalized completely positive definite functions $a=(a^k)_{k\in \mathbb{N}_0} \in ZM_{cb}A(\mathbb{F} O_F)$ and Borel probability measures $\mu$ on $[-N,N]$ given by
\[a^k = \frac{\int_{[-N,N]}\mu_k(s)d\mu(s)}{\mu_k(N)} \qquad (k \in \mathbb{N}_0). \]
\end{thm}
We are almost ready to prove Theorem \ref{approx-prop-fqg} in the unimodular case. The last ingredient we need is a remarkable result of Freslon, which gives polynomial bounds on the growth of the $M_{cb}A(\mathbb{F} O_F)$-norm of the sequence of finite rank central projections $p_k \in Z L^\infty(\mathbb{F} O_F)$ corresponding to the irreducible representation $U^k$ of $O^+_F$. The following theorem should be thought of as a quantum group analogue of Buchholz's operator-valued Haagerup inequality on free groups \cite{Bu99} and the resulting cb-norm estimates for projections onto the space of convolution operators on free groups supported on words of a fixed length \cite[Section 9]{Pi03}.
\begin{thm}[\cite{Fr12}] \label{freslon-cbnorm}
Let $N \ge 3$, $F \in \text{GL}_N(\mathbb{C})$ with $F \bar F = \pm 1$, and let $0 < q <1$ be such that $\text{Tr}(F^*F) = q+q^{-1}$. Let $p_k \in ZL^\infty(\mathbb{F} O_F)$ be the central projection corresponding to the $k$-th irreducible representation of $O^+_F$. Then there is a constant $C(q) > 0 $ so that
\[\|p_k\|_{M_{cb}A(\mathbb{F} O_F)} \le C(q)k^2 \qquad (k \ge 1).\]
\end{thm}
\begin{proof}[Proof of Theorem \ref{approx-prop-fqg} (unimodular case)]
We can assume $N \ge 3$, since $N=2$ corresponds to $SU_{\pm 1}(2)$, which are centrally amenable. For $2 < s_0 <s < N$, consider the Dirac measure on $[-N,N]$ supported at $s$. From Theorem \ref{Khintchine}, we obtain a net of normalized completely positive definite functions $(a_s)_{s \in (s_0 <s < N)} \subset ZM_{cb}A(\mathbb{F} O_F)$ such that
\[a_s^k = \frac{\mu_k(s)}{\mu_k(N)} \qquad (k \in \mathbb{N}_0).\]
From the structure of the Chebychev polynomials $(\mu_k)_k$ it is possible to show (see \cite[Proposition 4.4]{Br12}) that there is a constant $D > 0$ such that
\begin{align}\label{growth}0 < a^k_s \le D\Big(\frac{s}{N}\Big)^k \qquad (k \in \mathbb{N}_0, \ s_0 < s < N).
\end{align} Since we also have that $\lim_{s \to N}a_s = 1$ pointwise on $\mathbb{N}_0$, $(a_s)_{s_0 < s < N}$ is a bounded approximate identity for $C_0(\mathbb{F} O_F)$. Therefore $\mathbb{F} O_F$ has the central Haagerup property.
To prove that $Z\Lambda_{cb}(\mathbb{F} O_F) = 1$, the idea is to use a standard technique of truncating the net $(a_s)_s \subset ZM_{cb}A(\mathbb{F} O_F)$ to obtain a net of finitely supported central elements of $A(\mathbb{F} O_F)$ with the following properties: (1) they still converge pointwise to $1$, and (2) we still have control over their $M_{cb}A(\mathbb{F} O_F)$-norms.
To this end, for each $s_0 < s < N$ and $n \in \mathbb{N}$ define
\[a_{s,n} = \sum_{k=0}^n a^s_k p_k \in ZA(\mathbb{F} O_F).\]
Then we have from Theorem \ref{freslon-cbnorm} and inequality \eqref{growth}
\[\|a_s - a_{s,n}\|_{M_{cb}A(\mathbb{F} O_F)} \le \sum_{k=n+1}^\infty a^k_s\|p_k\|_{M_{cb}A(\mathbb{F} O_F)} \le \sum_{k=n+1}^\infty D\Big(\frac{s}{N}\Big)^k C(q)k^2.\]
Hence \[\lim_{n \to \infty} \|a_s - a_{s,n}\|_{M_{cb}A(\mathbb{F} O_F)} \to 0 \qquad (s_0 < s < N).\]
Thus the net $(a_s)_{s_0 < s < N}$ can be approximated in the cb-multiplier norm by elements of $A(\mathbb{F} O_F)$. It is then a routine exercise to extract a net $(a_i)_{i \in I}$ from the family $(a_{s,n})_{s,n}$ such that $a_i \to 1$ pointwise and $\limsup_i \|a_i\|_{M_{cb}A(\mathbb{F} O_F)} = \sup_s\|a_s\|_{M_{cb}A(\mathbb{F} O_F)} = 1$. In particular, $Z \Lambda_{cb}(\mathbb{F} O_F)=1$.
\end{proof}
\subsection{Theorem \ref{approx-prop-fqg} in the non-unimodular case}
We now consider a general $\mathbb{F} O_F$ with $F \in \text{GL}_N(\mathbb{C})$, $N \ge 2$ and $F\bar F \in \R1$. We start by recalling Corollary \ref{mon-equiv-SU}, which says that there is a unique
$q \in [-1,1] \backslash \{0\}$ such that the compact dual $O^+_F$ is monoidally equivalent to $SU_q(2)$. Therefore it suffices by Theorem \ref{thm:mon-equiv-ap} to show that $\widehat{SU_q(2)}$ has the central Haagerup property and $Z\Lambda_{cb}(\widehat{SU_q(2)}) = 1$. But from the arguments in the previous section in the unimodular case, Theorem \ref{thm:mon-equiv-ap} already tells us that this is true for certain values of $q$. More precisely, we have
\begin{cor}[\cite{Fr13}] \label{kactononkac}
Let $q \in [-1,1] \backslash \{0\}$ be such that $q+q^{-1} \in \mathbb{Z}$. Then $SU_q(2)$ is monoidally equivalent to a Kac type $O^+_F$ quantum group for some $F \in \text{GL}_N(\mathbb{C})$ satisfying $F \bar F \in \mathbb{R} 1$ and $N = |q| + |q|^{-1}$. In particular, $\widehat{SU_q(2)}$ has the central Haagerup property and $\Lambda_{cb}(\widehat{SU_q(2)}) = 1$. Moreover, if $q \ne \pm 1$, the central Haageup property for $Z\widehat{SU_q(2)}$ is implemented by the net of central multipliers $(a_s)_{s_0 < s < |q| + |q|^{-1}} \in ZM_{cb}A(\widehat{SU_q(2)})$ given by
\[a_s^k = \frac{\mu_k(s)}{\mu_k(|q| + |q|^{-1})} \qquad (k \in \mathbb{N}_0).\]
\end{cor}
Interpreting the multipliers $(a_s)_s$ appearing in Corollary \ref{kactononkac} in terms of the states on $C(SU_q(2)) = C^u(SU_q(2))$ that they implement via the formula \[x \mapsto \psi_s(x) = \epsilon (L^{(a_s)}(x)) \qquad (x \in \mathcal O(SU_q(2))),\] this suggests that for any $q \in (-1,1)\backslash \{0\}$, there ought to exist states $(\psi_s)_{s_0 <s < |q|+|q|^{-1}} \subseteq C(SU_q(2))^*$ defined by
\begin{align}\label{voigt-state}
\psi_s(u_{ij}^k) = \frac{\delta_{ij}\mu_k(s)}{\mu_k(|q|+|q|^{-1})} \qquad (\text{where } U^k = [u_{ij}^k] \text{ is the $k$th irrep. of $SU_q(2)$}).
\end{align}
It turns out that such states do indeed exist, as was shown by De Commer, Freslon and Yamashita in \cite{dCFY}. The main idea is to consider some $\ast$-representations of $C(SU_q(2))$ constructed by Voigt \cite{Vo11} in his study of the Baum-Connes conjecture for free orthogonal quantum groups. Let us outline the main ideas, referring the reader to \cite{dCFY, Vo11} for the details.
We begin by considering some representations of $\mathcal O(SU_q(2))$. Given $z \in \mathbb{C}$, we can define an algebra homomorphism $\pi_z:\mathcal O(SU_q(2)) \to \text{End}(\mathcal O(SU_q(2)))$ by
\[
\pi_z(x)y = x_{(1)}y (f_{1-z}\star S(x_{(2)})) \qquad (x,y \in \mathcal O(SU_q(2))),
\]
where $f_{1-z}: \mathcal O(SU_q(2)) \to \mathbb{C}$ is the Woronowicz character and $f_{1-z}\star x = (\iota \otimes f_{1-z}) \Delta(x)$.
Using $\pi_z$, we define a family of linear functionals \[\omega_z: \mathcal O(SU_q(2)) \to \mathbb{C}; \quad
\omega_z(x) = \varphi(\pi_z(x)1) = \langle \pi_z(x)1|1 \rangle_{L^2(SU_q(2))}. \]
A direct calculation (see \cite[Proposition 10]{dCFY}) shows that for any matrix element $u_{ij}^k \in \mathcal O(SU_q(2))$ of the $k$th irreducible representation of $SU_q(2)$, we have
\begin{align*}
\omega_z(u_{ij}^{k}) &= \varphi(\pi_z(u_{ij}^{k})1) = \varphi((u_{ij}^{k})_{(1)}(f_{1-z}\star S((u_{ij}^{k})_{(2)})) ) \\
&= \frac{\delta_{ij}\mu_k(|q|^{\bar z} +|q|^{-\bar z})}{\mu_{k}(|q| +|q|^{-1})}.
\end{align*}
In particular, when $z = t \in (-1,1)$, the functional $\omega_t$ has the desired form of \eqref{voigt-state} (just take $\psi_s = \omega_t$ with $s = |q|^t + |q|^{-t}$). Moreover, $\omega_t$ turns out to be a state in this case: In \cite[Section 4]{Vo11}, it is shown that the unital $\ast$-subalgebra
\[\mathcal O(SU_q(2)/\mathbb{T}) = \{x \in \mathcal O(SU_q(2)): (\iota \otimes \pi_{\mathbb{T}})\Delta (x) = x \otimes 1\}\]
of polynomial functions on the {\it standard Podles Sphere} $SU_q(2)/\mathbb{T}$ is invariant under each homomorphism $\pi_z$. (Here, $\pi_\mathbb{T}:C(SU_q(2)) \to C(\mathbb{T})$ is the canonical quotient map associated to the quantum subgroup $\mathbb{T}$). If $z = t \in (-1,1)$, \cite[Lemma 4.3]{Vo11} shows that we can moreover deform the $L^2(SU_q(2))$ inner product on $\mathcal O(SU_q(2)/ \mathbb{T})$ in a $t$-dependent way so that $\pi_t$ extends to $\ast$-representation of $C(SU_q(2))$ on $H_t$, the Hilbert space completion of $\mathcal O(SU_q(2)/\mathbb{T})$ with respect to this new inner product. From the structure of this deformed inner product, it follows (see \cite[Proposition]{dCFY}) that the following formula holds:
\[\omega_t(x) = \varphi(\pi_t(x)1) = \langle \pi_t(x)1|1 \rangle_{H_t} \qquad (x \in \mathcal O(SU_q(2))). \]
In particular, $\omega_t$ (and thus also $\psi_s$) is always a state. Conclusion: $\widehat{SU_q(2)}$ has the central Haagerup property.
To prove $Z\Lambda_{cb}(\widehat{SU_q(2)}) = 1$, the idea is exactly the same as in the unimodular case: we need to truncate our net of normalized completely positive definite functions coming from the central Haagerup property in a cb-norm controlled way. Now, if $q$ is such that $SU_q(2)$ happens to be monoidally equivalent to an $O^+_F$ for which Theorem \ref{freslon-cbnorm} can be applied, then our truncation argument in the unimodular case carries over verbatim. For the remaining range of values of $q$ (i.e., $|q| $ near $1$), we need another approach.
The alternate approach is in the spirit of the classical paper of Pytlik and Szwarc \cite{PySz86}, where one wants to show that the family of normalized completely positive definite functions $(b_t)_{t \in (0,1)} \subset ZM_{cb}A(\widehat{SU_q(2)})$ associated to the above states $\omega_t$ can be analytically continued to an analytic family \[\mathcal S \owns z \mapsto b_z \in ZM_{cb}A(\widehat{SU_q(2)}),\]
where $\mathcal S \subset \mathbb{C}$ is a domain containing $(-1,1) \subset \mathbb{R}$. In \cite[Theorem 11]{dCFY}, this is shown to be possible. It is moreover shown there that for $|z|$ sufficiently small, $b_z^3 \in A(\widehat{SU_q(2)})$. But since the family $\mathcal S \owns z \mapsto b_z^3$ is analytic, it follows that \[b_z^3 \in \overline{A(\widehat{SU_q(2)}) }^{\|\cdot\|_{M_{cb}A(\widehat{SU_q(2)} )}} \qquad (z \in \mathcal S).\]
Using this last fact, together with the fact that $b_t^3 \in ZM_{cb}A(\widehat{SU_q(2)}) $ is a normalized completely positive definite function with $\lim_{t \to 1}b_t^3 = 1$, we easily deduce that $Z\Lambda_{cb}(\widehat{SU_q(2)}) = 1$.
\bibliographystyle{plain}
|
\section{Generalized Gross-Neveu model: $ 2+\varepsilon$ expansion }
\label{sec:GN-ren}
The minimal action of the Weyl fermions in $d$ dimensions can be rewritten in Fourier
space as
\begin{eqnarray} \label{eq:action-b}
S &=&
\sum\limits_{\alpha=1}^N \int_{k} \bar{\psi}_{\alpha}(-\mathbf{k})( \bm{\gamma} \mathbf{k} -
i \omega ) \psi_{\alpha}(\mathbf{k}) \nonumber \\
&& - \sum\limits_{\alpha,\beta=1}^N \sum\limits_{n,\vec{A}}\frac{\Delta_n}2 \int_{k_i} \ [\bar{\psi}_\alpha(\mathbf{k}_1) \gamma^{(n)}_{\vec{A}} \psi_\alpha (\mathbf{k}_2)] \nonumber \\
&& \ \ \ \ \ \times [\bar{\psi}_\beta(\mathbf{k}_3) \gamma^{(n)}_{\vec{A}} \psi_\beta (-\mathbf{k}_1-\mathbf{k}_2-\mathbf{k}_3)].
\end{eqnarray}
One can build up a perturbation theory in small disorder calculating all
correlation and vertex functions perturbatively in $\Delta_n$.
Each term can be represented as a Feynman diagram. In these diagrams the solid lines
stand for the the bare propagator
\begin{eqnarray}
&& \langle \bar{\psi}_\alpha(\mathbf{k,\omega}) \psi_{\beta}(-\mathbf{k},-\omega) \rangle_0 = \delta_{\alpha\beta}
\frac{\bm{\gamma} \mathbf{k} + i \omega}{k^2+\omega^2},
\end{eqnarray}
and the dashed line corresponds to one of the vertex $\frac12 \Delta_n$.
Note that the dashed line transmit only momenta but not frequency.
These terms turn out to be diverging in $d=2$ which is the lower critical dimension of the transition.
Simple scaling analysis shows that weak disorder is irrelevant for $d>2$.
To make the theory finite we use the dimensional regularization and compute all integrals in $d=2+\varepsilon_2$.
At the end we put $\varepsilon_2=1$.
To render the divergences we employ the minimal subtraction scheme and collect all poles in $\varepsilon_2 $ in the Z-factors:
$Z_{\psi}$, $Z_\omega$ and $Z_n$ so that the correlation function calculated with the renormalized action
\begin{eqnarray} \label{eq:action-r}
&& S_R =
\sum\limits_{\alpha=1}^N \int_{k} \bar{\psi}_{\alpha}(-\mathbf{k})( Z_\psi \bm{\gamma} \mathbf{k} -
Z_\omega i \omega ) \psi_{\alpha}(\mathbf{k}) \nonumber \\
&& - \sum\limits_{\alpha,\beta=1}^N \sum\limits_{n,A} \frac{ \mu^{-\varepsilon }\Delta_n}{K_d} \int_{k_i} \ [\bar{\psi}_\alpha \gamma^{(n)}_{A} \psi_\alpha ] \cdot [\bar{\psi}_\beta \gamma^{(n)}_{A} \psi_\beta ] \ \ \ \ \\
\end{eqnarray}
remain finite in the limit
$\varepsilon_2 \to 0$.
Here we have introduced the renormalized fermionic fields $\psi$, $\bar{\psi}$ and
the renormalized dimensionless coupling constants $\Delta_n$ on the mass scale $\mu$, which are related to the parameters by
\begin{eqnarray} \label{eq:Z-factors}
&&\mathring{\psi} = Z_{\psi}^{1/2} \psi, \ \ \ \mathring{\bar{\psi}} = Z_{\psi}^{1/2} \bar{\psi},\\
&&\mathring{\omega} = Z_\omega Z_{\psi}^{-1}\omega, \ \ \ \
\mathring{\Delta}_n =
\frac{2 \mu^{-\varepsilon }}{K_d} \frac{Z_n}{Z_{\psi}^{2}} \Delta_n,
\end{eqnarray}
where $K_d = 2\pi^{d/2}/((2\pi)^d\Gamma(d/2))$ is the area of the $d$-dimensional unite sphere divided by $(2\pi)^d$.
The renormalized and the bare vertex and Green functions are related by
\begin{eqnarray}
&&\mathring{\Gamma}^{(\mathcal{N})}(p_i, \mathring{\omega},\mathring{\Delta}) = Z_{\psi}^{-\mathcal{N}/2}{\Gamma}^{(n)}(p_i, \omega, \Delta, \mu),
\label{eq-gam1} \\ \label{eq-gam2}
&& \mathring{G}^{(\mathcal{N})}(p_i, \mathring{\omega},\mathring{\Delta}) = Z_{\psi}^{\mathcal{N}/2}G^{(n)}(p_i, \omega, \Delta, \mu),
\end{eqnarray}
where $\Delta$ stands for all $\Delta_n$.
Using that the bare functions $\mathring{\Gamma}^{(\mathcal{N})}$ and $\mathring{G}^{(n)}$ do not depend on the renormalization scale
$\mu$ we take the derivative of Eqs.~(\ref{eq-gam1}) and (\ref{eq-gam2}) with respect to $\mu$ and obtain the RG flow
equations for the renormalized Green and vertex functions:
\begin{eqnarray}
&&\left[\mu\frac{\partial}{\partial \mu} - \sum\limits_n \beta_n(\Delta)
\frac{\partial}{\partial \Delta_n} - \frac{\mathcal{N}}2 \eta_\psi(\Delta) \right. \nonumber \\
&& \ \ \ \ \ \ \ \ \ \left.
- \gamma (\Delta) \omega \frac{\partial}{\partial \omega}\right] {\Gamma}^{(n)}(p_i, \omega, \Delta)=0, \label{eq-rg1-1} \\
&&\left[\mu\frac{\partial}{\partial \mu} - \sum\limits_n \beta_n(\Delta)
\frac{\partial}{\partial \Delta} + \frac{\mathcal{N}}2 \eta_\psi(\Delta) \right. \nonumber \\
&& \ \ \ \ \ \ \ \ \ \left.
- \gamma (\Delta) \omega \frac{\partial}{\partial \omega}\right] G^{(n)}(p_i, \omega, \Delta)=0. \label{eq-rg1-2}
\end{eqnarray}
Here we have defined the scaling functions
\begin{eqnarray}
&&\beta_n(\Delta)= - \left.\mu\frac{\partial \Delta_n}{\partial \mu} \right|_{\mathring{\Delta}}, \label{eq-def-beta-fun} \\
&&\eta_\psi(\Delta)= - \sum\limits_n\beta_n(\Delta)\frac{\partial \ln Z_\psi}{\partial \Delta_n}, \\
&&\eta_\omega(\Delta)= - \sum\limits_n \beta_n(\Delta)\frac{\partial \ln Z_\omega}{\partial \Delta_n}, \\
&& \gamma(\Delta) = \eta_\omega(\Delta)- \eta_\psi(\Delta).
\end{eqnarray}
Dimensional analysis gives the following rescaling formulas
\begin{eqnarray}
\Gamma^{(\mathcal{N})}(p_i, \omega, \Delta, \mu) &=&\lambda^{-d+\mathcal{N}(d-1)/2} \nonumber \\
&& \times {\Gamma}^{(\mathcal{N})}(\lambda p_i, \lambda \omega, \Delta, \lambda \mu), \\
G^{(\mathcal{N})}(p_i, \omega, \Delta, \mu) &=&\lambda^{d(\mathcal{N}-1)-\mathcal{N}(d-1)/2} \nonumber \\
&& \times G^{(\mathcal{N})}(\lambda p_i, \lambda \omega, \Delta, \lambda \mu),
\end{eqnarray}
which can be rewritten in an infinitesimal form as
\begin{eqnarray}
&&\left[\mu\frac{\partial}{\partial \mu} + \sum\limits_i p_i \frac{\partial}{\partial p_i} + \omega \frac{\partial}{\partial \omega} \right. \nonumber \\
&&\ \ \ \ \ \ \left.
-d+\frac{\mathcal{N}(d-1)}2 \right] {\Gamma}^{(\mathcal{N})}(p_i, \omega, \Delta)=0, \label{eq-rg2-1}\\
&&\left[\mu\frac{\partial}{\partial \mu} + \sum\limits_i p_i \frac{\partial}{\partial p_i} + \omega \frac{\partial}{\partial \omega} \right. \nonumber \\
&&\ \ \ \ \ \ \left.
+d(\mathcal{N}-1)-\frac{\mathcal{N}(d-1)}2 \right] G^{(\mathcal{N})}(p_i, \omega, \Delta)=0. \nonumber \\ \label{eq-rg2-2}
\end{eqnarray}
Subtracting Eqs.~(\ref{eq-rg1-1}) and (\ref{eq-rg1-2}) from Eqs.~(\ref{eq-rg2-1}) and (\ref{eq-rg2-2}) we obtain
\begin{eqnarray}
&& \!\!\!\!\!\!\!\!\! \left[ \sum\limits_n \beta_n(\Delta)
\frac{\partial}{\partial \Delta_n}+\sum\limits_i p_i \frac{\partial}{\partial p_i} + (1+\gamma (\Delta)) \omega \frac{\partial}{\partial \omega} \right. \nonumber \\
&& \ \left.
-d + \frac{\mathcal{N}}2 \left[ d-1 + \eta_\psi(\Delta) \right] \right] {\Gamma}^{(\mathcal{N})}(p_i, \omega, \Delta)=0, \nonumber \\ \label{eq-rg3-1} \\
&& \!\!\!\!\!\!\!\!\! \left[\sum\limits_n \beta_n(\Delta)
\frac{\partial}{\partial \Delta_n} +\sum\limits_i p_i \frac{\partial}{\partial p_i} + (1+\gamma (\Delta))\omega \frac{\partial}{\partial \omega}
\right. \nonumber \\
&& \!\!\!\!\!\!\!\!\! \left. + d(\mathcal{N}-1) - \frac{\mathcal{N}}2 \left[ d-1 + \eta_\psi(\Delta) \right]
\right] G^{(\mathcal{N})}(p_i, \omega, \Delta)=0. \nonumber \\ \label{eq-rg3-2}
\end{eqnarray}
The solutions of Eqs.~(\ref{eq-rg3-1}) and (\ref{eq-rg3-2}) can be found by using the method of characteristics.
The characteristics, i.e. lines in the space of $p_i$, $\omega$, and $\Delta_n$, parameterized by auxiliary parameter $\xi$ which below will be identified with the correlation length, can be found from the equations
\begin{eqnarray}
&& \frac{d p_i (\xi)}{d \ln \xi} = p_i(\xi), \\
&& \frac{d \Delta_n (\xi)}{d \ln \xi} = \beta_n(\Delta(\xi)), \\
&& \frac{d \omega (\xi)}{d \ln \xi} = [1+\gamma(\Delta(\xi))] \omega(\xi),
\end{eqnarray}
with initial conditions $\Delta_n (1)=\Delta_n$, $p_i(1)=p_i$, and $\omega (1)=\omega$.
The solution of Eqs.~(\ref{eq-rg3-1}) and (\ref{eq-rg3-2}) then propagate along the characteristics
according to the equations
\begin{eqnarray}
&& \frac{d M_\mathcal{N}(\xi)}{d \ln \xi} = [-d + \frac{\mathcal{N}}2 (d-1 + \eta_\psi(\Delta(\xi)))] M_\mathcal{N}(\xi), \nonumber
\\
&& \frac{d H_\mathcal{N}(\xi)}{d \ln \xi} = \left[d(\mathcal{N}-1) - \frac{\mathcal{N}}2 ( d-1 + \eta_\psi(\Delta(\xi)) )\right] H_\mathcal{N}(\xi), \nonumber \\
\end{eqnarray}
with the initial conditions $M_n(1)=H_n(1)=1$. Thus the solutions of Eqs.~(\ref{eq-rg3-1}) and (\ref{eq-rg3-2})
satisfy
\begin{eqnarray}
&& {\Gamma}^{(\mathcal{N})}(p_i, \omega, \Delta)= M_\mathcal{N}(\xi) {\Gamma}^{(\mathcal{N})}( p_i(\xi),\omega(\xi), \Delta(\xi) ), \ \ \ \ \label{eq-rg4-1} \\
&& G^{(\mathcal{N})}(p_i, \omega, \Delta) = H_\mathcal{N}(\xi) G^{(\mathcal{N})}( p_i(\xi),\omega(\xi), \Delta(\xi) ). \ \ \ \ \ \label{eq-rg4-2}
\end{eqnarray}
We assume that the $\beta$-function have a fixed point (FP)
\begin{eqnarray}
\beta(\Delta^*)=0, \label{eq:fixpoint0}
\end{eqnarray}
with a single unstable direction $\delta = \Delta - \Delta^*$, \textit{i.e.} the stability matrix
\begin{equation}\label{smatrix}
\mathcal{M}_{nm}= \left. \frac{\partial \beta_n(\Delta)}{\partial \Delta_m}\right|_{\Delta^*},
\end{equation}
has only one positive eigenvalue $\lambda_1^{(+)}$ associated with the direction $\delta$.
Then the solutions~(\ref{eq-rg4-1}) and (\ref{eq-rg4-2}) in the vicinity of the
FP~(\ref{eq:fixpoint0}) can be rewritten as
\begin{eqnarray}
&&\!\!\!\!\!\!\! {\Gamma}^{(\mathcal{N})}(p_i, \omega,\delta)= \xi^{\mathcal{N} d_\psi - d } f_\mathcal{N}( p_i \xi,\omega \xi^z, \delta \xi^{1/\nu}), \ \ \ \\
&& \!\!\!\!\!\!\! G^{(\mathcal{N})}(p_i, \omega,\delta) = \xi^{d(\mathcal{N}-1) - n d_\psi} g_\mathcal{N}( p_i \xi ,\omega \xi^z, \delta \xi^{1/\nu}), \ \ \
\end{eqnarray}
where we defined the critical exponents $\nu$, $z$, $d_\psi$. The parameter
$\xi$ can be identified with the correlation length that gives the critical exponent
for the correlation length
\begin{eqnarray}
\xi \sim \delta^{-\nu}, \ \ \ \ \frac1{\nu}=\lambda_1^{(+)},
\end{eqnarray}
and the dynamic dynamic critical exponent
\begin{eqnarray}
\omega \sim k^z, \ \ \ \ \ z = 1+\gamma(\Delta^*).
\end{eqnarray}
The anomalous dimension of the fields $\psi$ and $\bar{\psi}$ reads
\begin{eqnarray}
d_\psi = \frac12[d-1+\eta_\psi(\Delta^*)].
\end{eqnarray}
Note, that the exponent $\eta_\psi$ characterizes the scaling behavior of the two-point function
\begin{eqnarray}
G^{(2)}(p) = \overline{\left\langle\bar{\psi}(p) \psi(-p)\right\rangle} \sim p^{-1+\eta_\psi(\Delta^*)},
\end{eqnarray}
which can be viewed as the momentum distribution of fermions at the transition.
\subsection{Critical exponents to three-loop order}
To renormalize the theory we use the minimal substraction scheme
\begin{eqnarray}
&& Z_{\psi} \mathring{\Gamma}^{(2)}(p,\omega=Z_\omega Z_{\psi}^{-1}\mu,\mathring{\Delta}(\Delta)) = \mathrm{finite}, \ \ \ \ \\
&& Z^2_{\psi} \mathring{\Gamma}^{(4)}_n(p_i=0, \omega=Z_\omega Z_{\psi}^{-1} \mu,\mathring{\Delta}(\Delta)) = \mathrm{finite}, \ \ \
\end{eqnarray}
where $\mathring{\Delta}(\Delta)$ is given by Eq.~(\ref{eq:Z-factors}) and $\mathring{\Gamma}^{(4)}_n$ is the renormalized vertex $V_n$.
The three-loop corrections to the vertex $V_0$ have been many times discussed in the literature in the context of the GN model~\cite{Zinn-Justin:1986SM}. The corresponding $\beta_0$-function defined in Eq.~(\ref{eq-def-beta-fun}) reads
\begin{eqnarray}
\beta_0 &=& - \varepsilon_2 \Delta_0 - 2 \Delta_0^2 (N-2) - 4 \Delta_0^3 (N-2) \nonumber \\
&& + 2 \Delta_0^4 (N-2) (N-7), \label{eq-sm-beta1}
\end{eqnarray}
where we kept the dependence on $N$.
The 24 diagrams derived from the diagram (b) shown in Fig.~\ref{fig:diagrams-GN} by permutation of the dashed line ends which were
neglected in Ref.~\cite{Roy:2016SM} generate the vertex $V_3$ \cite{Vasilev:1997SM}.
Other diagrams which one has to take into account in calculation to order of $\varepsilon_2^3$ are the diagrams (c)-(e) shown in Fig.~\ref{fig:diagrams-GN}.
These diagrams with lines corresponding to $V_0$ and $V_3$ contribute to $V_4$ and with lines corresponding to $V_0$ and $V_3$
contribute to $V_4$. Since the contributions of the diagrams (b) are of order $\Delta_0^4$ one may naively conclude that while
$\Delta_0$ is of order $\varepsilon_2$, the two over vertices $\Delta_3$ and $\Delta_4$ are of order $\varepsilon_2^3$.
Indeed, the corresponding $\beta$-functions
\begin{eqnarray}
\beta_3 &=&-\varepsilon_2 \Delta_3 + a \Delta_0^4 +16 \Delta_0 \Delta_4 + 8 \Delta_0\Delta_3. \label{eq-sm-beta2} \\
\beta_4 &=& -\varepsilon_2 \Delta_4 -4 \Delta_0\Delta_3 -12 \Delta_0 \Delta_4, \label{eq-sm-beta3}
\end{eqnarray}
have the fixed point
\begin{eqnarray}
\Delta_0^*&=&\frac{\varepsilon_2 }{4-N} -\frac{\varepsilon_2 ^2}{2(2-N)^2} + \frac{(1+N)\varepsilon_2 ^3}{8(2-N)^3} + O(\varepsilon_2^4), \ \ \ \label{eq-sm-fp1-n} \\
\Delta_3^*&=&\frac{a \varepsilon_2 ^3 (N-8)}{16 N (N-6) (N-2)^3}+O\left(\varepsilon_2 ^4\right), \label{eq-sm-fp2-n} \\
\Delta_4^*&=& \frac{a \varepsilon_2 ^3}{8 N (N-6) (N-2)^3}+O\left(\varepsilon_2 ^4\right) \label{eq-sm-fp3-n},
\end{eqnarray}
which has non analytic behavior in the limit $N \to 0$. Taking first the limit $N \to 0$ in the $\beta$-functions
one finds the fixed point
\begin{eqnarray}
\Delta_0^*&=&\frac{\varepsilon_2 }{4} -\frac{\varepsilon_2 ^2}{8} + \frac{\varepsilon_2 ^3}{64} + O(\varepsilon_2^4), \ \ \ \label{eq-sm-fp1} \\
\Delta_3^*&=&\frac{a \varepsilon_2 ^2}{96}-\frac{23 a \varepsilon_2 ^3}{1152}+O\left(\varepsilon_2 ^4\right), \label{eq-sm-fp2} \\
\Delta_4^*&=& -\frac{a \varepsilon_2 ^2}{384}+\frac{49 a \varepsilon_2 ^3}{9216}+O\left(\varepsilon_2 ^4\right) \label{eq-sm-fp3},
\end{eqnarray}
similar to $\sqrt{\varepsilon}$ expansion for the diluted Ising model~\cite{Shalaev:1997SM}.
The stability of the FP can be described by the eigenvalues of the stability matrix $\frac{\partial \beta_i}{\partial \Delta_j}|_{\Delta^*},~i,j \in {0,3,4}$. Since one expects that the transition is controlled by an unstable IR FP, the stability matrix is expected
to have only one positive eigenvalue which is related to the critical exponent $1/\nu = \lambda_{1}^{(+)}$.
The stability eigenvalues read:
\begin{eqnarray}
&&\frac1{\nu} = \lambda_{1}^{(+)} = \varepsilon_4 +\frac{\varepsilon_4 ^2}{2}+\frac{3 \varepsilon_4 ^3}{8}+O\left(\varepsilon_4
^4\right), \label{eq:2-eps-eigenvalues3loop} \\
&& \lambda_{2}^{(-)} = -3 \varepsilon_4 +\varepsilon_4 ^2-\frac{\varepsilon_4 ^3}{8}+O\left(\varepsilon_4
^4\right), \\
&& \lambda_{3}^{(-)} = -\frac{\varepsilon_4 ^2}{2}+\frac{\varepsilon_4 ^3}{16}+O\left(\varepsilon_4
^4\right).
\end{eqnarray}
Only the first eigenvalue~\eqref{eq:2-eps-eigenvalues3loop} associated with a single instability direction is positive.
\begin{figure
\includegraphics[width=6.5cm]{diagrams-GN}
\caption{Diagrams entering the renormalization of the generalized GN action.
Solid lines stands for fermionic propagators and dashed lines for disorder vertices. }
\label{fig:diagrams-GN}
\end{figure}
The generation of vertices $\Delta_3$ and $\Delta_4$ at three-loop order might renormalize $\omega$ and thus give a correction to the
other critical exponents via diagrams of the type of diagram (a) of Fig.~\ref{fig:diagrams-GN}. The
combinatorial factor associated to this diagram is 2. The contribution will take the form ($n=3,4$):
\begin{align}
I_n & = \sum_{A=\{i_1,..,i_n\}} \gamma^{(n)}_A \int_{\bf k} \frac{\bm{\gamma} \mathbf{k} + i \omega}{k^2+\omega^2} \gamma^{(n)}_A \nonumber \\
& = \sum_{A=\{i_1,..,i_n\}} \gamma^{(n)}_A \gamma^{(n)}_A \int_{\bf k} \frac{i \omega}{k^2+\omega^2},
\label{eq:3loop-zdiag}
\end{align}
besides
\begin{align}
\gamma^{(n)}_A & = As[\gamma_{i_1}....\gamma_{i_n}] \nonumber \\
& = \epsilon^{i_1..i_n} \gamma_{i_1}....\gamma_{i_n},
\end{align}
where the set of indices is set and $\epsilon^{i_1..i_n}$ is the corresponding element of the n-th Levi Civita tensor. Therefore (no contraction on A is implied here):
\begin{align}
(\gamma^{(n)}_A)^2 & = (\epsilon^{i_1..i_n})^2 \gamma_{i_1}....\gamma_{i_n} \gamma_{i_1}....\gamma_{i_n} \nonumber \\
& = \gamma_{i_1}....\gamma_{i_n} \gamma_{i_1}....\gamma_{i_n} \nonumber \\
& = (-1)^{n-1} \gamma_{i_2}....\gamma_{i_n} (\gamma_{i_1})^2 \gamma_{i_2}....\gamma_{i_n} \nonumber \\
& = (-1)^{(n-1)!} \mathbb{I}
\end{align}
using the anticommutation relation $\{\gamma_\mu,\gamma_\nu\} = 2\delta_{\mu \nu} \mathbb{I}$ and assuming all indices $ i_1,...,i_n $ are distinct (otherwise, $\gamma^{(n)}_A$ vanishes trivially.) Note that since $n=3$ or $4$, $(-1)^{(n-1)!} = 1$.
Performing the sum in \eqref{eq:3loop-zdiag} thus yields:
\begin{align}
I_n & = \binom{d}{n} \int_{\bf k} \frac{i \omega}{k^2+\omega^2} \propto (d-2) \int \frac{d^d{\bf k}}{(2\pi)^d} \frac{i \omega}{k^2+\omega^2}\nonumber \\
\end{align}
For $n=3,4$, $\binom{d}{n} \propto (d-2) = O (\varepsilon)$, and thus the binomial coefficient cancels the pole in the integral, making the contribution $I_n = O(1)$ finite. At the end of the day we find that this diagram will give no contribution to the frequency renormalisation
and a fortiori to the $z$ exponent. Thus the critical exponents to three-loop order are given by
\begin{eqnarray}
z &=& 1+ \frac{\varepsilon_2}{2} -\frac{\varepsilon_2^2}{8} + \frac{3\varepsilon_2^3}{32} + O(\varepsilon_2^4) , \label{eq:exp-z}\\
\eta &=& -\frac{\varepsilon_2^2}{8} + \frac{3\varepsilon_2^3}{16} - \frac{25\varepsilon_2^4}{128} + O(\varepsilon_2^5). \\
d_\psi &=& \frac12[d-1+\eta_\psi] \nonumber \\
& = & \frac12 + \frac{\varepsilon}{2} -\frac{\varepsilon^2}{16} + \frac{3\varepsilon_2^3}{32} - \frac{25\varepsilon_2^4}{256} + O(\varepsilon_2^5).
\end{eqnarray}
To estimate numerical values of the exponents in $d=3$ we use direct evaluation at $\varepsilon_2=1$ (D), Pad\'{e} approximant P$[M/L]$
and Pad\'{e}-Borel resummation PB$[M/L]$.
We find $z=1.469$ (D), $z=1.429$ (P[2/1]) and $z=1.425$ (PB[2/1]);
$\eta=0.0625$ (D) to three loop and $\eta=-0.133$ (D) to four loop.
\subsection{Renormalization of composite operators}
We now discuss the renormalization of the composite operators
\begin{eqnarray}
\mathcal{O}_q (r) := \left(\bar{\psi}_\alpha(r){\psi}_\alpha(r)\right)^q. \label{eq:comp-oper}
\end{eqnarray}
which are related to the the deviation of the disorder distribution from the Gaussian
distribution. The bare scaling dimension of operators~(\ref{eq:comp-oper}) is $[\mathcal{O}_q ]=(d-1)q + O(\Delta)$.
To find their scaling dimension in the GN FP we introduce the $Z$-factors
\begin{eqnarray}
\mathring{\mathcal{O}}_q = \mathcal{Z}_{q} Z_\psi^{-q}{\mathcal{O}}_q. \label{eq:comp-oper-ren}
\end{eqnarray}
which has to render the divergence of the correlation functions involving operators~(\ref{eq:comp-oper}) .
To one loop order the diagrams contributing to the $Z_q$ factor are shown in Fig.~\ref{fig:diagrams4-composite}.
We find to one-loop order
\begin{eqnarray}
\mathcal{Z}_{q} = 1 + 2[q+q(q-1)]\frac{\Delta_0}{\varepsilon_2}. \label{eq:comp-oper-ren-Z}
\end{eqnarray}
The corresponding scaling function
\begin{eqnarray}
&&\eta_q(\Delta)= - \sum\limits_n \beta_n(\Delta)\frac{\partial \ln Z_q}{\partial \Delta_n},
\end{eqnarray}
gives the scaling dimension of the composite operators~(\ref{eq:comp-oper})
\begin{eqnarray}
&&[\mathcal{O}_q ]=(d-1+\eta_\psi)q - \eta_q(\Delta^*).
\end{eqnarray}
To one loop order this yields
\begin{eqnarray}
[\mathcal{O}_q] = (1+\varepsilon_2)q - \frac12 q^2 \varepsilon_2 +O(\varepsilon_2^2),
\end{eqnarray}
that is consistent with the conformal theory results of~\cite{Ghosh:2016SM}.
\begin{figure
\includegraphics[width=8cm]{composite-operator}
\caption{Diagrams renormalizing the composite operators (\ref{eq:comp-oper}). (a) is the bare vertex corresponding to a composite operator (\ref{eq:comp-oper}), (a) - (e) are the one-loop diagrams contributing to renormalization (\ref{eq:comp-oper-ren}): (b) $= q I_1$,
(c)$ = q(q-1)I_1$, (d)+(e) $= 0$, where $I_1$ is the one-loop integral. }
\label{fig:diagrams4-composite}
\end{figure}
In order to calculate the scaling behavior of the local DOS $\rho(\omega,\delta)$ it is enough to consider renormalization
of the composite operator $\mathcal{O}_1$. The corresponding $Z$ factor is not independent
and is related to $Z_\omega$ by
\begin{eqnarray}
\mathring{\mathcal{O}}_1 = Z_\omega Z_\psi^{-1} \mathcal{O}_1.
\end{eqnarray}
We can write the flow equation for the local DOS as
\begin{eqnarray}
&&\left[ \sum\limits_n \beta_n(\Delta)
\frac{\partial}{\partial \Delta_n} + (1+\gamma (\Delta)) \omega \frac{\partial}{\partial \omega} \right. \nonumber \\
&& \ \ \ \left.
- (d-1) + \eta_\omega(\Delta) - \eta_\psi(\Delta) \right]\rho(\omega,\Delta)=0. \label{eq:DOS1}
\end{eqnarray}
The solution of Eq.~(\ref{eq:DOS1}) in the vicinity of the FP~(\ref{eq:fixpoint0}) has the form
\begin{eqnarray}
&& \rho( \omega) = \xi^{z-d} \rho_0(\omega \xi^z, \delta \xi^{1/\nu}), \ \ \
\end{eqnarray}
with $z=[\mathcal{O}_1]$ given to three-loop order by (\ref{eq:exp-z}).
\section{Gross-Neveu-Yukawa Model: $4-\varepsilon$ expansion }
\label{sec:gny}
\subsection{Model}
The action for the $U(N)$ GNY model is given by
\begin{align}
S_{GNY} &= \int d^d r [ - {\bar \chi}_\alpha ( {\bm \gamma} \cdot {\bm \nabla} + \sqrt{g} \phi ) \chi_\alpha \nonumber \\
&+ \frac12 (\nabla \phi)^2 + \frac12 \mu \phi^2 + \frac{\lambda}{4\text{!}} \phi^4 ].
\label{eq:action-GNY-SM}
\end{align}
We are interested in the $N \rightarrow 0$ limit. In Fourier space ($- i\bm{\gamma} \cdot \bm{\nabla} \to \bm{\gamma} \cdot \mathbf{k}$) , the bare fermionic and bosonic propagators read
\begin{subequations}
\begin{align}
\langle \chi_\alpha ({\bf k}) \chi_\alpha (- {\bf k'}) \rangle = i \frac{ {\bm \gamma} \cdot {\bf k}}{k^2} \\
\langle \phi ({\bf q}) \phi(- {\bf q}) \rangle = \frac{1}{q^2 + \mu} .
\end{align}
\end{subequations}
\subsection{Renormalization}
\begin{figure}[htbp]
\includegraphics[width=7.5cm]{fig-diagrammes4-eps}
\caption{Diagrams entering the renormalization of the GNY action to one-loop order. Solid lines stand for the fermionic propagator
and dashed lines for the bosonic one. Diagrams from $\phi^4$ theory not shown.}
\label{fig:diagrams4-eps}
\end{figure}
We perform a perturbative expansion of correlation and vertex functions in the disorder parameters $g$ and $\lambda$.
Fig. \ref{fig:diagrams4-eps} shows the diverging diagrams in $d=4$ that involve fermionic-bosonic vertices. Other diverging diagrams come from the expansion in $\lambda$ and are known from the $\phi^4$ theory.
We use dimensional regularization in $d=4-\varepsilon_4$ and put $\varepsilon_4 = 1$ at the end of the day.
Following the minimal substraction scheme, we introduce the renormalization constants
$Z_\chi$, $Z_\phi$, $Z_\mu$, $Z_g$ and $Z_\lambda$. Calling $\Lambda$ the renormalization scale, the renormalized action reads:
\begin{align}
S &= \int d^d r [ - {\bar \chi}_\alpha (Z_\chi {\bm \gamma} \cdot {\bm \nabla} + \Lambda^{\varepsilon_4/2} \sqrt{g Z_g} \phi ) \chi_\alpha \nonumber \\
&+ \frac12 Z_\phi (\nabla \phi)^2 + \frac12 [\mu_c Z_\phi + \Lambda^2 Z_\mu ~ \delta \mu] \phi^2 + \Lambda^\varepsilon_4 Z_\lambda \frac{\lambda}{4\text{!}} \phi^4 ].
\end{align}
The renormalized fields are related to the bare ones through $\mathring{\chi} = Z_\chi^{1/2} \chi$, and $\mathring{\phi} = Z_\phi^{1/2} \phi$.
Similarly, we define the renormalized bosonic mass $\mathring{\mu} = \mu_c + \Lambda^2 Z_\mu Z_\phi^{-1} ~ \delta \mu$. The relations between bare and renormalized couplings read
$\mathring{g} = \Lambda^{\varepsilon_4} Z_g Z_\chi^{-2} Z_\phi^{-1} g $ and $\mathring{\lambda} = \Lambda^{\varepsilon_4} Z_\lambda Z_\phi^{-2} \lambda$, where we have introduced the renormalization scale $\Lambda$ to render the renormalized couplings dimensionless.
The bare and renormalized correlation and vertex functions are related as follows:
\begin{subequations}
\begin{align}
\mathring{\Gamma}^{(n,l)} & (p_i, q_j, \mathring{\mu}-\mu_c, \mathring{g} , \mathring{\lambda}) = \nonumber \\
& Z_\chi^{-n/2} Z_\phi^{-l/2} \Gamma^{(n,l)} (p_i, q_j, \delta\mu, g, \lambda, \Lambda ), \label{eq:sm-ren-gamma}\\
\mathring{G}^{(n,l)} & (p_i, q_j, \mathring{\mu}-\mu_c, \mathring{g} , \mathring{\lambda}) = \nonumber \\
& Z_\chi^{n/2} Z_\phi^{l/2} G^{(n,l)} (p_i, q_j, \delta \mu, g, \lambda, \Lambda ).
\end{align}
\end{subequations}
From Eq.~(\ref{eq:sm-ren-gamma}) we derive the RG flow equation for the vertex functions:
\begin{align}
& \left[ \Lambda \frac{\partial}{\partial \Lambda} - \beta_{g} \frac{\partial}{\partial g} - \beta_{\lambda} \frac{\partial}{\partial \lambda} - \frac n 2 \eta_\chi - \frac l 2 \eta_\phi \right. \nonumber \\
& \ \ \ \ \ \left. - \gamma_\mu \delta\mu \frac{\partial}{\partial \delta \mu} \right] \Gamma^{(n,l)} (p_i, q_j, \delta \mu, g, \lambda, \Lambda ) =0,
\label{eq:4-eps_rgeq-a}
\end{align}
with the scaling functions:
\begin{subequations}
\begin{align}
\beta_{g}(g,\lambda) & = - \left.\Lambda \frac{\partial g}{\partial \Lambda}\right|_{\mathring{g},\mathring{\lambda}}, \\
\beta_\lambda (g,\lambda) & = - \left. \Lambda \frac{\partial \lambda}{\partial \Lambda}\right|_{\mathring{g},\mathring{\lambda}}, \\
\eta_\chi(g,\lambda) & = - \sum_{u = \lambda,g} \beta_{u} \frac{\partial \ln Z_\chi}{\partial u}, \\
\eta_\phi(g,\lambda) & = - \sum_{u = \lambda,g} \beta_{u} \frac{\partial \ln Z_\phi}{\partial u}, \\
\eta_\mu(g,\lambda) & = - \sum_{u = \lambda,g} \beta_{u} \frac{ \partial \ln Z_\mu}{\partial u}, \\
\gamma_\mu (g,\lambda) & = 2 + \eta_\mu - \eta_\phi.
\end{align}
\end{subequations}
Besides, dimensional analysis gives
\begin{subequations}
\begin{align}
& \Gamma^{(n,l)} (p_i, q_j, \delta \mu, g , \lambda, \Lambda) = X^{- d + n (d-1)/2 + l (d-2)/2 } \nonumber \\
& \hspace{20mm} \times \Gamma^{(n,l)} (X p_i, X q_j, \delta \mu, g, \lambda, X \Lambda ), \label{eq:sp:gamma-dim-analys}\\
& G^{(n,l)} (p_i, q_j, \delta \mu, g , \lambda, \Lambda) = X^{(n + l -1) d - n(d-1)/2 - l(d-2)/2} \nonumber \\
& \hspace{20mm} \times G^{(n,l)} (X p_i, X q_j, \delta \mu, g, \lambda, X \Lambda ).
\end{align}
\end{subequations}
We rewrite the relation (\ref{eq:sp:gamma-dim-analys}) in an infinitesimal form as
\begin{align}
& \left[ \Lambda \frac{\partial}{\partial \Lambda} + p_i \frac{\partial}{\partial p_i} + q_j \frac{\partial}{\partial q_j} - d + \frac{n(d-1)}2 \right. \nonumber \\
& \ \ \ \ \ \ \ \ \left. + \frac{l(d-2)}2 \right] \Gamma^{(n,l)} (p_i, q_j, \delta \mu, g, \lambda, \Lambda ),
\label{eq:4-eps_rgeq-b}
\end{align}
Subtracting \eqref{eq:4-eps_rgeq-a} from \eqref{eq:4-eps_rgeq-b} to get rid of the derivative with respect to $\Lambda$ we obtain
\begin{eqnarray}
&& \!\!\!\!\!\!\!\!\!\!\!\!
\left[ \beta_{g} \frac{\partial}{\partial g} + \beta_{\lambda} \frac{\partial}{\partial \lambda} + p_i \frac{\partial}{\partial p_i} + q_j \frac{\partial}{\partial q_j} \right. \nonumber \\
\nonumber \\
&& + \gamma_\mu \delta\mu \frac{\partial}{\partial \delta \mu} - d + \frac{n}2 ( \eta_\chi + d - 1) \nonumber \\
&& \left.
+ \frac{l}2 ( \eta_\phi + d - 2) \right] \Gamma^{(n,l)} (p_i, q_j, \delta \mu, g, \lambda, \Lambda ) =0.
\label{eq:4-eps_rgeq-c}
\end{eqnarray}
The solutions to Eq.~\eqref{eq:4-eps_rgeq-c} can be found using the method of characteristics. These solutions propagate along
specific lines in the space of $p_i$, $q_j$, $\delta \mu$, $g$ and $\lambda$ called the characteristics. The characteristics are parametrized by
an auxiliary parameter $L$, which can be identified with a length scale; they are determined by the following set of RG flow equations:
\begin{subequations}
\begin{align}
\frac{d p_i(L)}{d \ln L} & = p_i(L), \\
\frac{d q_i(L)}{d \ln L} & = q_i(L), \\
\frac{d \delta\mu(L)}{d \ln L} & = \gamma_\mu \delta\mu(L), \\
\frac{d g (L)}{d \ln L} & = \beta_{g}(g(L)) ,\\
\frac{d \lambda (L)}{d \ln L} & = \beta_\lambda (\lambda(L)),
\end{align}
\label{eq:RG-4-eps}
\end{subequations}
with initial conditions $p_i(1) = p_i$, $q_j(1) = q_j$, $\delta\mu(1) = \delta\mu$, $g (1) = g$, $\lambda (1) = \lambda$.
Thus the solutions of \eqref{eq:4-eps_rgeq-c} satisfy
\begin{align}
\Gamma^{(n,l)} & (p_i,q_i,\delta \mu,g,\lambda) = \nonumber \\
& \mathcal{M}(L) \Gamma^{(n,l)} (p_i(L), q_i(L),\delta \mu(L),g(L),\lambda(L)).
\label{eq:4-eps_scale}
\end{align}
with
\begin{equation}
\frac{d \ln \mathcal{M}_{n,l}(L)}{d \ln L} = \frac{n}2 (\eta_\chi + d - 1) + \frac{l}2 ( \eta_\phi + d - 2) - d .
\end{equation}
In the vicinity of the critical point the RG flow parameter $L$ can be identified with the correlation length $\xi$ in \eqref{eq:RG-4-eps}, allowing one to calculate the critical exponents from the RG equations.
\subsection{Critical exponents}
Calculation of the one-loop diagrams shown in Fig.~\ref{fig:diagrams4-eps} in the limit $N \rightarrow 0$
gives~\cite{Zinn-Justin:1986SM}:
\begin{subequations}
\begin{align}
\Gamma^{(2,0)} & = \langle \bar{\chi} ({\bf k}) \chi(- {\bf k}) \rangle^{-1}
= i {\bm \gamma}{\bm p} Z_\chi + i {\bm \gamma}{\bm p} \frac{K_d}2\frac{g}{\varepsilon_4}, \\
\Gamma^{(2,1)} & = \langle \bar{\chi}({\bf k}_1) \chi({\bf k}_2) \phi({\bf p}) \rangle_{\mathrm{1PI}} = \sqrt{g Z_g} - g^{3/2} \frac{K_d}{\varepsilon_4}, \\
\Gamma^{(0,2)} & = \langle \phi({\bf k }) \phi(-{\bf k}) \rangle^{-1} = Z_\phi k^2 + Z_{\mu} \delta\mu - \frac{\lambda}2 \delta\mu \frac{K_d}{\varepsilon_4}, \\
\Gamma^{(0,4)} & = \langle \phi({\bf k }_1) \phi({\bf k}_2) \phi({\bf k}_3) \phi({\bf k}_4) \rangle_{\mathrm{1PI}} = Z_\lambda \lambda - \frac32 \lambda^2 \frac{K_d}{\varepsilon_4}.
\end{align}
\end{subequations}
To make these functions finite, we define the renormalization constants as follows:
\begin{subequations}
\begin{align}
Z_\chi & = 1 - \frac12 g \frac{K_d}{\varepsilon_4}, \\
Z_\phi & = 1, \\
Z_{\mu} & = 1 + \frac{\lambda}{2} \frac{K_d}{\varepsilon_4}, \\
Z_g & = 1 + 2 g \frac{K_d}{\varepsilon_4}, \\
Z_\lambda & = 1 + \frac32 \lambda \frac{K_d}{\varepsilon_4}.
\label{eq:GNY-counterterms}
\end{align}
\end{subequations}
It is convenient to include $K_d/2$ in the redefinition of $g$ and $\lambda$. The $\beta$-functions read
\begin{subequations}
\begin{align}
\beta_g (g,\lambda) & =\varepsilon_4 g - 6 g^2 \\
\beta_\lambda (g,\lambda) & =\varepsilon_4 \lambda - 3 \lambda^2,
\end{align}
\end{subequations}
The FP solution is given by
\begin{equation}
g_* = \frac{\varepsilon_4}{6},~\lambda_* = \frac{\varepsilon_4}{3}.
\end{equation}
The total RG flow in the three parameter space is shown in Fig.\ref{fig:3DFlow-SM}.
\begin{figure}
\includegraphics[width=75mm]{3DFlow-nobox}
\caption{
\label{fig:3DFlow-SM}
Schematic representation of the RG flow for the $U(N)$ GNY model~(\ref{eq:action-GNY-SM}) in the three-parameter space: $g$, $\lambda$ and $\mu$.
The transition is described by a fixed point of the flow
($g^{*}$, $\lambda^*$, $\mu_c$) which possesses only one unstable direction along $\mu$. On one side of the transition, the flow towards large $\mu$ corresponds to the semi-metallic phase with vanishing density of states at the band crossing and a Gaussian distribution of disorder. On the other side
the flow towards small $\mu$ drives the system towards a diffusive metal with a finite density of states and non-Gaussian distribution of disorder. }
\end{figure}
The critical length exponent is defined by:
\begin{equation}
\xi \sim \delta\mu^{-\nu}, \label{eq:sm-GN-nu}
\end{equation}
thus from the RG equations \eqref{eq:RG-4-eps}, identifying the parameter $L$ to the correlation length $\xi$ in the region near the critical point we get
\begin{equation}
\nu^{-1} =\gamma_\mu = 2 - \lambda_* = 2 - \frac{\varepsilon_4}3 + O(\varepsilon_4^2).
\end{equation}
The anomalous dimension of the fermionic field reads
\begin{align}
d_\chi & = \frac12 ( d - 1+ \eta_\chi) = \frac12 ( d-1 + g_*) \nonumber \\
& = \frac32 - \frac5{12} \varepsilon_4 + O(\varepsilon_4^2).
\end{align}
When $\delta \mu$ is negative, the scalar field acquires a finite expectation values which generates in turn a fermion mass $m_\chi = \sqrt{g} \langle \phi \rangle$.
This fermionic mass is analogous to a frequency and thus scales with the correlation length like $m_\chi \sim \xi^{-z}$.
Besides, the correlations of the scalar field are determined by a $\phi^4$ field theory; it is known that close to the transition the order parameter of this theory scales like
\begin{align}
\langle \phi \rangle \sim \delta\mu^\beta \sim \xi^{-\beta/\nu},
\end{align}
where $\nu$ is given by (\ref{eq:sm-GN-nu}).
Besides, from the RG analysis we have established that the \
dimensionful coupling constant $\sqrt{g}$ flows towards the FP as $\sqrt{g} \sim \xi^{-\varepsilon_4/2} \sqrt{g^*}$. This leads to
\begin{equation}
m_\chi \sim \sqrt{g} \langle \phi \rangle \sim \xi^{-\varepsilon_4/2-\beta/\nu},
\end{equation}
and therefore we get $z = \varepsilon_4/2 + \beta/\nu $.
Moreover, the exponent $\beta$ is related to the exponent $\nu$ through the scaling relation
\begin{equation}
\nu d = 2\beta + (2-\eta_\phi) \nu.
\end{equation}
From \eqref{eq:GNY-counterterms} we know $Z_\phi = 1$, which gives $\eta_\phi=O(\varepsilon_4^2)$. Hence we get
\begin{equation}
\beta = \nu \frac{2-\varepsilon_4}2,
\end{equation}
and finally we find for the critical dynamic exponent $z$ to one loop order:
\begin{equation}
z = \frac{\varepsilon_4}2 + \frac{2-\varepsilon_4}2 = 1 + O (\varepsilon_4^2) .
\end{equation}
The two-loop order contribution can be calculated using the two loop expression of $2-\eta_\phi$~\cite{Karkkainen:1994SM}:
\begin{equation}
2 - \eta_\phi = 2-\frac{\varepsilon_4^2}{54} +O(\varepsilon_4^3),
\end{equation}
which gives
\begin{align}
z & = \frac{\varepsilon_4}2 + \frac{\beta}{\nu} = \frac{\varepsilon_4}2 + \frac12 (d - 2 + \eta_\phi) \nonumber \\
& = 1 + \frac{\varepsilon_4^2}{108} + O(\varepsilon_4^3).
\end{align}
\subsection{Instanton solutions}
We now show the existence of localized instanton solutions to the GNY action in the limit $N \to 0$ that can give a non-zero contribution
to the zero-energy DOS in the semimetallic phase, similar to that found for the GN model in Ref.~\cite{Nandkishore:2014SM}.
Following \cite{Falco:2009SM} we start by rewriting the average DOS at the Dirac point directly in $d=3$ in the form:
\begin{equation}
\langle \rho (E=0) \rangle_V = \frac1{L^3} \int D[V,\chi,\Psi,\Upsilon] \exp[-S],
\end{equation}
where $\Psi(x)$ is a Lagrange multiplier field selecting solutions to the Dirac equation and $\Upsilon$ is a Lagrange multiplier enforcing normalization of $\Psi(x)$ and the action is given by
\begin{align}
S = & \int d^3x \left[ (\nabla \phi)^2 + \mu \phi(x)^2 + \frac{\lambda}{4!} \phi(x)^4 \right] \nonumber \\
& - \int d^3x~ \Psi^\dagger(x) ( {\bm \sigma} \cdot \nabla + \sqrt{g} \phi ) \chi(x) \nonumber \\
& + \Upsilon \left[ \int d^3x~ \chi^\dagger(x)\chi(x) - 1 \right], \label{eq:sm-action-instanton}
\end{align}
where ${\bm \sigma} = \sigma_x,\sigma_y,\sigma_z $ are the Pauli matrices.
We now look for a saddle-point solution to the classical equations of motion. To obtain the latter
we vary the action~(\ref{eq:sm-action-instanton}) with respect to $\phi$, $\chi$, $\chi^\dagger$, $\Psi^\dagger$, and $\Upsilon$.
\begin{center}
\begin{figure}
\includegraphics[width=7.5cm]{fig-instanton}
\caption{ The instanton wavefunction components ($f_1$: blue, continuous, $f_2$: red, dashed)
\eqref{eq:instanton-wf} and scalar field $\phi$ (black, dot-dashed) solution of \eqref{eq:instanton-a}-\eqref{eq:instanton} computed numerically using the expansion \eqref{eq:instanton-expansion}-\eqref{eq:instanton-effective-pot}
for $A=1$, $\mu=-0.1$, $\lambda=0.001$, $g=0.001$, and $\Psi_0=1$ . \label{fig:instantons-1}}
\end{figure}
\end{center}
This yields
\begin{equation}
- \nabla^2 \phi(x) + \mu \phi (x) + \frac{\lambda}{3!} \phi(x)^3 = \sqrt{g} \Psi^\dagger(x) \chi(x), \label{eq-sm-ceofm1}
\end{equation}
\begin{equation}
\Psi^\dagger(x)[{\bm \sigma} \cdot \nabla + \sqrt{g} \phi(x)] = 0,
\end{equation}
\begin{equation}
[{\bm \sigma} \cdot \nabla + \sqrt{g} \phi(x)]\chi(x) = 0,
\end{equation}
\begin{equation}
\int d^3x~ \chi^\dagger(x) \chi(x) = 1,
\end{equation}
\begin{equation}
\Upsilon ~\chi(x) = 0. \label{eq-sm-ceofm10}
\end{equation}
From Eq.~(\ref{eq-sm-ceofm10}) it follows that $\Upsilon=0$ and thus we can take $\Psi^\dagger(x) = \Psi_0 \chi^\dagger (x)$ where $\Psi_0$ is a real number.
Since the disorder distribution is isotropic it is naturally to assume that the DOS is dominated by a spherically symmetric
saddle-point solution. This drastically simplifies the solution of the classical equations of motion (\ref{eq-sm-ceofm1})-(\ref{eq-sm-ceofm10}) since they can be reduced to the problem of a Dirac particle in
a self-consistent central potential.
The solution to this problem can be factorized in the radial and angular parts according to
\begin{equation}
\chi = f_1(r) \varphi^- -f_2(r) \varphi^+,
\label{eq:instanton-wf}
\end{equation}
where $\varphi^\pm$ are two-component spinors with total angular momentum $j$, angular momentum along $z$ $j_z$ and orbital angular momentum $l^\pm = j \mp 1/2$. We have:
\begin{align}
{\bm \sigma} \cdot {\bm \nabla} f_i (r) \varphi^\pm = & {\bm \sigma} \cdot {\hat {\bf r}} \left( \partial_r - \frac{ {\bm \sigma} \cdot {\bf L}}{ r} \right) f_i (r) \varphi^\pm \nonumber \\
= & \left( \partial_r + \frac{1-\kappa}r \right) f_i (r) ~ {\bm \sigma} \cdot {\hat {\bf r}}~ \varphi^\pm \nonumber \\
= & \left( \partial_r + \frac{1-\kappa}r \right) f_i (r) \varphi^\mp,
\end{align}
with $\kappa = \pm (j+1/2) = \pm 1$ for the lowest level $j=1/2$. Thus we get the following system:
\begin{eqnarray}
&& \!\!\! \partial_r f_2(r) = f_1(r) \sqrt{g} \phi(r), \label{eq:instanton-a}
\end{eqnarray}
\begin{eqnarray}
&& \!\!\! (\partial_r + \frac2r) f_1(r) = f_2(r) \sqrt{g} \phi(r),
\end{eqnarray}
\begin{eqnarray}
&& \!\!\! -(\partial_r^2 + \frac2r \partial_r - \mu) \phi + \frac{\lambda}{3!} \phi^3 = \sqrt{g} \Psi_0 \left[ |f_1|^2 + |f_2|^2 \right]. \ \ \ \
\label{eq:instanton}
\end{eqnarray}
The large $r$ expansion of the Eqs.~(\ref{eq:instanton-a})-(\ref{eq:instanton}) gives the following asymptotic behavior
\begin{eqnarray}
f_1(r) & = &\frac{A}{r^2} + \frac{A^5 g^2}{30 \mu^2} \frac1{r^8} + O(\frac{A^9 g^4}{\mu^4 r^{14}}), \label{eq:instanton-expansion} \\
f_2 (r) & = &- \frac{A^3 g}{5\mu} \frac1{r^5} - 7 \frac{A^7 g^3}{550 \mu^3} \frac1{r^{11}} + O(\frac{A^{11} g^5}{\mu^5 r^{17}}),
\\
\phi (r) & = &\frac{A^2 \sqrt{g}}{\mu} \frac1{r^4} + O(\frac{A^6 g^{5/2}}{\mu^3 r^{10}}),
\label{eq:instanton-effective-pot}
\end{eqnarray}
which depends on a single free parameter $A$.
A typical solution to Eqs.~(\ref{eq:instanton-a})-(\ref{eq:instanton}) obtained numerically using the asymptotic
behavior~(\ref{eq:instanton-expansion})-(\ref{eq:instanton-effective-pot}) is shown in Fig.\ref{fig:instantons-1}.
The wave function and disorder distribution both exhibit a singular behavior at $r=0$ and thus require a regularization~\cite{Nandkishore:2014SM}.
Moreover, to obtain the full instanton contribution to the DOS (using either GN or GNY models) one has to expand around
the instanton solution and calculate the corresponding Gaussian integral which gives a prefactor to the exponential behavior. It is known
that in the case of quadratic dispersion this prefactor can be expressed in the form of a ratio of two functional determinants which
diverges in $d>1$. Thus, in this case the instanton solution requires renormalization~\cite{Falco:2015SM}.
The regularization and renormalization of the instanton solution in the case of a Dirac particles in disordered potential is an
interesting open question which is left for the future.
|
\section{Introduction}
While viable inflationary models of the early Universe have succeeded in the quantitative description of the main, smooth part of the primordial
power spectrum (PPS) of scalar (adiabatic density) perturbations $P_{\rm S}(k)$, and have even predicted its observed slope $n_{\rm S}(k)$
for some simplest variants of them, the complete shape of the primordial power spectra has not been established beyond doubt.
It is natural to expect small corrections to this smooth behavior which reflect new and subtle physical effects occurring during inflation. Indeed, some
relatively small features ($\lesssim 10\%$) have already been noticed in the low-$\ell$ ($\ell\lesssim 40$) multipole region of CMB fluctuations since
WMAP, which may result from tiny localized features in the PPS.
With the release of the new Planck temperature and polarization data~\cite{Planck:2015Like,Planck:2015Param,PLA}, we are now in a unique position to examine the existence
of primordial features in the scalar perturbations to a great precision over a wide range of cosmological scales. We had indication
of primordial features in all releases of WMAP temperature data~\cite{Peiris:2003ff,Komatsu:2008hk,Komatsu:2010fb,Hinshaw:2012aka} and Planck 2013-2015 temperature data~\cite{Planck:2013Inf,Planck:2015Inf}
and we have knowledge about the location and the types of the features from reconstructions of the primordial power spectrum. Inflaton potentials,
addressing the primordial features were proposed along the lines, that provided notable, if not significant, improvement in fit compared to the standard power law
primordial spectrum from slow roll inflation. A large scale scalar suppression, a dip near $2\times10^{-3} {\rm Mpc^{-1}}$, oscillations around
$0.02~{\rm Mpc^{-1}}$ and around $0.06~{\rm Mpc^{-1}}$ wavenumbers, are few notable features from Planck 2013 temperature anisotropy data~\cite{Hazra:2014PPSPlanck}.
Interestingly, the necessity of the large scale suppression of scalar power became significant when BICEP2 B-mode signal were considered primordial~\cite{BICEP2:Detection,BICEP2:datasets}.
The effect of tensor perturbation at the large scale temperature anisotropies aggravated (at more than 3$\sigma$ level~\cite{Hazra:2014RuleOut}) the already existing issue of lack of
low-multipole power of temperature anisotropies (1$\sigma$ support of power suppression from Planck 2013 TT data)~\cite{Hazra:2013BroadRecon}. We proposed Whipped Inflation (WI)~\cite{Hazra:2014WI} and Wiggly
Whipped Inflation (WWI)~\cite{Hazra:2014WWI} that could address all the above mentioned issues with the primordial power spectra and provided significant improvement in fit compared to
canonical slow roll models of inflation. Today, with the BICEP2 signal being consistent to dust polarization, one can anticipate a dramatic decrease in significance
of the large scale suppression. Also the large magnitude of quadratic potential, used in WI and
WWI makes it unsuitable for recent BICEP2/KECK-Planck joint constraints on the primordial B-modes (tensor-to-scalar ratio $r<0.07$ at 95\% confidence level~\cite{BICEP2Planck:2015joint,Array:2015xqh}).
However, note that the significance of the Wiggles in the WWI are not altered by change in the BICEP2 results since they were designed to address features
in Planck temperature data.
Since we now have tighter constraints in background and primordial cosmology from the new temperature and polarization data, it is important to revisit
the status of features in the light of new data. It is also appropriate to modify the WWI potential that can satisfy present bounds on B-mode and dust polarization.
If temperature and polarization votes for similar features it definitely shall increase its likelihood, hence equipped with Planck polarization data covering the largest
cosmological scales it is definitely worth having a joint constraints on features. Change in constraints on the background cosmological parameters in the presence
of features is another aspect that should be investigated. In this paper, we explore the above mentioned. It is clearly not possible to examine every primordial feature within
a single framework of potential. However, we show that our WWI potential is capable of producing a large variety of features that are discussed in the literature, which makes it extremely
efficient to hunt different types of features required by different datasets in the same framework.
The paper is organized as follows: In section~\ref{sec:scenario} we review the WWI potential in its old and modified form. We demonstrate to what extent WWI potential
can offer a variety of features in the PPS. In section~\ref{sec:num} we discuss the methodology, {\it i.e.} a brief outline of our numerical analysis. We provide our best fit results and
constraints on the model in section~\ref{sec:results}. We also present the bispectra from the best fits to the data in arbitrary triangular configurations of wavenumbers. We conclude
in section~\ref{sec:conclusions}.
\section{Wiggly Whipped Inflationary Scenario}~\label{sec:scenario}
The basic construction of the Wiggly Whipped Inflation potential followed the form :
\begin{equation}
V({\phi})=V_{S}+\gamma V_{R},~\label{eq:equation-WWI-basic}
\end{equation}
where, $V_{R}$ represents the part of the potential that introduces a moderate fast roll and $V_{S}$ is chosen to be nearly flat, allowing a strict slow roll.
Here, $\gamma$ determines how fast the inflaton starts rolling initially and in a way it determines the extent of deviation from slow-roll and thereby
deviations from featureless PPS. Similar types of phase transition potential have been discussed widely in the literature~\cite{S92,Linde:1998OpenInf,Linde:1999OpenInf2,CPKL03,JSS08,JSSS09,Bousso:2013uia}.
$V_{R}$ contains a Theta step function ensuring the termination of moderate fast roll after certain field value. Thereafter the scalar field emerges to the strict slow roll regime.
In this framework, the potential and/or its derivatives can contain discontinuities.
The WWI model was introduced in~\cite{Hazra:2014WWI} in order to address two major issues. Firstly with the BICEP2 B-mode
signal, assuming to be primordial, we ruled out the power law form of scalar power spectrum with more than 3$\sigma$ confidence~\cite{Hazra:2014RuleOut}. We needed
a strong suppression of scalar power at large scales and at the same time large tensors with low non-Gaussianities. Whipped inflation~\cite{Hazra:2014WI} that
simply uses a smooth transition from a moderate fast roll potential ($V_{R}$) to a strict slow roll inflation ($V_{S}$, assumed to be a quadratic potential) meets
all the above criteria. On top of that in order to address the primordial features indicated by Planck temperature anisotropies, we introduced a discontinuity in the potential (or in its
derivatives, keeping a continuous potential) at the transition which provided significant improvement in likelihood to the Planck and assumed primordial B-mode from BICEP2 compared to the
power law PPS. The scalar PPS, containing
a Whipped shaped tail at large scales and Wiggles (oscillatory behavior), the Inflaton potential was referred as Wiggly Whipped Inflation potential. The BICEP-2 B-mode signal being consistent to dust polarization,
do not favor a large field quadratic model since it produce large primordial B-modes. At the same time the significance of the requirement of a large scale scalar suppression
reduces. However, since we now have polarization data from Planck, it is important to test the WWI model features with the new datasets. In this paper we use the
WWI potential with a modification to the slow roll part of the potential.
\subsection{The potentials}
We use the WWI model in a modified form. Since BICEP2 B-mode signal is consistent with dust polarization power spectra
from Planck, a model with high tensor-to-scalar ratio ($r\sim0.2$) is not supported by the data anymore. Hence we use a
lower potential in the slow roll part. We use two different classes of WWI potential here, in the first case we have discontinuity
in the potential while in the second case the potential is continuous with a discontinuity in its slope. Both of them belong to the WWI form, but
to distinguish them conveniently, we refer the first potential as WWI and the second as WWI$'$.
Note that since the tensor perturbations depends on the scale of the potential, in order to get a lower
tensors we need to scale down the potential. Here, our basic structure of the potential remains same and provided below in Eq.~\ref{eq:equation-WWI}
\begin{equation}
V({\phi})=V_{i} \left(1-\left(\frac{\phi}{\mu}\right)^{p}\right)+\Theta(\phi_{\rm T}-\phi)\gamma V_{i}\left((\phi_{\rm T}-\phi)^{q}+\phi_{01}^q\right)~\label{eq:equation-WWI}
\end{equation}
The slow roll potential $V_{i} \left(1-\left(\frac{\phi}{\mu}\right)^{p}\right)$ depends on 2 parameters, namely the potential at $\phi=0$ ($V_{i}$) and $\mu$.
The spectral index ($n_{\rm s}$) and the tensor-to-scalar ratio ($r$) of the PPS generated at the end of inflation depends on $\mu$ and the power $p$.
We chose the value $p=4$ and $\mu=15~{\rm M_{PL}}$ such that $n_{\rm s}\sim0.96$ and $r\sim{\cal O}(10^{-2})$ (as in~\cite{Efstathiou:2006ak,Hazra:2010Step}).
The fast roll part of the potential remains identical to the original WWI potential. $\phi_{T}$ denotes the field value where the transition from the moderate
fast-roll to complete slow-roll occurs. $\phi_{01}$ is the extent of discontinuity and $\gamma$ is the slope that provides the deviation of slow-roll at the onset
of inflation. $\phi_{01}=0$ reduces the potential to Whipped Inflation form where the potential and its derivatives are continuous upto $(q-1)$'th derivatives.
The Heaviside Theta function $\Theta(\phi_{\rm T}-\phi)$ appearing in the potential can be modeled by a Tanh step ($1+{\tanh}[{(\phi-\phi_{\rm T})}/{\Delta}]$)
or Error function step ($1+{\rm erf}[{(\phi-\phi_{\rm T})}/{\Delta}]$) and the width of the step ($\Delta$) can be used as a free parameter. The scalar field
starts from the bottom of a quadratic potential (for $q=2$) and transits to the strict slow roll potential at $\phi=\phi_{\rm T}$. The initial deviation from slow-roll
introduces a whip shaped suppression of power at large and intermediate scales in the PPS. $\phi_{01}>0$ creates a temporary sharp departure from slow-roll and
generates wiggly features locally or extending a large range in cosmological scales depending on the sharpness of transition, $\Delta$. Hereafter, throughout the paper,
with WWI, we shall refer to the potential in Eq.~\ref{eq:equation-WWI}.
Using similar formalism, we investigate another potential that is continuous, but has discontinuities in its derivatives. The potential can be expressed as,
\begin{equation}
V({\phi})=\Theta(\phi_{\rm T}-\phi) V_{i} \left(1-\exp\left[-\alpha\kappa\phi\right]\right)+\Theta(\phi-\phi_{\rm T}) V_{ii}\left(1-\exp\left[-\alpha\kappa(\phi-\phi_{01})\right]\right)~\label{eq:equation-WWI'}
\end{equation}
Here $V_{i} \left(1-\exp\left[-\alpha\kappa\phi\right]\right)$ is the slow roll part of the potential which is present at $\phi\le\phi_{\rm T}$ and for higher field values,
$V_{ii}\left(1-\exp\left[-\alpha\kappa(\phi-\phi_{01})\right]\right)$ represents the moderate fast roll part.
Note that we have used the $\alpha$-attractor potential~\cite{alpha_attractor} to construct a model providing deviations from slow roll. $V_{ii}$ is related to $V_{i}$ by our demand of continuity in the potential.
$\kappa^2=8\pi G$ (which we equate to 1 in our convention). $\alpha$ denotes the slope of the potential and we have fixed it to be $\sqrt{2/3}$ that corresponds to the $R+R^2$ inflationary model~\cite{Starobinsky:1980te}
in the Einstein frame. It produces $r\sim 4\times10^{-3}$ for the slow roll part of the spectra. Hence $\phi_{\rm T}$ and $\phi_{01}$ are the only two extra parameters in our potential compared to the strict slow
roll part of the potential. Note that this potential can also have a discontinuity if $V_{ii}$ is treated as a free parameter, but since we have already incorporated potential discontinuities in Eq.~\ref{eq:equation-WWI}, in
Eq.~\ref{eq:equation-WWI'} we fix $V_{ii}$ through the continuity. We denote this potential with WWI$'$ to indicate that the potential has a discontinuity only in its derivatives.
The primordial feature generated from WWI$'$ is very similar to the original Starobinsky-1992 model~\cite{S92} but here, the scalar PPS has the asymptotic value of $n_{\rm S}\sim0.96$.
\subsection{Classes of features in WWI}
CMB angular power spectrum data since WMAP have indicated hints towards possible deviations from the standard power law PPS. Reconstructions of the PPS
directly from the angular power spectrum data~\cite{Hazra:2014PPSPlanck,reconstruction-all} have been very useful in order to highlight different
locations and types of these deviations (features) in the PPS. Using Planck 2013 data, it has been demonstrated~\cite{Hazra:2014PPSPlanck} that the
temperature anisotropy power spectrum indicates a suppression at large scale power, few localized oscillations around $\ell\sim22$, $\ell\sim250-300$ and $\ell\sim750-850$.
Non-standard inflationary models have been proposed in order to generate different kinds of features in the PPS. Below we enlist a non-exhaustive discussion of primordial features.
\begin{figure*}[!htb]
\begin{center}
\resizebox{210pt}{160pt}{\includegraphics{ps_cutoff-eps-converted-to.pdf}}
\resizebox{210pt}{160pt}{\includegraphics{ps_localized_feature-eps-converted-to.pdf}}
\resizebox{210pt}{160pt}{\includegraphics{ps_non_local-eps-converted-to.pdf}}
\resizebox{210pt}{160pt}{\includegraphics{ps_mixed-eps-converted-to.pdf}}
\end{center}
\caption{\footnotesize\label{fig:feature-range} [Top left] : PPS with suppression at large scale scalar power. We show that WWI (Eq.~\ref{eq:equation-WWI}) can generate a exponential cutoff (blue) type
PPS and also a PPS with a form of a step (red). [Top right] Here WWI produces localized oscillations. Note that a dip or a period of oscillation around $0.002~{\rm Mpc}^{-1}$
improves the fit to the data near multipole $\ell=22$. [Bottom left] Wide features and non-local oscillations. The angular power spectra are affected for a wide range of multipoles.
[Bottom right] We provide a combination of different features. The large range of features that are addressed by WWI models, makes WWI a suitable candidate to search for generic
features in the primordial power spectrum. A typical step like suppression, accompanied by wiggles in the PPS is generated by the WWI$'$ (Eq.~\ref{eq:equation-WWI'}) and is plotted here as well.}
\end{figure*}
1. Large scale power suppression : The primordial power spectrum with a large scale suppression can be modeled by a broken power law with different spectral index at different scales, or
a Tanh step (both these types have been discussed in~\cite{Hazra:2013BroadRecon}) or an exponential cutoff~\cite{Shafieloo:2004}. These types of features are generated
when the scalar field changes its kinetic energy in the first few {\it e-folds}. Starting from Starobinsky-1992 model~\cite{S92},
different models~\cite{Bousso:2013uia,Bousso:2014jca,Hazra:2014WI} have been
proposed that offers such scenarios. A brief halt in inflation caused by an inflection point~\cite{Jain:2009PI} in the potential can produce a sharp cut-off at the large scale
primordial power. Open inflation~\cite{Linde:1998OpenInf,Linde:1999OpenInf2}, radiation dominated epoch prior to inflation~\cite{Powell:2006RADDOM} also known
to provide such class of spectra. In Fig.~\ref{fig:feature-range}
we plot such type of spectra (in the top left) that are obtained from WWI (Eq.~\ref{eq:equation-WWI}). Note that in these cases, large value of the parameter $\gamma$
allows an initial faster roll and hence provide a cutoff at the large-intermediate scales.
2. Localized oscillation : Near multipoles $\ell\sim22$ and $\ell\sim40$ a dip and a bump {\it w.r.t.} the angular power spectra from power law were noticed since
WMAP data. This feature patterns in a localized scale and requires wiggles in the PPS only at certain wavenumbers, keeping the other part of the PPS nearly scale
invariant. Presence of {\it a step in the inflaton potential}~\cite{Hazra:2010Step,step-models} provides a momentary departure from the slow roll and generate localized
oscillations in the PPS. It has been shown that around $k\sim0.002~{\rm Mpc^{-1}}$ and $k\sim0.004~{\rm Mpc^{-1}}$, a period of oscillation can improve the match to the
data compared to the power law model. However, because of uncertainties owing to the cosmic variance, the existence of such large scale feature
have never been (possibly never will be) established beyond doubt. Top right of Fig.~\ref{fig:feature-range} shows similar features from WWI. Here, $\gamma$ is small such that
the fast roll part is not significant and we get nearly same tilt both at large and small scales. However, due to $\phi_{01}$, the field goes through a momentary
departure from slow roll and generates localized oscillations.
3. Non-local features : Features that extend over a wide range of cosmological scales are termed as non-local features. A slow roll potential modulated with sinusoidal
oscillations or presence of discontinuities in the potential and its derivatives give rise to such non local features. Effects of oscillations in the inflation potential~\cite{oip}
have been studied before with WMAP and Planck datasets. In~\cite{Hazra:2014WWI} we discussed the features generated by the WWI model and demonstrated
that non-local features in the WWI can provide similar improvement in fit as local feature models. Fig.~\ref{fig:feature-range}, bottom left plot demonstrates such features
from WWI. Here, $\gamma$ and $\phi_{01}$ are small and hence we do not observe any cutoff and the amplitude of the features are small as well. However, in these models $\Delta<<1$
and hence the field experience a sharp transition and a wide range of modes that leaves Hubble scale after the epoch of transition imprints oscillations in the PPS.
Wiggly Whipped Inflation, interestingly, is capable of generating all the above features and hence, we present it as the model offering a wide variety of
scenarios, within the framework of canonical scalar field Lagrangian. Fig.~\ref{fig:feature-range} bottom right plot represent two such PPS that offers the aforementioned classes
of features in combination, all arising from the same potential. Note that in the same plot we present the feature obtained from WWI$'$. The discontinuity in the derivative
of the potential leads to a power spectrum that offers a step shaped suppression at larger scales, followed by wiggles at the smaller scales. In this case, the
amplitude of oscillations decrease as we probe smaller scales.
\section{Essential numerical details}~\label{sec:num}
We use the publicly available code BI-spectra and Non-Gaussianity Operator, {\tt BINGO}~\cite{Hazra:2012BINGO,Sreenath:2014BINGO2} to generate
the power spectra and the bispectra from WWI and WWI$'$. We solve the background equation using a initial value of the field $\phi$ to ensure enough ($\sim70$ {\it e-folds}) inflation.
${\rm d}\phi/{\rm d} t$ is fixed assuming initial slow-roll condition ($3 H {\rm d}\phi/{\rm d} t=-{\rm d} V(\phi)/{\rm d} \phi$). Whenever necessary, we model the Theta function discontinuities
in the potential and delta function in its derivatives with a Tanh step and its derivatives respectively. Note that other representations of Theta function and
delta function can also be used in this context. Following standard methodology we fix the initial scale factor by assuming the $k=0.05~{\rm Mpc}^{-1}$ mode leaves
the Hubble radius 50 {\it e-folds} before the end of inflation.
We use publicly available {\tt CAMB}~\cite{cambsite,Lewis:1999bs} and {\tt COSMOMC}~\cite{cosmomcsite,Lewis:2002ah} in order to calculate the angular power spectra from
our models and compare them with the data. Note that, we modify {\tt CAMB} in order to use {\tt BINGO} to calculate the primordial scalar and tensor power spectra. Along with
the baryon density ($\Omega_{\rm b} h^2$), cold dark matter density ($\Omega_{\rm CDM} h^2$), the ratio of the sound horizon to the angular diameter distance at decoupling ($\theta$)
and the optical depth ($\tau$), we allow the WWI parameters, namely
$V_{i}$, $\gamma$, $\phi_{01}$ and $\phi_{\rm T}$ to vary. We treat the WWI parameters as semi slow-parameters (like the amplitude $A_{\rm S}$, spectral index $\rm n_{\rm S}$,
tensor-to-scalar ratio ($r$) in case of power law PPS). Note that we also allow the width ($\Delta$) of the Theta function to vary along with other potential parameters.
For WWI$'$ we treat $V_i$ and $\phi_{\rm T}$ and $\phi_{01}$ as potential variables.
In order to obtain the best fits we use Powell's BOBYQA (Bound Optimization BY Quadratic Approximation) method of iterative minimization~\cite{powell}. Since WWI offers
a wide variety of features, it is extremely difficult to converge to a global minima starting from a particular region of potential parameter space. Most of the times
the method settles to a local minima which represent particular features that provide a better fit to a subset of the complete Planck datasets. We approach this problem
in 2 steps. First, we use temperature, polarization datasets separately and in combination so that we can identify the primordial features supported by the individual and
complete Planck and BICEP2/KECK datasets. Secondly, in each of the above cases we perform MCMC analyses and locate distinguishable features that provide better fit to the
corresponding datasets compared to power law model. We use the WWI potential parameters and corresponding background cosmological parameters corresponding to the located
features and use them as starting points of BOBYQA minimization. Using this rigorous search, we are able to obtain the local minima and possibly the global minima for the individual and
complete datasets.
We use CMB temperature and polarization data from Planck-2015 public release datasets and likelihoods. In order to understand the necessity of features in the PPS indicated from temperature
and polarization data, we use the temperature and polarization likelihood from Planck separately and in combination. For high-$\ell$ temperature and polarization data,
we use the Plik likelihood that covers the multipoles $\ell=30-2508$ for TT and $30-1996$ for TE and EE.
For the low-$\ell$ part of the spectra ($\ell=2-29$) we use 2 different likelihoods in different cases. When we use only TT likelihood at high-$\ell$, we use
commander based likelihood for low-$\ell$ TT. In this paper we denote this likelihood as lowT. We use lowTEB likelihood at low-$\ell$ whenever we use EE data at high-$\ell$
or temperature and polarization likelihood in combination. The low-$\ell$ polarization uses the 70GHz LFI full mission (except second and fourth surveys) data. We use
TTTEEE likelihood and use TT + TE + EE likelihood at high-$\ell$ to track down the improvement in likelihood (compared to the power law model of the PPS) in
complete and individual datasets. The high-$\ell$ likelihood uses 100, 143 and 217 GHz half mission maps. Throughout our analyses, we have varied all the required
nuisance and calibration parameters for the corresponding likelihoods. We also use priors on nuisance and calibration parameters as have been discussed in~\cite{Planck:2015Like,PLC}.
Since the WWI and WWI$'$ potentials generate fine oscillations in the PPS, in order to take into account the effects of all the oscillations in the angular power spectra, we calculate the
angular power spectra at all $\ell$'s to avoid interpolation. Moreover, the accuracy that we use for our analyses, ensures the transfer function is well sampled in
$k$-space that resolves the finest oscillations in the PPS in the convolution integration. We also make use of the unbinned Plik likelihood for high-$\ell$ TT and TTTEEE datasets
in order to obtain the best fits. However, we use the binned Plik likelihood for the MCMC analyses since we did not find noticeable differences in the best fit $\Delta\chi^2$
and the cosmological parameters and calculation of unbinned likelihood is significantly slower compared to the binned ones.
For completeness too, since unbinned data for CMB polarization alone is not publicly available, in order to compare the individual constraints on the cosmological parameters,
binned likelihood are more useful in our MCMC runs.
We use BICEP2-Keck likelihood from the joint BICEP2/Keck and Planck analysis (BKP)~\cite{BICEP2Planck:2015joint}. In our analyses we use 5 bandpowers in the range $\ell=20-200$,
from 150 GHz band of BICEP2/Keck and 217GHz and 353GHz bands of Planck. Needless to mention we compute and use the tensor power spectra from the models when we confront the models
with BKP datasets.
We use the BINGO-2.0 version in order to evaluate the bispectra in equilateral and arbitrary triangular configurations. We have defined ${\rm M_{Pl}}^2=1/(8\pi G)=1$ and
used $\hslash=c=1$ throughout the paper.
\section{Results and discussions}\label{sec:results}
We present our results in this section in the following manner. We provide the best fit primordial power spectrum that we obtained from the complete
datasets. We tabulate the best likelihood values and the improvement in fit from WWI compared to the power law PPS. Using the best fit values,
we compare the best fit angular power spectra of temperature and polarization and their cross-correlations {\it w.r.t} best fit Planck baseline models.
We compare the background parameter constraints from the MCMC analyses against different datasets.
Finally, we present the bispectra in equilateral and arbitrary triangular configurations for the best fit models.
\subsection{Primordial scalar perturbation power spectrum}
We have demonstrated that WWI offers a wide variety of features in the primordial power spectra. When compared against the
Planck datasets we locate 4 local minima that provide substantial better fit compared to the power law power spectra. We categorize the local regions
of parameter space by their distinct nature of features. We denote them as WWI-[a,b,c,d]. We plot these models in left and middle plots of Fig.~\ref{fig:psk}. Here we should
stress that WWI-[a,b,c,d] that are plotted in the figure, in strict sense, are the local minima of the TTTEEE + lowTEB + BKP combination. WWI-a is actually the global
minima of TT + lowT dataset, that turns out to be a local minima of the complete datasets. WWI-c on the other hand, has characteristic wide scale oscillations that provide
improvement in fit to the high-$\ell$ EE data. We search for the local and global minima in other individual and combination of datasets in the vicinity of WWI-[a,b,c,d]. Obviously the best fit
parameters will be different when we compare different datasets. However, the broad shape of that particular feature remains similar in all the datasets since
the WWI potential parameters do not change significantly. The right plot of the same figure provide best fit PPS from WWI$'$ when compared
against T, E and combined datasets.
\begin{figure*}[!htb]
\begin{center}
\resizebox{142pt}{105pt}{\includegraphics{ps_1-eps-converted-to.pdf}}
\resizebox{142pt}{105pt}{\includegraphics{ps_2-eps-converted-to.pdf}}
\resizebox{142pt}{105pt}{\includegraphics{ps_II-eps-converted-to.pdf}}
\end{center}
\caption{\footnotesize\label{fig:psk} Wiggly Whipped Inflation : Best fit primordial power spectra. Left plot contains the local minima, WWI-[a,b] and the middle plot contains WWI-[c,d] respectively (Eq.~\ref{eq:equation-WWI}).
WWI-a provides $\sim7-8$ improvement in $\chi^2$ fit to temperature only data. Mainly the improvement comes from the large scale power suppression and $\ell\sim22-40$ region.
WWI-c provides $\sim10$ improvement in $\chi^2$ fit to polarization data. This improvement comes from low-$\ell$ TEB and high-$\ell$ E data.
WWI-b provides $\sim11$ and WWI-d provides more than $13$ improvement to combined temperature and polarization datasets. In this case, most of the improvement comes owing to the inability of
baseline model in fitting the temperature and polarization datasets in a combination. The plot at the right contains the WWI$'$ (Eq.~\ref{eq:equation-WWI'}) best fit when compared with T, E and combined datasets. For the combined
datasets, WWI$'$ provides nearly 12 improvement in the fit compared to power law best fit.}
\end{figure*}
\subsection{Best fit results}
In Table~\ref{tab:bestfits} we tabulate the best fit $-2\log{\rm [Likelihood]}$ for the WWI-[a,b,c,d], WWI$'$ and the Planck baseline model.
Each row block of the table contains the data combination that we used. From top to bottom we provide the analyses for TT + lowT,
EE + lowTEB, TTTEEE + lowTEB, TT + TE + EE + lowTEB + BICEP-Keck-Planck dust, TTTEEE + lowTEB + BICEP-Keck-Planck dust, unbinned TT + lowT and unbinned TTTEEE + lowTEB. Here we should
mention again that the cosmological parameters do change when compared with different combination of datasets. For example WWI-a that is
a local minima of TTTEEE + lowTEB (as plotted in Fig.~\ref{fig:psk}) is not {\it strictly} a local minima of TT + lowT (the position and the amplitude
of the features may vary a little). We find with a marginal change in parameter values, a PPS very close in shape to WWI-a, represent the global
best fit to TT + lowT datasets. Since Powell's minimizer algorithm converges to a point very close to the parameter space of the starting point
in our cases, we searched for the local best fit to TTTEEE + lowTEB + BKP data in the vicinity of the global best fit to TT + lowT data.
Hence, WWI-[a,b,c,d] represent a class of resembling PPS that can possibly be a global best fit to a subset of data but in strict sense they
are local best fit to the TTTEEE + lowTEB + BKP dataset. Obtained potential parameters, $\left[\ln (10^{10} V_i),~\phi_{01},~\gamma,~\phi_{\rm T},~\ln(\Delta)\right]$ for TTTEEE + lowTEB + BKP
are $\left[1.73,~0.099,~0.019,~7.88,~-4.47~\right]$ (for WWI-a), $\left[1.75,~0.019,~0.04,~7.91,-7.07~\right]$ (for WWI-b),
$\left[1.72,~0.039,~0.02,~7.91,-6.04~\right]$ (for WWI-c) and $\left[1.76,~0.017,~0.03,~7.91,-11~\right]$ (for WWI-d). However, since WWI$'$ has a unique shaped PPS, in Fig.~\ref{fig:psk} we plotted the obtained global best fits of
TT + lowT, EE + lowTEB and TTTEEE + lowTEB + BKP datasets. For WWI$'$, the obtained values for inflaton potential parameters, $\left[\ln (10^{10} V_i),~\phi_{01},~\phi_{\rm T}\right]$, are
$\left[0.344,~0.12,~4.5\right]$ (for TT + lowT), $\left[0.295,~0.11,~4.5\right]$ (for EE + lowTEB) and $\left[0.282,~0.11,~4.51\right]$ (for TTTEEE + lowTEB + BKP).
\renewcommand{\arraystretch}{1.1}
\begin{table*}[!htb]
\footnotesize{
\begin{center}
\vspace{4pt}
\begin{tabular}{| c | c | c | c | c | c | c |}
\hline\hline
\multicolumn{7}{|c|}{\bf Individual likelihoods comparison}\\
\hline
Individual & Baseline & WWI-a & WWI-b & WWI-c & WWI-d & WWI$'$\\
likelihood & & $\Delta_{\rm DOF}=4$&$\Delta_{\rm DOF}=4$ & $\Delta_{\rm DOF}=4$&$\Delta_{\rm DOF}=4$ &$\Delta_{\rm DOF}=2$\\
\hline
TT &761.1 &762 &761.9 &762.8 &762.8 & 762.4\\
lowT &15.4 &8.2 & 13.4 &12.1 &13 & 10.2\\
Total &778.1 &772.1 (-6) &777 (-1.1) &777 (-1.1) &778.4 (0.3) & 775 (-3.1)\\
\hline
EE &751.2 & 748.8 & 747.2 & 748.6 &750.2 & 746.8\\
lowTEB &10493.6&10490 &10495.6 &10492.4&10495.7 &10492.2\\
Total &11248.8&11241.8 (-7)&11246.2 (-2.6) &11244.5 (-4.3)&11249.3 (0.5) &11242.3 (-6.5) \\
\hline
TTTEEE &2431.7 &2432.7 & 2422.6 & 2427.8&2421.7 &2426.5\\
lowTEB &10497 &10490.8& 10495.1 &10493.4&10495.3 &10492.7\\
Total &12935.6&12929.5 (-6.1)& 12924.2 (-11.4) &12927.6 (-8)&12923.4 (-12.2) &12925.2 (-10.4) \\
\hline
TT & 764.5 &763.6 &762.2 &764.4 &762.9 &762.8 \\
EE & 753.9 &754.8 &750.5 &750.8 &750.8 &751\\
TE & 932 &933.4 &928.7 &929.2 &927 &928.8\\
lowTEB & 10498.4&10490.4&10495.8 &10493.7&10495.6&10492.4\\
BKP & 41.6 &42 &42 &42.6 &41.8 &42.9\\
Total &12997 &12991 (-6) &12985.9 (-11.1) &12987.2 (-9.8)&12985 (-12) &12985.1 (-11.9) \\
\hline
TTTEEE & 2431.7&2432.8 &2421.4 &2426.7 &2421 &2425.7 \\
lowTEB &10498.5&10490.5 &10495.5&10493.6 &10495.8 &10492.6\\
BKP & 41.6 &42 &42.7 &42 &41.9 &42.5 \\
Total &12978.3&12971.3 (-7) &12967.3 (-11)&12968.6 (-9.7) &12965 (-13.3) &12968.6 (-9.7) \\
\hline
TT (bin1) &8402.1&8404.1 &8403.9 & 8405.2 &8402.1 &8401.9 \\
lowT &15.4 &8.3 &13.3 & 11.9 &13.2 &10.3\\
Total &8419.6&8414.7 (-4.9)&8419.5 (-0.1) & 8419.8 (0.2) &8418.1 (-1.5) &8414.4 (-5.2) \\
\hline
TTTEEE (bin1) &24158.2&24158.6 &24149 &24155 &24148.4 & 24151.5\\
lowTEB &10497.6&10490.3 &10493.4 &10493.6 &10495.3 & 10492.7\\
Total &34661.9&34655.3 (-6.6)&34650.5 (-11.4)&34654.4 (-7.5) &34649.5 (-12.4)& 34650.6 (-11.3) \\
\hline\hline
\end{tabular}
\end{center}
\caption{~\label{tab:bestfits} Best fit parameters for the Wiggly Whipped Inflaton potential. We provide the best fit
$-2 \ln [{\rm Likelihood}]$ for individual datasets for the types of WWI features denoted as WWI-[a,b,c,d], WWI$'$ and that are plotted
in Fig~\ref{fig:psk}. Note that while WWI-[a,c] provide local best fits to individual and combined data, WWI-[b,d] seem to capture
the global best fit by providing improvement in all datasets (TT, TE, EE) in a combination when compared with power law model.
The best fit background, WWI potential and nuisance parameters do indeed change to certain extent in different dataset combinations
but the primordial power spectrum remains very similar to the ones plotted in Fig.~\ref{fig:psk} for all dataset combinations.
In case of WWI$'$, note that the similar shaped features (shown in right plot of Fig.~\ref{fig:psk}) are able to address both temperature and polarization data
individually and in combinations. We find similar improvement in fit ($\Delta\chi^2\sim-12$ compared to power law) as in WWI-[b,d] but here only with
2 extra parameters. WWI$'$ also provides the best likelihood to the high -$\ell$ EE data. We do not provide in the table, the ${\cal O}(1)$ improvement in fits
from the priors in the nuisance and calibration parameters. $\Delta_{\rm DOF}$ indicates extra parameters used in the models compared to its featureless potential
parameters. The `bin1' in the last 2 row blocks indicate unbinned high-$\ell$ data have been used to obtain the best fits. In the `Total' row we also
provide the differences in the $\chi^2$ {\it w.r.t} the power law model within brackets.}}
\end{table*}
\begin{figure*}[!htb]
\begin{center}
\resizebox{420pt}{320pt}{\includegraphics{clTT_1-eps-converted-to.pdf}}
\resizebox{210pt}{160pt}{\includegraphics{clTE_1-eps-converted-to.pdf}}
\resizebox{210pt}{160pt}{\includegraphics{clEE_1-eps-converted-to.pdf}}
\end{center}
\caption{\footnotesize\label{fig:clab} Wiggly Whipped Inflation : Angular power spectra for temperature and polarization anisotropies. Theoretical
predictions from WWI-a and WWI-b are provided. These are the best fits to TTTEEE + lowTEB + BKP datasets. Power law baseline best fit is provided in
black. Power spectra from WWI and the data, that residual from power law best fit model are plotted below to highlight the features in the angular power spectra.}
\end{figure*}
\begin{figure*}[!htb]
\begin{center}
\resizebox{420pt}{320pt}{\includegraphics{clTT_2-eps-converted-to.pdf}}
\resizebox{210pt}{160pt}{\includegraphics{clTE_2-eps-converted-to.pdf}}
\resizebox{210pt}{160pt}{\includegraphics{clEE_2-eps-converted-to.pdf}}
\end{center}
\caption{\footnotesize\label{fig:clcd} Wiggly Whipped Inflation : Angular power spectra for temperature and polarization anisotropies. Theoretical
predictions from WWI-c and WWI-d are provided. These are the best fits to TTTEEE + lowTEB + BKP datasets. Power law baseline best fit is provided in
black. Power spectra from WWI and the data, that are residual from power law best fit model are plotted below to highlight the features in the angular power spectra.}
\end{figure*}
\begin{figure*}[!htb]
\begin{center}
\resizebox{420pt}{320pt}{\includegraphics{clTT_II-eps-converted-to.pdf}}
\resizebox{210pt}{160pt}{\includegraphics{clTE_II-eps-converted-to.pdf}}
\resizebox{210pt}{160pt}{\includegraphics{clEE_II-eps-converted-to.pdf}}
\end{center}
\caption{\footnotesize\label{fig:clprime} Wiggly Whipped Inflation : Angular power spectra for temperature and polarization anisotropies. Theoretical
predictions from WWI$'$ are provided. These are the best fits to TT + lowT, EE+lowTEB and TTTEEE + lowTEB + BKP datasets. Power law baseline best fit is provided in
black. Power spectra from WWI$'$ and the data, that are residual from power law best fit model are plotted below to highlight the features in the angular power spectra.}
\end{figure*}
Note that WWI-a provides improvement in likelihood principally for the lowT case. The localized feature similar to step in the inflaton potential
around $k\sim0.002~{\rm Mpc}^{-1}$ ($\ell\sim22$) along with the large scale power suppression provides around 8 improvement in fit to the lowT and lowTEB data.
WWI-a also provide some improvement to the high-$\ell$ EE data. WWI-c which introduces wiggles in the PPS ranging $k\sim0.002-0.05~{\rm Mpc}^{-1}$
provide a moderate improvement to lowT and lowTEB datasets due to the large scale suppression but fails to capture the $\ell\sim22$ dip in TT. However
the wide wiggles provide a better fit to EE and TE-datasets. WWI-b and WWI-d, though having wiggles in the primordial power spectrum, do not provide
better fit to individual temperature or polarization data at high-$\ell$ but they provide an overall better fit when the complete datasets are compared with.
Note that when we combine T and E datasets (TT + TE + EE + lowTEB + BICEP-Keck-Planck dust), the baseline model provide worse fit to TT and EE datasets
{\it w.r.t.} their values when compared individually (TT + lowT and EE + lowTEB). This difference arise because temperature data favors a low baryon
density ($\Omega_{\rm b}{h^2}\sim 0.0222$) while E-polarization goes for a higher ($\Omega_{\rm b}{h^2}\sim 0.024$) value. The baseline model can not trade-off for the baryon
density and hence settles for marginal worse likelihood to all datasets. We carry out the analysis with TT + TE + EE + lowTEB + BICEP-Keck-Planck dust specifically
to point out the individual worse fits. WWI-b and WWI-d manage to compensate for the baseline worse fit by providing wiggles in the PPS that can address EE data better
than the baseline model {\it but} at the same time keeping a low baryon density ($\Omega_{\rm b}{h^2}\sim 0.023$). These two spectra represent the global fit to the overall Planck
data. Though the WWI-b and WWI-d are indistinguishable from the likelihood values, they have distinct signatures in the power spectrum. The Wiggles in WWI-b damps down at
$k\sim0.1~{\rm Mpc}^{-1}$ while WWI-d continues to be present at smaller scales. The difference between these two spectra becomes evident when we calculate the three point
correlations {\it i.e.} the bispectra. It is interesting that WWI$'$, with a fixed shaped power spectra feature, is able to provide better fit to the temperature and polarization
data individually and also in combinations. In Fig.~\ref{fig:psk} (the right plot), we have provided the best fit PPS for WWI corresponding to the best fit provided in Table~\ref{tab:bestfits}.
We find $\Delta\chi^2\sim-5$ and $-6$ {\it w.r.t.} power law when we compare WWI$'$ with the temperature and polarization data respectively. This particular best fit do not provide
better fit to the high-$\ell$ TT data but the best fit for EE quoted in the Table is in very good agreement with the high-$\ell$ EE data. However, the position and amplitude of the
oscillations required to have a good agreement with TT and EE data are different. When we compare WWI$'$ with the combined datasets,
the best fit provides $\Delta\chi^2\sim -12$ compared to power law with only 2 extra parameters. Breakdown of likelihood in temperature and polarization data is very similar to
WWI-[b,d] case, and we find, WWI$'$ with the PPS plotted in Fig.~\ref{fig:psk} is providing improvement to T and E datasets combinations from Planck since the best fit likelihoods
from power law are degraded from the individual best fit values when temperature and polarization data are used separately.
In Fig.~\ref{fig:clab} and~\ref{fig:clcd}, we plot the best fit angular power spectra for temperature and polarization anisotropies from WWI and
the Planck data. In both the plots we have also plotted the best fit results from the power law PPS in black. In each plot, the bottom panel
represent the data and the power spectra residual to the power law best fit. The left panel in each plot captures the multipoles $\ell=2-29$
and are plotted in log scales while the right panels display high-$\ell$ ($\ell=30-2508$ in case of TT and $\ell=30-1996$ in case of EE and TE) data
and best fit results in linear scale. Note that, here the plotted results correspond to the best fit obtained against TTTEEE + lowTEB + BKP datasets.
Improvement in fit from the low-$\ell$ data is evident the suppression in the low-$\ell$ residual plot in all the best fits. Only WWI-a is able to
address the dip around $\ell=22$ in a convincing manner. WWI-[b,c,d] introduce features at the high-$\ell$ as well. However, note that, at high-$\ell$
we do not get notable improvement in fit when we use temperature and polarization data separately, but in a joint analysis, WWI-b and WWI-d
interestingly provide a noticeable better fit. In Fig.~\ref{fig:clprime} we plot the best fit angular power spectra and their residuals from power law best fit
for WWI$'$. Unlike WWI model, in this plot we plot best fit from TT + lowT, EE + lowTEB and TTTEEE + lowTEB + BKP. For the TT plot we show best fit
from TT + lowT and TTTEEE + lowTEB + BKP and for EE plot we present EE + lowTEB and TTTEEE + lowTEB + BKP best fits. For TE, we provide all three best fit
power spectra. In the TE plot, the EE best fit is seen to be not fitting the data well at the acoustic peaks. Since high-$\ell$ temperature data is not used in obtaining
EE + lowTEB best fit, we can expect some disagreement between temperature and polarization best fits. The large mismatch here might point out a systematic tension between T and E
data. However, we should note that the statistical uncertainties in the EE data is substantially larger than the TT data and differences in their best fits
can just be an artifact of statistical fluctuations.
\subsection{Change in the parameter constraints}
The presence of features change the background parameter constraints from power law. Due to the features in WWI and WWI$'$ we find certain changes in the background
parameters. The power suppression at large angular scales shifts the reionization optical depth to a higher value $\tau\sim0.07-0.1$ compared to the power law case.
Here, we present only the change in the baryon density constraints. In Fig.~\ref{fig:omegabh2}, we plot the marginalized likelihoods from power law, WWI and WWI$'$.
Note that for WWI, in arriving at the constraints, we have fixed the width of the transition ($\Delta$) to the best fit value of WWI-d using the prior knowledge
that WWI-d is able to provide an agreement to the baryon density mismatch from power law PPS. In this figure the each plot at the top corresponds to different models
that compares the difference in the likelihood of baryon density from T, E and from their combination. For power law model we note that the peak of the likelihood
from EE + lowTEB is significantly far away from the tail of TT + lowT data. When feature models are used for the PPS, we find that the likelihoods shift towards lower
baryon density for EE (from $\Omega_{\rm b}{h^2}\sim 0.024$ for power law to $\sim0.023$ for WWI and WWI$'$). In the bottom panel of the same figure, we provide
the same likelihoods but plotted for T, E and complete datasets for comparison. We find that the shift is not substantial given the large standard deviation of EE
data but it definitely reduces the tension to a fair extent that amounts to more than 13 improvement in $\chi^2$ values. Furthermore, we find that the tension is not relaxed
due to the increase in degeneracy owing to the extra parameters. In fact, we find that for WWI and WWI$'$ models, the likelihood of baryon density is sharper than the
power law. The standard deviation of the likelihood in WWI and WWI$'$ models are 0.0012 and 0.001 respectively while for power law it is 0.0014. Qualitatively the
likelihoods can also be compared in middle plot of lower panel of Fig.~\ref{fig:omegabh2}. Here, we would like to point out that since our slow roll part of the potential generates a spectral tilt
$\sim 0.96$ (the asymptotic or average $n_{\rm S}$ at small scales), in our analyses we do not have the scope of marginalizing over the spectral tilt and hence the comparison
with power law case, where we marginalize over $n_{\rm S}$, might not be complete. Hence we obtained the standard deviation of the baryon density from the power law model
against EE + lowTEB dataset by fixing the $n_{\rm S}=0.964$, which turns out to be $\sim0.001$, similar to WWI$'$ model. For the complete datasets used in the analyses, we find that the standard
deviations for WWI and WWI$'$ are similar to the power law case as well. In other words, we find tighter (or at least similar) constraints on background parameters with WWI and
WWI$'$ models as in power law model.
\begin{figure*}[!htb]
\begin{center}
\resizebox{142pt}{105pt}{\includegraphics{omegabh2_PLAW-eps-converted-to.pdf}}
\resizebox{142pt}{105pt}{\includegraphics{omegabh2_WWI-eps-converted-to.pdf}}
\resizebox{142pt}{105pt}{\includegraphics{omegabh2_WWII-eps-converted-to.pdf}}
\resizebox{142pt}{105pt}{\includegraphics{omegabh2_TT-eps-converted-to.pdf}}
\resizebox{142pt}{105pt}{\includegraphics{omegabh2_EE-eps-converted-to.pdf}}
\resizebox{142pt}{105pt}{\includegraphics{omegabh2_TEB-eps-converted-to.pdf}}
\end{center}
\caption{\footnotesize\label{fig:omegabh2}The likelihood of baryon density obtained from temperature and polarization data from Planck-2015. In the top panel, in each plot,
we provide the likelihood obtained assuming the power law, WWI and WWI$'$ form of inflation. Note that when we assume power law PPS, the E polarization data favors a large baryon
density ($\Omega_{\rm b}{h^2}\sim 0.024$) which looks having slight disagreement with the temperature data. WWI and WWI$'$ reduce this disagreement with a mean value of baryon density
$\Omega_{\rm b}{h^2}\sim 0.023$. In the bottom panel we compare the results from different inflationary models and power law PPS for different datasets. The shift of the likelihoods
to the lower baryon density is evident for WWI and WWI$'$ models for EE + lowTEB data.}
\end{figure*}
Though we are getting substantial improvement in likelihood compared to power law with both WWI and WWI$'$, we fail to get more than 1$\sigma$ evidence for these features. The maximum
likelihoods show significant deviation from featureless case in both the models, but when marginalized, we find that the best fits have low marginalized probabilities. Hence, with Planck
data we are unable to rule out featureless primordial power spectrum even though feature models agree with the data substantially better.
At this point we should mention that WWI-[b,c,d] and WWI$'$ contain features around the BAO scales and hence the matter power spectra can be constrained using the
large scale structure data as well. With the upcoming data from DESI~\cite{DESI}, SDSS-IV, e-BOSS~\cite{SDSSIV,eboss}, Euclid~\cite{Euclid} we expect to
verify the existence of these features. Also with the upcoming low-$\ell$ EE data from Planck HFI is supposed to provide stronger constraints on the large scale suppression. We expect to
revisit Wiggly Whipped Inflation with the future datasets from CMB and LSS.
\subsection{Non-Gaussianity}
\begin{figure*}[!htb]
\begin{center}
\resizebox{180pt}{100pt}{\includegraphics{F_NL_eq_WWI_a-eps-converted-to.pdf}}
\resizebox{180pt}{100pt}{\includegraphics{F_nl_2d_WWI_a_GNU-eps-converted-to.pdf}}
\resizebox{180pt}{100pt}{\includegraphics{F_NL_eq_WWI_b-eps-converted-to.pdf}}
\resizebox{180pt}{100pt}{\includegraphics{F_nl_2d_WWI_b_GNU-eps-converted-to.pdf}}
\resizebox{180pt}{100pt}{\includegraphics{F_NL_eq_WWI_c-eps-converted-to.pdf}}
\resizebox{180pt}{100pt}{\includegraphics{F_nl_2d_WWI_c_GNU-eps-converted-to.pdf}}
\resizebox{180pt}{100pt}{\includegraphics{F_NL_eq_WWI_d-eps-converted-to.pdf}}
\resizebox{180pt}{100pt}{\includegraphics{F_nl_2d_WWI_d_GNU-eps-converted-to.pdf}}
\resizebox{180pt}{100pt}{\includegraphics{F_NL_eq_WWIP-eps-converted-to.pdf}}
\resizebox{180pt}{100pt}{\includegraphics{F_nl_2d_WWI_p_GNU-eps-converted-to.pdf}}
\end{center}
\caption{\footnotesize\label{fig:fnl} Wiggly Whipped Inflation : $f_{\rm NL}$ in equilateral (left) and in arbitrary triangular configurations (right). From top to
bottom we plot the $f_{\rm NL}$ from WWI-a, WWI-b, WWI-c, WWI-d and WWI$'$ respectively (see, Fig.~\ref{fig:psk} for corresponding PPS). Note that while WWI-a provides $f_{\rm NL}\sim{\cal O} (1)$ which has a localized
feature around $k\sim0.002~{\rm Mpc^{-1}}$, features that extend over wide range of cosmological scales with larger frequency,
generate higher non-Gaussianity with $f_{\rm NL}\sim{\cal O} (100-1000)$.}
\end{figure*}
It has been extensively discussed in the literature that inflation models, that offers departure from slow roll inflation to generate features in the PPS,
also produce non-negligible non-Gaussianities~\cite{chen,Flauger-amm,Martin:2011sn,Hazra:2012BINGO,Sreenath:2014BINGO2,Arroja:2011,Arroja:2012,Martin:2014}.
In this paper we shall only consider the three point
correlations, {\it i.e.} the bispectra generated by WWI and WWI$'$. We use {\tt BINGO-2.0}~\cite{Hazra:2012BINGO,Sreenath:2014BINGO2,BINGOPAGE} to calculate the bispectra
for our best fit models. We have used the best fit potential parameters
from TTTEEE + lowTEB + BKP for WWI-[a,b,c,d] and WWI$'$ and compute the $f_{_{\rm NL}}$. We calculate all the terms contributing to the $f_{_{\rm NL}}$ arising from the interaction Hamiltonian,
cubic in order of the curvature perturbations~\cite{maldacena-2003,chen,Martin:2011sn}. Since WWI best fit PPS do have features with high frequency
and high amplitude, any slow-roll approximation in the
bispectrum integral shall underestimate the actual value of the $f_{_{\rm NL}}$. Using {\tt BINGO}, we do not make any approximation and evaluate the bispectrum integral
numerically. We calculate local $f_{_{\rm NL}}$ for our models derived from bispectrum ${\cal B}_{_{\mathrm{S}}}({\bf k}_{1},{\bf k}_{2},{\bf k}_{3})$,
\begin{eqnarray}
f_{_{\rm NL}}({\bf k}_{1},{\bf k}_{2},{\bf k}_{3})=-\frac{10}{3}\, (2\,\pi)^{-4}\; (2\,\pi)^{9/2}\;
k_{1}^3\, k_{2}^3\,k_{3}^3\; {\cal B}_{_{\mathrm{S}}}({\bf k}_{1},{\bf k}_{2},{\bf k}_{3})\nonumber\\
\times\left[k_1^{3}\; {P}_{_{\mathrm{S}}}(k_2)\;
{P}_{_{\mathrm{S}}}(k_3)
+{\mathrm{two~permutations}}\right]^{-1},
\end{eqnarray}
where ${P}_{_{\mathrm{S}}}(k)$ denotes the primordial power spectra.
In Fig.~\ref{fig:fnl} we plot the $f_{_{\rm NL}}$ for the best fit potentials obtained from TTTEEE + lowTEB + BKP datasets. To the left we plot the $f_{_{\rm NL}}$ in equilateral
limit ($k_1=k_2=k_3=k$). To the right we plot the 2D heat map of the $f_{_{\rm NL}}$. The 2D $f_{_{\rm NL}}$ are plotted as a function of $k_3/k_1$ and $k_2/k_1$. The top left corner of
the triangular configurations represent the squeezed limit ($k_2,k_3<<k_1$) and the top right corner represent the equilateral limit. $k_1$ in WWI-[a,b,c,d] and WWI$'$ are chosen
to be $\sim2\times10^{-3}$, $0.03$, $0.015$, $0.14$ and $0.14$ ${\rm Mpc^{-1}}$ respectively. For the first three cases, we chose the mode $k_1$ by locating the scale where $f_{_{\rm NL}}$ becomes
maximum and for WWI-d and WWI$'$ we chose it to be a smaller scale since for a sharp transition in the potential (or in its derivative in the latter case), the $f_{_{\rm NL}}$ is expected to diverge
linearly with wavenumber and a smaller $k_1$ is expected to capture the profile of the $f_{_{\rm NL}}$ in the 2D heat map.
Note that the WWI-a and WWI-c generates $|f_{_{\rm NL}}|\sim1-10$. Because of the presence of a broad step in the potential and wide frequency oscillations, the slow-roll parameters
$\epsilon_{i+1}={\rm d} \ln \epsilon_{i+1}/{\rm d} N$ and their derivatives are not large enough
to produce large three point correlations of curvature perturbations. While WWI-b, WWI-d and WWI$'$ generates higher $f_{_{\rm NL}}$ because of sharper transition from moderate fast roll to
slow-roll potential ($\Delta \rightarrow 0$). We would like to point a crucial difference between WWI-b and WWI-d (or WWI$'$) models. When we compare these two models with the temperature and polarization
angular power spectrum, we find similar improvement in fit compared to power law models. Hence, to the power spectra level, these models are nearly indistinguishable. On the other hand
the WWI-b generates $f_{_{\rm NL}}$ which converges at small scales but WWI-d and WWI$'$ diverges with the wavenumbers because the singularity at $\phi=\phi_{\rm T}$ (although, note that in strict
numerical sense, we have to model the singularity by a transition width very close to zero). The divergent $f_{_{\rm NL}}$ arising from the instantaneous transition and the effect of smoothing
the discontinuity were studied before in literature~\cite{Arroja:2011,Arroja:2012,Martin:2014}\footnote{We find $\Delta\sim10^{-3}$ for WWI-b and $10^{-5}$ for WWI-d}.
As we have pointed out before, the WWI$'$ essentially generates same power spectra as generated by Starobinsky-1992 model of inflation with a spectral tilt of $\sim0.96$.
The bispectra generated from this model, hence are of same shape as discussed in literature~\cite{Arroja:2011,Martin:2011sn,Arroja:2012,Martin:2014,Sreenath:2014BINGO2}.
To evaluate the $f_{_{\rm NL}}$ from the WWI$'$ model we have used a Theta function and Delta function with a width similar to WWI-d model, to reproduce the effects of the singularity in the
second derivative of the potential. We should emphasize that an instantaneous transition is not a realistic situation and there has to be a finite width associated with the
transition in the potential that ensures the convergence of bispectra at small scales (as have been emphasized in~\cite{Martin:2014}).
From the analyses above we can state that feature models, that are indistinguishable from the likelihood {\it w.r.t.} the power spectra data, a joint estimation of PPS and
bispectra can quantitatively be able to distinguish between the models, and also provide extra significance for primordial features in the data, if present.
There are hints of oscillatory bispectra from Planck 2015 analysis~\cite{Planck:2015NG} and hence it is important to confront the WWI features with Planck bispectra data.
We emphasize that since WWI, within its single framework, offers a wide variety of features, this model will be extremely useful in comparing different features.
In this context, note that, the models considered in our paper, oscillations in the PPS at large $k$ do not play a significant role for the explanation of the features
in the CMB multipole spectrum apart from WWI-d. This is in agreement with the conclusion of~\cite{Fergusson:2014tza} that there is no statistically significant signs of
sharp oscillatory features in the CMB spectrum and bispectrum.
\section{Conclusions}\label{sec:conclusions}
We show that the WWI framework that was introduced mainly to explain the BICEP2 and the Planck 2013 results in a single theoretical model, can
explain a wide variety of the primordial features that are obtained from direct reconstructions using CMB angular power spectrum data and also that
are motivated from high energy theories. Without using different theoretical models of inflation, WWI allows us to generate different types of PPS features
in a single framework, making it extremely suitable for primordial feature hunt and also in constraining background cosmological parameters marginalized over
a multitude of inflationary scenarios.
In this paper we confront Wiggly Whipped inflaton potentials allowing deviations from strict slow roll, against the latest Planck 2015 angular power
spectrum data of temperature and polarization and BICEP2/Keck B-mode polarization data.
WWI offers a simple transition from moderate fast roll to strict slow roll potential with/without the presence of a discontinuity/jump in the potential at the field value of phase transition.
We also present WWI$'$, in which a smooth part of the inflaton potential is that of the $R+R^2$ inflationary model~\cite{Starobinsky:1980te} in the Einstein frame (and which in turn represents a
special case of the $\alpha$-attractor model~\cite{alpha_attractor}) and discontinuity appears only in the first derivative of the potential.
We have been able to generate a wide class of primordial features that have been discussed in the literature within the frameworks of WWI and WWI$'$. Owing to the flexibility of the
WWI model, we are able to locate the local minima and possibly the global minima in the parameter space of WWI using only one potential. From the individual and joint analyses of T, E and B
polarization we identify 4 distinguishable features in the primordial power spectra of WWI, namely WWI-[a,b,c,d]. In all the
cases we notice a common pattern, the large scale suppression of scalar perturbation spectra. WWI-a provides a dip in the power spectra at $2\times10^{-3}{\rm Mpc^{-1}}$, which has been
discussed extensively in the literature and we find that this feature, apart from providing an improved fit to the low-$\ell$ likelihood (mostly from around $\ell=22$), also provides
better fit to high-$\ell$ EE datasets. We find localized oscillations (but wider than WWI-a) within $0.002-0.05~{\rm Mpc^{-1}}$ that particularly agree with high-$\ell$ EE data
compared to power law best fit. WWI-a and WWI-c represent the primordial features that attempts to address the features in the individual CMB angular power spectra data, that are not addressed
by standard power law PPS. Apart from these best fits, our analysis with WWI offers a third kind of feature where the primordial power spectrum do not provide notable improvement
in fit when compared with TT and EE angular power spectrum individually, but provides more than 13 improvement in fit in $\chi^2$ compared to power law {\it w.r.t.} the complete dataset. Similarly
in WWI$'$, using only 2 extra parameters in the inflation potential, we find $\sim12$ improvement in $\chi^2$ fit. WWI$'$ offers a step like suppression at larger scales accompanied by oscillations
at smaller scales. In this model we show that primordial feature of a fixed shape, by changing its location and amplitude, can match both temperature, polarization data separately and also in a combined analysis.
We find that the standard baseline model of cosmology prefers higher baryon density ($\Omega_{\rm b}h^2\sim0.024$ as best fit value) for E-mode polarization which is more than 3$\sigma$ away from the TT best fit.
Though the uncertainties in the E-mode polarization data are much larger than temperature data and the distance to the best fit values from EE data can not be trusted in a statistically robust analysis,
we demonstrated that TT, EE Likelihood decrease in a joint analysis.~\footnote{Here we would like to mention that this difference can also be an artifact of the systematics in the Planck polarization and
temperature data~\cite{Addison:2015wyg}} WWI-b, WWI-d and WWI$'$ offering high frequency oscillations extending over a large range of cosmological scales
offer a scenario that fits the TT, EE, EE and lowTEB data better in joint analyses, keeping the best fit value of $\Omega_{\rm b}h^2\sim0.0222$, as demanded by temperature
power spectrum. Both WWI and WWI$'$ show a shift in baryon density to a lower value ($\Omega_{\rm b}h^2\sim0.023$) when compared with EE dataset.
We find that WWI-b and WWI-d represent the global best fit to the complete Planck 2015 and BICEP2/KECK datasets. The fundamental difference between WWI-b and WWI-d is in the
sharpness of their transition from moderate fast roll to the complete slow roll regime. These two features are
indistinguishable from the angular power spectrum analyses but we show that they have very distinct bispectra signatures. It is possible to find stringent constraints on the WWI and WWI$'$ models
upon joint analyses with CMB power spectra and bispectra data.
\section*{Acknowledgments}
DKH and GFS acknowledge Laboratoire APC-PCCP, Universit\'e Paris Diderot and Sorbonne Paris Cit\'e (DXCACHEXGS)
and also the financial support of the UnivEarthS Labex program at Sorbonne Paris Cit\'e (ANR-10-LABX-0023 and ANR-11-IDEX-0005-02).
DKH would like to thank the hospitality of Cluster Computing Center (through the support from the DOE HEP's Forum on Computational Excellence)
and Berkeley Center for Cosmological Physics, LBL, Berkeley and Princeton University where a part of the work has been carried out. AS would like to acknowledge
the support of the National Research Foundation of Korea (NRF-2016R1C1B2016478). AAS was partially supported by the grant RFBR 14-02-00894 and by the Russian
Government Program of Competitive Growth of Kazan Federal University.
|
\section{Introduction}
In black-hole (BHXB) and neutron-star (NSXB) X-ray binaries, the transfer
of mass from the optical companion to the compact star is mediated by an
optically thick, geometrically thin, accretion disk. Because the mass
to radius ratio in neutron stars and black holes is similar\footnote{A 5$~{\rm M}_\odot$
black hole has a Schwarzchild radius of $\sim$ 15 km.}, the characteristic
velocities and dynamical time scales near the compact object are also
similar ($v\sim 0.5c$, $t \simless 1$ ms). Thus, it is not
surprising that BHXB and NSXB show very similar X-ray spectral and
temporal characteristics
\citep{vanderklis94,maccarone03,vanderklis05,barret04,disalvo06,munoz14}.
It is generally accepted that the X-ray spectral continuum of BHXB and
NSXB results from the sum of two main components: one thermal component
which follows a blackbody distribution and another non-thermal component
that is best described as a power law. The thermal component dominates at
lower energies, hence it is referred to as the soft component, while the
power-law component extends up to a few hundred keV and it is often known as
the "hard tail" \citep{barret04,mclintock06,lin07,done07,church12}.
In BHXB, the soft thermal component is attributed to the emission from
the innermost, hotter regions of the accretion disk, while the hard power
law results from Comptonization of low-energy photons by energetic
electrons. The physical origin of the Comptonizing medium could be an
advection-dominated accretion flow \citep{narayan94,esin97}, a low angular
momentum accretion flow \citep{ghosh11,garain12}, or a radio jet
\citep{band86,georganopoulos02,reig03, giannios05, markoff05}.
In NSXB, the emission spectrum is more complex owing to the presence of a
hard surface. Thus below $\sim$ 10 keV, NSXB may display two soft
components: a single-temperature blackbody from the neutron star surface
or from the boundary layer between the disk and the neutron star and a
multi-temperature blackbody from the accretion disk
\citep{mitsuda84,gilfanov03}. The temperature of the neutron star photons
is expected to be higher ($kT\sim 2-3$ keV) than that of the disk soft
photons ($kT\sim 1$ keV), because the effective area of the neutron star
surface is more compact than that of the accretion disk
\citep{farinelli07}. Above $\sim$ 20 keV, the most prominent spectral
feature in the X-ray spectrum of NSXB is a power-law distribution (the hard
tail), that usually extends up to 200--300 keV without evidence for a
cutoff
\citep{disalvo00,disalvo01,iaria01,damico01,farinelli05,migliari07,ding11}.
Both kinds of binary systems exhibit distinct spectral states in response
to changes in the mass accretion rate, for example, in the course of an
outburst. The main of these states are generally referred to as soft and
hard, depending on whether the bulk of the luminosity is radiated below or
above a few keV, respectively.
Another common property to both BHXB and NSXB is the presence of radio jets
\citep{fender04,migliari06}. Optically thick, compact, steady radio
emission is detected during the low (in the 1-20 keV band) X-ray lumonisity
hard state, while optically thin radio flares occur during transitions from
the hard to the soft sate. The jet is quenched in the soft state, when
the X-ray luminosity approaches the Eddington limit
\citep{paizis06,migliari07,miller-jones10}.
A useful way to describe the rich phenomenology exhibited by these systems
in the X-ray band is the hardness-luminosity diagram (HLD). In the HLD, BHXB
and NSXB display distinct curves, along which the sources move smoothly
as the mass accretion rate varies.
In Kylafis \& Belloni (2015), a physical inerpretation was offered for the
phenomenology of BHXB along the q-shaped curve in the HLD.
In NSXB, the shape of the curve in the HLD devides them
into $Z$ sources and {\it atoll} sources
\citep{hasinger89,vanderklis06}. The three branches that form the Z-shaped
HLD are called horizontal (HB), normal (NB), and flaring (FB) branches.
The HB corresponds to the hardest X-ray state, while the FB is the soft
state. In the case of atoll NSXB, the branches are called island and banana
branches, with subcategories such as extreme island and lower and upper
banana. The extreme island branch corresponds to the hardest spectral sate.
The hard tail has been detected in almost all the currently known Z
sources \citep{disalvo02} and many of the atoll sources
\citep{paizis06}.
The association of the compact radio jet and the X-ray hard tail with a
specific region of the source in the HLD is well documented in BHXB
\citep{fender09} and NSXB \citep{paizis06,migliari06}. Both the radio
emission and the strength of the hard tail become weaker at higher
accretion rates. Radio and hard X-rays show the strongest intensity in the
hard states of BHXB, the horizontal branch of $Z$-NSXB
\citep{hjellming90b,migliari07,dai07}, and the island or lower banana
states in {\it atoll}-NSXB \citep{migliari03}. In BHXB, radio emission is
absent in the soft sate. In some NSXB, radio emission may be still detected
in the softer states, albeit highly reduced \citep{homan04}.
While there is a general consensus that the radio emission is produced by a
compact jet, the physical origin of the hard tail is not well understood.
It naturally arises from Comptonization, however, the details of how the
electrons acquire their energy (thermal, non-termal, or bulk motion) and
the source of seed photons (neutron-star surface, inner accretion disk,
synchrotron photons) remain unclear
\citep{barret00,disalvo06,paizis06,lin07,markoff05,farinelli09,revnivtsev14}.
A weak hard tail contributing a small percentage of the total flux
($\simless 3$\%) is also detected in the soft state, when the jet is
quenched or absent \citep[see, e.g.,][and references therein]{mclintock06}.
In this state, the power-law tail is thought to come from Comptonization in
non-thermal flares above and below the thermal disk
\citep{Poutanen97,gierlinski99}.
In a series of papers
\citep{reig03,giannios04,giannios05,kylafis08,reig15}, we showed that
Compton upscattering of soft photons from the accretion disk in the jet can
explain a number of observational relations between the spectral and timing
parameters in the hard state of BHXB. Our results clearly demonstrate that
jets play a central role in all the observed phenomena, not only in the
radio emission. Motivated by the similarities of the X-ray spectral
continuum in BHXB and NSXB and {\it especially by the clear connection of
the presence of a hard tail with a radio-loud state of the source}, we
investigate whether our model can also account for the spectral properties
of NSXB.
Our objective in this paper is to demonstrate that
emission from the neutron-star surface plus Comptonization in a jet
can reproduce the observed spectra of bright NSXB, namely a soft thermal
component that is well described by a blackbody distribution with $kT=2-3$
keV and a hard tail that follows a power-law with photon index in the range
1.8--3 with no evidence for a cutoff up to 200--300 keV.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{./fig1.eps}
\end{center}
\caption[]{Emergent spectra from the Monte Carlo simulations. The models
plotted correspond to a fixed $\gamma_{\rm min}=3.34$ ($v_{\perp}=0.52c$)
and optical depth $\tau=0.1,0.25,0.5,0.75,1.0,1.25,1.5$. The input source of photons
follows a blackbody distribution with $kT=2.5$ keV.
The normalization of the input blackbody
was chosen so that the hard component contributes
about 5\% of the 0.1-300 keV luminosity. In the energy range covered by
current missions ($<300$ keV) the hard tail shows no break.}
\label{fig1}
\end{figure}
\section{The model}
\label{model}
The model that we have used in this work is identical to that used in
\citet{reig15}. For our Monte Carlo code we follow
\citet{pozdnyakov83}. Photons from a blackbody distribution of
temperature $kT_{bb}$ are injected at the
base of the jet with an upward isotropic distribution. As the photons
travel through the medium, they experience Compton scatterings with the
spiraling electrons. In each scattering, the photons gain
on average energy from the bulk
motion (i.e. $v_{\parallel}$ and $v_{\perp}$) of the electrons.
Comptonization can
occur everywhere in the jet. The optical depth to electron scattering, the
energy change and the new direction of the photons after scattering are
computed using the corresponding relativistic expressions. Each model is
run for $10^7$ photons. More details of how the code works can be found in
\citet{kylafis08}.
Since we are not interested in reproducing the radio
spectrum \citep[see][]{giannios05}, we have assumed mono-energetic
electrons in the jet with Lorentz factor equal to the smallest in the
distribution, namely
\begin{equation}
\label{lorenzt}
\gamma_{\rm min}=\frac{1}{\sqrt{1-(v_{\parallel}^2+v_{\perp}^2)/c^2}},
\end{equation}
\noindent where $v_\parallel = v_0 =$ constant
is the terminal velocity of the jet and $v_\perp$ is the
smallest peprendicular velocity of the electrons in the lab frame.
The flow velocity in the jet is given by
\begin{equation}
\label{accel}
v_{\parallel}(z) =
\begin{cases}
v_0 ~ (z/z_1)^p & \text{if } 0< z \leqslant z_1\\
v_0 & \text{if } z> z_1,
\end{cases}
\end{equation}
\noindent where $z_1$ and $p$ are parameters. In other words, the jet has an
acceleration region of thickness $z_1$, beyond which the flow has
constant velocity $v_0$.
For a parabolic jet, i.e. one whose radius at height $z$ is
$R(z)=R_0(z/z_0)^{1/2}$, the electron density is obtained from the
continuity equation and for $z>z_1$ it is inversely proportional to $z$.
The fixed parameters of our models and their reference values are: the
radius $R_0 = 50 \, R_{\rm NS}$ of the base of the jet, where $R_{\rm
NS}=1.25 \times 10^6$ cm is the radius of the neutron star, the distance
$z_0 = 1 \, R_{\rm NS}$ of the bottom of the jet from the neutron
star center, the
height $H=10^5 \, R_{\rm NS}$ of the jet, the terminal velocity $v_0= 0.8
c$ of the jet, the thickness $z_1= 5 \, R_{\rm NS}$ of the acceleration
zone, the exponent $p=1/2$, and the temperature $kT_{bb} = 2.5$ keV of the
soft-photon input.
The parameters of our model that we have varied are the Thomson optical
depth $\tau_{\parallel}$ along the axis of the jet and the minimum Lorentz
factor $\gamma_{\rm min}$. Since $v_0$ is a constant in our models, the
variation of $\gamma_{\rm min}$ is equivalent to a variation in
$v_{\perp}$.
Because the jet is relativistic, the results also depend on the angle
$\theta$ of observation with respect to the jet axis. Here we consider an
intermediate range of observing angles $0.2 < \cos \theta < 0.6$.
Practically, for the Monte Carlo simulation this means that we count only
photons that leave the jet in this range of angles.
\begin{figure}[!t]
\centering
\includegraphics[width=8cm]{./fig2.eps}
\caption{Photon index and cutoff energy as functions of $\tau_{\parallel}$
and $\gamma_{\rm min}$. On the left panels, each point represents a
calculation with the same $\gamma_{\rm min}=3.34$ ($v_{\perp}=0.52c$)
and different $\tau_{\parallel}$. On the right panels, each point
represents a calculation with the same $\tau_{\parallel}=1$ and different
$\gamma_{\rm min}$. The scale of the Y-axis was left the same to
faciliate the comparison.}
\label{fig2}
\end{figure}
\begin{table}
\caption{Results of our Monte Carlo simulations. Different models
correspond to different values of the optical depth ($\tau_{\parallel}$)
and perpendicular component of the electron velocity ($\gamma_{\rm min}$).
All the remaining parameters of the model were fixed at the reference
values (see Sect.~\ref{model}).}
\label{simu}
\begin{center}
\begin{tabular}{cccc}
\hline \hline
$\tau_{\parallel}$ &$\gamma_{\rm min}$/$v_{\perp}$ (c) &$\Gamma$ &$E_{\rm c}$ (keV) \\
\hline
0.10 &3.015/0.50 &3.30 &133 \\
0.10 &3.164/0.51 &3.31 &172 \\
0.10 &3.341/0.52 &3.28 &223 \\
0.10 &3.556/0.53 &3.26 &314 \\
0.10 &3.824/0.54 &3.27 &571 \\
0.25 &3.015/0.50 &2.82 &151 \\
0.25 &3.164/0.51 &2.82 &195 \\
0.25 &3.341/0.52 &2.81 &253 \\
0.25 &3.556/0.53 &2.79 &355 \\
0.25 &3.824/0.54 &2.79 &566 \\
0.50 &3.015/0.50 &2.44 &173 \\
0.50 &3.164/0.51 &2.44 &219 \\
0.50 &3.341/0.52 &2.43 &288 \\
0.50 &3.556/0.53 &2.42 &392 \\
0.50 &3.824/0.54 &2.41 &588 \\
0.75 &3.015/0.50 &2.21 &187 \\
0.75 &3.164/0.51 &2.20 &238 \\
0.75 &3.341/0.52 &2.20 &309 \\
0.75 &3.556/0.53 &2.19 &422 \\
0.75 &3.824/0.54 &2.18 &604 \\
1.00 &3.015/0.50 &2.01 &195 \\
1.00 &3.164/0.51 &2.02 &251 \\
1.00 &3.341/0.52 &2.01 &322 \\
1.00 &3.556/0.53 &2.01 &438 \\
1.00 &3.824/0.54 &2.00 &628 \\
1.25 &3.015/0.50 &1.88 &205 \\
1.25 &3.164/0.51 &1.89 &262 \\
1.25 &3.341/0.52 &1.89 &344 \\
1.25 &3.556/0.53 &1.88 &461 \\
1.25 &3.824/0.54 &1.88 &665 \\
1.50 &3.015/0.50 &1.73 &210 \\
1.50 &3.164/0.51 &1.74 &268 \\
1.50 &3.341/0.52 &1.74 &347 \\
1.50 &3.556/0.53 &1.73 &457 \\
1.50 &3.824/0.54 &1.73 &640 \\
\hline
\hline
\end{tabular}
\end{center}
\end{table}
\section{Results}
\label{results}
The spectral parameters relevant to the present study are the photon index
of the hard-tail component and the cutoff energy. We have investigated the
dependence of these two quantities on the choice of two parameters of our
model, namely the optical depth and the electron velocity (see
Table~\ref{simu}).
Figure \ref{fig1} shows the emerging spectra that result from our
simulations using the reference values of the parameters reported in
Sect.~\ref{model} with $\gamma_{\rm min}=3.34$ ($v_{\perp}=0.52c$) and
various optical depths $\tau_{\parallel}$ = 0.1, 0.25, 0.5, 0.75, 1.0,
1.25, and 1.5. As it can be seen, although the spectra roll over above a
certain energy, this energy is well above 200-300 keV. Thus, up to 200 -
300 keV, the spectra are essentially pure power-laws with photon index
between 1.8--3, as observed. Figure \ref{fig2} shows the
photon index $\Gamma$ and the cutoff energy $E_{\rm c}$ as functions of
optical depth $\tau_{\parallel}$ and $\gamma_{\rm min}$ (or equivalently
$v_{\perp}$). The scale of the vertical axis was kept the same in the two
plots to better assess the differences. The cutoff energy has a weak
dependence on $\tau_{\parallel}$ for the range of optical depths
considered, but it is strongly dependent of $\gamma_{\rm min}$ . This is
expected because the cutoff is mainly determined by the energetics of the
electrons \citep{giannios05}. In contrast, the photon index strongly
depends on $\tau_{\parallel}$, but only weakly on $\gamma_{\rm min}$. The
reason is that the slope of the spectrum is very sensitive to the number of
scatterings, which increases with optical depth.
Because of the significant flow velocity in the jet, the up-scattered
photons escape preferentially in the forward direction. Thus we expect the
results to depend also on the angle $\theta$ of observation with respect to
the jet axis. The hardest spectra are found at small observation angles, as
these are the photons that have suffered the most energetic scatterings. The
cutoff energy increases as the observation angle decreases, but stabilizes
at intermediate angles. However, once the angle range is fixed, the
parameters follow a similar trend with optical depth and Lorentz gamma
factor, irrespective of the range considered.
We have also computed the time lag of hard photons in the energy range
7--40 keV with respect to softer photons in the range 2--7 keV. The
time-lag spectrum, i.e. the time lag as a function of Fourier frequency, is
shown in Fig.~\ref{fig3}. The spectra follow a power-law dependence on
Fourier frequency $t_{\rm lag} \sim \nu^{-\beta}$, with $\beta =
1.0\pm0.1$. This slope is somewhat steeper than the average slope measured
in black-hole binaries, $t_{\rm lag}\sim \nu^{-\alpha}$ with $\alpha =
0.7\pm0.1$ \citep{nowak99,pottschmidt03,cassatella12}, but agrees with the
results reported by \citet{olive01}. We have not found a significant
dependence of the shape of the lag spectrum with either the optical depth
or $\gamma_{\rm min}$.
\begin{figure}[!t]
\centering
\includegraphics[width=8cm]{./fig3.eps}
\caption{Time-lag spectrum that results from the model with
$\tau_{\parallel}=1$ and $\gamma_{\rm min}=3.34$. All the remaining parameters
of the model were fixed at the reference values (see Sect.~\ref{model}).}
\label{fig3}
\end{figure}
\section{Discussion}
We have run Monte Carlo simulations of the Compton upscattering of soft
photons in an outflow moving at relativistic velocity (representing a radio
jet) and computed the emerging X-ray spectrum. The input parameters of the
model have been selected so that the results match what is seen in
bright NSXB. Assuming a blackbody input source of photons with $kT$=2.5
keV, our model reproduces very well the observed 1-300 keV X-ray spectrum
in the harder states, a power law continuum with a photon index $\Gamma$ in
the range 1.8--3 without evidence for a cutoff up to 200--300 keV.
The main difficulty of the models that seek to reproduce the
high-energy X-ray spectral
continuum in NSXB lies in explaining the absence of a cutoff, because the
spectrum resulting from Comptonization is expected to roll over at a certain
energy (an effect known as Compton recoil), when the energy of the photons
approaches that of the electrons. Various models have been put forward to
explain the absence of a cutoff in the hard tails: bulk motion
Comptonization \citep{bradshaw03,paizis06,farinelli07,farinelli08}, Comptonization by a
hybrid--thermal-non thermal corona
\citep{gierlinski99,ozel00,zdziarski01,farinelli05,revnivtsev14}, or
synchrotron emission from the electrons of a jet \citep{markoff05}. Because
of the limited sensitivity of current detectors at energies above $\sim$100
keV, statistically significant detections are given up to $\sim$200 keV,
typically. The spectral continuum could in principle still roll over at a
certain energy provided that this cutoff energy is larger than the limit
energy of the detectors\footnote{From the point of view of the energetics,
if the slope of the power law is $\Gamma \leq 2$, then the spectrum must
show a rollover at certain energy to prevent a diverging luminosity.}. Recently,
\citet{revnivtsev14} reported the highest lower limit on the cutoff energy
that has been detected so far: $E_{\rm c}> 330$ keV, at a $2\sigma$
significance level.
This relatively high lower limit of the cutoff energy was used by
\citet{revnivtsev14} to rule out the bulk-motion Comptonization model. The
reason is that the cutoff energy in thermal and bulk-motion Comptonization
depends on the temperature and the velocity of the Comtonizing electrons,
respectively. For realistic values of $kT_e < 50$ keV and assuming
free-fall velocity $v_{\rm ff} \approx 0.5 c$
onto a neutron star of canonical mass ($1.4~{\rm M}_\odot$) and
radius ($10^6$ cm), the cutoff energy is expected
to be below 200 keV. Moreover, the radiation pressure expected to result
from the emission of
the neutron-star surface may reduce the infall velocity to less than
$0.2c$, decreasing the bulk motion Comptonization effect \citep{farinelli08}.
Observations collected over the past decade show that there exists a clear
connection between the X-ray spectral states and the radio emission in
NSXB. Similarly to BHXB, the NSXB are radio loud in the hard state,
usually displaying a flat spectrum indicative of a relativistic jet. The
radio emission is quenched when the source enters the soft state in most
(but not all) cases
\citep{penninx88,hasinger90,hjellming90b,oosterbroek94,migliari06,migliari07,miller-jones10}.
Motivated by the disputed origin of the hard X-ray tail in NSXB and by the
clear relationship between the presence of this hard component and radio
emission from the source in the form of a jet, we have investigated whether
the observed X-ray spectrum of NSXB can be reproduced by our jet model. Our
goal has been to reproduce the observed spectrum above $\sim$20 keV,
namely, a power-law distribution with a photon index in the range 1.8--3
and the absence of a high-energy cutoff up to an observed energy of 300
keV. In Sect.~\ref{results}, we demonstrated that such a range of photon
indices can be obtained with an optical depth varying in the range $0.1
\le \tau_{\parallel} \le 1.5$ (Table~\ref{simu}). The lack of a cutoff in
the observed spectra translates into cutoff energies above 200 keV in our
models. These energies are obtained when the electrons move with
$\gamma_{\rm min} \simmore 3.2$ ($v_{\perp} \simmore 0.52 c$, for
$v_{\parallel} = 0.8 c$). Note that, in principle, and contrary to the bulk
motion Comptonization model, there is no restriction on the velocity
components of the electrons in the jet, provided that the parallel
component is significantly larger than the perpendicular component.
Therefore, although we selected models that give cutoff energies above but
not too far from 300 keV, we could increase our limit on $E_{\rm c}$ simply
by increasing $\gamma_{\rm min}$ (see Fig~\ref{fig2}). Note that such high
$\gamma_{\rm min}$ is possible, since \citet{fomalont01b}, who studied the
evolution of the radio emission in Sco X--1, found that $v > 0.95 c$.
\subsection{Comparison with black-hole binaries}
There are two major differences between the jets in NSXB and BHXB: {\em i)}
the radio luminosities of neutron-star jets are typically $\sim$30 times
lower than those of black-hole jets at comparable X-ray luminosities
\citep{fender01, migliari03} and {\em ii)} radio emission is not completely
quenched in the soft state of some NSXB \citep{migliari04}, as opposed to
their non-detection in BHXB, when they are in the same state.
Both of these differences have been addressed successfully in
\citet{kylafis12}, where the formation and the destruction of jets is
explained using the Cosmic Battery
\citep{contopoulos98,contopoulos06,christodoulou08}.
That the radio luminosity is significantly lower in neutron-star
jets than in black-hole ones can be understood as follows:
while the radio luminosity scales as $\dot m M$ for both types
of compact object (here, $\dot m$ is the mass accretion rate
in units of the Eddington value and $M$ is the mass of the compact object),
the X-ray luminosity
due to the ADAF-like inner accretion flow scales as $\dot m M$ for
neutron stars and as $\dot m^2 M$ for black holes
\citep{narayan97, migliari06, abramowicz13}.
Thus, for comparable X-ray luminosities, the radio luminosity of
neutron-star jets is $\sim \dot m$ times smaller than that of black-hole ones.
The radio emission is quenched in the soft state of BHXB, because the
Cosmic Battery works very inefficiently when the accretion flow is in the
form of a Shakura-Sunyayev type \citep{shakura73}. On the other hand, in
low-mass NSXB with weak magnetic field ($B\sim 10^{8}-10^{9}$ G), the
Cosmic Battery can work moderately efficiently due to the ``spreading
layer'' \citep{inogamov10} that forms on the surface of the neutron star,
provided that there is a significant difference between the spin frequency
of the neutron star and the Keplerian frequency of the disk.
The findings from our calculations are consistent with the characteristics
of neuron-star and black-hole jets discussed above. We have found that
in NSXB the required optical depths are in the range $0.1 \le
\tau_{\parallel} \le 1.5$, while in BHXB they are in the range $1 \le
\tau_{\parallel} \le 10$ \citep{reig03,kylafis08, reig15}. Because the
optical depth is proportional to density and so is the radio emission, the
lower optical depth in NSXB is consistent with the observational fact that
the radio emission of the jet in NSXB is significantly smaller than that in
BHXB.
To explain the fact that BHXB exhibit a high-energy cutoff in the X-ray
spectrum while NSXB do not, we were forced to assume that $\gamma_{\rm
min}$ in the jets of NSXB is larger than in BHXB. In particular, we found
from our calculations that for BHXB $\gamma_{\rm min} \simless 2.4$
\citep{reig15}, while for NSXB $\gamma_{\rm min} \simmore 3$. It will be
interesting to see if this is confirmed by observations.
In a recent paper \citep{koljonen15}, an important anti-correlation was
found between the break frequency $\nu_b$ in the jet spectrum (spectral
energy distribution) and the power-law index $\Gamma$ of the hard X-ray
spectrum for BHXB and AGN. If the X-ray spectrum is produced in the jet by
inverse Compton scattering \citep{reig03, giannios04, giannios05,
kylafis08, reig15}, then this anticorrelation can be understood at least
qualitatively. As the jet weakens, $\tau_{\parallel}$ decreases, and as a
result $\Gamma$ increases \citep{reig15} and $\nu_b$ decreases. If our
model is correct, i.e., the hard X-ray spectrum is produced in the jet by
inverse Comptonization, then we {\em predict that NSXB as a class will
exhibit lower values of $\nu_b$ than BHXB}.
In our model, time or phase lags result from the random walk of the
photons in the scattering medium. Hard photons scatter on average
more times than softer ones
before escaping the Comptonizing medium, i.e. the jet, hence they
are delayed with respect to the softer ones. While the amplitude of the lags
depends on the optical depth and the dimensions of the jet
\citep{giannios04,kylafis08}, the slope of the time-lag spectrum
appears to be rather insensitive to those parameters or the Lorentz factor.
In practice, the slope is determined for given energy ranges of the soft
and hard photons and calculated for a given frequency range. Although most of
our models produce slopes $\sim$ 1, smaller indices, similar to those of
BHXB, could be obtained by different combinations of frequency and energy
ranges. Given the lack of publications reporting Fourier time-lag spectra
of NSXB, we cannot conclusively establish whether the different slopes that
our calculations suggest corresponds to a distinguishing feature between
NSXB and BHXB.
\section{Conclusion}
We have run Monte Carlo simulations of the Compton upscattering of soft
photons in an outflow moving at relativistic velocity (representing a radio
jet) and computed the emerging X-ray spectrum of bright NSXB. Our Monte
Carlo simulations reproduce the observed spectral features: a power-law
energy distribution at energies above 20 keV with photon index in the range
1.8--3 and the absence of a high energy cutoff up to $\sim$300 keV. We
demonstrate that the hard tail detected in these syetms can be explained by
Comptonization in the jet. The observed connection between the hard X-ray
tail and the radio emission in low-mass X-ray binaries with neutron star
companions supports this scenario. We explain the differences in the
spectral and timing parameters between BHXB and NSXB by assuming that the
electron population in the jets of NSXB have higher velocities and lower
densities than in BHXB.
\begin{acknowledgements}
This research has been supported in part by the
"RoboPol" project, which is implemented under the "ARISTEIA" Action of the
"OPERATIONAL PROGRAM EDUCATION AND LIFELONG LEARNING" and is co-funded by
the European Social Fund (ESF) and National Resources.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
One aspect of the classical theory of Sturm-Liouville problems is concerned with the existence of trace formulae relating,
for instance, a regularized (infinite) sum of eigenvalues to the potential. Given the problem
\begin{equation}
\label{stlibdd}
\left\{
\begin{array}{l}
-u''(x) + q(x) u(x) = \lambda u(x)\eqskip
u(0) = u(\pi) = 0
\end{array},
\right.
\end{equation}
the classical example due to Gelfand and Levitan in $1953$~\cite{gele} reads as follows
\begin{equation}
\label{classtr}
\sum_{k=1}^{\infty}\left[\lambda_{k}-k^2-\frac{\ds 1}{\ds \pi}\dint_{0}^{\pi} q(t)\dx{t}\right]=
\frac{\ds 1}{\ds 2\pi}\dint_{0}^{\pi} q(t)\dx{t} - \frac{\ds q(0) + q(\pi)}{\ds 4},
\end{equation}
where $\lambda_{k} = \lambda_{k}(q)$ denotes the $k^{\rm th}$ eigenvalue corresponding to the potential $q$.
From the Weyl asymptotics for the eigenvalues of~(\ref{stlibdd}), namely,
\begin{equation}
\label{asymp}
\lambda_{k} = k^2 + \frac{\ds 1}{\ds \pi}\dint_{0}^{\pi} q(t)\dx{t} + {\rm{O}}(k^{-2}),
\end{equation}
we have the convergence of the series on the left-hand side of~\eqref{classtr}.
This and other examples of this type of identity may be found in the books by Levitan and Sargsjan~\cite{lesa,lesa2} (see also~\cite{hakr})
and have more recently been shown to have extensions to the case of the perturbed harmonic oscillator on the whole real line -- see~\cite{puso}
and the references therein.
The main purpose of this paper is to show that these trace identities are limiting situations of inequalities satisfied by finite sums
of eigenvalues. As we shall see below, in the case of equation~(\ref{stlibdd}) and the corresponding trace formula~(\ref{classtr}),
if we expand the potential $q(x)$ in a Fourier series
\begin{equation}
\label{fourier}
q(x) = \frac{\ds q_{0}}{\ds 2} + \sum_{k=1}^{\infty} q_{k} \cos(k x),
\end{equation}
where
\begin{equation}
\label{eq:fourier-int}
q_{k} = \frac{\ds 2}{\ds \pi}\dint_{0}^{\pi}q(x) \cos(kx) \dx{x}, \;\; k=0,1,\dots,
\end{equation}
we actually have
\begin{equation}
\label{eq:ex}
\sum_{k=1}^{n}\left[\lambda_{k}-k^2-\frac{\ds 1}{\ds \pi}\dint_{0}^{\pi} q(t)\dx{t}\right]\leq
-\frac{\ds 1}{\ds 2}\sum_{k=1}^{n} q_{2k},
\end{equation}
(see Theorem~\ref{th:finitetracedir}) and it is not very difficult to check that the right-hand side converges to the right-hand
side of~(\ref{classtr}). We also show that there exists no lower bound for the left-hand side of~\eqref{eq:ex}
depending on a finite number of Fourier coefficients alone -- see Remark~\ref{nolowerbound}. It is, however, possible to use our results in combination with an argument given in~\cite{dikii} to obtain an alternative proof of~\eqref{classtr} and we explore this approach in Section~\ref{sec:trace}.
Such inequalities, including for other problems such as the case of Neumann boundary conditions and $N$-dimensional
flat tori, in fact follow from elementary test function arguments. The general principle is that if the potential $q$ is written out
as a Fourier series as in \eqref{fourier}, the eigenfunctions of the corresponding zero-potential problem (in one dimension, sines and
cosines) behave well when tested against the resulting series. The resulting sharp inequalities in terms of the Fourier coefficients
of $q$, which are completely explicit if $q$ is known, also contain interesting special cases under various less explicit assumptions
on $q$ (cf.~Theorems~\ref{th:app} and \ref{torus}).
We remark that the case of (non-periodic) Schr\"odinger operators on the whole line may be addressed via similar methods and this
will appear in another paper~\cite{frke}.
Although the underlying principle is quite simple, surprisingly, its application in this context appears to be new and we have not been able
to find any similar results in the literature. The earliest results regarding \emph{finite} sums of eigenvalues seem to be those for sums of
reciprocals of eigenvalues of the Laplacian, typically on inhomogeneous membranes, such as the classical work of P\'olya and Schiffer in
$1954$~\cite[Chap.~III]{posc}; see also, for example, \cite{laug,lapu}. In higher dimensions, there are bounds for sums of eigenvalues and
spectral zeta functions (cf.~\eqref{speczeta}) of the Laplacian based on the geometry of the underlying domain, such as in \cite{dittm,lasi1,lasi}.
For Schr\"odinger operators, besides the vast literature on Lieb--Thirring inequalities (which are concerned with rather different issues
and bounds from \eqref{eq:ex}; see, e.g., \cite{lawe}), there are well-established bounds on the number of eigenvalues less that a given
positive constant as in \cite{liya}, which also cover the case of Schr\"odinger operators on domains; see also \cite{flm} and the references
therein. However, these are all of a fundamentally different nature from \eqref{eq:ex}, typically involving either estimates purely in terms
of eigenvalues, eigenfunctions and/or geometric or dimensional quantities, or else integral
expressions for $q$ such as of the form $\int_\Omega (q+\alpha)^{N/2}_- \dx{x}$ for an arbitrary given constant $\alpha \geq 0$ and
dimension $N \geq 3$, as in \cite{liya}.
This paper is organized as follows. Section~\ref{sec:prelim} lays out the general setting and notation. In Section~\ref{sec:finite}, we
consider summation bounds of the form of~\eqref{eq:ex} for eigenvalues of Schr\"odinger operators on finite intervals subject to either
Dirichlet or Neumann boundary conditions (see Theorems ~\ref{th:finitetracedir} and~\ref{th:finitetraceneu}, respectively). We also sum
the one-dimensional Dirichlet and Neumann bounds in Theorem~\ref{finitetracecomb}, obtaining a bound
independent of the Fourier coefficients of $q$, and a generalization to ``zeta function''-type bounds, where powers of eigenvalues are
considered (see Theorem~\ref{th:sumpower}).
We then apply Theorems ~\ref{th:finitetracedir} and~\ref{th:finitetraceneu} to a particular class of potentials in Section~\ref{sec:app} (see Theorem~\ref{th:app}),
obtaining a simplified eigenvalue bound in dimension one under additional assumptions on the potential $q$, which include the case where
$q$ is convex. We show that one can essentially trivially obtain a similar bound on flat tori
in Section~\ref{sec:torus}.
In Section~\ref{sec:trace} we consider the relationship between our finite bounds and (known and potential) trace formulae. We show in particular how our one-dimensional results (Theorems ~\ref{th:finitetracedir} and~\ref{th:finitetraceneu})
can be used to replace some of the arguments of Diki\u{\i}'s proof \cite{dikii} of the Gelfand--Levitan formula \eqref{classtr}, which are arguably more natural than their counterparts in \cite{dikii}.
The final Sections~\ref{sec:zeta} and~\ref{sec:equality} are appendiceal, the former giving generic result which allows our bounds to be generalized
to powers of eigenvalues (i.e.~zeta functions), and the latter proving the sharpness of (the finite versions of) our inequalities: roughly speaking,
equality can be achieved in the inequalities for finite sums only if the potential is constant.
\section{Notation and preliminaries\label{sec:prelim}}
We will consider the general Schr\"odinger eigenvalue equation
\begin{equation}
\label{lapln}
-\Delta u(x) + q(x) u(x) = \lambda u(x),
\end{equation}
where $\Delta = \sum_{i=1}^N \frac{\partial^2}{\partial x_i^2}$ is the Laplace operator on either a finite interval ($N=1$) or an $N$-dimensional torus.
In either case we understand the problem in the usual weak sense. If $q\equiv 0$, \eqref{lapln} reduces to the usual
Laplacian eigenvalue problem, i.e.~the Helmholtz equation. For general $q \not\equiv 0$, unless otherwise specified we
make the standing assumption throughout the paper that $q$ may be expanded as an absolutely convergent Fourier series in terms of the eigenfunction
of the Helmholtz equation (cf.~\eqref{fourier}).
Under these assumptions the operator associated with \eqref{lapln} admits a discrete sequence of eigenvalues
$\lambda_1(q) < \lambda_2(q) \leq \ldots \to \infty$, which we repeat according to multiplicities. We will often abbreviate $\lambda_n(q)$ as
$\lambda_n$ if there is no danger of confusion, and we will write $\mu_n(q)$ (or $\mu_n$) instead of $\lambda_n(q)$ for the Neumann eigenvalues.
An (arbitrary) eigenfunction associated with $\lambda_n(q)$ or $\mu_n(q)$ will be denoted by $\varphi_n$. For the case of a zero potential,
that is, the ordinary Laplacian or Helmholtz
equation, we will in general write $\lambda_n(0)$ (or $\mu_n(0)$ as appropriate) for the $n^{\rm th}$ ordered eigenvalue and $\psi_n$ for any
corresponding eigenfunction.
For either a bounded interval or a torus, which we for now denote generically by $\Omega$, given a potential $q:\Omega\to\R$ and a test function $\phi \in V = H^1_0(\Omega)$ or $H^1(\Omega)$ as appropriate, we denote by
\begin{equation}
\label{rayleighsum}
\mathcal{R}[q,\phi]:= \frac{Q(\phi)}{\|\phi\|_2^2}:= \frac{\dint_\Omega |\nabla\phi(x)|^2\dx{x}+\dint_\Omega q(x)\phi^2(x)\dx{x}}
{\dint_\Omega \phi^2(x)\dx{x}}
\end{equation}
the Rayleigh quotient associated with $q$ at $\phi$, where $Q(\phi)=Q(\phi,\phi)$ is the bilinear form associated with the Schr\"odinger operator.
A standard generalization of the usual min-max formula for the eigenvalues
of \eqref{lapln} states that if $\phi_1,\ldots,\phi_n$ is a collection of $n \geq 1$ such test functions mutually orthogonal in $L^2(\Omega)$, then
\begin{equation}
\label{sumprinciple}
\sum_{k=1}^{n} \lambda_k(q) \leq \sum_{k=1}^{n} \mathcal{R}[q,\phi_k]
\end{equation}
(see, e.g., \cite{band}). If $q\equiv 0$ and we consider \eqref{lapln} with Dirichlet boundary conditions on $(0,\pi)$, then $\lambda_n(0) = n^2$
with associated eigenfunction $\psi_n(x) = \sin(nx)$, meaning in particular that if $s>1/2$, then the infinite sum of powers of eigenvalues
\begin{displaymath}
\sum_{k=1}^\infty \lambda_k^{-s}(0) = \sum_{k=1}^\infty \frac{1}{k^{2s}} = \zeta(2s)
\end{displaymath}
equals the Riemann zeta function, which we write as $\zeta(2s) =: \zeta_0(s)$. By way of analogy we define, as standard, the generalized
(or spectral) zeta function associated with the potential $q$ as
\begin{equation}
\label{speczeta}
\zeta_q(s) := \sum_{k=1}^\infty \lambda_k^{-s}(q).
\end{equation}
Finally, we denote by $\langle \,.\,,\,.\, \rangle$ the usual inner product in $L^2(\Omega)$, and by $\|\,.\,\|_p$ the $L^p$-norm on
$\Omega$, $1\leq p \leq \infty$.
\section{Summation bounds on finite intervals}
\label{sec:finite}
We first treat one-dimensional Schr\"odinger operators on the interval $(0,\pi)$. This means we can write the potential $q$ as a cosine Fourier series
as in~\eqref{fourier}, where we note in particular that $q_0/2$ is the average value of $q$,
\begin{displaymath}
\frac{q_0}{2} = \dashint_0^\pi q(x)\dx{x} := \frac{1}{\pi}\dint_0^\pi q(x)\dx{x}.
\end{displaymath}
Our starting point is the following bound for the Dirichlet problem.
\begin{theorem}
\label{th:finitetracedir}
If in the problem \eqref{stlibdd} the potential $q$ admits the expansion \eqref{fourier}, then its eigenvalues $\lambda_k = \lambda_k(q)$ satisfy
\begin{equation}
\label{finitetracedir}
\sum_{k=1}^{n}\left(\lambda_{k}-k^2-\frac{\ds q_{0}}{\ds 2}\right)\leq
-\frac{\ds 1}{\ds 2}\sum_{k=1}^{n} q_{2k}.
\end{equation}
for all $n\geq 1$. Equality for any $n\geq 1$ implies $q$ is constant. As $n$ goes to infinity, the right-hand side of the above inequality converges to
\begin{equation}
\label{tracedir}
\frac{1}{2}\,\dashint_0^\pi q(x)\dx{x}- \frac{\ds q(0) + q(\pi)}{\ds 4}.
\end{equation}
\end{theorem}
\begin{proof}
Using $\psi_k(x)=\sin(kx)$, $k=1,\ldots,n$, as test functions in \eqref{rayleighsum} we have
\begin{displaymath}
\begin{split}
\sum_{k=1}^{n}\lambda_{k} & \leq \sum_{k=1}^{n}\frac{\ds \dint_{0}^{\pi} k^2\cos^{2}(kx) +
q(x)\sin^{2}(k x)\dx{x}}{\ds \dint_{0}^{\pi} \sin^{2}(k x)\dx{x}}\\
& = \sum_{k=1}^{n}\left[ k^2 + \frac{\ds 2}{\ds \pi}\dint_{0}^{\pi} q(x)\sin^{2}(k x)\dx{x}\right].
\end{split}
\end{displaymath}
By expanding $q$ as in \eqref{fourier}, and using the identities $\sin^{2}(k x)=(1-\cos(2kx))/2$ and
$\cos(jx) \cos(2kx) = (\cos(jx+2kx)+\cos(jx-2kx))/2$, we obtain
\begin{displaymath}
\begin{split}
\dint_{0}^{\pi} q(x)\sin^{2}(k x)\dx{x} & = \frac{\ds 1}{\ds 2}\dint_{0}^{\pi} \left[\frac{\ds q_{0}}{\ds 2} +
\sum_{j=1}^{\infty} q_{j} \cos(j x)\right]\left[1-\cos(2kx)\right]\dx{x}\\
& = \frac{\ds \pi}{\ds 4}(q_{0} - q_{2k}).
\end{split}
\end{displaymath}
Replacing this in the expression above yields
\begin{displaymath}
\sum_{k=1}^{n}\lambda_{k} \leq \sum_{k=1}^{n} \left[k^2 + \frac{\ds q_{0} - q_{2k}}{\ds 2}\right]
\end{displaymath}
as desired. If we evaluate the Fourier series for $q$ at $x=0$ and at $x=\pi$ and add the resulting sums we see that
\begin{equation}
\label{potentialvalue}
q(0) + q(\pi) = q_{0} + 2\sum_{k=1}^{\infty}q_{2k},
\end{equation}
showing that the limit of the sum on the right-hand side of inequality \eqref{finitetracedir} is given by~(\ref{tracedir}). The statement in case of equality follows from Theorem~\ref{th:uniqueness}.
\end{proof}
\begin{remark}\label{nolowerbound}
(i) It is not possible to find a \emph{lower} bound for $\sum_{k=1}^n \lambda_k$ depending only on the first $2n$ Fourier coefficients of $q$ (or indeed, the first $m$ for any fixed $m\geq 0$) and the eigenvalues $k^2$, as the following simple example shows. For arbitrary $n\geq 1$, if we let $q(x)=t\cos(2n+2)x$, where $t>0$ is taken very large, then we have $q_{2n+2}=t$, while $q_k=0$ for all other $k\geq 0$. In this case, using an argument as in the proof of Theorem~\ref{th:finitetracedir} with $\sin(x), \ldots,\sin(n-1)x$, $\sin(n+1)x$ as our $n$ test functions, we see
\begin{displaymath}
\sum_{k=1}^n \lambda_k \leq \sum_{k=1}^{n-1}k^2 + (n+1)^2 - \frac{t}{2} \longrightarrow -\infty
\end{displaymath}
if we let $t \to \infty$, even though $q_0=\ldots=q_{2n}=0$. We could construct a similar example which also satisfies $q(0)=q(\pi)=0$ by taking,
for example, $q(x)=t\cos(2n+2)x - t\cos(2n+4)x$. It seems that any lower bound would have to take into account a quantity such as
$\sup_{n\in\N}|q_{2n}|$ or $\sup_{x\in(0,\pi)}|q(x)|$, or else only be valid asymptotically (e.g.~via the inclusion of an $O(n^s)$-type error
term). It is easy to construct analogous examples for the other problems we will consider.
(ii) The inequality \eqref{finitetracedir} is valid without convergence of the Fourier series \eqref{fourier}, provided only that the
coefficients $q_k$
given by \eqref{eq:fourier-int} are well defined.
\end{remark}
There is a direct analogue of Theorem~\ref{th:finitetracedir} for the Neumann problem.
\begin{theorem}
\label{th:finitetraceneu}
For all $n\geq 0$,
\begin{equation}
\label{finitetraceneu}
\sum_{k=0}^{n}\left(\mu_{k}-k^2-\frac{\ds q_{0}}{\ds 2}\right)\leq
\frac{\ds 1}{\ds 2}\sum_{k=1}^{n} q_{2k}.
\end{equation}
Equality for some $n\geq 0$ implies that $q$ is constant. The right-hand side of the above inequality converges to
\begin{equation*}
\label{traceneu}
-\frac{1}{2}\,\dashint_0^\pi q(x)\dx{x} + \frac{\ds q(0) + q(\pi)}{\ds 4}
\end{equation*}
as $n$ goes to infinity.
\end{theorem}
\begin{proof}
The proof is exactly the same as for Theorem~\ref{th:finitetracedir}, except that we use $\psi_k(x)=\cos(kx)$, $k=0,1,\ldots,n$ as test
functions. We omit the details.
\end{proof}
As in the Dirichlet case, a classical trace formula analogous to \eqref{classtr} (see e.g. \cite[Sec.~1.14]{lesa2}) implies there is equality in the
limit in \eqref{finitetraceneu} as $n \to \infty$.
By combining our separate estimates for Dirichlet and Neumann eigenvalues, we can simplify the resulting sums. This effectively corresponds
to considering the eigenvalues of the circle; cf.~Theorem~\ref{torus}. In this case, however, the result is an immediate consequence of
Theorems~\ref{th:finitetracedir} and \ref{finitetraceneu}.
\begin{theorem}
\label{finitetracecomb}
For all $n \geq 1$,
\begin{equation}
\label{eq:finitetracecomb}
\sum_{k=1}^n \left[\frac{\lambda_k+\mu_k}{2} - k^2 - \dashint_0^\pi q(x)\dx{x}\right]
+\frac{1}{2}\left[\mu_0-\dashint_0^\pi q(x)\dx{x}\right] \leq 0.
\end{equation}
with equality for any $n\geq 1$ implying $q$ is constant.
\end{theorem}
\begin{remark}
\label{rem:finitetracecomb}
As in the Dirichlet and Neumann cases, the Weyl asymptotics imply that the left-hand side of \eqref{eq:finitetracecomb} converges as $n \to \infty$,
and by combining the separate trace formulae for $\lambda_k$ and $\mu_k$ we see there is again equality in the limit.
\end{remark}
This combination of boundary conditions also allows us to obtain the following ``zeta function"-type bound.
\begin{theorem}
\label{th:sumpower}
Suppose that $\mu_0>0$. Then for all $n\geq 1$ and $s>0$,
\begin{displaymath}
\sum_{k=1}^n \lambda_k^{-s}+\sum_{k=0}^n \mu_k^{-s} \geq 2\sum_{k=1}^n
\left[k^2 + \dashint_0^\pi q(x)\dx{x}\right]^{-s} + \left[\dashint_0^\pi q(x)\dx{x}\right]^{-s}.
\end{displaymath}
\end{theorem}
If $s>1/2$, then both sides of the above inequality converge as $n \to \infty$. The proof
is a variant of that of Theorem~\ref{th:zeta} and is therefore delayed until Section~\ref{sec:zeta}.
\section{An application to a particular class of potentials on the interval}
\label{sec:app}
Here we give an application, or special case, of Theorems~\ref{th:finitetracedir} and \ref{th:finitetraceneu}. As in Section~\ref{sec:finite}, we denote by
$\lambda_k$ and $\mu_k$ the ordered Dirichlet and Neumann eigenvalues associated with $q$, respectively.
\begin{theorem}
\label{th:app}
Suppose that $q$ admits the expansion \eqref{fourier} and is absolutely continuous on $(0,\pi)$.
\begin{itemize}
\item[(i)] If $q'(x) \leq q'(\pi-x)$ a.e. on $(0,\frac{\pi}{2})$, then for all $n \geq 1$,
\begin{displaymath}
\sum_{k=1}^n \left(\lambda_k- k^2 - \dashint_0^\pi q(x) \dx{x}\right) \leq 0.
\end{displaymath}
\item[(ii)] If $q'(x) \geq q'(\pi-x)$ a.e. on $(0,\frac{\pi}{2})$, then for all $n \geq 0$,
\begin{displaymath}
\sum_{k=0}^n \left(\mu_k - k^2 - \dashint_0^\pi q(x) \dx{x}\right) \leq 0.
\end{displaymath}
\item[(iii)] Under the assumptions of (i), if in addition $\int_0^\pi q(x) \dx{x} \leq 0$, then for all $s>1/2$ we also have
\begin{displaymath}
\sum_{k=1}^n \lambda_k^{-s} \geq \sum_{k=1}^n k^{-2s}
\end{displaymath}
for all $n \geq 1$, and $\zeta_q(s) \geq \zeta(2s)$.
\end{itemize}
Equality in any of the above finite inequalities implies that $q$ is constant on $(0,\pi)$.
\end{theorem}
\begin{remark}
The assumptions of the Dirichlet case (i) are always satisfied by convex potentials, i.e.~potentials $q$ for which $q''(x) \geq 0$ a.e.~in $(0,\pi)$, while concave potentials, i.e.~with $q''(x) \leq 0$ a.e., always satisfy the Neumann condition~(ii).
\end{remark}
\begin{proof}[Proof of Theorem~\ref{th:app}]
(i) Fix $n \geq 1$. Recalling that $\dashint_0^\pi q(x)\dx{x} = q_0/2$ and using Theorem~\ref{th:finitetracedir}, we only need to show that
the right-hand side of \eqref{finitetracedir} is non-positive. Recalling the definition of $q_{2k}$ and integrating by parts,
\begin{displaymath}
\begin{split}
q_{2k} = \frac{2}{\pi} \int_0^\pi q(x) \cos(2kx)\dx{x} &= \frac{2}{\pi} \int_0^\frac{\pi}{2}
\left[q(x)+q(\pi-x)\right] \cos(2kx)\dx{x}\\
& = -\frac{1}{\pi k}\int_0^\frac{\pi}{2} \left[q'(x)-q'(\pi-x)\right] \sin(2kx)\dx{x}.
\end{split}
\end{displaymath}
Summing over $k$ and noting that $\frac{d}{dx} \cos^2(kx) /k = -\sin(2kx)$, this means
\begin{displaymath}
-\frac{1}{2}\sum_{k=1}^n q_{2k} = -\frac{1}{2\pi}\int_0^\frac{\pi}{2} \left[q'(x)-q'(\pi-x)\right] \frac{d}{dx}\left(
\sum_{k=1}^n \frac{\cos^2(kx)}{k^2}\right)\dx{x}.
\end{displaymath}
It is known that $\sum_{k=1}^n \cos^2(kx)/k^2$ is decreasing on $(0,\pi/2)$ for every $n \geq 1$ (cf. \cite[pp.~322--3]{laug}). Our assumptions on
$q$
therefore imply that the above integrand is positive for almost all $x \in (0,\pi/2)$, and thus
\begin{displaymath}
\sum_{k=1}^n \left(\lambda_k - k^2 - \frac{q_0}{2}\right) \leq -\frac{1}{2} \sum_{k=1}^n q_{2k} \leq 0.
\end{displaymath}
Equality for some $n \geq 1$ means that
\begin{displaymath}
0 = \sum_{k=1}^n \left(\lambda_k - k^2 -\frac{q_0}{2}\right) \leq - \frac{1}{2} \sum_{k=1}^n q_{2k} \leq 0,
\end{displaymath}
so that every inequality is an equality. In this case there is also equality in Theorem~\ref{th:finitetracedir}, and so $q$ is constant.
(ii) Applying Theorem~\ref{th:finitetraceneu} in place of Theorem~\ref{th:finitetracedir}, we obtain
\begin{displaymath}
\begin{split}
\sum_{k=0}^n \left(\mu_k - k^2 - \frac{q_0}{2}\right) &\leq \frac{1}{2}\sum_{k=1}^n q_{2k} \\
&\leq\frac{1}{2\pi}\int_0^\frac{\pi}{2} \left[q'(x)+q'(\pi-x)\right]\frac{d}{dx}\left(\sum_{k=1}^n\frac{\cos^2(kx)}{k^2}\right)\dx{x}.
\end{split}
\end{displaymath}
This time the integrand is negative almost everywhere. The case of equality follows in the same way as in (i).
(iii) This follows from Theorem~\ref{th:zeta} with $a_k = b_k = k^2$.
\end{proof}
\section{Summation bounds for the circle and flat torus}
\label{sec:torus}
In the case of $N$-dimensional flat tori we can obtain an especially simple bound which may be seen as the natural multi-dimensional generalization of Theorem~\ref{finitetracecomb}. By writing the zero-potential eigenfunctions as complex exponentials instead of sines and cosines, as is more natural on a torus, the proof becomes essentially trivial.
In order to proceed, we shall need some notation. We will denote by $\mathbb{T}$ an $N$-dimensional flat torus, $N\geq 1$, spanned by linearly independent vectors $v_1,\ldots,v_N \in \R^N$.
That is, if we define a lattice $\Gamma \subset \R^N$ as
\begin{displaymath}
\Gamma = \{n_1 v_1+\ldots+n_N v_N: n_i \in \mathbb{Z}, \,i=1\,\ldots,N\}.
\end{displaymath}
and an action of $\Gamma$ on $\R^N$ by $\gamma(x):=\gamma+x$, $\gamma\in\Gamma$, $x\in\R^N$, then our torus is given by
$\mathbb{T} = \R^N / \Gamma$. If $N=1$ then of course $\mathbb{T}$ is the circle.
We define the vectors $w_1,\ldots,w_N \in \R^N$ by $(w_j,v_k) = \delta_{jk}$, the Kronecker delta, where $( \,.\, , \,.\, )$ is the usual inner
product on $\R^N$. Denote by $W$ the matrix whose $j^{\rm th}$ row is given by the vector $w_j$. Then for each $\alpha =
(\alpha_1,\ldots,\alpha_N) \in \Z^N$, we may define an eigenfunction $\psi_\alpha$ of the zero-potential problem \eqref{lapln} with $q=0$ on the
manifold without boundary $\mathbb{T}$ by
\begin{displaymath}
\psi_\alpha(x):= e^{2\pi i \alpha^{\textrm T} W x};
\end{displaymath}
if we denote the $(m,n)^{\rm th}$ entry of $W$ by $w_{mn}$, then the associated eigenvalue $\lambda_\alpha = \lambda_\alpha(0)$ is given by
\begin{displaymath}
\lambda_\alpha = 4\pi^2 \sum_{n=1}^N \left(\sum_{m=1}^N \alpha_m w_{mn}\right)^2.
\end{displaymath}
In the one-dimensional case, we have just one vector $v=2\pi a \in \R$ (without loss of generality $v\geq 0$), in this case just one vector
$w =v^{-1} > 0$, and the eigenfunctions become $4\psi_n = e^{inx/a}$, with $\lambda_n = (n/a)^2$, $n \in \Z$.
Ordering the eigenvalues $\lambda_\alpha$ as an increasing sequence $\{\lambda_k(0)\}_{k\in\Z}$ with $\lambda_0(0)=0$ (corresponding to $\alpha = 0$), for each $k\geq 1$ there exists $\alpha=\alpha(k)$ for which $\lambda_k(0) = \lambda_\alpha$. The exact relationship between $\alpha$ and $k$ depends on the $v_j$, and to the best of our knowledge there is no known explicit formula for this for arbitrary $v_j$. We will order the eigenvalues $\lambda_k(q)$ of \eqref{lapln} with potential $q$ on $\mathbb{T}$ similarly.
\begin{theorem}
\label{torus}
For any integrable $q$ and for all $n \geq 0$,
\begin{equation}
\label{finitetracetor}
\sum_{k=0}^n \left[\lambda_k(q)-\lambda_k\left(\dashint_{\mathbb{T}}q(x)\dx{x}\right) \right]\leq 0.
\end{equation}
\end{theorem}
\begin{proof}
We use the $\psi_\alpha$ as test functions in the Rayleigh quotient \eqref{rayleighsum}: for any $\alpha \in \Z^N$, we have
\begin{displaymath}
\mathcal{R}[q,\psi_\alpha] = \lambda_\alpha + \frac{\dint_{\mathbb{T}} q(x) |\psi_\alpha(x)|^2\dx{x}}{\dint_{\mathbb{T}} |\psi_\alpha(x)|^2\dx{x}} =\lambda_\alpha+\dashint_{\mathbb{T}}q(x) \dx{x},
\end{displaymath}
since obviously $|\psi_\alpha(x)|=1$ for all $x \in \mathbb{T}$ and all $\alpha\in \Z^N$. Choosing the first $n$ of these functions and using the principle \eqref{sumprinciple} gives us the inequality.
\end{proof}
\begin{theorem}
\label{th:toruspower}
Suppose that $\lambda_0(q)>0$. Then for all $n\geq 1$ and $s>0$,
\begin{displaymath}
\sum_{k=0}^n \lambda_k^{-s}(q) \geq\sum_{k=0}^n\left[\lambda_k\left(
\dashint_{\mathbb{T}}q(x)\dx{x}\right)\right]^{-s}.
\end{displaymath}
\end{theorem}
\begin{proof}
Since $0<\lambda_0(q)\leq \dashint_{\mathbb{T}}q(x)\dx{x}$ by Theorem~\ref{torus}, we may apply Theorem~\ref{th:zeta}, from which the conclusion
follows immediately.
\end{proof}
\section{On the associated trace formulae}
\label{sec:trace}
We have already observed that there is equality in the limit as $n \to \infty$ in Theorems~\ref{th:finitetracedir},~\ref{th:finitetraceneu} and~\ref{finitetracecomb}, since we have convergence to the classical trace formulae of Gelfand--Levitan type. What is interesting is that the theorems from Section~\ref{sec:finite} allow us to
obtain a new (part of a) proof of the trace formulae. Our starting point is a paper by Diki\u{\i}~\cite{dikii}, who gave an alternative proof of the trace formula which is, in some sense, more natural that that in~\cite{gele}, which was based on a study of the asymptotics of the associated Green's functions.
This proof involves a two-part argument, which in our notation is as follows. If we denote the ordered eigenfunctions associated with the zero potential by $\psi_k$ and those associated with $q$ by $\varphi_k$, and assuming without loss of generality that the mean value $q_0$ of $q$ is zero, Diki\u{\i} proved using trigonometric identities, integration by parts and a manipulation of the resulting sums that
\begin{equation}
\label{eq:dikii1}
\sum_{k=1}^n \left(\mathcal{R}[q,\psi_k] - k^2\right) \longrightarrow -\frac{q(0)+q(4)}{2}
\end{equation}
as $n \to \infty$. The second part of the proof consists in using eigenvalue and eigenfunction asymptotics to show that
\begin{equation}
\label{eq:dikii2}
\lim_{n\to\infty} \sum_{k=1}^n \left(\mathcal{R}[q,\varphi_k]-\mathcal{R}[q,\psi_k]\right)=0.
\end{equation}
If we denote by $H = -\Delta +q$ the Schr\"odinger operator on $L^2(0,\pi)$ associated with $q$ (and the Dirichlet boundary condition), then we may
formally rewrite \eqref{eq:dikii2} as
\begin{equation}
\label{eq:basisform}
\lim_{n \to \infty} \sum_{k=1}^n \left(\langle \varphi_k,H\varphi_k\rangle - \langle \psi_k,H\psi_k\rangle\right) = 0.
\end{equation}
In some sense we can think of \eqref{eq:basisform} as asserting that the
action of $H$ is invariant with respect to a ``change of basis'' from $\{\varphi_k\}$ to $\{\psi_k\}$; we note that the arguments in \cite{dikii} were
phrased in these terms, and did not involve the use of Rayleigh quotients. Indeed, for $k,m \geq 1$,
we shall write $a^k_m = \langle \psi_k, \varphi_m \rangle$, so that $\psi_k = \sum_{m=1}^\infty a^k_m \varphi_m$ and $\varphi_m = \sum_{k=1}^\infty a^k_m \psi_k$, with $\sum_{k=1}^\infty (a^k_m)^2 = \sum_{m=1}^\infty (a^k_m)^2 = 1$ under the normalization
$\|\psi_k\|_2 = \|\varphi_m\|_2 = 1$. Then the sum \eqref{eq:dikii2} is equivalent to
\begin{equation}
\label{eq:dikiiestsum}
\begin{split}
\sum_{k=1}^\infty \sum_{m=1}^\infty & \lambda_k(a^k_m)^2 - \sum_{k=1}^n \sum_{m=1}^\infty \lambda_m (a^k_m)^2
=\\ &\sum_{k=1}^n \sum_{m=n+1}^\infty \lambda_k \left((a^m_k)^2 - (a^k_m)^2\right)
+ \sum_{k=1}^n \sum_{m=n+1}^\infty (\lambda_k - \lambda_m)(a^k_m)^2.
\end{split}
\end{equation}
Diki\u{\i} showed that for the problem on the interval the two sums on the right-hand side of \eqref{eq:dikiiestsum} tend to zero as $n \to \infty$ by using the basic asymptotic estimates $\lambda_k = k^2 + O(1)$ and $\varphi_k = \psi_k + O(k^{-1})$, the latter holding uniformly in $x \in (0,\pi)$ (see, e.g., \cite{lesa2}).
The proof of Theorem~\ref{th:finitetracedir}, whose conclusion says exactly that $\sum_{k=1}^n (\mathcal{R}[q,\varphi_k]-\mathcal{R}[q,\psi_k]) \leq 0$
for all $n \geq 1$, also gives an alternative proof of \eqref{eq:dikii1} along the way.
In a sense this is
more natural than the proof in~\cite{dikii}, at least when taken together with \eqref{eq:dikii2}, since it computes the finite sum
$\sum_{k=1}^n \mathcal{R} [q,\psi_k]$ explicitly in terms of $q$.
\begin{remark}
It would be interesting to know if such arguments might also work in higher dimensions, where there are next to no known trace formulae. We will
not explore this here, but as an example we remark that equality in the limit as $n \to \infty$ in Theorem~\ref{torus} is equivalent
to~\eqref{eq:dikii2} holding on the torus. The difficulty in using Diki\u{\i}'s idea in higher dimensions is that the asymptotic behaviour of
the $\lambda_k$ and $\varphi_k$ changes; for example, in two dimensions, we now have $\lambda_k \sim k$, not $k^2$. This makes it harder to
obtain effective bounds on the sums in \eqref{eq:dikiiestsum}.
\end{remark}
\section{Generalization to power bounds and zeta functions}
\label{sec:zeta}
Here we will prove a theorem from which the power bound generalizations stated in earlier sections, such as Theorem~\ref{th:app} (iii), will follow
immediately; it was inspired by, and based upon, a similar result and argument in \cite[Sec.~4]{laug}. We also give the very similar proof of Theorem~\ref{th:sumpower}.
\begin{theorem}
\label{th:zeta}
Suppose the sequences $(\lambda_k)_{k\geq 1}$ and $(a_k)_{k\geq 1}$ are positive and that $(a_k)_{k\geq 1}$ is non-decreasing in $k\geq 1$. Suppose also that the sequence $(b_k)_{k\geq 1}$ satisfies $\sum_{k=1}^m \lambda_k \leq \sum_{k=1}^m b_k$ for all $m\geq 1$. Then for all $s>0$ and all $n\geq 1$ we have
\begin{equation}
\label{zeta}
\sum_{k=1}^n (\lambda_k)^{-s} \geq \sum_{k=1}^n \left( (s+1)(a_k)^{-s}-s(a_k)^{-s-1}b_k\right).
\end{equation}
If the sequence $(b_k)_{k\geq 1}$ is itself positive and non-decreasing in $k\geq 1$, then the right-hand side of \eqref{zeta} is maximized when $a_k=b_k$ for all $1\leq k\leq n$.
\end{theorem}
Notice that if $a_k=b_k$ for all $k\geq 1$, then \eqref{zeta} simplifies to $\sum_{k=1}^n (\lambda_k)^{-s} \geq \sum_{k=1}^n (b_k)^{-s}$.
\begin{proof}
We will use the notation $[y]_+$, $y\in\R$, to mean the expression taking on the value $y$ if $y\geq 0$ and zero otherwise; $[f(x)]_{g(x)\geq y}$
will
represent $f(x)$ if $g(x) \geq y$ and zero otherwise. We start with the following identity, valid for $\lambda > 0$,
\begin{equation}
\label{powerrep}
\lambda^{-s} = s(s+1)\dint_0^\infty \alpha^{-s-2}[\alpha-\lambda]_+\dx{\alpha}
\end{equation}
For $n\geq 1$, $s>0$ arbitrary, applying this to both $\lambda_k$ and $a_k$, we have
\begin{displaymath}
\begin{split}
\sum_{k=1}^n \left(\lambda_k^{-s}-a_k^{-s}\right) &= s(s+1)\dint_0^\infty \alpha^{-s-2}
\sum_{k=1}^n \left([\alpha-\lambda_k]_+-[\alpha-a_k]_+\right)\dx{\alpha}\\
&\geq s(s+1)\dint_0^\infty \alpha^{-s-2}\sum_{k=1}^n [a_k-\lambda_k]_{\alpha\geq a_k}\dx{\alpha}\\
&\geq s(s+1)\dint_0^\infty \alpha^{-s-2}\sum_{k=1}^n [a_k-b_k]_{\alpha\geq a_k}\dx{\alpha}\\
&=\sum_{k=1}^n s(s+1)(a_k-b_k)\dint_{a_k}^\infty \alpha^{-s-2}\dx{\alpha},
\end{split}
\end{displaymath}
which after simplification and rearrangement gives us \eqref{zeta}. (Note that we needed the sequence $a_k$ to be weakly increasing to justify the third line above.) For the maximizing property we consider each term on the right-hand side of \eqref{zeta} as a function of $a_k$ by setting $g_k(a_k):= (s+1)(a_k)^{-s}-s(a_k)^{-s-1}b_k$. Differentiating in $a_k$ shows that $g_k$ reaches its unique maximum when $a_k=b_k$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{th:sumpower}]
Keeping the notation from the proof of Theorem~\ref{th:zeta} and using \eqref{powerrep}, we have
\begin{displaymath}
\begin{split}
\sum_{k=1}^n \lambda_k^{-s} &+\sum_{k=0}^n \mu_k^{-s}
=s(s+1)\dint_0^\infty \alpha^{-s-2}\left(\sum_{k=1}^n [\alpha-\lambda_k]_+ +\sum_{k=0}^n
[\alpha-\mu_k]_+\right)\dx{\alpha}\\
&\geq s(s+1)\dint_0^\infty \alpha^{-s-2}\left(\sum_{k=1}^n [2\alpha-\lambda_k-\mu_k]_{\alpha \geq k^2+\frac{q_0}{2}}+
[\alpha-\mu_0]_{\alpha\geq \frac{q_0}{2}}\right)\dx{\alpha}.\\
\end{split}
\end{displaymath}
For each fixed $\alpha\geq q_0/2+1$, the sum in the latter integral is from $k=1$ to some $m=m(\alpha) \leq n$; if $\alpha \in [q_0/2, q_0/2+1)$, then the bracketed term reduces to $\alpha-\mu_0$, and otherwise it is zero. This means that for each fixed $\alpha\geq q_0/2+1$ we may apply Theorem~\ref{finitetracecomb} (or Theorem~\ref{th:finitetraceneu} with $n=0$ if $\alpha \in [q_0/2, q_0/2+1)$) to obtain
\begin{displaymath}
\sum_{k=1}^{m(\alpha)} [2\alpha-\lambda_k-\mu_k]_{\alpha \geq k^2+\frac{q_0}{2}}
+[\alpha-\mu_0]_{\alpha\geq \frac{q_0}{2}}\leq \sum_{k=1}^{m(\alpha)} [2\alpha-2k^2-q_0]_+ +[\alpha-q_0/2]_+
\end{displaymath}
for each $\alpha>0$. Substituting this back into the above expression for $\sum_{k=1}^n \lambda_k^{-s} +\sum_{k=0}^n \mu_k^{-s}$ and
applying \eqref{powerrep} in the other direction yields the theorem.
\end{proof}
\section{The case of equality}
\label{sec:equality}
Finally, we will prove the sharpness of our inequalities, in the sense that equality for some $n\geq 1$ in a bound of the form
\eqref{sumprinciple} forces the potential $q$ to be a constant. This stems from the fact that the only functions that can minimize the Rayleigh
quotient expression \eqref{rayleighsum} are sums of eigenfunctions of the corresponding equation. Although we doubt this is new, we do not know
of any explicit reference in the literature and so give a proof here.
We suppose we have the equation \eqref{lapln} in any one of the cases considered and two different potentials $q_1, q_2\in L^\infty(\Omega)$.
We denote by $V = H^1_0 (\Omega)$ or $H^1(\Omega)$ the appropriate Hilbert space. We will write $\lambda_k(q_2)$, $k\geq 1$ for the eigenvalues of the problem associated with $q_2$, ordered by increasing size and repeated according to multiplicities, and $\psi_k$ for the corresponding eigenfunctions which form an orthonormal basis of $L^2(\Omega)$.
\begin{lemma}
\label{lemma:equality}
Suppose that for some $n\geq 1$,
\begin{equation}
\label{eq}
\sum_{k=1}^{n} \lambda_k(q_1) = \sum_{k=1}^n \mathcal{R}[q_1,\psi_k].
\end{equation}
Then there exist eigenfunctions $\varphi_k$, $k\geq 1$ corresponding to $\lambda_k(q_1)$ (ordered by increasing magnitude) such that
\begin{equation}
\label{span}
\spn\{\psi_1,\ldots,\psi_n\} = \spn\{\varphi_1,\ldots,\varphi_n\}
\end{equation}
in $L^2(\Omega)$, that is, each $\psi_k$ may be expressed as a finite linear combination $\psi_k(x)=\sum_{m=1}^n a^k_m\varphi_m(x)$ for
suitable constants $a^k_m \in\R$.
\end{lemma}
Of course, if $\lambda_n(q_1)$ is not simple, then we need to choose the right eigenfunction(s) $\varphi_n$ (and possibly $\varphi_{n-1},\ldots,\varphi_{n-m}$) in the corresponding eigenspace. We also note that this is really an abstract result which is true for any two positive, self-adjoint operators on a (real) Hilbert space, and in particular valid in greater generality.
In general, however, it does not seem so easy to prove that $q_1-q_2$ is constant in $\Omega$ (following from $a^k_m = \delta_{km}$ in Lemma~\ref{lemma:equality}, so that the $\psi_k$ are directly eigenfunctions of $q_1$). Here, we will deal only with the case of an interval with Dirichlet or Neumann boundary conditions, and with $q_2=0$. A key element of our proof, as in Section~\ref{sec:finite}, is the fact that products of eigenfunctions of the zero potential, namely sines and cosines, are mutually orthogonal in the relevant $L^2$-space. We expect the same idea should work on the $N$-dimensional
torus, since the eigenfunctions are complex exponentials, but the argument is complicated by various issues related to the multiplicity of the eigenvalues, and we do not explore it here. So we now return to having $q_2\equiv 0$ and labelling $q_1$ as $q$, given by \eqref{fourier}.
\begin{theorem}
\label{th:uniqueness}
Suppose that for some $n\geq 1$, there is equality in \eqref{finitetracedir}, respectively \eqref{finitetraceneu}. Then $q(x)$ is constant in $(0,\pi)$ with eigenfunctions given by $\varphi_k(x)=\sin(kx)$, $k\geq 1$, and $\varphi_k(x)=\cos(kx)$, $k\geq 0$, in the Dirichlet and Neumann cases, respectively.
\end{theorem}
\begin{proof}[Proof of Lemma~\ref{lemma:equality}]
For each $k\geq 1$, as in Section~\ref{sec:trace} we write $\psi_k = \sum_{m=1}^\infty a_m^k \varphi_m$, where
$a_m^k = \langle \psi_k,\varphi_m \rangle$ and the $\varphi_k$ are, for the meantime, an arbitrary set of eigenfunctions for $q_1$,
in the sense that we allow an arbitrary decomposition of any eigenspace of dimension $\geq 2$.
The orthonormality relations imply $\sum_{m=1}^\infty a_m^k a_m^l = \delta_{kl}$. Using \eqref{eq} and denoting by $Q_1$ the bilinear
form associated with $q_1$ (cf~\eqref{rayleighsum}) we have
\begin{displaymath}
\sum_{k=1}^n \lambda_k(q_1) = \sum_{k=1}^n \mathcal{R}[q_1,\psi_k]
= \sum_{k=1}^n Q_1 \left( \sum_{m=1}^\infty a_m^k \varphi_m, \sum_{l=1}^\infty a_l^k \varphi_l \right).
\end{displaymath}
Since everything converges, and since $Q_1(\varphi_m,\varphi_l) = \delta_{ml} \lambda_m (q_1)$, this reduces to
\begin{equation}
\label{weightesums}
\sum_{k=1}^n \lambda_k(q_1) = \sum_{k=1}^n \sum_{m=1}^\infty (a_m^k)^2 \lambda_m (q_1)
= \sum_{m=1}^\infty \left( \sum_{k=1}^n (a_m^k)^2\right) \lambda_m (q_1).
\end{equation}
Since the functions $\varphi_m = \sum_{k=1}^\infty a_m^k \psi_k$ are also normalized, we have $\sum_{k=1}^n (a_m^k)^2 \leq 1$ for all $m\geq 1$
and $\sum_{m=1}^\infty\left(\sum_{k=1}^n (a_m^k)^2\right)=n$. Hence the only way we can have equality in \eqref{weightesums}
is if the coefficient of $\lambda_m (q_1)$ in the sum on the right-hand side is zero whenever $\lambda_m (q_1) > \lambda_n (q_1)$, which
means by definition of the $a_m^k$ that $\spn_{L^2(\Omega)}\{\psi_k\}_{k=1}^n$ is contained in the union of the eigenspaces associated with
$\lambda_1(q_1),\ldots,\lambda_n(q_1)$ (which may be more than $n$-dimensional if $\lambda_n$ is not simple). However, since
$\spn_{L^2(\Omega)}\{\psi_k\}_{k=1}^n$ is $n$-dimensional, we can find a decomposition of the eigenspace associated with $\lambda_n(q_1)$
so that $\spn_{L^2(\Omega)}\{\psi_k\}_{k=1}^n$ is equal to the span of the corresponding first $n$ eigenfunctions for $q_1$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{th:uniqueness}]
We will give the proof in detail only for the Dirichlet case \eqref{finitetracedir}; it is elementary, though tedious, to work
through the (very similar) details in the Neumann case \eqref{finitetraceneu}, and so we will briefly sketch the proof of the Neumann case afterwards. Supposing that equality holds in~\eqref{finitetracedir} for some $n\geq 2$ (the case $n=1$ being trivial), denoting by $\varphi_k$, $k=1,\ldots, n$ the first $n$ eigenfunctions associated with $q$, without loss of generality chosen and numbered so that the conclusion of Lemma~\ref{lemma:equality} holds, since $\varphi_k(x) \neq 0$ almost everywhere in $(0,\pi)$, for each $k\geq 1$ we may write
\begin{displaymath}
q(x) = \lambda_k(q)+\frac{\varphi_k''(x)}{\varphi_k(x)},
\end{displaymath}
which is valid pointwise almost everywhere. In particular, this means that for all $j,k=1,\ldots,n$ and almost all $x\in (0,\pi)$,
\begin{equation}
\label{eigenfndiff}
\frac{\varphi_k''(x)}{\varphi_k(x)}=C_{k,j} +\frac{\varphi_j''(x)}{\varphi_j(x)},
\end{equation}
where $\R \ni C_{k,j}=\lambda_k(q)-\lambda_j(q)$. Since by Lemma~\ref{lemma:equality} we have
\begin{displaymath}
\varphi_k(x)=\sum_{m=1}^n a_m^k \sin(mx)
\end{displaymath}
for appropriate $a_m^k \in \R$, after rearranging, we may rewrite \eqref{eigenfndiff} explicitly as
\begin{multline*}
\left[\sum_{m=1}^n m^2 a_m^k \sin(mx)\right]\left[\sum_{l=1}^n a_l^j \sin(lx)\right]
= \\ -C_{k,j}\left[\sum_{m=1}^n a_m^k \sin(mx)\right]\left[\sum_{l=1}^n a_l^j \sin(lx)\right]+
\left[\sum_{m=1}^n a_m^k \sin(mx)\right]\left[\sum_{l=1}^n l^2 a_l^j \sin(lx)\right]
\end{multline*}
which in turn may be rewritten as
\begin{multline}
\label{cosequality}
\sum_{l,m=1}^n m^2 a_m^k a_l^j \left[\cos(l-m)x -\cos(l+m)x \right]= -C_{k,j}
\sum_{l,m=1}^n a_m^k a_l^j \, \cdot \\ \cdot \left[\cos(l-m)x -\cos(l+m)x \right]+
\sum_{l,m=1}^n l^2 a_m^k a_l^j \left[\cos(l-m)x -\cos(l+m)x \right].
\end{multline}
We observe that \eqref{cosequality} is a sum of the form
\begin{displaymath}
\sum_{p=0}^{2n} b_p \cos(px)=0,
\end{displaymath}
where the coefficients $b_p$ are obtained by summing all the relevant coefficients of $\cos(l-m)x$ with $|l-m|=p$ and $\cos(l+m)x$ with $l+m=p$. Since this holds for almost every $x\in(0,\pi)$, orthogonality of the (finite) family $\cos(px)$ implies that $b_p=0$ for all $p\geq 0$; we will use this to show that no combination of coefficients $a_m^k$ other than $a_m^k = c \delta_{mk}$ (with $c$ a normalizing constant) can satisfy \eqref{eigenfndiff}, and hence no non-constant $q$ is possible.
To do so we make a particular choice of $j,k$: without loss of generality, we may assume there exist two distinct eigenfunctions
$\varphi_j$, $\varphi_k$ such that $a_n^j, a_n^k \neq 0$, that is, both have non-zero $L^2$-projection onto $\spn\{\sin(nx)\}$; otherwise,
\eqref{span}
forces $\varphi_i(x)=\sin(nx)$ for some $1\leq i\leq n$, and by comparing the two eigenfunction equations of the form \eqref{stlibdd} that $\sin(nx)$
must therefore satisfy, a routine argument shows that $q$ must be constant. So let us assume we have our $\varphi_j$ and $\varphi_k$
and consider the equation for $b_{2n}$. That is, equating coefficients of $\cos(2nx)$ in \eqref{cosequality},
\begin{displaymath}
n^2 a_n^k a_n^j = -C_{k,j} a_n^k a_n^j + n^2 a_n^k a_n^j.
\end{displaymath}
Since $a_n^j, a_n^k \neq 0$, this implies $C_{k,j}=0$, that is, $\lambda_k(q)=\lambda_j(q)$. We claim that in this case
\begin{equation}
\label{equalratios}
\frac{a_m^k}{a_n^k} = \frac{a_m^j}{a_n^j}
\end{equation}
for all $m=1,\ldots,n-1$. Let us first show why this proves the theorem. Assuming \eqref{equalratios} holds,
and writing $a_m^k = c_m a_n^k$, $a_m^j = c_m a_n^j$ for $c_m \in \R$,
\begin{displaymath}
\sum_{m=1}^n c_m^2(a_n^k)^2 = \sum_{m=1}^n (a_m^k)^2 = \|\varphi_k\|_2^2 = \|\varphi_j\|_2^2
= \sum_{m=1}^n (a_m^j)^2 = \sum_{m=1}^n c_m^2(a_n^j)^2,
\end{displaymath}
implying $a_n^k = a_n^j$. \eqref{equalratios} now implies inductively that $a_m^k=a_m^j$ for all $m=1,\ldots,n$, that is, $\varphi_k = \varphi_j$,
contradicting our assumption that $\varphi_k$ and $\varphi_j$ were two distinct eigenfunctions with non-trivial projection onto $\spn\{\sin(nx)\}$.
The only possibility is therefore that $\sin(nx)$ is itself an eigenfunction associated with $q(x)$, which implies $q$ is constant, as can be seen directly from the equation \eqref{stlibdd}.
It remains to prove \eqref{equalratios}. We will proceed by induction on $p$ from $2n$ down to $n+1$, equating the coefficients of $\cos(l+m)x$,
$l+m=p$ in~\eqref{cosequality} in order to obtain \eqref{equalratios} for $a_{p-n}^k$, $a_{p-n}^j$. Observe that for each $p=n+1,\ldots,2n$,
\eqref{equalratios} reduces in this case to
\begin{equation}
\label{reduceddiff}
\sum (m^2-l^2)a_m^k a_l^j = 0,
\end{equation}
where the sum is over all $l,m=1,\ldots,n$ such that $l+m=p$. When $p=2n-1$, this says
\begin{displaymath}
(n-1)^2 a_{n-1}^k a_n^j+n^2 a_n^k a_{n-1}^j = n^2 a_{n-1}^k a_n^j+(n-1)^2 a_n^k a_{n-1}^j,
\end{displaymath}
which upon rearrangement gives \eqref{equalratios} for $n-1$. Suppose now that \eqref{equalratios} holds for $p=2n-1,\ldots,n+i+1$, $i \geq 1$.
Taking\eqref{reduceddiff} for $p=n+i$ and dividing through by $a_n^k a_n^j$, we have
\begin{displaymath}
\sum_{m=i}^n m^2 \frac{a_m^k}{a_n^k}\frac{a_{n+i-m}^j}{a_n^j}=
\sum_{l=i}^n l^2\frac{a_{n+i-l}^k}{a_n^k}\frac{a_l^j}{a_n^j}.
\end{displaymath}
Using the induction hypothesis that \eqref{equalratios} holds for $m=i+1,\ldots,n$, we see we can cancel all but the first terms in the above
equality,
leaving
\begin{displaymath}
i^2 \frac{a_i^k}{a_n^k}\frac{a_n^j}{a_n^j} = i^2 \frac{a_n^k}{a_n^k}\frac{a_i^j}{a_n^j}.
\end{displaymath}
Hence \eqref{equalratios} holds for $i$, proving our claim.
Finally, let us remark that in the case of Neumann boundary conditions \eqref{finitetraceneu}, \eqref{eigenfndiff} is unchanged, while the expression for $\varphi_k$ is now
\begin{displaymath}
\varphi_k(x) = \sum_{m=0}^n a_m^k \cos(mx),
\end{displaymath}
meaning \eqref{cosequality} is the same, only with expressions of the form $[\cos(l-m)x+\cos(l+m)x]$ replacing $[\cos(l-m)x -\cos(l+m)x]$, and with
summation from $m=0$ to $n$ rather than $1$ to $n$, meaning our induction proceeds down to $p=n$. Otherwise, the argument is the same.
\end{proof}
\subsection*{Acknowledgements}
J.B.K. was supported by a grant within the scope of the Funda\c c\~{a}o para a Ci\^{e}ncia e Tecnologia's project PTDC/\ MAT/\ 101007/2008
and by a fellowship of the Alexander von Humboldt Foundation, Germany. P.F. was partially supported by FCT's project
PTDC/\ MAT-CAL/\ 4334/2014.
|
\section{Introduction}
\noindent The theory of numbers gives a mathematical object which is related with this work, is the classical {\it Lagrange spectrum} (cf. \cite{CF}), which we describe in the following: Given an irrational number $\alpha$, according to Dirichlet's theorem the inequality $\left|\alpha-\frac{p}{q}\right|<\frac{1}{q^2}$ has infinitely many rational solutions $\frac{p}{q}$.
Markov and Hurwitz improved this result (cf. \cite{CF}), proving that, for all irrational $\alpha$, the inequality $\left|\alpha-\frac{p}{q}\right|<\frac{1}{\sqrt{5}q^2}$ has infinitely many rational solutions $\frac{p}{q}$.
\noindent Meanwhile, for a fixed irrational $\alpha$ better results can be expected. We associate, to each $\alpha$, its best constant of approximation (Lagrange value of $\alpha$), given by \begin{eqnarray*}k(\alpha)&=&\sup\left\{k>0:\left|\alpha-\frac{p}{q}\right|<\frac{1}{kq^2} \ \text{has infinite rational solutions $\frac{p}{q}$ }\right\}\\
\end{eqnarray*}
The set $L=\{k(\alpha):\alpha\in \mathbb{R}\setminus \mathbb{Q} \ \text{and} \ k(\alpha)<\infty\}$
is known as the \textit{{Lagrange spectrum}}.
\noindent Many people had study the structure of this object, the principal results is due to M. Hall \cite{Hall} (cf. also \cite{CF}), he proved that $L\supset[6,+\infty)$, so the Lagrange spectrum contains a whole half-line - such a half-line is known as a \textit{Hall's ray} of the Lagrange spectrum, also G. Freiman (cf. \cite{F} and \cite{CF}) determined the precise beginning of Hall's ray (the biggest half-line contained in $L$), which is
$$
\frac{2221564096 + 283748\sqrt{462}}{491993569} \cong 4,52782956616\dots\,.
$$
\noindent Another interesting set related to diophantine approximations is the classical \textit{Markov spectrum} defined by (cf. \cite{CF})
\begin{equation*}\label{SMClassic
M=\left\{\inf_{(x,y)\in \mathbb{Z}^{2}\setminus(0,0)}|f(x,y)|^{-1}:f(x,y)=ax^2+bxy+cy^2 \ \text{with} \ b^2-4ac=1\right\}.
\end{equation*}
\noindent Both the Lagrange and Markov spectrum have a dynamical interpretation, that is of interest for our work.
\noindent Let $\Sigma=({\mathbb{N}^*})^{\mathbb{Z}}$ and $\sigma\colon \Sigma \to \Sigma$ the shift defined by $\sigma((a_n)_{n\in\mathbb{Z}})=(a_{n+1})_{n\in\mathbb{Z}}$. If $f\colon \Sigma \to \mathbb{R}$ is defined by $f((a_n)_{n\in\mathbb{Z}})=\alpha_{0}+\beta_{0}=[a_0,a_1,\dots]+[0,a_{-1},a_{-2},\dots]$ (in continued fractions), then
$$L=\left\{\limsup_{n\to\infty}f(\sigma^{n}(\underline{\theta})):\underline{\theta}\in \Sigma\right\} \\ \ \text{and} \ \
M=\left\{\sup_{n\in\mathbb{Z}}f(\sigma^{n}(\underline{\theta})):\underline{\theta}\in \Sigma\right\}.$$
\noindent This last interpretation, in terms of shift, admits a natural generalization of Lagrange and Markov spectrum in a most general context ( hyperbolic dynamics context, which is the focus of this work).\\
\noindent We define the Lagrange and Markov dynamical spectrum as follows. Let $\varphi \colon M\to M$ be a diffeomorphism with $\Lambda\subset M$ a invariant set, \emph{i.e.}, $\varphi(\Lambda)=\Lambda$. Let $f\colon M\to \mathbb{R}$ be a continuous real function, then the \textit{Lagrange Dynamical Spectrum} associated to $(f,\Lambda)$ is defined by
$$\displaystyle L(f,\Lambda)=\left\{\limsup_{n\to\infty}f(\varphi^{n}(x)):x\in \Lambda\right\},$$
and the \textit{Markov Dynamical Spectrum} associate to $(f,\Lambda)$ is defined by
$$M(f,\Lambda)=\left\{ \sup_{n\in \mathbb{Z}}f(\varphi^{n}(x)):x\in \Lambda\right\}.$$
\ \\
In this context S. Roma\~na and C. Moreira (cf. \cite{RM}) showed that for $2$-dimensional manifold we have the following \\
\ \\
\noindent {\bf Theorem}\cite{RM} \emph{Let $\Lambda$ be a horseshoe associated to a $C^2$-diffeomorphism $\varphi$ such that $HD(\Lambda)>1$. Then there is, arbitrarily close to $\varphi$, a diffeomorphism $\varphi_{0}$ and a $C^{2}$-neighborhood $W$ of $\varphi_{0}$ such that, if $\Lambda_{\psi}$ denotes the continuation of $\Lambda$ associated to $\psi\in W$, there is an open and dense set $H_{\psi}\subset C^{1}(M,\mathbb{R})$ such that for all $f\in H_{\psi}$, we have
\begin{equation*}
int \ L(f,\Lambda_{\psi})\neq\emptyset \ \text{and} \ int \ M(f,\Lambda_{\psi})\neq\emptyset,
\end{equation*}
where $int A$ denotes the interior of $A$.}\\
\ \\
Here by horseshoe we mean a compact, locally maximal, transitive hyperbolic invariant set of saddle type (and so it contains a dense subset of periodic orbits).\\
\ \\
\noindent There is also a geometric interpretation of the classical Lagrange spectrum (cf. \cite{CF}). Consider the modular group, $SL(2,\mathbb{Z})$, that is, the set of all $2\times 2$ integer matrices with determinant equal to one, and $PSL(2,\mathbb{Z})$ the projectivization of $SL(2,\mathbb{Z})$. Given any $V\in SL(2,\mathbb{Z})$, $V=\bigl(\begin{smallmatrix}
a&b\\ c&d
\end{smallmatrix} \bigr)$ we define the associated transformation by $V(z)=\frac{az+b}{cz+d}$. However for an irrational number $\alpha$ the Lagrange value of $\alpha$ is
$$k(\alpha)=\sup \{k:|V(\infty)-V(\alpha)|^{-1}=|q(q\alpha-p)|\leq k^{-1} \ \text{for infinitely many} \ V\in SL(2,\mathbb{Z})\}.$$
\noindent Let $\mathbb{H}^{2}$ be of upper half-plane model of the real hyperbolic plane, with the Poincar\'e metric, and let $N:=\mathbb{H}^{2}/PSL(2,\mathbb{Z})$ the modular orbifold. Let $e$ be an end of $N$ ( cf. \cite{HP1} and \cite{PP}), define \textit{the asymptotic height spectrum of the pair $\left(N,e\right)$} by
$$LimsupSp(N,e)=\left\{\limsup_{t\to\infty}ht_{e}(\gamma(t)): \gamma \in SN\right\},$$
\noindent where
\begin{equation}\label{Height}
ht_{e}(x)=\displaystyle\lim_{t\to+\infty}d(x,\Gamma(t))-t
\end{equation}
is the height associated to the end $e$ of $N$, being $\Gamma$ a ray that defines the end $e$, and $SN$ denotes the unitary tangent bundle of $N$.\\
\noindent Using the latter interpretation of the Lagrange spectrum, the asymptotic height spectrum $LimsupSp(N,e)$ of the modular orbifold $N$ is the image of the Lagrange spectrum by the map $t\to \log \frac{t}{2}$ (see for instance [\cite{HP1}, theorem 3.4]).
The geometric interpretation of Freiman's result in our context is that $LimsupSp(N,e)$ contains the maximal interval $[\mu,+\infty)$ with
$$\mu= \log \left(\frac{2221564096+283748\sqrt{462}}{2\cdot491993569}\right)\approxeq 0,817095519650396598... \ .$$
\noindent Similar results were obtained by A. Haas and C. Series (cf. \cite{HS}) to the quotient of $\mathbb{H}^{2}$ by a fuchsian group of $SL(2,\mathbb{R})$. In the same year, Andrew Haas \cite{Haas} obtained results in this direction for hyperbolic Riemann surfaces. Then 11 years later, in 1997, Thomas A. Schmidt and Mark Sheigorn (cf. \cite{SS}) proved that Riemann surfaces have a Hall's ray in every cusp.
In 2012, P. Hubert, L. Marchese and C. Ulcigrai (cf.\cite{HMU}) showed the existence of Hall's ray in the context of Teichm\" uller dynamics, more precisely for moduli surfaces. Recently, in 2014, M. Artigiani, L. Marchese, C. Ulcigrai (cf. \cite{AMU}) showed than Veech surfaces also have a Hall's ray.\\
\ \\
Observe that all results mentioned above are on surfaces, which in the geometrical cases they all have negative constant curvature. Let $M$ be a complete connected Riemannian manifold with sectional curvature at most $-1$ dimension grater than or equal to $3$, and let $e$ be an end of it; the associated \textit{Lagrange and Markov Spectra} are defined respectively by
$$LimsupSp(M,e)=\left\{\limsup_{t\to\infty}ht_{e}(\gamma(t)): \gamma \in SM\right\}$$ \text{and}
$$MaxSp(M,e)=\left\{\sup_{t\in \mathbb{R}}ht_{e}(\gamma(t)): \gamma \in SM\right\},$$
where $\gamma(t)$ is the geodesic such that $\gamma(0)=\gamma\in SM$.\\
In this case, J. Parkkonen and F. Paulin \cite{PP}, using purely geometric arguments showed the following theorems:\\
\noindent {\bf Theorem}\cite{PP} \emph{If $M$ has finite volume, dimension $n\geq 3$ and $e$ is an end of $M$, then $Max Sp(M,e)$ contains the interval $[4.2,+\infty]$.}\\
\ \\
\noindent {\bf Theorem}\cite{PP}({The Ubiquity of Hall's rays}) \emph{If $M$ has finite volume, dimension $n\geq 3$ and $e$ is and end of $M$, then $LimsupSp(M,e)$ contains the interval $[6.8,+\infty]$.}\\
\ \\
\noindent The above geometrical interpretation can be seen from the point of view of dynamics as: Let $X$ be a complete vector field on a manifold $M$ and consider $X^{t} \colon M\to M$ the flow associated to $X$. Let $f\colon M\to \mathbb{R}$ be a continuous real function, then the \textit{Lagrange Dynamical Spectrum} associated to $(f,X)$ is defined by
$$\displaystyle L(f,X)=\left\{\limsup_{t\to\infty}f(X^{t}(x)):x\in M \right\},$$
\noindent
and the \textit{Markov Dynamical Spectrum} associate to $(f,X)$ is defined by
$$M(f,X)=\left\{ \sup_{t\in \mathbb{R}}f(X^{t}(x)):x\in X\right\}.$$
In this context S. Roma\~na and C. Moreira (cf. \cite{RM2}) proved that if $N$ be a complete noncompact surface $N$ with metric $g_0$ such that the Gaussian curvature is bounded between two negative constants, the Gaussian volume is finite and if $\phi_0$ denotes the vector field in $SN$ (the unitary tangent bundle) defining the geodesic flow of the metric $g_0$. Let $\mathfrak{X}^{1}(SN)$ be the space of $C^{1}$ vector field. Then \\
\ \\
\noindent \textbf{Theorem}\cite{RM2} \emph{Arbitrarily close to $\phi_0$ there is an open set $\mathcal{V}\subset \mathfrak{X}^{1}(SN)$ such that for any $X\in \mathcal{V}$ we have $$int\, M(f,X)\neq \emptyset \ \text{and} \ int\, L(f,X)\neq \emptyset$$
for any $f$ in a dense and $C^{2}$-open subset $\mathcal{U}_X$ of $C^{2}(SN,\mathbb{R})$.
Moreover, the above statement holds persistently: for any $Y\in \mathcal{V}$, it holds for any $(f,X)$ in a suitable neighborhood of $\mathcal{U}_Y\times \{Y\}$ in $C^{2}({SN,\mathbb{R}})\times \mathfrak{X}^{1}(SN)$.}\\
\noindent \textbf{Theorem}\cite{RM2} \emph{Arbitrarily close to $g_0$, there is an open set $\mathcal{G}$ such that, for any $g\in \mathcal{G}$, there is a dense and $C^{2}$-open subset $\mathcal{H}_g\subset C^{2}(S^{g}N,\mathbb{R})$ such that
$$int\, M(f,\phi_{g})\neq \emptyset \ \ \text{and} \ \ int\, L(f,\phi_{g})\neq \emptyset \ \ \ \text{for all} \ \ f\in \mathcal{H}_g,$$
where $\phi_{g}$ is the vector field defining the geodesic flow of the metric $g$ and $S^{g}N$ is the unitary tangent bundle of the metric $g$.}
\\
\ \\
\noindent \textbf{Theorem }\cite{RM2}\,\emph{Let $M$ be as above and $e$ an {end} of $M$. Arbitrarily close to $g_0$ there is an open set $\mathcal{G}$ of metrics such that for any $g\in \mathcal{G}$,
$$int\, M(h^{g}_{e}\circ \pi,\phi_{g})\neq \emptyset \ \ \text{and} \ \ int\, L(h^{g}_{e}\circ \pi,\phi_{g})\neq \emptyset,$$
where $h^{g}_{e}$ is the height function (see (\ref{Height})) associated to $e$ with the metric $g$ .
}\\
\\
\noindent Since the set of Anosov flows is an open set, then for all geodesic flow associated with a metric in $\mathcal{G}$ and for all flow associated to vector field in $\mathcal{V}$ are still Anosov flows. Moreover, the Liouville measure associated to the metric $g_0$ is invariant for the flow, so the this geodesic flow is conservative.
The geodesic flow has some fundamental properties that can be used to prove the last two theorems above (cf.\cite[section 3]{RM2}), by example the stable and unstable foliations are $C^1$ (cf. \cite{MHP}) and by the result of S. G. Dani (cf. \cite{D} and \cite{D1}) the set of point in $SN$ whose orbit under the geodesic flow is bounded has Hausdorff dimension equal to $3$. It is important to note that, in general these properties are not true for all Anosov flows. Nevertheless, in this work we used an Urba\'nski result (cf. \cite{Urbanski}) to construct hyperbolic sets with large Hausdorff dimension for conservative Anosov flows and conservative Anosov diffeomorphism. Moreover, the main difficult in this paper is to show one separation Lemma (see Lemma \ref{L7}) using only the $C^{0}$ stable and unstable foliations and produce small conservative perturbations of the flow such that we can obtain the property $V$ (see subsection \ref{SIRCS}). \\
\ \\
We consider a three-dimensional closed and connected $C^{\infty}$ Riemannian manifold $M$ endowed with a volume-form, if $m$ denotes the measure associated to it that we call \textit{Lebesgue measure}.
Let $\mathfrak{X}_{w}^{r}(M)$ be the space of conservative $C^{r}$ vector field, then
\ \\
\ \\
\noindent \textbf{Theorem A.}\,\emph{Let $\phi$ be a vector field, such that ${\phi}^{t}$ is an conservative Anosov flow which is a suspension (modulo time scale change by a constant factor) of the a conservative Anosov diffeomorphism $\varphi$ of a compact $C^{2}$ submanifold of codimension one in $M$. Then, arbitrarily close to $\varphi$ there is an open set $\mathcal{W}$ of $C^2$-conservative Anosov diffeomorphims such that for any $\psi\in \mathcal{W}$ we have
$$int\, M(f,\psi^t)\neq \emptyset \ \text{and} \ int\, L(f,\psi^t)\neq \emptyset$$
for any $f$ in a dense and $C^{2}$-open subset $\mathcal{U}_{\psi}$ of $C^{2}(M,\mathbb{R})$, where $\psi^{t}$ is the suspension flow associated to $\psi\in \mathcal{W}$.}\\
\ \\
\noindent \textbf{Theorem B.}\,\emph{Let $\phi$ be a vector field, such that ${\phi}^{t}$ is a conservative Anosov flow, then arbitrarily close to $\phi$ there is an open set $\mathcal{W}\subset \mathfrak{X}_{w}^{1}(M)$ such that for any $X\in \mathcal{W}$ we have
$$int\, M(f,X)\neq \emptyset \, \text{and}\, \, int\, L(f,X)\neq \emptyset$$
for any $f$ in a dense and $C^{2}$-open subset $\mathcal{U}_X$ of $C^{2}(M,\mathbb{R})$.
Moreover, the above statement holds persistently: for any $Y\in \mathcal{W}$, it holds for any $(f,X)$ in a suitable neighborhood of $\mathcal{U}_Y\times \{Y\}$ in $C^{2}({M,\mathbb{R}})\times \mathfrak{X}_{w}^{1}(M)$. Here $int A$ denotes the interior of $A$.}
\section{Preliminaries}
\noindent The flow $\phi^{t}$ is an Anosov flow, if there exists a continuous splitting $TM=E^{s}\oplus \phi\oplus E^{u}$, invariant under the derivative of the flow $D\phi$ on $TM$, such that ${\phi}$ is the subbundle spanned by the direction of geodesic flow, $D\phi$ exponentially expands $E^{u}$, and $D\phi$ exponentially contracts $E^{s}$, that is, there are constants $C,c>0$, $\lambda>1$ such that
\begin{eqnarray}\label{EEEE}
\left|D\phi^{t}(x)\right|&\geq& c\lambda^{t}\left|x\right| \ \ \ \ \ \text{if} \ \ \ x\in E^{u} \ \ \ \text{and} \ \ \ t\geq 0, \\ \nonumber \\
\left|D\phi^{t}(x)\right|&\leq& C\lambda^{-t}\left|x\right| \ \ \text{if} \ \ \ x\in E^{s} \ \ \ \ \text{and} \ \ \ t\geq 0.
\end{eqnarray}
\noindent The subbundles $E^s$ and $E^u$ are known to be uniquely integrable. They are tangent to
the strong stable foliation $W^s$ and strong unstable foliation $W^{u}$. The center stable and center unstable foliations are denoted respectively by $W^{cs}$ and $W^{cu}$ (cf. \cite{K}).\\
\noindent The main examples of Anosov flows are geodesic flows on unit tangent
bundles of compact Riemannian manifolds of negative curvature, and suspensions of Anosov diffeomorphisms.\\
Let $f\colon N \to N$ be an Anosov diffeomorphism of a compact manifold
$N$ (this means that there is an $f$-invariant splitting $TN = E^{s}\oplus E^{u}$ which
satisfies conditions analogous to (1) and (2) above) and consider the flow
$\eta_{t}$ on $N\times \mathbb{R}$ given by $\eta_{t}(x,s) = (x, s+t)$. The suspension of $f$ is the flow
induced by $\eta_t$ on the manifold obtained from $N\times \mathbb{R}$ by making the identifications
$(x,s)\sim(f(x),s+1)$.\ \\
\\
Denoted by $\mathcal{M}$ the set of invariant probability measure. An Anosov flow (or Anosov diffemorphism) is conservative, if there is a invariant probability measure $\tilde{m}$, which is absolutely continuous with respect to normalized Lebesgue measure.
\section{Hyperbolic set with larger Hausdorff dimension}\label{HSLHD}
In this section we will prove that for conservative Anosov flows and conservative Anosov diffemorphisms, there exists hyperbolic set that having Hausdorff dimension close to $3$ and $2$ respectively. More precisely:
\begin{Le}\label{Le1} Let $\phi^{t}$ be a conservative Anosov flow, then there exists a hyperbolic set $\Lambda\subsetneq M$ such that
$HD({\Lambda})\ \ \text{is close to} \ \ 3,$
where $HD(A)$ is the Hausdorff dimension of set $A$.
\end{Le}
\begin{Le}\label{Le1'}
If $\psi$ is a conservative diffeomorphism on the $2$-torus, then there is hyperbolic set $\Lambda$ with Hausdorff dimension close $2$.
\end{Le}
\noindent For the proof of these Lemmas, we use the following theorem.\\
\ \\
\textbf{Theorem}[Urba\'nski, \cite{Urbanski}]\,\emph{If $M$ is a compact Riemannian manifold and $\varphi \colon M \to M$ is a transitive Anosov diffeomorphism, then the Hausdorff dimension of the set of points with non dense (full) orbit under $\varphi$ equals dim $M$. The same statement is proved for Anosov flows.}
\begin{proof}[\bf{Proof of Lemma \ref{Le1}}]
Consider $\{x_k\}$ an enumerable and dense set, then for each $m\in \mathbb{N}$, we define the set $A_{m}^{k}:=M\setminus B_{\frac{1}{m}}(x_k)$, where $B_{\frac{1}{m}}(x_k)$ is the ball of center $x_k$ and radius $\frac{1}{m}$ and $\tilde{A}_{m}^{k}= \displaystyle\bigcap_{t\in\mathbb{R}}\phi^{t}(A_{m}^{k})$. This set is compact, invariant and therefore hyperbolic set for $\phi^{t}$. Then\\
\textit{Claim:} If $\mathcal{ND}$ is the set of points with non-dense orbit under $\phi^{t}$, then
$$\mathcal{ND}\subset \displaystyle \bigcup_{m\geq 1} \bigcup_{k}\tilde{A}_{m}^{k}.$$
\textit{Proof of Claim:} Let $x\in \mathcal{ND}$, then there is an open set $U\subset M$ such that the orbit of $x$, $O(x)$ does not intersects $U$, \emph{i.e.} $O(x)\cap U=\emptyset$. Thus, there are $m\geq 1$ and $x_k$ such that $B_{\frac{1}{m}}(x_k)\subset U$, therefore $\phi^{t}(x)\notin B_{\frac{1}{m}}(x_k)$ or $\phi^{t}(x)\in M\setminus B_{\frac{1}{m}}(x_k)$ for all $t\in \mathbb{R}$, this implies that $x\in \tilde{A}_{m}^{k}.$ Thus we concluded the proof of claim.\\
The Urba\'nski's Theorem implies that $HD(\mathcal{ND})=3$, then by the previous Claim, $$3=HD(\displaystyle \bigcup_{m\geq 1} \bigcup_{k}\tilde{A}_{m}^{k})=\displaystyle \sup_{k,m}\tilde{A}_{m}^{k}.$$
So, there are $m$ and $k$ such that $HD(\tilde{A}_{m}^{k})$ is very close to $3$.
\end{proof}
\noindent The proof of Lemma \ref{Le1'} is similar.
\section{Poincar\'e Map and Hyperbolicity}\label{PMCS}
In this section we construct a hyperbolic set with Hausdorff dimension close to 2 for
Poincar\'e map associated for some cross-section (relate to compact hyperbolic set $\Lambda$ given in the Lemma \ref{Le1}) for the flow. The arguments are similar to \cite[section 3]{RM2}. From now on, let us fixed the hyperbolic set given in Lemma \ref{Le1}.\\
\ \\
Now, since our flow is a conservative flow, in particular transitive, then the following result due to J. Plante characterized the flow in terms of stable and unstable manifolds. \\
\ \\
\noindent \textbf{Theorem}\,[Plante]\, \, \textit{For a compact manifold, let $\phi^{t}$ be a transitive Anosov flow. Then there are exactly two possibilities:
\begin{enumerate}
\item[(i)] Each strong stable and each strong unstable manifold is dense in $M$, or
\item[(ii)] $\phi^{t}$ is the suspension (modulo time scale change by a constant factor) of the Anosov diffeomorphism of a compact $C^{1}$ submanifold of codimension one in $M$.
\end{enumerate}
}
\noindent In the second case the bundle $E^{s}\oplus E^{u}$ is integrable and if the flow is conservative then \cite[Theorem 3.7]{Plante} implies that the flow topologically equivalent to the suspension of a linear toral automorphism.
\noindent The condition $ii)$ will be used for the proof of Theorem $A$.
\subsection{Good Croos-Sections}
By the Plante's Theorem, we can assume that the flow is not suspension, \emph{i.e.} $E^{s}\oplus E^{u}$ is not integrable.\\
Let $\mathcal{F}^{s}$ be the strong stable foliations and $\mathcal{F}^{u}$ the strong unstable foliations, this is $\mathcal{F}^{i}(x)=W^{i}(x)$ for $i=s, u$, are continuous foliations of dimension one (not necessari-ly $C^1$-foliations).
\begin{Defi} A $C^{0}$-surface $S$ is transverse to the flow $\phi^t$, if there are $\theta, r >0$ such that for every $z\in S$ the cone $C_z$ of angle $\theta$ centered in $\phi(z)$ with vertex at the point $z$ satisfies $C_z\cap B_{r}(z)\cap S=\{z\}$.
\end{Defi}
\begin{Le}\label{Le5}
Let $x \in M$ and $L$ be a $C^{1}$-embedded curve of dimension one, containing x and transverse to the foliation $\mathcal{F}^{u}$, then the set
\begin{center}
$\displaystyle S_{L}:=\bigcup_{z\in L}\mathcal{F}^{s}(z)$
\end{center}
contains a surface $S_{x}$ that is $C^{0}$-embedded, which contains $x$ in the interior and if $L$ is transverse to the foliation $W^{cs}$ then, $S_{x}$ is transverse to the flow.
\end{Le}
\noindent The proof of this Lemma is similar to \cite[Lemma 1]{RM2}.
\noindent In particular, taking $L=W^{u}_{\epsilon}(x)$ with $\epsilon$ given by the stable and unstable manifolds theorem, we call $S_{x}:=\Sigma_{x}$. Note that an analogous Lemma holds for the foliation $\mathcal{F}^{s}$.
\begin{R}\label{R1}
Without loss of generality, we can assume that $\Sigma_{x}$ is homeomorphic to the square $[0,1]\times[0,1]$. Put $\Sigma_{x}=\Sigma$, with the horizontal lines $[0,1]\times \eta$ being mapped to stable sets $W^{s}(y,\Sigma_{x})=W^{ss}(y)\cap \Sigma_{x}$. The stable-boundary $\partial^{s}\Sigma$ is the image of $[0,1]\times\left\{0,1\right\}$, the unstable-boundary $\partial^{u}\Sigma$ is the image of $\left\{0,1\right\}\times [0,1]$. Therefore, we have the following definition.
\end{R}
\begin{Defi}\label{D1GCS}
A cross sections is said \textbf{$\boldsymbol{\delta}$-Good Cross-Section} for some $\delta>0$, if satisfies the following:
\begin{center}
$d(\Lambda \cap \Sigma, \partial^{u}\Sigma)>\delta$
and \ \
$d(\Lambda \cap \Sigma, \partial^{s}\Sigma)>\delta$
\end{center}
where $d$ is the intrinsic distance in $\Sigma$, (cf. Figure \ref{fig:f7}).\\
A cross-section which is $\delta$-Good Cross-Section for some $\delta>0$ is said a \textbf{Good Cross-Section-GCS}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.4\textwidth]{f7.pdf}
\caption{Good Cross-Section}
\label{fig:f7}
\end{figure}
\end{Defi}
\noindent It follows easily that
\begin{R}\label{R2'}
Let $\Sigma$ be a $\delta$-Good Cross-Section, then given $0<\delta'<\delta$ there is a $\delta'$-Good Cross-Section $\Sigma'\subset int(\Sigma)$ and such that $\partial \Sigma'\cap \partial \Sigma=\emptyset$. Therefore, from now on we can assume that if two GCS has nonempty intersection, then their interiors have nonempty intersection.
\end{R}
\noindent Analogously to \cite[Lemma 5]{RM2} we have that
\begin{Le}\label{Le6}
Let $x\in \Lambda$, then there is $\delta>0$ and a $\delta$-Good Cross-Section $\Sigma$ at $x$.
\end{Le}
\noindent The $\delta$-Good Cross-Section given in the previous Lemma it is constructed using the fact that for transitive Anosov flows and for any hyperbolic set $\Lambda\subsetneq M$, we have the property that for any $x\in \Lambda$ there are points $x_{s,u}^{+}\notin \Lambda$ and $x_{s,u}^{-} \notin \Lambda$ in distinct connected components of $W^{s,u}(x)-\left\{x\right\}$, respectively (see \cite[Lemma 3 and Lemma 4]{RM2}). Moreover, the section $\Sigma\subset \Sigma_{x}$, where $\Sigma_{x}$ is as the Lemma \ref{Le5} which is saturated by the foliation $\mathcal{F}^{s}$.
\begin{R}\label{R3}
In the proof of Lemma \ref{Le6} (cf. \cite[Lemma 5]{RM2}), can be observed that there is a family of GCS $\{\Sigma_{x}^{k}\}_{k}$ as above $(\Sigma_x:=\Sigma^{0}_x)$, with the property that the unstable boundary of $\Sigma^{k}_x$ is sufficiently close of $W^{u}_{\epsilon}(x)\cap \Sigma_x$, and also a family of GCS $\{\widetilde{\Sigma}_{x}^{k}\}_{k}$ as above $(\Sigma_x:=\widetilde{\Sigma}^{0}_x)$, with the property that the stable boundary of $\Sigma^{k}_x$ is sufficiently close of $W^{s}_{\epsilon}(x)\cap \Sigma_x$.
\end{R}
\noindent This kind of cross-sections has good properties. Before showing of the properties of these sections, remember that they are $C^{0}$ sections, so we will need one definition of $C^{0}$ transversal intersection.
\begin{Defi}\label{D2}
We say that a continuous curve $\xi\subset \mathbb{R}^2$ is $\theta$-transverse in neighborhood of radius $r$ to $1$-dimensional foliation $\mathcal{F}\subset \mathbb{R}^2$, if for any $z\in \xi\cap \mathcal{F}$ there is a cone $C$ with vertex at the point $z$ such that $\zeta \cap B(z,r)\subset C$ and the angle $\angle(v, T_{z}\mathcal{F}(z))\geq \theta$ for every tangent vector $v$ at the point $z$ contained in the cone $C$.
\end{Defi}
\noindent As the section $\Sigma_x$ is saturated by the foliation $\mathcal{F}^{s}$, then call $h_{x}$ the homeomorphism given in Remark \ref{R1}, then we have the following definition
\begin{Defi}\label{D3}
We say that a continuous curve $\zeta\subset \Sigma_x$ is transverse to foliation $\mathcal{F}^{s}\cap \Sigma_x$, if there are $\theta$ and $r$ such that \, $h_{x}(\zeta)$ is $\theta$-transverse in neighborhood of radius $r$ to foliation and $h_x(\mathcal{F}^{s}\cap \Sigma_x)$.
\end{Defi}
\begin{Pro}\label{P3}
Given $x,y \in \Lambda$, such that there is a $C^{0}$-curve $\zeta\subset int(\Sigma_{x})\cap int(\Sigma_{y})$. If $\zeta$ intersects transversely to foliation $\mathcal{F}^{s}$, then $int(\Sigma_{x}) \cap int(\Sigma_{y})$ is an open set of $\Sigma_{x}$ and $\Sigma_{y}$.
\end{Pro}
\begin{proof}
Since $\zeta\subset int(\Sigma_{x})\cap int(\Sigma_{y})$ a $C^{0}$-curve transverse to $\mathcal{F}^{s}$. Then for all $z\in \zeta$, there are $x^{\prime} \in W^{uu}_{\epsilon}(x)$ and $y^{\prime} \in W^{uu}_{\epsilon}(y)$ such that $z\in W^{s}(x^{\prime})\cap \Sigma_{x}$ and $z\in W^{s}(y^{\prime})\cap \Sigma_{y}$, thus $W^{s}(x^{\prime})=W^{s}(y^{\prime})$. Therefore, there is $\delta>0$ such that the set
\begin{center}
$\displaystyle B=\bigcup_{z\in \zeta}W^{s}_{\delta}(z)\subset int(\Sigma_{x}) \cap int(\Sigma_{y})$.
\end{center}
Thus, we concluded the proof of proposition.
\end{proof}
\subsubsection{Separation of GCS}\label{Separation of GCS}
\noindent By Lemma \ref{Le6}, at each point of $x\in \Lambda$, we can find a Good Cross-Section $\Sigma_{x}$. Since $\Lambda$ is a compact set, then for $\gamma>0$, there are a finite number of points $x_{i}\in \Lambda$, $i=1,\dots,l$ such that
\begin{eqnarray}\label{E6GCS}
\displaystyle \Lambda\subset \bigcup^{l}_{i=1}\phi^{(-\frac{\gamma}{2},\frac{\gamma}{2})}(int{\Sigma_{i}})\subset \bigcup^{l}_{i=1}\phi^{(-\gamma,\gamma)}(int{\Sigma_{i}})=\bigcup^{l}_{i=1}U_{\Sigma_{i}},
\end{eqnarray}
where $\Sigma_{i}:=\Sigma_{x_i}$.
\ \\
In this section we prove that the $\{\Sigma_{i}\}_{i}$ can be taken pairwise disjoint.\\
Note that if two sections $\Sigma_i$, $\Sigma_j$ has nonempty intersection, then by Proposition \ref{P3} we have
\begin{itemize}
\item[1.] $\Sigma_i\cap\Sigma_j$ is totally transverse to foliation $\mathcal{F}^{s}$, \emph{i.e.} for any $x\in int\,\Sigma_i\cap int\,\Sigma_j$, there is a curve $\xi_x\subset int \,\Sigma_i\cap int\,\Sigma_j$ which is transverse to $\mathcal{F}^{s}$, then the Proposition \ref{P3} implies that $int\,\Sigma_i\cap int\,\Sigma_j$ is an open set of $\Sigma_i$ and $\Sigma_j$.
\item[2.]$\Sigma_i\cap\Sigma_j$ does not intersect totally transverse to foliation $\mathcal{F}^{s}$, here can happen two cases:
\begin{itemize}
\item[i)] For all point $x \in \Sigma_i\cap\Sigma_j$ there is not a curve $\xi_x\subset int \,\Sigma_i\cap int\,\Sigma_j$ transverse to $\mathcal{F}^{s}$, this implies $\Sigma_i\cap\Sigma_j$ does not contains open sets of $\Sigma_i$ and $\Sigma_j$.
\item[ii)] $\Sigma_i\cap\Sigma_j$ contains open sets of $\Sigma_i$ and $\Sigma_j$ and point as in i).
\end{itemize}
\end{itemize}
\textbf{Remark:}\textit{\,Note that, in the geodesic flow case as \cite{RM2}, $\pi^{s}_i(\Sigma_i\cap \Sigma_j)= \{W^{s}_{\epsilon}(x)\cap W^u_{\epsilon}(x_i)\colon x\in \Sigma_i \cap \Sigma_j\}$ is a finite set, because in this case the GCS are $C^1$ and transverse.}
\ \\
The following Lemma show that if two GCS are as in 2\,ii), then a small translation in the time in one of two sections makes that they are in the conditions 2\,i).
\begin{Le}\label{L7'0}
Let $\Sigma_i$, $\Sigma_j$ be as in (\ref{E6GCS}) in the condition \text{2\,ii)}, then there is $t'$ small such $\phi^{t'}(\Sigma_i)$ and $\Sigma_j$ are as in 2\,i).
\end{Le}
\begin{proof}
Otherwise, if for all $t$ small we have that $int \, (\phi^{t}(\Sigma_i)\cap int \, \Sigma_j)$ contains an open set of $\phi^{t}(\Sigma_i)$ and $\Sigma_j$, then there is a nondegenerate interval $I^{j}_{t}\in W_{\epsilon}^{u}(x_j)\subset \Sigma_j$ and a nondegenerate interval $I^i_t\subset {W_{\epsilon}}^{u}(x_i)\subset \Sigma_i$
such that the set
\begin{equation}\label{EQ5'}
\Delta_t:=\bigcup_{z\in \phi^{t}(I^{i}_{t})}W^s_{\epsilon}(z)\cap \bigcup_{w\in I^{i}_{t}}W^s_{\epsilon}(w)
\end{equation}
contains an open set of $\phi^{t}(\Sigma_i)$ and $\Sigma_j$.\ \\
\ \\
\textit{Claim:} The family of interval $I^{j}_{t}$ are pairwise disjoint. Indeed, if there is $x\in I^{j}_{t}\cap I^j_{t'}\subset W^{u}_{\epsilon}(x_j)$ with $t\neq t'$ then by (\ref{EQ5'}) there are $y\in I^{i}_{t}$ and $z\in I^{i}_{t'}$ such that $\phi^{t}(y), \, \phi^{t'}(z)\in W^{s}_{\epsilon}(x)$.
Therefore, since $t,t'$ are small
$$\{ \phi^{t'}(z)\}\subset W^{cu}_{loc}(\phi^{t}(y))\cap W^{s}_{\epsilon}(\phi^{t}(y))$$ this last set, by the stable and unstable theorem, it is contained in the orbit of $y$, $\mathcal{O}(y)$.
Thus if $\phi^{\eta}(\phi^t(y))=\phi^{t'}(z)\in W^{s}_{\epsilon}(x)$ and $\phi^{t}(y)\in W^{s}_{\epsilon}(x)$, then $\eta=0$ and $\phi^t(y)=\phi^{t'}(z)$, so as $z,y\in W^{u}_{\epsilon}(x_i)$ we have that $t=t'$ and $z=y$ which is a contradiction, so we concluded the claim.
\ \\
\noindent This claim implies that the family of nondegenerate intervals $\{I^{j}_{t}\}_t$ of $W^u_{\epsilon}(x_j)$ is pairwise disjoint, which can not happen because $\epsilon$ is finite.
\end{proof}
\noindent This Lemma implies that we can assume that $\Sigma_i$ and $\Sigma_j$ are as in 1) and 2\,i). Let's understand what happen in the case 2\,i).\\
In the case 2\,i) $\Sigma_i \cap \Sigma_j$ is a family $\Gamma_{ij}$ of curves $\gamma$ and by construction each element $\gamma\in \Gamma_{ij}$ is a leaf of the foliation $\mathcal{F}^{s}\cap \Sigma_i$. Remember that $\Sigma_i=\Sigma_{x_i}$, then consider the projection $\Pi^{s}_i\colon \Sigma_i\to W^u_{\epsilon}(x_i)$ along to $\mathcal{F}^{s}$.\\%which is a continuous function.\\
\ \\
\textit{Claim:} $\pi^{s}_i(\Sigma_i\cap \Sigma_j)= \{W^{s}_{\epsilon}(x)\cap W^u_{\epsilon}(x_i)\colon x\in \Sigma_i \cap \Sigma_j\}$ is a compact set. Indeed, we need only to show that $\pi^{s}_i(\Sigma_i\cap \Sigma_j)$ is a closed set. Let $x_n\in \pi^{s}_i(\Sigma_i\cap \Sigma_j)$, such that $x_n \to x$, then there is $y_n\in \Sigma_i\cap \Sigma_j$ such that $W^{s}_{\epsilon}(y_n)\cap W^u_{\epsilon}(x_i)=x_n$, since $\Sigma_i\cap \Sigma_j$ is compact, then we can assume that $y_{n_k} \to y\in \Sigma_i\cap \Sigma_j$, moreover by continuity of foliation $\mathcal{F}^s$ we have that $W^{s}_{\epsilon}(y_{n_k})\cap W^u_{\epsilon}(x_i)\to \pi^{s}_{i}(y)$ and $W^{s}_{\epsilon}(y_{n_k})\cap W^u_{\epsilon}(x_i)=x_{n_k}\to x$, so $x=\pi^{s}_{i}(y)\in \pi^{s}_i(\Sigma_i\cap \Sigma_j)$.\\
\ \\
By transversality, given $\delta>0$, then for each $x \in\pi^{s}_i(\Sigma_i\cap \Sigma_j)$, we have that
$$\phi^{\delta}(W^s_{\epsilon}(x)\cap \Sigma_i)\cap \Sigma_j=\emptyset,$$
and by continuity we have that there is $U_x$ neighborhood of $W^s_{\epsilon}(x)\cap \Sigma_i$ in $\Sigma_i$ such that
\begin{equation}\label{eq-sep}
\phi^{\delta}(U_x\cap \Sigma_i)\cap \Sigma_j=\emptyset,
\end{equation}
\noindent namely, there is an interval centered in $x$, $I_{x}\subset W^{u}_{\epsilon}(x_i)$ such that $U_x=\bigcup_{z\in I_x} W^{s}_{\epsilon}(z)\cap \Sigma_i$.\\
If $\mathcal{F}^{s}(x)\cap \Sigma_{i}\cap \Lambda=\emptyset$ for some $x\in\pi^{s}_i(\Sigma_i\cap \Sigma_j)$, then since $\Lambda$ is a compact set there is an open set $V_{x}$ containing $\mathcal{F}^{s}(x)\cap \Sigma_i$ and $V_{x}\cap\Lambda=\emptyset$. Therefore, $\Sigma_{i}$ is subdivided into two GCS $\Sigma_{i}^{1}$ and $\Sigma_{i}^{2}$, such that $\Sigma_{i}^{r}$ and $\Sigma_j$ intersect as 2\,i) for $r=1,2$.\\
The above implies that, without loss of generality, we can assume that for any $x\in\pi^{s}_i(\Sigma_i\cap \Sigma_j)$ there is $p_{x}\in \mathcal{F}^{s}(x)\cap \Sigma_i\cap \Lambda$.
\begin{Le}\label{L8GCS}
If $\Sigma_i$ and $\Sigma_j$ are two GCS as in the condition 2i). Given $\delta>0$, $0<\delta<\frac{\gamma}{2}$ with $\gamma$ as in (\ref{E6GCS}) for $x\in \pi^{s}_i(\Sigma_i\cap \Sigma_j)$,
there is a $GCS$ $\widetilde{\Sigma}_{x}\subset U_x\cap{\Sigma_i}$ containing $\mathcal{F}^{s}(x)\cap \Sigma_i$, such that $\Sigma_{i}$ is subdivided into three GCS disjoint, including $\widetilde{\Sigma}_{x}$. Denoted by $\Sigma_{i}^{\#}$ the set of complementary sections in the subdivision above of $\Sigma_{i}$, then
\begin{enumerate}
\item $\phi^{\delta}(\widetilde{\Sigma}_{x})\cap \Sigma_j=\emptyset$
\item $ \Lambda\cap \phi^{(-\frac{\gamma}{2},\frac{\gamma}{2})}(int(\Sigma_i))\subset \Lambda \cap \left(\phi^{\delta}(int(\widetilde{\Sigma}_{x}))\cup \bigcup_{\Sigma\in \Sigma_{i}^{\#}}\phi^{(-\frac{\gamma}{2},\frac{\gamma}{2})}(int(\Sigma))\right).$
\end{enumerate}
\end{Le}
\begin{proof}
\noindent We can assume that for any $x\in\pi^{s}_i(\Sigma_i\cap \Sigma_j)$ there is $p_{x}\in \mathcal{F}^{s}(x)\cap \Sigma_i\cap \Lambda$. Consider $\mathcal{F}^{u}_{loc}(p_{x})$, then by Remark \ref{R3} we can find open sets $V_{p_x}^{+}$ and $V_{p_x}^{-}$ in each side of $\mathcal{F}^{u}_{loc}(p_{x})\setminus\{p_x\}$ sufficiently close to $\mathcal{F}^{s}_{loc}(p_{x})$ and diameter sufficiently large with $V_{p_x}^{\pm}\cap \Lambda=\emptyset$. Denoted by $\widetilde{V}_{p_x}^{\pm}$ the projection by the flow of $V_{p_x}^{\pm}$ over $\Sigma_{i}$, respectively.
Therefore, by Remark \ref{R3} we can take $\widetilde{V}_{p_x}^{\pm}$ such that $\widetilde{V}_{p_x}^{\pm}\cap \Sigma_{i}\subset U_x$ and $\widetilde{V}^{\pm}_{p_x}$ crosses $\Sigma_{i}$. Using $\widetilde{V}^{\pm}_{p_x}$ we can construct the GCS $\widetilde{\Sigma}_{x}$ such that $\widetilde{\Sigma}_{x}\subset U_x$ and by (\ref{eq-sep}) satisfies 1).\\
To prove item 2) note simply that $\delta<\frac{\gamma}{2}$ and $\Lambda\cap \phi^{(-\frac{\gamma}{2},\frac{\gamma}{2})}(int(\Sigma_i))=\phi^{(-\frac{\gamma}{2},\frac{\gamma}{2})}(\Lambda\cap int(\Sigma_{i}))$, which is a consequence of $\Lambda$ be invariant by the flow.
\end{proof}
\noindent As the GCS $\widetilde{\Sigma}_{x}$ obtained in this Lemma is contained in $U_x\cap{\Sigma_i}$, then there is an interval centered in $x$, $\widetilde{I}_x\subset I_x\subset W^{u}_{\epsilon}(x_i)$ such that
$$\widetilde{\Sigma}_x = \displaystyle\bigcup_{z\in\widetilde{I}_x}W^{s}_{\epsilon}(z)\cap \Sigma_i.$$
Now $\pi^{s}_i(\Sigma_i\cap \Sigma_j)\subset \bigcup \widetilde{I}_{x}$ and by last claim, we have that there is a finite set of points in $\pi^{s}_i(\Sigma_i\cap \Sigma_j)$ let's say $\{x^1,\dots,x^m\}$ such that
$$\pi^{s}_i(\Sigma_i\cap \Sigma_j)\subset \bigcup_{r=1}^{m} \widetilde{I}_{x^r}.$$
Thus by first part of Lemma \ref{L8GCS} hold that
$$\phi^{\delta}(\widetilde{\Sigma}_{x^{r}})\cap \Sigma_j=\emptyset \ \ \ \text{for} \ \ \ r=1,\dots,m \ \ \text{and} \ \ \phi^{\delta}(\widetilde{\Sigma}_{x^{r}})\cap\phi^{\delta}(\widetilde{\Sigma}_{x^{r'}})=\emptyset \ \ \text{for} \ \ r\neq r'.$$
In other words, we have the following Lemma, which is similar to \cite[Lemma 8]{RM2}
\begin{Le}\label{L9GCS}
If $\Sigma_i$ and $\Sigma_j$ are two GCS as in the condition 2i). Given $0<\delta<\frac{\gamma}{2}$ with $\gamma$ as in (\ref{E6GCS}) there are $GCS$ $\widetilde{\Sigma}_{x^r}\subset U_{x^r}$ containing $\mathcal{F}^{ss}(x^r)\cap \Sigma_x$, such that $\Sigma_{i}$ is subdivided into $2m+1$ GCS disjoint, including $\widetilde{\Sigma}_{x^r}$ for $r\in\{1,\dots,m\}$. Denoted by $\Sigma_{i}^{\#}$ the complement of the set $\{\widetilde{\Sigma}_{x^r}\}_{r=1}^{m}$ in the subdivision above of $\Sigma_{i}$, then
\begin{enumerate}
\item $\phi^{\delta}(\widetilde{\Sigma}_{x^r})\cap \Sigma_j=\emptyset$ for all $r\in\{1,\dots,m\}$ and $\Sigma'\cap \Sigma=\emptyset$ for any $\Sigma\in \Sigma_{x}^{\#}$.
\item $\phi^{\delta}(\widetilde{\Sigma}_{x^r})\cap\phi^{\delta}(\widetilde{\Sigma}_{x^{r'}})=\emptyset$ for $r\neq r'$ and
$\phi^{\delta}(\widetilde{\Sigma}_{x^r})\cap \Sigma_{i}=\emptyset$ for all $r\in\{1,\dots,m\}$.
\item $\Lambda\cap \phi^{(-\frac{\gamma}{2},\frac{\gamma}{2})}(int(\Sigma_i))\subset \Lambda \cap \left(\bigcup^{m}_{r=1}\phi^{(-\gamma,\gamma)} \left(\phi^{\delta}(int(\widetilde{\Sigma}_{x^r}))\right)\cup \bigcup_{\Sigma\in \Sigma_{i}^{\#}}\phi^{(-\frac{\gamma}{2},\frac{\gamma}{2})}(int(\Sigma))\right).$
\end{enumerate}
\end{Le}
\noindent This last Lemma together with the proof of \cite[Lemma 7, Lemma 9 Lemma 10]{RM2} we have that
\begin{Le}\label{L7}
There are GCS $\widetilde{\Sigma}_{i}$ such that
\begin{equation}\label{eq6}
\displaystyle\Lambda\subset \bigcup^{m(l)}_{i=1}\phi^{(-2\gamma,2\gamma)}(int(\widetilde{\Sigma}_i))
\end{equation}
with $\widetilde{\Sigma}_i\cap \widetilde{\Sigma}_j=\emptyset$.
\end{Le}
\noindent The sections $\widetilde{\Sigma}_i$ given in previous Lemma are as the Lemma \ref{Le6}, that is, they are $C^{0}$-GCS and saturated by the foliation $\mathcal{F}^{s}$. Since $C^{\infty}$ is dense in $C^{0}$, we can assume without loss of generality, that there are sections $\Sigma_{i}$ than are $C^{\infty}$-GCS and (\ref{eq6}) is satisfied.
\subsection{Hyperbolicity of Poincar\'e Maps}\label{Hyp of PM}
Let $\Xi=\bigcup_{i=1}^{m(l)} \Sigma_{i}$ be finite union of cross-sections to the flow $\phi^{t}$ without loss of generality, assume that $m(l)=l$ and since the sections are pairwise disjoint, then sometimes we consider $\Xi=\{\Sigma_1,\cdots, \Sigma_l\}$. Let ${\mathcal{R}}\colon \Xi \to \Xi$ be a Poincar\'e map or the map of first return to $\Xi$, ${\mathcal{R}}(y)=\phi^{t_{+}(y)}(y)$, where $t_{+}(y)$ correspond to the first time that the orbits of $y\in \Xi$ encounter $\Xi$.
\ \\
\ \\
The splitting $E^{s}\oplus \phi\oplus E^{u}$ over neighborhood $U_{0}$ of $\Lambda$ defines a continuous splitting $E^{s}_{\Sigma}\oplus E^{u}_{\Sigma}$ of the tangent bundle $T\Sigma$ with $\Sigma\in \Xi$, defined by
\begin{eqnarray}\label{eq7}
E^{s}_{\Sigma}(y)=E^{cs}_{y}\cap T_{y}\Sigma \ and \ E^{u}_{\Sigma}(y)=E^{cu}_{y}\cap T_{y}\Sigma
\end{eqnarray}
where $E_{y}^{cs}=E^{s}_y\oplus \left\langle \phi(y)\right\rangle$ and $E_{y}^{cu}=E^{u}_y\oplus \left\langle {\phi}(y)\right\rangle$.\\
\ \\
\noindent Analogously as in \cite[Proposition 1]{RM2}, we show that since $\Sigma_i$ is GCS for all $i$, then for a sufficiently large iterated of ${\mathcal{R}}$, ${\mathcal{R}}^{n}$, then (\ref{eq7}) define a hyperbolic splitting for transformation ${\mathcal{R}}^{n}$, at least restricted to $\Lambda$ and since $\Lambda\cap \Xi$ is invariant for $\mathcal{R}$, then $\Lambda\cap \Xi$ is hyperbolic for $\mathcal{R}$, and
$$\Lambda\cap \Xi\subset \bigcap_{n\in \mathbb{Z}}{\mathcal{R}}^{-n}(\bigcup^{l}_{i=1}\Sigma_{i}).$$
The Lemma 13 and Lemma 14 in \cite{RM2} implies that
$$HD\left( \bigcap_{n\in \mathbb{Z}}{\mathcal{R}}^{-n}(\bigcup^{l}_{i=1}\Sigma_{i})\right) \geq HD\left(\Lambda \cap (\bigcup^{l}_{i=1}\Sigma_{i})\right) \geq HD(\Lambda)-1$$ and thus
by Lemma \ref{Le1} $HD(\Lambda\cap\Xi)$ is close to $2$.
\section{Proof of Theorem A and Theorem B}
Let $\varphi\colon N\to N$ be the $C^{2}$-conservative Anosov diffeomorphim by which $\phi^t$ is suspension (modulo time scale change by constant factor), then by Lemma \ref{Le1'} there is a hyperbolic set $\Lambda$ for $\varphi$ with $HD(\Lambda)$ close to $2$.
Given $F\in C^{0}(M,\mathbb{R})$, we can define the function $maxF_{\phi}\colon N\to\mathbb{R}$ by
$$max F_{\phi}(x):=\max_{-1\leq t \leq 1}F(\phi^{t}(x)).$$
Note that this definition depends on the flow $\phi^{t}$, or equivalently the vector field $\phi$. Note also that $max F_{\phi}$ is always a continuous function, but even if $F$ is $C^{\infty}$, $max F_{\phi}$ can be only a continuous function. The Lemma 18 in \cite{RM2}) gives some ``differentiability" to $max F_{\phi}$ at least for $F\in C^{2}(M,\mathbb{R})$.
\begin{Le}\label{Lmax0}
There exists a dense $\mathcal{B}_{\phi}\subset C^{\infty}(M,\mathbb{R})$ and $C^{2}$-open such that given $\beta>0$, then for any $F\in \mathcal{B}_{\phi}$ there are sub-horseshoe $\Lambda_{F}^{s,u}$ of $\Lambda$ with $HD(K_{F}^{s})\geq HD(K^{s})-\beta$, $HD(K_{F}^{u})\geq HD(K^{u})-\beta$ and a Markov partition $R_{F}^{s,u}$ of $\Lambda_{F}^{s,u}$, respectively, such that the function $max F_{\phi}|_{N\cap R_{F}^{s,u}}\in C^{1}(N\cap R_{F}^{s,u},\mathbb{R})$, where $K_{F}^{s,u}$, $K^{s,u}$ are the stable and unstable Cantor set, respectively.
\end{Le}
\begin{R}\label{RLmax}
Observe that if $g$ is $C^{2}$-close to $\varphi$ and $\phi_{g}$ is the suspension flow for $g$, then hold that
\begin{equation}\label{eqLmax}
max F_{\phi_g}|_{N\cap R_{F}^{s,u}}\in C^{1}(N\cap R_{F}^{s,u},\mathbb{R}).
\end{equation}
\end{R}
\subsection{Proof of Theorem A}
The proof of Theorem A is a consequence of Lemma \ref{Le1'} and \cite[Main Theorem]{RM}.
\begin{proof}[\bf{Proof of Theorem A}]
Let $\varphi\colon N\to N$ be the $C^{2}$-conservative Anosov diffeomorphim by which $\phi^t$ is suspension (modulo time scale change by constant factor), then by Lemma \ref{Le1'} there is a hyperbolic set $\Lambda$ for $\varphi$ with $HD(\Lambda)$ close to $2$, then by \cite[Main Theorem]{RM} we have that there is arbitrarily close to $\varphi$, neighborhood $\mathcal{W}\subset Diff_{c}^{2}(N)$ (the space of $C^{2}$-conservative diffeomorphism on $N$) \footnote{The perturbation in \cite[Main Theorem]{RM} can be make conservative cf. \cite[section 4.3]{MY1}. } such that if $\Lambda_{g}$ denotes the continuation of $\Lambda$ associated to $g \in \mathcal{W}$, ($g$ is still Anosov diffemorphism), then
there is an open and dense set $H_g\subset C^1(N,\mathbb{R})$ such that for all $h\in H_g$, we have
$$int L(h,\Lambda_g) \neq \emptyset \ \ \text{and} \ \ int M(h,\Lambda_g) \neq \emptyset,$$
where $L(h,\Lambda_g)$ e $M(h,\Lambda_g)$ are the Lagrange and Markov dynamical spectrum respectively (see section 1).\\
For $g\in \mathcal{W}$, we denote by $\phi_{g}$ the vector field associated to the suspension flow of $g$.
Let $B_{\phi_{g}}$ be as in the Lemma \ref{Lmax0} and let $F\in B_{\phi_{g}}$ then using local coordinates with respect to the vector field $\phi_{g}$, we can find $j\in C^2(N,\mathbb{R})$ such that
\begin{equation}\label{E1PT1}
max F_{\phi_{g}}|_{N\cap (R_{F}^{s}\cup R_{F}^{u})}(x_1,x_2,x_3)+j(x_1,x_2)\in H_g.
\end{equation}
Put $h(x_1,x_2,x_3)=F(x_1,x_2,x_3)+j(x_1,x_2)$, then $max h_{\phi_g}=max F_{\phi_g}+j\in H_g$.
Therefore, since $M(max h_{\phi_g},{\Lambda}_g)=\left\{\sup_{n\in\mathbb{Z}}max h_{\phi_g}({g}^{n}(x)):x\in\Lambda_g\right\}\subset M(h,\phi_{g})$.
So, $$int M(h,\phi_g)\neq\emptyset.$$
Analogously, $L(max h_{\phi_g},{\Lambda}_g)\subset L(h,\phi_g)$, therefore $int L(h,\phi_g)\neq\emptyset$.
\end{proof}
\section{Proof of Theorem B}
\noindent To prove the Theorem B use the results of section \ref{PMCS} and \cite{RM} and a construction for obtain the property $V$ (cf. section \ref{SIRCS} and \cite[subsection 4.2]{RM2}).\\
\ \\
\noindent In the section \ref{PMCS} it was proven that there are a finite number of $C^{\infty}$-GCS, $\Sigma_{i}$ pairwise disjoint and such that
the Poincar\'e map $\mathcal{R}$ (map of first return) of $\Xi:=\bigcup^{l}_{i=1}\Sigma_{i}$ $$\mathcal{R} \colon \Xi\to \Xi$$
satisfies:
\begin{itemize}
\item $\bigcap_{n\in \mathbb{Z}}\mathcal{R}^{-n}(\Xi):=\Delta$ is hyperbolic set for $\mathcal{R}$.
\item $\displaystyle HD\left(\Delta\right)\sim 2$.
\end{itemize}
\noindent Now let's construct a family of perturbations of $\phi$, such that the perturbation produced in $\mathcal{R}$ has the desired property $V$.
\begin{R}
Given a vector field $X$ close to $\phi$, then the flow of $X$ still defines a Poincar\'e
map $\mathcal{R}_{X}$ defined in the same cross-section where is defined $\mathcal{R}$.
\end{R}
\subsubsection{Family of Perturbations}\label{Realiz of the Pert}
The central goal in this section is to do a small conservative perturbation of $\phi$ with good properties.
\begin{Le}\label{Le 6.2}
Given $\mathcal{V}$ a $C^{r}$-neighborhood of $\phi$ and $p\in \Delta$, suppose that $p\in \Sigma$, with $\Sigma=\Sigma_i$ for some $i$ and $U$ a neighborhood of $\phi^{\frac{t_{+}(p)}{2}}(p)$, there exists conservative vector field $X\in \mathcal{V}$ such that:
\begin{enumerate}
\item[1.]$X\equiv\phi$ outside of $U$.
\item[2.] There is $\tau>0$ such that $X\equiv\phi$ outside of a subset of $U$ of the form
$X^{[0,\tau]}(\Sigma_0)=\{X^{t}(x):x\in \Sigma_0, \, 0<t<\tau\}$, where $\Sigma_0$ is a neighborhood of $\phi^{\frac{t_{+}(p)}{2}}(p)$ in $\phi^{\frac{t_{+}(p)}{2}}(\Sigma)$.
\item[3.] If $\mathcal{R}_{X}$ is the Poincar\'e associate to $X$ over $\Xi$, then $\mathcal{R}_{X}(p)\neq\mathcal{R}(p)$.
\end{enumerate}
\end{Le}
\noindent The proof of this lemma is a immediate consequence of Lemmas below. The first is about conservative trivialization and the second is about local conservative perturbations.
\begin{Le}{\cite[Lemma ${3.4}$ (Conservative flowbox theorem)]{Bessa1}}\label{Le 6.3}
Let $X\in \mathfrak{X}^{r}_{\omega}(M)$, $p$ be a regular point of the vector field and $\Sigma$ a transversal section to $X$ which contains $p$, then there exists a $C^{\infty}$-coordinate system $\alpha\colon U\subset M \to \mathbb{R}^3$ with $\alpha(p)=0$ and such that
\begin{enumerate}
\item[1.] $\alpha_{*}X=(0,0,1)$.
\item[2.]$\alpha_{*}\omega=dx\wedge dy \wedge dz.$
\item[3.]$\alpha(U\cap \Sigma)\subset\{z=0\}.$
\end{enumerate}
\end{Le}
\noindent Let $B_{\delta}(x,y)\subset \mathbb{R}^2$ be the open ball of center $(x,y)$ and ratio $\delta$. Similarly $\overline{B}_{\delta}(x,y)$ denotes the closed ball. If $C$ is the cylinder $\partial B_{\delta}(x,y)\times [0,h]\subset \mathbb{R}^3$ and $0<\beta<\delta$, we define of neighborhood of $C$ as
$$A_{\beta}(C)=\left( B_{\delta+\beta}(x,y)\setminus \overline{B}_{\delta-\beta}(x,y)\right) \times [0,h]\subset \mathbb{R}^3.$$
and call it \textit{cylinder ring} with center at $C$ and radius $\beta$.
\begin{Le}\cite[Lemma $4.1$]{CandO}\label{Le 6.4}
Let $X\colon \mathbb{R}^3\to \mathbb{R}^3$ be the constant vector field defined by $X(x,y,z)=(0,0,1)$. Consider the cylinder $C=\partial B_{\delta}(0,0)\times [0,h]\subset \mathbb{R}^3$, $\delta>0$, $h>0$, and points $p\in \partial B_{\delta}(0,0)\times \{0\}$ and $q\in \partial B_{\delta}(0,0)\times \{h\}$. Let $\theta$ be the angle between the vector $p-(0,0,0)$ and $q-(0,0,h)$. Given $0<\beta<\delta$ there exists a $C^{\infty}$ vector field $Z$ on $\mathbb{R}^3$ with the following properties
\begin{enumerate}
\item[1.] $Z$ preserves the canonical volume form $dx\wedge dy \wedge dz$;
\item[2.] $Z\equiv X$ outside the cylinder ring $A_{\beta}(C)$;
\item[3.]The positive orbit of $p$, with respect to $Z$, contains $q$;
\item[4.] Given $r\in \mathbb{N}$ and $\epsilon>0$, if $|\theta|$ is small enough, then ${\Vert Z-X\Vert}_{r}< \epsilon$, where $\Vert\cdot \Vert$ denotes the $C^r$ norm on the set of $C^r$ vector fields.
\begin{figure}[hbtp]
\centering
\includegraphics[scale=0.5]{Figure2.png}
\caption{The motion of the orbit }
\end{figure}
\end{enumerate}
\end{Le}
\noindent With the help of these last lemmas we can make the proof of Lemma \ref{Le 6.3}.
\begin{proof}[Proof of Lemma \ref{Le 6.2}]
By Lemma \ref{Le 6.3} we can consider a coordinates systems $\alpha\colon U'\to V$ in a neighborhood $U'\subset U$ of $\phi^{\frac{t_{+}(p)}{2}}(p)$, with $\alpha(p)=(0,0,0)$, $\alpha_{*}\phi=(0,0,1)$, $\alpha_{*}\omega=dx\wedge dy\wedge dz$ and $\alpha(\phi^{\frac{t_{+}(p)}{2}}(\Sigma)\cap U')\subset \{z=0\}$. Let $\beta , \delta >0$, $0<h<1$ and $q\in\{z=0\}$ such that
the solid cylinder $\overline{B}_{\beta+\delta}(q)\times [0,h]\subset V$ and $(0,0)\in \partial B_{\delta}(q)$.\\
Consider now the cylinder ring $A_{\beta}(C)$ defined by $\beta , \delta , h$ and $q$. Let $\theta$ be a small angle, and let $q'\in \partial B_{\delta}(q)\times \{h\}$ such that the angle between $(0,0,0)-q$ and $q'-(q,h)$ is equal to $\theta$. Now we may apply the perturbation Lemma \ref{Le 6.4} at the
cylinder $C=\partial B_{\delta}(q)\times[0, h]$ to join $0$ to $q'$ and obtain a vector field $Z$ on $V$ such that:
\begin{enumerate}
\item[a.] $Z$ preserves the canonical volume form;
\item[b.] $Z\equiv (0,0,1)$ in $V\setminus A_{\beta}(C)$;
\item[c.] The positive orbit of $0$, with respect to $Z$, contains $q'$.
\end{enumerate}
Let us define the vector field $X$ in $M$ in the following way: $X\equiv \phi$ outside of $U'$ and
$X=\alpha_{*}(Z)$ in $U'$. Note that $X$ is $C^{r}$, satisfies item 1, and taking $\theta$
sufficiently small, we may assume $X\in \mathcal{V}$.\\
In order to prove item 2, we consider $\Pi_0\subset V\cap\{z=0\}$ a compact neighborhood of
the origin contained in $\alpha(\phi^{\frac{t_{+}(p)}{2}}(\Sigma)\cap U'))$. Then just take $\Sigma_0=\alpha^{-1}(\Pi_{0})$ and $\tau=\sup \{t>0: \alpha(X^{t}(x))\in \Pi_{0}\times [0,h], x\in \Sigma_0\}$. The item 3 is an immediate consequence of item b) and c) above.
\end{proof}
\noindent The proof of Lemma \ref{Le 6.2}, implies that para $p\in \Delta$ and for every small $\theta$ there is $X_{\theta}^{p}\in \mathfrak{X}^{r}_{\omega}(M)$ such that the poincar\'e map ${\mathcal{R}}_{X_{\theta}^{p}}$ associated to $X_{\theta}^{p}$ satisfies $\mathcal{R}_{X_{\theta}^{p}}(p)\neq \mathcal{R}(p)$.\\
In particular, if $p\in \Sigma\cap \Delta$, then
\begin{equation}\label{E1-Sec6}
{\mathcal{R}}_{X_{\theta}^{p}}({W^{s}}_{loc,\mathcal{R}}(q,\Sigma))\cap \mathcal{R}({W^{s}}_{loc,\mathcal{R}}(q,\Sigma))=\emptyset,
\end{equation}
\noindent for $q\in \Sigma\cap \Delta$ close to $p$, where $W^{s}_{loc,\mathcal{R}}(q,\Sigma)$ is the local stable manifold associated to $\mathcal{R}$.\\
Note that if $\theta=0$, then $X_{\theta}=\phi$.
\noindent As $\Delta$ is a compact hyperbolic set, then there are a finite number of point in $\Delta$, let's say $p_1,\dots, p_n$ and $U_i$ neighborhood of $\phi^{\frac{t_{+}(p_i)}{2}}$ as Lemma \ref{Le 6.2} pairwise disjoint respectively, such that the projection over $\Xi$ along the flow $\phi^{t}$ of $\displaystyle\bigcup_{i}^{n}U_i$ contains a small Markov partition of $\Delta$.\\
So, we can define the $C^{r}$ vector field $X_{\theta}\in \mathfrak{X}^{r}_{\omega}(M)$ by
\[ X_{\theta} = \left\{\begin{array}{lll}
\ \ X_{\theta}^{p_i} & \mbox{$\text{if} \ \ x\in U_i$};\\
\ \ & \mbox{ \ } \\
\ \ \phi & \mbox{$ \text{otherwise} $}\end{array} \right..\]
\noindent As $\theta$ is small, then of flow of $X_{\theta}$ is still a conservative Anosov.\\ Consider now the map $\Phi^{\theta}(x):=\mathcal{R}^{-1}\circ \mathcal{R}_{X_{\theta}}(x)$ defined in a small Markov partition of $\Delta$. Then by equation (\ref{E1-Sec6}), the map $\mathcal{R}_{\theta}:=\mathcal{R}\circ \Phi^{\theta}$ satisfies
\begin{equation}\label{E2-Sec6}
{\mathcal{R}}_{\theta}({W^{s}}_{loc,\mathcal{R}}(q,\Sigma))\cap \mathcal{R}({W^{s}}_{loc,\mathcal{R}}(q,\Sigma))=\emptyset.
\end{equation}
\begin{R}\label{R1-Sec6}
Then by equation $(\ref{E2-Sec6})$ the perturbation of $\mathcal{R}$ given by ${\mathcal{R}}^{\theta}$ satisfies the condition on the family of perturbation to get the property $V$ (cf. section \ref{SIRCS} and \cite[subsection 4.2]{RM2}).
\end{R}
\subsection{Regaining the Spectrum}\label{RS}
\noindent Given $F\in C^{0}(M,\mathbb{R})$, we can define the function $maxF_{\phi}\colon \Xi\to\mathbb{R}$ by
$$max F_{\phi}(x):=\max_{t_{-}(x)\leq t \leq t_{+}(x)}F(\phi^{t}(x))$$
\noindent where $t_{-}(x),t_{+}(x)$ are such that $\mathcal{R}^{-1}(x)=\phi^{t_{-}(x)}(x)$ and $\mathcal{R}(x)=\phi^{t_{+}(x)}(x)$.\\
\noindent Note that this definition depends on the flow $\phi^{t}$, or equivalently the vector field $\phi$. Note also that $max F_{\phi}$ is always a continuous function, but even if $F$ is $C^{\infty}$, $max F_{\phi}$ can be only a continuous function. In what follows we give some ``differentiability" to $max F_{\phi}$ at least for $F\in C^{2}(M,\mathbb{R})$. Thus, similar to Lemma \ref{Lmax0}, we have
\begin{Le}\label{Lmax}
There exists a dense $\mathcal{B}_{\phi}\subset C^{\infty}(M,\mathbb{R})$ and $C^{2}$-open such that given $\beta>0$, then for any $F\in \mathcal{B}_{\phi}$ there are sub-horseshoe $\Delta_{F}^{s,u}$ of $\Delta$ with $HD(K_{F}^{s})\geq HD(K^{s})-\beta$, $HD(K_{F}^{u})\geq HD(K^{u})-\beta$ and a Markov partition $R_{F}^{s,u}$ of $\Delta_{F}^{s,u}$, respectively, such that the function $max F_{\phi}|_{\Xi\cap R_{F}^{s,u}}\in C^{1}(\Xi\cap R_{F}^{s,u},\mathbb{R})$, where $K_{F}^{s}$, $K^{s}$ are regular Cantor sets that describe the geometry transverse of the unstable of $W^{u}(\Delta_{F}^{s}), \, W^{u}(\Delta)$ respectively, and $K_{F}^{u}$, $K^{u}$ are regular Cantor sets that describe the geometry transverse of the stable foliations of $W^{s}(\Delta_{F}^{u}), \, W^{s}(\Delta)$ respectively (cf. subsection \ref{sec EMAH}).
\end{Le}
\noindent The proof of Lemma \ref{Lmax} can be found in \cite[Lemma 18]{RM2}.
\begin{R}\label{R4}
\noindent Let $x\in int(\Sigma)$ with $\Sigma\in \Xi$ such that $\mathcal{R}(x)=\phi^{t_{+}(x)}(x)\in int(\Xi)$, by The Large Tubular Flow Theorem, there exists a neihgborhood $U_{x}\subset \Sigma$ of $x$, a diffeomorphism $\varphi\colon U_{x}\times (-\epsilon,t_{+}(x)+\epsilon)\to \varphi(U_{x}\times (-\epsilon,t_{+}(x)+\epsilon))\subset M$ such that $D\varphi_{(z,t)}(0,0,1)=\phi(\varphi(z,t))$ for $(z,t)\in U_{x}\times (-\epsilon, t_{+}(x)+\epsilon)$. Moreover, as the elements of the Markov partition are disjoint, has small diameter and $\Delta$ is compact, then, we can suppose that there is a finite number an open set $U_{x_i}$ such that $U_{x_i}\cap U_{x_j}=\emptyset$ and $\Delta\subset \bigcup U_{x_i}$ for some $x_i\in \Delta$. Denote $\varphi_{i}\colon U_{x_{i}}\times (-\epsilon,t_{+}(x_{i})+\epsilon)\to \varphi_{i}\left(U_{x_i}\times (-\epsilon,t_{+}(x_{i})+\epsilon)\right)\subset M$ such that $(D\varphi_{i})_{(z,t)}(0,0,1)=\phi(\varphi_{i}(z,t))$.
\end{R}
\noindent Keeping the notation of the previous Lemma we have:
\begin{C}\label{Cmax}
The above property is robust in the following sense: If $X$ is a vector field $C^{1}$-close to $\phi$, then $\mathcal{B}_{\phi}=\mathcal{B}_{X}$ and for any $F\in \mathcal{B}_{X}$, holds that $maxF_{X}\in C^{1}(\Xi\cap R_{F}^{s,u},\mathbb{R})$.
\end{C}
\noindent In the proof of Theorem B, we will use the following proposition found in (\cite[pg. 21]{MY1}).
\begin{Pro}\label{propMY1pag21} Let $\Lambda$ be a horseshoe and let $L\subset \Lambda$ an invariant proper subset of $\Lambda$. Then, for all $\epsilon>0$, there is a sub-horseshoe $\widetilde{\Lambda}\subset \Lambda$ such that $\widetilde{\Lambda}\cap L=\emptyset$ and $$HD(\widetilde{K})\geq HD(K)-\epsilon,$$
\noindent where, $K$, $\widetilde{K}$ are of regular cantor set that describe the geometry transverse of the stable foliation $W^{s}(\Lambda), \ W^{s}(\widetilde{\Lambda})$, respectively (cf. subsection \ref{sec EMAH}).\\
\end{Pro}
\begin{proof}[\bf Proof of Theorem B]
Let $F\in \mathcal{B}_{\phi}$, $\Delta_{F}^{s,u}$ and $R_{F}^{s,u}$ as in Lemma \ref{Lmax} with $$HD(K_{F}^{s,u})\geq HD(K^{s,u})-\beta.$$
Put $L=\Delta_{F}^{s}\cap \Delta_{F}^{u}\subset \Delta_{F}^{s}$ a $\mathcal{R}$-invariant set, then by Proposition \ref{propMY1pag21} (applied to $\mathcal{R}^{-1}$), there is a sub-horseshoe $_{1}{\Delta}_{F}^s$ of $\Delta_{F}^{s}$ such that $_{1}{\Delta}_{F}^s\cap L=\emptyset$, thus $_{1}{\Delta}_{F}^s\cap\Delta_{F}^{u}=\emptyset$. Moreover,
$$HD(_{1}{K}_{F}^s)\geq HD(K_{F}^s)-\beta \geq HD(K^{s})-2\beta,$$
\noindent where $_{1}{K}_{F}^{s}$ is of regular cantor set that describe the geometry transverse of the unstable foliation $W^{u}(_{1}{\Delta}^{s}_{F})$.
Define the sub-horseshoe $\Delta_{F}$ of $\Delta$ by $\Delta_{F}:= \ _{1}{\Delta}_{F}^{s}\cup\Delta^{u}_{F}$. Since $\ _{1}{\Delta}_{F}^{s}$ and $\Delta^{u}_{F}$ are disjoint, we used the notation $\Delta_{F}:=(_{1}{\Delta}^{s}_{F},\Delta^{u}_{F})$, put $R_{F}:=R_{F}^{s} \cup R_{F}^{u}$ and consider the open and dense set, as in \cite[Theorem 1]{RM}
$${H}_{1}(\mathcal{R},\Delta_{F})=\left\{f\in C^{1}(\Xi\cap R_F,\mathbb{R}):\#M_{f}(\Delta_{F})=1 \ \ for \ \ z\in M_{f}(\Delta_{F}), \ D\mathcal{R}_{z}(e_{z}^{s,u})\neq 0\right\},$$
where $e_{z}^{s,u}$ are unit vectors in $E^{s,u}_{\Xi}(z)$ respectively.\\
Let $f\in {H}_{1}(\mathcal{R},\Delta_{F})$, then there is a unique $z_{f}\in M_{f}(\Delta_{F})$.
Since $_{1}{\Delta}^{s}_{F}\cap\Delta^{u}_{F}=\emptyset$, we can suppose that $z_f\in$ $_{1}{\Delta}^{s}_{F}$. Thus as in \cite[section 4]{RM}, let $_{1}\widetilde{\Delta}_{F}^{s}$ sub-horseshoe of $_{1}{\Delta}^{s}_{F}$ such that $HD(_{1}\widetilde{\Delta}^{s}_{F})$ is close to $ HD(_{1}\Delta^{s}_{F})$ and $z_f\notin$ $_{1}\widetilde{\Delta}^{s}_{F}$, then put
$$\widetilde{\Delta}_{F}=(_{1}\widetilde{\Delta}^{s}_{F},{\Delta}^{u}_{F}).$$
\noindent Moreover, since $HD(_{1}\widetilde{K}_{F}^{s})$ is close to $ HD(_{1}{K}_{F}^{s})$, $HD(K^{s})+HD(K^{u})=HD(\Delta)\sim 2$ and $\beta$ is small, then $$HD(_{1}\widetilde{K}_{F}^{s})+HD(K^{u}_{F})>1,$$
\noindent where $_{1}\widetilde{K}_{F}^{s}$ is the regular Cantor set that describe the geometry transverse of the unstable foliation $W^{u}(_{1}\widetilde{\Delta}^{s}_{F})$ (cf. subsection \ref{sec EMAH}).\\
\noindent Hence, by [MY1] it is sufficient perturb $_{1}\widetilde {\Delta}^{s}_{F}$ as in \cite[subsection 4.2]{RM2} to obtain property $V$ (cf. subsection \ref{SIRCS}). Then by Remark \ref{R1-Sec6}, there is a small $\theta$, such that the pair $(\mathcal{R}^{\theta},\widetilde{\Delta}^{\theta}_{F})$ has the property $V$ (see definition \ref{Property V})
where $\widetilde{\Delta}^{\theta}_{F}=(_{1}\widetilde{\Delta}_{\theta},{\Delta}^{u}_{F})$ and $_{1}\widetilde{\Delta}_{\theta}$ is the continuation of the hyperbolic set $_{1}\widetilde{\Delta}^{s}_{F}$ for $\mathcal{R}^{\theta}$, so by \cite[Main Theorem]{RM} we have that
$$int \,M(f,\widetilde{\Delta}^{\theta}_{F})\neq \emptyset \ \ \text{and} \ \ int\, L(f,\widetilde{\Delta}^{\theta}_{F})\neq \emptyset,$$
\noindent for any $f\in{H}_{1}(\mathcal{R},\Delta_F)$. Now by Corollary \ref{Cmax} the function $max F_{X_{\theta}}|_{\Xi\cap (R_{F}^{s}\cup R_{F}^{u})}$ is $C^{1}$, using local coordinates as in Remark \ref{R4} respect to the field $X_{\theta}$, we can find $g\in C^2(\Xi,\mathbb{R})$ such that
\begin{equation}\label{E1PT1}
max F_{{X_{\theta}}}|_{\Xi\cap (R_{F}^{s}\cup R_{F}^{u})}(x_1,x_2,x_3)+g(x_1,x_2)\in H_{1}(\mathcal{R},\Delta_{F}).
\end{equation}
Put $h(x_1,x_2,x_3)=F(x_1,x_2,x_3)+g(x_1,x_2)$, then $max h_{X_{\theta}}=max F_{{X_{\theta}}}+g\in {H}_{1}(\mathcal{R},\Delta_F)$.
Therefore, since $M(max h_{X_{\theta}},{\widetilde{\Delta}}^{\theta}_{F})=\left\{\sup_{n\in\mathbb{Z}}h({\mathcal{R}}^{n}_{{\theta}}(x)):x\in\widetilde{\Delta}^{\theta}_{F}\right\}\subset M(h,X_{\theta})$.
So, $$int M(h,{X_{\theta}})\neq\emptyset.$$
Analogously, $L(max h_{X_{\theta}},{\widetilde{\Delta}}^{\theta}_{F})\subset L(h,{X_{\theta}})$, therefore $int\,L(h,{X_{\theta}})\neq\emptyset$.
\end{proof}
\section{Appendix}\label{Ap}
\subsection{Regular Cantor Sets}\label{RCS}
Let $\mathbb{A}$ be a finite alphabet, $\mathbb{B}$ a subset of $\mathbb{A}^{2}$, and $\Sigma_{\mathbb{B}}$ the subshift of finite type of $\mathbb{A}^{\mathbb{Z}}$ with allowed transitions $\mathbb{B}$. We will always assume that $\Sigma_{\mathbb{B}}$ is topologically mixing, and that every letter in $\mathbb{A}$ occurs in $\Sigma_{\mathbb{B}}$.
\indent An {\it expansive map of type\/} $\Sigma_{\mathbb{B}}$ is a map $g$ with the following properties:
\begin{itemize}
\item[(i)] the domain of $g$ is a disjoint union
$\displaystyle\bigcup_{\mathbb{B}}I(a,b)$. Where for each $(a,b)$,\,\, $I(a,b)$ is a compact subinterval of $I(a) := [0,1]\times\{a\}$;
\item[(ii)] for each $(a,b) \in \mathbb{B}$, the restriction of $g$ to $I(a,b)$ is a smooth diffeomorphism onto $I(b)$
satisfying $|Dg(t)| > 1$ for all $t$.
\end{itemize}
\noindent The {\it regular Cantor set\/} associated to $g$ is the maximal invariant set
$$K = \bigcap_{n\ge0} g^{-n}\bigg(\bigcup_{\mathbb{B}} I(a,b)\bigg).$$
\noindent Let $\Sigma^+_{\mathbb{B}}$ be the unilateral subshift associated to $\Sigma_{\mathbb{B}}$. There exists a unique homeomorphism $h\colon \Sigma^{+}_{\mathbb{B}} \to K$ such that
$$
h(\underline{a}) \in I(a_0), \text{ for } \underline{a} = (a_0,a_1,\dots) \in \Sigma^+_{\mathbb{B}}
\ \ and \ \
h\circ\sigma =g \circ h,$$
where $\sigma^{+}\colon \Sigma_{\mathbb{B}}^{+} \to \Sigma_{\mathbb{B}}^{+}$, is defined as follows $\sigma^{+}((a_{n})_{n\geq 0})=(a_{n+1})_{n\geq0}$.
\subsection {Expanding Maps Associated to a Horseshoe}\label{sec EMAH}
\noindent Let $\Lambda$ be a horseshoe associate a $C^{2}$-diffeomorphism $\varphi$ on the a surface $M$ and consider a finite collection $(R_{a})_{a\in\mathbb{A}}$ of disjoint rectangles of $M$, which are a Markov partition of $\Lambda$. Put the sets
$$W^{s}(\Lambda,R)=\bigcap_{n\geq0}\varphi^{-n}(\bigcup_{a\in \mathbb{A}}R_{a}),$$
$$W^{u}(\Lambda,R)=\bigcap_{n\leq0}\varphi^{-n}(\bigcup_{a\in \mathbb{A}}R_{a}).$$
There is a $r>1$ and a collection of $C^{r}$-submersions $(\pi_{a}:R_{a}\rightarrow I(a))_{a\in\mathbb{A}}$, satisfying the following property:\\
\ \\
If $z,z^{\prime}\in R_{a_{0}}\cap \varphi^{-1}(R_{a_{1}})$ and $\pi_{a_{0}}(z)=\pi_{a_{0}}(z^{\prime})$, then we have $$\pi_{a_{1}}(\varphi(z))=\pi_{a_{1}}(\varphi(z^{\prime})).$$
\noindent In particular, the connected components of $W^{s}(\Lambda,R)\cap R_{a}$ are the level lines of $\pi_{a}$. Then we define a mapping $g^{u}$ of class $C^{r}$ (expansive of type $\Sigma_{\mathbb{B}}$) by the formula
$$g^{u}(\pi_{a_{0}}(z))=\pi_{a_{1}}(\varphi(z))$$
\noindent for $(a_{0},a_{1})\in \mathbb{B}$, $z\in R_{a_{0}}\cap\varphi^{-1}(R_{a_{1}})$.
The regular Cantor set $K^{u}$ defined by $g^{u}$, describes the geometry transverse of the stable foliation $W^{s}(\Lambda,R)$.
Analogously, we can describe the geometry transverse of the unstable foliation $W^{u}(\Lambda,R)$, using a regular Cantor set $K^{s}$ define by a mapping $g^{s}$ of class $C^{r}$ (expansive of type $\Sigma_{\mathbb{B}}$).\\
\noindent Also, the horseshoe $\Lambda$ is locally the product of two regular Cantor sets $K^{s}$ and $K^{u}$. So, the Hausdorff dimension of $\Lambda$, $HD(\Lambda)$ is equal to
$HD(K^{s}\times K^{u})$, but for regular Cantor sets, we have that $HD(K^{s}\times K^{u})=HD(K^{s})+HD(K^{u})$. Thus $HD(\Lambda)=HD(K^{s})+HD(K^{u})$ (cf. \cite[chap 4]{PT}).
\subsection{Intersections of Regular Cantor Sets and Property $V$}\label{SIRCS}
\noindent Let $r$ be a real number $> 1$, or $r=+\infty$. The space of $C^r$ expansive maps of type $\Sigma$ (cf. section \ref{RCS}), endowed with the $C^r$ topology, will be denoted by $\Omega_\Sigma^r$\,. The union $\Omega_\Sigma = \displaystyle\bigcup_{r>1} \Omega_\Sigma^r$ is endowed with the inductive limit topology.\\
\noindent Let $\Sigma^- = \{(\theta_n)_{n\leq 0}\,, (\theta_i,\theta_{i+1})
\in \mathbb{B} \text{ for } i < 0\}$. We equip $\Sigma^-$ with the following
ultrametric distance: for $\underline{\theta} \ne \underline{\widetilde\theta} \in \Sigma^-$, set
\[ d(\underline{\theta},\underline{\widetilde\theta}) = \left\{ \begin{array}{lll}
\ \ \ 1 & \mbox{if \ $\theta_0 \ne \widetilde{\theta}_0$};\\
& \\
|I(\underline{\theta} \wedge \underline{\widetilde\theta})|& \mbox{otherwise}\end{array} \right., \]
\noindent where $\underline{\theta} \wedge \underline{\widetilde\theta} = (\theta_{-n},\dots,\theta_0)$ if
$\widetilde\theta_{-j} = \theta_{-j}$ for $0 \le j \le n$ and $\widetilde\theta_{-n-1}
\ne \theta_{-n-1}$\,.
\\
\indent Now, let $\underline{\theta} \in \Sigma^-$; for $n > 0$, let $\underline{\theta}^n = (\theta_{-n},\dots,\theta_0)$, and let $B(\underline{\theta}^n)$ be the affine map from
$I(\underline{\theta}^n)$ onto $I(\theta_0)$ such that the diffeomorphism $k_n^{\underline{\theta}}
= B(\underline{\theta}^n) \circ f_{\underline{\theta}^n}$ is orientation preserving.\\
\noindent We have the following well-known result (cf. \cite{Su}):
\\
\noindent{\bf Proposition}. \textit{Let} $r \in (1,+\infty)$, $g \in \Omega_\Sigma^r$.
\begin{enumerate}
\it{
\item For any $\underline{\theta} \in \Sigma^-$, there is a diffeomorphism $k^{\underline{\theta}} \in \text{Diff}_+^{\ r}(I(\theta_0))$ such that $k_n^{\underline{\theta}}$ converge to $k^{\underline{\theta}}$ in $\text{Diff}_{+}^{\ r'}(I(\theta_0))$, for any $r'< r$, uniformly in $\underline{\theta}$. The convergence is also uniform in a neighborhood of $g$ in $\Omega_\Sigma^r$\,.
\item If $r$ is an integer, or $r = +\infty$,\,\,\,$k_n^{\underline{\theta}}$ converge to $k^{\underline{\theta}}$ in $\text{Diff}_+^r(I(\theta_0))$. More precisely, for every $0 \leq j \leq r-1$, there is a constant $C_j$ (independent on $\underline{\theta}$) such that
$$
\left| D^j \, \log \, D \left[k_n^{\underline{\theta}} \circ (k^{\underline{\theta}})^{-1}\right](x)\right| \leq C_j|I(\underline{\theta}^n)|.
$$
It follows that $\underline{\theta} \to k^{\underline{\theta}}$ is Lipschitz in the following sense: for $\theta_0 = \widetilde\theta_0$, we have
$$
\left|D^j \, \log \, D\big[k^{\underline{\widetilde\theta}} \circ (k^{\underline{\theta}})^{-1}\big](x)\right| \leq C_j\,d(\underline{\theta}, \underline{\widetilde\theta}).
$$
}
\end{enumerate}
\ \\
\noindent Let $r \in (1,+\infty]$. For $a \in \mathbb{A}$, denote by ${\cal{P}}^{r}(a)$ the space of $C^r$-embeddings of $I(a)$ into $\mathbb{R}$, endowed with the $C^r$ topology. The affine group $Aff(\mathbb{R})$ acts by composition on the left on ${\cal{P}}^r(a)$, the quotient space being denoted by $\overline{\cal{P}}^r(a)$. We also consider ${\cal{P}}(a) = \displaystyle\bigcup_{r>1} {\cal{P}}^r(a)$ and $\overline{\cal{P}}(a) = \displaystyle\bigcup_{r>1} \overline{\cal{P}}^r(a)$, endowed with the inductive limit topologies.
\begin{R}\label{r=1}
In \cite{MY} is considered ${\cal{P}}^{r}(a)$ for $r \in (1,+\infty]$, but all the definitions and results involving ${\cal{P}}^{r}(a)$ can be obtained considering $r\in [1,+\infty]$.
\end{R}
\noindent Let $\mathcal{A} =(\underline{\theta}, A)$, where $\underline{\theta} \in \Sigma^-$ and $A$ is now an {\it affine\/} embedding of $I(\theta_0)$ into $\mathbb{R}$. We have a canonical map
\begin{eqnarray*}
\cal{A} & \to & {\cal{P}}^r = \bigcup_{\mathbb{A}} {\cal{P}}^r(a)\\
(\underline{\theta},A) &\mapsto & A\circ k^{\underline{\theta}} \ \ (\in {\cal{P}}^r(\theta_0)).
\end{eqnarray*}
\noindent Now assume we are given two sets of data $(\mathbb{A},\mathbb{B},\Sigma,g)$, $({\mathbb{A}}',{\mathbb{B}}',\Sigma',g')$ defining regular Cantor sets $K$, $K'$.\\
We define as in the previous the spaces $\mathcal{P} = \displaystyle\bigcup_{\mathbb{A}}{\mathcal{P}}(a)$ and ${\cal{P}}'= \displaystyle\bigcup_{{\mathbb{A}}'} {\cal{P}}(a')$.
\noindent A pair $(h,h')$,
$(h \in {\cal{P}}(a), h'\in {\cal{P}} '(a'))$ is called a {\it smooth configuration\/} for $K(a)=K\cap I(a)$, $K'(a')=K'\cap I(a')$. Actually, rather than working in the product $\cal{P} \times {\cal{P}}'$, it is better to go to the quotient $Q$ by the diagonal action of the affine group $Aff(\mathbb{R})$. Elements of $Q$ are called {\it smooth relative configurations\/} for $K(a)$, $K'(a')$.
\noindent We say that a smooth configuration $(h,h') \in {\cal{P}}(a)\times {\cal{P}}(a')$ is
\begin{itemize}
\item {\it linked\/} if $h(I(a)) \cap h'(I(a')) \ne \emptyset$;
\item {\it intersecting\/} if $h(K(\underline{a})) \cap h'(K(\underline{a}')) \ne \emptyset$, where $K(\underline{a})=K\cap I(\underline{a})$ and $K(\underline{a}')=K\cap I(\underline{a}')$;
\item {\it stably intersecting\/} if it is still intersecting when we perturb it in $\cal{P}\times\cal{P}'$, and we perturb $(g,g')$ in $\Omega_\Sigma \times \Omega_{\Sigma'}$\,.
\end{itemize}
\noindent All these definitions are invariant under the action of the affine group, and therefore make sense for smooth relative configurations.\\
\noindent As in previous, we can introduce the spaces $\cal{A}$, ${\cal{A}}'$ associated to the limit geometries of $g$,\,\,$g'$, respectively. We denote by $\cal{C}$ the quotient of $\cal{A}\times{\cal{A}}'$ by the diagonal action on the left of the affine group. An element of $\cal{C}$, represented by $(\underline\theta,A) \in \cal{A}$,\,\, $(\underline{\theta}', A') \in {\cal{A}}'$, is called a relative configuration of the limit geometries determined by $\underline{\theta}$, $\underline{\theta}'$. We have canonical maps
\begin{eqnarray*}
\cal{A}\times{\cal{A}}'&\to & \cal{P}\times{\cal{P}}'\\
\cal{C} &\to & Q
\end{eqnarray*}
which allow to define linked, intersecting, and stably intersecting configurations at the level of $\cal{A}\times{\cal{A}}'$ or $\cal{C}$.\\
\noindent{\bf Remark}: For a configuration $((\underline{\theta}, A), (\underline{\theta}',A'))$ of limit geometries, one could also consider the {\it weaker\/} notion of stable intersection, obtained by considering perturbations of $g$, $g'$ in $\Omega_\Sigma \times \Omega_{\Sigma'}$ and perturbations of $(\underline{\theta},A)$, $(\underline{\theta}',A')$ in $\cal{A}\times{\cal{A}}'$. We do not know of any example of expansive maps $g$, $g'$, and configurations $(\underline{\theta},A)$, $(\underline{\theta}',A')$ which are stably intersecting in the weaker sense but not in the stronger sense.\\
\noindent We consider the following subset $V$ of $\Omega_\Sigma \times \Omega_{\Sigma'}$\,. A pair $(g, g')$ belongs to $V$ if for any $[(\underline{\theta},A), (\underline{\theta}',A')] \in \cal{A} \times {\cal{A}}'$ there is a translation $R_t$ (in $\mathbb{R}$) such that $(R_t\circ A \circ k^{\underline{\theta}}, A'\circ k^{\prime\underline{\theta}'})$ is a stably intersecting configuration.\\
\begin{Defi}\label{Property V}
We say that a pair $(\psi, \Lambda)$, where $\Lambda$ is a horseshoe for $\psi$, has the property $V$ if the stable and unstable cantor sets has the property $V$ in the above sense.
\end{Defi}
\noindent \textbf{Sergio Augusto Roma\~na Ibarra}\\
Universidade Federal do Rio de Janeiro\\
Av. Athos da Silveira Ramos 149, Centro de Tecnologia \ - Bloco C \ - Cidade Universit\'aria \ - Ilha do Fund\~ao, cep 21941-909 \\
Rio de Janeiro-Brasil\\
E-mail: <EMAIL> \\
\ \\
\bibliographystyle{alpha}
|
\section{Introduction}
Photons propagating in a polarizable medium can resonantly interact with an optically active resonance and a part of the photonic energy thus resides in the matter degrees of freedom, resulting in a dispersive reduction of the group velocity.
Quantum descriptions of such systems can be achieved through macroscopic approaches where one starts from the Maxwell equations treating the matter by a dispersive dielectric function $\epsilon\left(\omega\right)$. A modified orthonormality between the fields is derived to achieve the expected commutation relations with the final results dependant on the group velocity in the dielectric \cite{Santos95, Milonni95}. Such approaches can describe dissipative systems by addition into the Maxwell equations of noise currents preserving the commutation relations \cite{Knoll92}, allowing to describe the system dynamics by recourse to the Green's function of the classical scattering problem \cite{Gruner96, Matloob95,Matloob96}. The consistency of the approach can be shown in the homogeneous \cite{Matloob99} and general inhomogeneous case \cite{Scheel98}. Quantization of the electromagnetic field in inhomogeneous systems is particularly relevant to the study of quantum effects in surface plasmon polaritons \cite{Altewischer02,Fasel06,Chang06,Kolesov09,Tame13,DiMartino14} and other analogous excitations living at the interface between materials with different optical properties, like surface phonon polaritons \cite{Borstel74} surface excitons \cite{Lagois78}, and Tamm states \cite{Kavokin05}. In the study of surface plasmon polaritons the quantum properties of the matter degrees of freedom are usually of little interest, and macroscopic approaches that do not explicitly account for them are very successful. Elson and Ritchie where the first to quantize the electromagnetic field at the surface of a metal described by a Drude model in the absence of dissipation \cite{Elson71}. This work was then extended to more general dielectric constants \cite{Archambault10,Ballester10}.
A different, microscopic approach to the interaction of light with matter in solid-state physics was initially pioneered by Hopfield, that considered the matter degrees of freedom as bosonic fields coupled to the photons \cite{Hopfield58}. The normal modes of the coupled light-matter system, usually referred to as polaritons, are then found as linear superpositions of the creation and annihilation operators of the original light and matter fields. In this way it is possible to microscopically derive results postulated in macroscopic approaches \cite{Huttner91,Drummond99}. Losses can be taken into consideration by coupling the system to a broadband reservoir \cite{Huttner92}. Furthermore the Hopfield approach can be made fully covariant \cite{Belgiorno15} and equivalence can also be shown with the macroscopic Green's function approaches under the appropriate choice of dielectric function \cite{Suttorp04}.
The Hopfield method though, has the advantage of treating on equal footing light and matter, thus becoming the tool of choice in the domain of quantum polaritonics, where nonlinear processes depending upon the matter component of the excitations are of paramount importance \cite{Carusotto13,DeLiberato09,DeLiberato13}. Once the polaritonic operators have been obtained as linear superpositions of light and matter fields the Hopfield transformation can be inverted, allowing to express the fields describing the microscopic degrees of freedom as linear superpositions of polaritonic operators. Arbitrary nonlinearities, usually stemming from terms nonlinear in the matter field can then be naturally expressed as scattering terms between the polaritonic normal modes, allowing to investigate coherent and non-coherent polaritonic scattering processes.
Although plasmons were initially quantized by Bohm and Pines as microscopic bosonic degrees of freedom \cite{Bohm53}, Hopfield approaches to the study of surface plasmon polaritons have appeared only recently \cite{Todorov14}, using ad-hoc methods that make it difficult to invert the Hopfield transformation.
To the best of our knowledge, no general theory of Hopfield diagonalization in inhomogeneous materials has been formulated, and this becomes a pressing issue as inhomogeneous polaritonic systems characterised by extremely localised resonances \cite{Caldwell13,Kim14}, and thus potentially by large nonlinear effects, become commonplace. The necessity of real-space approaches in polaritonics has also recently been highlighted in Ref. \onlinecite{Elistratov16}, proving that an inhomogeneous Hopfield theory is necessary to study sub-healing length details in polaritonic condensates.
In this work we introduce a real-space Hopfield approach to the study of non-magnetic polarizable materials, which allows to determine invertible Hopfield transformations for generic geometries. For sake of simplicity we will limit ourselves to the isotropic case, the extension to the anisotropic one not presenting any conceptual difficulty.
As an alternative to the Green's function and path integral approaches previously employed we are able to use a straightforward modal analysis in the non-dissipative case to derive the polaritonic wavefunctions, and we can normalize the resulting modes imposing bosonic commutation relations, without need to recur to flux normalization \cite{Huttner91}.
This paper is organised as follows: in Sec. II we introduce the general real-space Hopfield approach in the non-dissipative case. While for surface plasmon polaritons a non-dissipative treatment is accurate only at the qualitative level,
this is not true for surface phonon polaritons, where quality factors in excess of $100$ for localised resonances are nowadays commonly achieved \cite{Caldwell13,Chen14,Gubbin16}. The increasing interest in those systems \cite{Greffet02,Hillenbrand02,Shen09,Holmstrom12,Li13,Ming13,Caldwell14,Feng15} and the need for a simple and powerful tool to study them has been one of the primary motivations for the present work. In Sec. III, in order to give a physical instantiation of the formal theory, we explicitly apply it to the analytically manageable case of the interface between vacuum and a polar dielectric, sustaining both propagative and surface excitations \cite{Xu90}.
In Sec. IV we extend the model to include losses, showing how we recover an equivalent formulation to the Green function approach of Ref. \onlinecite{Suttorp04} and the results of Ref. \onlinecite{Huttner92} in the homogeneous case.
\section{Real-space Hopfield theory for non-dissipative materials}
\subsection{Formulation of the Problem}
The coupled light-matter system can be described by the Power-Zienau-Wooley Lagrangian density \cite{Todorov14}
\begin{align}
\label{eq:MinCoupLag}
\mathcal{L}_0\left(\bf{r}\right)&=\int \mathrm{d}\mathbf{r}\,\,\biggr[\frac{\epsilon_0}{2}\mathrm{E}\left(\bf{r}\right)^2-\frac{1}{2\mu_0}\mathrm{B}\left(\bf{r}\right)^2 \\&\nonumber+\frac{\rho\left(\bf{r}\right)}{2}\dot{\mathrm{X}}\left(\bf{r}\right)^2 -\frac{\rho\left(\bf{r}\right) \omega_{\mathrm{T}}\left(\bf{r}\right)^2}{2}\mathrm{X}\left(\bf{r}\right)^2 \\ & -\kappa \left(\bf{r}\right) {\mathbf{X}}\left(\bf{r}\right)\cdot \dot{\mathbf{A}}\left(\bf{r}\right)-\mathrm{U}\left(\bf{r}\right)\nabla\left[\kappa \left(\bf{r}\right)\mathbf{X}\left(\bf{r}\right)\right]\biggr],\nonumber
\end{align}
where the electromagnetic field is described by the electric and magnetic components $\mathbf{E}\left(\mathbf{r}\right), \mathbf{B}\left(\mathbf{r}\right)$, related to the vector and scalar potentials $\mathbf{A}\left(\mathbf{r}\right)$ and $\mathrm{U}\left(\mathbf{r}\right)$ by
\begin{align}
\mathbf{E}\left(\mathbf{r}\right)&=-\nabla \mathrm{U}\left(\mathbf{r}\right) -\dot{\mathbf{A}}\left(\mathbf{r}\right),\\
\mathbf{B}\left(\mathbf{r}\right)&=\nabla \times \mathbf{A}\left(\mathbf{r}\right),
\end{align}
and the field $\bf{X}\left(\mathbf{r}\right)$ describes the degrees of freedom of the matter resonance with a space-dependent transverse frequency $\omega_{\mathrm{T}}\left(\bf{r}\right)$ and density $\rho\left(\bf{r}\right)$. The function $\kappa \left(\bf{r}\right)$ describes the spatially inhomogeneous light-matter coupling.
Matter is present only inside the regions where $\rho\left(\bf{r}\right)\neq0$ and thus, in order not to burden the notation, we will in the following assume that all the integrals involving matter degrees of freedom extend only inside those regions.
In the rest of this paper the spatial dependence of all variables will be suppressed where not necessary. Notice that the Lagrangian in \Eq{eq:MinCoupLag} not only models the coupling of light with microscopic harmonic degrees of freedom as excitons or phonons, but it has been shown to correctly describe plasmonic excitations in the limit of vanishing $\omega_{\mathrm{T}}$ \cite{Todorov14}.
The canonical momenta can now be calculated as
\begin{align}
\boldsymbol{\Pi}&=\frac{\delta {\mathcal{L}_0}}{\delta \dot{\mathbf{A}}}=\epsilon_0 \dot{\mathbf{A}}-\left[\kappa \mathbf{X}\right]^{\mathrm{T}},\\
\mathbf{P}&=\frac{\delta {\mathcal{L}_0}}{\delta \dot{\mathbf{X}}}=\rho\dot{\mathbf{X}},
\end{align}
where $\delta$ is a functional derivative and the superscript $\mathrm{T}$ refers to the transverse component of the field. Introducing the longitudinal frequency
\begin{align}
{\omega}_{\mathrm{L}}^2&=\omega_{\mathrm{T}}^2+\frac{\kappa ^2}{\epsilon_0 \rho},
\end{align}
we can obtain from the Lagrangian in \Eq{eq:MinCoupLag} the Hamiltonian\begin{align}
\label{eq:PZWH}
\mathcal{H}_0&=\int \mathrm{d}\mathbf{r} \left[\frac{\mathrm{D}^2}{2\epsilon_0}+\frac{\mu_0 \mathrm{H}^2}{2}+\frac{\mathrm{P}^2}{2{\rho}}
+\frac{{\rho}{\omega}_{\mathrm{L}}^2\mathrm{X}^2}{2}-\frac{\kappa}{\epsilon_0} \mathbf{X\cdot D}\right],
\end{align}
where we expressed the electromagnetic variables in terms of the electric displacement $\mathbf{D}=-\mathbf{\Pi}$ and magnetic field $\mathbf{H}=\mathbf{B}/{\mu_0}$.
This Hamiltonian can be quantized by imposition of canonical commutation relations, whose non-zero components read
\begin{align}
\label{eq:comm1}
\left[\mathbf{X}(\mathbf{r}),\mathbf{P}(\mathbf{r}') \right]&=i\hbar\boldsymbol{\delta}(\mathbf{r-r'}),\\
\left[\mathbf{D}(\mathbf{r}),\mathbf{A}(\mathbf{r}') \right]&=i\hbar\boldsymbol{\delta}^{\mathrm{T}}(\mathbf{r-r'}),
\end{align}
where $\boldsymbol{\delta}^{\mathrm{T}}(\mathbf{r-r'})$ is the transverse delta function and $\hbar$ is Planck's constant. From \Eq{eq:comm1} we can derive the further commutation relation between electric displacement and magnetic field
\begin{equation}
\label{eq:comm2}
\left[\mathbf{D}(\mathbf{r}),\mathbf{H}(\mathbf{r}') \right]= i\frac{\hbar}{\mu_0} \nabla' \times \boldsymbol{\delta}(\mathbf{r-r'}),
\end{equation}
where the prime on $\nabla$ denotes it acting upon primed variables.
Being the Hamiltonian quadratic in the fields, in the spirit of the original Hopfield paper \cite{Hopfield58}, we look for normal modes of the system in the form of polaritonic operators, linear superpositions of the light and matter microscopic fields, weighted by arbitrary space-dependent coefficients
\begin{equation}
\label{eq:Kexp}
\mathcal{K}=\int \mathrm{d}\mathbf{r} \left[\boldsymbol{\alpha}\cdot \mathbf{D} +\boldsymbol{\beta}\cdot \mathbf{H} +\boldsymbol{\gamma}\cdot \mathbf{P}+\boldsymbol{\eta}\cdot \mathbf{X}\right],
\end{equation}
where we will refer to $\boldsymbol{\alpha},\;\boldsymbol{\beta},\;\boldsymbol{\gamma},\;\boldsymbol{\eta}$ as real-space Hopfield coefficients and to $\mathcal{K}$ as the annihilation operator of a polariton mode. The choice of the four microscopic operators we use in the definition of the polariton operators in \Eq{eq:Kexp} is somehow arbitrary, and we could have used as well creation and annihilation operators for the light and matter fields, as in the original Hopfield paper \cite{Hopfield58}, or conjugate variables $\boldsymbol{\Pi}$ and $\boldsymbol{\mathrm{A}}$ also for the electromagnetic field.
In order for the polariton operator to diagonalise the Hamiltonian $\mathcal{H}_0$ it must obey the Heisenberg equation
\begin{equation}
\label{eq:eigenvalues}
\left[ \mathcal{K}, \mathcal{H}_0\right] = \hbar\omega \mathcal{K}.
\end{equation}
Using \Eqs{eq:PZWH}{eq:Kexp} we obtain
\begin{equation}
\begin{split}
\left[ \mathcal{K},\mathcal{H}_0\right]&=i\hbar\int \mathrm{d}\mathbf{r}\, \nabla\times\boldsymbol{\alpha} \cdot\mathbf{H}
-c^2\nabla\times\boldsymbol{\beta}\cdot\mathbf{D}
\\ & \quad +\kappa \,c^2\nabla\times\boldsymbol{\beta}\cdot \mathbf{X} -{\rho}{\omega}_{\mathrm{L}}^2\boldsymbol{\gamma}\cdot \mathbf{X} +\frac{\kappa }{\epsilon_0}\boldsymbol{\gamma}\cdot \mathbf{D}
+\frac{1}{{\rho}}\boldsymbol{\eta}\cdot \mathbf{P}\\&=\hbar
\omega \int \mathrm{d}\mathbf{r}\, \boldsymbol{\alpha}\cdot \mathbf{D}+\boldsymbol{\beta}\cdot \mathbf{H} +\boldsymbol{\gamma}\cdot \mathbf{P}+\boldsymbol{\eta}\cdot \mathbf{X},
\end{split}
\end{equation}
which, equating the coefficients of the different operators, can be restated as the eigensystem
\begin{align}
\label{eq:diffsys1}
\omega\boldsymbol{\alpha}&=-ic^2\nabla\times
\boldsymbol{\beta}+i\frac{\left[\kappa \boldsymbol{\gamma}\right]^{\mathrm{T}}}{\epsilon_0},\\
\label{eq:diffsys2}
\omega\boldsymbol{\beta}&=i\nabla\times
\boldsymbol{\alpha},\\
\label{eq:diffsys3}
\omega\boldsymbol{\gamma}&=i\frac{\boldsymbol{\eta}}{{\rho}},
\\
\label{eq:diffsys4}
\omega\boldsymbol{\eta}&=i\kappa \,c^2\nabla\times
\boldsymbol{\beta}-i{\rho}{\omega}_{\mathrm{L}}^2\boldsymbol{\gamma}.
\end{align}
We have thus transformed the operator-valued \Eq{eq:eigenvalues} into a set of equations, equivalent to Maxwell equations in matter, over the Hopfield coefficients.
Such a system can be formally solved yielding discrete eigenmodes
\begin{align}
\label{eq:Psi}
\ket{\boldsymbol{\Psi}_{{n}} }&=\left(\boldsymbol{\alpha}_{{n}} ,\boldsymbol{\beta}_{{n}} ,\boldsymbol{\gamma}_{{n}} ,\boldsymbol{\eta}_{{n}} \right),
\end{align}
and the relative eigenfrequencies $\omega_{{n}}$.
Bamba and Ogawa showed that polarizable matter is stable against superradiant phase transitions both in the homogeneous and inhomogeneous cases \cite{Bamba14}, implying that \Eqs{eq:diffsys1}{eq:diffsys4} do not present zero energy solutions. In particular they proved that, under suitable continuity conditions, an inhomogeneous system is stable if the condition of stability for the corresponding point-wise homogeneous system is verified. In the homogeneous case, by evaluating \Eq{eq:diffsys1} and \Eq{eq:diffsys4} for $\omega=0$, we infer the stability condition $\omega_{\mathrm{T}}>0$. In the following we will thus consider a transverse frequency everywhere positive, although possibly arbitrarily small.
From direct inspection we can verify that if $\ket{\boldsymbol{\Psi}_{{n}} }$ in \Eq{eq:Psi} is solution of \Eqs{eq:diffsys1}{eq:diffsys4} then
\begin{align}
\label{eq:Psibar}
\ket{\boldsymbol{\Psi}_{\bar{{n}}} }&=\left(\bar{\boldsymbol{\alpha}}_{{n}} ,\bar{\boldsymbol{\beta}}_{{n}} ,\bar{\boldsymbol{\gamma}}_{{n}} ,\bar{\boldsymbol{\eta}}_{{n}} \right),
\end{align}
is also solution with eigenvalue $\omega_{\bar{{n}}} =-\omega_{{n}} $, that is for each positive energy solution there exist a negative energy one such that $\mathcal{K}_{{n}} ^{\dagger}=\mathcal{K}_{\bar{{n}}}$.
This is a general feature, that remains valid also if we chose a different representation for the polaritonic operators (e.g., representing them as linear superpositions of creation and annihilation operators), even if in this case the coefficients of the negative energy solutions will in general not be the complex conjugate of the positive energy ones. This bipartition of the solutions into positive and negative energy subspaces is fundamental for our interpretation of the positive (negative) energy polariton operators as annihilation (creation) operators for the relative excitations.
In order to determine the eigenmodes of \Eqs{eq:diffsys1}{eq:diffsys4} it is convenient to introduce the novel variable
\begin{align}
\label{eq:novel}
\boldsymbol{\theta}=\boldsymbol{\alpha}+i\frac{\left[\kappa \boldsymbol{\gamma}\right]^{\mathrm{L}}}{\omega\epsilon_0},
\end{align}
where $\mathrm{L}$ denotes the longitudinal component, that allows us to restate \Eqs{eq:diffsys1}{eq:diffsys4}, as the wave equation
\begin{align}
\label{eq:waveeq}
\nabla \times \nabla \times \boldsymbol{\theta} = \frac{\omega^2 \epsilon\left(\omega\right)}{c^2} \boldsymbol{\theta},
\end{align}
where
\begin{align}
\epsilon\left(\omega\right) = \frac{{\omega}_{\mathrm{L}}^2 - \omega^2}{\omega_{\mathrm{T}}^2 - \omega^2},
\end{align}
is the (generally space-dependent) dielectric function of a lossless dielectric composed of Lorentz oscillators with ${\omega}_{\mathrm{L}}$ and $\omega_{\mathrm{T}}$ as longitudinal and transverse resonant frequencies.
We recognise in \Eq{eq:waveeq} the electromagnetic wave equation in an inhomogeneous, non-magnetic medium, that allows us to solve the differential problem in terms of $\boldsymbol{\theta}$ with the usual methods employed in classical electromagnetism. The coefficients of the polariton operator can then be calculated from \Eq{eq:novel} and \Eqs{eq:diffsys1}{eq:diffsys4} as
\begin{align}
\boldsymbol{\alpha} & = \boldsymbol{\theta}^{\mathrm{T}},\\
\boldsymbol{\beta} & = \frac{i }{\omega} \nabla \times \boldsymbol{\theta},\\
\boldsymbol{\gamma} &= i\frac{\kappa\omega}{{\rho}\left(\omega_{\mathrm{T}}^2 - \omega^2\right)} \boldsymbol{\theta},\\
\boldsymbol{\eta} &= \frac{\kappa \omega^2}{\omega_{\mathrm{T}}^2 - \omega^2} \boldsymbol{\theta}.
\label{eq:betaexpression}
\end{align}
Still we notice that \Eq{eq:waveeq} is generally not in the form of an eigenvalue problem, due to the simultaneous frequency and position dependence of the dielectric function, making it simpler to reason using the eigensystem in \Eqs{eq:diffsys1}{eq:diffsys4}.
Once the eigenmodes have been determined, they can be normalized requiring the polariton operators to obey
\begin{align}
\label{eq:BosonComm}
\left[\mathcal{K}_{{m}}, \mathcal{K}_{{n}}^{\dagger}\right] &=\delta_{\mathrm{m,n}}\,\mathrm{sgn}(\omega_{{n}}),
\end{align}
where $\mathrm{sgn}(\omega)=\tfrac{\omega}{\lvert\omega\rvert}$ is the sign function. Such an equation reduces to the standard bosonic commutation relation if we restrict the indexes ${m}$ and ${n}$ only over the positive energy solutions.
In term of the Hopfield coefficients \Eq{eq:BosonComm} reads
\begin{align}
\label{eq:BosonCommCoeff}
&i\hbar\int \mathrm{d}\mathbf{r}\,\frac{1}{\mu_0}\boldsymbol{\alpha}_{m}\cdot\nabla\times\bar{\boldsymbol{\beta}}_{n}-\frac{1}{\mu_0}\nabla\times{\boldsymbol{\beta}}_{m}\cdot\bar{\boldsymbol{\alpha}}_{n}
\nonumber+\boldsymbol{\eta}_{m}\cdot\bar{\boldsymbol{\gamma}}_{n}
-\boldsymbol{\gamma}_{m}\cdot\bar{\boldsymbol{\eta}}_{n}\\&=\hbar\epsilon_0 \int \mathrm{d}\mathbf{r}
\frac{\epsilon(\omega_{{m}})\omega_{{m}}^2-\epsilon(\omega_{{n}})\omega_{{n}}^2}{\omega_{{m}}-\omega_{{n}}}\boldsymbol{\theta}_{{m}}\cdot\bar{\boldsymbol{\theta}}_{{n}}=\delta_{\mathrm{m,n}}\,\mathrm{sgn}(\omega_{{n}}),
\end{align}
that in the case ${m}={n}$ may conveniently be expressed as
\begin{equation}
\label{eq:bosononorm}
\hbar\omega_{{n}}\epsilon_0 \int \mathrm{d}\mathbf{r}\,\epsilon\left(\omega_n\right) \nu(\omega_{{n}})\, \boldsymbol{\theta}_{{n}}\cdot\bar{\boldsymbol{\theta}}_{{n}}=\mathrm{sgn}(\omega_{{n}}),
\end{equation}
where the function $\nu(\omega)$ is given by
\begin{equation}
\label{eq:Vrat}
\nu(\omega)=1+ \frac{1}{\epsilon\left(\omega\right)} \frac{\partial \left[\epsilon(\omega)\omega\right]}{\partial\omega} = \frac{\mathrm{v}^\mathrm{G}\left(\omega\right)}{\mathrm{v}^{\mathrm{P} }\left(\omega\right)},
\end{equation}
which is equivalent to the ratio of the local group and phase velocities $\mathrm{v}^\mathrm{G}\left(\omega\right)$ and $\mathrm{v}^{\mathrm{P}} \left(\omega\right)$ \cite{Ruppin02}.
\subsection{Expressions of the microscopic fields}
We can now express the microscopic light and matter fields as linear combinations of the polaritonic modes
\begin{align}
\label{eq:elecfieldexp}
\mathbf{D}&=\sum_{n} \mathbf{f}_{n}^\mathrm{D}\mathcal{K}_{n}
=\sum'_{n}\left[ \mathbf{f}_{n}^\mathrm{D}\mathcal{K}_{n}+\bar{\mathbf{f}}_{{n}}^\mathrm{D}\mathcal{K}^{\dagger}_{n}\right],\\
\label{eq:magfieldexp}
\mathbf{H}&=\sum_{n} \mathbf{f}_{n}^\mathrm{H}\mathcal{K}_{n}=\sum'_{n} \left[\mathbf{f}_{n}^\mathrm{H}\mathcal{K}_{n}+\bar{\mathbf{f}}_{{n}}^\mathrm{H}\mathcal{K}^{\dagger}_{n}\right],\\
\label{eq:momfieldexp}
\mathbf{P}&=\sum_{n} \mathbf{f}_{n}^{\mathrm{P}} \mathcal{K}_{n}=\sum'_{n} \left[\mathbf{f}_{n}^\mathrm{P}\mathcal{K}_{n}+\bar{\mathbf{f}}_{{n}}^\mathrm{P}\mathcal{K}^{\dagger}_{n}\right],\\
\label{eq:polfieldexp}
\mathbf{X}&=\sum_{n} \mathbf{f}_{n}^\mathrm{X}\mathcal{K}_{n}=\sum'_{n} \left[\mathbf{f}_{n}^\mathrm{X}\mathcal{K}_{n}+\bar{\mathbf{f}}_{{n}}^\mathrm{X}\mathcal{K}^{\dagger}_{n}\right],
\end{align}
where the primed sums are intended to be only on positive energy solutions. Expansion coefficients for the positive energy solutions can be calculated using the expansion of the polariton operators in \Eq{eq:Kexp} and the commutator relations in \Eqs{eq:comm1}{eq:comm2} as
\begin{align}
\label{eq:ModeMag}
\mathbf{f}_{n}^\mathrm{D}&=\left[\mathbf{D},\mathcal{K}_{n}^{\dagger}\right]=-i\frac{\hbar}{\mu_0}\nabla\times\bar{\boldsymbol{\beta}}_{{n}},\\
\mathbf{f}_{n}^\mathrm{H}&=\left[\mathbf{H},\mathcal{K}_{n}^{\dagger}\right]=i\frac{\hbar}{\mu_0}\nabla\times\bar{\boldsymbol{\alpha}}_{n},\\
\mathbf{f}_{n}^{\mathrm{P}} &=\left[\mathbf{P},\mathcal{K}_{n}^{\dagger}\right]=- i\hbar \bar{\boldsymbol{\eta}}_{n},\\
\mathbf{f}_{n}^\mathrm{X}&=\left[\mathbf{X},\mathcal{K}_{n}^{\dagger}\right]= i\hbar \bar{\boldsymbol{\gamma}}_{{n}}.
\end{align}
\subsection{Orthonormality and Completeness}
It remains to verify that our procedure yields a complete set of solutions in the form of \Eq{eq:Kexp}, satisfying \Eq{eq:BosonComm}.
The polariton normalization in \Eq{eq:BosonCommCoeff} offers the natural inner product over the space of Hopfield coefficients
$\ket{\boldsymbol{\Psi}}=\left(\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\gamma},\boldsymbol{\eta}\right)$, as orthonormality would assure the different polariton modes respect the required commutator relations. We thus define the inner product between two wavefunctions
$\ket{\boldsymbol{\Psi}'}$ and $\ket{\boldsymbol{\Psi}}$ as
\begin{align}
\label{eq:innerproduct}
\langle\braket{\boldsymbol{\Psi}'}{\boldsymbol{\Psi}}\rangle=\,
i\hbar\int \mathrm{d}\mathbf{r}\biggr[&\frac{1}{\mu_0}\boldsymbol{\alpha}\cdot \nabla\times\bar{\boldsymbol{\beta}}'-\frac{1}{\mu_0}\nabla\times{\boldsymbol{\beta}}\cdot\bar{\boldsymbol{\alpha}}'\\&
+\boldsymbol{\eta}\cdot\bar{\boldsymbol{\gamma}}'
-\boldsymbol{\gamma}\cdot\bar{\boldsymbol{\eta}}'\biggr].\nonumber
\end{align}
From \Eq{eq:BosonCommCoeff} such an inner product is not positive definite and thus the vector space of the $\ket{\boldsymbol{\Psi}}$ does not form an Hilbert space over the scalar product in \Eq{eq:innerproduct}. Still it forms a Kre\v\i n space \cite{Dritschel96} with signature operator
\begin{align}
\eta=&\sum'_{n} \left[\ket{\boldsymbol{\Psi}_{{n}} }\rangle\langle\bra{\boldsymbol{\Psi}_{{n}} }-\ket{\boldsymbol{\Psi}_{\bar{{n}}} }\rangle\langle\bra{\boldsymbol{\Psi}_{\bar{{n}}} }\right],
\end{align}
allowing us to recover the structure of an Hilbert space over the inner product
\begin{align}
\label{eq:competenessformal}
\braket{\boldsymbol{\Psi}'}{\boldsymbol{\Psi}}=&\langle\bra{\boldsymbol{\Psi}'}\eta\ket{\boldsymbol{\Psi}}\rangle.
\end{align}
Recasting the eigensystem in \Eqs{eq:diffsys1}{eq:diffsys4} in operatorial form as
\begin{align}
\mathcal{B}_0\ket{\boldsymbol{\Psi}}&=\omega \ket{\boldsymbol{\Psi}},
\end{align}
we have to impose $\mathcal{B}_0$ to be self-adjoint with respect to an inner product defined in \Eq{eq:competenessformal}, that is
\begin{align}
\label{eq:self}
\braket{\boldsymbol{\Psi}'}{\mathcal{B}_0 \boldsymbol{\Psi}}
- \braket{\mathcal{B}_0 \boldsymbol{\Psi}'}{\boldsymbol{\Psi}} &=0.
\end{align}
Using \Eqs{eq:diffsys1}{eq:diffsys4} and \Eqs{eq:innerproduct}{eq:competenessformal} we can transform \Eq{eq:self} into
\begin{align}
\label{eq:bcon1}
\int \mathrm{d}\mathbf{r} \left[\boldsymbol{\alpha}\cdot\nabla\times\nabla\times\bar{\boldsymbol{\alpha}}'
-\bar{\boldsymbol{\alpha}}'\cdot\nabla\times\nabla\times{\boldsymbol{\alpha}}\right]&=0.
\end{align}
Imposing the condition $\nabla\cdot{\boldsymbol{\alpha}}=0$, that from \Eq{eq:ModeMag} is just the transversality of the electric displacement field $\nabla\cdot{\boldsymbol{\mathrm{D}}}=0$, and using Green's second identity, we can put \Eq{eq:bcon1} in the form
\begin{align}
\label{eq:boundarycondition}
\int_{\partial \mathcal{V}} \mathrm{d}S\,\left[
\boldsymbol{\alpha}\cdot \left(\mathbf{n}\cdot\nabla\right) \bar{\boldsymbol{\alpha}}'-\bar{\boldsymbol{\alpha}}' \cdot\left(\mathbf{n}\cdot\nabla\right){\boldsymbol{\alpha}}\right]=0,
\end{align}
where $\partial \mathcal{V}$ is the surface of the quantization volume $\mathcal{V}$, $\mathrm{d}S$ the infinitesimal surface element, and $\mathbf{n}$ the unit vector normal to it. It is possible to satisfy \Eq{eq:boundarycondition} imposing proper homogeneous boundary conditions for $\boldsymbol{\alpha}$ on $\partial \mathcal{V}$ and under such conditions the system in \Eqs{eq:diffsys1}{eq:diffsys4} is self-adjoint, allowing us to find a complete set of solutions in the form of the polaritonic operators defined in \Eq{eq:Kexp} respecting bosonic commutation relations in \Eq{eq:BosonComm}.
We can at this point write the completeness relation that, for arbitrary $\ket{\boldsymbol{\Psi}}$ and $\ket{\boldsymbol{\Psi}'}$, reads
\begin{eqnarray}
\bra{\boldsymbol{\Psi}}\left[ \sum_n \ket{\boldsymbol{\Psi}_n}\bra{\boldsymbol{\Psi}_n}\right]\ket{\boldsymbol{\Psi}'}= \braket{\boldsymbol{\Psi}}{\boldsymbol{\Psi}'},
\end{eqnarray}
leading to expressions between the Hopfield coefficients of the form
\begin{align}
\label{eq:completenessexplicit}
\sum_n'\biggr[\boldsymbol{\alpha}_{{n}}(\mathbf{r'})\cdot\nabla\times\bar{\boldsymbol{\beta}}_{{n}}(\mathbf{r})
-\bar{\boldsymbol{\alpha}}_{{n}}(\mathbf{r'})\cdot\nabla\times{\boldsymbol{\beta}}_{{n}}(\mathbf{r})\biggr]=i\frac{\mu_0}{\hbar}\delta^{\mathrm{T}}(\mathbf{r-r'}),
\end{align}
allowing us to verify the consistence of our procedure by explicitly calculating the commutators we used to quantize the theory
\begin{align}
\left[\mathbf{D}(\mathbf{r}),\mathbf{H}(\mathbf{r'}) \right]=& \frac{\hbar^2}{\mu_0^2}\sum'_{{n}}\biggr[
\nabla\times\bar{\boldsymbol{\beta}}_{{n}}(\mathbf{r})\cdot
\nabla'\times{\boldsymbol{\alpha}}_{{n}}(\mathbf{r'})\\&\quad\quad\;
-\nabla\times{\boldsymbol{\beta}}_{{n}}(\mathbf{r})\cdot
\nabla'\times{\bar{\boldsymbol{\alpha}}}_{{n}}(\mathbf{r'})\biggr]\nonumber\\
=&\frac{\hbar^2}{\mu_0^2}\sum'_{{n}}\nabla'\times\biggr[\nonumber
\nabla\times\bar{\boldsymbol{\beta}}_{{n}}(\mathbf{r})\cdot
{\boldsymbol{\alpha}}_{{n}}(\mathbf{r'})\\&\quad\quad\quad\quad\;\,\,
-\nabla\times{\boldsymbol{\beta}}_{{n}}(\mathbf{r})\cdot
{\bar{\boldsymbol{\alpha}}}_{{n}}(\mathbf{r'})\biggr]\nonumber\\
=&i\frac{\hbar}{\mu_0}\nabla'\times\delta^{\mathrm{T}}(\mathbf{r-r'}),\nonumber
\end{align}
with equivalent expressions arising for the other coefficients and leading to the other commutators between the microscopic fields.
The Hamiltonian in \Eq{eq:PZWH} can thus be put in the diagonal form
\begin{align}
\mathcal{H}_0=&\sum'_{{n}} \hbar\omega_{{n}}\, \mathcal{K}_{{n}}^{\dagger} \mathcal{K}_{{n}}.
\end{align}
\section{Application: Phonon Polaritons at a Planar Interface}
\begin{figure}
\includegraphics[width=0.45\textwidth]{QuantFig}
\caption{\label{fig:Fig1} Dispersions of the different classes of solutions of a planar interface between vacuum and a polar dielectric.
The dispersion of photons in vacuum (red) and of bulk phonon polaritons in the dielectric (blue) are plotted as a function of the three-dimensional wavevector $k$, while the dispersion of the surface modes (yellow) is plotted as a function of the two-dimensional in-plane wavevector $k_{\parallel}$. Representative field profiles for TM bulk and surface solutions are represented as insets.}
\end{figure}
In order to illustrate the method developed in the previous section, we will apply it to the well-known case of a planar interface between vacuum and a polar dielectric described by the Lorentz dielectric function, where the coupling of the electromagnetic field with the transverse optical phonons gives rise to both bulk and surface phonon polariton excitations \cite{Xu90}. Assuming the surface to lie in the $x-y$ plane, the air-dielectric interface, sketched in Fig.~\ref{fig:Fig1}(a), will be described by the dielectric function
\begin{align}
\label{epsvd}
\nonumber
\epsilon\left(\omega,z>0\right) &= 1,\\
\epsilon\left(\omega,z<0\right) &= \epsilon_{\mathrm{L}}\left(\omega\right)= \frac{{\omega}_{\mathrm{LO}}^2- \omega^2}{\omega_{\mathrm{TO}}^2 -\omega^2},
\end{align}
where ${\omega}_{\mathrm{LO}}$ and ${\omega}_{\mathrm{TO}}$ are the longitudinal and transverse optical phonon frequencies, linked between them by the Lyddane-Sachs-Teller relation \cite{Lyddane41}.
The eigenmodes and the relative eigenfrequencies can be calculated solving \Eq{eq:waveeq} with the dielectric function in \Eq{epsvd}, leading to both bulk and surface solutions. The bulk, propagative ones, impinging upon the surface from each side, are indexed by their polarization ($\mathrm{TM}$ or $\mathrm{TE}$) and three dimensional wavevector in the medium of origin, $\mathbf{k}=(\mathbf{k}_{\parallel},k_z)$, with $k_z>0$ for waves coming from vacuum ($\mathrm{v}$) and $k_z<0$ for waves coming from the dielectric. The former ones need also an extra index over the two bulk phonon polariton branches, lower ($\mathrm{l}$) and upper ($\mathrm{u}$) polaritons, existing in the dielectric for each value of $\mathbf{k}$. The surface solutions instead form a single branch of surface phonon polaritons, indexed by the two-dimensional in-plane wavevector $\mathbf{k}_{\parallel}$, and they only exist for $ck_{\parallel}\geq \omega_{\mathrm{TO}}$. The dispersion of the different modes can be found in Fig.~\ref{fig:Fig1}.
We can thus write the general expression of a polariton operators
\begin{align}
\label{eq:Kexpspecific}
\mathcal{K}_{\mathbf{k}^{\sigma}}^{\sigma} =&\int \mathrm{d}\mathbf{r} \biggr[ \boldsymbol{\alpha}_{\mathbf{k}^{\sigma}}^{\sigma} \cdot \mathbf{D} +\frac{i}{\omega_{\mathbf{k}^{\sigma}}} \nabla \times \boldsymbol{\alpha}_{\mathbf{k^{\sigma}}}^{\sigma} \cdot \mathbf{H} \\ &\nonumber + i\frac{\kappa \omega_{\mathbf{k}^{\sigma}}}{\rho\left(\omega_{\mathrm{TO}}^2 - \omega_{\mathbf{k}^{\sigma}}^2\right)} \boldsymbol{\alpha}_{\mathbf{k}^{\sigma}}^{{\sigma}} \cdot \mathbf{P} + \frac{\kappa \omega_{\mathbf{k}^{\sigma}}^2}{\left(\omega_{\mathrm{TO}}^2 - \omega_{\mathbf{k}^{\sigma}}^2\right)} \boldsymbol{\alpha}_{\mathbf{k}^{\sigma}}^{\sigma} \cdot \mathbf{X}\biggr],
\end{align}
where $\sigma=\lbrack$TMv, TMl, TMu, TEv, TEl, TEu, S$\rbrack$ indexes the different classes of solutions, and $\mathbf{k}^{\sigma}$ is the relative two- or three-dimensional wavevector.
In the following we will explicitly consider only the surface phonon polariton modes ($\sigma=$S) and a single kind of bulk propagative solution ($\sigma=$TEv, the $\mathrm{TE}$ polarised modes incident from the vacuum side of the interface), as the other solutions lead to very similar expressions. Introducing the orthogonal basis vectors
$\hat{\mathbf{e}}_{z}$, $\hat{\mathbf{e}}_{\parallel}$, and $\hat{\mathbf{e}}_{\perp}$, oriented along the $z$ axis and in the $x-y$ plane respectively parallel and perpendicular to $\mathbf{k}_{\parallel}$, we can write the positive energy solutions of \Eq{eq:waveeq} for the TEv modes as
\begin{align}
\label{eq:prop}
\boldsymbol{\theta}^{\mathrm{TEv}}_{\mathbf{k}}(\mathbf{r}_{\parallel},z>0) &= \mathrm{N}^{\mathrm{TEv}}_{\mathbf{k}} e^{-i \mathbf{k}_{\parallel}\cdot\mathbf{r}_{\parallel} } \left(e^{i k_{z} z} + r^{\mathrm{TEv}}_{\mathbf{k}} e^{-i k_{z} z}\right) \hat{\mathbf{e}}_{\perp},\\
\boldsymbol{\theta}^{\mathrm{TEv}}_{\mathbf{k}}(\mathbf{r}_{\parallel},z<0) &= \mathrm{N}^{\mathrm{TEv}}_{\mathbf{k}} t^{\mathrm{TEv}}_{\mathbf{k}} e^{-i \mathbf{k}_{\parallel}\cdot\mathbf{r}_{\parallel}} e^{i \sqrt{\epsilon_{\mathrm{L}}\left(\omega_{\mathbf{k}}\right) k^2 - k_{\parallel}^2} z} \hat{\mathbf{e}}_{\perp},\nonumber
\end{align}
where the Fresnel coefficients read
\begin{align}
r^{\mathrm{TEv}}_{\mathbf{k}} &= \frac{k_{z} - \sqrt{\epsilon_{\mathrm{L}}\left(\omega_{\mathbf{k}}\right){k}^2-{k}_{\parallel}^2}}{k_{z} + \sqrt{\epsilon_{\mathrm{L}}\left(\omega_{\mathbf{k}}\right){k}^2-{k}_{\parallel}^2}},\\
t^{\mathrm{TEv}}_{\mathbf{k}} &= \frac{2 k_{z}}{k_{z} + \sqrt{\epsilon_{\mathrm{L}}\left(\omega_{\mathbf{k}}\right){k}^2-{k}_{\parallel}^2}}.
\end{align}
For the S modes we have instead
\begin{align}
\label{eq:surf}
\boldsymbol{\theta}^{\mathrm{S}}_{\mathbf{k}_{\parallel}}(\mathbf{r}_{\parallel},z>0) &=\mathrm{N}^{\mathrm{S}}_{\mathbf{k}_{\parallel}} \biggr[ \frac{1}{\sqrt{\epsilon_{\mathrm{L}}\left(\omega_{\mathbf{k}_{\parallel}}\right)}} \hat{\mathbf{e}}_{\parallel} + \hat{\mathbf{e}}_z \biggr] \\ & \quad \times e^{- i \mathbf{k}_{\parallel}\cdot \mathbf{r}_{\parallel}} e^{-k_{\parallel}\sqrt{\frac{-1}{\epsilon_{\mathrm{L}}\left(\omega_{\mathbf{k}_{\parallel}}\right)}} z},\nonumber\\
\boldsymbol{\theta}^{\mathrm{S}}_{\mathbf{k}_{\parallel}}(\mathbf{r}_{\parallel},z<0) &= \frac{\mathrm{N}^{\mathrm{S}}_{\mathbf{k}_{\parallel}}}{\epsilon_{\mathrm{L}}\left(\omega_{\mathbf{k}_{\parallel}}\right)} \biggr[- \sqrt{\epsilon_{\mathrm{L}}\left(\omega_{\mathbf{k}_{\parallel}}\right)} \hat{\mathbf{e}}_{\parallel} + \hat{\mathbf{e}}_z \biggr]\nonumber \\ & \quad \times e^{- i \mathbf{k}_{\parallel}\cdot \mathbf{r}_{\parallel}} e^{k_{\parallel}\sqrt{-\epsilon_{\mathrm{L}}\left(\omega_{\mathbf{k}_{\parallel}}\right)} z}.\nonumber
\end{align}
In \Eq{eq:prop} and \Eq{eq:surf} the $\mathrm{N}^{\sigma}_{\mathbf{k}^{\sigma}}$ coefficients are the normalization of the different modes, that can be fixed by plugging the relevant coefficients into \Eq{eq:Kexpspecific}, and then using \Eq{eq:bosononorm}. This yields
\begin{align}
\mathrm{N}^{\mathrm{TEv}}_{\mathbf{k}} =& \sqrt{\frac{1}{2 \epsilon_0 \hbar\omega_{\mathbf{k}} \mathcal{V}}},\\
\mathrm{N}^{\mathrm{S}}_{\mathbf{k}_{\parallel}} =& \sqrt{\frac{k_{\parallel}}{\epsilon_0\hbar\omega_{\mathbf{k}_{\parallel}} \mathcal{A}} }\left[1 - \frac{\nu_{\mathrm{L}}\left(\omega_{\mathbf{k}_{\parallel}}\right)}{\epsilon_{\mathrm{L}}\left(\omega_{\mathbf{k}_{\parallel}}\right)} \right]^{-\frac{1}{2}}\nonumber\\&\times
\left[\frac{1}{\sqrt{-\epsilon_{\mathrm{L}}\left(\omega_{\mathbf{k}_{\parallel}}\right)}} + \sqrt{-\epsilon_{\mathrm{L}}\left(\omega_{\mathbf{k}_{\parallel}}\right)}\right]^{-\frac{1}{2}} ,
\end{align}
where $\mathcal{A}$ is the area of the surface and $\nu_{\mathrm{L}}\left(\omega\right)$ the function obtained using $\epsilon_{\mathrm{L}}\left(\omega\right)$ into \Eq{eq:Vrat}.
We can finally recover the expressions of the microscopic fields by inverting the Hopfield transformation using \Eqs{eq:elecfieldexp}{eq:polfieldexp}, allowing us not only to calculate physical observables, but also to naturally express nonlinear interactions in terms of scattering between polariton operators. In the specific example we are considering here nonlinearities can arise due to phonon anharmonicity, leading to $N^\mathrm{{th}}$ order nonlinear interaction Hamiltonians that can be written in the general form
\begin{align}
\label{eq:HNL}
\mathcal{H}_{\mathrm{NL}}&=\int \mathrm{d}\mathbf{r}\, \sum_{j_1\cdots j_N=1}^3\Phi^{j_1\cdots j_{N}}\prod_{{l}=1}^{N} \mathrm{X}_{j_l},
\end{align}
where the $j$s index space coordinates and values of the nonlinear tensor $\Phi$ for some reference structure can be found in the literature at least up to the fourth order \cite{Vanderbilt85}.
Inverting \Eq{eq:Kexpspecific} and using \Eq{eq:betaexpression} we can express the matter field as
\begin{align}
\label{eq:Xspecific}
\boldsymbol{\mathrm{X}}&=\sum_{{\sigma}}\int \mathrm{d}\mathbf{k}^{\sigma} \frac{\kappa\hbar\omega_{\mathbf{k}^{\sigma}} }{\rho(\omega_{\mathrm{TO}}^2-\omega_{\mathbf{k}^{\sigma}}^2 )}\bar{\boldsymbol{\theta}}^{\sigma}_{\mathbf{k^{\sigma}}}\mathcal{K}^{\sigma}_{\mathbf{k}^{\sigma}},
\end{align}
where $\sigma$ runs over all the different classes of solutions as in \Eq{eq:Kexpspecific}, and including both positive and negative energy solutions.
Substituting \Eq{eq:Xspecific} into the nonlinear Hamiltonian in \Eq{eq:HNL} we thus obtain the interaction Hamiltonian written as scattering terms between the polaritonic operators
\begin{align}
\mathcal{H}_{\mathrm{NL}}&=\sum_{{\sigma}_1\cdots {\sigma}_N}
\prod_{l=1}^N \left[\int \mathrm{d}\mathbf{k}^{{\sigma}_l}\right] \Xi^{{\sigma}_1\cdots {\sigma}_N}_{\mathbf{k}^{\sigma_1}\cdots \mathbf{k}^{\sigma_N}} \prod_{l=1}^N \mathcal{K}^{{\sigma}_l}_{\mathbf{k}^{{\sigma}_l}},
\end{align}
where the scattering coefficients $\Xi$ are obtained performing the relevant integrals over the space variables.
\section{Extension to the dissipative case}
The model described in Sec. II can be extended to include dissipative effects on the same line of the original Huttner and Barnett paper \cite{Huttner92}. This is achieved by coupling the matter field to a continuum bath of harmonic oscillators modelling the continuum in which the matter energy can be dissipated.
Given the continuous character of the resulting spectrum, this will come at the cost of renouncing to the simple modal solution of \Eq{eq:waveeq} and we will have to solve instead a non-homogeneous wave equation. Still it is important to show that in this case our method remains viable, effectively recovering a result equivalent to Ref. \onlinecite{Suttorp04}, and reducing to Ref. \onlinecite{Huttner92} in the homogeneous case.
We thus consider the total Lagrangian $\mathcal{L} = \mathcal{L}_0 + \mathcal{L}_\text{B}$ where $\mathcal{L}_0$, from \Eq{eq:MinCoupLag}, describes the non-dissipative system and
\begin{align}
\mathcal{L}_{\text{B}} = \int_0^{\infty} \mathrm{d} \zeta \left[ \frac{\rho\dot{\mathrm{Y}}_{\zeta}^2}{2} - \frac{\rho \zeta^2\mathrm{Y}_{\zeta}^2}{2} - \upsilon_{\zeta} \mathbf{X} \cdot \dot{\mathbf{Y}}_{\zeta}\right],
\end{align}
the bath of harmonic oscillators $\mathrm{Y}_{\zeta}$, indexed by their frequency $\zeta$, and coupled to the matter mode by the coupling $\upsilon_{\zeta}$. Analogously to what done in Sec. II we will assume that all the integrals over the bath degrees of freedom extend only in the regions where $\rho\neq0$.
The canonical momenta for the bath operators are found as
\begin{align}
&\mathbf{Q}_{\zeta} = \frac{\delta \mathcal{L}_{\text{B}}}{\delta \dot{\mathbf{Y}}_{\zeta}} = \rho\dot{\mathbf{Y}}_{\zeta} - \upsilon_{\zeta}\mathbf{X}.
\end{align}
The corresponding total Hamiltonian will thus be given by $\mathcal{H} = \mathcal{H}_0 + \mathcal{H}_\text{B}$ with
\begin{align}
\mathcal{H}_{\text{B}} = \int \mathrm{d}\mathbf{r} \int_{0}^{\infty} \mathrm{d}\zeta \left[\frac{\rho \zeta^2\mathrm{Y}_{\zeta}^2}{2} + \frac{\mathrm{Q}_{\zeta}^2}{2 \rho} + \frac{\upsilon_{\zeta}}{\rho} \mathbf{Q}_{\zeta} \cdot \mathbf{X} + \frac{\upsilon_{\zeta}^2\mathrm{X}^2}{2 \rho} \right],
\label{HB}
\end{align}
where the last term in \Eq{HB} can be included into $\mathcal{H}_0$ introducing the renormalized longitudinal frequency
\begin{equation}
\tilde{\omega}_{\mathrm{L}}^2 = \omega_{\mathrm{T}}^2 + \frac{\kappa ^2}{\epsilon_0 \rho} + \int_0^{\infty} \mathrm{d} \zeta\, \frac{\upsilon_{\zeta}^2}{2 \rho^2} ,
\end{equation}
that includes the static shift from coupling to the bath.
The full Hamiltonian $\mathcal{H}$ is quantized by imposition of commutation relations whose non-zero elements are those in \Eqs{eq:comm1}{eq:comm2}, in addition to
\begin{equation}
\left[\mathbf{Y}_{\zeta}\left(\mathbf{r}\right), \mathbf{Q}_{\zeta'}\left(\mathbf{r'}\right)\right] = i \hbar\boldsymbol{\delta}\left(\mathbf{r}- \mathbf{r}'\right)\delta\left(\zeta-\zeta'\right).
\end{equation}
The polaritonic operators are now defined as
\begin{align}
\label{nonumber}
\mathcal{K}&=\int \mathrm{d}\mathbf{r} \biggr[\boldsymbol{\alpha}\cdot \mathbf{D} +\boldsymbol{\beta}\cdot \mathbf{H} +\boldsymbol{\gamma}\cdot \mathbf{P}+\boldsymbol{\eta}\cdot \mathbf{X} \\&\quad+ \int_0^{\infty} \mathrm{d} \zeta \left(\boldsymbol{\chi}_{\zeta}\cdot\mathbf{Q}_{\zeta} + \boldsymbol{\xi}_{\zeta}\cdot\mathbf{Y}_{\zeta} \right)\biggr],
\label{eq:Kexploss}
\end{align}
and solving the Heisenberg equation
\begin{equation}
\left[\mathcal{K}, \mathcal{H}\right] = \hbar\omega \mathcal{K},
\label{eq:CommFull}
\end{equation}
we obtain the system of equations
\begin{align}
\label{eq:Ndiffsys1}
\omega\boldsymbol{\alpha}&=-ic^2\nabla\times
\boldsymbol{\beta}+i\frac{\left[\kappa \boldsymbol{\gamma}\right]^{\mathrm{T}}}{\epsilon_0},\\
\label{eq:Ndiffsys2}
\omega\boldsymbol{\beta}&=i\nabla\times
\boldsymbol{\alpha},\\
\label{eq:Ndiffsys3}
\omega\boldsymbol{\gamma}&=i\frac{\boldsymbol{\eta}}{{\rho}},
\\
\label{eq:Ndiffsys4}
\omega\boldsymbol{\eta}&=i\kappa \,c^2\nabla\times
\boldsymbol{\beta}-i{\rho}{\omega}_{\mathrm{L}}^2\boldsymbol{\gamma}
+i\frac{\upsilon_{\zeta}}{\rho}\boldsymbol{\xi}_{\zeta},\\
\label{eq:Ndiffsys5}
\omega\boldsymbol{\chi}_{\zeta}&=i\frac{\boldsymbol{\xi}_{\zeta}}{\rho}-i\frac{\upsilon_{\zeta}}{\rho}
\boldsymbol{\gamma}\\
\label{eq:Ndiffsys6}
\omega\boldsymbol{\xi}_{\zeta}&=-i\rho\zeta^2 \boldsymbol{\chi}_{\zeta}.
\end{align}
Such a system may be solved using the method originally due to Fano \cite{Fano56}, by using \Eq{eq:Ndiffsys6} to eliminate $\boldsymbol{\chi}_{\zeta}$ and then writing the bath operator $\boldsymbol{\xi}_{\zeta}$ as a function of $\boldsymbol{\gamma}$ as
\begin{equation}
\label{eq:ydef}
\boldsymbol{\xi}_{\zeta} = \mathcal{P}\left[\frac{\upsilon_{\zeta} \zeta^2 }{ \zeta^2-\omega^2}\boldsymbol{\gamma}\right] + \mathbf{y}\left(\omega\right) \delta\left(\zeta-\omega\right),
\end{equation}
where $\mathcal{P}$ refers to the principal part and $\mathbf{y}\left(\omega\right)$ is a frequency- and space-dependent function to be determined. Analogously to what done in Sec. II we can restate the system in \Eqs{eq:Ndiffsys1}{eq:Ndiffsys6} as an inhomogeneous wave equation
\begin{equation}
\label{eq:waveeqloss}
\nabla\times \nabla\times \boldsymbol{\theta} - \frac{\omega^2 \tilde{\epsilon}\left(\omega\right)}{c^2} \boldsymbol{\theta} =i\omega\, \mathbf{j}\left(\omega\right),
\end{equation}
where
\begin{align}
\tilde{\epsilon}\left(\omega\right) = \frac{\tilde{\omega}_{\mathrm{L}}^2 - \omega^2- \mathcal{F}\left(\omega\right)}{\omega_{\mathrm{T}}^2 - \omega^2 - \mathcal{F}\left(\omega\right)},
\label{eq:LossyDieFun}
\end{align}
is the complex dielectric function with
\begin{equation}
\mathcal{F}\left(\omega\right)=\mathcal{P}\left[\int_{0}^{\infty} \mathrm{d}\zeta \frac{\upsilon_{\zeta}^2\zeta^2}{{\rho^2}\left(\zeta^2-\omega^2\right)}\right],
\end{equation}
which is of the form derived by Wubs and Suttorp \cite{Suttorp04}, and
the source current $\mathbf{j}\left(\omega\right)$ relates to the function $\mathbf{y}\left(\omega\right)$ as
\begin{equation}
\mathbf{j}\left(\omega\right) = \frac{\upsilon_{\omega}}{{\kappa\rho}c^2}\left[\tilde{\epsilon}\left(\omega\right)-1\right]\mathbf{y}\left(\omega\right).
\end{equation}
Solutions of \Eq{eq:waveeqloss} can now be determined for any value of $\omega$, leading to a continuous spectrum of solutions indexed by the frequency $\omega$ where the function $\mathbf{y}\left(\omega\right)$ takes the role of the normalization factor. In order to fix such a function we impose again the bosonic commutation relation for the polaritonic operators, in a continuous version of \Eq{eq:BosonComm}
\begin{align}
\left[\mathcal{K}_{\omega}, \mathcal{K}_{\omega'}^{\dagger}\right] & =
\delta({\omega-\omega'})\mathrm{sgn}(\omega').
\label{eq:bosonnorm2}
\end{align}
Following exactly the same steps as in Sec. II, extending it to wavevectors of the form
\begin{align}
\ket{\tilde{\boldsymbol{\Psi}}}=\left(\boldsymbol{\alpha},\boldsymbol{\beta},\boldsymbol{\gamma},\boldsymbol{\eta},\boldsymbol{\chi}_{\zeta},\boldsymbol{\xi}_{\zeta} \right),
\end{align}
we can write the total Hamiltonian in diagonal form as
\begin{align}
\mathcal{H}=&\hbar\int_{\omega>0} \mathrm{d}\omega \,\mathcal{K}_{\omega}^{\dagger} \mathcal{K}_{\omega},
\end{align}
where the integral extends over the positive energy spectrum of \Eqs{eq:Ndiffsys1}{eq:Ndiffsys6}.
\section{Conclusion}
We introduced a real-space Hopfield approach to the diagonalization of polarizable media, able to extend to inhomogeneous materials the standard machinery used in the field of quantum polaritonics. Our approach allows us to obtain explicit expressions for the quantum light and matter microscopic fields as a function of the polaritonic operators. Natural applications of this method are in the study of quantum nonlinear processes in inhomogeneous systems, where the microscopic nonlinearity, usually known as a nonlinear function of the matter field, can be expressed as scattering terms between polaritonic operators. We thus expect our theory will be an important tool in the developing fields studying quantum properties of plasmons and other surface excitations, where extremely small mode volumes can lead to strong nonlinear effects.
\section{Acknowledgments}
S.A.M. acknowledges support from EPSRC programme grants EP/L024926/1 and EP/M013812/1, plus ONR Global, the Royal Society, and the Lee-Lucas Chair in Physics.
S.D.L. is Royal Society Research Fellow and he acknowledges support from EPSRC grant EP/M003183/1.
|
\section{Introduction}
This note is devoted to the study of valuations on star bodies. A valuation is a function $V$, defined on a class of sets, with the property that
$$
V(A\cup B)+V(A\cap B)=V(A)+V(B).
$$
As a generalization of the notion of measure, valuations have become a relevant area of study in Convex Geometry. In fact, this notion played a critical role in M. Dehn's solution to Hilbert's third problem, asking whether an elementary definition for volume of polytopes was possible. See, for instance, \cite{Lu1}, \cite{Lu2} and the references there included for a broad vision of the field.
In \cite{Had}, H. Hadwiger provided a characterization of continuous rotationally and translationally invariant valuations on convex bodies as linear combinations of the quermassintegrals. In \cite{Alesker}, S. Alesker studied the valuations on convex bodies which are only rotationally invariant.
Valuations on convex bodies belong to the Brunn-Minkowski Theory. This theory has been extended in several important ways, and in particular, to the dual Brunn-Minkowski Theory, where convex bodies, Minkowski addition and Hausdorff metric are replaced by star bodies, radial addition and radial metric, respectively. The dual Brunn-Minkowski theory, initiated in \cite{Lut_mv1}, has been broadly developed and successfully applied to several areas, such as integral geometry, local theory of Banach spaces and geometric tomography (see \cite{DGP}, \cite{Gabook} for these and other applications). In particular, it played a key role in the solution of the Busemann-Petty problem \cite{Ga1}, \cite{Ga2}, \cite{Zh}.
D. A. Klain initiated in \cite{Klain96}, \cite{Klain97} the study of rotationally invariant valuations on a certain class of star sets, namely those whose radial function is $n$-th power integrable.
In \cite{Vi}, the second named author initiated the study of valuations on star bodies. In this note we continue this study showing that every continuous valuation $V:\mathcal S_0^n\longrightarrow \mathbb R$ on the $n$-dimensional star bodies can be decomposed as the difference of two positive continuous valuations.
The precise result is
\begin{teo}\label{main}
Let $V:\mathcal S_0^n\longrightarrow \mathbb R$ be a radial continuous valuation on the $n$-dimensional star bodies $\mathcal S_0^n$ such that $V(\{0\})=0$. Then, there exist two radial continuous valuations $V^+, V^-:\mathcal S_0^n\longrightarrow \mathbb R_+$ such that $V^+(\{0\})=V^-(\{0\})=0$ and such that
$$
V=V^+-V^-.
$$
Moreover, if $V$ is rotationally invariant, so are $V^+$ and $V^-$.
\end{teo}
With this structural result at hand, the study of continuous valuations on star bodies reduces to the simpler case of positive continuous valuations.
As an application, we can complete the main result of \cite{Vi}. In that paper, {\em positive} rotationally invariant continuous valuations $V$ on the star bodies of $\mathbb R^n$, satisfying that $V(\{0\})=0$ are characterized by an integral representation as in Corollary \ref{representacion} below. The question of whether a similar description is valid for the case of real-valued (not necessarily positive or negative) continuous rotationally invariant valuations was left open.
Now, Theorem \ref{main} allows us to give a positive answer to this question:
\begin{corolario}\label{representacion}
Let $V:\mathcal S_0^n\longrightarrow \mathbb R$ be a rotationally invariant radial continuous valuation on the $n$-dimensional star bodies $\mathcal S_0^n$. Then, there exists a continuous function $\theta:[0,\infty) \longrightarrow \mathbb R$ such that, for every $K\in \mathcal S_0^n$,
$$V(K)=\int_{S^{n-1}} \theta(\rho_K(t)) dm(t),$$
where $\rho_K$ is the radial function of $K$ and $m$ is the Lebesgue measure on $S^{n-1}$ normalized so that $m(S^{n-1})=1$.
Conversely, let $\theta:\mathbb R^+\longrightarrow \mathbb R$ be a continuous function. Then the application $V:\mathcal S_0^n \longrightarrow \mathbb R$ given by
$$V(K)=\int_{S^{n-1}} \theta(\rho_K(t)) dm(t)$$ is a radial continuous rotationally invariant valuation.
\end{corolario}
\smallskip
As in \cite{Vi}, the function $\theta$ in Corollary \ref{representacion} is nothing but $\theta(\lambda)=V(\lambda S^{n-1})$.
\smallskip
The rest of the paper is structured as follows:
In Section \ref{sectionnotation} we describe our notation and some known facts that we will need.
In Section \ref{results} we prove Theorem \ref{main} and Corollary \ref{representacion}
\section{Notation and known facts}\label{sectionnotation}
A set $L\subset \mathbb R^n$ is a {\em star set} if it contains the origin and every line through $0$ that meets $L$ does so in a (possibly degenerate) line segment. Let $\mathcal S^n$ denote the set of the star sets of $\mathbb R^n$.
Given $L\in \mathcal S^n$, we define its {\em radial function} $\rho_L$ by
$$
\rho_L(t)= \sup \{c\geq 0 \, : \, ct\in L\},
$$
for each $t\in\mathbb R^n$. Clearly, radial functions are completely characterized by their restriction to $S^{n-1}$, the euclidean unit sphere in $\mathbb R^n$, so from now on we consider them defined on $S^{n-1}$.
A star set $L$ is called a {\em star body} if $\rho_L$ is continuous. Conversely, given a positive continuous function $f:S^{n-1}\longrightarrow \mathbb R^+=[0,\infty)$ there exists a star body $L_f$ such that $f$ is the radial function of $L_f$.
We denote by $\mathcal S_0^n$ the set of $n$-dimensional star bodies and we denote by $C(S^{n-1})^+$ the set of positive continuous functions on $S^{n-1}$.
Given two sets $K,L\in \mathcal S^n$, we define their {\em radial sum} as the star set $K\tilde{+}L$ whose radial function is $\rho_K+\rho_L$. Note that $K\tilde{+}L\in \mathcal S_0^n$ whenever $K,L\in \mathcal S_0^n$.
The dual analog for the Hausdorff metric of convex bodies is the so called {\em radial metric}, which is defined by
$$
\delta(K,L)=\inf\{\lambda\geq 0 : K\subset L\tilde{+} \lambda B_n, L\subset K\tilde{+} \lambda B_n\},
$$
where $B_n$ denotes the euclidean unit ball of $\mathbb R^n$. It is easy to check that
$$
\delta(K,L)=\|\rho_K-\rho_L\|_\infty.
$$
\smallskip
An application $V:\mathcal S_0^n\longrightarrow \mathbb R$ is a {\em valuation} if for any $K,L\in\mathcal S_0^n$,
$$
V(K\cup L)+V(K\cap L)=V(K)+ V(L).
$$
It is clear that a linear combination of valuations is a valuation.
\smallskip
Given two functions $f_1, f_2\in C(S^{n-1})^+$, we denote their maximum and minimum by
$$
(f_1\vee f_2)(t)=\max \{f_1(t), f_2(t)\},
$$
$$
(f_1\wedge f_2)(t)=\min \{f_1(t), f_2(t)\}.
$$
Given two star bodies $K,L$, both $K\cup L$ and $K\cap L$ are star bodies, and it is easy to see that
$$
\rho_{K\cup L}=\rho_K\vee \rho_L, \hspace{1cm} \rho_{K\cap L}=\rho_K\wedge \rho_L.
$$
With this notation, a valuation $V:\mathcal S_0^n\rightarrow \mathbb R$ induces a function $\tilde V:C(S^{n-1})^+\rightarrow \mathbb R$ given by
$$
\tilde V(f)=V(L_f),
$$
where $L_f$ is the star body whose radial function satisfies $\rho_{L_f}=f$. If $V$ is continuous, then $\tilde V$ is continuous with respect to the $\|\cdot\|_\infty$ norm in $C(S^{n-1})^+$ and satisfies
$$
\tilde V(f)+\tilde V(g)=\tilde V(f\vee g)+\tilde V(f\wedge g)
$$
for every $f,g\in C(S^{n-1})^+$. Conversely, every such function $\tilde V$ induces a radial continuous valuation on $\mathcal S_0^n$.
Given $A\subset S^{n-1}$, we denote the closure of $A$ by $\overline{A}$.
Given a function $f:S^{n-1} \longrightarrow \mathbb R$, we define the support of $f$ by $$supp(f)=\overline{\{t\in S^{n-1} \mbox{ such that } f(t)\not = 0\}},$$ and for any set $G\subset S^{n-1}$, we will write $f\prec G$ if $\mbox{supp}(f)\subset G$. Throughout, $1\!\!1:S^{n-1}\longrightarrow \mathbb R$ denotes the function constantly equal to 1.
For completeness, we state now a result of \cite{Vi} which will be needed later.
\begin{lema}\label{split}\cite[Lemmata 3.3 and 3.4]{Vi}
Let $\{G_i: i\in I\}$ be a family of open subsets of $S^{n-1}$. Let $G=\cup_{i\in I} G_i$. Then, for every $i\in I$ there exists a function $\varphi_i: G \longrightarrow [0,1]$ continuous in $G$ verifying $\varphi_i\prec G_i$ and such that $\bigvee_{i\in I} \varphi_i=1\!\!1$ in $G$. Moreover, let $f\in C(S^{n-1})^+$ verify $f\prec G$. Then, for every $i\in I$, the function $f_i=\varphi_if$ belongs to $ C(S^{n-1})^+$. Also, $f_i\prec G_i$ and $\bigvee_{i\in I} f_i=f$. In particular, for every $i\in I$, $0\leq f_i\leq f$.
\end{lema}
\section{The results}\label{results}
To prove Theorem \ref{main} we will need to control the maximum value of $V$ on certain sets. The first step in this direction is to show that $V$ is {\em bounded on bounded sets}:
We say that a valuation $V:\mathcal S^n_0\longrightarrow \mathbb R$ is bounded on bounded sets if for every $\lambda>0$ there exists a real number $K>0$ such that, for every star body $L\subset \lambda B_n$, $|V(L)|\leq K.$
Equivalently, $V$ is bounded on bounded sets if for every $\lambda>0$ there exists $K>0$ such that for every $f\in C(S^{n-1})^+$ with $\|f\|_\infty\leq \lambda$ we have $\tilde{V}(f)\leq K$.
\begin{lema}\label{l:bbs}
Every radial continuous valuation $V:\mathcal S^n_0\longrightarrow \mathbb R$ is bounded on bounded sets.
\end{lema}
\begin{proof}
We reason by contradiction. If the result is not true, there exists $\lambda>0$ and a sequence $(f_i)_{i\in \mathbb N}\subset C(S^{n-1})^+$, with $\|f_i\|_\infty\leq \lambda$ for every $i\in \mathbb N$ and such that $|\tilde{V}(f_i)|\rightarrow +\infty$.
Consider the function $$\theta:\mathbb R^+\longrightarrow \mathbb R$$ defined by $$\theta(c)=\tilde{V}(c 1\!\!1).$$ The continuity of $\tilde{V}$ implies that $\theta$ is continuous. Therefore, $\theta$ is uniformly continuous on $[0,\lambda]$. In particular, it is bounded on that interval. Therefore, there exists $M>0$ such that, for every $c\in [0,\lambda]$, $$|\tilde{V}(c 1\!\!1)|\leq M.$$
We define inductively two sequences $(a_j)_{j\in \mathbb N}, (b_j)_{j\in \mathbb N}\subset \mathbb R^+$: Define first $a_0=0$, $b_0=\lambda$. Consider $c_0=\frac{a_0+b_0}{2}$.
We note that $$\tilde{V}(f_i\vee c_01\!\!1) +\tilde{V}(f_i\wedge c_01\!\!1)=\tilde{V}(f_i) + \tilde{V}(c_01\!\!1).$$
Since $|\tilde{V}(c_01\!\!1)|\leq M$ and $|\tilde{V}(f_i)|\rightarrow +\infty$, we know that there must exist an infinite set $\mathbb M_1\subset \mathbb N$ such that for $i\in \mathbb M_1$ either $|\tilde{V}(f_i\vee c_01\!\!1)|\rightarrow +\infty$ or $|\tilde{V}(f_i\wedge c_01\!\!1)|\rightarrow +\infty$ as $i$ grows to $\infty$. In the first case, we set $a_1=c_0$, $b_1=\lambda$ and $f^1_i=f_i\vee c_01\!\!1$. In the second case, we set $a_1=0$ and $b_1=c_0$ and $f^1_i=f_i\wedge c_01\!\!1$. Now we define $c_1=\frac{a_1+ b_1}{2}$ and proceed similarly.
Inductively, we construct two sequences $(a_j), (b_j)\subset \mathbb R^+$, a decreasing sequence of infinite subsets $\mathbb M_j\subset \mathbb N$, and sequences $(f^j_i)_{i\in \mathbb M_j}\subset C(S^{n-1})^+$ such that, for every $j\in\mathbb N$,
$$
|a_j-b_j|=\frac{\lambda}{2^j},
$$
and for every $i\in\mathbb M_j$, for every $t\in S^{n-1}$,
$$
a_j\leq f^j_i(t)\leq b_j,
$$
and with the property that
$$
\lim_{ i\rightarrow \infty} |\tilde{V}(f^j_i)|=+\infty.
$$
Passing to a further subsequence we may assume without loss of generality that, for every $i\in \mathbb N$, $$|\tilde{V}(f_i^i)|\geq i.$$
Call $d=\lim_i a_i$. If we consider now the sequence $(f_i^i)_{i\in \mathbb N}\subset C(S^{n-1})^+$, we have that $$\|f_i^i-d1\!\!1\|_\infty \rightarrow 0$$ but $$|\tilde{V}(f_i^i)|\geq i,$$
in contradiction to the continuity of $\tilde{V}$ at $d1\!\!1$.
\end{proof}
We thank the anonymous referee of \cite{Vi} for suggesting a procedure very similar to this as an alternative reasoning to show a statement in that paper.
\smallskip
In the rest of this note we will repeatedly use the fact that $S^{n-1}$ is a compact metric space. We will write $d$ to denote the euclidean metric in $S^{n-1}$.
We need to define an additional concept for our next result:
Given any set $A\subset S^{n-1}$, and $\omega>0$, let us consider the {\em outer parallel band} around $A$ defined by
$$
A_\omega=\{t\in S^{n-1} : 0<d(t, A)<\omega\}.
$$
Note that, for every $A\subset S^{n-1}$ and $\omega>0$, $A_\omega$ is an open set.
In our next result we use the fact that $V$ is bounded on bounded sets to control $V$ on these bands. In \cite{FK} these outer parallel bands are called {\em rims}, and they are used for similar purposes to ours.
\begin{lema}\label{rims}
Let $V:\mathcal S^n_0\rightarrow \mathbb R$ be a radial continuous valuation. Let $A\subset S^{n-1}$ be any set and $\lambda\in \mathbb R^+$.
$$
\lim_{\omega\rightarrow 0} \sup\{|\tilde{V}(f)|: \, f\prec A_\omega, \, \|f\|_\infty\leq \lambda\}=0.
$$
\end{lema}
\begin{proof}
We reason by contradiction. Suppose the result is not true. Then there exist $A\subset S^{n-1}$, $\lambda \in \mathbb R^+$, $\epsilon>0$, a sequence $(\omega_i)_{i\in \mathbb N}\subset \mathbb R$ and a sequence $(f_i)_{i\in \mathbb N}\subset C(S^{n-1})^+$ such that $\lim_{i\rightarrow \mathbb N} \omega_i=0$ and, for every $i\in \mathbb N$,
\begin{itemize}
\item $\omega_i>0$
\item $f_i\prec A_{\omega_i}$
\item $\|f_i\|_\infty\leq \lambda$
\item $|\tilde{V}(f_i)|\geq \epsilon.$
\end{itemize}
Therefore, there exists an infinite subset $I\subset \mathbb N$ such that either $\tilde{V}(f_i) >\epsilon$ for every $i\in I$ or $\tilde{V}(f_i) <\epsilon$ for every $i\in I$. So, we assume without loss of generality that $\tilde{V}(f_i) >\epsilon$ for every $i\in I$. The case $\tilde{V}(f_i) <\epsilon$ is totally analogous.
Consider $f_1$. Using the continuity of $\tilde{V}$ at $f_1$, we get the existence of $\delta>0$ such that for every $g\in C(S^{n-1})^+$ with $\|f-g\|_\infty<\delta$,
$$
|\tilde{V}(f)-\tilde{V}(g)|\leq \frac{\epsilon}{2}.
$$
Since $f_1$ is uniformly continuous and $f_1(t)=0$ for every $t\in A\subset S^{n-1}\backslash A_{\omega_1}$, there exists $0<\rho<\omega_1$ such that, for every $t\in S^{n-1}$ with $d(t, A)<\rho$, $f_1(t)<\delta$.
We consider the disjoint closed sets
$$
C_1=\{t\in S^{n-1} : d(t, A)\leq \frac{\rho}{2}\}
$$
and
$$
C_2=f_1^{-1}\left([\delta, \lambda]\right).
$$
By Urysohn's Lemma, we can consider a continuous function $\psi_1$ with $\psi_{1|_{C_1}}=0$, $\psi_{1|_{C_2}}=1$ and $0\leq \psi_1(t)\leq 1$ for every $t\in S^{n-1}$. We consider now the function $\psi_1 f_1\in C(S^{n-1})^+$. On the one hand, $\|f_1-\psi_1 f_1\|_\infty\leq \delta$ and, therefore, $$|\tilde{V}(\psi_1 f_1)|\geq \left||\tilde{V}(f_1)|-|\tilde{V}(f_1)-\tilde{V}(\psi_1 f_1)|\right|> \epsilon-\frac{\epsilon}{2}=\frac{\epsilon}{2}.$$
On the other hand, $\psi_1 f_1\prec A_{\omega_1}\backslash A_{\frac{\rho}{2}}$. Now, we can choose $\omega_{i_2}< \frac{\rho}{2}$ and we can reason similarly as above with the function $f_{i_2}$.
Inductively, we construct a sequence of functions $(\psi_j f_{i_j})_{j\in \mathbb N}\subset C(S^{n-1})^+$ with disjoint support such that $\tilde{V}(\psi_j f_{i_j})>\frac{\epsilon}{2}$. Noting that
$$
\tilde{V}\left(\bigvee_j \psi_j f_{i_j}\right)=\sum_j \tilde{V}(\psi_j f_{i_j}),
$$
and that
$$
\|\bigvee_j \psi_j f_{i_j}\|_\infty\leq\lambda,
$$
we get a contradiction with the fact that $V$ is bounded on bounded sets.
\end{proof}
Now we can prove Theorem \ref{main}.
\begin{proof}[Proof of Theorem \ref{main}]
Let $V:\mathcal S_0^n\longrightarrow \mathbb R$ be as in the hypothesis. For every $f\in C(S^{n-1})^+$, we define
$$
\tilde{V}^+(f)=\sup\{\tilde{V}(g): \, 0\leq g \leq f\},
$$
and we consider the application $\tilde{V}:\mathcal S_0^n\longrightarrow \mathbb R$ defined by
$V^+(K)=\tilde{V}^+(\rho_K)$.
Assume for the moment that $V^+$ is a radial continuous valuation. In that case, the result follows easily:
First we note that it follows from $\tilde V(0)=0$ that $V^+(\{0\})=0$ and that, for every $f\in C(S^{n-1})^+$, one has $\tilde V^+(f)\geq 0$. Therefore, $V^+(K)\geq 0$ for every $K\in \mathcal S_0^n$.
We define next $V^-=V^+-V$. Clearly, $V^-$ is a radial continuous valuation and $V^-(\{0\})=0$.
By the definition of $ V^+$, it follows that, for every $K\in \mathcal S_0^n$, one has $V(K)\leq V^+(K)$. Thus, $V^-(K)\geq 0$.
Now, trivially $$V=V^+-V^-.$$
In addition, if $V$ is rotationally invariant, then clearly $V^+$ and $V^-$ are so.
\smallskip
Therefore, we will finish if we show that
$V^+$ is a radial continuous valuation. Let us prove it.
First, we see that it is a valuation. Let $f_1, f_2\in C(S^{n-1})^+$. We have to check that
\begin{equation}\label{igualdad}
\tilde{V}^+(f_1\vee f_2)+ \tilde{V}^+(f_1\wedge f_2) = \tilde{V}^+(f_1)+ \tilde{V}^+(f_2).
\end{equation}
Fix $\epsilon>0$. We choose $0\leq g_1\leq f_1$ such that $\tilde{V}^+(f_1)\leq \tilde{V}(g_1)+\epsilon$, and $0\leq g_2\leq f_2$ such that $\tilde{V}^+(f_2)\leq \tilde{V}(g_2)+\epsilon$.
Then, $$\tilde{V}^+(f_1)+ \tilde{V}^+(f_2)\leq \tilde{V}(g_1)+ \tilde{V}(g_2)+ 2\epsilon = \tilde{V}(g_1\vee g_2) + \tilde{V}(g_1\wedge g_2) + 2\epsilon\leq $$ $$\leq \tilde{V}^+(f_1\vee f_2) + \tilde{V}^+(f_1\wedge f_2) + 2\epsilon, $$
where the last inequality follows from the fact that $0\leq g_1\vee g_2\leq f_1\vee f_2$, $0\leq g_1\wedge g_2\leq f_1\wedge f_2$. Since $\epsilon>0$ was arbitrary, this proves one of the inequalities in \eqref{igualdad}.
For the other one, fix again $\epsilon>0$. We choose $0\leq g \leq f_1\vee f_2 $ such that $\tilde{V}^+(f_1\vee f_2)\leq \tilde{V}(g)+\epsilon$, and $0\leq h \leq f_1\wedge f_2 $ such that $\tilde{V}^+(f_1\wedge f_2)\leq \tilde{V}(h)+\epsilon$.
Let us consider the sets
$$
A=\{ t\in S^{n-1} : f_1(t)\geq f_2(t)\}
$$
and
$$
B=\{ t\in S^{n-1} : f_1(t)< f_2(t)\}.
$$
Let $\lambda=\|f_1\vee f_2\|_\infty$. According to Lemma \ref{rims}, there exists $\omega_1>0$ such that, for every $f\prec A_{\omega_1}$ with $\|f\|_\infty\leq\lambda$ we have $|\tilde{V}(f)|\leq \epsilon$.
Since $\tilde{V}$ is continuous at $g$, there exists $\delta>0$ such that, $|\tilde{V}(g)-\tilde{V}(g')|<\epsilon$ for every $g'$ such that $\|g-g'\|_\infty< \delta$. We define $g'=(g-\frac{\delta}{2})\vee 0$. Then, for every $t\in A$, it follows that
$$
g'(t)=\max\big\{g(t)-\frac{\delta}{2},0\big\}\leq g(t)\leq (f_1\vee f_2)(t)=f_1(t).
$$
Now, we can apply the uniform continuity of $g'$ and $f_1$ to find $\omega_2$ such that for every $t,s\in S^{n-1}$, if $|t-s|<\omega_2$, then $|f_1(t)-f_1(s)|<\delta/4$ and $|g'(t)-g'(s)|<\delta/4$. In particular, this implies that for every $t\in A_{\omega_2}$, $g'(t)\leq f_1(t)$.
Let $\omega=\min\{\omega_1, \omega_2\}$, and let
$$
J(A, \omega)=A\cup A_\omega
$$
be the open $\omega$-outer parallel of the closed set $A$.
Note that $S^{n-1}=J(A, \omega)\cup B$, where both $J(A, \omega)$ and $B$ are open sets. Moreover, we clearly have $J(A, \omega)\cap B=A_\omega$.
We consider the functions $\varphi_1\prec J(A,\omega)$, $\varphi_2\prec B$ associated to the decomposition $S^{n-1}=J(A, \omega)\cup B$ by Lemma \ref{split}. Then $\varphi_1\vee \varphi_2=1\!\!1$. Let us define $g'_1=\varphi_1 g'$, $g'_2=\varphi_2 g'$, $h_1=\varphi_1 h$, $h_2=\varphi_2 h$ as in Lemma \ref{split}.
A simple verification yields \begin{itemize}
\item $g'=g'_1\vee g'_2$, $h=h_1\vee h_2$,
\item $g'_1\wedge g'_2\prec A_\omega$, $h_1\wedge h_2\prec A_\omega$,
\item $g'_1\wedge h_2\prec A_\omega$, $h_1\wedge g'_2\prec A_\omega$,
\item $0\leq g'_1\vee h_2 \leq f_1$,
\item $0\leq g'_2\vee h_1\leq f_2$.
\end{itemize}
Therefore, we get
\begin{eqnarray*}
\tilde{V^+}(f_1\vee f_2)&+& \tilde{V}^+(f_1\wedge f_2) \leq \tilde{V}(g)+ \tilde{V}(h) + 2\epsilon \leq \tilde{V}(g')+ \tilde{V}(h) + 3\epsilon \\
&=&\tilde{V}(g'_1) + \tilde{V}(g'_2) - \tilde{V}(g'_1\wedge g'_2) + \tilde{V}(h_1) + \tilde{V}(h_2) - \tilde{V}(h_1\wedge h_2)+3\epsilon \\
&\leq &\tilde{V}(g'_1) + \tilde{V}(h_2)+ \tilde{V}(g'_2) +\tilde{V}(h_1)+5\epsilon\\
&=&\tilde{V}(g'_1\vee h_2)+ \tilde{V}(g'_1\wedge h_2)+ \tilde{V}(g'_2\vee h_1)+\tilde{V}(g'_2\wedge h_1)+ 5\epsilon\\
&\leq & \tilde{V}(g'_1\vee h_2)+ \tilde{V}(g'_2\vee h_1) + 7\epsilon\leq \tilde{V^+}(f_1)+ \tilde{V}^+(f_2)+7\epsilon.
\end{eqnarray*}
Again, since $\epsilon>0$ was arbitrary, this finishes the proof of \eqref{igualdad}.
\smallskip
Let us see now that $\tilde{V}^+$ is continuous. Let us consider $f_0\in C(S^{n-1})^+$ and take $\epsilon>0$. There exists $g_0\in C(S^{n-1})^+$ with $0\leq g_0 \leq f_0$ such that $\tilde{V}^+(f_0)\leq \tilde{V}(g_0)+\epsilon$.
Since $\tilde{V}$ is continuous at $f_0$ and $g_0$, there exists $\delta>0$ such that for every $f, g \in C(S^{n-1})^+$ with $\|f_0-f\|_\infty<\delta$ and $\|g_0-g\|<\delta$, we have $|\tilde{V}(f_0)-\tilde{V}(f)|<\epsilon$ and $|\tilde{V}(g_0)-\tilde{V}(g)|<\epsilon$.
Let now $f\in C(S^{n-1})^+$ be such that $\|f_0-f\|_\infty<\delta$. Pick $g\in C(S^{n-1})^+$ with $0\leq g \leq f$ such that $\tilde{V}^+(f)\leq \tilde{V}(g) + \epsilon$.
Note that $\|g_0\wedge f - g_0\|<\delta$ and $\|g\vee f_0 - f_0\|<\delta$. Then, we have
$$
\tilde{V}^+(f)\geq \tilde{V}(g_0\wedge f) \geq \tilde{V}(g_0)-\epsilon\geq \tilde{V}^+(f_0)-2\epsilon,
$$
and
\begin{eqnarray*}
\tilde{V}^+(f)&\leq& \tilde{V}(g) + \epsilon =\tilde{V}(g\wedge f_0) + \tilde{V}(g\vee f_0) - \tilde{V}(f_0) +\epsilon \\
&\leq& \tilde{V}(g\wedge f_0) + |\tilde{V}(g\vee f_0) - \tilde{V}(f_0) |+ \epsilon \leq \tilde{V}^+(f_0) + 2\epsilon.
\end{eqnarray*}
Hence, $$|\tilde{V}^+(f_0)-\tilde{V}^+(f)|<2\epsilon$$ and $\tilde{V}^+$ is continuous as claimed.
\end{proof}
\begin{remark}
The same proof shows that every continuous ``valuation'' $\tilde{V}:C(K)\longrightarrow \mathbb R$, where $K$ is a metrizable compact space, can be written as a difference of two positive continuous valuations.
\end{remark}
The proof of Corollary \ref{representacion} is now immediate.
\begin{proof}[Proof of Corollary \ref{representacion}]
Let $V:\mathcal S_0^n\longrightarrow \mathbb R$ be a rotationally invariant radial continuous valuation. We decompose it as $V=V^+-V^-$ as in Theorem \ref{main}. According to \cite[Theorem 1.1]{Vi}, there exist two continuous functions $\theta^+, \theta^-:[0,\infty) \longrightarrow \mathbb R$ such that, for every $K\in \mathcal S_0^n$,
$$V(K)=V^+(K)-V^-(K) =\int_{S^{n-1}} \theta^+(\rho_K(t)) dm(t)- \int_{S^{n-1}} \theta^-(\rho_K(t)) dm(t).$$
We define now $\theta=\theta^+-\theta^-$ and the first part of the result follows.
The converse statement had already been proved in \cite[Theorem 1.1]{Vi} (for that implication, the positivity is not needed).
\end{proof}
|
\section{Introduction}
The Schur multiplier $M(G)$ of a group $G$ was introduced by Schur \cite{IS1} in 1904 on the study of projective representation of groups.
For $p$-groups $G$ of order $p^n$, Green \cite{JG} gave an upper bound $p^{\frac{1}{2}n(n-1)}$ for order of the Schur Multiplier $M(G)$. So we have $|M(G)|=p^{\frac{1}{2}n(n-1)-t(G)}$, for some $t(G) \geq 0$. Now the question comes in our mind that whether it is possible to classify the structure of all $p$-groups $G$ by the order of the Schur multiplier $M(G)$, i.e., when $t(G)$ is known. Several authors have already answered this question. They classified the groups of order $p^n$ for $t(G) \leq 5$ in \cite{BY,ZH,EG,PN3,PN1}. The structure of $p$-groups with $t(G) = 6$ has been determined in \cite{SHJ}. In the present paper, we classify
the structure of all non-abelian finite $p$-groups when $t(G) = 6$, i.e. $|M(G)|=p^{\frac{1}{2}n(n-1)-6}$. Our method is quite different to that of \cite{SHJ}. We have stated some structural results of group $G$ with the assumtion $t(G)=6$.
By $ES_p(p^3)$ and $ES_{p^2}(p^3)$ we denote extra-special $p$-groups of order $p^3$ having exponent $p$ and $p^2$ respectively. By $ES_p(p^5)$ and $ES_{p^2}(p^5)$ we denote extra-special $p$-groups of order $p^5$ having exponent $p$ and $p^2$ respectively. By $\mathbb{Z}_p^{(k)}$ we denote $\mathbb{Z}_p \times \mathbb{Z}_p \cdots \times \mathbb{Z}_p$($k$ times).
James \cite{RJ} classified all $p$-groups of order $p^n$ for $n \leq 6$ upto isoclinism which are denoted by $\Phi_k$. We use his notation throughout this paper.
In this paper we prove the following result.
\begin{thm} (Main Theorem)
Let $G$ be a non-abelian $p$-group of order $p^n$ with $|M(G)|=p^{\frac{1}{2}n(n-1)-6}$.\\
If $p$ is odd, then $G$ is isomorphic to\\
(i) $ES_p(p^3) \times \mathbb{Z}_p^{(5)}$.\\
(ii) $\Phi_2(21^4)a = ES_{p^2}(p^3) \times \mathbb{Z}_p^{(3)}$.\\
(iii) $\Phi_2(21^4)b= \Phi_2(211)b \times \mathbb{Z}_p^{(2)}$, \\
where $\Phi_2(211)b = \langle{\alpha,\alpha_1,\alpha_2,\gamma \mid [\alpha_1,\alpha]=\gamma^p=\alpha_2, \alpha^p=\alpha_1^p=\alpha_2^p=1\rangle}$.\\
(iv) $\Phi_5(21^4)a = ES_{p^2}(p^5) \times \mathbb{Z}_p$.\\
(v) $\Phi_5(1^6)=ES_p(p^5) \times \mathbb{Z}_p$.\\
(vi) $\Phi_5(21^4)b=\langle{\alpha_1,\alpha_2,\alpha_3,\alpha_4,\beta,\gamma \mid [\alpha_1,\alpha_2]=[\alpha_3,\alpha_4]=\gamma^p=\beta,\alpha_i^p=\beta^p=1(i=1,2,3,4)\rangle}$.\\
(vii) $\Phi_4(1^6)=\Phi_4(1^5) \times \mathbb{Z}_p$, \\
where $\Phi_4\left(1^5\right)= \langle{\alpha,\alpha_1,\alpha_2,\beta_1,\beta_2 \mid [\alpha_i,\alpha]=\beta_i, \alpha^p=\alpha_i^p=\beta_i^p=1 (i=1,2)\rangle}$.\\
(viii) $\Phi_2(2111)c=\Phi_2(211)c \times \mathbb{Z}_p$,\\
where $\Phi_2(211)c = \langle{\alpha,\alpha_1,\alpha_2 \mid [\alpha_1,\alpha]=\alpha_2, \alpha^{p^2}=\alpha_1^p=\alpha_2^p=1\rangle}$.\\
(ix) $\Phi_2(2111)d=ES_p(p^3) \times \mathbb{Z}_{p^2}$.\\
(x) $\Phi_3(1^5)=\Phi_3(1^4) \times \mathbb{Z}_p$, \\
where $\Phi_3(1^4) = \langle{\alpha,\alpha_1,\alpha_2,\alpha_3 \mid [\alpha_i,\alpha]=\alpha_{i+1},\alpha^p=\alpha_i^{(p)}=\alpha_3^p=1(i=1,2)\rangle}$.\\
(xi) $\Phi_7(1^5)=\langle{\alpha,\alpha_1,\alpha_2,\alpha_3,\beta \mid [\alpha_i,\alpha]=\alpha_{i+1},[\alpha_1,\beta]=\alpha_3, \alpha^p=\alpha_1^{(p)}=\alpha_{i+1}^p=\beta^p=1 (i=1,2)\rangle}$.\\
(xii) $\Phi_2(31)=\langle{\alpha,\alpha_1,\alpha_2 \mid [\alpha_1,\alpha]=\alpha^{p^2}=\alpha_2, \alpha_1^p=\alpha_2^p=1\rangle}$.
If $p=2$, then $G$ is isomorphic to\\
(xiii) $D_8 \times \mathbb{Z}_2^{(4)}$, \\
(xiv) $\langle a,b,c,d,e \mid [d, c] = [e, b] = a^2, a^4=b^2=c^2=d^2=e^2=1\rangle$,\\
(xv) $\mathbb{Z}_2 \times \langle a,b,c,d \mid [b,c]=[a,d]=b^2,a^2=b^4=c^2=d^2=1\rangle$,\\
(xvi) $\mathbb{Z}_2 \times \langle a,b,c,d \mid b^2=c^2, [b,c]=[a,d]=b^2,a^2=b^4=c^4=d^2=1\rangle$,\\
(xvii) $\mathbb{Z}_2 \times \mathbb{Z}_2 \times X $ where $X=\langle{a,b,c \mid [b,c]=a^2,[a,b]=[a,c]=1,a^4=b^2=c^2=1\rangle}$\\
(xviii) $Q_8 \times \mathbb{Z}_2^{(3)}$, \\
(xix) $\mathbb{Z}_2^{(4)} \rtimes \mathbb{Z}_2$, \\
(xx) $\langle a, b, c \mid [a, c] = [b, c] = 1, (ba)^2 = (ab)^2, ba^2 = a^2b, a^4 = b^2 = c^2 = 1 \rangle$,\\%$\mathbb{Z}_2 \times((\mathbb{Z}_4 \times \mathbb{Z}_2) \rtimes \mathbb{Z}_2)$ \\
(xxi) $QD_{16}$, \\
(xxii) $Q_{16}$, \\
(xxiii) $\langle{a,b \mid [a,b]=a^4,a^8=b^2=1\rangle}$,\\
(xxiv) $\langle a,b,c \mid [a, c] = a^2, [b, c] = b^2,a^4 = b^4=c^2 = 1 \rangle$,\\
where $D_n$ denotes Dihedral group of order $n$, $QD_n$ denotes QuasiDihedral groups of order n and $Q_n$ denotes Quaternion group of order $n$.
\end{thm}
\section{Preliminaries}
In this section we list following results which are used in our proof.
\begin{thm}(see \cite[Theorem 4.1]{MRRR})\label{J}
Let $G$ be a finite group and $K$ a central subgroup. Set $A = G/K$. Then
$|M(G)||G'\cap K|$ divides $|M(A)| |M(K)| |A \otimes K|$
\end{thm}
\begin{prop}(see \cite[Proposition 2.4]{MRR})\label{J1}
Let $G$ be a finite nilpotent group of class $c \geq 2$. Then
$|\gamma_c(G)||M(G)| \leq |M(G/\gamma_c(G))||G/Z_{c-1}(G) \otimes \gamma_c(G)|$
\end{prop}
\begin{thm}(see \cite[Corollary 4.16]{BT})\label{SHH}
Let $G$ be an extra special p-group of order $p^{2n+1}$.\\
(i) If $n \geq 2$, then $|M(G)| = p^{2n^2-n-1}.$\\
(ii) If $n = 1$, then the Schur multiplier of $D_8 , Q_8 , ES_p(p^3)$ and $ES_{p^2}(p^3)$ are isomorphic to $\mathbb{Z}_2, 1, \mathbb{Z}_p \times \mathbb{Z}_p$ and $1$ respectively.
\end{thm}
The following result follows from \cite{KO} for $|G'|=p$ and from \cite[page. 4177]{EG} for $|G'|=p^2$.
\begin{thm}\label{SHHH}
For non-abelian $p$-groups $G$ of order $p^4$ with $|G'|=p$,
$M(\Phi_2(211)a) \cong \mathbb{Z}_p \times \mathbb{Z}_p$, $M(\Phi_2(1^4)) \cong \mathbb{Z}_p^{(4)}$, $M(\phi_2(31)) \cong {1}$, $M(\Phi_2(22)) \cong \mathbb{Z}_p$, $M(\Phi_2(211)b) \cong \mathbb{Z}_p \times \mathbb{Z}_p$, $M(\Phi_2(211)c) \cong \mathbb{Z}_p \times \mathbb{Z}_p$ \\
and For $|G'|=p^2$,\\
$M(\Phi_3(211)a) \cong \mathbb{Z}_p$, $M(\Phi_3(211)b_r) \cong \mathbb{Z}_p$, $M(\Phi_3(1^4)) \cong \mathbb{Z}_p \times \mathbb{Z}_p$
\end{thm}
The following three lemmas follow from \cite[Main theorem]{PN}.
\begin{lemma}\label{5}
Let $G$ be a non-abelian $p$-group of order $p^n$ with $|M(G)|=p^{\frac{1}{2}n(n-1)-6}$. Then $n \leq 8$.
\end{lemma}
\begin{lemma}\label{6}
Let $G$ be a non-abelian $p$-group of order $p^8$ with $|M(G)|=p^{\frac{1}{2}n(n-1)-6}$. Then $G \cong ES_p(p^3) \times \mathbb{Z}_p^{(5)}$.
\end{lemma}
\begin{lemma}\label{m3}
There is no $p$-group $G$ of order $p^n(n \geq 6)$ with $|G'| \geq p^3$ and $|M(G)|=p^{\frac{1}{2}n(n-1)-6}$.
\end{lemma}
\begin{lemma}\label{7}
There is no non-abelian $p$-group $G$ of order $p^7$ with $|M(G)|=p^{\frac{1}{2}7(7-1)-6}=p^{15}$.
\end{lemma}
\begin{proof}
It follows from \cite[Theorem 21]{PN2}. \hspace{5.4cm} $\hfill\square$
\end{proof}
\begin{lemma}\label{8}
Let $G$ be a non-abelian $p$-group of order $p^4$ with $|M(G)|=p^{\frac{1}{2}4(4-1)-6}=1$. Then $G \cong \Phi_2(31)$.
\end{lemma}
\begin{proof}
This result follows from Theorem \ref{SHHH}. \hspace{5 cm} $\hfill\square$
\end{proof}
Hence from the discussion above it is clear that we have to study groups of order $p^5$ and $p^6$.
\section{Groups of order $p^5$}
In this section we characterize groups of order $p^5$ which have Schur multiplier of order $p^4$.
\begin{lemma}
There is no non-abelian $p$-group of order $p^5$ with $|G'|= p^3$ and $|M(G)|=p^{\frac{1}{2}5(5-1)-6}=p^4$.
\end{lemma}
\begin{proof}
Notice that nilpotency class of $G$ is either $3$ or $4$.
Let $G$ be of class $4$. Then $G$ lie in the isoclinism classes $\Phi_9$ or $\Phi_{10}$. By Proposition \ref{J1} and Theorem \ref{SHHH} we have $|M(G)| \leq p^3$.
Now assume that the nilpotency class of $G$ is 3. Then it follows that $G$ lie in $\Phi_6$ in \cite{RJ}. In this case take any subgroup $K \subset Z(G) \cap G'$ of order $p$. By Theorem \ref{J} we have $|M(G)|p \leq |M(G/K)|p^2$. Here $G/K$ is of order $p^4$ with $|(G/K)'|=p^2$.
Hence it follows from Theorem \ref{SHHH} that $|M(G)| \leq p^3$. This concludes the proof. $\hfill \square$
\end{proof}
Before we proceed to the next result, we explain a method by Blackburn and Evens \cite{BE} for computing Schur multiplier of $p$-groups of class $2$ with $G/G'$ is elementary abelian.
Here $G/G'$ and $G'$ are elementary abelian of order $p^3$ and $p^2$ respectively. We can consider $G/G'$ and $G'$ as vector spaces over $GF(p)$, denote by $V, W$ respectively. Let $v_1,v_2 \in V$ such that $v_i=g_iG', i \in \{1,2\}$ and take $(v_1,v_2)=[g_1,g_2]$.
Let $X_1$ be the subspace of $V \otimes W$ spanned by all \\
\centerline{$v_1 \otimes (v_2,v_3) + v_2 \otimes (v_3,v_1) + v_3 \otimes (v_1,v_2)$}
Consider a map $f:V \rightarrow W$ given by $f(gG')=g^p$, $g \in G$. We denote by $X_2$ the subspace spanned by all $v \otimes f(v)$ for $v \in V$ and take $X=X_1+X_2$. Now consider a homomorpism $\sigma: V\wedge V \rightarrow V \otimes W/X$ given by\\
$\sigma(v_1 \wedge v_2)=v_1 \otimes f(v_2)+ {p \choose 2}v_2 \otimes (v_1,v_2)+X$.
Then there exists an abelian group $M^*$ having a subgroup $N$ for which \\
\centerline{$1 \rightarrow V\otimes W/X \rightarrow M^* \xrightarrow{\xi} V \wedge V \rightarrow 1$}\\
is exact, where $N \cong V \otimes W/X, M^*/N \cong V \wedge V$ and $(\sigma\xi)(m)=m^p$ for all $m \in M^*$.
\begin{thm}\label{B}\cite{BE}
With the notation above, consider a homomorphism $\rho:V \wedge V \rightarrow W$ given by\\
\centerline{$\rho(v_1 \wedge v_2)=(v_1,v_2)$ for all $v_1,v_2 \in V$.}\\
Denote by $M$, the subgroup of $M^*$ containg $N$ for which $M/N$ corresponds to $ker \rho$. Then $M(G) \cong M$
\end{thm}
\begin{lemma}\label{2}
Let $G$ be a non-abelian $p$-group of order $p^5$ with $|G'|=p^2$ and $|M(G)|=p^{\frac{1}{2}5(5-1)-6}=p^4$. Then $G \cong \Phi_3(1^5), \Phi_7(1^5)$.
\end{lemma}
\begin{proof}
Groups of order $p^5$ with $|G'|=p^2$ lie in the isoclinism classes $\Phi_3, \Phi_4, \Phi_7$ or $\Phi_8$ \cite{RJ}.
In isoclinism class $\Phi_3$, some groups are direct product of its subgroups. So for them we can easily compute $M(G)$ and we see $|M(\Phi_3(1^5))|=p^4$. Now we consider other groups. We have $|M(G)|p \leq |M(G/\gamma_3(G))|p^2$ using Proposition \ref{J1}. Observe that $|M(G/\gamma_3(G))| \leq p^2$ except $\Phi_3(2111)c$. Therefore $|M(G)| \leq p^3$ except $\Phi_3(2111)c$. For $G \cong \Phi_3(2111)c$ by Theorem \ref{J} (taking $K=Z(G)$) it follows that $|M(G)| \leq p^3$.
In class $\Phi_4$, every group $G$ has nilpotency class 2 with $G/G'$ elementary abelian. By Theorem \ref{B} $|M(G)|/|N|=|V \wedge V|/|W|$ and thus one can show that $G=\Phi_4(1^5)$ has $|M(G)|=p^6$ but for all other groups $G$ in $\Phi_4$, $|M(G)| \leq p^3$.
In class $\Phi_7$, for the groups $G \cong \Phi_7(2111)a, \Phi_7(2111)b_r,\Phi_7(2111)c$ we can choose a normal subgroup $K$ such that $G/K$ is cyclic with $|M(K)|=p, |K'|=p^2$. Hence $|M(G)| \leq p^3$ by \cite[Theorem 3.1]{MRR}.
Now by \cite[Theorem 2.3.10]{GK} we have the following sequence is exact.\\
\centerline{$\operatorname{Hom} (Z, \mathbb{C}^*) \xrightarrow{Tra} M(G/Z) \xrightarrow{inf} M(G) \xrightarrow{\delta} G/G' \otimes Z$}\\
In particular for $G \cong \Phi_7(1^5)$, taking $Z=Z(G)$, $Im(Tra) \cong ker(Inf) \cong G' \cap Z \cong \mathbb{Z}_p$. We can see in \cite{RJ}, $G$ is capable as $E/Z(E) \cong G$ for groups $E$ of order $p^6$ in isoclinism class $\Phi_{30}$. Thus it follows from \cite[Corollary 2.5.8 and 2.5.10]{GK} that $\delta$ is not trivial map. So $|ker(\delta)|=|Im(Inf)|=\frac{|M(G/Z)|}{p}=p^3 < |M(G)|$. So $p^4 \leq |M(G)|$ and by \cite[Theorem 3.1]{MRR} (taking $K=<\alpha,\alpha_1, \alpha_2,\alpha_3>$) $|M(G)| \leq p^4$.\\
Hence we conclude $|M(\Phi_7(1^5))|=p^4$.
Finally consider the class $\Phi_8$, consisting of only one group $\Phi_8(32)$. Observe that $|M(\Phi_8(32))| \leq p^3$ by Theorem \ref{J} (taking $K=Z(G)$). \hspace{4cm} $\hfill\square$
\end{proof}
\begin{lemma}\label{1}
Let $G$ be a non-abelian $p$-group of order $p^5$ with $|G'|=p$ and $|M(G)|=p^{\frac{1}{2}5(5-1)-6}=p^4$. Then $G \cong \Phi_2(2111)c, \Phi_2(2111)d$.
\end{lemma}
\begin{proof}
Groups of order $p^5$ with $|G'|=p$ lie in the isoclinism classes $\Phi_2$ and $\Phi_5$ \cite{RJ}.\\ The isoclinism class $\Phi_5$ consists of extra-special $p$-groups which have Schur multiplier of order $p^5$ by Theorem \ref{SHH}. Now we consider the class $\Phi_2$. Since $|M(G)|=p^4$, by Theorem \ref{J} we have $G/G' \cong \mathbb{Z}_{p^2} \times \mathbb{Z}_p \times \mathbb{Z}_p$ or $\mathbb{Z}_p^{(4)}$. Now it follows that $G$ is either isomorphic to $\Phi_2(311)b, \Phi_2(221)c$ or $G$ is direct product of its subgroups.
If $G \cong \Phi_2(311)b, \Phi_2(221)c$, then by Theorem \ref{J} $|M(G)| \leq p^3$ for suitable $K$. If $G$ is direct product of its subgroups, then $|M(G)|=p^4$ only for $\Phi_2(2111)c, \Phi_2(2111)d$. $\hfill\square$
\end{proof}
\section{Groups of order $p^6$ and Proof of Main theorem}
In this section we characterize groups of order $p^6$ having Schur multiplier of order $p^9$ and prove the Main Theorem.
\begin{lemma}
Let $G$ be a non-abelian $p$-group of order $p^6$ with $|M(G)|=p^{\frac{1}{2}6(6-1)-6}=p^9$. Then $G/G'$ is elementary abelian.
\end{lemma}
\begin{proof}
From Lemma \ref{m3} we have $|G'| \leq p^2$. It follows from \cite{RJ} that if $|G'|=p$, then $G/Z(G)$ is isomorphic to $\mathbb{Z}_p \times \mathbb{Z}_p$ or $\mathbb{Z}_p^{(4)}$. If $G' \cong \mathbb{Z}_p \times \mathbb{Z}_p$ then $G/Z(G)$ is isomorphic to $ES_p(p^3), \mathbb{Z}_p^{(3)}, ES_p(p^3) \times \mathbb{Z}_p$, $ES_p(p^3) \times \mathbb{Z}_p^{(2)}$ or $\mathbb{Z}_p^{(4)}$. If $G' \cong \mathbb{Z}_{p^2}$, then again by \cite{RJ} $G/Z(G)$ is isomorphic to $\Phi_2(22)$ or $\mathbb{Z}_{p^2} \times \mathbb{Z}_{p^2}$.
In all the above cases using \cite[Proposition 1]{EW} we conclude that if $G/G'$ is not elementary abelian, then $|M(G)| < p^9$, which is not our case. Hence the result follows. \hspace{4 cm} $\hfill\square$
\end{proof}
\begin{lemma}\label{3}
Let $G$ be a non-abelian $p$-group of order $p^6$ with $|G'|=p$ and $|M(G)|=p^9$, then $G$ is isomorphic to one of the following: $\Phi_2(21^4)a, \Phi_2(21^4)b, $ $\Phi_5(21^4)a, \Phi_5(1^6), \Phi_5(21^4)b$.
\end{lemma}
\begin{proof}
It follows from \cite{RJ} that $G$ lie in the isoclinism classes $\Phi_2$ or $\Phi_5$. By the preceeding lemma we have $G/G'$ is elementary abelian of order $p^5$. So in the isoclinism class $\Phi_2$ we have to check $\Phi_2(21^4)a, \Phi_2(21^4)b, \Phi_2(21^4)c, \Phi_2(21^4)d, \Phi_2(1^6)$ and in the isoclinism class $\Phi_5$ we have to check $\Phi_5(21^4)a, \Phi_5(1^6), \Phi_5(21^4)b$.
If $G \cong \Phi_5(21^4)b$, then by Theorem \ref{J} (taking $K=Z(G)$) we have $|M(G)| \leq p^9$ and by \cite[Corollary 3.2]{MRRR} $|M(G)| \geq p^9$. Hence $\Phi_5(21^4)b$ has Schur multiplier of order $p^9$.
All other above groups are direct product of its subgroups. So it is easy to see that among them $\Phi_2(21^4)a, \Phi_2(21^4)b, \Phi_5(21^4)a, \Phi_5(1^6)$ have Schur multiplier of order $p^9$. \hspace{10.3 cm}$\hfill\square$
\end{proof}
\begin{lemma}
There is no non-abelian $p$-group of order $p^6$ with $G'\cong \mathbb{Z}_{p^2}$ and $|M(G)|=p^9$.
\end{lemma}
\begin{proof}
From \cite{RJ} it suffices to consider isoclinism classes $\Phi_8, \Phi_{14}$. Now using Theorem \ref{J} (taking $K=Z(G)$) we can easily show that $|M(G)| < p^9$. \hspace{2 cm}$\hfill\square$
\end{proof}
\begin{lemma}
Let $G$ be a non-abelian $p$-group of order $p^6$ with $G' \cong \mathbb{Z}_p \times \mathbb{Z}_p $ and $|M(G)|=p^9$. Then $Z(G)$ is of exponent $p$ and $G'\subseteq Z(G)$.
\end{lemma}
\begin{proof}
Let the exponent of $Z(G)$ be $p^k(k \geq 2)$ and $K$ be a cyclic central subgroup of order $p^k$. Then using Theorem \ref{J} and \cite{PN}, we have\\
\centerline{$|M(G)| \leq p^{-1}|M(G/K)||G/K \otimes K| \leq p^{-1}p^{\frac{1}{2}(n-3)(n-4)+1}p^{(n-3)}$}. \\
When $n=6$ it gives $|M(G)| < p^9$, which is a contradiction. Hence $Z(G)$ is of exponent $p$.
Now assume that $G' \nsubseteq Z(G)$. Then it follows that $|Z(G)| \leq p^3$.
If $|Z(G)|=p$, then $G$ is of nilpotency class 3. Now by Proposition \ref{J1}, we get $|M(G)| < p^9$. If $|Z(G)|=p^2$, then by the assumption there is a central subgroup $K$ of order $p$ such that $G' \cap K=1$ and $(G/K)'=p^2$. By Theorem \ref{J}, $|M(G)| \leq |M(G/K)|p^3$. This is possible only when $|M(G/K)| \geq p^6$. Now by Theorem \cite{PN, PN3} we see that there is no such $G/K$ of order $p^5$ such that $|M(G/K)| \geq p^6$ with $(G/K)'=p^2$ and $Z(G/K)=p$. Finally if $|Z(G)|=p^3$, then consider a central subgroup of order $p^2$ such that $G' \cap K=1$. Then $|M(G)| \leq |M(G/K)|p^4$ by Theorem \ref{J}. Therefore it follows from Theorem \ref{SHHH} $|M(G)| < p^9$. \hspace{5cm} $\hfill\square$
\end{proof}
\begin{lemma}\label{4}
Let $G$ be a non-abelian $p$-group of order $p^6$ with $G'\cong \mathbb{Z}_p \times \mathbb{Z}_p$ and $|M(G)|=p^{\frac{1}{2}6(6-1)-6}=p^9$. Then $G$ is isomorphic to $\Phi_4(1^6)$.
\end{lemma}
\begin{proof}
By preceeding lemma, $G' \subseteq Z(G)$.
Now consider a central subgroup $K$ in $G'$ of order $p$. By Theorem \ref{J} we have $|M(G)| p \leq |M(G/K)|p^4$. This is possible only when $|M(G/K)| \geq p^6$. Since $(G/K)' \cong \mathbb{Z}_p$ so $G/K \cong ES_p(p^3) \times \mathbb{Z}_p^{(2)}$ by \cite{PN}.
This tells that $G$ is of exponent $p$. So we have to study the groups of exponent $p$ in the isoclinism classes $\Phi_3, \Phi_4, \Phi_7, \Phi_{22}$ which are $\Phi_3(1^6), \Phi_4(1^6), \Phi_7(1^6), \Phi_{22}(1^6)$.
If $G \cong \Phi_{22}(1^6)$, then by \cite[Theorem 3.1]{MRR} (taking $K=<\alpha,\alpha_1,\alpha_2, \alpha_3,\beta_1>$) we have $|M(G)| < p^9$. Other groups are direct product of its subgroups. Hence it is easy to see that only $\Phi_4(1^6)$ has Schur multiplier of order $p^9$. \hspace{3 cm}$\hfill\square$
\end{proof}
We are now ready to prove our Main Theorem.
\subsection{Proof of Main Theorem}
Let $G$ be a group of order $p^n$ ($p$ odd) with $|M(G)|=p^{\frac{1}{2}n(n-1)-6}$. By Lemma \ref{5}, it follows that $n \leq 8$. If $|G|=p^8$, then the assertion $(i)$ of the main theorem follows from Lemma \ref{6}. By Lemma \ref{7} it follows that there is no non-abelian $p$-group of order $p^7$ with $|M(G)|=p^{\frac{1}{2}7(7-1)-6}=p^{15}$. So the problem reduces to studying groups of order $p^4$, $p^5$ and $p^6$.
If $|G|=p^4$, then the assertion $(xii)$ follows from Lemma \ref{8}.
Now we assume $|G|=p^5$. Then assertions $(viii), (ix)$ follows from Lemma \ref{1} and $(x), (xi)$ follows from Lemma \ref{2}. Now consider groups of order $p^6$. Then the assertions $(ii), (iii), (iv), (v), (vi)$ follow from Lemma \ref{3} and $(vii)$ follows from \ref{4}.
For the case $p=2$, the classification follows from computation using HAP package \cite{HAP} of GAP \cite{GAP}. $\hfill\square$
\\
\\
{\bf Acknowledgement}:
I am grateful to my supervisor Manoj K. Yadav for his guidance, motivation and discussions. I wish to thank the Harish-Chandra Research Institute, the Dept. of Atomic Energy, Govt. of India, for providing excellent research facility.
|
\section{Introduction}
\label{sec:intro}
The nearest neighbor (NN) relations and their properties have been extensively studied in various fields,
such as probability and statistics (\cite{bickel:1983}), computer science (\cite{yao:1997},
and ecology (\cite{clark:1954}).
Based on the NN relations,
NN graphs and digraphs are constructed and related graph quantities/invariants are widely studied
(\cite{yao:1997}, \cite{penrose:2001}, and \cite{kozakova:2006}).
We consider NN digraphs and quantities based on their arcs (i.e., directed edges).
In a NN digraph, $D=(V,A)$, the vertices are data points in $\mathbb{R}^d$,
and there is an arc from vertex $u$ to vertex $v$ (i.e., $(u,v) \in A$) if $v$ is a NN of $u$.
We call a pair of vertices $u,v$ as a \emph{reflexive NN pair},
if $v$ is a NN of $u$ and vice versa (i.e.,$ \{ (u,v),(v,u)\} \subset A$) (\cite{clark:1955} and \cite{cox:1981}).
If both $(u,w)$ and $(v,w)$ are in $A$ for some $w\in V$ (i.e., $u$ and $v$ share a NN), then $(u,w)\& (v,w)$ are called \emph{shared NNs}.
Notice that altough $w$ is the shared vertex, arcs $(u,w)$ and $(v,w)$ are called shared NNs in literature (see, \cite{dixon:1994}).
The vertices in a reflexive NN pair are also called isolated NNs (\cite{pickard:1982}),
mutual NNs (\cite{schiling:1986}) or biroot (\cite{yao:1997}).
The NN digraph is also referred as the NN graph in literature (e.g., \cite{yao:1997});
but, since the NN relation is not symmetric,
we opt to use ``NN digraph" which reflects this asymmetry.
Also, the underlying graph of a NN digraph
(an underlying graph of a digraph is obtained by replacing each arc with an (undirected) edge,
disallowing multiple edges between two vertices (\cite{chartrand:1996})) is sometimes referred to as the NN graph
(see, e.g., \cite{friedman:1983}, \cite{penrose:2001}).
Since in any (undirected) graph,
the relation defining the edges is symmetric (i.e., each edge is symmetric),
reflexivity is not an interesting property for undirected graphs.
Number of reflexive and shared NN pairs in a NN digraph is of importance in various fields.
For example,
in spatial data analysis,
the distributions of the tests based on nearest neighbor contingency tables depend on these two quantities (\cite{dixon:1994} and \cite{ceyhan:cell2008}),
when the underlying pattern of the points is from a spatial distribution
(e.g., from homogeneous Poisson process (HPP) or a binomial process).
Moreover, neighbor sharing type quantities such as $Q_n$ are also of interest for the problem of estimating the intrinsic dimension of a data set (see, \cite{brito:2013}).
In our analysis,
we consider the special case of $d=1$ (i.e., one dimensional data),
and study the case when the random sample of size $n$ is obtained from uniform distribution over an interval.
We denote the total number of reflexive and shared NN pairs in the corresponding sample as $R_n$ and $Q_n$, respectively.
The quantity $R_n$ could be of interest for inferential purposes as well,
since it is a measure of mutual (symmetric) spatial dependence between points,
which might indicate a special and/or stronger form of clustering of data points.
For instance, a simple test based on the proportion of the number of
reflexive pairs to the sample size was presented by \cite{dacey:1960}
to interpret the degree of regularity or clustering of the locations of towns alongside a river.
However, the methodology of \cite{dacey:1960} ignores the randomness (and hence uncertainty)
in the value of $R_n$ and hence is not reliable.
The exact distribution of $R_n$ can be computed for finite values of $n$ and hence,
would make possible the use of $R_n$ in exact inference
for testing such one-dimensional clustering.
NN relations, such as reflexivity and neighbor sharing, are studied by many authors.
\cite{enns:1999} provide $\E(R_n)=n/3$ for $n\geq 3$, $\Var (R_n)=2n/45$ for $n\geq 5$ and a recurrence relation giving
the exact pmf of $R_n$ for finite $n$,
whereas the results in \cite{schiling:1986} yield
$\E(Q_n)/ n \rightarrow 1/4$ as $n \rightarrow \infty$.
For the number of reflexive pairs, we approach to the problem in the same way as \cite{enns:1999},
but we drive the mean, the variance and the recurrence relation by a different approach.
Further, we obtain mean and variance of $Q_n$ and compute the asymptotic distribution of both $R_n$ and $Q_n$,
which are novel contributions of this article to the literature.
We provide preliminary results in Section \ref{sec:prelim} where
we convert our problems into random permutations by using interchangeability of uniform spacings.
We derive means and variances of $R_n$ and $Q_n$ together with a recurrence relation giving
the exact pmf of $R_n$ in Section \ref{sec:mean-var}.
The asymptotic results (such as SLLN and CLT) for $R_n$ and $Q_n$ are presented in Section \ref{sec:asy-res},
and discussion and conclusions are provided in Section \ref{sec:disc-conc}.
\section{Preliminaries}
\label{sec:prelim}
The number of reflexive pairs and the number of shared neighbors in the data is invariant under translation and scaling,
since both depend only on the ordering of the pairwise distances of the data points.
Therefore, without loss of generality, we may only consider the uniform distribution over the interval $(0,1)$ (denoted $U(0,1)$).
A NN of a point is one of the ``closest'' points
with respect to some distance or dissimilarity measure.
We will employ the usual Euclidean distance in our analysis.
Observe that under uniform distribution,
the Lebesgue measure of the set of points which have more than one NN
is zero and therefore we may assume that each point has a unique NN with probability 1.
In a sample of size $n$ from $U(0,1)$,
recall that a pair of points is called \emph{reflexive}, if each one is the NN of the other,
and we denote the total number of reflexive pairs as $R_n$, and
a pair of points in the sample is called \emph{shared NN}, if they have the same NN (sharing the NN) and we denote the total number of shared NNs as $Q_n$.
Let $\{ U_1,U_2,\dots , U_n \}$ be a random sample of size $n$ from the uniform distribution $U(0,1)$.
On the real line,
there is a nice ordering structure for the data which we exploit in our results.
Let $ U_{(1)}, U_{(2)},\dots , U_{(n)}$ be the order statistics of $\{ U_1,U_2,\dots , U_n \}$.
Denote the spacings between the order statistics as $D_i:=U_{(i+1)}-U_{(i)}$ for $1\leq i \leq n-1$ with $D_0:=U_{(1)}$.
\begin{lemma}\label{lem:obsrn}
For $n\geq 3$
$$R_n={\bf 1}_{ \{D_1<D_2 \} }+\sum_{i=2}^{n-2} {\bf 1}_{ \{ D_{i}<\min\{ D_{i+1}, D_{i-1}\} \} }+{\bf 1}_{ \{ D_{n-1}<D_{n-2} \} },$$
and for $n\geq 4$
$$Q_n={\bf 1}_{ \{D_2<D_3 \} }+\sum_{i=2}^{n-3} {\bf 1}_{ \{ D_{i}< D_{i-1}, D_{i+1}<D_{i+2} \} }+{\bf 1}_{ \{ D_{n-2}<D_{n-3} \} },$$
where ${\bf 1}_{A}$ is the indicator for the event $A$.
\end{lemma}
\begin{proof}
First observe that the NNs of $U_{(1)}$ and $U_{(n)}$ are always $U_{(2)}$ and $U_{(n-1)}$, respectively.
Therefore,
$\{ U_{(1)}, U_{(2)} \}$ is a reflexive pair if and only if $D_1<D_2$ and, similarly, $\{ U_{(n-1)}, U_{(n)} \}$ is reflexive if and only if $D_{n-1}<D_{n-2}$.
Also note that, for each $2\leq i \leq n-1$, the NN of $U_{(i)}$ is either $U_{(i-1)}$ (if $D_{i-1}<D_i$) or $U_{(i+1)}$ (if $D_i<D_{i-1}$).
Thus, for $2\leq i,j\leq n-1$,
the pair $\{U_{(i)}, U_{(j)}\}$ with $i<j$ is reflexive if and only if $j=i+1$ and $D_{i}$ is less than both $D_{i-1}$ and $D_{i+1}$.
So, we obtain the first identity in the Lemma \ref{lem:obsrn}.
For the representation of $Q_n$, in a similar manner,
one can easily see that $U_{(i)}$ and $U_{(j)}$ ($i<j$) have the same NN only if $j=i+2$ and the common NN is $U_{(i+1)}$,
and obtain the desired result.
\end{proof}
As the quantities $R_n$ and $Q_n$ depend on the ordering of the spacings,
we focus on the distribution of the spacings.
By elementary probability arguments (e.g., Jacobian density theorem) it follows that the joint density of the spacings $(D_0,D_1,\dots , D_{n-1})$ is
\begin{align}\label{eq:unispdens}
f_S(d_0,d_1,\dots, d_{n-1})=n! {\bf 1}_{ \{d_0+d_1+\cdots +d_{n-1}<1\}} {\bf 1}_{ \{ \min \{d_0,d_1, \dots ,d_{n-1} \}>0\} }
\end{align}
with the understanding that $\{d_0,d_1, \dots, d_{n-1}>0\}=\{d_0>0, d_1>0, \dots, d_{n-1}>0\}$.
By \eqref{eq:unispdens} it is clear that the spacings $D_1,\dots ,D_{n-1}$ are interchangeable and hence
$P(D_{\sigma (1)}< \cdots < D_{\sigma (n-1)})=P(D_1<\cdots <D_{n-1})$ for any permutation $\sigma$ in $P_{n-1}$,
where $P_{n-1}$ is the permutation group on $\{1,2,\dots , n-1\}$.
In other words, every ordering of the spacings $D_1,\dots ,D_{n-1}$ is equally likely to occur.
Let $\sigma$ be chosen uniformly at random from $P_{n-1}$.
Define the events $A_1=\{\sigma(1)<\sigma(2)\}$, $A_{n-1}=\{\sigma(n-1)<\sigma(n-2)\}$,
$A_i=\{\sigma(i)<\sigma(i-1), \sigma(i)< \sigma(i+1) \}$ for all $2\leq i \leq n-2$,
and the events $B_1={\{\s(2)<\s(3)\}}$, $B_{n-2}={\{\s(n-2)<\s(n-3)\}}$ and
$B_i={\{\s(i)<\s(i-1), \s(i+1)<\s(i+2)\}}$ for each $2\leq i \leq n-3$.
Then, by Lemma \ref{lem:obsrn} and the interchangeability of the spacings we have
\begin{align}
R_n\stackrel{d}{=}\sum_{i=1}^{n-1} {\bf 1}_{A_i} \text{ and } Q_n\stackrel{d}{=}\sum_{i=1}^{n-2} {\bf 1}_{B_i}, \label{eq:equiv}
\end{align}
where $\stackrel{d}{=}$ denotes equality in distribution.
Therefore, throughout of this paper, we consider $\sum_{i=1}^{n-1} {\bf 1}_{A_i}$ and $\sum_{i=1}^{n-2} {\bf 1}_{B_i}$
for the probabilistic results for $R_n$ and $Q_n$, respectively.
\section{Some Probabilistic Results for $R_n$ and $Q_n$}
\label{sec:mean-var}
In this section, we derive the means and variances of $R_n$ and $Q_n$, and present a recurrence relation for the exact distribution of $R_n$.
\subsection{Mean and Variance of $R_n$}
In a digraph $D$, a \emph{weakly connected component} is a maximal subdigraph of $D$ in which there is a path from every vertex to every other vertex in the underlying graph of $D$.
\cite{enns:1999} call a weakly connected component of a digraph as \emph{society} and examine the number of societies in a uniform data of size $n$ in one dimension.
By the simple observation that each society contains exactly one reflexive pair,
they convert the problem into the number of reflexive pairs and focus on the ranking of the spacings.
Considering the spacing with the largest length, they derive a recurrence relation and obtain
$\E(R_n)=n/3$ for $n\geq 3$ and $\Var (R_n)=2n/45$ for $n\geq 5$ by using generating functions.
We verify their results by following the idea in \cite{romik:2011}.
We obtain the mean and variance of $R_n$ by computing those of $\sum_{i=1}^{n-1} {\bf 1}_{A_i}$.
The random variable $\sum_{i=1}^{n-1} {\bf 1}_{A_i}$ is closely related to the length of the longest alternating subsequence in a random permutation
(see, e.g., \cite{romik:2011}, \cite{houdre:2010}).
For a sequence of pairwise distinct real numbers $x_1,\dots ,x_n$, a subsequence $x_{i_1}, \dots , x_{i_k}$ with
$1\leq i_1< \cdots <i_k \leq n$ is called \emph{alternating} if it satisfies
\begin{align*}
x_{i_1}>x_{i_2} < x_{i_3} > \cdots x_{i_k}.
\end{align*}
Note that there may be more than one alternating subsequence with the maximal length.
For instance, the sequence $6,4,1,3,5,2$ has seven longest alternating subsequences, particularly
$(6,1,3,2)$, $(6,1,5,2)$, $(6,4,5,2)$, $(6,3,5,2)$, $(4,1,3,2)$, $(4,1,5,2)$ and $(4,3,5,2)$.
Let the random variable $L_n^{as}$ be the maximal length of an alternating subsequence of $\tau(1),\dots , \tau(n)$, where $\tau$ is a uniformly random permutation from $P_n$.
Also, for $2\leq k \leq n-1$, $x_k$ is called \emph{local minimum} (resp. \emph{local maximum}) if $x_k<\min \{x_{k-1}, x_{k+1} \}$
(resp. $x_k>\max \{x_{k-1}, x_{k+1} \}$) (\cite{romik:2011}).
Note that the sum $\sum_{i=2}^{n-2} {\bf 1}_{A_i}$ is the number of local minima in $\sigma(1),\dots , \sigma(n-1)$,
where $\sigma$ is a uniformly random permutation from $P_{n-1}$.
\cite{romik:2011} shows that $L_n^{as}$ is equal to $1+\bf{1}_{ \{ \tau(1)>\tau(2)\}}$ plus the number of local minimums and local maximums in a random permutation, and provides $\E(L_n^{as})=2n/3+1/6$ and $\Var(L_n^{as})=8n/45-13/180$ (which are also computed in \cite{stanley:2008} and \cite{houdre:2010} in different ways).
Notice that the number of local minima and number of local maxima differ by at most one,
and hence the number of local minima is about half of $L_n^{as}$.
Therefore, we have $\E(R_n)/n\rightarrow 1/3$ and $\Var(R_n)/n \rightarrow 2/45$ as $n\rightarrow \infty$.
In fact, the limits $1/3$ and $2/45$ are actually attained for every $n\geq 5$.
\begin{thm}\label{thm:meanvarrn}
For a random sample of size $n$ from $U(0,1)$, the mean and the variance of the number of reflexive pairs, $R_n$, is $n/3$ (for $n\geq 3$) and $2n/45$ (for $n\geq 5$), respectively.
\end{thm}
\begin{proof}
By \eqref{eq:equiv} it suffices to derive the mean and the variance of $\sum_{i=1}^{n-1} {\bf 1}_{A_i}$.
We first compute the mean.
Clearly, $\E( {\bf 1}_{A_1})=P(A_1)=P(\s(2)<\s(1))=1/2$ and similarly by symmetry $\E( {\bf 1}_{A_{n-1}})=1/2$.
For $2\leq i \leq n-2$, we easily get $\E( {\bf 1}_{A_i})=P(A_i)=P(\s(i)<\min\{\s(i-1),\s(i+1)\})=1/3$.
Thus, for $n\geq 3$ we obtain
\[
\E(R_n)=\E \left(\sum_{i=1}^{n-1} {\bf 1}_{A_i} \right)= \sum_{i=1}^{n-1} \E({\bf 1}_{A_i}) =\frac{1}{2}+(n-3)\frac{1}{3} +\frac{1}{2}=\frac{n}{3}.
\]
For the variance of $R_n$, we derive the covariances of ${\bf 1}_{A_i}$'s given in the following matrix:
\[
\left( \Cov({\bf 1}_{A_i}, {\bf 1}_{A_j}) \right)_{i,j =1}^{n-1}=
\left( \begin{array}{cccccccccc}
\vspace{.2cm}
\frac{1}{4} & \frac{-1}{6} & \frac{1}{24} & 0 & 0 & 0 & 0 & 0 & \cdots & 0\\
\vspace{.2cm}
\frac{-1}{6} & \frac{2}{9} & \frac{-1}{9} & \frac{1}{45} & 0 & 0 & 0 & 0 & \cdots & 0\\
\vspace{.2cm}
\frac{1}{24} & \frac{-1}{9} & \frac{2}{9} & \frac{-1}{9} & \frac{1}{45} & 0 & 0 & 0 & \cdots & 0\\
\vspace{.2cm}
0 & \frac{1}{45} & \frac{-1}{9} & \frac{2}{9} & \frac{-1}{9} & \frac{1}{45} & 0 & 0 & \cdots & 0\\
\vspace{.2cm}
0 & 0 & \frac{1}{45} & \frac{-1}{9} & \frac{2}{9} & \frac{-1}{9} & \frac{1}{45} & 0 & \cdots & 0\\
\vspace{.2cm}
\vdots & \vdots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \vdots \\
\vspace{.2cm}
0 & 0 & \cdots & 0 & \frac{1}{45} & \frac{-1}{9} & \frac{2}{9} & \frac{-1}{9} & \frac{1}{45} & 0 \\
\vspace{.2cm}
0 & 0 & \cdots & & 0 & \frac{1}{45} & \frac{-1}{9} & \frac{2}{9} & \frac{-1}{9} & \frac{1}{24} \\
\vspace{.2cm}
0 & 0 & \cdots & & & 0 & \frac{1}{45} & \frac{-1}{9} & \frac{2}{9} & \frac{-1}{6} \\
\vspace{.2cm}
0 & 0 & \cdots & & & & 0 & \frac{1}{24} & \frac{-1}{6} & \frac{1}{4} \\
\end{array}
\right).
\]
First notice that the events $A_i$ and $A_j$ are independent whenever $|i-j|>2$,
since each $A_i$ only depends on the ordering of $\s(i-1), \s(i)$ and $\s(i+1)$.
Thus, we get $\Cov({\bf 1}_{A_i}, {\bf 1}_{A_j})=0$ if $|i-j|>2$.
Remaining covariances on the diagonal strip $|i-j|\leq 2$ are computed as follows.
By symmetry assume $i\leq j$.
For $i=j$, one can easily have
\begin{align}
\Cov({\bf 1}_{A_i}, {\bf 1}_{A_i})=\Var({\bf 1}_{A_i})=P(A_i) (1-P(A_i))=\begin{cases} 1/4 & i=1 \text{ or\ } n-1,\\ 2/9 & 2\leq i \leq n-2. \end{cases} \label{eq:varAi}
\end{align}
Next, we compute the off diagonal terms on the strip $|i-j|\leq 2$.
For $i=1$, we have
\begin{align}
\Cov({\bf 1}_{A_1}, {\bf 1}_{A_2})&=P(A_1 \cap A_2)-\frac{1}{2} \cdot \frac{1}{3}
=P(\s(1)<\s(2), \s(2)<\s(1), \s(2)<\s(3))-\frac{1}{6} \nonumber\\
&= 0-\frac{1}{6}=\frac{-1}{6}, \label{eq:covA1A2}
\end{align}
since the event $\{ \s(1)<\s(2), \s(2)<\s(1), \s(2)<\s(3) \}$ can not occur
and
\begin{align}
\Cov({\bf 1}_{A_1}, {\bf 1}_{A_3})&=P(A_1 \cap A_3)-\frac{1}{2} \cdot \frac{1}{3}
=P(\s(1)<\s(2)> \s(3)<\s(4))-\frac{1}{6} \nonumber \\
&= \frac{5}{24}-\frac{1}{6}=\frac{1}{24}. \label{eq:covA1A3}
\end{align}
where $5/24$ comes from the fact that there are 5 alternating permutations of order 4.
By symmetry, we also have
\begin{align}
\Cov({\bf 1}_{A_{n-2}}, {\bf 1}_{A_{n-1}})=-1/6 \text{ and }
\Cov({\bf 1}_{A_{n-3}}, {\bf 1}_{A_{n-1}})=1/24. \label{eq:covAn-1}
\end{align}
When $j=i+1$, for each $2\leq i \leq n-3$ we have
\begin{align}
\Cov({\bf 1}_{A_i}, {\bf 1}_{A_{i+1}})&=\Cov({\bf 1}_{A_2}, {\bf 1}_{A_{3}}) =P(A_2 \cap A_{3})-\frac{1}{3} \cdot \frac{1}{3} \nonumber \\
&=P(\s(2)<\s(1), \s(2)<\s(3), \s(3)<\s(2), \s(3)<\s(4))-\frac{1}{9} = 0-\frac{1}{9}=\frac{-1}{9}, \label{eq:covA2A3}
\end{align}
since the event $\{ \s(2)<\s(1), \s(2)<\s(3), \s(3)<\s(2), \s(3)<\s(4) \}$ can not occur.
And, finally, when $j=i+2$, for each $2 \leq i \leq n-4$ we obtain
\begin{align}
\Cov({\bf 1}_{A_i}, {\bf 1}_{A_{i+2}})&=\Cov({\bf 1}_{A_2}, {\bf 1}_{A_{4}}) =P(A_2 \cap A_{4})-\frac{1}{3} \cdot \frac{1}{3} \nonumber \\
&=P(\s(1)> \s(2)<\s(3)> \s(4)<\s(5))-\frac{1}{9} \nonumber \\
&= \frac{16}{120}-\frac{1}{9}=\frac{1}{45}. \label{eq:covA2A4}
\end{align}
where $16/120$ comes from the fact that there are 16 alternating permutations of order 5.
Therefore, for each $n\geq 5$, by combining the equations in \eqref{eq:varAi}-\eqref{eq:covA2A4} we obtain
\begin{align*}
\Var(R_n)&=\Var \left(\sum_{i=1}^{n-1} {\bf 1}_{A_i} \right)=\sum_{i,j=1}^{n-1} \Cov( {\bf 1}_{A_i}, {\bf 1}_{A_j}) \\
&=2 \cdot \frac{1}{4} +(n-3)\frac{2}{9} +4 \cdot \frac{-1}{6}+ 4 \cdot \frac{1}{24}+ 2(n-4)\frac{-1}{9} + 2(n-5)\frac{1}{45} \\
&= \frac{2n}{45}.
\end{align*}
\end{proof}
\subsection{Mean and Variance of $Q_n$}
The mean and variance of $Q_n$ can be derived in a similar manner.
\begin{thm}\label{thm:meanvarqn}
For a random sample of size $n$ from $U(0,1)$, the mean and the variance of the number of shared NNs, $Q_n$,
is $n/4$, for $n\geq 4$, and $19n/240$, for $n\geq 7$, respectively.
\end{thm}
\begin{proof}
Again by \eqref{eq:equiv}, we compute the mean and the variance of $\sum_{i=1}^{n-2} {\bf 1}_{B_i}$.
Clearly,
\[
\E({\bf 1}_{B_i})=P(B_i)=\begin{cases} 1/2 & i=1 \ {\text or\ } n-2,\\ 1/4 & 2\leq i \leq n-3, \end{cases}
\]
and hence, for every $n\geq 4$, we get
\[
\E(Q_n)=\E \left(\sum_{i=1}^{n-2} {\bf 1}_{B_i}\right)= \sum_{i=1}^{n-2} \E({\bf 1}_{B_i})=\frac{1}{2}+(n-4)\frac{1}{4} +\frac{1}{2}=\frac{n}{4}.
\]
For the variance, we compute the covariances of ${\bf 1}_{B_i}$'s in the following matrix:
\[
\left( \Cov({\bf 1}_{B_i}, {\bf 1}_{B_j}) \right)_{i,j =1}^{n-2}=
\left( \begin{array}{ccccccccccccc}
\vspace{.2cm}
\frac{1}{4} & 0 & \frac{-1}{8} & \frac{1}{24} & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots & 0\\
\vspace{.2cm}
0 & \frac{3}{16} & \frac{-1}{80} & \frac{-1}{16}& \frac{1}{48} & 0 & 0 & 0 & 0 & 0 & 0 & \cdots & 0 \\
\vspace{.2cm}
\frac{-1}{8} & \frac{-1}{80} & \frac{3}{16} & \frac{-1}{80} & \frac{-1}{16} & \frac{1}{48} & 0 & 0 & 0 & 0 & 0 & \cdots & 0 \\
\vspace{.2cm}
\frac{1}{24} & \frac{-1}{16} & \frac{-1}{80} & \frac{3}{16} & \frac{-1}{80} & \frac{-1}{16} & \frac{1}{48} & 0 & 0 & 0 & 0 & \cdots & 0\\
\vspace{.2cm}
0 & \frac{1}{48} & \frac{-1}{16} & \frac{-1}{80} & \frac{3}{16} & \frac{-1}{80} & \frac{-1}{16} & \frac{1}{48} & 0 & 0 & 0 & \cdots & 0 \\
\vspace{.2cm}
0 & 0 & \frac{1}{48} & \frac{-1}{16} & \frac{-1}{80} & \frac{3}{16} & \frac{-1}{80} & \frac{-1}{16} & \frac{1}{48} & 0 & 0 & \cdots & 0 \\
0 & 0 & 0 & \frac{1}{48} & \frac{-1}{16} & \frac{-1}{80} & \frac{3}{16} & \frac{-1}{80} & \frac{-1}{16} & \frac{1}{48} & 0 & \cdots & 0 \\
\vspace{.2cm}
\vdots & \vdots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \ddots & \vdots \\
\vspace{.2cm}
0 & 0 & \cdots & & 0 & \frac{1}{48} & \frac{-1}{16} & \frac{-1}{80} & \frac{3}{16} & \frac{-1}{80} & \frac{-1}{16} & \frac{1}{48} & 0 \\
\vspace{.2cm}
0 & 0 & \cdots & & & 0 & \frac{1}{48} & \frac{-1}{16} & \frac{-1}{80} & \frac{3}{16} & \frac{-1}{80} & \frac{-1}{16} & \frac{1}{24} \\
\vspace{.2cm}
0 & 0 & \cdots & & & & 0 & \frac{1}{48} & \frac{-1}{16} & \frac{-1}{80} & \frac{3}{16} & \frac{-1}{80} & \frac{-1}{8} \\
\vspace{.2cm}
0 & 0 & \cdots & & & & & 0 & \frac{1}{48} & \frac{-1}{16} & \frac{-1}{80} & \frac{3}{16} & 0 \\
\vspace{.2cm}
0 & 0 & \cdots & & & & & & 0 & \frac{1}{24} & \frac{-1}{8} & 0 & \frac{1}{4}
\end{array} \right).
\]
Note that the events $B_i$ and $B_j$ are independent whenever $|i-j|>3$,
since each $B_i$ only depends on the ordering of $\s(i-1), \s(i), \s(i+1)$ and $\s(i+2)$.
Therefore, we have $\Cov({\bf 1}_{B_i}, {\bf 1}_{B_j})=0$ if $|i-j|>3$.
Remaining covariances (i.e., the entries in the diagonal strip $|i-j|\leq 3$) are computed as follows.
By symmetry suppose $i\leq j$.
When $i=j$, one can easily obtain the main diagonal terms
\begin{align}
\Cov({\bf 1}_{B_i}, {\bf 1}_{B_i})=\Var({\bf 1}_{B_i})=P(B_i) (1-P(B_i))=\begin{cases} 1/4 & i=1 \ \text{ or\ } n-2,\\ 3/16 & 2\leq i \leq n-3. \end{cases} \label{eq:varBi}
\end{align}
We next compute the off diagonal terms on the strip.
For $i=1$, we have
\begin{align}
\Cov({\bf 1}_{B_1}, {\bf 1}_{B_2})&=P(B_1 \cap B_2)-\frac{1}{2} \cdot \frac{1}{4}
=P(\s(2)<\s(3), \s(2)<\s(1), \s(3)<\s(4))-\frac{1}{8} \nonumber \\
&= P(\s(2)<\min\{\s(1),\s(3),\s(4)\}, \s(3)<\s(4))-\frac{1}{8} = \frac{1}{4} \cdot \frac{1}{2}-\frac{1}{8}=0, \label{eq:covB1B2}
\end{align}
\begin{align}
\Cov({\bf 1}_{B_1}, {\bf 1}_{B_3}) &=P(B_1 \cap B_3)-\frac{1}{2} \cdot \frac{1}{4}
=P(\s(2)<\s(3), \s(3)<\s(2), \s(4)<\s(5))-\frac{1}{8} \nonumber \\
&=0-\frac{1}{8}=\frac{-1}{8}, \label{eq:covB1B3}
\end{align}
since the event $\{ \s(2)<\s(3), \s(3)<\s(2), \s(4)<\s(5) \}$ can not occur
and
\begin{align}
\Cov({\bf 1}_{B_1}, {\bf 1}_{B_4})&=P(B_1 \cap B_4)-\frac{1}{2} \cdot \frac{1}{4}
=P(\s(2)<\s(3), \s(4)<\s(3), \s(5)<\s(6))-\frac{1}{8} \nonumber \\
&= P(\max\{\s(2),\s(4)\}<\s(3), \s(5)<\s(6))-\frac{1}{8} = \frac{1}{3} \cdot \frac{1}{2}-\frac{1}{8}=\frac{1}{24}. \label{eq:covB1B4}
\end{align}
By symmetry, we get
\begin{align}
\Cov({\bf 1}_{B_{n-3}}, {\bf 1}_{B_{n-2}})=0, \Cov({\bf 1}_{B_{n-4}}, {\bf 1}_{B_{n-2}})=-1/8 \text{ and }
\Cov({\bf 1}_{B_{n-5}}, {\bf 1}_{B_{n-2}})=1/24. \label{eq:covBn-2}
\end{align}
When $j=i+1$ and $2\leq i \leq n-4$, we obtain
\begin{align}
\Cov({\bf 1}_{B_{i}}, {\bf 1}_{B_{i+1}})&=\Cov({\bf 1}_{B_{2}}, {\bf 1}_{B_{3}})=P(B_2 \cap B_{3})-\frac{1}{4} \cdot \frac{1}{4} \nonumber \\
&=P(\s(2)<\s(1), \s(3)<\s(4), \s(3)<\s(2), \s(4)<\s(5))-\frac{1}{16} \nonumber \\
&=P(\s(3) < \min\{\s(1),\s(2), \s(4), \s(5)\}, \s(2)<\s(1), \s(4)<\s(5)) -\frac{1}{16} \nonumber \\
&=\frac{1}{5} \cdot \frac{1}{2} \cdot \frac{1}{2}-\frac{1}{16}=\frac{-1}{80}. \label{eq:covB2B3}
\end{align}
Similarly, if $j=i+2$ and $2\leq i \leq n-5$ we have
\begin{align}
\Cov({\bf 1}_{B_{i}}, {\bf 1}_{B_{i+2}})&=\Cov({\bf 1}_{B_{2}}, {\bf 1}_{B_{4}})=P(B_2 \cap B_{4})-\frac{1}{4} \cdot \frac{1}{4} \nonumber \\
&=P(\s(2)<\s(1), \s(3)<\s(4), \s(4)<\s(3), \s(5)<\s(6))-\frac{1}{16} \nonumber \\
&=0-\frac{1}{16}=\frac{-1}{16}, \label{eq:covB2B4}
\end{align}
since the event $\{ \s(2)<\s(1), \s(3)<\s(4), \s(4)<\s(3), \s(5)<\s(6) \}$ can not occur.
Finally, for the case $j=i+3$ and $2\leq i \leq n-6$, we get
\begin{align}
\Cov({\bf 1}_{B_{i}}, {\bf 1}_{B_{i+3}})&=\Cov({\bf 1}_{B_{2}}, {\bf 1}_{B_{5}})=P(B_2 \cap B_{5})-\frac{1}{4} \cdot \frac{1}{4} \nonumber \\
&=P(\s(2)<\s(1), \s(3)<\s(4), \s(5)<\s(4), \s(6)<\s(7))-\frac{1}{16} \nonumber \\
&=P(\max\{\s(3), \s(5)\}<\s(4), \s(2)<\s(1), \s(6)<\s(7)) -\frac{1}{16} \nonumber \\
&=\frac{1}{3} \cdot \frac{1}{2} \cdot \frac{1}{2}-\frac{1}{16}=\frac{1}{48}, \label{eq:covB2B5}
\end{align}
and therefore, for every $n\geq 7$, by combining the equations in \eqref{eq:varBi}-\eqref{eq:covB2B5} we obtain
\begin{align*}
\Var(Q_n)&=\Var \left(\sum_{i=1}^{n-2} {\bf 1}_{B_i} \right)=\sum_{i,j=1}^{n-2} \Cov( {\bf 1}_{B_i}, {\bf 1}_{B_j}) \\
&=2 \cdot \frac{1}{4} +(n-4)\frac{3}{16} +4 \cdot \frac{-1}{8}+ 4 \cdot \frac{1}{24}+ 2(n-5)\frac{-1}{80} + 2(n-6)\frac{-1}{16} + 2(n-7)\frac{1}{48} \\
&= \frac{19n}{240}.
\end{align*}
\end{proof}
\subsection{A Recurrence Relation for the Exact Distribution of $R_n$}
\label{sec:recursive-reln-Rn}
Recall that for a sequence of pairwise distinct real numbers $x_1,\dots ,x_n$, we say $x_k$ is a local minimum if $x_k$ is less than its neighbors (i.e., $x_{k-1}$ and $x_{k+1}$) for $2\leq k \leq n-1$.
Let us also consider $x_1$ (resp. $x_n$) as a local minimum if $x_1<x_2$ (resp. $x_n<x_{n-1}$).
Then, notice that $\sum_{i=1}^{n-1} {\bf 1}_{A_i}$ is exactly the number of local minima in a uniformly random permutation from $P_{n-1}$.
Let $p(n,k)$ denote $P(R_n=k)$ and set $p(1,0)=p(2,1)=1$.
Also, let $m(n,k)$ be the number of permutations in $P_n$ with exactly $k$ local minimums.
Notice that $p(n,k)=m(n-1,k)/(n-1)!$.
Since the term 1 in the sequence is always a local minimum and any two local minimums are not adjacent,
we see that $p(n,0)=0$ for $n\geq 2$ and $p(n,k)=0$ whenever $k>n/2$.
Any permutation in $P_n$ can be uniquely obtained by increasing each element of a permutation in $P_{n-1}$ by one and inserting the element 1 in one of the possible $n$ places.
In this process, inserting the element 1 into the sequence does not effect the number of local minimums if it is placed next to a local minimum, and otherwise, increases the number of local minimums by one.
Therefore, we obtain
\begin{align*}
m(n,k)=2k\cdot m(n-1,k)+(n-2(k-1))\cdot m(n-1,k-1),
\end{align*}
since any two local minimums are not adjacent.
Thus, as $m(n-1,k)=p(n,k)(n-1)!$, we have
\begin{align}
\label{eq:rec}
p(n+1,k)=\frac{2k}{n} p(n,k)+\frac{n-2k+2}{n} p(n,k-1),
\end{align}
for every $n\geq 2$.
Therefore, the exact pmf of $R_n$ can be computed for any $n\geq 3$ by using the recursion given in \eqref{eq:rec}.
\cite{enns:1999} consider the index of the spacing with the largest length (i.e., index $i$ such that $D_i=\max \{D_1,D_2,\dots ,D_{n-1}\}$) and derive the following recurrence relation
\begin{align*}
p(n,k)= \frac{2}{n-1} p(n-1,k)+ \sum_{i=2}^{n-2} \sum_{j=1}^{k-1}\frac{ p(i,j) p(n-i,k-j)}{n-1},
\end{align*}
for every $n\geq 4$.
Then, using generating functions, they obtain the relation in \eqref{eq:rec}.
\section{Asymptotic Results for $R_n$ and $Q_n$}
\label{sec:asy-res}
In this section,
we prove SLLN results and CLTs for both $R_n$ and $Q_n$ as $n \rightarrow \infty$.
Observe that neither ${\bf 1}_{A_1}, {\bf 1}_{A_2}, \dots , {\bf 1}_{A_{n-1}}$ nor ${\bf 1}_{B_1}, {\bf 1}_{B_2}, \dots , {\bf 1}_{B_{n-2}}$ is an i.i.d. sequence.
However, both have a nice structure which allows SLLN and CLT results to follow.
\begin{definition}
A sequence of random variables $X_1,X_2, \dots , X_n$
is said to be \emph{$m$-dependent} if the random variables $(X_{1},X_2, \dots , X_{i})$ and $(X_{j}, X_{j+1}, \dots , X_{n})$
are independent whenever $j-i>m$.
\end{definition}
Since each $A_i$ (resp. $B_i$) only depends on the ordering of $\s(i-1), \s(i)$ and $\s(i+1)$ (resp. $\s(i-1), \s(i), \s(i+1)$ and $\s(i+2)$), it is clear to see that the sequence ${\bf 1}_{A_1}, {\bf 1}_{A_2}, \dots , {\bf 1}_{A_{n-1}}$
(resp. ${\bf 1}_{B_1}, {\bf 1}_{B_2}, \dots , {\bf 1}_{B_{n-2}}$) is 2-dependent (resp. 3-dependent).
For the asymptotic results,
note that we can ignore the random variables ${\bf 1}_{A_{1}}$ and ${\bf 1}_{A_{n-1}}$ (resp. ${\bf 1}_{B_{1}}$ and ${\bf 1}_{B_{n-2}}$),
since their contribution to the summand $\sum_{i=1}^{n-1} {\bf 1}_{A_{i}}$ (resp. $\sum_{i=1}^{n-2} {\bf 1}_{B_{i}}$) is negligible in the limit as $n$ goes to infinity.
Therefore, to obtain asymptotic results for $R_n$ (resp. $Q_n$) it suffices to consider $\sum_{i=2}^{n-2} {\bf 1}_{A_{i}}$ (resp. $\sum_{i=2}^{n-3} {\bf 1}_{B_{i}}$)
which is the sum of identically distributed 2-dependent (resp. 3-dependent)
indicator random variables with mean $1/3$ (resp. $1/4$).
For $m$-dependent identically distributed sequences, the SLLN extends in a straightforward manner by just partitioning the summand in $m+1$ sums of i.i.d. subsequences, and hence,
we obtain SLLN results for both $R_n$ and $Q_n$.
\begin{thm}\label{thm:slln}
{\rm (SLLN for $U(0,1)$ data)} For a random sample of size $n$ from $U(0,1)$, we have ${R_n}/{n} \xrightarrow{a.s.} {1}/{3}$ and ${Q_n}/{n} \xrightarrow{a.s.} {1}/{4}$ as $n\rightarrow \infty$, where $\xrightarrow{a.s.}$ denotes almost sure convergence.
\end{thm}
The asymptotic normality of the random variables we consider is due to the well-known results on the sequence of $m$-dependent identically distributed and bounded random variables (e.g., see \cite{hoeffding:1948}, \cite{chung:1974}).
\begin{thm}{\rm (CLT for $U(0,1)$ data)}\label{thm:cltrnqn}
For a random sample of size $n$ from $U(0,1)$, we have
$$
\frac{R_n-n/3}{\sqrt{2n/45}} \xrightarrow{\mathcal{L}} \mathcal{N}(0,1) \text{ and }
\frac{Q_n-n/4}{\sqrt{19n/240}} \xrightarrow{\mathcal{L}} \mathcal{N}(0,1),$$
as $n\rightarrow \infty$,
where $\xrightarrow{\mathcal{L}}$ denotes the convergence in law and $\mathcal{N}(0,1)$ is the standard normal distribution.
\end{thm}
\begin{remark}
\label{rem:u-stat}
{\bf Is $R_n$ or $Q_n$ a U-statistic?}
At first glance (a scaled form of) $R_n$ and $Q_n$ might look like a $U$-statistic of degree 2
with symmetric kernels as we can write them as
\begin{align}\label{equstat}
\frac{R_n}{ {n \choose 2}}=\frac{1}{ {n \choose 2}} \sum_{1\leq i< j \leq n} {\bf 1}_{A(i,j)} \text{ and }
\frac{Q_n}{ {n \choose 2}}=\frac{1}{ {n \choose 2}} \sum_{1\leq i< j \leq n} {\bf 1}_{B(i,j)}
\end{align}
where $A(i,j)$ is the event that $\{X_i,X_j\}$ is a reflexive pair and
$B(i,j)$ is the event that $X_i$ and $X_j$ is share a NN.
If $R_n$ and $Q_n$ were $U$-statistics,
then asymptotic normality of both would follow by the general CLT for $U$-statistics (\cite{hoeffdingUstat:1948}).
However, the kernels (${\bf 1}_{A(i,j)}$ and ${\bf 1}_{B(i,j)}$) do not only depend on $X_i$ and $X_j$, but to all data points.
Hence the kernels are not of degree 2 but of degree $n$.
Although,
for $U$-statistics,
the degree can be equal to the sample size,
it should be a fixed quantity, $m$.
So $m \le n$ allows $m=n$, but this would be for small samples,
and as $n$ increases,
$m$ should stay fixed which is not the case here.
So, neither $R_n$ nor $Q_n$ is a $U$-statistic of finite (fixed) degree,
hence this approach would not work in proving the CLT for $R_n$ and $Q_n$.
\end{remark}
\begin{remark}\label{rem:rd}
{\bf Asymptotic Behavior of $R_n$ and $Q_n$ in Higher Dimensions.}
The results in \cite{henze:1987} and \cite{schiling:1986} imply that $\E(R_n)/n\rightarrow r(d)$ and $\E(Q_n)/n\rightarrow q(d)$ as $n\rightarrow \infty$,
where $r(d)$ and $q(d)$ are constants which only depend on the dimension $d$,
whenever the underlying distribution has an a.e. continuous density in $\mathbb{R}^d$
(i.e., $r(d)$ and $q(d)$ (somewhat unexpectedly) do not depend on the continuous distribution).
We have $r(1)=1/3$, $r(2)=3\pi /(8\pi +3\sqrt{3})\approx 0.3108$, $r(3)=8/27$, and in general,
\[
r(d) =
\begin{cases}
\displaystyle \left[ 3+\sum_{k=1}^m \frac{1\cdot 3 \cdots (2k-1)}{2\cdot 4 \cdots (2k)} \left(\frac{3}{4} \right)^k \right ]^{-1} & \text{if } d=2m+1, \\
\displaystyle \left[ \frac{8}{3}+\frac{\sqrt{3}}{\pi} \left (1+\sum_{k=1}^{m-1} \frac{2\cdot 4 \cdots (2k)}{3\cdot 5 \cdots (2k+1)} \left(\frac{3}{4} \right)^k \right) \right ]^{-1} & \text{if } d=2m,
\end{cases}
\]
(see, e.g., \cite{pickard:1982}).
On the other hand, the exact value of $q(d)$ is known only for $d=1$, $q(1)=0.25$.
For $d>1$, we only have empirical approximations, for example,
$q(2)\approx 0.315 $, $q(3) \approx 0.355$, $q(4)\approx 0.38$ and $q(5)\approx 0.4$
(\cite{schiling:1986}).
\end{remark}
\begin{remark}
Some other quantities based on NN digraph are of interest in the literature.
Notice that even though each point has a unique NN, it is not necessarily the NN of precisely one point.
Let $Q_{j,n}$ be the number of points in the data which are NN of exactly $j$ other points.
The quantities $Q_{j,n}$'s are used in tests for spatial symmetry (see, \cite{ceyhan:2014}).
Also in \cite{enns:1999}, $Q_{0,n}, Q_{1,n}$ and $Q_{2,n}$ correspond to the number of \emph{lonely}, \emph{normal} and \emph{popular} individuals in a population of size $n$, respectively.
Moreover, the fraction of points serving as NN to precisely $j$ other points (i.e., $Q_{j,n}/n$) is studied by many authors
(e.g., see \cite{clark:1955}, \cite{henze:1987} and \cite{newmanRT:1983}).
Clearly, in one dimension, a point is NN to at most two other points and hence $Q_{j,n}=0$ for every $j\geq 3$.
Double counting arguments for the number of vertices and the number of arcs give
$n=Q_{0,n}+Q_{1,n}+Q_{2,n}$ and $n=0\cdot Q_{0,n}+1\cdot Q_{1,n}+2\cdot Q_{2,n}$, respectively.
On the other hand, one can easily see $Q_n=\sum_{j\geq 0} {j \choose 2}Q_{j,n}=Q_{2,n}$ and obtain
$Q_n=Q_{0,n}=Q_{2,n}=(n-Q_{1,n})/2$.
Thus, for each $j=0,1,2$, we have SLLN and CLT results for $Q_{j,n}$ together with the exact values of its mean and variance
using the results on $Q_n$.
\end{remark}
\section{Discussion and Conclusions}
\label{sec:disc-conc}
In this article,
we study the probabilistic behavior of the number of reflexive nearest neighbors (NNs), denoted $R_n$,
and the number of shared NNs, denoted $Q_n$,
for one dimensional uniform data.
$R_n$ and $Q_n$ can also be viewed as graph invariants for the NN digraph with vertices being the data points,
and arcs being inserted from a point to its NN.
In particular,
we provide the means and variances of both $R_n$ and $Q_n$, and derive SLLN and CLT results for both of the quantities under the same settings.
We also present a recursive relation for the probability mass function (pmf)
of $R_n$,
which can provide the exact distribution of $R_n$ (by computation for finite $n$).
Recall that the results we obtain for $R_n$ (the mean, the variance and the recurrence relation)
are in agrement with the ones in \cite{enns:1999}.
However, our derivation of the results are different from theirs and their method is not applicable for $Q_n$.
This work lays the foundation for the study of (number) reflexive NN pairs, shared NN pairs
related invariants of NN digraphs in higher dimensions
which would be more challenging due to the lack of ordering of the data points in multiple dimensions.
Another potential research direction is that
the results can also be extended to data from non-uniform distributions in one or multiple dimensions.
Our Monte Carlo simulations suggest that CLT results for both $R_n$ and $Q_n$ seem to hold and
$\Var(R_n)/n$ and $\Var(Q_n)/n$ converge to $\sigma_R^2(d)$ and $\sigma_Q^2(d)$, respectively,
whenever the underlying process is a distribution on $\mathbb{R}^d$ with an a.e. continuous density,
where $\sigma_R^2(d)$ and $\sigma_Q^2(d)$ are constants which only depend on the dimension $d$.
Expectations are handled in \cite{henze:1987} and \cite{schiling:1986}, see Remark \ref{rem:rd}.
Notice that even in case of $d=1$,
we can not apply the method used in the paper when the distribution of the sample is not uniform since
we lose the interchangeability of the spacings.
The number of reflexive NN pairs was also used in inferential statistics in literature.
For example, \cite{dacey:1960} used it to test the clustering of river towns in US.
However, \cite{dacey:1960} ignored the uncertainty due to the randomness in $R_n$
and compared the observed $R_n$ values to its expected value
to declare clustering or regularity of the towns.
This methodology was also criticized by \cite{pinder:1975}
who proposed an alternative method based on the average NN distance (and its empirical pdf) for the same type of inference.
But,
$R_n$ can be employed in exact inference using its exact pmf for testing such one-dimensional patterns
for small $n$ (as the exact distribution depends on the distribution of the data).
However, by the above discussion on $r(d)$ and $\sigma_R^2(d)$,
for data from any continuous distribution,
$R_n$ would converge to the same normal distribution as $n$ goes to infinity.
Hence, testing spatial clustering/regularity based on the asymptotic approximation of $R_n$ is not appropriate
(hence not recommended),
as it would have power equal to the significance level of the test in the limit
under any continuous alternative as well as under the null pattern (i.e., under uniformity of the points).
On the other hand,
if the convergence in probability of $R_n/n$ and $Q_n/n$ to some constants (regardless of the distribution of continuous data)
is established in all dimensions,
then this would be a desirable property for removing the restrictions of NN tests
which are conditional on $R_n$ and $Q_n$, (e.g., tests of \cite{dixon:1994})
in their asymptotic distribution.
The types of convergence for $R_n/n$ and $Q_n/n$ for data in higher dimensions are topics of ongoing research.
\section*{Acknowledgments}
EC was supported by the European Commission under the Marie Curie International Outgoing Fellowship Programme
via Project \# 329370 titled PRinHDD.
|
\section{Introduction}
The three--parameter Mittag--Leffler function was introduced by Prabhakar \cite{Prabhakar} in the form
\begin{equation} \label{M1}
E_{\alpha, \beta}^\gamma(z) = \sum_{k \geq 0} \frac{(\gamma)_k}{\Gamma(\alpha k+\beta)}
\frac{z^k}{k!}, \qquad \min\{ \Re(\alpha), \Re(\beta), \Re(\gamma)\}>0;\, z \in \mathbb C.
\end{equation}
Here and in what follows
\[ (\lambda)_{\mu} = \dfrac{\Gamma(\lambda+\mu)}{\Gamma(\lambda)} =
\begin{cases}
1 & \mu = 0;\, \lambda \in \mathbb C \setminus \{0\}\\
\lambda(\lambda+1) \cdots (\lambda + n-1) & \mu=n \in \mathbb N;\; \lambda \in \mathbb C
\end{cases} \,; \]
it being understood conventionally that $(0)_0:= 1$, where the latter case stands for the standard familiar Pochhammer symbol
(or shifted factorial). For comprehensive treatment of $E_{\alpha, \beta}^\gamma(z)$ consult e.g. \cite{Tomovski}.
For $\gamma =1$, we recover with \eqref{M1} the two--parameter Mittag--Leffler function $E_{\alpha, \beta }(z)$ defined by
\[ E_{\alpha, \beta }(z) = \sum_{k \geq 0} \frac{z^k}{\Gamma(\alpha k + \beta)},
\qquad \Re (\alpha )>0, \beta \in \mathbb{C};\,z\in \mathbb C \, . \]
Moreover, the case $\beta = \gamma =1$ yields the classical Mittag--Leffler function
\[ E_{\alpha }(z)=\sum_{k\geq 0}\frac{z^{k}}{\Gamma (\alpha k+1)},\qquad \Re(\alpha )>0;\,z\in \mathbb{C}. \]
\section{Fractional Poisson process with Prabhakar's three--parameter Mittag--Leffler function}
The Mittag-Leffler function appeared as residual waiting time between events in renewal processes already in the sixties of the
past century, namely processes with properly scaled thinning out the sequence of events in a power law renewal process. Such a process
in essences is a fractional Poisson process \cite{Beghin}, \cite{Laskin1}, \cite{Laskin2}, \cite{Mainardi}, \cite{Politi}
which in the recent years has become subject of intensive research. The telegraph process is a finite--velocity one dimensional
random motion of which various probabilistic features are known.
The fractional extensions of the telegraph process $\{ \mathscr T_\alpha(t) \colon t \geq 0\}$, whose changes of
direction are related to the fractional Poisson process $\{ \mathscr N_\alpha(t) \colon t \geq 0\}$ having distribution \cite{Beghin}
\[ \mathbb P(\mathscr N_\alpha(t) = k) = \frac{\lambda^k}{E_\alpha(\lambda t^\alpha)} \frac{t^{\alpha k}}{\Gamma(\alpha k+1)},
\qquad k \in \mathbb{N}_{0}:=\mathbb{N}\cup \{0\};\,t\geq 0. \]
However, we point out that the fractional Poisson process resulting in $\{ \mathscr N_{\alpha, \beta}(t) \colon t\geq 0\}$, defined by
the two--parameter Mittag--Leffler function $E_{\alpha, \beta}(\lambda t^\alpha)$ was studied by Herrmann \cite{Herrmann}. The
related distribution reads
\[ \mathbb P(\mathscr N_{\alpha, \beta}(t) = k) = \frac{\lambda^k}{E_{\alpha, \beta}(\lambda t^\alpha)}
\frac{t^{\alpha k}}{\Gamma (\alpha k + \beta)},\qquad k\in \mathbb{N}_{0};\,t\geq 0, \]
for which the related raw moments are obtained in terms of the Bell polynomials, see \cite[p. 5]{Herrmann}.
Here we introduce a more general fractional Poisson process $\{\mathscr N_{\alpha, \beta}^\gamma(t) \colon t\geq 0\}$ defined by the
Prabhakar (or in another words three--parameter Mittag--Leffler) function $E_{\alpha, \beta}^\gamma(\lambda t^\alpha)$
having distribution
\begin{equation} \label{Y0}
\mathbb P(\mathscr N_{\alpha, \beta}^\gamma(t) = k) = \frac{\lambda^k}{E_{\alpha, \beta}^\gamma(\lambda t^\alpha)}
\frac{(\gamma)_k\, t^{\alpha k}}{k!\, \Gamma(\alpha k+\beta)}, \qquad
k \in \mathbb N_0;\, t \geq 0.
\end{equation}
At this point it is worth to mention that there is a correspondence between the non--homogeneous Poisson process
$\{ \mathscr N(t) \colon t \geq 0\}$ with intensity function $\lambda \alpha t^{\alpha-1}$, that is
\[ \mathbb P( \mathscr N(t) = k) = {\rm e}^{-\lambda t^\alpha} \, \frac{(\lambda t^\alpha)^k}{\Gamma(k+1)},
\qquad \lambda, \alpha>0; k \in \mathbb N_0\, ,\]
and the Prabhakar--type fractional Poisson process $\{\mathscr N_{\alpha, \beta}^\gamma(t) \colon t\geq 0\}$. This connection is
exposed in the following result.
\begin{proposition} Let $\min\{ \alpha, \beta, \gamma, \lambda\}>0$ and $t \geq 0$. Then
\[ \mathbb P(\mathscr N_{\alpha, \beta}^\gamma(t) = k) = \dfrac{\dfrac{(\gamma)_k}{\Gamma(\alpha k + \beta)}\,
\mathbb P( \mathscr N(t) = k)}{\sum\limits_{n \geq 0} \dfrac{(\gamma)_n}{\Gamma(\alpha n + \beta)}\,
\mathbb P( \mathscr N(t) = n)}, \qquad k \in \mathbb N_0\,,\]
where $\mathscr N(t)$ is a non--homogeneous Poisson process with intensity function $\lambda \alpha t^{\alpha-1}$.
\end{proposition}
\begin{proof} Rewriting \eqref{Y0} into
\[ \mathbb P(\mathscr N_{\alpha, \beta}^\gamma(t) = k) = \frac{\dfrac{(\gamma)_k}{\Gamma(\alpha k + \beta)}\,
\dfrac{(\lambda t^\alpha)^k}{k!}\, {\rm e}^{-\lambda t^\alpha}}{\sum\limits_{n \geq 0}
\dfrac{(\gamma)_n}{\Gamma(\alpha n + \beta)}\, \dfrac{(\lambda t^\alpha)^n}{n!}\, {\rm e}^{-\lambda t^\alpha}}\,, \]
we arrive at the statement.
\end{proof}
In accordance with \eqref{Y0} we shall study the probability distribution associated with the Prabhakar function
$E_{\alpha, \beta}^\gamma(\lambda t^\alpha)$ letting $\lambda = 1$ throughout for the sake of simplicity. So, consider a non--negative
random variable $X$ on a standard probability space $(\Omega, \mathscr F, \mathbb P)$ having the fractional
Poisson--type distribution
\[ \mathbb P_{\alpha, \beta}^\gamma(k) = \mathbb P(X = k) = \frac1{E_{\alpha, \beta}^\gamma(t^\alpha) }
\frac{(\gamma)_k \,t^{\alpha k}}{k!\, \Gamma(\alpha k+\beta)}, \qquad
k \in \mathbb N_0;\, t \geq 0, \]
where remains $\min\{\alpha, \beta, \gamma\} >0$. Being $\sum_{k \geq 0} \mathbb P_{\alpha, \beta}^\gamma(k) = 1$, the rv $X$ i
s well defined. This correspondence we quote in the sequel $X \sim {\rm ML}(\alpha, \beta, \gamma)$.
Let us remind that the factorial moment of the rv $X$ of order $s \in \mathbb N$ is given by
\[ \Phi_s = \mathbb E X(X-1) \cdots (X-s+1) = (-1)^s\, \mathbb E(-X)_s
= \frac{{\rm d}^s}{{\rm d}t^s}\left(\mathbb E\,t^X\right)\Big|_{t=1}\,,\]
provided the moment generating function $M_X(t) = \mathbb E\, t^X$ there exists in some neighborhood of $t =1$ together will all
its derivatives up to the order $s$. By virtue of the Vi\`ete--Girard formulae for expanding $X(X-1) \cdots (X-s+1)$ we obtain
\[ \Phi_s = \sum _{r=1}^s (-1)^{s-r} \, e_r\, \mathbb EX^r\,;\]
here $e_r$ represents elementary symmetric polynomials:
\[ e_r = e_r(\ell_1, \cdots, \ell_r) = \sum_{1 \le \ell_1< \cdots <\ell_r \le s-1}
\ell_1 \cdots \ell_r, \qquad r = \overline{0,s-1}.\]
\begin{theorem}
For all $\min\{ \Re(\alpha), \Re(\beta), \Re(\gamma)\}>0$ the $s$--th raw moment of the rv $X \sim {\rm ML}(\alpha, \beta, \gamma)$
reads as follows
\begin{equation} \label{N3}
\mathbb E X^s = \frac1{E_{\alpha, \beta}^\gamma(t^\alpha) } \sum_{j=0}^s (\gamma)_j\,{s \brace j} \,t^{\alpha j}\,
E_{\alpha, \alpha j+\beta}^{\gamma+j}(t^\alpha), \qquad s \in \mathbb N_0,\, t \geq 0.
\end{equation}
Moreover, the $s$--th factorial moment
\begin{equation} \label{N5}
\Phi_s = \frac1{E_{\alpha, \beta}^\gamma(t^\alpha) } \sum _{r=1}^s (-1)^r \, e_r\,
\sum_{j=0}^r (\gamma)_j\,{r \brace j} \,t^{\alpha j}\, E_{\alpha, \alpha j+\beta}^{\gamma+j}(t^\alpha)\,.
\end{equation}
\end{theorem}
\begin{proof} Recall the connection between the raw and the factorial moments of a rv:
\[ \mathbb E X^s = \sum_{j=0}^s (-1)^j\,{s \brace j} \mathbb E(-X)_j, \qquad
{s \brace j} = \frac1{j!} \sum_{m=0}^j (-1)^{j-m} \binom{j}{m} m^s\,,\]
where the curly braces denote Stirling numbers of the second kind. As to the stated result we need only the direct calculation:
\begin{align*}
\mathbb E X^s &= \sum_{j=0}^s (-1)^j\, {s \brace j} \mathbb E(-X)_j
= \sum_{j=0}^s (-1)^j\,{s \brace j} \sum_{k \geq 0} (-k)_j \, \mathbb P_{\alpha, \beta}^\gamma(k) \\
&= \frac1{E_{\alpha, \beta}^\gamma(t^\alpha) } \sum_{j=0}^s (-1)^j\,{s \brace j} \sum_{k \geq 0}
\frac{(-k)_j \,(\gamma)_k \,t^{\alpha k}}{k!\, \Gamma(\alpha k+\beta)} \\
&= \frac1{E_{\alpha, \beta}^\gamma(t^\alpha) \, \Gamma(\gamma)} \sum_{j=0}^s {s \brace j}\, t^{\alpha j}\,
\sum_{k \geq j} \frac{\Gamma(\gamma+(k-j)+j) \,t^{\alpha(k-j)}}{(k-j)!\,
\Gamma(\alpha (k-j)+ \alpha j+\beta )} \\
&= \frac1{E_{\alpha, \beta}^\gamma(t^\alpha) } \sum_{j=0}^s \frac{\Gamma(\gamma+j)}{\Gamma(\gamma)} \,
\,{s \brace j} \,t^{\alpha j} \,\sum_{k \geq 0} \frac{(\gamma+j)_k \,t^{\alpha k}}
{k!\, \Gamma(\alpha k+ \alpha j+\beta)} \,,
\end{align*}
which is the statement \eqref{N3}. The derivation of \eqref{N5} is now obvious.
\end{proof}
To obtain the fractional order moments we need the so-called extended Hurwitz--Lerch Zeta (HLZ) function introduced by Gupta {\it et al.}
\cite{GGOS} (see also \cite[p. 491, Eq. (1.20); p. 503, Eq. (6.2)]{Srivastava}) $\Phi^{(\rho, \sigma, \kappa)}_{\lambda, \mu;
\nu}(z, s, a)$ is given by
\begin{equation} \label{N6}
\Phi^{(\rho, \sigma, \kappa)}_{\lambda, \mu; \nu}(z, s, a) = \sum_{n \geq 0}
\frac{(\lambda)_{\rho n}\, (\mu)_{\sigma n}}{n!\, (\nu)_{\kappa n}}\, \dfrac{z^n}{(n+a)^s}\,,
\end{equation}
where $\lambda, \mu \in \mathbb C;\, a, \nu \in \mathbb C \setminus \mathbb Z_0^-$; $\rho, \sigma, \kappa >0$; $\kappa-\rho-\sigma+1>0$
when $s, z \in \mathbb C$; $\kappa-\rho-\sigma = -1$ and $s \in \mathbb C$ when $|z|< \delta = \rho^{-\rho} \sigma^{-\sigma}
\kappa^\kappa$; while $\kappa-\rho-\sigma = -1$ and $\Re(s + \nu - \lambda - \mu) >1$ when $|z|= \delta$. Specially by letting
$\sigma \to 0$ in \eqref{N6} we arrive at the generalized HLZ function $\Phi^{(\rho, 0, \kappa)}_{\lambda, \mu; \nu}(z, s, a) \equiv
\Phi^{(\rho, \kappa)}_{\,\,\lambda; \nu}(z, s, a)$ which is of importance one for our next result.
\begin{theorem} Let $X \sim {\rm ML}(\alpha, \beta, \gamma)$. For all $\min\{\alpha, \beta, \gamma\}>0$ and for all $s \geq 0$
we have
\begin{equation} \label{N7}
\mathbb E X^s = \frac{\gamma\,t^\alpha}{E_{\alpha, \beta}^\gamma(t^\alpha)\, \Gamma(\alpha+\beta)}\,
\Phi^{(1, \alpha)}_{\gamma+1; \alpha+\beta}(t^\alpha, 1-s, 1)\,.
\end{equation}
\end{theorem}
\begin{proof} By definition, for all $s>0$ we have
\[ \mathbb E X^s = \frac1{E_{\alpha, \beta}^\gamma(t^\alpha)} \sum_{n \geq 1} n^s \,
\frac{(\gamma)_n \,t^{\alpha n}}{n!\, \Gamma(\alpha n+\beta)}\,, \]
since the therm zeroth vanishes. In turn {\it mutatis mutandis}
\begin{align*}
\mathbb E X^s &= \frac1{E_{\alpha, \beta}^\gamma(t^\alpha)} \sum_{n \geq 1} \,
\frac{n^{s-1}\, (\gamma)_n\, t^{\alpha n}}{(n-1)!\, \Gamma(\alpha n+\beta)} \\
&= \frac{\gamma\,t^\alpha}{E_{\alpha, \beta}^\gamma(t^\alpha)} \sum_{n \geq 0} \,
\frac{(\gamma+1)_n\, t^{\alpha n}}{n!\, \Gamma(\alpha n+\alpha+\beta)\,(n+1)^{1-s}} \\
&= \frac{\gamma\,t^\alpha}{E_{\alpha, \beta}^\gamma(t^\alpha)\, \Gamma(\alpha+\beta)}\,
\Phi^{(1, \alpha)}_{\gamma+1; \alpha+\beta}(t^\alpha, 1-s, 1)\,.
\end{align*}
Now, being $\lambda = \gamma+1, \nu=\alpha+\beta, z=t^\alpha, s \mapsto 1-s, \rho = 1, \kappa = \alpha$ and $a=1$, by applying the
convergence constraints for $\Phi^{(\rho, \kappa)}_{\,\,\lambda; \nu}(z, s, a)$ listed around \eqref{N6}, we finish the proof
of the assertion.
\end{proof}
\begin{remark} It is worth to mention that Herrmann \cite[p. 5, Eq. (3.10)]{Herrmann} has been derived the raw integer order moments
for the two--parameter Mittag--Leffler distributed rv which is covered with our $Y \sim {\rm ML}(\alpha, \beta, 1)$ distribution,
in the form
\[ \mathbb EY^n = \frac1{E_{\alpha, \beta}(t)}\, \left( t\, \frac{\rm d}{{\rm d}t}\right)^n\, E_{\alpha, \beta}(t),
\qquad n \in \mathbb N_0\, . \]
Recalling that $E_{\alpha, \beta}^1(t) = E_{\alpha, \beta}(t)$, our relations \eqref{N3} and \eqref{N7} cover this result. Indeed, taking
$s=1, \gamma=1, t \mapsto t^\frac1\alpha$ in \eqref{N3} we have
\[ \mathbb EX = t\, \frac{E_{\alpha, \alpha+\beta}^2(t)}{E_{\alpha, \beta}(t)}\,.\]
On the other hand, since
\[ \left(E_{\alpha, \beta}(t)\right)' = \sum_{n \geq 1} \frac{n\, t^{n-1}}{\Gamma(\alpha n+\beta)}
= \sum_{n \geq 1} \frac{(2)_{n-1} t^{n-1}}{(n-1)!\,\Gamma(\alpha (n-1) + \alpha + \beta)}
= E_{\alpha, \alpha+\beta}^2(t)\,,\]
we see that $\mathbb EX \equiv \mathbb EY$.
Next, by setting $s=1, \gamma=1, t \mapsto t^\frac1\alpha$ formula \eqref{N7} becomes
\[ \mathbb E X = t\,\frac{\Phi^{(1, \alpha)}_{2; \alpha+\beta}(t, 0, 1) }{E_{\alpha, \beta}(t)\, \Gamma(\alpha+\beta)}
= \frac{t}{E_{\alpha, \beta}(t)\, \Gamma(\alpha+\beta)}\,
\sum_{n \geq 0} \frac{(2)_n t^n}{n!\,(\alpha+\beta)_{\alpha n}}.\]
The rest is obvious. \hfill $\Box$
\end{remark}
\begin{remark}
In fact, Theorem 2 is the fractional (in $s$) counterpart of the first claim \eqref{N3} of Theorem 1. However, the advantage of
preceding is that it consists from a finite $s+1$--term linear combination of three parameter Mittag--Leffler functions of
Prabhakar type. Specifying now $s \in \mathbb N_0$ in \eqref{N7}, we obtain a new set of summation formulae for a special class
of extended HLZ functions $\Phi^{(1, \alpha)}_{\gamma+1; \alpha+\beta}(t^\alpha, 1-s, 1)$ in terms of related Prabhakar function.
\hfill $\Box$
\end{remark}
\begin{corollary} For all $\min\{\alpha, \beta, \gamma\}>0$ and for all $s \in \mathbb N_0$ we have
\[ \Phi^{(1, \alpha)}_{\gamma+1; \alpha+\beta}(t^\alpha, 1-s, 1) = \frac{\Gamma(\alpha+\beta)}{\gamma\,t^\alpha}\,
\sum_{j=0}^s (\gamma)_j\,{s \brace j} \,t^{\alpha j}\, E_{\alpha, \alpha j+\beta}^{\gamma+j}(t^\alpha)\,,
\qquad t >0.\]
\end{corollary}
\section{Functional upper bounds for a Tur\'anian difference built by Prabhakar function}
Our next goal is to derive a Tur\'an type inequality for the Prabhakar function. In this purpose let us introduce the
generalized Tur\'anian difference
\[ \Delta_T(t) = \left[E_{\alpha, \alpha+\beta}^{\gamma+1}(t^\alpha) \right]^2
- E_{\alpha, \beta}^\gamma(t^\alpha)\,E_{\alpha, 2\alpha+\beta}^{\gamma+2}(t^\alpha),\qquad t>0\, ,\]
which was built with respect to all three parameters in the initial Mittag--Leffler function $E_{\alpha, \beta}^\gamma(t^\alpha)$. To
establish a functional upper bound for $\Delta_T(t)$ we study firstly the first two raw moments, using Theorem 2.
The resulting expressions are
\begin{align}
\mathbb E X &= \frac{\gamma\,t^\alpha}{E_{\alpha, \beta}^\gamma(t^\alpha)\, \Gamma(\alpha+\beta)}\,
\Phi^{(1, \alpha)}_{\gamma+1; \alpha+\beta}(t^\alpha, 0, 1) \notag \\
&= \frac{\gamma\,t^\alpha\, E_{\alpha, \alpha+\beta}^{\gamma+1}(t^\alpha)}
{E_{\alpha, \beta}^\gamma(t^\alpha)} \\ \label{N9}
\mathbb E X^2 &= \frac{\gamma\,t^\alpha}{E_{\alpha, \beta}^\gamma(t^\alpha)\, \Gamma(\alpha+\beta)}\,
\Phi^{(1, \alpha)}_{\gamma+1; \alpha+\beta}(t^\alpha, 1-s, 1) \notag \\
&= \frac1{E_{\alpha, \beta}^\gamma(t^\alpha)} \sum_{n \geq 1} \,
\frac{[(n-1)+1]\, (\gamma)_n\, t^{\alpha (n-1+1)}}{(n-1)!\, \Gamma(\alpha (n-1)+\alpha+\beta)} \notag \\
&= \frac{\gamma\,t^\alpha}{E_{\alpha, \beta}^\gamma(t^\alpha)} \left[ E_{\alpha, \alpha+\beta}^{\gamma+1}(t^\alpha)
+ (\gamma+1) t^\alpha \,E_{\alpha, 2\alpha+\beta}^{\gamma+2}(t^\alpha) \right]\, .
\end{align}
Making use of the basic moment property $\mathbb EX^2 \geq (\mathbb EX)^2$ valid for all finite second order moment random variables,
we deduce:
\[ E_{\alpha, \beta}^\gamma(t^\alpha)\, \left[ E_{\alpha, \alpha+\beta}^{\gamma+1}(t^\alpha)
+ (\gamma+1) t^\alpha \,E_{\alpha, 2\alpha+\beta}^{\gamma+2}(t^\alpha) \right]
\geq \gamma\,t^\alpha\, \left[E_{\alpha, \alpha+\beta}^{\gamma+1}(t^\alpha) \right]^2\,, \]
which is equivalent to
\begin{equation} \label{N9}
E_{\alpha, \beta}^\gamma(t^\alpha)\, \left[ E_{\alpha, \alpha+\beta}^{\gamma+1}(t^\alpha)
+ t^\alpha \,E_{\alpha, 2\alpha+\beta}^{\gamma+2}(t^\alpha) \right]
\geq \gamma\,t^\alpha\, \Delta_T(t) \, .
\end{equation}
\begin{lemma} For all $\min\{\alpha, \beta, \gamma\}>0$ we have
\begin{equation} \label{N10}
E_{\alpha, \beta}^\gamma(t^\alpha) \leq \frac1{\Gamma_0}\, \frac1{(1-t^\alpha)^\gamma}\,, \qquad t\in (0,1)\, ,
\end{equation}
where $\Gamma_0 = \min\limits_{n \in \mathbb N_0} \Gamma(\alpha n+\beta)$.
Moreover, for all $\alpha \geq 1,\, \beta \geq t_0,\, \gamma>0$, where $(t_0, \Gamma(t_0))$ is the unique positive minimum of
the $\Gamma$ function, there holds
\begin{equation} \label{N11}
E_{\alpha, \beta}^\gamma(t^\alpha) \leq \frac1{\Gamma(\beta)} \, {}_1F_1(\gamma; \beta; t^\alpha)\,, \qquad t>0\, .
\end{equation}
Here ${}_1F_1(a; b; z)$ stands for the Kummer's confluent hypergeometric function
\[ {}_1F_1(a; b; z) = \sum_{k \geq 0} \frac{(a)_k}{(b)_k} \frac{z^k}{k!}\, . \]
\end{lemma}
\begin{proof} As to the proof of \eqref{N10} we only have to remind that in the right half--plane there is the unique
$\min \Gamma(t_0) \approx 0.885603$ at $t_0 \approx 1.46163$, say. (Obviously we cannot guarantee that $x_0 \in \{ \alpha k+\beta
\colon k \in \mathbb N_0\}$). So
\[ E_{\alpha, \beta}^\gamma(t^\alpha) \leq \frac1{\min\limits_{k \in \mathbb N_0} \Gamma(\alpha k+\beta)}
\sum_{k \geq 0} (\gamma)_k \frac{t^{\alpha k}}{k!}
= \frac1{\Gamma_0}\, \frac1{(1-t^\alpha)^\gamma},\qquad t\in (0,1)\, . \]
When $\alpha \geq 1$ and $\beta \geq t_0$, then obviously $\Gamma(\alpha k+\beta) \geq \Gamma(\beta),\, k \in \mathbb N_0$. Then
\[ E_{\alpha, \beta}^\gamma(t^\alpha) \leq \frac1{\Gamma(\beta)} \,\sum_{k \geq 0} \frac{(\gamma)_k}{(\beta)_k} \frac{z^k}{k!}\,,\]
which coincides with \eqref{N11}.
\end{proof}
Now, combining \eqref{N9}, \eqref{N10} and/or \eqref{N11} we arrive at
\begin{theorem} For all $\min\{\alpha, \beta, \gamma\}>0$ and for all $t \in (0,1)$ we have
\[ \Delta_T(t) \leq \frac1{\gamma \Gamma_0\, t^\alpha (1-t^\alpha)^{2\gamma+1}}\,
\left[ \frac1{\Gamma_1} + \frac{t^\alpha}{\Gamma_2\, (1-t^\alpha)}\right], \]
where $\Gamma_p = \min\limits_{n \in \mathbb N_p} \Gamma(\alpha n+\beta),\, p=0,1,2$; $\mathbb N_q = \{ q, q+1, q+2, \cdots \},
q \in \mathbb N_0$.
Moreover, for all $\alpha \geq 1,\, \beta \geq t_0,\, \gamma>0$ and for all $t>0$ there holds
\begin{align*}
\Delta_T(t) &\leq \frac{{}_1F_1(\gamma; \beta; t^\alpha)}{\gamma\,\Gamma(\beta)\, t^\alpha} \,
\left[ \frac1{\Gamma(\alpha+\beta)}\,{}_1F_1(\gamma+1; \alpha+\beta; t^\alpha) \right. \\
&\qquad + \left. \frac{t^\alpha }{\Gamma(2\alpha+\beta)}\,{}_1F_1(\gamma+2; 2\alpha+\beta; t^\alpha)\right]\,,
\end{align*}
where $(t_0, \Gamma(t_0))$ is the positive minimum of the $\Gamma$--function.
\end{theorem}
\section{Toward Laguerre type inequality results}
The Laguerreian difference (see e.g. the classical paper \cite{Skovgaard} or the recent works by Baricz) concerns the
subsequent derivatives of certain entire function $F$ of special type. It is
\[ \Delta_L^n (t) = [F^{(n)}(t)]^2 - F^{(n-1)}(t)\,F^{(n+1)}(t), \qquad n \in \mathbb N\,.\]
The estimate $\Delta_L^n (t) \geq 0, \, t\in I \subseteq \mathbb R$ we call Laguerre inequality of order $n$.
In turn, taking $t \mapsto t^{\frac1\alpha}$ in the Prabhakar type model $X \sim {\rm ML}(\alpha, \beta, \gamma)$ we get a
transformed random variable $X_1$, say. Since the characteristic function (CHF) of $X_1$ reads
\[ \varphi_t(x) = \mathbb E {\rm e}^{{\rm i}x X_1} = \frac{E_{\alpha, \beta}^\gamma\left(t{\rm e}^{{\rm i}x}\right)}
{E_{\alpha, \beta}^\gamma(t)}\,, \qquad (t, x) \in \mathbb R_+ \times \mathbb R \,,\]
which implies
\begin{align} \label{O1}
\mathbb EX_1 &= -{\rm i}\, \frac{\partial \varphi_t(x)}{\partial x}\Big|_{x=0}
= -\frac{\rm i}{E_{\alpha, \beta}^\gamma(t)}\, \frac{\partial}{\partial x}\, E_{\alpha, \beta}^\gamma
\left(t{\rm e}^{{\rm i}x}\right)\Big|_{x=0} \notag \\
&= \gamma t \, \frac{E_{\alpha, \alpha+\beta}^{\gamma+1}(t)}{E_{\alpha, \beta}^\gamma(t)}
= \frac{t}{E_{\alpha, \beta}^\gamma(t)}\,\left(E_{\alpha, \beta}^\gamma(t)\right)'\,,
\end{align}
from which we infer
\begin{equation} \label{O2}
\frac{\partial}{\partial x}\, E_{\alpha, \beta}^\gamma\left(t{\rm e}^{{\rm i}x}\right)\Big|_{x=0}
= {\rm i}\,t\,\left(E_{\alpha, \beta}^\gamma(t)\right)'\,.
\end{equation}
Similarly we have
\begin{align} \label{O3}
\mathbb EX_1^2 &= -\frac{\partial^2 \varphi_t(x)}{\partial x^2}\Big|_{x=0}
= - \frac1{E_{\alpha, \beta}^\gamma(t)}\,
\frac{\partial^2}{\partial x^2}E_{\alpha, \beta}^\gamma\left(t{\rm e}^{{\rm i}x}\right) \Big|_{x=0} \notag \\
&= \frac1{E_{\alpha, \beta}^\gamma(t)} \left\{ \gamma t\, E_{\alpha, \alpha+\beta}^{\gamma+1}(t)
+ \gamma(\gamma+1)t^2\, E_{\alpha, 2\alpha+\beta}^{\gamma+2}(t)\right\}\,,
\end{align}
and since
\begin{equation} \label{X1}
\left(E_{\alpha, \beta}^\gamma(t)\right)'' = \gamma(\gamma+1) \, E_{\alpha, 2\alpha+\beta}^{\gamma+2}(t)\,,
\end{equation}
we conclude
\begin{equation} \label{O4}
\frac{\partial^2}{\partial x^2}E_{\alpha, \beta}^\gamma\left(t{\rm e}^{{\rm i}x}\right) \Big|_{x=0}
+ t\,\left(E_{\alpha, \beta}^\gamma(t)\right)' + t^2\,\left(E_{\alpha, \beta}^\gamma(t)\right)'' = 0 \, .
\end{equation}
Now, we are ready to formulate our next set of results.
\begin{theorem} For all $\min\{\alpha, \beta, \gamma\}>0$ and $t>0$ there holds the first order Laguerre inequality:
\begin{equation} \label{O5}
\Delta_L(t) \equiv \left[\frac{\partial}{\partial x}E_{\alpha, \beta}^\gamma\left(t{\rm e}^{{\rm i}x}\right) \Big|_{x=0}\right]^2
- E_{\alpha, \beta}^\gamma(t) \cdot \frac{\partial^2}{\partial x^2}E_{\alpha, \beta}^\gamma
\left(t{\rm e}^{{\rm i}x}\right) \Big|_{x=0} \geq 0,
\end{equation}
Moreover, for the Laguerre difference the following functional upper bounds hold
\begin{equation} \label{O6}
\Delta_L(t) \leq \begin{cases}
\dfrac{\gamma}{ \Gamma_0\Gamma_1\, t(1-t)^{2\gamma+1}} \hspace{2.9cm}
\min\{\alpha, \beta, \gamma\}>0,\, t\in (0,1) \\ \\
\dfrac{\gamma\,{}_1F_1(\gamma; \beta; t)\,{}_1F_1(\gamma+1; \alpha+\beta; t)}
{t\,\Gamma(\beta)\, \Gamma(\alpha+\beta)} \qquad \alpha \geq 1,\, \beta \geq t_0;\, t>0
\end{cases}\,,
\end{equation}
where $\Gamma_0, \Gamma_1, t_0$ remain unchanged from {\rm Theorem 3}.
\end{theorem}
\begin{proof} By the moment inequality $\mathbb EX_1^2 \geq \big[\mathbb EX_1\big]^2$, \eqref{O1} and \eqref{O3} we deduce the
first assertion of the theorem. Next, replacing the partial derivatives in \eqref{O5} with the derivatives in $t$ from \eqref{O2}
and \eqref{O4} we immediately get
\begin{equation} \label{X2}
E_{\alpha, \beta}^\gamma(t) \left[ \left(E_{\alpha, \beta}^\gamma(t)\right)'
+ t\,\left(E_{\alpha, \beta}^\gamma(t)\right)''\right] \geq t\,\left[ \left(E_{\alpha, \beta}^\gamma(t)\right)' \right]^2\,,
\end{equation}
or equivalently
\[ t\,\Delta_L(t) \leq E_{\alpha, \beta}^\gamma(t) \cdot \left(E_{\alpha, \beta}^\gamma(t)\right)'
= \gamma \, E_{\alpha, \beta}^\gamma(t) \cdot E_{\alpha, \alpha+\beta}^{\gamma+1}(t)\, .\]
Now, applying the estimates \eqref{N10}, \eqref{N11} (having in mind to use $t \mapsto t^{\frac1\alpha}$) to the
right--hand--side expression respectively, we arrive at the second statement \eqref{O6} of the theorem.
\end{proof}
\begin{remark}
We point out that \eqref{N9} is in fact equivalent to \eqref{X2}. \hfill $\Box$
\end{remark}
Finally it is worth to mention two recurrence properties of the Prabhakar function's $t$--derivatives.
\begin{proposition} For all $\min\{\alpha, \beta, \gamma\}>0$ and $t>0$ there hold:
\begin{align*}
\alpha\gamma \,t\, E_{\alpha, \alpha+\beta+1}^{\gamma+1}(t) &= E_{\alpha, \beta}^\gamma(t) - \beta\, E_{\alpha, \beta+1}^\gamma(t)\\
\alpha^2 \gamma (\gamma+1)\,t^2\, E_{\alpha, 2\alpha+\beta+2}^{\gamma+2}(t) &= E_{\alpha, \beta}^\gamma(t)
- (\alpha+2\beta+1)\, E_{\alpha, \beta+1}^\gamma(t)\\
&\qquad + (\alpha+\beta+1)(\beta+1)\, E_{\alpha, \beta+2}^\gamma(t)\, .
\end{align*}
\end{proposition}
\begin{proof} Indeed, by direct calculation we show the identity
\[ \alpha \,t\, \left(E_{\alpha, \beta+1}^\gamma(t)\right)' = E_{\alpha, \beta}^\gamma(t) - \beta\, E_{\alpha, \beta+1}^\gamma(t).\]
Since $\big(E_{\alpha, \beta+1}^\gamma(t)\big)' = \gamma E_{\alpha, \alpha+\beta+1}^{\gamma+1}(t)$ (see \eqref{O1}), we have
\[ \alpha \gamma\,t\, E_{\alpha, \alpha+\beta+1}^{\gamma+1}(t) = E_{\alpha, \beta}^\gamma(t) - \beta\, E_{\alpha, \beta+1}^\gamma(t).\]
Differentiating this equation, previously adapted with $\beta \mapsto \beta+1$ and multiplying with $\alpha t>0$, taking into
account \eqref{X1}, we get
\begin{align*}
\alpha^2 \gamma\, t^2\,\left(E_{\alpha, \alpha+\beta+2}^{\gamma+1}(t)\right)'
&= \alpha^2 \gamma(\gamma+1)\,t^2\, E_{\alpha, 2\alpha+\beta+2}^{\gamma+2}(t) \\
&= E_{\alpha, \beta}^\gamma(t) - (\alpha+2\beta+1)\, E_{\alpha, \beta+1}^\gamma(t) \\
&\qquad + (\alpha+\beta+1)(\beta+1)\, E_{\alpha, \beta+2}^\gamma(t) \,,
\end{align*}
which finishes the proof of the second recurrence.
\end{proof}
\section*{Acknowledgement}
\v Zivorad Tomovski acknowledges NWO grant 2015/2016, Department of Applied Mathematics, TU Delft.
|
\section{Introduction} \label{sec:intro}
One path to addressing the hierarchy problem is to consider extensions of the standard model that lessen the degrees of divergence
of loop integrals. Historically, supersymmetry has been the most popular approach of this type. Loop diagrams involving the supersymmetric partners of ordinary particles cancel
the quadratic divergence of the Higgs boson squared mass that would otherwise be present. The surviving dependence on any high mass scales in the theory is only logarithmic, so that
extreme fine tuning is avoided. In Lee-Wick theories~\cite{lworig}, in particular the Lee-Wick Standard Model~\cite{lwsm}, more convergent loop diagrams are assured by the introduction of
higher-derivative kinetic terms that yield propagators that fall off more quickly with momentum. However, a propagator whose inverse is a higher-order polynomial in the momentum will have additional
poles. This fact is reflected in an auxiliary field formulation of Lee-Wick theories in which higher-derivative terms are absent, but additional field are present that correspond to these Lee-Wick partner states~\cite{lwsm}. Diagrams involving the Lee-Wick partner particles serve to cancel unwanted quadratic divergences, and hence play a role similar to the partner particles in supersymmetric theories.
Among the scenarios with partner particles that address the hierarchy problem, Lee-Wick theories are particularly unusual. The partner states in Lee-Wick theories have wrong-sign kinetic and mass
terms, requiring special rules to be applied so that the theory has a chance at a sensible interpretation~\cite{clop}. However, such states need not appear in all theories with with higher-derivative quadratic
terms~\cite{nopoles}. Given the possibility of applications in addressing the hierarchy problem~\cite{Biswas:2014yia}, it is well motivated to consider higher-derivative theories in which the complications of Lee-Wick theories might
be avoided altogether.
As an example of the type of theory of interest here, consider
\begin{equation}
{\cal L} = - \frac{1}{2} \phi \, \hat{F}(\Box)^{-1} (\Box+m^2) \, \phi - \frac{\lambda}{4!} \, \phi^4 \,\,\, ,
\label{eq:toy}
\end{equation}
where $\phi$ is a real scalar field, $\Box \equiv \partial_\mu \partial^\mu$ and the momentum-space propagator is given at tree-level by
\begin{equation}
\tilde{D}_F(p) = \frac{i \, \hat{F}(-p^2)}{p^2 -m^2 + i \epsilon} \,\,\, .
\label{eq:toy2}
\end{equation}
If $\hat{F}(-p^2)$ is an entire function, then there will be no additional poles in Eq.~(\ref{eq:toy2}), aside from the one at $p^2=m^2$.
If $\hat{F}$ is a transcendental function (rather than a polynomial, which is also entire), then we can find forms that drop off at large momentum. In particular, we will
focus on the simple choice
\begin{equation}
\hat{F}(\Box) = \exp(-\eta\stackrel{}{\Box}^n) \,\, ,
\label{eq:toy3}
\end{equation}
where $\eta>0$ is a coupling constant, and $n$ is a positive, even integer. (We restrict ourselves to even $n$ so that $\hat{F}$ provides a
convergence factor in either Euclidean or Minkowski space.) This theory is nonlocal. The consequences of nonlocal modifications
of the quadratic terms in the Lagrangian were discussed as early as the 1950's~\cite{Pais:1950za}, but have periodically met a resurgence of
interest~\cite{Barnaby:2007yb,Tomboulis:1997gg,Modesto:2011kw,Biswas:2011ar,Biswas:2013cha,Modesto:2014xta,Modesto:2014lga,Biswas:2014yia,Modesto:2015foa,Addazi:2015ppa,Tomboulis:2015gfa}.
Motivated by the infinite-derivative Lagrangians obtained in string field theory~\cite{sft} and $p$-adic string theory~\cite{padic}, nonlocal theories of the general
type of interest here were studied as possible models of inflation~\cite{Barnaby:2007yb}. More recently, the possibility that such nonlocal quadratic terms could provide an avenue for quantizing gravity has also been discussed~\cite{Tomboulis:1997gg,Modesto:2011kw,Biswas:2011ar,Biswas:2013cha,Modesto:2014xta,Modesto:2014lga}. Of particular motivation here is the work of Ref.~\cite{Biswas:2014yia} which applies nonlocal modifications of the quadratic terms to parts of the standard model itself and considers some aspects of the phenomenology. A more extensive list of background references on nonlocal field theories and their applications can be found in that work.
Ref.~\cite{Biswas:2014yia}, like most phenomenological studies of proposed modifications to the standard model, ultimately focuses on scattering
processes, which reflect the overlap of asymptotic states defined in the far past and far future. In the context of Lee-Wick theories, it was pointed out
by Grinstein, O'Connel and Wise (GOW) that the distinctive acausal features of the theory could be studied by considering the time-dependence of
the scattering processes via a wave-packet analysis conducted in the semi-classical limit~\cite{Grinstein:2008bg}, as we discuss in more detail later.
The trajectories of wave packets can be used to define the apparent production and decay points of an exchanged resonance, and the dependence
of the amplitude on the ordering of these events evaluated. GOW worked with a theory of real scalar fields with O($N$) symmetry, where the unitarity
of the theory could be demonstrated to all orders in perturbation theory in the large $N$ limit. As argued by
Coleman~\cite{acoleman}, the existence of a unitary $S$-matrix implies that observable acausality does not lead to logical paradoxes in scattering
experiments, since there is a unitary evolution of initial states to final states. The question that we wish to study in the present work is how the
approach and conclusions of GOW are modified if one instead assumes a theory with an infinite-derivative modification of the quadratic terms, one
that does not introduce additional poles with wrong-sign residues in the propagator at tree-level.
We proceed largely by analogy, first addressing the issue of unitarity in a specific nonlocal extension of the O($N$) model studied by GOW. Unitarity in nonlocal theories has been discussed in a more general context in Refs.~\cite{Tomboulis:1997gg,Tomboulis:2015gfa,Alebastrov:1973vw} and in a different context in Ref.~\cite{Addazi:2015ppa}. What we gain by working in the large $N$ limit of the O($N$) model is that unitarity can be studied in explicit detail, to all orders in perturbation theory, via a one-loop calculation. In addition, the intermediate steps and final conclusions can be readily compared to those of
Ref.~\cite{Grinstein:2008bg}. The reader who is familiar with the phenomenological work of Ref.~\cite{Biswas:2014yia} will recall that the authors define their nonlocal theory via Euclidean correlation functions that are analytically continued in their external momentum to Minkowski space. If one were to attempt to quantize the theory in Minkowski space directly, one would find that unitary is violated. The calculation that we present in Sec.~\ref{sec:unitarity} will make clear why this is the case. We then turn to the issue of causality in Sec.~\ref{sec:causality}. It is generally expected that the nonlocal theories having the form
shown in Eq.~(\ref{eq:toy}) have field commutators that do not vanish at space-like separation~\cite{Tomboulis:2015gfa}. We show that this is indeed the case in the specific O($N$) model defined in Sec.~\ref{sec:unitarity} by an explicit calculation. What our consideration of unitarity and causality demonstrate up to this point is that the theory of interest may show signs of acausality in scattering experiments without logical inconsistency, in the sense discussed by Coleman. To address this further, we turn to the scattering of wave packets in the latter half of Sec.~\ref{sec:causality}, and show that there is a non-vanishing amplitude for acausal orderings of production and decay vertices for exchanged resonances. Unlike the Lee-Wick case, where the resonance is a Lee-Wick partner with wrong-sign kinetic and mass terms, the resonances in this case are ordinary particles. In Sec.~\ref{sec:conc} we summarize our conclusions.
The explicit calculations that we present in this work, as well as the discussion of the nonlocal O($N$) model and the detailed application of the approach of Ref.~\cite{Grinstein:2008bg} to similar theories, have not appeared in the
literature previously. These may serve as a useful complement to more formal treatments that anticipate the qualitative features of some of our results. Moreover, the explicit examples and calculations that we present may resonate
with a wider audience of model-builders who are interested in phenomenological applications relevant to TeV-scale physics, an exploration that has been quite limited thus far~\cite{Biswas:2014yia}.
\section{Unitarity} \label{sec:unitarity}
\subsection{Preliminaries}
In the absence of higher-derivative modifications, we work with a theory of $N$ real scalar fields with the Lagrangian density
\begin{equation}
{\cal L} = \frac{1}{2} \partial_\mu \phi^a \partial^\mu\phi^a -\frac{1}{2} m^2 \,\phi^a\phi^a -\frac{1}{8} \lambda_0 (\phi^a\phi^a)^2 \,\,\, .
\label{eq:on1}
\end{equation}
This theory has an O($N$) global symmetry, with the index $a$ running from $1$ to $N$. The theory has a sensible $N \rightarrow \infty$ limit, {\em i.e.}, there are no Feynman
diagrams that grow as positive powers of $N$, if the coupling $\lambda_0$ scales as $1/N$. (For a pedagogical discussion, see Ref.~\cite{aspects}.) It is convenient to redefine
the coupling $\lambda_0 \equiv \lambda/N$, so that the $N$ dependence of a given amplitude is explicit. Following Ref.~\cite{Grinstein:2008bg}, the theory in Eq.~(\ref{eq:on1}) is
equivalent to
\begin{equation}
{\cal L} = \frac{1}{2} \partial_\mu \phi^a \partial^\mu\phi^a -\frac{1}{2} m^2 \,\phi^a\phi^a + \frac{N}{2 \lambda} \sigma^2 -\frac{1}{2} \sigma \phi^a \phi^a \,\,\, ,
\label{eq:on2}
\end{equation}
where $\sigma$ is an auxiliary field; this can be verified by substitution of the auxiliary field's equation of motion into Eq.~(\ref{eq:on2}). The advantage of working
with the auxiliary field formulation is that it makes counting of powers of $N$ transparent, since each $\sigma$ propagator scales as $1/N$. For example, the self-energy function
for the $\sigma$ field, $\Sigma_0 (p^2)$, receives it's leading order contribution from a $\phi^a$ loop, and scales as $N$. Following the sign conventions of
Ref.~\cite{Grinstein:2008bg}, the full $\sigma$ propagator is given by
\begin{equation}
\tilde{D}(p^2) = \frac{i}{1/\lambda_0} + \frac{i}{1/\lambda_0} \left(i \Sigma_0 \right) \frac{i}{1/\lambda_0}
+ \frac{i}{1/\lambda_0} \left( i \Sigma_0 \right) \frac{i}{1/\lambda_0} \left(i \Sigma_0 \right) \frac{i}{1/\lambda_0} + \cdots \,\,\, ,
\end{equation}
which can be re-summed to
\begin{equation}
\tilde{D}(p^2) = \frac{\lambda}{N} \frac{i}{1+\lambda \, \Sigma(p^2)} \,\,\, ,
\label{eq:dprop}
\end{equation}
where $\Sigma_0(p^2) \equiv N \,\Sigma(p^2)$, so that the $N$-scaling of Eq.~(\ref{eq:dprop}) is explicit. All corrections to $\Sigma(p^2)$ that
are higher than one-loop are suppressed by additional factors of $1/N$, by virtue of the additional $\sigma$ propagators. Hence, if one is
interested in only the leading-order behavior of $\Sigma(p^2)$, one only needs to compute a one-loop diagram\footnote{At leading
order there is also a one-loop $\sigma$ tadpole diagram, but it can be eliminated by a shift in the auxiliary field and a redefinition of the $\phi^a$ mass~\cite{aspects}.}.
At leading order in $1/N$, two-into-two scattering in the auxiliary field formulation corresponds to the $s$-, $t$- and $u$- channel exchanges of
the auxiliary field, with the dressed propagator given by Eq.~(\ref{eq:dprop}). All other loop corrections to the scattering amplitude involve additional
$\sigma$ propagators and are sub-leading in the $1/N$ expansion. It follows that the scattering amplitude is given by
\begin{equation}
{\cal M}(ab\rightarrow cd) = -\frac{\lambda}{N} \left[ \frac{\delta_{ab} \delta_{cd}}{1 + \lambda \Sigma(s)}
+ \frac{\delta_{ac} \delta_{bd}}{1 + \lambda \Sigma(t)}
+\frac{ \delta_{ad} \delta_{bc} }{1 + \lambda \Sigma(u)} \right]
\,\,\, ,
\label{eq:scamp1}
\end{equation}
where $s$, $t$ and $u$ are the usual Mandelstam invariants. As reviewed in Ref.~\cite{Grinstein:2008bg}, Eq.~(\ref{eq:scamp1}) can be used to demonstrate the unitarity of
the theory at leading order in $1/N$ and at all orders in perturbation theory.
Our present interest is how this calculation is altered when there is a nonlocal modification to Eq.~(\ref{eq:on1}), of either the form
\begin{equation}
{\cal L} =
- \frac{1}{2} \phi^a \, \hat{F}(\Box)^{-1} (\Box+m^2) \, \phi^a -\frac{1}{8} \lambda_0 (\phi^a\phi^a)^2 \,\,\, ,
\label{eq:nloc2}
\end{equation}
or
\begin{equation}
{\cal L} = - \frac{1}{2} \phi^a \, (\Box+m^2) \, \phi^a
-\frac{1}{8} \lambda_0 \left[(\hat{F}^{1/2}\phi^a)( \hat{F}^{1/2}\phi^a)\right]^2 \,\,\, .
\label{eq:nloc1}
\end{equation}
Here $\hat{F}$ is the differential operator defined in Eq.~(\ref{eq:toy3}), with $\eta>0$ and $n$ and even positive integer, and
$\hat{F}^{1/2} \equiv \exp(-\frac{1}{2} \eta\!\stackrel{}{\Box}^n)$. We choose $n$ even
so that the factors of $\hat{F}$ lead to improved convergence of loop integrals in momentum space, regardless of whether we assume
a Euclidean or Minkowski metric. We compare each possibility in the following subsection, for the simplest choice of $n=2$, which we assume henceforth.
Eqs.~(\ref{eq:nloc2}) and (\ref{eq:nloc1}) are related by a nonlocal field redefinition and give the same results for scattering
amplitudes. Working with Eq.~(\ref{eq:nloc1}), the factors of $\hat{F}^{1/2}$ acting on internal lines reproduce the momentum dependence of the propagator that
one obtains from Eq.~(\ref{eq:nloc2}); the factors of $\hat{F}^{1/2}$ acting on external lines each give a factor of $\exp(-\eta \, m^4/2)$, matching the wave function renormalization factors in the
scattering amplitudes obtained from Eq.~(\ref{eq:nloc2}). For definiteness, we will examine the case
where $a=b \neq c=d$ so that only the $s$-channel amplitudes is relevant. Then the scattering amplitude takes the form
\begin{equation}
{\cal M} = - \frac{\lambda}{N} \frac{e^{-2 \eta \, m^4}}{1 + \lambda\, \Sigma(s)} \, \delta_{ab} \delta_{cd} \,\,\, ,
\label{eq:ourm}
\end{equation}
where the constant exponential factor is due to the higher-derivative operator acting on the external lines, and where $\Sigma(s)$ now includes the
effects of $\hat{F}$ on the $\phi^a$ propagator.
\subsection{Minkowski Space}
We show in this section that the theory defined in Minkowski space by Eq.~(\ref{eq:nloc1}) violates unitarity. The self-energy function $\Sigma(p^2)$ is
given by
\begin{equation}
\Sigma(p^2) = -\frac{i}{2} \int\frac{d^4k}{(2 \pi)^4} \frac{\exp\{-\eta \, (k+p/2)^4\}\exp\{-\eta \,(k-p/2)^4\}}{[(k-p/2)^2 -m^2 + i \epsilon] [(k+p/2)^2-m^2 + i \epsilon]} \,\,\, .
\label{eq:sigmink}
\end{equation}
Unitarity implies the operator relation $i (T^\dagger-T) = T^\dagger T$, where the $T$-matrix is related to the $S$-matrix by $S=1+i T$. One can derive a condition on
scattering amplitudes by taking matrix elements of both sides of this expression and including an appropriate insertion of a complete set of intermediate states. Working at leading
order in the $1/N$ expansion, this procedure was carried out in the O(N) model in Ref.~\cite{Grinstein:2008bg}, and the derivation is not altered by the presence of the
additional momentum space suppression factors in the numerator of Eq.~(\ref{eq:sigmink}). One finds~\cite{Grinstein:2008bg}
\begin{eqnarray}
2 \mbox{ Im}&\!\!{\cal M}& \!\!(k_1,a; k_2,b \rightarrow k_1',c;k_2',d) =
\sum_{e,f} I_{e,f} \int \frac{d^3 q_1}{(2 \pi)^3} \frac{d^3 q_2}{(2 \pi)^3} \frac{1}{2 E_1}\frac{1}{2 E_2} (2 \pi)^4 \delta^{(4)}(q_1 + q_2 -p) \nonumber \\ &&
{\cal M}(k_1,a; k_2,b \rightarrow q_1,e ; q_2, f) \, {\cal M}^*(k_1', c; k_2' , d \rightarrow q_1, e; q_2, f) \,\,\, ,
\label{eq:opthm}
\end{eqnarray}
where the identical particle factor $I_{e,f}=1/2$ if $e=f$ and $1$ otherwise. The left-hand-side of this expression follows immediately from Eq.~(\ref{eq:ourm}):
\begin{equation}
LHS = \frac{\lambda^2}{N} \left[2 \,e^{-2 \, \eta \, m^4} \mbox{ Im }\Sigma(s) \right] \frac{1}{|1+\lambda\, \Sigma(s)|^2} \,\delta_{ab} \delta_{cd}\,\,\, .
\label{eq:lhs}
\end{equation}
The right-hand-side of Eq.~(\ref{eq:opthm}) includes only two-particle intermediate states, which provide the leading contribution in the large $N$ limit. After
substitution of Eq.~(\ref{eq:ourm}), the necessary integral evaluation is identical to that of the two-body Lorentz-invariant phase space factor. The result is
\begin{equation}
RHS = \frac{\lambda^2}{N} \left[ \frac{1}{16 \pi} e^{-4 \, \eta \, m^4} \sqrt{1-\frac{4 m^2}{s}} \,\right] \frac{1}{|1+\lambda\, \Sigma(s)|^2} \,\delta_{ab} \delta_{cd} \,\,\, .
\label{eq:rhs}
\end{equation}
When $\eta=0$, the quantities in brackets in Eqs.~(\ref{eq:lhs}) and (\ref{eq:rhs}) coincide, as can be seen either from an elementary one-loop calculation~\cite{Grinstein:2008bg}, or by examining the $\eta \rightarrow 0$ limit of the numerical calculation that we are about to describe. When
$\eta \neq 0$, these quantities differ. After exploring the source of the discrepancy, we show how it is avoided by defining the theory as an
analytic continuation of correlation functions defined in Euclidean space.
It is easiest to see that Eqs. (\ref{eq:lhs}) and (\ref{eq:rhs}) do not agree when $\eta \neq 0$ by showing that $\mbox{Im } \Sigma (p^2)$
no longer has a functional form proportional to $\sqrt{1- 4\, m^2/ s}$. To confirm this claim most quickly, we simply evaluate the imaginary part of Eq.~(\ref{eq:sigmink}) numerically, working in the center-of-mass frame, where $\vec{p}=0$; we perform the $k^0$ integral exactly along the
real axis with $\epsilon$ finite and evaluate the limit as $\epsilon \rightarrow 0$. Note that the imaginary part of the loop integral is finite, even when
$\eta$ is vanishing. It is convenient to write Eq.~(\ref{eq:sigmink}) in the following form:
\begin{equation}
\mbox{Im } \Sigma =-\frac{1}{4 \pi^3} \int_0^\infty dk \int_0^\infty dk^0 \left\{ k^2 g(k^0,k)
\frac{f_+(k^0,k) f_-(k^0,k) - \epsilon^2}{\left[f_+(k^0,k)^2+\epsilon^2\right] \left[f_-(k^0,k)^2+\epsilon^2\right]} \right\} \,\,\, ,
\end{equation}
where $k \equiv |\vec{k}|$,
\begin{equation}
f_\pm(k^0,k) \equiv (k^0 \pm p^0/2)^2-k^2-m^2 \,\,\, ,
\end{equation}
and
\begin{equation}
g(k^0,k) \equiv \exp\{-\eta [(k^0 + p^0/2)^2-k^2]^2-\eta [(k^0 - p^0/2)^2-k^2]^2 \} \,\,\, .
\end{equation}
The integration can be performed using symbolic mathematics code (we used MAPLE~\cite{maple}), provided care is taken in dealing with the points on the real
$k^0$ axis that would be singularities in the $\epsilon \rightarrow 0$ limit. For $\epsilon$ small but non-zero, the growth of the integrand around these points
are taken into account by singularity handling routines in MAPLE that are invoked automatically by breaking up the region of $k^0$ integration into intervals that are terminated at these points. We then have no difficulty obtaining numerically convergent results. In Fig.~\ref{fig:ufig}, we show the result for $\mbox{Im } \Sigma$ as a
function of the center-of-mass energy, working in units where $m=1$, for $\eta=0$ and an example where $\eta \neq 0$. The line with long dashes shows the
expectation for $\mbox{Im } \Sigma$ following from the analytic result of the one-loop calculation in the $\eta=0$ case,
\begin{equation}
\mbox{Im } \Sigma (s) = \frac{1}{32 \pi} \sqrt{1-\frac{4 m^2}{s}} \theta(s -4 m^2) \,\,\, ,
\end{equation}
where $\theta$ is the Heaviside step function. This agrees with the numerical result for $\eta=0$, given by the open circular points in Fig.~\ref{fig:ufig}.
However, the results are not proportional to the same functional form in $s$ for the case where $\eta \neq 0$. One would not suspect that the disagreement is the
result of a numerical artifact, since the extra exponential factor in the integrand in the case where $\eta \neq 0$ is smooth and serves primarily to
truncate the domain of integration.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.5\textwidth]{ufig.eps}
\caption{ Imaginary part of the self-energy functions $\Sigma(s)$ as a function of $\sqrt{s}$. The open circular points indicate the results of the direct
numerical evaluation of Eq.~(\ref{eq:sigmink}), for the cases where $\eta=0$ and $\eta=0.3$. The long dashed line gives
the $\eta=0$ expectation, proportional to $(1-4 \, m^2 /s)^{1/2}$ for $s >4 m^2$. The solid points are
the result of Eqs.~(\ref{eq:res1}), after a numerical evaluation of the second term, as discussed in the text.}
\label{fig:ufig}
\end{center}
\end{figure}
To further verify this result, let us now do the calculation in a different way. Imagine we evaluate the $k^0$ integral in $\Sigma$ by closing a semi-circular contour in the lower-half complex plane. In ordinary, local
quantum field theories, the integral along the semi-circular contour would vanish as the radius of the contour is taken to infinity. In the present theory, this is not the case; the numerator of the loop integral becomes
$\exp[-2 \, \eta\, (k^0)^4]$ far from the origin, which blows up in directions where $\mbox{Re } (k^0)^4 <0$. Hence, let us decompose
\begin{equation}
\Sigma = \Sigma_p - I_C \,\,\,\, \mbox{ and } \,\,\,\, 2 \, \mbox{Im} \, \Sigma = 2\, \mbox{Im} \, \Sigma_p - 2\, \mbox{Im} \, I_C \,\,\, ,
\label{eq:decomp}
\end{equation}
where $\Sigma_p$ is $-2 \pi i$ times the residues of the poles contained within the contour and $I_C$ is the clockwise integral along the semi-circular portion. Since
\begin{equation}
\Sigma_p(p^0) = \frac{1}{2} \int \frac{d^3 k}{(2 \pi)^3} \frac{1}{2 E_{\vec{k}}\,p^0} \left[ \frac{N(k^0 = E_{\vec{k}} - p^0/2)}{2 E_{\vec{k}} - p^0}
-\frac{N(k^0 = E_{\vec{k}} + p^0/2)}{2 E_{\vec{k}} + p^0} \right] \,\,\, ,
\label{eq:ppart}
\end{equation}
where $N$ represents the numerator of the integrand of Eq.~(\ref{eq:sigmink}) and $E^2_{\vec{k}}\equiv |\vec{k}|^2+m^2$, it is clear
for $0<p^0<2 m$ on the real axis that $\Sigma_p(p^0) = \Sigma_p(p^{0*})^*$. Since $\Sigma_p$ is an analytic function of $p^0$ off the real axis, this can be analytically continued, from which it follows that
\begin{equation}
\mbox{Disc} \, \Sigma_p = 2 \, i \, \mbox{Im} \,\Sigma_p (p^0+i \epsilon) \,\,\, ,
\label{eq:discim}
\end{equation}
where the discontinuity is about a cut on the real $p^0$ axis,
\begin{equation}
\mbox{Disc}\, \Sigma_p = \lim_{\epsilon \rightarrow 0} \left[ \Sigma_p(p^0+i \epsilon) -\Sigma_p(p^0-i\epsilon) \right] \,\,\, .
\end{equation}
We may express
\begin{eqnarray}
\Sigma_p(p^0+i \epsilon) &=&\frac{1}{8 \pi^2} \int d E_{\vec{k}} \frac{ |\vec{k}|}{p^0 + i \epsilon} \nonumber \\
&& \times
\left[ \frac{N(k^0 = E_{\vec{k}} - p^0/2 -i \epsilon/2)}{2 E_{\vec{k}} - p^0-i \epsilon}-\frac{N(k^0 = E_{\vec{k}} + p^0/2+i \epsilon/2)}{2 E_{\vec{k}} + p^0+i \epsilon} \right] \,\,\, .
\end{eqnarray}
There can be poles in the $E_{\vec{k}}$ integration that blow up at most as $1/\epsilon$; hence, we only need expand what remains
to order $\epsilon$. Defining $N' \equiv \partial N / \partial k^0$, we find
\begin{equation}
\begin{split}
\mbox{Disc}\, \Sigma_p & = - \frac{1}{8\pi^2} \lim_{\epsilon \rightarrow 0} \int d E_{\vec{k}} \frac{|\vec{k}|}{p^0} \\
& \left\{ N(k^0 = k^0_-) \left[\frac{1}{p^0-2 E_{\vec{k}} +i \epsilon}- \frac{1}{p^0-2 E_{\vec{k}} - i \epsilon}\right] \right. \\
&+ N(k^0 = k^0_+) \left[\frac{1}{p^0+2 E_{\vec{k}} +i \epsilon}- \frac{1}{p^0+2 E_{\vec{k}} - i \epsilon}\right] \\
& +\left[\frac{-i}{2} N'(k^0 = k^0_-)-\frac{i}{p^0}N(k^0 = k^0_-)\right]
\left[\frac{\epsilon}{p^0-2 E_{\vec{k}} +i \epsilon}+ \frac{\epsilon}{p^0-2 E_{\vec{k}} - i \epsilon}\right] \\
& \left.+\left[\frac{i}{2} N'(k^0 = k^0_+)-\frac{i}{p^0}N(k^0 = k^0_+)\right] \left[\frac{\epsilon}{p^0+2 E_{\vec{k}} +i \epsilon}+ \frac{\epsilon}{p^0+2 E_{\vec{k}} - i \epsilon}\right] \right\}
\end{split}
\end{equation}
where $k^0_\pm = E_{\vec{k}} \pm p^0/2$. We can now take the $\epsilon \rightarrow 0$ limits of the quantities in square brackets, using
\begin{equation}
\lim_{\epsilon \rightarrow 0} \left[\frac{1}{y+i \epsilon} - \frac{1}{y-i \epsilon} \right] = -2 \pi i\, \delta(y)
\,\,\, \mbox{ and } \,\,\,
\lim_{\epsilon \rightarrow 0} \left[\frac{\epsilon}{y+i \epsilon} + \frac{\epsilon}{y-i \epsilon} \right] = 2 \pi y \, \delta(y) \,\,\, .
\end{equation}
We see that the third and fourth terms in the curly braces are proportional to $(p^0 \pm 2 E_{\vec{k}}) \delta(p^0 \pm 2 E_{\vec{k}})$,
so that they vanish after integration. Since $p^0>0$, the surviving term is given by
\begin{equation}
\mbox{Disc}\, \Sigma_p = i \frac{1}{4 \pi} \int d E_{\vec{k}}\, \frac{|\vec{k}|}{p^0} \, \delta(p^0 - 2 E_{\vec{k}}) \, N(k^0=E_{\vec{k}}-p^0/2)
\end{equation}
It is straightforward to confirm that the same result is obtained by making the conventional Cutkosky replacements in
the original integral for $\Sigma_p$
\begin{equation}
\frac{1}{(k \pm p/2)^2-m^2 + i \epsilon} \rightarrow -2 \pi i \,\delta([k \pm p/2]^2-m^2)
\end{equation}
so that
\begin{equation}
i \, \mbox{Disc} \, \Sigma_p = \frac{1}{2} \int \frac{d^4 k }{(2 \pi)^4} N(k) (-2 \pi i)^2 \delta([k + p/2]^2-m^2)) \delta([k - p/2]^2-m^2) \,\,\, .
\end{equation}
Changing variables, introducing an additional integral, and writing out the numerator factor $N$, this is equivalent to
\begin{equation}
\begin{split}
i \, \mbox{Disc} \, \Sigma_p &= \\
-\frac{1}{2} &\int \frac{d^4 q_1}{(2 \pi)^4} \frac{d^4 q_2}{(2 \pi)^4} \, e^{(-\eta \, q_1^4-\eta \,q_2^4)} \,
(2 \pi) \delta(q_1^2-m^2) \, (2 \pi) \delta(q_2^2-m^2) \\
& (2 \pi)^4 \delta^{(4)}(q_1+q_2-p) \,\,\, ,
\end{split}
\end{equation}
which integrates to
\begin{equation}
i \, \mbox{Disc} \, \Sigma_p = -\frac{1}{16 \pi} e^{-2 \,\eta\, m^4} \sqrt{1-\frac{4 m^2}{s}} \,\,\, .
\end{equation}
From Eqs.~(\ref{eq:decomp}) and (\ref{eq:discim}), it follows that we can write the quantity in square brackets from
the left-hand-side of our unitarity relation, Eq.~(\ref{eq:lhs}), as
\begin{equation}
\left[ 2 \,e^{-2 \, \eta \, m^4} \mbox{ Im }\Sigma(s) \right]= \frac{1}{16 \pi} e^{-4 \,\eta\, m^4} \sqrt{1-\frac{4 m^2}{s}} - 2\, e^{-2 \,\eta\, m^4}
\, \mbox{Im} \, I_C \,\,\,.
\label{eq:res1}
\end{equation}
The first term agrees with the desired form of the quantity in square brackets in Eq.~(\ref{eq:rhs}); it follows that the violation of unitarity stems entirely
from the non-vanishing of the integral $I_C$ along the semi-circular contour.
We can verify that Eq.~(\ref{eq:res1}) is correct by evaluating the imaginary part of $I_C$ and comparing the result for
$\mbox{Im} \,\Sigma$ with what we obtained previously in Fig.~\ref{fig:ufig}. Notice that if we were to push all the poles on the real $k^0$ axis to
the upper half-plane, then $I_C$ would be given by the negative of the integral along the real axis. Hence, we may identify
\begin{equation}
I_C = \frac{i}{2} \int \frac{d^4 k}{(2 \pi)^4} \frac{N(k)}{\left[(k^0-p^0/2 - i \epsilon)^2 - E_{\vec{k}}^2\right]\left[(k^0+p^0/2-i \epsilon)^2- E_{\vec{k}}^2 \right]} \,\,\, .
\label{eq:icf2}
\end{equation}
The point is that Eq.~(\ref{eq:icf2}) can be evaluated numerically in exactly the same way as the integral in Eq.~(\ref{eq:sigmink}) that we described earlier. The result for
$\mbox{Im}\,\Sigma$ computed from Eq.~(\ref{eq:res1}) using the numerical evaluation of Eq.~(\ref{eq:icf2}) is indicated by the solid points shown in Fig.~\ref{fig:ufig}: they are in complete agreement
with our previous direct evaluation of $\mbox{Im}\,\Sigma$ in the case where $\eta \neq 0$.
To understand this result, it is useful to consider how the calculation might have proceeded had we started by evaluating the discontinuity of
Eq.~(\ref{eq:sigmink}) directly using Cutkosky's formula~\cite{Cutkosky:1960sp}. It is straightforward to check that the discontinuity computed in this way
would reconcile Eq.~(\ref{eq:lhs}) and Eq.~(\ref{eq:rhs}) only if $\mbox{Disc}\, \Sigma = 2 \, i \,\mbox{Im} \,\Sigma$. However, this relation is not justified
in the present case since $\Sigma$ cannot be shown to satisfy the Schwartz reflection principle $\Sigma(p^0) = \Sigma(p^{0*})^*$ when $\eta \neq 0$.
The reflection principle requires that there be a segment along the real $p^0$ axis over which $\Sigma$ is purely real; in the case where $\eta \neq 0$
it is not possible to prove that such a region exists and our numerical results shown in Fig.~\ref{fig:ufig} strongly suggest that the opposite is true. In
the Appendix, we show in more detail how the violation of the Scwartz reflection principle can be directly related to the non-vanishing of contour integrals,
like $I_C$, at large radius in the complex plane.
Another starting point~\cite{Tomboulis:1997gg} for attempts to demonstrate unitarity is the Largest Time Equation (LTE)~\cite{Veltman:1963th}. We simply note here that this approach
cannot be consistently applied to the present problem. As discussed by Anselmi~\cite{Anselmi:2016fid}, derivation of the LTE requires two
assumptions: (1) the vertices of the theory are localized time and (2) the propagator in position space is of the form
$\theta(x^0) g_+(x) + \theta(-x^0) g_-(x)$, where $\theta$ is the step function. If nonlocality appears in the vertices of the theory, then assumption (1) is violated. If a field redefinition is used to move the
nonlocality to the propagators, then assumption (2) is violated due to the appearance of additional terms in the propagator that are proportional to derivatives of $\delta(x^0)$. (The explicit form of the
propagator can be found in Ref.~\cite{Tomboulis:2015gfa}.) The subsequent derivation of the LTE described in Ref.~\cite{Anselmi:2016fid} fails. Hence, we say nothing further
about this approach.
\subsection{Euclidean Space}
We have discussed in the previous subsection how unitarity is violated if we attempt to formulate the theory of interest directly in Minkowski space. If correlation functions are
defined in Euclidean space and analytically continued in the external momenta to Minkowski space, unitarity is preserved. This might be expected since the theory quantized via
a Euclidean functional integral automatically satisfies reflection positivity. The way that the calculation of the previous section is modified is as follows: The Euclidean version of
$\Sigma$ corresponds to Eq.~(\ref{eq:sigmink}) with the $k^0$ integration taken along the imaginary axis, and with Euclidean external momentum $p^0=i p^0_E$. In other words,
the starting point is the path that one would obtain with a Wick rotation if it were justified in a Minkowski-space formulation of the theory. Now, close the contour with a semi-circle
in the right half plane, so that
\begin{equation}
2\, \mbox{Im} \, \Sigma = 2\, \mbox{Im}\, \Sigma_p - 2\, \mbox{Im}\, I_C' \,\,\, ,
\label{eq:decomp2}
\end{equation}
where $I_C'$ is the integral over the semi-circular path, and $\Sigma_p$ again is determined by the residues of the poles contained within the closed contour. While $\mbox{Im}\, I_C$ in our previous calculation was non-vanishing, we now show that $\mbox{Im}\, I_C'=0$. Let us write $I_C' = -i \int dk^0 X(k^0)$, where $X(k^0)$ is given by Eq.~(\ref{eq:sigmink}) with $\epsilon$ set to zero and $k^0$ placed on the desired semi-circle, $k^0 = \rho \exp( i \theta)$ for $-\pi/2 \leq \theta \leq \pi/2$. In the
center-of-mass frame where $\vec{p}=0$, it is straightforward to check that $X(k^0)$ is also a function of ${p^0}^2={-{p^0_E}^2}$, which is real; it follows immediately that
$X(k^0)^* = X(k^{0*})$. Since $dk^0 = i k^0 d\theta$,
\begin{equation}
2\, i \, \mbox{Im}\, I_C'=I_C' - I_C'^*= \lim_{\rho \rightarrow \infty} \int_{-\pi/2}^{\pi/2} d\theta \left[ k^0 X(k^{0}) - k^{0*} X(k^{0*}) \right] \,\,\, ,
\end{equation}
which vanishes; this can be seen by changing variables $\theta \rightarrow -\theta$, and noting that $k^0(-\theta) =k^{0*}$, indicating that the $d\theta$ integral is equal to its negative. The surviving term in Eq.~(\ref{eq:decomp2}) is the same function of $p^0$ that reconciled the left- and right-hand-sides of our unitarity relation in the previous section.
By Lorentz invariance, the result holds in any other reference frame in which the scattering process is measured. Hence, we have verified that the large $N$ scattering amplitudes of interest in the present context are unitary provided that the theory is defined as in Ref.~\cite{Biswas:2014yia}, via the analytic continuation to Minkowski-space external momentum of correlation functions defined in a Euclidean field theory. We will assume that correlation functions are computed in this way in the discussion that follows.
\section{Causality} \label{sec:causality}
Nonlocal theories of the type studied here were known long ago to violate causality~\cite{Pais:1950za}. In general, the commutator of fields at space-like
separation is expected to be non-vanishing for theories where $\hat{F}$ is an entire, transcendental function~\cite{Tomboulis:2015gfa}. We demonstrate this
in the case where $m=0$ in the unitary theory discussed in the previous section, a limit in which we can explicitly evaluate the commutator. We will then turn to
scalar theories with similar nonlocal modifications and consider how acausality affects the time-dependence of scattering amplitudes, following the general approach of Ref.~\cite{Grinstein:2008bg}.
\subsection{Commutator}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.5\textwidth]{complot.eps}
\caption{The commutator $C(\rho) = \langle 0 | [\phi(x) , \phi(y)] | 0 \rangle$, as a function of
$\rho = (x^0-y^0)^2-|\vec{x}-\vec{y}|^2$ at space-like separation $\rho<0$, for $m=0$ and $\eta=1$.}
\label{fig:com}
\end{center}
\end{figure}
The Feynman-propagator $D_F(x-y)$ is identified with the two-point correlation function $\langle 0 | T \phi(x) \phi(y) | 0 \rangle$. If we
strictly assume that $x^0>y^0$, then we may write the commutator
\begin{equation}
\langle 0 | [\phi(x) , \phi(y)] | 0 \rangle = D_F(x-y) - D_F(x-y)^* \,\,\,\,\,\,\,\,\,\, (x^0>y^0) \,\,\, .
\label{eq:dfdf}
\end{equation}
Working with the form of the theory in which the higher-derivative operator appears in the quadratic terms for $\phi$, it
follows immediately that\footnote{Since we work here with the lowest-order propagator, our prescription of starting
with a Euclidean correlation function and continuing to Minkowski space in the external momentum simply gives us the
usual momentum-space propagator. All subsequent Fourier transforms are, of course, in Minkowski space.}
\begin{equation}
D_F(x-y) = \int \frac{d^4 k}{(2 \pi)^4} \frac{i \, e^{-\eta \, k^4}}{k^2 -m^2 +i \epsilon} e^{-i k \cdot (x-y)} \,\,\, .
\label{eq:dfdef}
\end{equation}
Because the factor of $e^{-\eta \, k^4}$ blows up in certain directions in the complex $k^0$ plane, as indicated earlier, the usual procedure
of closing the integration contour in the lower half plane is not useful; instead we directly evaluate the $k^0$ integral along the real axis, deviating
by small semi-circles below and above the poles at $k^0 = -E_{\vec{k}}$ and $+E_{\vec{k}}$, respectively. Hence, we
may write $D_F(x-y) = I_{2C} + I_{PV}$, where $I_{2C}$ is the result from the semi-circle integrations while $I_{PV}$ is the
remaining principal value integral. The latter can be reduced to a one-dimensional integral in the case $m=0$, which we can
numerically evaluate. We find
\begin{equation}
I_{2C} = \frac{1}{4 \pi^2 |\vec{r}|} \int_0^\infty dk \sin(k\, |\vec{r}|) \cos(k \,r^0) \,\,\, ,
\end{equation}
where we define $r=x-y$ and $k=|\vec{k}|$. The remaining principal value integral is identical to one considered in the computation of the retarded
propagator for this theory in Ref.~\cite{Pais:1950za}, and is given by
\begin{equation}
I_{PV} = \frac{-i}{\pi^3} \frac{\partial}{\partial \rho} \left[ \mbox{sign}(\rho) \int_0^\infty \frac{dy}{y} \exp(-\eta y^4/\rho^2)
\left[K_0(y)+\frac{\pi}{2} Y_0(y)\right] \right] \,\,\, .
\label{eq:ipv}
\end{equation}
where $\rho = {r^0}^2-|\vec{r}|^2$. Restricting to the case where $r^0 < |\vec{r}|$, it follows from Eqs.~(\ref{eq:dfdf}) and (\ref{eq:ipv})
that
\begin{equation}
\langle 0 | [\phi(x) , \phi(y)] | 0 \rangle = \frac{4 \, i\, \eta}{\pi^3 \rho^3} \int_0^\infty dy \, y^3 \exp(-\eta \,y^4/\rho^2)
\left[K_0(y)+\frac{\pi}{2} Y_0(y)\right] \,,\,\,\,\,\,\,\, (0< r^0 < |\vec{r}|) \,\,\, ,
\label{eq:comres}
\end{equation}
where $K_0$ and $Y_0$ are Bessel functions. Eq.~(\ref{eq:comres}) is nonvanishing, as is shown in Fig.~\ref{fig:com}. We do not consider the case
$m \neq 0$, since the necessary two-dimensional numerical integration is much more tedious but no more illuminating.
One might wonder how this calculation of the commutator relates to a similar calculation in the formulation of the theory where
the nonlocality appears only in the interaction terms, Eq.~(\ref{eq:nloc1}). The unstated assumption is that the theory presented
in this form results from a field redefinition starting with the theory in which only the quadratic terms are modified, Eq.~(\ref{eq:nloc2}).
With nonlocal interactions arising in this way, operators that correspond to observables are built out of the ``smeared" fields,
$\hat{F}^{1/2} \phi(x)$, and it is the commutator of these objects that is the physically relevant quantity to evaluate at space-like separation.
This gives precisely the same integral to evaluate as in Eq.~(\ref{eq:dfdef}), with a different origin for the momentum dependence in the
numerator.
\subsection{Acausal Vertex Ordering}
The question we now wish to address is how acausality manifests itself in the time-dependence of scattering processes. We allow ourselves
to stray from the O($N$) model in this subsection to consider theories of a single real scalar field with the same modification of their quadratic
terms as Eq.~(\ref{eq:nloc2}), but with different interaction terms. This will allow us to illustrate the effects of interest most clearly; the
generalization to the O($N$) model that we previously considered will be clear by analogy. We consider two examples, following
the general approach of Ref.~\cite{Grinstein:2008bg}:
{\em Particle Production by a Source.} Consider a local theory of a real scalar field which includes a coupling to a classical source,
\begin{equation}
{\cal L }_{int} = \phi(x) j(x) \,\,\, ,
\end{equation}
where ${\cal L}_{int}$ is the interaction Lagrangian. We wish to study $\langle \psi_{out} | \Omega \rangle$, the amplitude for the source to create an
outgoing wave-packet state from the vacuum, where
\begin{equation}
| \psi_{out} \rangle = \int d^4 x' g(x'-y') \, \phi(x') \, | \Omega \rangle \,\,\, .
\end{equation}
Here we follow the convections of Ref.~\cite{Grinstein:2008bg} where primed coordinates correspond to ``out" states. The function $g(x')$ is chosen so that the outgoing wave packet is localized about the space-time point $y'$ at some
time long after the source is turned off, and its four-momentum is localized about $p'$. By the choice of this function, we can determine the position of the wave-packet at any earlier time when the source is turned on. For a source localized within a small region about the spacetime origin, we first show that the amplitude vanishes if the wave-packet's trajectory extrapolates back to the origin at a time substantially earlier than $t=0$, as one would expect for a causal process. We then consider how this conclusion changes given the chosen nonlocal modification of this theory.
The amplitude $\langle \psi_{out} | \Omega\rangle$ may be written
\begin{eqnarray}
\langle \psi_{out} | \Omega\rangle &=& \int d^4 x' g^*(x'-y') \langle \Omega | \phi(x') | \Omega \rangle \nonumber \\
&=& i \int d^4 y \, j(y) \int \frac{d^4 p}{(2 \pi)^4} \frac{i}{p^2-m^2+i \epsilon} \, \tilde{g}(p)^* \, e^{-i p \cdot (y'-y)} \nonumber \\
&\equiv& i \int d^4 y \, j(y) \, I(y'-y) \,\,\,,
\end{eqnarray}
where $\tilde{g}(p)$ is the Fourier transform $\tilde{g}(p)= \int \!d^4 x \, g(x)\, e^{i p\cdot x}$. As we discussed earlier, all momentum-space
correlation functions are defined via analytic continuation from a Euclidean theory; all subsequent calculations,
including Fourier transforms, are then performed in Minkowski space. The integral $I(y'-y)$ can be re-expressed using a Schwinger parameter,
\begin{equation}
I(\Delta y) = \frac{1}{\hbar} \int_0^\infty ds \int \frac{d^4 p}{(2 \pi)^4} \, e^{i s (p^2-m^2 + i \epsilon)/\hbar} \,
\tilde{g}(p)^* \, e^{-i p\cdot \Delta y/\hbar} \,\,\, ,
\label{eq:schwing}
\end{equation}
where $\Delta y \equiv y'-y$, and we have temporarily restored the $\hbar$ dependence. As in Ref.~\cite{Grinstein:2008bg},
if the relevant distance scales (in this case $\Delta y$) are large compared to all characteristic inverse masses and inverse momenta, then we
are justified in using the stationary phase approximation, since this limit is equivalent to taking $\hbar \rightarrow 0$ in Eq.~(\ref{eq:schwing}).
Evaluating the $d^4 p$ integral in this way (and resuming our convention that $\hbar=1$) yields
\begin{equation}
I(\Delta y) = \frac{i}{16 \pi^2} \int_0^\infty ds\, \frac{1}{s^2} \, \tilde{g}(\frac{\Delta y}{2 s})^* \, e^{-i [ \Delta y^2/(4s) + s m^2]} \,\,\,.
\end{equation}
By evaluating the $ds$ integral in the same way one finds
\begin{equation}
I(\Delta y) = \frac{\sqrt{i}}{4 \sqrt{2} \pi^{3/2}} \frac{m^{1/2}}{(\Delta y^2)^{3/4}} \tilde{g}(m \frac{\Delta y}{\sqrt{\Delta y^2}})^* e^{-i m \sqrt{\Delta y^2}} \,\,\, ,
\end{equation}
leading finally to
\begin{equation}
\langle \psi_{out} | \Omega \rangle = \frac{1}{2} \left(\frac{i}{2\pi}\right)^{3/2} \int d^4 y \, \frac{m^{1/2}}{[(y'-y)^2]^{3/4}} \, j(y) \,
\tilde{g} \!\left( m \frac{y'-y}{\sqrt{(y'-y)^2}} \right)^* e^{-i m \sqrt{(y'-y)^2}} \,\,\, .
\label{eq:theamp}
\end{equation}
By construction, the function $\tilde{g}$ only has support in the region where
\begin{equation}
m \frac{(y' - y)} {\sqrt{(y'-y)^2}} \approx p' \,\,\, ,
\end{equation}
which limits the possible values of $y$ that contribute to the integral. Let us assume a $\tilde{g}$ in which $\vec{y} \approx 0$ for some $y^0 \ll 0$. If $j(y)$
is strongly localized about the spacetime origin, for example a delta function source $j(y) \propto \delta^{(4)}(y)$, the integral in Eq.~(\ref{eq:theamp}) vanishes. The ``production vertex"
for the outgoing wave packet, which is identified spatially with the origin, cannot occur before the time at which the source excites the system.
The conclusion is different if we introduce a nonlocal coupling to the source following our earlier prescription
\begin{equation}
{\cal L }_{int}= [\hat{F}^{1/2} \phi(x)] j(x) \,\,\, .
\end{equation}
This case is simple to understand since we can integrate by parts, and recover a theory of the original form, but with a ``smeared" source,
\begin{equation}
j(y)_s=\hat{F}^{1/2} j(y) = \int d^4 x \, \epsilon(y-x) \, j(x) \,\,\,\,\, \mbox{ where } \,\,\,\,\, \epsilon(y-x) = \int \frac{d^4 k}{(2 \pi)^4} e^{-\eta \, k^4/2} \, e^{i k\cdot(y-x)} \,\,\,.
\end{equation}
Assuming the example where $j(y)=c_0 \, \delta^{(4)}(y)$, where $c_0$ is a coupling, consider the time-dependence of $j(y)_s$ near the spatial origin
\begin{eqnarray}
j(y^0,\vec{y}=0)_s &=& \frac{c_0}{2 \pi^3} \int_0^\infty dk^0 \int_0^\infty dk\, k^2 \exp[-\eta\, ({k^0}^2-k^2)^2/2] \cos(k^0 y^0) \nonumber \\
&=& \frac{c_0}{8 \sqrt{2} \pi^3} \int_0^\infty dk^0 \, {k^0}^3 e^{-\eta \, k0^4/4} \cos[k^0 y^0] \left[ K_{1/4}({k^0}^4)-K_{3/4}({k^0}^4) \right] \,\,\, ,
\label{eq:kint}
\end{eqnarray}
where $K_i$ is a Bessel function of the second kind, of order $i$. Unlike the original $j(x)$, this function is no longer localized in time at
$t=0$. The second line of Eq.~(\ref{eq:kint}) can be evaluated numerically and the results are shown in Fig.~\ref{fig:js}. This result implies that
there is a common region with $y^0 \ll 0$ and $\vec{y} \approx \vec{0}$ where the functions $j$ and $\tilde{g}$ in Eq.~(\ref{eq:theamp}) both have support; the
overlap $\langle \psi_{out} | \Omega \rangle$ is therefore generally nonvanishing. One concludes that there is a non-vanishing probability that
the wave packet appears to emerge from the position of a spatially localized source at a time before the system has
been excited by the source.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.5\textwidth]{jsplot.eps}
\caption{Time-dependence of the smeared source function at the origin $j(t)_s \equiv j(y^0=t,\vec{y}=\vec{0})_s$ for $\eta=1$ and $c_0=1$.}
\label{fig:js}
\end{center}
\end{figure}
{\em Two-into-two scattering.} The previous example is perhaps the simplest illustration of how the smearing of interaction positions can lead to the apparent
acausal ordering of scattering events. In the more phenomenologically relevant case of two-into-two scattering, similar results can be obtained.
We use the term ``apparent" since the spacetime positions of the production and the subsequent decay of a resonance, for example, are inferred from
the wave packet positions and momenta long before and after the interaction has occurred. Nothing meaningful can be said about the system directly at intermediate times,
since no measurements are made during this period. We now consider how the wave-packet analysis of two-into-two scattering discussed
in the context of Lee-Wick theories in Ref.~\cite{Grinstein:2008bg} is modified in the present context.
We consider the same free theory of a real scalar field $\phi$, and introduce couplings to two additional real scalar fields $\psi$ and $\chi$, that would otherwise have no
higher-derivative couplings. In this case we assume
\begin{equation}
{\cal L}_{int} = \frac{1}{2} g_\chi (\hat{F}^{1/2} \phi) \chi^2 + \frac{1}{2} g_\psi (\hat{F}^{1/2} \phi) \psi^2 \,\,\, ,
\label{eq:moddef}
\end{equation}
Again, this is consistent with the assumption that we start with a theory in which the higher-derivative operators appear in the $\phi$
quadratic terms only, and that these terms have been subsequently put in canonical form by a field redefinition. We do not consider
doing the same for the $\psi$ and $\chi$ fields to simplify the analysis; there is no reason to expect that this choice will affect our conclusions
qualitatively. We consider the scattering process $\chi \chi \rightarrow \psi \psi$. Setting the problem up in the way that we have is convenient since the
nonlocality affects the propagator but not the external lines, which allows us to immediately carry over most of the wave-packet analysis of
Ref.~\cite{Grinstein:2008bg} without modification. Let us briefly recapitulate the key steps in this approach.
We assume incoming and outgoing wavepacket states given by
\begin{align}
| \psi_{in} \rangle & = \int d^4 x_1 d^4 x_2 f_1(x_1-y_1) f_2(x_2-y_2) \phi(x_1) \phi(x_2) | \Omega \rangle \,\,\ ,\\
| \psi_{out} \rangle & = \int d^4 x'_1 d^4 x'_2 g_1(x'_1-y'_1) g_2(x'_2-y'_2) \phi(x'_1) \phi(x'_2) | \Omega \rangle \,\,\,,
\end{align}
where the functions $f_i$ and $g_i$ define the wavepackets. These are chosen so that in the process of interest, we can
specify well-defined production and decay vertices for the resonance, in this case associated with the field $\phi$,
exchanged in the $s$-channel. To be more explicit, the functions $f_i$ are chosen so that the initial wavepackets are localized
about the space-like separated points $y_1$ and $y_2$, respectively, and have momenta peaked at
$p_1= m_\chi v_1$ and $p_2 = m_\chi v_2$. A production vertex can be defined at point $z_0$, where
\begin{equation}
\frac{z_0-y_1}{\tau_1} = v_1 \,\,\,\,\, \mbox{ and } \,\,\,\,\, \frac{z_0-y_2}{\tau_2}=v_2 \,\,\, ,
\end{equation}
with $ \tau_i^2 \equiv (z_0 - y_i)^2$. Similarly, the functions $g_i$ are chosen so that the final wavepackets are localized
about the space-like separated points $y'_1$ and $y'_2$, respectively, and have momenta peaked at
$p'_1= m_\psi v'_1$ and $p'_2 = m_\psi v'_2$. A decay vertex can be defined at point $z_0'$ where
\begin{equation}
\frac{y'_1-z'_0}{\tau'_1} = v'_1 \,\,\,\,\, \mbox{ and } \,\,\,\,\, \frac{y'_2-z'_0}{\tau'_2}=v'_2 \,\,\, ,
\end{equation}
with ${\tau'}_i^2 \equiv (y'_i-z'_0)^2$. Defining $w^\mu \equiv {z'}_0^\mu -z_0^\mu$, the authors of Ref.~\cite{Grinstein:2008bg}
determined how the amplitude $\langle \psi_{out} | \psi_{in} \rangle$ depends on $w^0$ and showed in Lee-Wick
theories that the amplitude is non-vanishing for an acausal ordering of the vertices. The key intermediate steps are these: the
amplitude of interest can be written in the form
\begin{equation}
\langle \psi_{out} | \psi_{in} \rangle = \int \frac{d^4 q}{(2 \pi)^4} \tilde{F}(q) \tilde{G}(q) \Gamma^{(4)}_s(q^2) \,\,\, ,
\end{equation}
where $\Gamma^{(4)}_s(q^2)$ is the momentum-space four-point function for the $s$-channel process of interest, with propagators truncated from the external lines, and
\begin{align}
\tilde{F}(q) & = \int d^4 z \, e^{i z \cdot q} \, I_1(z) I_2 (z) \,\,\,\,\, \mbox{ with}\\
I_i(z) & = \int \frac{d^4 k_i}{(2 \pi)^4} \, e^{i k_i \cdot (y_i -z)} \tilde{f}_i(k_i) \, \frac{i}{k_i^2 - m^2 +i \epsilon} \,\,\,,
\label{eq:spints}
\end{align}
where $\tilde{f}_i(k)$ are the Fourier transforms of the incoming wave packet functions. We do not display the analogous expressions for $G(q)$, corresponding to the outgoing
wave packet states. In the same limit described in our earlier example involving a classical source, the momentum and $z$ integrals in Eq.~(\ref{eq:spints}) can be evaluated
in the stationary phase approximation, leading to a result of the form
\begin{equation}
\langle \psi_{out} | \psi_{in} \rangle \simeq \int \frac{d^4 q}{(2 \pi)^4} \, e^{-i q \cdot (z_0'-z_0)} \, \hat{F}(q)\, \hat{G}(q) \, \Gamma^{(4)}_s(q^2)
\,\,\, ,
\label{eq:gowres}
\end{equation}
where the functions $\hat{F}$ and $\hat{G}$ have localized support at $q \approx p_1+p_2$ and $q \approx p'_1+p'_2$, respectively.
We study the nonlocal theory of interest using Eq.~(\ref{eq:gowres}) as the starting point. Hence, from Eq.~(\ref{eq:moddef})
it follows that
\begin{equation}
\langle \psi_{out} | \psi_{in} \rangle = \int \frac{d^4 q}{(2 \pi)^4} \, e^{-i q \cdot w}
\left[\frac{- i \,g_\chi \, g_\psi \, e^{-\eta \, q^4} }{q^2-m_\phi^2+i \, m\, \Gamma} \right] \,
\hat{F}(q)\, \hat{G}(q) \,\,\,,
\label{eq:ourap}
\end{equation}
where $\Gamma$ is the $\phi$ decay width. Defining the Fourier transform
\begin{equation}
\epsilon'(x)=\int\frac{d^4 q}{(2\pi)^4} \, e^{-\eta \,q^4} \, e^{i q \cdot x} \,\,\,\, ,
\end{equation}
the amplitude can be written as
\begin{equation}
\langle \psi_{out} | \psi_{in} \rangle = \int d^4 x \, \epsilon(x-w) I(x) \,\,\, ,
\label{eq:epi}
\end{equation}
where
\begin{equation}
I(x) = \frac{\sqrt{i} g_\phi g_\chi}{{8 \pi^{3/2}}} \frac{m_\phi^{1/2}}{(\sqrt{x^2})^{3/2}} e^{-i \, m_\phi \sqrt{x^2}} \,
e^{-\Gamma \sqrt{x^2}/2} \hat{F}(m_\phi \frac{x}{\sqrt{x^2}}) \, \hat{G}(m_\phi \frac{x}{\sqrt{x^2}}) \,\,\, .
\label{eq:iox}
\end{equation}
As in the case of the ``ordinary resonance" discussed in Ref.~\cite{Grinstein:2008bg}, as well as in our previous example,
$I(x)$ is derived by exponentiating the propagator denominator using a Schwinger parameter and
then integrating using the stationary phase approximation. Note that we cannot apply this approximation to Eq.~(\ref{eq:ourap})
directly since we require that the nonlocal length scale $\eta^{1/4}$ to be comparable to the vertex separation;
one cannot then assume that the real exponential prefactor is a slowly varying function of the momentum relative to the complex phase
factor. Eq.~(\ref{eq:iox}) coincides with the ordinary resonance result when $\eta=0$, in which case $\epsilon'$ becomes a four-dimensional
delta function. In that limit, the argument of Ref.~\cite{Grinstein:2008bg} is the following: in the center-of-mass frame, $\hat{F}$ and $\hat{G}$ only have support
where $x^0=w^0>0$ and $\vec{x}=\vec{w} \approx0$. Hence the amplitude is only nonvanishing for the causal
ordering of the production and decay vertices. (In the Lee-Wick case, the result is the opposite.) Making the same
assumptions here, one concludes only that $x^0$ must be greater than zero; the amplitude may be nonvanishing, for
example, when $w^0<0$ and $\vec{w}=0$, since $\epsilon'(x-w)$ is no longer a delta function. This can be verified by noting
that the function $\epsilon'$ differs from the function $\epsilon$ that we have previously studied by the replacement
$\eta \rightarrow 2 \eta$. Since the $x$ integral
is dominated by the region where $\vec{x}\approx 0$, we can evaluate $\epsilon(x-w)$ for the choice $\vec{x}=\vec{w}=0$,
where $\vec{w}=0$ corresponds to the case in which the production and decay vertices are spatially coincident. Since $\epsilon'$
is non-vanishing for $x^0>0$ and $w^0<0$ we conclude that $\epsilon'$, $\hat{F}$ and $\hat{G}$ have common regions of support,
so that Eq.~(\ref{eq:epi}) is generally non-vanishing. Hence, there is a non-vanishing amplitude for an acausal ordering of the production
and decay vertices. The effect emerges in a very different way than in the Lee-Wick theories. In that case, a crucial sign flip in the propagator of the
Lee-Wick resonance leads to a change from $w^0$ to $-w^0$ in comparison to the ordinary resonance case. The sign flip
affects the sign of the width appearing in one of the exponential factors in the amplitude, leading to the interpretation that
the exponential decay is happening as the Lee-Wick resonance propagates backward in time from the decay to production
vertex. Here, however, the form of $I(x)$ corresponds to propagation forward in time over the time-like interval $x$. The
nonlocality in the theory leads to a spatial smearing of the interaction points so that one no longer identifies $x$ with the physical
spacetime separation of the extrapolated decay and production vertices.
\section{Conclusions} \label{sec:conc}
We have considered unitarity and causality in a theory where quadratic terms are modified by higher-derivative terms of
infinite order, chosen so as not to induce additional poles in the propagator at tree level. We have studied unitarity at leading order in the large $N$ limit of the
scalar O($N$) model for Euclidean and Minkowski space formulations of the theory. We have verified that a unitary theory
is obtained from Euclidean correlation functions that are analytically continued in their external momenta to Minkowski space,
but not when correlation functions are formulated in Minkowski space directly. In the same theory, we verified the non-vanishing of field
commutators at space-like separation by an explicit calculation. We then studied the time-dependence of scattering amplitudes in similar theories
using a wave-packet approach employed by others~\cite{Grinstein:2008bg} in studying Lee-Wick theories. We found that the apparent acausal ordering of
decay and production vertices of resonances was a common feature in these theories. Unlike Lee-Wick theories, this effect would be present in tree-level resonant
exchanges for all the states in the theory that are subject to modified quadratic terms and would make solutions to the hierarchy problem based on this idea phenomenologically distinct from others
that have been proposed.
\begin{acknowledgments}
We thank Josh Erlich for valuable discussions. This work was supported by the NSF under Grant PHY-1519644, and is dedicated to the memory of
David A. Carone.
\end{acknowledgments}
|
\section{Introduction and statement of results}
It is well known that the cyclic homology of any commutative (DG) algebra $A$ has a natural decomposition
\begin{equation}
\la{HCcomm}
\overline{\mathrm{HC}}_\bullet(A) \,\cong\, \bigoplus_{p=1}^{\infty}\, \mathrm{HC}_\bullet^{(p)}(A)\ ,
\end{equation}
which (in the case when $A$ is smooth) arises from a Hodge-style truncation of the de Rham complex of $A$. This decomposition is usually called the Hodge (or $\lambda$-)decomposition of $ \mathrm{HC}_\bullet(A)$, and it can indeed be linked to classical Hodge theory in algebraic geometry (see \cite{BV}, \cite{GS}, \cite{L1} or \cite{L}).
In \cite{bfprw1}, we found that a direct sum decomposition similar to \eqref{HCcomm} exists for the universal enveloping algebra $ \U \mathfrak{a} $ of any (DG) Lie algebra $ \mathfrak{a}\,$:
\begin{equation}
\la{hodgeds}
\overline{\mathrm{HC}}_{\bullet}(\U\mathfrak{a}) \,\cong\,\bigoplus_{p=1}^{\infty}\, \mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})\ \text{.}
\end{equation}
The direct summands of \eqref{hodgeds} appeared in \cite{bfprw1} as domains of certain (derived) character maps $\,{\rm{Tr}}_{\mathfrak{g}}(\mathfrak{a}):\,\mathrm{HC}_\bullet^{(p)}(\mathfrak{a}) \to \mathrm H_\bullet(\mathfrak{a}, \mathfrak{g})\,$ with values in representation homology of $ \mathfrak{a} $ in
a finite-dimensional Lie algebra $ \mathfrak{g} $; in terms of these
maps, we gave a new homological interpretation of the (strong) Macdonald Conjecture for a
reductive Lie algebra $ \mathfrak{g} $ (see \cite{bfprw1}, Section~9).
The purpose of this paper is two-fold. First, we study general properties of the Hodge-type decomposition \eqref{hodgeds} for an arbitrary Lie algebra $ \mathfrak{a} $. We extend the construction of \cite{bfprw1} to the Hochschild homology of $\U\mathfrak{a} $
and show how the resulting homology theories $ \mathrm{HC}_\bullet^{(p)}(\mathfrak{a}) $ and $ {\rm{HH}}_\bullet^{(p)}(\mathfrak{a}) $ generalize the operadic construction of cyclic/Hochschild homology for Lie algebras
due to Getzler-Kapranov \cite{GK}.
Furthermore, we give a natural topological interpretation of the decomposition
\eqref{hodgeds} in terms of Frobenius operations on $S^1$-equivariant homology
$ \mathrm H^{S^1}_{\bullet}(\LL X, \mathbb{Q}) $ of the free loop space of a simply connected
topological space $X$.
Our second goal is to clarify the relation of \eqref{hodgeds}
to a canonical derived Poisson structure on $ \U\mathfrak{a} $ introduced in \cite{BCER}.
This Poisson structure comes from a cyclic pairing on the Koszul
dual coalgebra of $ \mathfrak{a} $, and our key observation is that the corresponding Lie
bracket on cyclic homology of $ \U\mathfrak{a} $ preserves the Hodge filtration associated to
\eqref{hodgeds}. The motivation for studying such cyclic Poisson
structures comes from topology: it is known
that the Chas-Sullivan Lie algebra of a simply connected closed
manifold $M$ is an example of a derived Poisson algebra associated with a natural
cyclic pairing on the Lambrechts-Stanley model of $M$ (see \cite{BCER}, Section~5.5). In this way, our results reveal a new general property of string topology Lie algebras. We actually expect that the Chas-Sullivan Lie bracket preserves the Hodge
decomposition \eqref{hodgeds} ({\it cf.} Conjecture~\ref{conj1} in Section~\ref{Sect4}).
We now proceed with a detailed discussion of results of the paper. We begin by reviewing the derived functor construction of the Hodge decomposition \eqref{hodgeds} given in \cite{bfprw1}.
Let $k$ be a field of characteristic $0$. Given a Lie algebra $ \mathfrak{a} $ over $k$, we consider
the symmetric ad-invariant $k$-multilinear forms on $ \mathfrak{a} \,$ of a (fixed) degree $ p \ge 1 $. Every such form is induced from the universal
one: $\,\mathfrak{a} \times \mathfrak{a} \times \ldots \times \mathfrak{a} \to \lambda^{(p)}(\mathfrak{a}) \,$, which takes its values in the space $\,\lambda^{(p)}(\mathfrak{a})\,$ of coinvariants of the adjoint representation of $ \mathfrak{a} $ in $ {\rm{Sym}}^p(\mathfrak{a})\,$.
The assignment $\,\mathfrak{a} \mapsto \lambda^{(p)}(\mathfrak{a})\,$ defines a (non-additive) functor on the category of Lie algebras that extends in a canonical way to the category of DG Lie algebras:
\begin{equation}
\la{lam}
\lambda^{(p)}:\,\mathtt{DGLA}_k \xrightarrow{} \mathtt{Com}_k \ ,\quad \mathfrak{a} \mapsto {\rm{Sym}}^p(\mathfrak{a})/[\mathfrak{a}, {\rm{Sym}}^p(\mathfrak{a})]\ .
\end{equation}
The category $ \mathtt{DGLA}_k $ has a natural model structure (in the sense of Quillen \cite{Q1}), with
weak equivalences being the quasi-isomorphisms of DG Lie algebras. The corresponding homotopy (derived) category $ {\mathtt{Ho}}(\mathtt{DGLA}_k) $ is obtained from $ \mathtt{DGLA}_k $ by localizing at the class of
weak equivalences, i.e. by formally
inverting all the quasi-isomorphisms in $ \mathtt{DGLA}_k $. The functor \eqref{lam}, however,
does {\it not} preserve quasi-isomorphisms and hence does not descend to the homotopy category
$ {\mathtt{Ho}}(\mathtt{DGLA}_k) $. To remedy this problem, one has to replace $\,\lambda^{(p)}\,$ by its (left) derived functor
\begin{equation}
\la{Llam}
\boldsymbol{L}\lambda^{(p)}:\,{\mathtt{Ho}}(\mathtt{DGLA}_k) \to {\mathtt D}(k)\ ,
\end{equation}
which takes its values in the derived category $ {\mathtt D}(k) $
of $k$-complexes. We write $\,\mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})\,$ for the homology of $\, \boldsymbol{L}\lambda^{(p)}(\mathfrak{a}) \,$ and call it the {\it Lie-Hodge homology} of $ \mathfrak{a} $.
For $ p = 1 $, the functor $ \lambda^{(1)} $ is just abelianization of
Lie algebras; in this case, the existence of $\, \boldsymbol{L}\lambda^{(1)} \,$ follows from
Quillen's general theory (see \cite[Chapter~II, \S 5]{Q1}), and
$\, \mathrm{HC}^{(1)}_{\bullet}(\mathfrak{a}) \,$ coincides (up to shift in degree)
with the classical Chevalley-Eilenberg homology $ \mathrm H_\bullet(\mathfrak{a}, k) $ of the Lie algebra $ \mathfrak{a} $.
For $p=2$, the functor $ \lambda^{(2)} $ was introduced by Drinfeld \cite{Dr}; the existence of $ \boldsymbol{L}\lambda^{(2)} $ was established by Getzler and Kapranov \cite{GK} who
suggested that $
\mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a}) $ should be viewed as an (operadic) version of cyclic homology for Lie algebras.
One of the key results of \cite{GK} is the existence of Connes' type periodicity sequence
for Lie cyclic homology, which (in the notation of \cite{GK}, see Section~\ref{GKcomp} below)
reads
\begin{equation}
\la{GKconnes}
\ldots \,\to\, \mathrm{HA}_{n-1}(\mathtt{Lie}, \mathfrak{a})\,\to\, {\mathrm{HB}}_{n-1}(\mathtt{Lie}, \mathfrak{a}) \,\to\, \mathrm{HC}_{n-1}(\mathtt{Lie}, \mathfrak{a}) \,\to\,\mathrm{HA}_{n-2}(\mathtt{Lie}, \mathfrak{a}) \,\to\,
\ldots
\end{equation}
Now, for an arbitrary $ p \ge 1 $, the existence of the derived functor $ \boldsymbol{L}\lambda^{(p)} $ was
proven in \cite{bfprw1} (see {\it op. cit.}, Theorem~7.1), using Brown's Lemma and some homotopical arguments from \cite{BKR}.
To construct the direct sum decomposition \eqref{hodgeds}
we observe that each $ \lambda^{(p)} $ comes together with a natural transformation to the composite functor
$ \U_\n := (\,\mbox{--}\,)_\n \circ \,\U:\,\mathtt{DGLA}_k \to \mathtt{DGA}_{k/k} \to \mathtt{Com}_k $, where
$\, (\,\mbox{--}\,)_\n $
denotes the cyclic functor $\,R \mapsto R/(k+[R,R]) \,$ on the category of (augmented) associative DG algebras. The natural transformations $\,\lambda^{(p)} \to \U_\n \,$ are induced by the symmetrization maps
\begin{equation}
\la{symfun}
{\rm{Sym}}^p(\mathfrak{a}) \to \U\mathfrak{a}\ ,\quad x_1 x_2 \ldots x_p\, \mapsto\,
\frac{1}{p!}\,\sum_{\sigma \in {\mathbb S}_p}\, \pm \,x_{\sigma(1)} \cdot x_{\sigma(2)} \cdot
\ldots \cdot x_{\sigma(p)}\ ,
\end{equation}
which, by the Poincar\'e-Birkhoff-Witt Theorem, assemble to an isomorphism of DG $\mathfrak{a}$-modules
$\,{\rm{Sym}}_k(\mathfrak{a}) \cong \U \mathfrak{a} \,$. From this, it follows that $\,\lambda^{(p)} \to \U_\n \,$ assemble to an isomorphism of functors
\begin{equation}
\la{eqv1}
\bigoplus_{p=1}^{\infty} \lambda^{(p)} \,\cong \, \U_\n \ .
\end{equation}
On the other hand, by a theorem of Feigin and Tsygan
\cite{FT} (see also \cite{BKR}), the functor $ (\,\mbox{--}\,)_\n $ has a left
derived functor $ \boldsymbol{L}(\,\mbox{--}\,)_\n:\, {\mathtt{Ho}}(\mathtt{DGA}_{k/k}) \to {\mathtt D}(k) $ that computes
the reduced cyclic homology $\,\overline{\mathrm{HC}}_\bullet(R)\,$ of an associative algebra
$ R \in \mathtt{DGA}_{k/k} $. Since $ \U $ preserves quasi-isomorphisms and maps cofibrant
DG Lie algebras to cofibrant DG associative algebras, the isomorphism \eqref{eqv1}
induces an isomorphism of derived functors from $ {\mathtt{Ho}}(\mathtt{DGLA}_k) $ to $ {\mathtt D}(k) $:
\begin{equation}
\la{eqv2}
\bigoplus_{p=1}^{\infty}\, \boldsymbol{L}\lambda^{(p)}\, \cong\, \boldsymbol{L}(\,\mbox{--}\,)_\n \circ \,\U \ .
\end{equation}
At the level of homology, \eqref{eqv2} yields the direct decomposition \eqref{hodgeds} ({\it cf.}~\cite[Theorem 7.2]{bfprw1}).
As explained in \cite{bfprw1}, the existence of \eqref{eqv2} is related to the fact that $ \U\mathfrak{a} $ is a cocommutative Hopf algebra, and in a sense, the Lie Hodge decomposition \eqref{hodgeds} is Koszul
dual to the classical Hodge decomposition \eqref{HCcomm} for commutative algebras.
In the present paper, we extend the above derived functor construction to the Hochschild homology of $\, \U\mathfrak{a} \,$:
\begin{equation}
\la{hhoch}
\overline{{\rm{HH}}}_{\bullet}(\U\mathfrak{a}) \,\cong\,\bigoplus_{p=0}^{\infty}\, {\rm{HH}}^{(p)}_{\bullet}(\mathfrak{a})
\ ,
\end{equation}
and relate \eqref{hhoch} to the Hodge decomposition \eqref{hodgeds} of the
cyclic homology of $ \U\mathfrak{a} $. More precisely, we prove the following result ({\it cf.}
Theorem~\ref{hodgesbi} and Theorem~\ref{connesvsgksbi} in Section~\ref{S2}).
\begin{theorem}
\la{T1}
For any (DG) Lie algebra $\mathfrak{a}$, the Connes periodicity sequence for $ \U\mathfrak{a} $
decomposes into a natural direct sum of Hodge components, each of which is itself a long exact sequence:
\begin{equation*}
\ldots \, \to\, \mathrm{HC}_{n-1}^{(p+1)}(\mathfrak{a}) \,\xrightarrow{B} \,
{{\rm{HH}}}_n^{(p)}(\mathfrak{a}) \,\xrightarrow{I}\, \mathrm{HC}_n^{(p)}(\mathfrak{a}) \,\xrightarrow{S} \,
\mathrm{HC}_{n-2}^{(p+1)}(\mathfrak{a}) \,\to\, \ldots
\end{equation*}
For $p=1$, the above exact sequence is precisely the Getzler-Kapranov exact
sequence \eqref{GKconnes}.
\end{theorem}
Next, we explain a topological meaning of the decompositions \eqref{hodgeds} and \eqref{hhoch}. Our starting point is a fundamental theorem
of Quillen \cite{Q2} that assigns to every $1$-connected topological space
$X$ a DG Lie algebra $ \mathfrak{a}_X $ over $\mathbb{Q}$, called a Lie model of $X$. The Lie algebra $ \mathfrak{a}_X $ determines the rational homotopy type of $X$ and thus determines any homotopy invariant of $X$ defined over $ \mathbb{Q} $. In particular, it is known \cite{M} that the universal enveloping algebra
$ \U \mathfrak{a}_X $ is quasi-isomorphic to the singular chain algebra
$ C_\bullet(\Omega X, \mathbb{Q}) $ of the based loop space $ \Omega X $
of $X$, while the Hochschild and cyclic homology of $ \U \mathfrak{a}_X $ are
isomorphic respectively to the rational homology and rational $S^1$-equivariant homology of the free loop space $ \LL X $ of $X$ (see \cite{Go, J} and also \cite{JM}):
\begin{equation}
\la{jonesiso}
{\rHH}_{\bullet}(\mathcal U\mathfrak{a}_X) \,\cong\,
\overline{\mathrm H}_{\bullet}(\LL X, \mathbb{Q}) \ ,\qquad
\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a}_X) \,\cong\,
{\rH}^{S^1}_{\bullet}(\LL X, \mathbb{Q}) \ .
\end{equation}
In view of \eqref{jonesiso}, it is natural to ask if the Hodge decompositions
\eqref{hodgeds} and \eqref{hhoch} for $\mathfrak{a}_X$ can be interpreted in terms of homology of the free loop space $ \LL X$.
To answer this question we recall that $ \LL X := {\rm Map}(S^1, X) $
carries a natural circle action (induced by the action of $ S^1 $ on itself).
Considering the $n$-fold covering of the circle:
$\, S^1 \to S^1 $, $ \, e^{i\theta} \mapsto e^{i n \theta}\,$, we may
define, for each $ n \ge 0 $, $\mathbb{Q}$-linear operations on the homology
of $\LL X$ (see Section~\ref{Sect4.1} for details):
\begin{equation*}
\Phi_X^n:\, {\rH}_{\bullet}(\LL X, \mathbb{Q}) \to {\rH}_{\bullet}(\LL X, \mathbb{Q})
\ ,\qquad \tilde{\Phi}_X^n :\, {\rH}^{S^1}_{\bullet}(\LL X, \mathbb{Q})
\to {\rH}^{S^1}_{\bullet}(\LL X, \mathbb{Q})\ .
\end{equation*}
called the Frobenius operations. Now, let ${\rH}^{(p)}_{\bullet}(\LL X, \mathbb{Q}) $ and $ {\rH}^{S^1,\,(p)}_{\bullet}(\LL X, \mathbb{Q}) $ denote the common eigenspaces of these
Frobenius operations ($ \Phi_X^n $ and $ \tilde{\Phi}_X^n $, respectively), corresponding to the eigenvalues $ n^p $ for all $ n \ge 0 $. Then, we have the following theorem ({\it cf.} Theorem~\ref{top2} in
Section~\ref{Sect4}), which is our second main result in the paper.
\begin{theorem}
\la{T2}
For each $p\ge 0$, there are natural isomorphisms
\begin{equation*}
{\rHH}^{(p)}_{\bullet}(\mathfrak{a}_X) \cong {\rH}^{(p)}_{\bullet}(\mathcal L X, \mathbb{Q}) \ ,\qquad
\overline{\mathrm{HC}}^{(p)}_{\bullet}(\mathfrak{a}_X) \cong
{\rH}^{S^1, \,(p-1)}_{\bullet}(\LL X, \mathbb{Q})
\end{equation*}
compatible with the isomorphisms \eqref{jonesiso}.
\end{theorem}
We prove Theorem~\ref{T2} by dualizing the classical Hodge decomposition \eqref{HCcomm}
for Sullivan's commutative DG algebra model of $ X $ constructed in \cite{BFG}.
In Section~\ref{Sect3}, we turn to derived Poisson structures. This notion was introduced in \cite{BCER} as a natural homological extension of the notion of an $\mathrm H_0$-Poisson structure proposed by Crawley-Boevey \cite{CB}.
Roughly speaking, a Poisson structure on an (augmented)
DG algebra $A$ is a DG Lie algebra structure on the cyclic space
$\, A_\n := A/(k+[A,A])\,$ induced by derivations of $A$ (see Section~\ref{Sect3.1}
for a precise definition). We introduce a category of Poisson DG algebras,$\, \mathtt{DGPA}_k \,$,
together with an appropriate class of weak equivalences. Ideally, one would like to make
$ \mathtt{DGPA}_k $ a (closed) model category in order to have a well-behaved homotopy category $ {\mathtt{Ho}}(\mathtt{DGPA}_k) $ to consider homotopy invariant structures and derived functors on $ \mathtt{DGPA}_k $. Although we do not achieve this goal in the present paper, we observe (see Proposition~\ref{homcat}) that $ \mathtt{DGPA}_k $ has a weaker property of being a (saturated) {\it homotopical category} in the sense of Dwyer-Hirschhorn-Kan-Smith \cite{DHKS}. Thanks to general results of \cite{DHKS}, this property still allows one to define a homotopy category $ {\mathtt{Ho}}(\mathtt{DGPA}_k) $ that has good formal properties and supports a meaningful theory of derived functors on $ \mathtt{DGPA}_k $. Having defined the homotopy category $ {\mathtt{Ho}}(\mathtt{DGPA}_k)$ of Poisson DG algebras, we then define a {\it derived Poisson algebra} to be simply an object\footnote{For technical reasons, we will also assume that the associative DG algebra $A$ on which we put a Poisson structure is cofibrant as an object of $ \mathtt{DGA}_{k/k} $.} of $\, {\mathtt{Ho}}(\mathtt{DGPA}_k)$. The key result here is Proposition~\ref{lieonft} which says that the cyclic
homology of any derived Poisson algebra $A$ carries a well-defined bracket $\,\{\,\mbox{--}\,,\mbox{--} \,\}: \overline{\mathrm{HC}}_\bullet(A) \times \overline{\mathrm{HC}}_\bullet(A)
\to \overline{\mathrm{HC}}_\bullet(A) \,$, making $ \overline{\mathrm{HC}}_\bullet(A) $ a (graded) Lie algebra.
In Section~\ref{Sect3.1}, following \cite{BCER}, we consider a particular class of derived Poisson structures on $ \U\mathfrak{a} $ that arise from a cyclic pairing on the Koszul dual coalgebra
of $ \mathfrak{a} $. We call such derived Poisson structures {\it cyclic}. The following theorem clarifies the relation between cyclic Poisson structures and the Hodge decompostion \eqref{hodgeds} of $ \U\mathfrak{a} $ ({\it cf.} Theorem~\ref{thodgefiltration} in Section~\ref{Sect3}).
\begin{theorem}
\la{T3}
The Lie bracket on $ \overline{\mathrm{HC}}_\bullet(\U\mathfrak{a}) $ induced by a cyclic derived Poisson structure
on $ \U\mathfrak{a} $ preserves the {\rm Hodge filtration:}
\begin{equation*}
F_p\overline{\mathrm{HC}}_{\bullet}(\U\mathfrak{a})\,:=\,
\bigoplus_{r \leq p+2} \mathrm{HC}^{(r)}_{\bullet}(\mathfrak{a}) \ ,
\end{equation*}
thus making $ \overline{\mathrm{HC}}_\bullet(\U\mathfrak{a}) $ a filtered Lie algebra. Moreover, in general, we have
$$
\{\mathrm{HC}^{(1)}_{\bullet}(\mathfrak{a}), \,\mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})\} \,\subseteq \,
\mathrm{HC}^{(p-1)}_{\bullet}(\mathfrak{a}) \quad \mbox{and}\quad
\{\mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a}),\, \mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})\} \subseteq \, \mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})\ ,
\quad \forall\,p \geq 1\ .
$$
In particular, $\mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a})$ is a Lie subalgebra of $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})$ and $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})$ is a graded Lie module over $\mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a})$.
\end{theorem}
We illustrate Theorem~\ref{T3} in a number of explicit examples, which include Abelian, unimodular and necklace Lie algebras (see Section~\ref{Sect3.5}). In particular, in the case of necklace Lie algebras, we show that
the Lie bracket on $ \overline{\mathrm{HC}}_\bullet(\U\mathfrak{a}) $ induced by a cyclic Poisson structure does {\it not} preserve the Hodge decomposition \eqref{hodgeds}. Thus, the result of Theorem~\ref{T3} cannot be strengthened in the obvious way.
Finally, we apply the results of Theorem~\ref{T2} and Theorem~\ref{T3} to the string topology Lie algebra of Chas and Sullivan \cite{ChS}. We recall
that Chas and Sullivan have shown that the $S^1$-equivariant homology
$ {\rH}^{S^1}_{\bullet}(\LL M, \mathbb{Q}) $ of the free loop space of any
smooth compact oriented manifold $M$ carries a natural Lie algebra
structure. Their construction uses the transversal intersection product of chains which is difficult to realize algebraically
in general ({\it cf.} \cite{CEG}). However, if $M$ is simply connected,
a theorem of Lambrechts and Stanley \cite{LS} provides a
finite-dimensional commutative DG algebra model
for $M$, whose linear dual coalgebra is Koszul
dual to Quillen's Lie model $ \mathfrak{a}_M $. The Lambrechts-Stanley
algebra comes with a natural cyclic pairing which yields a cyclic pairing
on the dual coalgebra. Now, it turns out that the associated cyclic Poisson
structure on $ \U \mathfrak{a}_M $ induces a Lie bracket on $ \overline{\mathrm{HC}}_\bullet(\U\mathfrak{a}_M) $ that corresponds (under the isomorphism \eqref{jonesiso})
precisely to the Chas-Sullivan bracket on $ {\rH}^{S^1}_{\bullet}(\LL M, \mathbb{Q}) $. As mentioned above, this fact was the main motivation for us to introduce and study derived Poisson structures in general (see \cite{BCER}).
Our last theorem in this Introduction gathers together some properties of the Chas-Sullivan Lie algebras arising from results of the present paper.
({\it cf.} Theorem~\ref{tstringhomology} in Section~\ref{Sect4}).
\begin{theorem}
\la{T4}
Let $M$ be a simply connected smooth oriented closed manifold of dimension $d$.
$(i)$
The string topology Lie algebra of $M$ is filtered as a Lie algebra, with Lie
bracket of degree $ 2-d $ preserving the following {\rm Hodge filtration}
$$
F_p \rH^{S^1}_{\bullet}(\LL M, \mathbb{Q}) \,:=\, \bigoplus_{q \leq p+1} \rH^{S^1,\, (q)}_{\bullet}(\LL M, \mathbb{Q})\,\text{.}
$$
\noindent
$(ii)$ The homology $ \rH_{\bullet}(\LL M, \mathbb{Q})$ of the free loop space $ \LL M$ is filtered as a Lie module over the string topology Lie algebra of $M$ with respect to the following Hodge filtration
$$ F_p\overline{\mathrm H}_{\bullet}(\mathcal L M, \mathbb{Q}) \,:=\, \bigoplus_{q \leq p+2} \overline{\mathrm H}^{(q)}_{\bullet}(\mathcal L M, \mathbb{Q})\,\text{.} $$
\noindent
$(iii)$ The string topology Lie bracket restricts to the first Hodge component $\, {\rH}^{S^1,\,(1)}_{\bullet}(\mathcal L M, \mathbb{Q})$, making it a Lie algebra. Further, $ {\rH}^{S^1}_{\bullet}(\mathcal L M, \mathbb{Q})$ is a graded Lie module over $\overline{\mathrm H}^{S^1,\,(1)}_{\bullet}(\mathcal L M, \mathbb{Q})$ with the grading given by the Hodge decomposition of $\,\overline{\mathrm H}^{S^1}_{\bullet}(\mathcal L M, \mathbb{Q})$.
\end{theorem}
We believe that, unlike Theorem~\ref{T3}, Theorem~\ref{T4} may be strengthened. In particular, we expect that the Chas-Sullivan Lie bracket of a closed $d$-dimensional manifold preserves not only the Hodge filtration but actually the Hodge decomposition \eqref{hodgeds}, thus making the string topology Lie algebra
a {\it graded} Lie algebra with respect to $p$-degree:
$$
\{\overline{\mathrm H}^{S^1,\,(p)}_{\bullet}(\mathcal L M, \mathbb{Q}), \ \overline{\mathrm H}^{S^1,\,(q)}_{\bullet}(\mathcal L M, \mathbb{Q})\}\, \subseteq\, \overline{\mathrm H}^{S^1,\, (p+q-1)}_{\bullet}(\mathcal L M, \mathbb{Q})\ , \quad \forall\,p,q \ge 1\ .
$$
This is part of a more general Conjecture~\ref{conj1} stated in Section~\ref{Sect4}.
\subsection*{Notation}
Throughout this paper, we denote by $\mathtt{DGA}_{k/k}$ (resp., $\mathtt{DGCA}_{k/k}$) the category of augmented, non-negatively graded DG algebras (resp., commutative DG algebras). The category of non-negatively graded coaugmented, conilpotent DG coalgebras (resp., cocommutative DG coalgebras) will be denoted by $\mathtt{DGC}_{k/k}$ (resp., $\mathtt{DGCC}_{k/k}$). The category of non-negatively graded DG Lie algebras will be denoted by $\mathtt{DGLA}_k$.
For $R \in\mathtt{DGA}_{k/k}$, let $R_\n\,:=\, R/(k+[R,R])$. For an $R$-bimodule $M$, let $M_\n\,:=\, M/[R,M]$. Let $\Omega^1R$ denote the kernel of the multiplication map $R \otimes R \xrightarrow{} R$ (where $R \otimes R$ is equipped with the outer $R$-bimodule structure). The DG $R$-bimodule $\Omega^1R$ represents the complex of derivations $\underline{\mathrm{Der}}(R,M)$. In particular, the universal derivation $\partial\,:\, R \xrightarrow{} \Omega^1R$ is given by $r \mapsto r \otimes 1-1 \otimes r$.
\section{Hodge decomposition for universal enveloping algebras}
\la{S2}
In this section, we construct a Hodge decomposition of Hochschild homology for the universal enveloping algebras of DG Lie algebras. The corresponding Hodge components are defined in terms of derived functors, similar to the definition of the cyclic Lie-Hodge homology given in~\cite{bfprw1}. The main results of this sections (Theorem~\ref{hodgesbi} and
Theorem~\ref{connesvsgksbi}) explain the relation of our construction to the earlier work of Getzler and Kapranov \cite{GK}.
\subsection{Hodge decomposition of Hochschild homology}
\la{ns1.1}
Let $\mathfrak{a}$ be a DG Lie algebra, and let $V$ be a right DG $\mathfrak{a}$-module.
Using the action map on $V$:
$$
V \otimes \mathfrak{a} \to V\ ,\quad v\otimes x \mapsto v \cdot x\ ,
$$
we define a complex of vector spaces, $\, \theta(\mathfrak{a},V) \,$, by factoring $V \otimes \mathfrak{a}$ modulo the relations
\begin{equation*}
\la{relunat}
v \otimes [x,y] \,=\, v \cdot x \otimes y- (-1)^{|x||y|} v \cdot y \otimes x \ ,\quad \forall\,v \in V\ ,\ \forall\, x,\,y \in \mathfrak{a}\ .
\end{equation*}
The action map $\,V \otimes \mathfrak{a} \to V\,$ then factors through $\theta(\mathfrak{a}, V)$, giving a canonical
morphism of complexes
\begin{equation}
\la{betav}
\beta_V :\, \theta(\mathfrak{a}, V) \xrightarrow{} V \,\text{.}
\end{equation}
For example, if $V=k$ is the trivial $\mathfrak{a}$-module, then
$\,\theta(\mathfrak{a}, k)\,\cong\,\mathfrak{a}/[\mathfrak{a},\mathfrak{a}]\,$, and \eqref{betav} is the zero map. On the other hand,
for $V\,=\, \mathcal U\mathfrak{a}$ equipped with the (right) adjoint action of $\mathfrak{a}\,$, the complex $\theta(\mathfrak{a}, \U\mathfrak{a})$ can be identified with $\Omega^1(\U\mathfrak{a})_\n\,$, and the map \eqref{betav} is given by
\begin{equation}
\la{betau}
\beta_{\U}:\, \Omega^1(\mathcal U\mathfrak{a})_\n \to \U\mathfrak{a}\ ,\quad
[\alpha \otimes y] \mapsto [\alpha,y]\ ,
\end{equation}
where $\alpha\,\in\,\mathcal U\mathfrak{a}$, $\,y\,\in\,\mathfrak{a}$, and $[\alpha \otimes y ]$ denotes the class of $\alpha \otimes y$ in $\Omega^1(\mathcal U\mathfrak{a})_\n$.
Now, fix an integer $\,p\ge1 \,$ and let $ V = {\rm{Sym}}^p(\mathfrak{a}) $ be the $p$-th symmetric power of $ \mathfrak{a} $ equipped with the right adjoint action. Write
$ \beta_p(\mathfrak{a}):\,\theta(\mathfrak{a}, {\rm{Sym}}^p(\mathfrak{a})) \to {\rm{Sym}}^p(\mathfrak{a}) $ for the corresponding
map \eqref{betav} and define
$$
\phi^{(p)}(\mathfrak{a}) := \mathrm{Cone}[\beta_p(\mathfrak{a})]\ .
$$
Note that the functor $\lambda^{(p)}$ introduced in \eqref{lam} is given by
$ \lambda^{(p)}(\mathfrak{a})\,=\,{\rm{Coker}}[\beta_p(\mathfrak{a})]\,$; hence, there is a natural transformation of functors
%
\begin{equation} \la{natI}
\mathcal I :\, \phi^{(p)} \to \lambda^{(p)} \ ,
\end{equation}
defined by the canonical map $\,\mathrm{Cone}[\beta_p(\mathfrak{a})] \to {\rm{Coker}}[\beta_p(\mathfrak{a})]\,$.
Next, for $p=0$, we set $\,\phi^{(0)}(\mathfrak{a})\,:=\, \theta(\mathfrak{a}, k)[1]\,$ and define $\phi(\mathfrak{a})$ to be the cone of the map \eqref{betau} composed with the natural projection:
$$
\phi(\mathfrak{a}) := \mathrm{Cone}[\,\Omega^1(\mathcal U\mathfrak{a})_\n \xrightarrow{\beta_{\U}} \U\mathfrak{a} \,\twoheadrightarrow\, \overline{\U\mathfrak{a}}\,] \ .
$$
The proof of the following proposition is similar to that of~\cite[Theorem~7.1]{bfprw1}:
we give details for reader's convenience.
\begin{proposition} \la{lderived}
The functors $\phi^{(p)}$ and $\phi$ have (total) left derived functors $\boldsymbol{L}\phi^{(p)}$ and $\,\boldsymbol{L}\phi\,$ from ${\mathtt{Ho}}(\mathtt{DGLA}_k)$ to $\,{\mathtt D}(k)$.
\end{proposition}
\begin{proof}
First, observe that for any commutative DG algebra $B$, there are natural maps of complexes
$$
\phi^{(p)}(\mathfrak{a} \otimes B) \xrightarrow{} \phi^{(p)}(\mathfrak{a}) \otimes B\quad ,\quad
\phi(\mathfrak{a} \otimes B) \xrightarrow{} \phi(\mathfrak{a}) \otimes B\
$$
obtained by extending scalars in the above constructions.
Next, by \cite[Proposition~B.2]{BKR}, if $\mathcal{L}$ is a cofibrant DG Lie algebra,
two morphisms $f,g\,:\,\mathcal{L} \xrightarrow{} \mathfrak{a}$ are homotopic in $ \mathtt{DGLA}_k $ iff there is a DG Lie algebra
homomorphism $h\,:\,\mathcal{L} \xrightarrow{} \mathfrak{a} \otimes \Omega$ such that $h(0)=f$ and $h(1)=g$.
(We call such a homomorphism a {\it homotopy} from $f$ to $g$.). Here $\Omega\,:=\,\Omega(\mathbb A^1_k)$ is the de Rham algebra of the affine line (equipped with homological grading), and $ h(a)$ denotes
the composite map of $ h $ with $ {\rm{Id}} \otimes \mathrm{ev}_a$, where
$\mathrm{ev}_a\,:\,\Omega \xrightarrow{} k$ is the evaluation map at point $ a \in {\mathbb A}^1_k $.
Now, if $\mathcal{L}$ is a cofibrant DG Lie algebra and $f,g :\,\mathcal{L} \xrightarrow{} \mathfrak{a}$ are homotopic, with $h$ being a homotopy from $f$ to $g$, one has the map $ H:\,\phi(\mathcal{L}) \xrightarrow{} \phi(\mathfrak{a}) \otimes \Omega $ given by the composition
$$
\phi(\mathcal{L}) \xrightarrow{\phi(h)} \phi(\mathfrak{a} \otimes \Omega) \to \phi(\mathfrak{a}) \otimes \Omega\ .
$$
A straightforward verification shows that $ H(0)=\phi(f)$ and $ H(1)=\phi(g)$. Thus, if $f,g\,:\, \mathcal{L} \xrightarrow{} \mathfrak{a}$ are homotopic and $\mathcal{L}$ is cofibrant, then $\phi(f)$ and $\phi(g)$ are homotopic as morphisms of complexes.
If $f\,:\,\mathcal{L} \stackrel{\sim}{\to} \mathcal{L}'$ is a weak equivalence between two cofibrant objects
in $\mathtt{DGLA}_k$, by Whitehead's Theorem (see~\cite[Lemma~4.24]{DS}), there exists a map $g\,:\, \mathcal{L}' \xrightarrow{} \mathcal{L}$, such that $fg$ and $gf$ are homotopic to the identities of $ \mathcal{L}' $ and $\mathcal{L}$, respectively.
It follows that $\phi(fg)$ and $\phi(gf)$ are homotopic to the identities of $\phi(\mathcal{L}')$ and $\phi(\mathcal{L})$, and therefore $\phi(f)$ is a quasi-isomorphism. Hence the functor $\phi$ takes weak equivalences between cofibrant objects to weak equivalences. A similar argument
shows that the functors $\phi^{(p)}$ take weak equivalences between cofibrant objects to weak equivalences. The existence of $\,\boldsymbol{L}\phi\,$ and $\,\boldsymbol{L}\phi^{(p)}\,$ follows now from Brown's Lemma in abstract homotopy theory ({\it cf.}~\cite[Lemma 9.9]{DS}).
\end{proof}
Now, with Proposition~\ref{lderived}, we may define
\begin{equation*}
\la{hhodge}
{\rm{HH}}^{(p)}_{\bullet}(\mathfrak{a})\,:=\,
\mathrm H_{\bullet}[\boldsymbol{L}\phi^{(p)}(\mathfrak{a})] \ , \quad p\ge 0 \ ,
\end{equation*}
and state first main result of this section:
\begin{theorem} \la{hodgehoch}
For any DG Lie algebra $\mathfrak{a}$, there is a functorial direct sum decomposition
\begin{equation}
\la{hodgeds1}
\overline{{\rm{HH}}}_{\bullet}(\mathcal U\mathfrak{a}) \,\cong\,\bigoplus_{p=0}^{\infty} {\rm{HH}}^{(p)}_{\bullet}(\mathfrak{a})\,\text{.}
\end{equation}
\end{theorem}
By the Poincar\'e-Birkhoff-Witt Theorem (see, e.g., \cite{Q2}, Appendix~B, Theorem~2.3), the symmetrization maps \eqref{symfun} give a natural isomorphism of right DG $\mathfrak{a}$-modules $\,
\bigoplus_{p=0}^{\infty} {\rm{Sym}}^p(\mathfrak{a}) \cong \U\mathfrak{a}\,$. This isomorphism
induces an isomorphism of functors $\, \bigoplus_{p=0}^{\infty} \phi^{(p)} \cong \phi\,$,
which, in turn, induces an isomorphism of the corresponding derived functors
from ${\mathtt{Ho}}(\mathtt{DGLA}_k)$ to ${\mathtt D}(k)\,$:
\begin{equation*}
\la{Lphis}
\bigoplus_{p=0}^{\infty} \boldsymbol{L}\phi^{(p)} \,\cong\,\boldsymbol{L}\phi\,\text{.}
\end{equation*}
To prove Theorem~\ref{hodgehoch} it thus suffices to prove the following proposition.
\begin{proposition}
\la{lrhh}
For any $\mathfrak{a}\,\in\,\mathtt{DGLA}_k$, there is a natural isomorphism $ \mathrm H_{\bullet}[\boldsymbol{L}\phi(\mathfrak{a})]\,\cong\, \overline{{\rm{HH}}}_{\bullet}(\mathcal U\mathfrak{a})$.
\end{proposition}
\begin{proof}
Let $C\,\in\,\mathtt{DGCC}_{k/k}$ be a cocommutative coalgebra Koszul dual to $\mathfrak{a}\,\in\,\mathtt{DGLA}_k$.
Assume that $\bar{C}$ is concentrated in strictly positive homological degrees. Then $\,R\,:=\,\boldsymbol{\Omega}(C) \stackrel{\sim}{\xrightarrow{}} \U\mathfrak{a}\,$ is a cofibrant resolution of $ \U\mathfrak{a} $ in $\mathtt{DGA}_{k/k}$ and $\mathcal L\,:=\, \boldsymbol{\Omega}_{\mathtt{Comm}}(C) \stackrel{\sim}{\xrightarrow{}} \mathfrak{a} $ is a cofibrant resolution
of $\mathfrak{a} $ in $\mathtt{DGLA}_k$. Since $\, R\,\cong\,\,\ U\mathcal L\,$, we have $\theta(\mathcal L, R)\,\cong\,\Omega^1R_\n$. Now, let $ \beta(\mathcal L) $ denote the map \eqref{betau}, with $ \mathfrak{a} $ replaced by $ \mathcal L $ and composed with the natural projection:
$$
\beta(\mathcal L):\, \Omega^1R_\n \xrightarrow{} \bar{R}\,,\,\,\,\,\, \alpha \otimes v \mapsto [\alpha,v]\,\text{.}
$$
Then $\,\boldsymbol{L}\phi(\mathfrak{a}) \cong \mathrm{Cone}[\beta(\mathcal L)]\,$ in the derived category
${\mathtt D}(k)$. Hence, Proposition~\ref{lrhh} is a consequence of the following more general assertion: if $\,R \stackrel{\sim}{\xrightarrow{}} A\,$ be a semi-free resolution of $A$ in $\mathtt{DGA}_{k/k}$, then
$$
\overline{{\rm{HH}}}_{\bullet}(A)\,\cong\, \mathrm H_{\bullet}(\mathrm{Cone}[\beta(R)])\,\text{.}
$$
To prove this assertion, we will use the approach of \cite[Section 5]{BKR}, which, in turn, is based on Quillen's results \cite{Q}. First, we notice that $ R $ is isomorphic to the tensor algebra $TV$ of a graded vector space $V$; hence, there is an isomorphism of $R$-bimodules $\,I:\,R \otimes V \otimes R \xrightarrow{} \Omega^1R\,$ given by (see~\cite[Example 3.10]{Q} or~\cite[Section 2.3]{CEEY})
\begin{align*}
& (v_1,\ldots, v_{i-1}) \otimes v_i \otimes (v_{i+1}, \ldots ,v_n) \mapsto (v_1 ,\ldots v_{i-1},v_i) \otimes (v_{i+1}, \ldots, v_n) - (v_1,\ldots ,v_{i-1}) \otimes (v_i,v_{i+1} ,\ldots ,v_n) \,\text{.}
\end{align*}
The inverse map $I^{-1}$ induces an isomorphism of graded vector spaces
\begin{align} \la{omeganat}
\Omega^1R_\n\,\cong\, R \otimes V \,\text{.}
\end{align}
Under \eqref{omeganat}, the map $\bar{\partial}\,:\, \bar{R} \xrightarrow{} \Omega^1R_\n$ becomes\footnote{As explained in~\cite[Appendix A]{BR}, this map may be called the {\it cyclic de Rham differential}. Its kernel is $[\bar{R}, \bar{R}]$ (see {\it loc. cit}).}
\begin{equation}
\la{cyclicderham} \
\bar{\partial}(v_1, \ldots, v_m) \,=\, \sum_{i=1}^m (-1)^{(|v_1|+ \ldots+|v_i| )(|v_{i+1}|+\ldots +|v_m|)} (v_{i+1},\ldots ,v_m,v_1 ,\ldots ,v_{i-1}) \otimes v_i\,\text{.}
\end{equation}
Now, there is a first quadrant bicomplex (see \cite[(5.21)]{BKR})
\begin{equation}
\la{xc}
X^{+}(R)\,:=\,
[\, 0 \xleftarrow{} \bar{R} \xleftarrow{\beta} \Omega^1R_\n \xleftarrow{\bar{\partial}} \bar{R} \xleftarrow{\beta} \ldots\, ] \ , \end{equation}
for which the quotient map
$\,\mathrm{Tot}\,X^+(R) \stackrel{\sim}{\twoheadrightarrow} R_\n\,$
defined by the canonical projection from the first column is a quasi-isomorphism; thus,
$$
\mathrm H_{\bullet}[\mathrm{Tot}\,X^+(R)]\,\cong\, \overline{\mathrm{HC}}_{\bullet}(A)\,\text{.}
$$
By the above identifications, the total complex $\,\mathrm{Tot}\,X_2^+(R) $ of the sub-bicomplex $\,X_2^+(R)\,$ of $\,X^+(R)\,$ comprising of the first two columns is precisely
$ \mathrm{Cone}[\beta(R)]\,$. On the other hand, by \cite[(5.25)]{BKR}),
\begin{equation} \la{x2c} \mathrm H_{\bullet}[\mathrm{Tot}\,X_2^+(R)]\,\cong\, \overline{{\rm{HH}}}_{\bullet}(A)\,\text{.}\end{equation}
This proves the desired proposition.
\end{proof}
\subsection{Hodge decomposition of the Connes periodicity sequence}
\la{s1.3}
One of the fundamental properties of cyclic homology is the Connes periodicity exact sequence ({\it cf.}~\cite[2.2.13]{L}).
\begin{equation} \la{connessbi} \begin{diagram}[small] \ldots & \rTo^S & \overline{\mathrm{HC}}_{n-1}(A) & \rTo^{B} & \overline{{\rm{HH}}}_n(A) & \rTo^I & \overline{\mathrm{HC}}_n(A) & \rTo^S & \overline{\mathrm{HC}}_{n-2}(A) & \rTo & \ldots \end{diagram} \,\text{.}\end{equation}
This sequence involves two important operations on cyclic homology: the periodicity operator $S$ and the Connes differential $B$.
\begin{theorem}
\la{hodgesbi} Let $\mathfrak{a}\,\in\,\mathtt{DGLA}_k$. The Connes periodicity sequence for $\mathcal U\mathfrak{a}$ decomposes into a direct sum of Hodge components: the summand of Hodge degree $p$ is given by the long exact sequence
\begin{equation} \la{conneshodgep} \begin{diagram}[small] \ldots & \rTo^S & \mathrm{HC}_{n-1}^{(p+1)}(\mathfrak{a}) & \rTo^{B} & {{\rm{HH}}}_n^{(p)}(\mathfrak{a}) & \rTo^I & \mathrm{HC}_n^{(p)}(\mathfrak{a}) & \rTo^S & \mathrm{HC}_{n-2}^{(p+1)}(\mathfrak{a}) & \rTo & \ldots \end{diagram} \,, \end{equation}
with the map $I\,:\,{{\rm{HH}}}_n^{(p)}(\mathfrak{a}) \xrightarrow{} \mathrm{HC}_n^{(p)}(\mathfrak{a})$ induced on homologies by the natural transformation $\mathcal I\,:\, \boldsymbol{L}\phi^{(p)}(\mathfrak{a}) \xrightarrow{} \boldsymbol{L}\lambda^{(p)}(\mathfrak{a})$.
\end{theorem}
In order the prove Theorem~\ref{hodgesbi}, note that $R$ is freely generated by $V\,:=\,\bar{C}[-1]$ as a graded $k$-algebra. For notational brevity, let $R^{(p)}$ denote the image of ${\rm{Sym}}^p(\mathcal L)$ in $R\,\cong\, \mathcal U\mathcal L$ under the symmetrization map. Under the isomorphism~\eqref{omeganat}, the direct summand $\theta^{(p)}(\mathcal L)$ of $\Omega^1R_\n$ is identified with $R^{(p)} \otimes V$.
\begin{lemma} \la{partialbeta}
For any $p \geq 1$, $\bar{\partial}R^{(p)} \subseteq \theta^{(p-1)}(\mathcal L)$ and $\beta[\theta^{(p)}(\mathcal L)] \subseteq R^{(p)}$.
\end{lemma}
\begin{proof}
The inclusion $\bar{\partial}R^{(p)} \subseteq \theta^{(p-1)}(\mathcal L)$ follows immediately from Lemma~\ref{cycderhamhodge} proved in the appendix. The inclusion $\beta[\theta^{(p)}(\mathcal L)] \subseteq R^{(p)}$ is a consequence of $V \subseteq \mathcal L$ and $[{\rm{Sym}}^p(\mathcal L), \mathcal L] \subseteq {\rm{Sym}}^p(\mathcal L)$ in $\mathcal U\mathcal L$ (where we think of the symmetric powers of $\mathcal L$ as subcomplexes of $\mathcal U\mathcal L$ via the symmetrization map).
\end{proof}
We now proceed with
\begin{proof}[Proof of Theorem~\ref{hodgesbi}]
Let $X^+(R)$ (resp., $X_2^+(R)$) be as in~\eqref{xc} (resp.,~\eqref{x2c}). Recall from~\cite[Section 5]{BKR} that there is an exact sequence of bicomplexes
\begin{equation} \la{xcomplex} 0 \xrightarrow{} X_2^+(R) \xrightarrow{} X^+(R) \xrightarrow{} X^+(R)[2,0] \xrightarrow{} 0\,\text{.} \end{equation}
At the level of total complexes, this gives the exact sequence
\begin{equation} \la{xcomplextot} \begin{diagram}[small] 0 & \rTo & \mathrm{Tot}\,X_2^+(R) & \rTo^I & \mathrm{Tot}\,X^+(R) & \rTo^S & \mathrm{Tot}\,X^+(R)[2] & \rTo & 0 \end{diagram}\,,\end{equation}
which induces the Connes periodicity sequence on homologies. As an immediate consequence of Lemma~\ref{partialbeta}, we get a direct sum decomposition of bicomplexes
$$X^+(R) \,=\, \bigoplus_{p=0}^{\infty} X^{+,(p)}(\mathfrak{a})\,,$$
where
$$X^{+,(p)}(\mathfrak{a})\,:=\, [ \begin{diagram}[small] 0 & \lTo& {R}^{(p)} & \lTo^{\beta} & \theta^{(p)}(\mathcal L) & \lTo^{\bar{\partial}} & {R}^{(p+1)} & \lTo^{\beta} & \theta^{(p+1)}(\mathcal L) & \lTo^{\bar{\partial}} & {R}^{(p+2)} & \lTo^{\beta} & \ldots \end{diagram} ] \,\text{.}$$
In particular, $X_2^+(R) \,=\, \bigoplus_{p=0}^{\infty} X_2^{+,(p)}(\mathfrak{a})\,,$ where $X_2^{+, (p)}(\mathfrak{a})$ is the sub-bicomplex of $X^{+,(p)}(\mathfrak{a})$ comprising its first two columns. Note that
$$ \mathrm{Tot}\,X_2^{+, (p)}(\mathfrak{a}) \,\cong\, \boldsymbol{L}\phi^{(p)}(\mathfrak{a}) $$
in ${\mathtt D}(k)$, whence $\mathrm H_{\bullet}[\mathrm{Tot}\,X_2^{+, (p)}(\mathfrak{a})]\,\cong\,{\rm{HH}}^{(p)}_{\bullet}(\mathfrak{a})$. Further, since the composite map $\mathrm{Tot}\,X^+(R) \twoheadrightarrow R_\n $
is a quasi-isomorphism, $\mathrm{Tot}\,X^{+,(p)}(\mathfrak{a})$ is quasi-isomorphic to $\lambda^{(p)}(\mathcal L)$. It follows that $\mathrm H_{\bullet}[\mathrm{Tot}\,X^{+,(p)}(\mathfrak{a})]$ is isomorphic to $\mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})$.
In addition, the exact sequence~\eqref{xcomplex} decomposes as a direct sum of Hodge components for $p \geq 0$, with the $p$-th Hodge component given by
\begin{equation} \la{xcomplexhp} 0 \xrightarrow{} X_2^{+,(p)}(\mathfrak{a}) \xrightarrow{} X^{+,(p)}(\mathfrak{a}) \xrightarrow{} X^{+,(p+1)}(\mathfrak{a})[2,0] \xrightarrow{} 0\,\text{.} \end{equation}
At the level of total complexes, this gives a Hodge decomposition of the exact sequence~\eqref{xcomplextot}, with the summand in Hodge degree $p$ being
\begin{equation} \la{xcomplexhptot}\begin{diagram}[small] 0 & \rTo & \mathrm{Tot}\,X_2^{+,(p)}(\mathfrak{a}) & \rTo^I & \mathrm{Tot}\,X^{+,(p)}(\mathfrak{a}) & \rTo^S & \mathrm{Tot}\,X^{+,(p+1)}(\mathfrak{a})[2] & \rTo & 0 \end{diagram}\,\text{.}\end{equation}
The long exact sequence on homologies corresponding to~\eqref{xcomplexhptot} is~\eqref{conneshodgep}. Finally, since the projection $\mathrm{Tot}\,X^{+,(p)}(\mathfrak{a}) \twoheadrightarrow \lambda^{(p)}(\mathcal L)$ is a quasi-isomorphism, the natural transformation $\mathcal I\,:\, \boldsymbol{L}\phi^{(p)}(\mathfrak{a}) \xrightarrow{} \boldsymbol{L}\lambda^{(p)}(\mathfrak{a})$ is represented by the inclusion of complexes $I$ in~\eqref{xcomplexhptot}. This proves the desired theorem.
\end{proof}
When $p=0$, the column in degree $0$ of $X_2^{+,(p)}(\mathfrak{a})$, namely $R^{(p)}$ vanishes while the column of degree $1$ is isomorphic to $V\,=\,\bar{C}[-1]$. Thus, $X_2^{+,(0)}(\mathfrak{a}) \,\cong\, \bar{C}$, whence ${\rm{HH}}^{(0)}_{\bullet}(\mathfrak{a})\,\cong\, \mathrm H(\mathfrak{a};k)$, except in degree $0$ where it vanishes. Since $\mathrm{HC}^{(0)}(\mathfrak{a})$ vanishes and since $\mathrm{HC}^{(1)}_{\bullet}(\mathfrak{a})\,\cong\,\mathrm H_{\bullet+1}(\mathfrak{a};k)$, the long exact sequence~\eqref{conneshodgep} for $p=0$ becomes:
$$ \begin{diagram}[small] \ldots &\rTo&\mathrm H_n(\mathfrak{a};k) & \rTo^{\mathrm{Id}} & {\mathrm H}_n(\mathfrak{a};k) & \rTo& 0 & \rTo & \mathrm H_{n-1}(\mathfrak{a};k) & \rTo^{\mathrm{Id}} & \ldots \end{diagram} \,,$$
with the $B$ map being $\mathrm{Id}$ and the $S$ and $I$ maps vanishing.
\subsection{Comparison to the Getzler-Kapranov Lie cyclic homology}
\la{GKcomp}
We recall that for any cyclic operad $\mathcal P$ and any (DG) $\mathcal P$-algebra $A$, Getzler and Kapranov introduced the (operadic) $\mathcal P$-cyclic homology $\mathrm{HA}_{\bullet}(\mathcal P, A)$ as the homology of the left derived functor of the universal invariant bilinear form on $A$ (see~\cite[Section 4.7, Section 5]{GK}). In addition, they introduced the homologies $\mathrm{HB}_{\bullet}(\mathcal P, A)$ and $\mathrm{HC}_{\bullet}(\mathcal P, A)$ that form the long exact sequence (see~\cite[Section 5.8]{GK})
\begin{equation}
\la{gkseqop}
\ldots \,\to\, \mathrm{HA}_{n}(\mathcal P, A) \,\to\,
{\mathrm{HB}}_{n}(\mathcal P, A) \,\to\, \mathrm{HC}_{n}(\mathcal P, A) \,\to\,
\mathrm{HA}_{n-1}(\mathcal P, A)\,\to\, \ldots
\end{equation}
which we call the Connes periodicity sequence for (operadic) $\mathcal P$-cyclic homology.
It was shown in~\cite[Section 6.10]{GK} that
$$\mathrm{HB}_{n-1}(\mathtt{Lie}, \mathfrak{a})\,\cong\, \mathrm H_n(\mathfrak{a};\mathfrak{a})\,,\,\,\,\, \mathrm{HC}_{n-1}(\mathtt{Lie};\mathfrak{a})\,\cong\, \mathrm H_{n+1}(\mathfrak{a};k)\, \text{.}$$
Thus, $\mathrm{HC}_{n-1}(\mathtt{Lie};\mathfrak{a})\,\cong\,\mathrm{HC}^{(1)}_n(\mathfrak{a})$. Recall from~\cite[Section 5]{GK} that $\mathrm{HA}_{n-1}(\mathtt{Lie}, \mathfrak{a})\,\cong\,\mathrm{HC}^{(2)}_{n-1}(\mathfrak{a})$. Our final result in this section is:
\begin{theorem}
\la{connesvsgksbi}
For $p=1$, the Hodge component~\eqref{conneshodgep} of the Connes periodicity sequence coincides with the Connes periodicity sequence for (operadic) Lie cyclic homology (see~\cite[Section 6.10]{GK}).
\begin{equation}
\la{gkseq}
\ldots \,\to\, \mathrm{HA}_{n-1}(\mathtt{Lie}, \mathfrak{a})\,\to\, {\mathrm{HB}}_{n-1}(\mathtt{Lie}, \mathfrak{a}) \,\to\, \mathrm{HC}_{n-1}(\mathtt{Lie}, \mathfrak{a}) \,\to\,\mathrm{HA}_{n-2}(\mathtt{Lie}, \mathfrak{a}) \,\to\,
\ldots
\end{equation}
\end{theorem}
Before proving Theorem~\ref{connesvsgksbi}, we develop the necessary technical tools.
\subsubsection{}\la{s1.1}
Let $C\,\in\,\mathtt{DGCC}_{k/k}$ be Koszul dual to $\mathfrak{a}$. As shown by Quillen~\cite{Q}, $\boldsymbol{\Omega}(C)_\n\,\cong\, \overline{\mathrm{CC}}(C)[-1]$, where the right hand side is the reduced cyclic complex of $C$. Thus, $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})\,\cong\, \overline{\mathrm{HC}}_{\bullet+1}(C)$. Under this isomorphism, $\mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})$ is identified with the Hodge summand $\overline{\mathrm{HC}}_{\bullet+1}^{(p-1)}(C)$ coming from the Hodge decomposition of $\overline{\mathrm{HC}}_{\bullet}(C)$ which exists due to cocommutativity of $C$ (see~\cite[Proposition 7.4]{bfprw1}).
In particular, for the rest of this section, let $C$ be the Chevalley-Eilenberg coalgebra $\mathcal{C}(\mathfrak{a};k)$. The Hodge decomposition of $\overline{\mathrm{HC}}_{\bullet}(C)$ has an explicit description in terms of the de Rham coalgebra of $C$ (see~\cite[Section 2]{bfprw2}). This description is dual to that of the Hodge decomposition of the cyclic homology of a smooth commutative (in fact, symmetric) (DG) algebra $A$ in terms of the de Rham complex of $A$. Explicitly, let $\mathrm{DR}^{\bullet}(C)$ denote the mixed de Rham complex of $C$ (see~\cite[Section 2]{bfprw2} for the definition). The negative cyclic complex $\mathrm{CC}^{-}[\mathrm{DR}^{\bullet}(C)]$ has a Hodge decomposition due to cocommutativity of $C$. Let $\mathrm{CC}^{-,(p)}[\mathrm{DR}^{\bullet}(C)]$ denote the component with Hodge weight $p$. Then (see~\cite{bfprw2}, Proposition~2.2 and Theorem 2.4)
\begin{proposition} \la{formderham}
There are natural isomorphisms
$$\overline{\mathrm{HC}}_{\bullet}^{(p)}(C)\,\cong\, \mathrm H_{\bullet}(\mathrm{CC}^{-,(p)}[\mathrm{DR}^{\bullet}(C)])\,\cong\, \mathrm H_{\bullet+p}(\ker[d\,:\, \Omega^p_C \xrightarrow{} \Omega^{p-1}_C])\,,$$
where $d\,:\,\Omega^p_C \xrightarrow{} \Omega^{p-1}_C$ is the de Rham differential.
\end{proposition}
It is easy to see that $\Omega^p_C[-p]$ and $\mathcal{C}(\mathfrak{a};{\rm{Sym}}^p(\mathfrak{a}))$ are both isomorphic to ${\rm{Sym}}^p(\mathfrak{a}) \otimes \wedge \mathfrak{a}$ as graded vector spaces. Thus, $\mathrm{DR}^{\bullet}(C)$ is isomorphic to $\mathcal{C}(\mathfrak{a};{\rm{Sym}}(\mathfrak{a}))$, which is in turn isomorphic to the (reduced) de Rham algebra $\Omega^{\bullet}_{{\rm{Sym}}(\mathfrak{a})}$ as graded vector spaces. We complete this picture with the following
\begin{lemma} \la{chevalley}
Under the above isomorphism, the de Rham differential on $\mathrm{DR}^{\bullet}(C)$ is identified with the de Rham differential on $\Omega^{\bullet}_{\mathtt{Sym}(\mathfrak{a})/k}$ and the differential on $\Omega^p_C[-p]$ induced by the differential on $C$ is identified with the Chevalley-Eilenberg differential on $\mathcal{C}(\mathfrak{a};{\rm{Sym}}^p(\mathfrak{a}))$.
\end{lemma}
\begin{proof}
This lemma is formally dual to the following assertion: let $A\,:=\,\mathcal{C}^{\bullet}(\mathfrak{a};k)$ be the Chevalley-Eilenberg {\it cochain} complex of $\mathfrak{a}$ (which is isomorphic as a graded algebra to ${\rm{Sym}}(\mathfrak{a}^{\ast}[1])$. Then, $\Omega^p_A[-p]\,\cong\,\mathcal{C}^{\bullet}(\mathfrak{a};{\rm{Sym}}^p(\mathfrak{a}^{\ast}))$. For $p=1$, this assertion follows from a direct computation (see~\cite[Section 5.4]{L} for instance). For higher $p$, one notices that the natural map
${\rm{Sym}}^p_A(\Omega^1_A) \xrightarrow{} \mathcal{C}^{\bullet}(\mathfrak{a};{\rm{Sym}}^p(\mathfrak{a}^{\ast}))$ is compatible with differentials (it induces the $p$-fold cup product) and is an isomorphism of graded $k$-vector spaces.
\end{proof}
Thus the de Rham differential $d$ on $\Omega_{{\rm{Sym}}(\mathfrak{a})/k}$ (anti)commutes with the Chevalley-Eilenberg differential $\delta$ on $\mathcal{C}(\mathfrak{a};{\rm{Sym}}(\mathfrak{a}))$.
This makes $(\Omega^{\bullet}_{{\rm{Sym}}(\mathfrak{a})/k}, \delta, d)$ a mixed complex (since $d$ has degree $1$ while $\delta$ has degree $-1$) isomorphic to $\mathrm{DR}^{\bullet}(C)$. As a consequence of Proposition~\ref{formderham} and~\cite[Proposition 7.4]{bfprw1},
\begin{corollary} \la{liecyclic}
There is a natural isomorphism:
$$ \mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a}) \,\cong\, \mathrm H_{\bullet+1}[\ker(d\,:\,\mathcal{C}(\mathfrak{a};{\rm{Sym}}^{p-1}(\mathfrak{a})) \xrightarrow{} \mathcal{C}(\mathfrak{a};{\rm{Sym}}^{p-2}(\mathfrak{a}))[-1])] \,\text{.}$$
\end{corollary}
In particular, as shown in~\cite[Section~6.10]{GK},
$$ \mathrm{HC}^{(2)}_n(\mathfrak{a})\,\cong\, \mathrm H_{n+1}[\ker(d\,:\,\mathcal{C}(\mathfrak{a};\mathfrak{a}) \xrightarrow{} \mathcal{C}(\mathfrak{a};k)[-1])]\,\text{.} $$
As a consequence of \eqref{hodgeds} and Corollary~\ref{liecyclic}, we now obtain a different proof of the following result originally due to Kassel~\cite{K} (see also~\cite[Theorem 3.3.7]{L}).
\begin{corollary} \la{kassel}
$\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})$ is isomorphic to the cyclic homology of the mixed complex $(\Omega^{\bullet}_{{\rm{Sym}}(\mathfrak{a})}, \delta, d)$.
\end{corollary}
\begin{proof}
We have natural isomorphisms
\begin{eqnarray*}
\mathrm{HC}_{\bullet}(\Omega^{\bullet}_{{\rm{Sym}}(\mathfrak{a})}, \delta, d)\, &\cong &\, \mathrm H_{\bullet}[ \Omega^{\bullet}_{{\rm{Sym}}(\mathfrak{a})}/d(\Omega^{\bullet}_{{\rm{Sym}}(\mathfrak{a})}), \delta]\\
& \cong & \bigoplus_{p=1}^{\infty} \mathrm H_{\bullet}[\mathrm{coker}(d\,:\, \mathcal{C}(\mathfrak{a};{\rm{Sym}}^{p+1}(\mathfrak{a})) \xrightarrow{} \mathcal{C}(\mathfrak{a};{\rm{Sym}}^{p}(\mathfrak{a}))[-1])]\\
&\cong & \bigoplus_{p=1}^{\infty} \mathrm H_{\bullet+1}[\ker(d\,:\,\mathcal{C}(\mathfrak{a};{\rm{Sym}}^{p-1}(\mathfrak{a})) \xrightarrow{} \mathcal{C}(\mathfrak{a};{\rm{Sym}}^{p-2}(\mathfrak{a}))[-1])]\\
& \cong & \bigoplus_{p=1}^{\infty} \mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})\,\, (\text{by Corollary}~\ref{liecyclic})\\
& \cong & \overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})\quad (\text{by}~\eqref{hodgeds})\text{.}\\
\end{eqnarray*}
where the first isomorphism is due to the fact that $\Omega^{\bullet}_{{\rm{Sym}}(\mathfrak{a})}$ is acyclic with respect to the de Rham differential $d$ and the third isomorphism is induced by the de Rham differential.
\end{proof}
\noindent
\textbf{Remark.} The existence of a Hodge decomposition of $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})$ that is functorial in $\mathfrak{a}$ follows from Kassel's Corollary~\ref{kassel} and the acyclicity of $\Omega^{\bullet}_{{\rm{Sym}}(\mathfrak{a})}$ with respect to the de Rham differential, although this fact seems to have been left unnoticed in~\cite{K, L}. The new interesting fact discovered in~\cite{bfprw1} is the realization of the Hodge components $\mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})$ as homologies of the non-abelian derived functors of universal multilinear forms $\lambda^{(p)}$.
\subsubsection{Proof of Theorem~\ref{connesvsgksbi}} Theorem~\ref{connesvsgksbi} follows from the results of~\cite[Section~6.10]{GK} and the following proposition.
\begin{proposition} \la{compkassel}
Let $d^{(p)}\,:\,\mathcal{C}(\mathfrak{a};{\rm{Sym}}^{p}(\mathfrak{a})) \xrightarrow{} \mathcal{C}(\mathfrak{a};{\rm{Sym}}^{p-1}(\mathfrak{a}))[-1]$ be the de Rham differential. Let $I$ denote the map $d^{(p)}$ thought of as a map from $\mathcal{C}(\mathfrak{a};{\rm{Sym}}^{p}(\mathfrak{a}))$ to $\mathrm{Im}(d^{(p)})$. Then, for all $p \geq 1$, the homology long exact sequence arising from the short exact sequence of complexes
$$ \begin{diagram}[small] 0 & \rTo & \ker(d^{(p)}) & \rTo^B & \mathcal{C}(\mathfrak{a};{\rm{Sym}}^{p}(\mathfrak{a})) & \rTo^{I} & \ker(d^{(p-1)})[-1] & \rTo & 0 \end{diagram} $$
coincides with the exact sequence~\eqref{conneshodgep}. In particular, we have
$$
{\rm{HH}}^{(p)}_{\bullet}(\mathfrak{a})\,\cong\, \mathrm H_{\bullet}(\mathfrak{a};{\rm{Sym}}^p(\mathfrak{a}))\ ,
\ \forall\,p \ge 1 \ .
$$
\end{proposition}
\begin{proof}
The bicomplex $X^+(R)$ for $R=\boldsymbol{\Omega}(C)$ is formally dual to Tsygan's double complex (in columns of strictly negative degree) for the augmentation ideal $\bar{A}$ of a smooth commutative DGA $A\,\in\,\mathtt{DGCA}_{k/k}$ (with the latter given a bidegree shift of $[-1,1]$. By~\cite[1.4.5 and 2.2.16]{L}, the total complex of this bicomplex is canonically isomorphic to the total complex of Connes' reduced $(b,B)$-bicomplex $\mathcal B^{-}(A)_{\mathrm{red}}[-1,0]$ for the negative cyclic homology $A$. It is easy to see that this isomorphism is compatible with the corresponding $S$-$I$ short exact sequence of complexes for negative cyclic homology. Dually, if $C=\mathcal{C}(\mathfrak{a};k)$ (which is a symmetric coalgebra), $X^+(R)$ is canonically isomorphic to Connes' $(b,B)$-complex for $C$ with a shift of $[1,-1]$ (the isomorphism being compatible with the $I$-$S$ short exact sequence of total complexes. By the dual to the Hochschild-Kostant-Rosenberg theorem and the fact that the de Rham differential on $\Omega^{\bullet}_C$ corresponds to the $B$ differential under the dual HKR map, there is a dual HKR quasi-isomorphism of bicomplexes
$$ \mathcal B_{\mathrm{DR}}(C)[1,-1] \xrightarrow{} \ \mathcal B(C)_{\mathrm{red}}[1,-1]\,,$$
where
$$
\mathcal B_{\mathrm{DR}}(C)\,:=\,
\left[\ \oplus_p\, \Omega^p_C[-p] \,\xleftarrow{d}\, \oplus_p \,\Omega^{p+1}_C[-p]\, \xleftarrow{d} \,
\ldots \,\right]\,\text{.}
$$
Note that there is a natural Hodge decomposition $\, \mathcal B_{\mathrm{DR}}(C)\,\cong \,\bigoplus_p \mathcal B_{\mathrm{DR}}^{(p)}(C)\,$, where
$$
\mathcal B_{\mathrm{DR}}^{(p)}(C)\,:=\, \left[\,
\Omega^p_C[-p] \,\xleftarrow{d}\, \Omega^{p+1}_C[-p] \,\xleftarrow{d}\,
\ldots \,\right]\,\text{.} $$
Besides being compatible with the $I$-$S$ short exact sequence of total complexes, the (dual) HKR map is also compatible with Hodge decomposition. The exact sequence~\eqref{conneshodgep} is thus induced by the short exact sequence of complexes
$$
0 \to \Omega^p_C[-p] \,\xrightarrow{I}\, \mathrm{Tot}\, \mathcal B_{\mathrm{DR}}^{(p)}(C)
\,\xrightarrow{S}\, \mathrm{Tot} \, \mathcal B_{\mathrm{DR}}^{(p+1)}(C)[2] \to 0 \ \text{.}
$$
By Lemma~\ref{chevalley}, $\Omega^p_C[-p]\,\cong\,\mathcal{C}(\mathfrak{a};{\rm{Sym}}^p(\mathfrak{a}))$. Let $\varphi^{(p)}\,:\, \mathrm{Tot}\, \mathcal B_{\mathrm{DR}}^{(p)}(C) \xrightarrow{} \ker(d^{(p-1)})[-1]$ be the map given by $d^{(p)}$ on $\Omega^p_C[-p]$ and vanishing on all other direct summands. By acyclicity of $\Omega^{\bullet}_C$ with respect to $d$, $\varphi^{(p)}$ is a quasi-isomorphism. Thus, in the derived category $\mathcal D(k)$ of complexes of $k$-vector spaces, there is a commutative diagram where the rows are distinguished triangles and given vertical arrows are isomorphisms.
$$ \begin{diagram}
\mathrm{Tot} \, \mathcal B_{\mathrm{DR}}^{(p+1)}(C)[1] & \rTo^B & \Omega^p_C[-p] & \rTo^I & \mathrm{Tot}\, \mathcal B_{\mathrm{DR}}^{(p)}(C) & \rTo^S & \mathrm{Tot} \, \mathcal B_{\mathrm{DR}}^{(p+1)}(C)[2] \\
& & \dTo^{\mathrm{Id}} & & \dTo^{\varphi^{(p)}} & &\\
\ker(d^{(p)}) & \rTo^B & \Omega^p_C[-p] & \rTo^I & \ker(d^{(p-1)})[-1] & \rTo & \ker(d^{(p)})[1]\\
\end{diagram}$$
It follows that there is an arrow in $\mathcal D(k)$ completing the above diagram into an isomorphism of distinguished triangles. This proves the desired proposition.
\end{proof}
\section{Derived Poisson structures}
\la{Sect3}
In this section, we study derived Poisson structures on (the universal enveloping algebra of) a DG Lie algebra $\mathfrak{a}$. We focus on the case when the derived Poisson structure on $\mathcal U\mathfrak{a}$ arises from a cyclic pairing on a cocommutative DG coalgebra $C$ that is Koszul dual to $\mathfrak{a}$. Our main observation is that this derived Poisson structure is compatible with Lie Hodge decomposition of $\mathcal U\mathfrak{a}$.
\subsection{Derived Poisson algebras}
\la{Sect3.1}
We begin by reviewing the notion of a derived Poisson algebra introduced in ~\cite{BCER}. This notion is a higher homological extension of the notion of an ${\rm H}_0$-Poisson algebra proposed by Crawley-Boevey \cite{CB}.
\subsubsection{Definitions}
\la{Defs}
Let $ A$ be an (augmented) DG algebra. The space $\DER(A)$ of graded $k$-linear derivations of $A$
is naturally a DG Lie algebra with respect to the commutator bracket. Let $\DER(A)^\n $ denote the subcomplex of $\DER(A) $
comprising derivations with image in $\,k+[A,A] \subseteq A \,$. It is easy to see that $ \DER(A)^\n $ is a DG Lie ideal of $ \DER(A) $,
so that $\,\DER(A)_{\natural} := \DER(A)/\DER(A)^\n $ is a DG Lie algebra. The natural action of $ \DER(A) $ on $A$ induces a Lie algebra
action of $ \DER(A)_\n $ on the quotient space $ A_\n := A/(k+[A,A]) $. We write $\, \varrho: \DER(A)_\n \to \END(A_\n) \,$ for the
corresponding DG Lie algebra homomorphism.
Now, following \cite{BCER}, we define a {\it Poisson structure} on $A$ to be a DG Lie algebra structure on $\,A_\n \,$ such that the adjoint representation
$ \mbox{\rm ad}:\, A_\n \to \END(A_\n) $ factors through $ \varrho \,$: i.~e., there is a morphism of DG Lie algebras $\,\alpha :\, A_\n \xrightarrow{} \DER(A)_{\natural}\,$ such that $\,\mbox{\rm ad} = \varrho \circ \alpha \,$. It is easy to see that if $A$ is a commutative DG algebra, then a Poisson structure on $A$ is
the same thing as a (graded) Poisson bracket on $A$. On the other hand, if $A$ is an ordinary $k$-algebra (viewed as a DG algebra), then a Poisson structure on
$A$ is precisely an ${\rm H}_0$-Poisson structure in the sense of \cite{CB}.
Let $A$ and $B$ be two Poisson DG algebras, i.e. objects of $\mathtt{DGA}_{k/k}$ equipped with Poisson structures.
A {\it morphism} $\,f:\, A \xrightarrow{} B $ of Poisson algebras is then a morphism $ f: A \to B $ in $\mathtt{DGA}_{k/k} $ such that $ f_{\natural}:\, A_{\natural} \xrightarrow{} B_{\natural} $ is a morphism of DG Lie algebras. With this notion of morphisms, the Poisson DG algebras form a category which we denote $\mathtt{DGPA}_k $.
Note that $\mathtt{DGPA}_k $ comes with two natural functors: the forgetful functor $ U:\, \mathtt{DGPA}_k \to \mathtt{DGA}_{k/k} $ and the cyclic functor
$ (\,\mbox{--}\,)_\n : \mathtt{DGPA}_k \to \mathtt{DGLA}_k $. We say that a morphism $ f $ is a {\it weak equivalence} in $ \mathtt{DGPA}_k $
if $ Uf $ is a weak equivalence in $ \mathtt{DGA}_{k/k} $ and $ f_\n $ is a weak equivalence in $ \mathtt{DGLA}_k $; in other words, a weak equivalence in $\mathtt{DGPA}_k $
is a quasi-isomorphism of DG algebras, $ f: A \to B \,$, such that the induced map $ f_{\n}\,:\,A_\n \xrightarrow{} B_\n $ is a quasi-isomorphism of DG Lie algebras.
Although we do not know at the moment whether the category $ \mathtt{DGPA}_k $ carries a Quillen model structure (with weak equivalences specified above), it has a weaker property of being a homotopical category in the sense of Dwyer-Hirschhorn-Kan-Smith \cite{DHKS}. This still allows one to define a well-behaved homotopy category of Poisson algebras and consider derived functors on $ \mathtt{DGPA}_k $.
\subsubsection{Homotopical categories}
Recall ({\it cf.} \cite{DHKS}) that a {\it homotopical category} is a category $\mathscr{C}$ equipped with a class of morphisms $\mathscr{W} $ (called weak equivalences) that contains all identities
of $ \mathscr{C} $ and satisfies the following {\it 2-of-6 property}: for every composable triple of morphisms $ f,g,h $ in $\mathscr{C}$, if $ g f $ and $ h g $ are in $ \mathscr{W} $, then so are $ f, g, h $ and $h g f $. The 2-of-6 property formally implies, but is stronger than, the usual 2-of-3 property: for every composable pair of morphisms $ f, g $ in $\mathscr{C}$,
if any two of $f$, $g$ and $ gf $ are in $ \mathcal W $, so is the third.
The class of weak equivalences thus forms a subcategory which contains all objects and all isomorphisms of $ \mathscr{C} $.
The isomorphisms in any category satisfy the 2-of-6 property: indeed, if $ f,g,h $ is a composable triple such that $ g f $ and $ h g $ are isomorphisms, then $ g $ has the right inverse
$ f(gf)^{-1} $, which must also be a left inverse, since $ g $ is monic (because $ h g $ is an isomorphism); hence $ g $ and therefore also $f$, $h$ and $hgf$ are isomorphisms.
Thus, any category can be viewed as a homotopical category by taking the weak equivalences to be the isomorphisms\footnote{In \cite{DHKS}, these are called {\it minimal} homotopical categories.}.
Furthermore, any model category is a homotopical category. This follows from the important property
of model categories (see \cite[Proposition~5.1]{Q1}) that the class $ \mathscr{W} $ of weak equivalences is {\it saturated} in $ \mathscr{C} \,$: i.e, it comprises {\it all} the arrows of $ \mathscr{C}$ that become isomorphisms in the localized category
$ \mathscr{C}[ \mathscr{W}^{-1}] \,$. Since the isomorphisms
in $ \mathscr{C}[ \mathscr{W}^{-1}] \,$ satisfy the 2-of-6 property, it follows immediately
that the weak equivalences in a saturated category satisfy the 2-of-6 property.
The category $ \mathscr{C}[ \mathscr{W}^{-1}] \,$ is called the {\it homotopy category} of $ \mathscr{C} $ and usually denoted $ {\mathtt{Ho}}(\mathscr{C})$. It is a domain
(and target) of homotopical functors and other homotopical structures that constitute the homotopy theory associated to a model category $ \mathscr{C} $.
Now, a key observation of \cite{DHKS} is that a well-behaved homotopy category, including
a meaningful notion of derived functors, can be defined for any homotopical category in which the class of weak equivalences is saturated. Such categories are called in \cite{DHKS} the saturated homotopical categories.
After these preliminaries, we can state our proposition.
\begin{proposition}
\la{homcat}
$\,\mathtt{DGPA}_k$ is a saturated homotopical category.
\end{proposition}
\begin{proof}
It suffices to prove that the class of weak equivalences in $\,\mathtt{DGPA}_k$ is saturated. Let $ \gamma: \mathtt{DGPA}_k \to {\mathtt{Ho}}(\mathtt{DGPA}_k) $ denote the (formal) localization of $ \mathtt{DGPA}_k $ at the class of weak equivalences (similarly, abusing notation, we will write $ \gamma $ for the localizations of $ \mathtt{DGA}_{k/k} $ and $ \mathtt{DGLA}_{k} $.)
Since both the forgetful functor $U: \mathtt{DGPA}_k \to \mathtt{DGA}_{k/k} $ and the cyclic functor
$ (\,\mbox{--}\,)_\n: \mathtt{DGPA}_k \to \mathtt{DGLA}_k \to$ preserve weak equivalences, by the universal property of localization, they factor through $ \gamma $, inducing $ \gamma U:\,{\mathtt{Ho}}(\mathtt{DGPA}_k) \to {\mathtt{Ho}}(\mathtt{DGA}_{k/k}) $ and $\,\gamma (\,\mbox{--}\,)_\n : {\mathtt{Ho}}(\mathtt{DGPA}_k) \to {\mathtt{Ho}}(\mathtt{DGLA}_{k}) \,$. Now, if $ f $ is a morphism in $ \mathtt{DGPA}_k $
such that $ \gamma f $ is an isomorphism in ${\mathtt{Ho}}(\mathtt{DGPA}_k)$, then $ \gamma (U f) $ is an isomorphism in $ {\mathtt{Ho}}(\mathtt{DGA}_{k/k}) $ and
$ \gamma(f_\n) $ is an isomorphism in $ {\mathtt{Ho}}(\mathtt{DGLA}_{k}) $. Since both
$ \mathtt{DGA}_{k/k} $ and $ \mathtt{DGLA}_{k} $ are model categories,
the classes of their weak equivalences are saturated. It follows that
$ U f $ and $ f_\n $ are weak equivalences in $ \mathtt{DGA}_{k/k} $ and $ \mathtt{DGLA}_{k} $,
respectively. Then, by definition, $ f $ is a weak equivalence in $ \mathtt{DGPA}_k $, so
that the class of weak equivalences $ \mathtt{DGPA}_k $ is saturated.
\end{proof}
\subsubsection{Homotopy category of Poisson algebras}
If $ \mathscr{C} $ is a model category, there are two ways to define
a homotopy category of $ \mathscr{C} $: first, one can simply put
$\, {\mathtt{Ho}}(\mathscr{C}) := \mathscr{C}[\mathscr{W}^{-1}]\,$, or alternatively,
one can consider the full subcategory $ \mathscr{C}^{\rm c} $ of cofibrant/fibrant objects
in $ \mathscr{C} $ and then define $ {\mathtt{Ho}}(\mathscr{C}^{\rm c}) := \mathscr{C}^{\rm c}/\!\!\sim $
to be the quotient category of $ \mathscr{C}^{\rm c} $ modulo an
appropriate homotopy equivalence relation\footnote{The category
$ {\mathtt{Ho}}(\mathscr{C}^{\rm c}) $ is often called the {\it classical homotopy category}
of $ \mathscr{C} $.}. By (an abstract version of) Whitehead's Theorem, the
two definitions are equivalent, the equivalence
$ {\mathtt{Ho}}(\mathscr{C}^{\rm c}) \stackrel{\sim}{\to} {\mathtt{Ho}}(\mathscr{C}) $ being induced
by the natural functor $ \mathscr{C}^{\rm c} \,\hookrightarrow\, \mathscr{C} \to {\mathtt{Ho}}(\mathscr{C}) $.
Each definition has its advantages: the first one is more natural (it gives a characterization
of $ {\mathtt{Ho}}(\mathscr{C}) $ in terms of a universal property and shows that
$ {\mathtt{Ho}}(\mathscr{C}) $ depends only on the class of weak
equivalences in $ \mathscr{C} $)); the second is more concrete and accessible for computations
(it implies, in particular, that $ {\mathtt{Ho}}(\mathscr{C}) $ is a locally small
category if so is $\mathscr{C}$.)
For the homotopical category $\mathtt{DGPA}_k $, we
can also define a homotopy category in two ways. First, we can simply put
$$
{\mathtt{Ho}}(\mathtt{DGPA}_k) := \mathtt{DGPA}_k[\mathscr{W}^{-1}]\ ,
$$
where $ \mathscr{W} $ is the class of weak equivalences specified in Section~\ref{Defs}.
Proposition~\ref{homcat} then ensures that $ {\mathtt{Ho}}(\mathtt{DGPA}_k) $ has properties
similar to those of the homotopy category of a model category (see~\cite[Sect.~33.8]{DHKS}).
Alternatively, following \cite{BCER}, we can mimick the definition of the classical homotopy category of a model category and define\footnote{This category was denoted
$ {\mathtt{Ho}}^*(\mathtt{NCPoiss}_k) $ in \cite{BCER} and simply referred to as the homotopy
category of Poisson algebras.}
\begin{equation}
\la{defho}
{\mathtt{Ho}}(\mathtt{DGPA}^{\rm c}_k) := \mathtt{DGPA}^{\rm c}_k/\!\!\sim
\end{equation}
The objects in this quotient category are the homotopy classes of Poisson
algebras $A $ whose underlying DG algebras $ U(A) $ are cofibrant
as objects in $ \mathtt{DGA}_{k/k} $. The equivalence relation $\, \sim \,$
is based on the notion of $P$-homotopy (`polynomial homotopy') for Poisson
algebras introduced in \cite{BCER}. We recall that
two morphisms $\,f,g: A \xrightarrow{} B\,$ in $\mathtt{DGPA}_k $ are called {\it $P$-homotopic}
if there is a morphism $\, h: A \xrightarrow{} B \otimes \Omega \,$ such that $ h(0)=f $ and $ h(1)=g $, where $ \Omega = \Omega({\mathbb A}_k^1) $ is the de Rham algebra of the affine line and
$ B \otimes \Omega $ is given the structure of a Poisson DG algebra via the extension of
scalars from $ B $ ({\it cf.}~\cite{BCER}, Sect.~3.1). It is easy to check that the $P$-homotopy defines an equivalence relation on $\, \Hom_{\mathtt{DGPA}}(A,B)\,$ for any objects $A$ and $B$ in $ \mathtt{DGPA}_k $. Now, as in the case of model categories, we have
\begin{proposition}
\la{hoeq}
There is a natural functor $\, {\mathtt{Ho}}(\mathtt{DGPA}^{\rm c}_k) \to {\mathtt{Ho}}(\mathtt{DGPA}_k) \,$.
\end{proposition}
\begin{proof}
Recall that, for $ B\,\in\,\mathtt{DGPA}_k$, a Poisson DG algebra structure on
$ B \otimes \Omega $ is given by extension of scalars, using the natural isomorphism
$ (B \otimes \Omega)_\n \cong B_\n \otimes \Omega\, $ ({\it cf.}~\cite[Section 3.1.1]{BCER}).
Then, the inclusion $i\,:\, B \xrightarrow{} B \otimes \Omega$ is a weak equivalence of Poisson algebras: indeed, $\,Ui$ is a weak equivalence of DG algebras, since the de Rham algebra $\Omega$ is acyclic, and $i_\n$ is a weak equivalence of DG Lie algebras, since $i_\n$ can be identified with the inclusion $B_\n \hookrightarrow B_\n \otimes \Omega$. Thus, if $f,g\,:\, A \xrightarrow{} B$ are $P$-homotopic with a homotopy $h\,:\,A \xrightarrow{} B \otimes \Omega$, then $f=g=i^{-1} \circ h$ in ${\mathtt{Ho}}(\mathtt{DGPA}_k)$. This proves the desired proposition.
\end{proof}
\begin{definition}
By a {\it derived Poisson algebra} we mean a cofibrant associative DG algebra $A$ equipped with a
Poisson structure (in the sense of Defintion~\ref{Defs}), which is viewed up to weak equivalence,
i.e. as an object in $ {\mathtt{Ho}}(\mathtt{DGPA}_k) $.
\end{definition}
\vspace{1ex}
The above definition differs from that of \cite{BCER}, where
the derived Poisson algebras were simply defined to be the objects of
$ {\mathtt{Ho}}(\mathtt{DGPA}^{\rm c}_k) $. Proposition~\ref{hoeq} shows, however, that any
derived Poisson algebra in the sense of~\cite{BCER} gives naturally a derived Poisson
algebra in our current sense. The point is that all results of \cite{BCER} established
with the use of explicit $P$-homotopies can be strengthened and reproved in a more
natural way with a weaker notion of equivalence. For example, we have
\begin{proposition}
\la{lieonft}
The cyclic homology $ \overline{\mathrm{HC}}_{\bullet}(A) $ of any derived Poisson algebra $A$
carries a natural structure of a graded Lie algebra.
\end{proposition}
\begin{proof}
By definition, the functor $\,(\,\mbox{--}\,)_\n:\,
\mathtt{DGPA}^{\rm c}_k \,\hookrightarrow\, \mathtt{DGPA}_k \to \mathtt{DGLA}_k \,$ preserves weak equivalences,
and hence induces
\begin{equation}
\la{FTf}
(\,\mbox{--}\,)_\n: \, {\mathtt{Ho}}(\mathtt{DGPA}_k)^{\rm c} \to {\mathtt{Ho}}(\mathtt{DGLA}_k)\ ,
\end{equation}
where $ {\mathtt{Ho}}(\mathtt{DGPA}_k)^{\rm c} $ denotes the full subcategory of the homotopy
category $ {\mathtt{Ho}}(\mathtt{DGPA}_k) $ whose objects are cofibrant DG algebras.
Now, the image of $A$ under \eqref{FTf} is a DG Lie algebra $ A_\n $ whose underlying complex
computes the (reduced) cyclic homology of $A$. This is a consequence of \cite{BKR}, Theorem~3.1,
(which is essentially due to Feigin and Tsygan) that states that the functor
$ (\,\mbox{--}\,)_\n $ has a well-defined derived functor $ \boldsymbol{L}(\,\mbox{--}\,)_\n $ on the category
of DG algebras whose homology agrees with cyclic homology. On cofibrant DG algebras, the values of
$ (\,\mbox{--}\,)_\n $ and $ \boldsymbol{L}(\,\mbox{--}\,)_\n $ are naturally isomorphic, hence
$ \mathrm H_{\bullet}(A_\n) \cong \overline{\mathrm{HC}}_{\bullet}(A) $ for any cofibrant $A$. On the other hand,
since $A_\n$ is a DG Lie algebra, $\, \mathrm H_{\bullet}(A_\n) $ carries a graded Lie algebra structure. Identifying $ \mathrm H_{\bullet}(A_\n) $ with $ \overline{\mathrm{HC}}_{\bullet}(A) $ for a cofibrant $A$,
we get a graded Lie algebra structure on $ \overline{\mathrm{HC}}_{\bullet}(A) $ claimed by the proposition.
\end{proof}
Another important result of \cite{BCER} that holds for the derived Poisson algebras in
$ {\mathtt{Ho}}(\mathtt{DGPA}_k) $ and that motivates our study of these objects is the following
\begin{theorem}[{\it cf.} \cite{BCER}, Theorem~2]
\la{t3s2int}
If $A$ is a derived Poisson DG algebra, then, for any finite-dimensional vector space $ V $,
there is a unique graded Poisson bracket on the representation homology
$ \mathrm H_\bullet(A, V)^{\GL(V)} $ of $A$ in $V$, such that the derived character map
$\,
{\rm{Tr}}_V(A)_\bullet:\, \overline{\mathrm{HC}}_\bullet(A) \to \mathrm H_\bullet(A,V)^{\GL(V)} $
is a Lie algebra homomorphism.
\end{theorem}
We will not reprove Theorem~\ref{t3s2int} here; instead, in Section~\ref{s4}, we will give a generalization of this result to representation homology of Lie algebras.
\subsection{Cyclic pairings} \la{s2.1}
We now describe our basic construction of derived Poisson structures associated with cyclic coalgebras. Recall ({\it cf.}~\cite{GK}) that a graded associative $k$-algebra is called $n$-{\it cyclic} if it is equipped with a symmetric bilinear pairing $\langle \mbox{--},\mbox{--} \rangle\,:\, A \times A \xrightarrow{} k$ of degree $n$ such that
$$ \langle a, bc \rangle \,=\, \pm \langle ca, b \rangle \,,\,\,\,\,\, \forall\,\, a,b,c\,\in\,A\,$$
the signs being determined by the Koszul sign rule. Dually, a graded coalgebra $C$ is called $n$-{\it cyclic} if it is equipped with a symmetric bilinear pairing $ \langle \mbox{--},\mbox{--} \rangle\,:\, {C} \times {C} \xrightarrow{} k$ of degree $n$ such that
$$ \langle v', w\rangle v'' \,=\, \pm \langle v, w''\rangle w'\,,\,\,\,\,\,\, \forall\,\,v,w\,\in\,C,$$
where $v'$ and $v''$ are the components of the coproduct of $v$ written in the Sweedler notation. A DG coalgebra $C$ is $n$-cyclic if it is $n$-cyclic as a graded coalgebra and
$$ \langle du, v \rangle \pm \langle u, dv \rangle \,=\, 0\,,$$
for all homogeneous $u,v \,\in\,{C}$, i.e, if $\langle \mbox{--}, \mbox{--} \rangle \,:\, {C}[n] \otimes C[n] \xrightarrow{} k[n]$ is a map of complexes. {\it By convention, we say that $C\,\in\,\mathtt{DGC}_{k/k}$ is $n$-cyclic if $\bar{C}$ is $n$-cyclic as a non-counital DG coalgebra}.
Assume that $C\,\in\,\mathtt{DGC}_{k/k}$ is equipped with a cyclic pairing of degree $n$ and let $R\,:=\,\boldsymbol{\Omega}(C)$ denote the (associative) cobar construction of $C$. Recall that $R\,\cong\, T_k(\bar{C}[-1])$ as a graded $k$-algebra. For $v_1,\ldots,v_n\,\in\,\bar{C}[-1]$, let $(v_1, \ldots ,v_n)$ denote the element $v_1 \otimes \ldots \otimes v_n$ of $R$. By~\cite[Theorem~15]{BCER}, the cyclic pairing on $C$ of degree $n$ induces a double Poisson bracket of degree $n+2$ (in the sense of~\cite{VdB})
$$\{\!\{ \mbox{--}, \mbox{--}\}\!\}\,:\, \bar{R} \otimes \bar{R} \xrightarrow{} {R} \otimes {R}$$
given by the formula
\begin{align} \la{dpbr}
\begin{aligned}
&\{\!\{(v_1, \ldots, v_n), (w_1, \ldots, w_m)\}\!\}\,=\, \\
&\sum_{\stackrel{i=1,\ldots,n}{j = 1,\ldots, m}} \pm \langle v_i, w_j \rangle (w_1 ,\ldots, w_{j-1}, v_{i+1} ,\ldots, v_n)\otimes (v_1, \ldots, v_{i-1}, w_{j+1} ,\ldots, w_m) \,\text{.} \end{aligned} \end{align}
The above double bracket can be extended to $R \otimes R$ by setting $\{\!\{r,1\}\!\}\,=\, \{\!\{1,r\}\!\}=0$. Let $\{ \mbox{--}, \mbox{--}\}$ be the bracket associated to~\eqref{dpbr}:
\begin{equation} \la{bronr} \{ \mbox{--}, \mbox{--}\} \,:=\, \mu\,\circ\, \{\!\{ \mbox{--},\mbox{--}\}\!\}\,:\, {R} \otimes {R} \xrightarrow{} {R}\,,\end{equation}
where $\mu$ is the multiplication map on ${R}$. Let $\n\,:\, {R} \xrightarrow{} R_\n$ be the canonical projection and let $\{ \mbox{--}, \mbox{--}\}\,:\, \n \circ \{ \mbox{--}, \mbox{--}\}\,:\, {R} \otimes {R} \xrightarrow{} R_\n$. We recall that the bimodule ${R} \otimes {R}$ (with outer $R$-bimodule structure) has a double bracket (in the sense of~\cite[Defn. 3.5]{CEEY}) given by the formula
\begin{align*}
&\{\!\{ \mbox{--}, \mbox{--}\}\!\}\,\,:\,{R} \times ({R} \otimes {R}) \xrightarrow{} {R} \otimes ({R} \otimes {R}) \oplus ({R} \otimes {R}) \otimes {R}\,, \\
& \{\!\{r, p \otimes q\}\!\} \,:=\, \{\!\{r,p\}\!\} \otimes q \oplus (-1)^{|p|(|r|+n)} p \otimes \{\!\{ r,q\}\!\} \,\text{.}
\end{align*}
This double bracket restricts to a double bracket on the sub-bimodule $\Omega^1R$ of $R \otimes R$ (~\cite[Corollary~5.2]{CEEY}). Let $\{\mbox{--}, \mbox{--}\}\,:\, R \otimes \Omega^1R \xrightarrow{} \Omega^1$ be the map $\mu \circ \{\!\{ \mbox{--},\mbox{--}\}\!\}$, where $\mu$ is the bimodule action map and let $\{\mbox{--}, \mbox{--}\}\, :\, R \otimes \Omega^1R \xrightarrow{} \Omega^1R_{\n}$ denote the map $\n \circ \{ \mbox{--},\mbox{--}\}$.
The bracket $\{ \mbox{--},\mbox{--}\}\,:\, {R} \otimes {R} \xrightarrow{} R_\n$ descends to a DG $(n+2)$-Poisson structure on $R$. In particular, it descends to a (DG) Lie bracket $\{\mbox{--},\mbox{--}\}_{\n}$ on $R_\n$ of degree $n+2$. The restriction of the bracket~\eqref{bronr} to $\bar{R}$ induces a degree $n+2$ DG Lie module structure over $R_\n$ on $\bar{R}$ and the bracket $\{\mbox{--},\mbox{--}\}\,:\,R \otimes \Omega^1R \xrightarrow{} \Omega^1R_\n$ induces a degree $n+2$ DG Lie module structure over $R_\n$ on $\Omega^1R_\n$ (see~\cite[Proposition~3.11]{CEEY}). On homologies, we have (see~
\cite{CEEY}, Theorem~1.1 and Theorem~1.2)
\begin{theorem} \la{liestronhom}
Let $A\,\in\,\mathtt{DGA}_{k/k}$ be an augmented associative algebra Koszul dual to $C\,\in\,\mathtt{DGC}_{k/k}$. Assume that $C$ is $n$-cyclic. Then, \\
$(i)$ $\overline{\mathrm{HC}}_{\bullet}(A)$ has the structure of a graded Lie algebra (with Lie bracket of degree $n+2$).\\
$(ii)$ $\overline{{\rm{HH}}}_{\bullet}(A)$ has a graded Lie module structure over $\overline{\mathrm{HC}}_{\bullet}(A)$ of degree $n+2$.\\
$(iii)$ The maps $S,B$ and $I$ in the Connes periodicity sequence~\eqref{connessbi} are homomorphisms of degree $n+2$ graded Lie modules over $\overline{\mathrm{HC}}_{\bullet}(A)$.
\end{theorem}
The Lie bracket of degree $n+2$ on $\overline{\mathrm{HC}}_{\bullet}(A)$ that is induced by a $(n+2)$-Poisson structure on $R_{\n}$ as above is an example of a derived $(n+2)$-Poisson structure on $A$.\\
\noindent
\textbf{Convention.} Since we work with algebras that are Koszul dual to $n$-cyclic coalgebras, all Lie algebras that we work with have Lie bracket of degree $n+2$. Similarly, all Lie modules are degree $n+2$ Lie modules. We therefore, drop the prefix ``degree $n+2$" in the sections that follow. Following this convention, we shall refer to (derived) $(n+2)$-Poisson structures as (derived) Poisson structures.
\subsection{The Hodge filtration} \la{s2.2}
Consider the filtration on (the graded vector space) $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})$ given by
\begin{equation}
\la{eqfilthc}
F_p\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})\,:=\, \bigoplus_{r \leq p+2} \mathrm{HC}^{(r)}_{\bullet}(\mathfrak{a}) \,\text{.}\end{equation}
Let $\mathfrak{a}\,\in\,\mathtt{DGLA}_k$ be Koszul dual to $C\,\in\,\mathtt{DGCC}_{k/k}$, where $C$ is $n$-cyclic. The following theorem is one of our main results.
\begin{theorem} \la{thodgefiltration}
The derived Poisson bracket $\{ \mbox{--},\mbox{--}\}$ on $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})$ respects the filtration~\eqref{eqfilthc}.
Moreover,\\
\noindent
$(i)$ $\{ \mathrm{HC}^{(1)}_{\bullet}(\mathfrak{a}), \mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})\} \,\subseteq \, \mathrm{HC}^{(p-1)}_{\bullet}(\mathfrak{a})$ for all $p \geq 1$. In particular, $\{\mathrm{HC}^{(1)}_{\bullet}(\mathfrak{a}), \mathrm{HC}^{(1)}_{\bullet}(\mathfrak{a})\}=0$.\\
$(ii)$ $\{\mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a}), \mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})\} \subseteq \, \mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})$ for all $p \geq 1$.\\
In particular, $\mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a})$ is a Lie subalgebra of $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})$ and $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})$ is a Hodge weight graded Lie module over $\mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a})$.
\end{theorem}
Recall that $\mathcal L\,:=\, \boldsymbol{\Omega}_{\mathtt{Comm}}(C)$ gives a cofibrant resolution $\mathcal L \stackrel{\sim}{\xrightarrow{}} \mathfrak{a}$ of $\mathfrak{a}$ in $\mathtt{DGLA}_k$ and the (associative) cobar construction $R\,:=\,\boldsymbol{\Omega}(C)$ gives a cofibrant resolution $R \stackrel{\sim}{\xrightarrow{}} \mathcal U\mathfrak{a}$ of $\mathcal U\mathfrak{a}$ (with $R\,\cong\, \mathcal U\mathcal L$). By Theorem~\ref{liestronhom}, the $n$-cyclic pairing on $C$ induces a derived Poisson structure on $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})$. The proof of Theorem~\ref{thodgefiltration} is based on the following proposition. For notational brevity, let $V\,:=\, \bar{C}[-1]$.
\begin{proposition} \la{hodgebracketr}
Let $R^{(p)}\,:=\, {\rm{Sym}}^p(\mathcal L)$ as in Section~\ref{s1.3}. Let $\{ \mbox{--}, \mbox{--}\}$ be defined as in~\eqref{bronr}. \\
$(i)$ $\{ R^{(1)}, R^{(p)}\} \,\subseteq \, R^{(p-1)}$ for any $p \geq 1$. In particular, $\{R^{(1)}, R^{(1)}\}=0$.\\
$(ii)$ $ \{R^{(2)}, R^{(p)}\}\,\subseteq \, R^{(p)}$ for any $p \geq 1$.\\
$(iii)$ For $p,q>2$, $\{R^{(q)}, R^{(p)}\} \,\subseteq\, \bigoplus_{r \leq p+q-2} R^{(r)}$.
\end{proposition}
\begin{proof}
Let $\alpha\,\in\,R^{(q)}$ be homogeneous. By~\cite[Section 2.4]{VdB}, the map $\{\alpha , \mbox{--}\}\,:\, \bar{R} \xrightarrow{} R$ is a derivation of degree $|\alpha|+n+2$. By Lemma~\ref{derivations}, it suffices to show that $\{\alpha, V\}\,\subseteq \, {\rm{Sym}}^{q-1}(\mathcal L)$. Note that if $w\,\in\, V$, then by~\eqref{dpbr},
\begin{equation*} \{(v_1, \ldots, v_n), w\} \,=\, \sum_{i=1}^n \pm \langle v_i, w\rangle (v_{i+1} , \ldots, v_n, v_1, \ldots ,v_{i-1}) \,\text{.}\end{equation*}
It is easy to see that this coincides with the expression $\langle \bar{\partial}(v_1, \ldots, v_n), w \rangle$, where
$\langle \mbox{--}, \mbox{--} \rangle\,:\, \bar{R} \otimes V \xrightarrow{} R$ denotes the composite map
$$ \begin{diagram} \bar{R} \otimes V & \rTo^{\bar{\partial} \otimes {\rm{Id}}_V} & (R \otimes V) \otimes V & \rTo & R \otimes (V \otimes V) & \rTo^{{\rm{Id}}_V \otimes \langle \mbox{--}, \mbox{--} \rangle} & R \end{diagram} \,\text{.}$$
The unlabelled arrow in the above diagram swaps factors. Now, if $\alpha\,\in, R^{(q)}$, $\bar{\partial}(\alpha)\,\in\, {\rm{Sym}}^{q-1}(\mathcal L) \otimes V$ by Lemma~\ref{cycderhamhodge}. Thus, $\{\alpha, V\}\,\subseteq \, {\rm{Sym}}^{q-1}(\mathcal L)$ as desired.
\end{proof}
\begin{corollary} \la{hodgencpoiss} Let $\{\mbox{--},\mbox{--}\}_{\n}$ be the DG Poisson bracket on $R_{\n}$ as in Section~\ref{s2.1} above. Then,\\
$(i)$ $\{ R^{(1)}_\n, R^{(p)}_\n\}_{\n} \,\subseteq \, R^{(p-1)}_{\n}$ for any $p \geq 1$. In particular, $\{R^{(1)}_\n, R^{(1)}_\n\}_\n=0$.\\
$(ii)$ $ \{R^{(2)}_\n, R^{(p)}_{\n}\}_\n\,\subseteq \, R^{(p)}_\n$ for any $p \geq 1$.\\
$(iii)$ For $p,q>2$, $\{R^{(q)}_\n, R^{(p)}_\n\}_\n \,\subseteq\, \bigoplus_{r \leq p+q-2} R^{(r)}_\n$.
\end{corollary}
Corollary~\ref{hodgencpoiss} follows immediately from Proposition~\ref{hodgebracketr}. Theorem~\ref{thodgefiltration} follows immediately
from \eqref{hodgeds} and Corollary~\ref{hodgencpoiss}.
\subsection{Hodge filtration on Poisson modules} Let $\mathfrak{a},C, R, V$ and $\mathcal L$ be as in Section~\ref{s2.2}, with $C$ having a cyclic pairing of degree $n$. Recall that by Theorem~\ref{liestronhom}, $\overline{{\rm{HH}}}_{\bullet}(\mathcal U\mathfrak{a})$ is a Lie module over $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})$. In addition to~\eqref{eqfilthc}, we define the filtration on $\overline{{\rm{HH}}}_{\bullet}(\mathcal U\mathfrak{a})$
\begin{equation} \la{eqfilthh} F_p\overline{{\rm{HH}}}_{\bullet}(\mathcal U\mathfrak{a})\,:=\, \bigoplus_{r \leq p+2} {\rm{HH}}^{(r)}_{\bullet}(\mathfrak{a}) \,\text{.} \end{equation}
and let $\{\mbox{--}, \mbox{--}\}\,:\, \overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a}) \times \overline{{\rm{HH}}}_{\bullet}(\mathcal U\mathfrak{a}) \xrightarrow{} \overline{{\rm{HH}}}_{\bullet}(\mathcal U\mathfrak{a})$ denote the action map.
\begin{theorem} \la{pstrhoch1}
$\overline{{\rm{HH}}}_{\bullet}(\mathcal U\mathfrak{a})$ is a filtered Lie module, with filtration given by~\eqref{eqfilthh}. Moreover, for all $p \geq 0$,\\
$(i)$ $ \{\mathrm{HC}^{(1)}_{\bullet}(\mathfrak{a}) , {\rm{HH}}^{(p)}_{\bullet}(\mathfrak{a})\} \subseteq {\rm{HH}}^{(p-1)}_{\bullet}(\mathfrak{a})$. \\
$(ii)$ $\{\mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a}), {\rm{HH}}^{(p)}_{\bullet}(\mathfrak{a})\} \subseteq {\rm{HH}}^{(p)}_{\bullet}(\mathfrak{a})$.\\
In particular, $\overline{{\rm{HH}}}_{\bullet}(\mathcal U\mathfrak{a}) $ is a Hodge weight graded Lie module over $\mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a})$.
\end{theorem}
The restrictions of the maps $S,B$ and $I$ of the Connes periodicity sequence to the Hodge summand~\eqref{conneshodgep} give maps of graded vector spaces
\begin{align}
\la{sp1} & S\,:\, \mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a}) \xrightarrow{} \mathrm{HC}^{(p+1)}_{\bullet-2}(\mathfrak{a})\\
\la{bp1} & B\,:\, \mathrm{HC}^{(p+1)}_{\bullet}(\mathfrak{a}) \xrightarrow{} {\rm{HH}}^{(p)}_{\bullet+1}(\mathfrak{a})\\
\la{ip1} & I\,:\, {\rm{HH}}^{(p)}_{\bullet}(\mathfrak{a}) \xrightarrow{} \mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a}) \,\text{.}
\end{align}
If we equip $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})$ with filtration~\eqref{eqfilthc} and $\overline{{\rm{HH}}}_{\bullet}(\mathcal U\mathfrak{a})$ with filtration~\eqref{eqfilthh}, the maps $S,B$ and $I$ become filtered maps; more precisely, we have:
\begin{align}
\la{sp2} & S\,:\, F_{\bullet}\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a}) \xrightarrow{} F_{\bullet+1}\overline{\mathrm{HC}}_{\bullet-2}(\mathcal U\mathfrak{a})\\
\la{bp2} & B \,:\, F_{\bullet+1}\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a}) \xrightarrow{} F_{\bullet}\overline{{\rm{HH}}}_{\bullet+1}(\mathcal U\mathfrak{a})\\
\la{ip2} & I\,:\, F_{\bullet}\overline{{\rm{HH}}}_{\bullet}(\mathcal U\mathfrak{a}) \xrightarrow{} F_{\bullet}\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a})\,\text{.}
\end{align}
The first statement in the following theorem is a refinement of~\cite[Theorem~1.2]{CEEY}.
\begin{theorem} \la{pstrhoch2}
With definitions~\eqref{sp2}-\eqref{ip2}, the maps $S,B$ and $I$ become {\rm filtered} Lie module maps. Moreover, in the Hodge summand~\eqref{conneshodgep}, these maps are module maps over the Lie algebra $\mathrm{HC}^{(2)}_{\bullet}(\mathcal U\mathfrak{a})$.
\end{theorem}
\begin{corollary} \la{pstrGK}
The space $\mathrm{HA}_{\bullet}(\mathtt{Lie}, \mathfrak{a})$ has the structure of a graded Lie algebra, and $\mathrm{HB}_{\bullet}(\mathtt{Lie},\mathfrak{a})$ and $\mathrm{HC}_{\bullet}(\mathtt{Lie},\mathfrak{a})$ are graded Lie modules over $\mathrm{HA}_{\bullet}(\mathtt{Lie},\mathfrak{a})$. Further, the maps $S,B$ and $I$ in the Connes periodicity sequence for (operadic) Lie cyclic homology (see Theorem~\ref{connesvsgksbi}) are maps of Lie modules.
\end{corollary}
\begin{proof}
Indeed, by definition, $\mathrm{HA}_{\bullet}(\mathtt{Lie},\mathfrak{a})\,=\, \mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a})$. The remaining statements follow from Theorem~\ref{connesvsgksbi} and by putting $p=1$ in Theorem~\ref{thodgefiltration}, Theorem~\ref{pstrhoch1} and Theorem~\ref{pstrhoch2}.
\end{proof}
\subsubsection{Proof of Theorem~\ref{pstrhoch1}}
Recall from Section~\ref{s1.3} that $\Omega^1R_\n\,\cong\, R \otimes V$ and that there is a direct sum decomposition (of complexes) $\Omega^1R_\n \,\cong\, \bigoplus_{p \geq 0} \theta^{(p)}(\mathcal L)$. Further recall that the isomorphism $\Omega^1R_\n\,\cong\, R \otimes V$ identifies $\theta^{(p)}(\mathcal L)$ with $R^{(p)} \otimes V$, where $R^{(p)}\,:=\, {\rm{Sym}}^p(\mathcal L)$. Recall from Section~\ref{s2.1} that there is a bracket $\{\mbox{--},\mbox{--}\} \,:\, \bar{R} \times \Omega^1R_{\n} \xrightarrow{} \Omega^1R_{\n}$ inducing the structure of a DG Lie module over $R_\n$ on $\Omega^1R_{\n}$.
\begin{proposition} \la{hodgencpoissonforms} For any $p \geq 0$, the following inclusions hold:\\
$(i)$ $\{R^{(1)}, R^{(p)} \otimes V\} \subseteq R^{(p-1)} \otimes V$. In particular, $\{R^{(1)}, V\} =0$.\\
$(ii)$ $\{R^{(2)}, R^{(p)} \otimes V\} \subseteq R^{(p)} \otimes V$. Thus, $\Omega^1R_\n$ is a Hodge weight graded DG Lie module over $R^{(2)}_{\n}$.\\
$(iii)$ For any $m > 2$, $\{R^{(m)}, R^{(p)} \otimes V\} \subseteq \bigoplus_{ r \leq p+m-2} R^{(r)} \otimes V$.
\end{proposition}
\begin{proof}
Let $r\,\in\,R^{(m)}$, $q\,\in\, R^{(p)}$ and $v \,\in \,V$ be homogeneous. By~\cite[Lemma 5.5]{CEEY},
\begin{equation} \la{ebforms} \{r, q \otimes v\}_{\Omega^1R_\n} \,=\, \n \circ [ \{r,q\}\cdot v \otimes 1 - \{r,q\} \otimes v + (-1)^{(|r|+n)|q|} (q\cdot \{r,v\} \otimes 1 -q \otimes \{r,v\})]\,, \end{equation}
where $\Omega^1R$ on the right hand side is realized as a sub-bimodule of $R \otimes R$ equipped with the outer bimodule structure and where $\n\,:\,\Omega^1R \xrightarrow{} \Omega^1R_{\n}$ is the canonical projection. The first summand $\n \circ ( \{r,q\}\cdot v \otimes 1 - \{r,q\} \otimes v)$ of~\eqref{ebforms} is equal to $\{r,q\} \otimes v$ (after the identification of $\Omega^1R_\n$ with $R \otimes V$. The remaining (i.e, second) summand of~\eqref{ebforms} may be written as
$\n \circ [ q \cdot d\{r,v\}]$, where $d\,:\,R \xrightarrow{} \Omega^1R$ is the universal derivation.
If $m=1$, Proposition~\ref{hodgebracketr} implies that $\{r,q\}\,\in\, R^{(p-1)}$ and that $\{r,v\}\,\in\, k$. Thus, in this case, the second sumand vanishes while the first summand is in $R^{(p-1)} \otimes V$. This proves (i).
If $m=2$, Proposition~\ref{hodgebracketr} implies that $\{r,q\}\,\in\, R^{(p)}$ and that $\{r,v\}\,\in\, \mathcal L$. Thus, the first statement of (ii) will follow once we show that for any $\alpha\,\in\,\mathcal L$, $\n \circ [ q \cdot d\alpha] \,\in\, R^{(p)} \otimes V$. Note that if $\n \circ [ q \cdot d\alpha] \,\in\, R^{(p)} \otimes V$ for all $q\,\in\,R^{(p)}$ for a given homogeneous $\alpha \in \mathcal L$, then for any $v \,\in\,V$ homogeneous,
\begin{align*}
\n \circ ( q \cdot d[v,\alpha])\,&=\, \n \circ (q \cdot d[v\cdot \alpha -(-1)^{|v||\alpha|} \alpha \cdot v])
\,=\, \n \circ (q \cdot [dv \cdot \alpha +v \cdot d\alpha -(-1)^{|v||\alpha|} (d\alpha \cdot v+ \alpha \cdot dv)])\\
\,&=\, \n \circ ((-1)^{|\alpha|(|q|+|v|)} [\alpha \cdot q \cdot dv - (-1)^{|q||\alpha|} q \cdot \alpha \cdot dv] + [q \cdot v -(-1)^{|q||v|} v \cdot q] \cdot d\alpha) \\
\,&=\, \n \circ ((-1)^{|\alpha|(|q|+|v|)} [\alpha, q] \cdot dv + [q,v] \cdot d\alpha)\,\text{.}
\end{align*}
Since $[\alpha,q]$ and $[q,v]$ are in $R^{(p)}$, we have shown that $\n \circ ( q \cdot d[v,\alpha])$ is in $R^{(p)} \otimes V$ for all $q$. Thus, by induction on $n$, we show that $\n \circ ( q \cdot d[v,\alpha_n])$ is in $R^{(p)} \otimes V$ for $\alpha_n\,=\, [v_1,[v_2, \ldots [v_{n-1}, v_n] \ldots ]]$ for homogeneous $v_1,\ldots,v_n$ in $V$. This proves that $\n \circ [ q \cdot d\alpha] \,\in\, R^{(p)} \otimes V$ for all $q\,\in\,R^{(p)}$ for all $\alpha \in \mathcal L$ as desired. The second statement in (ii) follows from the first and the fact that the bracket on $\Omega^1R_\n$ descends to an DG Lie module structure on $\Omega^1R_\n$ over $R_\n$ (of which $R^{(2)}_{\n}$ is a DG Lie subalgebra).
Finally, for $m>2$, the first summand of~\eqref{ebforms}, which can be identified with $\{r,q\} \otimes v$, lies in $\bigoplus_{r \leq p+m-2} R^{(r)} \otimes V$, by Proposition~\ref{hodgebracketr}. Also note that $\{r,v\}\,\in\, R^{(m-1)}$, by Proposition~\ref{hodgebracketr}. Thus, the second summand $\n \circ [q \cdot d\{r,v\}]$ is of the form $\n \circ [q \cdot d\beta]$, where $\beta\,\in\,R^{(m-1)}$. This last expression is a $k$-linear combination of expressions of the form
$\n \circ [q \cdot \gamma \cdot d\alpha \cdot \phi] \,=\, \pm \n \circ [\phi \cdot q \cdot \gamma \cdot d\alpha]$, where $\alpha\,\in\,\mathcal L$, and $\gamma\,\in\,R^{(i)}$ and $\phi\,\in\,R^{(j)}$ with $i+j=m-2$. Now, for any such expression, $\phi \cdot q \cdot \gamma\,\in\, \bigoplus_{ r \leq p+m-2} R^{(r)}$.
That $ \n \circ [\phi \cdot q \cdot \gamma \cdot d\alpha]$ lies in $ \bigoplus_{ r \leq p+m-2} R^{(r)} \otimes V$ now follows from the computation we made while proving (ii).
\end{proof}
It follows from Proposition~\ref{hodgebracketr}, Proposition~\ref{hodgencpoissonforms} and~\cite[Theorem~5.3]{CEEY} that the statement of Proposition~\ref{hodgencpoissonforms} holds word for word with $R^{(p)} \otimes V$ replaced by $\mathrm{Tot}\,X_2^{+,(p)}$. Theorem~\ref{pstrhoch1} then follows on homologies.
\subsubsection{Proof of Theorem~\ref{pstrhoch2}}
Let $F_pX^+(R)$ be given by $\oplus_{ r \leq p+2} X^{+, (r)}(\mathfrak{a})$. By Proposition~\ref{hodgebracketr}, Proposition~\ref{hodgencpoissonforms} and~\cite[Theorem~5.3]{CEEY}, there is an exact sequence of filtered DG Lie modules over $R_\n$
$$\begin{diagram} 0 & \rTo & F_{\bullet}\mathrm{Tot}\,X_2^{+}(R) & \rTo^I & F_{\bullet}\mathrm{Tot}\,X^{+}(R)& \rTo^S & F_{\bullet+1}\mathrm{Tot}\,X^{+}(R)[2] & \rTo & 0 \end{diagram}\,\text{.}$$
This implies (i) on homologies. Again by Proposition~\ref{hodgebracketr}, Proposition~\ref{hodgencpoissonforms} and~\cite[Theorem~5.3]{CEEY}, the short exact sequence of complexes~\eqref{xcomplexhptot} is an exact sequence of DG Lie modules over $R^{(2)}_n$. This gives (ii).
\subsection{Examples}
\la{Sect3.5}
We illustrate the results of this section on several examples.
\subsubsection{Abelian Lie algebras} \la{s2.2.2} Let $\mathfrak{a}=V$ be an abelian Lie algebra in homological degree $0$. Then, $A\,:=\,\mathcal U\mathfrak{a}\,=\, {\rm{Sym}}(V)$. The Koszul dual cocommutative coalgebra is $C\,:=\,{\rm{Sym}}^c(V[1])$. By~\cite[Proposition~9.4]{CEEY}, $C$ has (upto scalars) a unique cyclic pairing $\langle \mbox{--},\mbox{--} \rangle$ that is of degree $-n$, where $n\,:=\,\dim_kV$. For a basis $\{v_1,\ldots,v_n\}$ of $V$, this pairing is uniquely determined by $\langle 1,v_1 \wedge \ldots \wedge v_n \rangle$.
In this case, there is a Hochschild-Kostant-Rosenberg (HKR) isomorphism $\overline{\mathrm{HC}}_{n}(A)\,\cong\, \Omega^n(A)/d\Omega^{n-1}(A)$. Further, there is an isomorphism of $A$-modules $\Psi\,:\, \Theta_p(A) \cong \Omega^{n-p}(A)$ given by $\Theta_p(\xi) \,=\, \iota_{\xi} \omega$, where $\omega$ is a fixed constant nonzero $n$-form and where $\iota_{\xi}$ is contraction by $\xi$. In this case, by~\cite[Corollary 9.6]{CEEY}, the derived Poisson structure (of degree $2-n$) on $\overline{\mathrm{HC}}_{\bullet}(A)$ has a very explicit description after identifying reduced cyclic homology with forms via the HKR as above:
\begin{equation} \la{bracketofdiffforms} \{\alpha, \beta\} \,=\, (-1)^{(n-|\alpha|-1)(n-|\beta|)} \iota_{\eta}d\alpha\,,\,\,\,\, \text{ where } \eta\,:=\, \Psi^{-1}(d\beta)\,\text{.}\end{equation}
Now, by~\cite[Section 2]{bfprw2}, $\mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a})$ is identified (via the HKR map) with $\Omega^{n,(p)}(A)/d\Omega^{n-1,(p+1)}(A)$, where $\Omega^{n,(p)}(A)$ is the space of $n$-forms on $A$ whose polynomial coefficients are homogeneous of weight $p$, i,e, ${\rm{Sym}}^p(V) \otimes \wedge^n(V)$. An easy computation using~\eqref{bracketofdiffforms} shows that $\{\mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a}), \mathrm{HC}^{(q)}_{\bullet}(\mathfrak{a})\}$ is actually contained in $\mathrm{HC}^{(p+q-2)}_{\bullet}(\mathfrak{a})$. Thus, in this case the derived Poisson Lie algebra from Theorem~\ref{thodgefiltration} becomes a graded (not just filtered) Lie algebra with respect to the Hodge decomposition.
\subsubsection{Necklace Lie algebras} Our next example is when $\mathfrak{a}\,=\,LV$, the free Lie algebra generated by an even dimensional vector space (concentrated in homological degre $0$) equipped with a symplectic form $\langle \mbox{--}, \mbox{--} \rangle\,:\, V \times V \xrightarrow{} k$. Taking $\bar{C}:= V[1]$ with $0$ coproduct, we see that $C$ is Koszul dual to $\mathfrak{a}$. Further, the symplectic form on $V$ can be viewed as a cyclic pairing of degree $-2$ on $\bar{C}$. The resulting derived Poisson bracket $\{\mbox{--},\mbox{--}\}$ equips $R_\n\,:=\, T_kV_\n$ with the structure of a Lie algebra. The Lie algebra $(R_{\n}, \{\mbox{--},\mbox{--}\})$ is isomorphic to the necklace Lie algebra (see~\cite{BL,G}) associated with a one vertex quiver having $\frac{1}{2}\dim_k V$ loops.
Theorem~\ref{thodgefiltration} implies that the necklace Lie bracket restricts to give a Lie bracket on the direct summand $\mathrm{HC}^{(2)}_0(\mathfrak{a})\,=\,{\rm{Sym}}^{(2)}(LV)_{\n}$ of $T_kV$. Further, the filtration defined in Theorem~\ref{thodgefiltration} makes the necklace Lie algebra a filtered Lie algebra. This example also shows that the filtered structure in Theorem~\ref{thodgefiltration} is the best that we can get in general: for example, in the necklace Lie algebra generated by a two dimensional symplectic vector space $V$ with $v,w \,\in\,V$ such that $\langle v,w\rangle =1$, we have
$$\{ [v^3]_{\n}, [w^3]_{\n}\} \,=\, 9\,[v^2w^2]_{\n}\,\text{.} $$
It is not difficult to verify that while the right hand side above is contained in $\oplus_{r \leq 4} {\rm{Sym}}^r(LV)_\n$, it is not contained in ${\rm{Sym}}^4(LV)_\n$.
Further, let $V_n\,=\,k^{2n}$ with the standard symplectic form. The symplectic form preserving inclusion $V_n \hookrightarrow V_{n+1}$ induces a homomorphism $L_n \xrightarrow{} L_{n+1}$ of necklace Lie algebras, where $L_n\,:=\, ([TV_n]_\n, \{\mbox{--},\mbox{--}\})$. It is easily seen that these maps are compatible with the filtration from Theorem~\ref{thodgefiltration}. Thus, $L_{\infty}\,:=\, \varinjlim_n L_n$ admits a filtration as in Theorem~\ref{thodgefiltration}. In particular, $L^{(2)}_{\infty}\,:=\, \varinjlim_n {\rm{Sym}}^2(LV_n)_\n$ is a Lie subalgebra of $L_{\infty}$.
Let $\mathcal P$ be a cyclic Koszul operad and let $W$ be a symplectic vector space with symplectic form $\omega$. Let $F_{\mathcal P}W$ be the free $\mathcal P$-algebra generated by $W$. Let $\mathrm{Der}(F_{\mathcal P}W, \omega)$ be the Lie algebra of $\mathcal P$-algebra derivations of $F_{\mathcal P}W$ that preserve $\omega$ (see~\cite[Section 6]{G} for example). We refer to $\mathrm{Der}(F_{\mathcal P}W, \omega)$ as the Lie algebra associated with (the symplectic vector space) $W$ for the operad $\mathcal P$. It is not difficult to see that ${\rm{Sym}}^2(LV_n)_\n$ is the Lie algebra associated with $V_n$ for the Lie operad while $L_n$ is the Lie algebra associated with $V_n$ for the associative operad. The Lie homology of $L_{\infty}$ is related to the cohomology of coarse moduli spaces of algebraic curves of fixed genus and fixed number of punctures while the homology of $L^{(2)}_{\infty}$ is related to spaces of outer automorphisms of free groups with punctures (see~\cite{Ko}).
\subsubsection{Unimodular Lie algebras} Let $\mathfrak{a}$ be a Lie algebra of finite dimension $n$ in homological degree $0$. The Chevalley-Eilenberg coalgebra $C\,:=\,\mathcal{C}(\mathfrak{a};k)$ is Koszul dual to $\mathfrak{a}$ and isomorphic to ${\rm{Sym}}^c(\mathfrak{a}[1])$ as graded vector spaces. For the pairing $\langle \mbox{--},\mbox{--} \rangle$ from Section~\ref{s2.2.2} to be compatible with differentials, it is necessary and sufficient that $\mathfrak{a}$ be {\it unimodular} i.e, that ${\rm{Tr}}(\mathrm{ad}(x))=0$ for all $x\,\in\,\mathfrak{a}$. This is the case, for instance, if $\mathfrak{a}$ is semisimple. Thus, for unimodular $\mathfrak{a}$, $\overline{\mathrm{HC}}_{\bullet}(\mathfrak{a})$ has a derived Poisson structure of degree $2-n$, to which Theorem~\ref{thodgefiltration} applies. Given that the graded $(2-n)$-Lie structure on $\boldsymbol{\Omega}(C)_{\n}$ is identical to that of Section~\ref{s2.2.2} (with the Chevalley-Eilenberg diferential being the only new ingredient), it is reasonable to expect that for any unimodular Lie algebra $\mathfrak{a}$, $\{\mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a}), \mathrm{HC}^{(q)}_{\bullet}(\mathfrak{a})\} \subseteq \mathrm{HC}^{(p+q-2)}_{\bullet}(\mathfrak{a})$ for all $p,q$.
\section{Topological applications}
\label{Sect4}
In this section, we will give a topological interpretation of Lie Hodge decompositions.
This interpretation is based on Quillen's famous theorem \cite{Q2} assigning to each simply connected topological space $ X $ a DG Lie algebra $\mathfrak{a}_X$ called a Lie model of $X$. The cyclic homology of $\, \U \mathfrak{a}_X $ can be identified with the $S^1$-equivariant homology of the free loop space $ \LL X$ of $ X $, and our main observation is that the Hodge
components $ \mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a}_X) $ correspond precisely to eigenspaces
of Frobenius maps under this identification.
For an application, we will look at the Chas-Sullivan Lie algebra of a simply connected
closed manifold and show that the corresponding Lie bracket respects the Hodge filtration, thus making the string topology algebra a {\it filtered} Lie algebra.
Throughout this section, we assume that $k= \mathbb{Q} $ and all homology and cohomology groups are taken with rational coefficients.
\subsection{Hodge decomposition of $S^1$-equivariant homology}
\la{Sect4.1}
Let $X$ be a 1-connected
topological space of finite rational type.
Recall ({\it cf.} \cite{FHT}) that one can associate to $ X $ a commutative
cochain DG algebra $ {\mathcal A}_X $, called a {\it Sullivan model} of $X$,
and a connected (chain) DG Lie algebra $ \mathfrak{a}_X $, called a {\it Quillen model} of $X$.
Each of these algebras is uniquely determined up to homotopy and each encodes the
rational homotopy type of $X$. The relation between
them is given by a DG algebra quasi-isomorphism
\begin{equation}
\la{SQ}
\mathcal{C}^{\bullet}(\mathfrak{a}_X;\mathbb Q) \stackrel{\sim}{\to} {\mathcal A}_X\ ,
\end{equation}
where $\, \mathcal{C}^{\bullet}(\mathfrak{a}_X;\mathbb{Q}) \,$ is the Chevalley-Eilenberg cochain complex of $ \mathfrak{a}_X $.
Now, let $\LL X$ denote the free loop space of $X$, i.e. the space of all continuous
maps $ S^1 \to X $ equipped with compact open topology. This space carries a natural
$ S^1$-action (induced by rotations of $ S^1 $), and one can define the $S^1$-equivariant
homology of $ \LL X $ via the Borel construction:
$$
\mathrm H^{S^1}_{\bullet}(\LL X) := \mathrm H_\bullet(ES^1 \times_{S^1} \LL X, \mathbb{Q})\ .
$$
We will use a {\it reduced} version of equivariant homology, which is defined by
$$
\rH^{S^1}_{\bullet}(\LL X) := {\rm{Ker}}[\,\mathrm H^{S^1}_{\bullet}(\LL X) \xrightarrow{\pi_*} \mathrm H_\bullet(BS^1)\,]\ ,
$$
where the map $ \pi_* $ comes from the natural homotopy fibration
\begin{equation}
\la{fibr}
\LL X \to ES^1 \times_{S^1} \LL X \xrightarrow{\pi} BS^1 \ .
\end{equation}
The following theorem is a well known result due to Goodwillie \cite{Go} and Jones \cite{J}
(see also \cite{JM}).
\begin{theorem}[\cite{J}]
\la{top1}
There are natural isomorphisms of graded vector spaces
\begin{equation*}
\la{ax}
\alpha_X:\, {\rHH}_{\bullet}(\mathcal U\mathfrak{a}_X) \xrightarrow{\sim} \overline{\mathrm H}_{\bullet}(\mathcal L X)
\ ,\qquad
\beta_X:\, \overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a}_X) \xrightarrow{\sim} {\rH}^{S^1}_{\bullet}(\LL X)
\end{equation*}
transforming the Connes periodicity sequence for $ \, \U\mathfrak{a}_X $ to the Gysin long exact sequence for the $S^1$-equivariant homology of $ \LL X $.
\end{theorem}
Next, for each integer $n \ge 0 $, we consider the $n$-fold covering of the circle:
$$
\omega^n\,:\,S^1 \xrightarrow{} S^1\ ,\quad e^{i\theta} \mapsto e^{in\theta}\ ,
$$
and denote by $\,\varphi_X^n\,;\,\mathcal L X \xrightarrow{} \mathcal L X\,$ the induced map on $ \LL X$. While the maps $\varphi_X^n$ are not equivariant with respect to the $S^1$-action on $ \LL X$, they fit into the commutative diagram ({\it cf.} \cite{BFG})
\begin{equation}\la{diag}
\begin{diagram}
\LL X & \rTo^{\varphi_X^n} & \LL X\\
\dTo & & \dTo\\
(ES^1 \times_{S^1} \mathcal L X)_\mathbb{Q} & \rTo^{\tilde{\varphi}_X^n} & (ES^1 \times_{S^1} \mathcal L X)_\mathbb{Q} \\
\dTo^{} & & \dTo_{}\\
(BS^1)_\mathbb{Q} & \rTo^{B\omega^n} & (BS^1)_\mathbb{Q}\\
\end{diagram}\ ,
\end{equation}
%
where the columns arise from the rationalization (Bousfield localization at zero) of the homotopy fibration \eqref{fibr}. The maps $ \varphi_X^n $ and $ \tilde{\varphi}_X^n $ in \eqref{diag} induce graded linear endomorphisms
\begin{equation*}
\Phi_X^n:\, {\rH}_{\bullet}(\LL X) \to {\rH}_{\bullet}(\LL X)
\ ,\qquad \tilde{\Phi}_X^n :\, {\rH}^{S^1}_{\bullet}(\LL X) \to {\rH}^{S^1}_{\bullet}(\LL X)\ ,
\end{equation*}
which are called the {\it power} or {\it Frobenius operations} on $ {\rH}_{\bullet}(\LL X) $ and $ {\rH}^{S^1}_{\bullet}(\LL X) $. We write
\begin{equation}\la{phi}
{\rH}^{(d)}_{\bullet}(\LL X)\, := \bigcap_{n \ge 0}\,{\rm{Ker}}(\Phi_X^n - n^d\,{\rm{Id}})\ ,\qquad
{\rH}^{S^1,\, (d)}_{\bullet}(\LL X)\, :=
\bigcap_{n \ge 0}\,{\rm{Ker}}(\tilde{\Phi}_X^n - n^d\,{\rm{Id}})
\end{equation}
for the common eigenspaces of these endomorphisms corresponding to the eigenvalues $ n^d $.
Now, the main result of this section can be stated as follows.
\begin{theorem}
\la{top2}
For each $p\ge 0$, there are natural isomorphisms
\begin{equation*}
\overline{{\rm{HH}}}^{(p)}_{\bullet}(\mathfrak{a}_X) \cong {\rH}^{(p)}_{\bullet}(\mathcal L X) \ ,\qquad
\overline{\mathrm{HC}}^{(p)}_{\bullet}(\mathfrak{a}_X) \cong
{\rH}^{S^1, \,(p-1)}_{\bullet}(\LL X)
\end{equation*}
given by restriction of the isomorphisms $ \alpha_X $ and $\beta_X$ of Theorem~\ref{top1}.
\end{theorem}
To prove Theorem~\ref{top2} we recall from \cite{bfprw1} the construction of a Lie Hodge decompostion in terms of (dual) Adams operations on the cobar construction of a cocommutative DG algebra. First, recall that by Adams
operations on a commutative algebra $A$ one usually means a
family $\{\psi^n\}_{n\ge 0}$ of algebra homomorphisms
$ \psi^n: A \to A $ satisfying the relations
\begin{equation}
\la{adams1}
\psi^1 = {\rm{Id}}\ ,\quad \psi^n \circ \psi^m = \psi^{nm}\ .
\end{equation}
Dually, for a cocommutative coalgebra $ A $,
we define Adams operations to be a
family $\{\psi^n\}_{n\ge 0}$ of {\it coalgebra} homomorphisms
$ \psi^n: A \to A $ satisfying \eqref{adams1}.
Now, if $ A $ is either a commutative or cocommutative Hopf algebra, there are natural Adams operations on $A$ defined by the formulas\footnote{For commutative Hopf algebras, these Adams operations are defined and studied in \cite[Section~4.5]{L}, while, for cocommutative Hopf algebras, they appear in \cite[Section~7]{bfprw1}.}
\begin{equation}
\la{adams2}
\psi^1 = {\rm{Id}}\ ,\quad \psi^n = \mu_n \circ \Delta^n\ ,\ n \ge 2\ ,
\end{equation}
where $\,\mu_n:\,A^{\otimes n} \to A\,$ is the $n$-fold product and $\,\Delta^n: A \to A^{\otimes n}\, $ is the $n$-fold coproduct
on $A$.
Next, recall that, if $ \mathfrak{a} $ is a DG Lie algebra, its Koszul dual is a cocommutative (conilpotent) DG coalgebra $C$. Associated to $C$
are two cobar constructions: the classical (Adams) cobar construction
$ \bOmega(C) $, which gives a cofibrant resolution of $ \U\mathfrak{a} $ in $ \mathtt{DGA}_{k/k} $, and the Lie cobar construction $ \bOmega_{\tt Comm}(C) $, which gives a cofibrant resolution of $ \mathfrak{a} $ in $ \mathtt{DGLA}_k $. The relation between these constructions is given by a canonical isomorphism $ \bOmega(C) \cong \U[ \bOmega_{\tt Comm}(C)] $, which shows that $ \bOmega(C) $ has a natural structure of a cocommutative DG Hopf algebra. Thus, $ \bOmega(C) $ can be equipped with a family $ \{\psi^n\}_{n\ge 0} $ of (coalgebra) Adams operations given by formulas \eqref{adams2}. It is shown in \cite{bfprw1} (see {\it op. cit.}, Section~7) that these operations on $ \bOmega(C) $ induce Adams operations on the (reduced) cyclic homology of $\, \U\mathfrak{a} $, and the construction of \cite{bfprw1} can be extended to define Adams operations on the Hochschild homology of $\, \U\mathfrak{a} $.
We denote these Adams operations on $\, \rHH_\bullet(\U \mathfrak{a}) \,$ and
$\, \overline{\mathrm{HC}}_\bullet(\U \mathfrak{a}) \,$ by $ \Psi^n $ and $ \tilde{\Psi}^n $, respectively.
Proposition~7.3 of \cite{bfprw1} then implies that the Hodge components $ {\rm{HH}}_\bullet^{(p)}(\mathfrak{a}) $ and $ \mathrm{HC}_\bullet^{(p)}(\mathfrak{a}) $ are the common
eigenspaces of $ \Psi^n $ and $ \tilde{\Psi}^n $ corresponding to the eigenvalues $ n^p $:
%
\begin{equation}
\la{psi}
{\rHH}_\bullet^{(p)}(\mathfrak{a})\, =
\bigcap_{n \ge 0}\,{\rm{Ker}}(\Psi^n - n^p\,{\rm{Id}})\ , \qquad
{\overline{\mathrm{HC}}}_\bullet^{(p)}(\mathfrak{a})\, =
\bigcap_{n \ge 0}\,{\rm{Ker}}(\tilde{\Psi}^n - n^p\,{\rm{Id}})\ .
\end{equation}
Now, let $ \Psi_X^n $ and $ \tilde{\Psi}_X^n $ denote the Adams operations on $\, \rHH_\bullet(\U \mathfrak{a}_X) $ and $\, \overline{\mathrm{HC}}_\bullet(\U\mathfrak{a}_X) $ coming from $\boldsymbol{\Omega}(C) $,
where $ C = \mathcal{C}_{\bullet}(\mathfrak{a}_X; \mathbb{Q}) $ is the Chevalley-Eilenberg chain complex of $ \mathfrak{a}_X $. Theorem~\ref{top2} follows immediately from \eqref{phi}, \eqref{psi} and the next key proposition.
\begin{proposition}
\la{intert}
For each $ n \ge 0 $, the isomorphism $\alpha_X $ intertwines the Adams operation $ \Psi_X^n $ with the Frobenius operation $ \Phi_X^n $, and the isomorphism $ \beta_X$ intertwines $ \tilde{\Psi}_X^n $ with
the $n$-th multiple of $ \tilde{\Phi}^n_X$, i.e.
$$
\alpha_X \circ \Psi^n_X \,=\, \Phi_X^n \circ \alpha_X \ ,\qquad
\beta_X \circ \tilde{\Psi}^n_X \, = \, n\,\tilde{\Phi}^n_X \circ \beta_X\ .
$$
\end{proposition}
\begin{proof}
We deduce Proposition~\ref{intert} from results of the paper \cite{BFG}. Theorem~B of that paper provides natural isomorphisms
\begin{equation} \la{abx}
a_X:\, {\rHH}_{- \bullet}({\mathcal A}_X) \xrightarrow{\sim} {\rH}^{\bullet}(\mathcal L X)
\ ,\qquad
b_X:\, \overline{\mathrm{HC}}_{- \bullet}({\mathcal A}_X) \xrightarrow{\sim} {\rH}_{S^1}^{\bullet - 1}(\LL X)\ ,
\end{equation}
relating the Hochschild and cyclic homology of a Sullivan model of
$ X $ to the (reduced) cohomology and $ S^1 $-equivariant cohomology of $ \LL X $. Since $ {\mathcal A}_X $ is a commutative DG algebra,
its Hochschild and cyclic homology carry natural Adams operations
which we denote by $ \Psi_n $ and $ \tilde{\Psi}_n $, respectively. On the other hand, the cohomology and $ S^1 $-equivariant cohomology of $ \LL X $ carry Frobenius operations induced by $ \varphi^n $: we denote these by $ \Phi_n:\, {\rH}^{\bullet}(\mathcal L X) \to {\rH}^{\bullet}(\mathcal L X) $ and $ \tilde{\Phi}_n:
{\rH}_{S^1}^{\bullet}(\LL X) \to {\rH}_{S^1}^{\bullet}(\LL X) $.
Now, part $(2)$ of \cite[Theorem~B]{BFG} says that
\begin{equation}
\la{ab3}
a_X \circ \Psi_n \,=\, \frac{1}{n}\, \Phi_n \circ a_X \ ,\qquad
b_X \circ \tilde{\Psi}_n \, = \, \tilde{\Phi}_n \circ b_X\ .
\end{equation}
To relate the isomorphisms \eqref{abx} to those of Theorem~\ref{top1}, we recall a theorem of Quillen~\cite{Q} identifying
\begin{equation} \la{quil}
{\rHH}_{\bullet}(\mathcal U\mathfrak{a}_X)\,\cong\, \overline{{\rm{HH}}}_{\bullet}[\mathcal{C}_\bullet(\mathfrak{a}_X;\mathbb Q)] \ ,
\qquad \overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a}_X) \,\cong\, \overline{\mathrm{HC}}_{\bullet+1}[\mathcal{C}_\bullet(\mathfrak{a}_X;\mathbb Q)]\ .
\end{equation}
Since $X$ is assumed to be of finite rational type, its Lie and Sullivan models $ \mathfrak{a}_X $ and $ {\mathcal A}_X $ are locally finite DG algebras, i.e. have finite-dimensional components in each homological degree. Dualizing
\eqref{SQ}, we then have a quasi-isomorphism of DG coalgebras $\, {\mathcal A}_X^* \stackrel{\sim}{\to} \mathcal{C}^{\bullet}(\mathfrak{a}_X;\mathbb{Q})^* \cong \mathcal{C}_{\bullet}(\mathfrak{a}_X;\mathbb{Q})\,$.
Combined with \eqref{abx} and \eqref{quil}, this quasi-isomorphism induces natural isomorphisms
\begin{eqnarray}
&& \overline{{\rm{HH}}}_{\bullet}(\mathcal U\mathfrak{a}_X) \,\cong\, \overline{{\rm{HH}}}_{\bullet}[\mathcal{C}_\bullet(\mathfrak{a}_X;\mathbb Q)] \,\cong\, {\rHH}_{-\bullet}({\mathcal A}_X)^{\ast} \,\xrightarrow{(a_X^{-1})^*}\, \overline{\mathrm H}^{\bullet}(\mathcal L X)^{\ast} \,\cong\, \overline{\mathrm H}_{\bullet}(\mathcal L X)\ , \la{mm1}\\
&&\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a}_X) \,\cong\, \overline{\mathrm{HC}}_{\bullet+1}[\mathcal{C}_\bullet (\mathfrak{a}_X;\mathbb Q)] \,\cong\, \overline{\mathrm{HC}}_{-\bullet-1}({\mathcal A}_X)^{\ast} \,\xrightarrow{(b_X^{-1})^*}\, \overline{\mathrm H}^{\bullet}_{S^1}(\mathcal L_X)^{\ast} \,\cong\, \overline{\mathrm H}^{S^1}_{\bullet}(\mathcal L X)\ ,\la{mm2}
\end{eqnarray}
where the star $(\,\mbox{--}\,)^{\ast}$ stands for graded linear duals. Since $a_X$ and $b_X$ are functorial and transform Connes' exact sequence on homology to the Gysin sequence for $S^1$-equivariant cohomology, the isomorphisms \eqref{mm1} and \eqref{mm2} are functorial and transform Connes' sequence to the Gysin sequence on $S^1$-equivariant homology. Thus \eqref{mm1} and \eqref{mm2} coincide with the isomorphisms $\alpha_X$ and $\beta_X$ of Theorem~\ref{top1}.
Now, by~\cite[Proposition 7.4]{bfprw1}, the direct summand $\mathrm{HC}_{\bullet}^{(p)}(\mathfrak{a}_X)$ of $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a}_X)$ is identified with $ \overline{\mathrm{HC}}_{-\bullet-1}^{(p-1)}({\mathcal A}_X)^{\ast}$ under the first two isomorphisms in \eqref{mm2}, and by \cite[Theorem~B]{BFG}, $\, \overline{\mathrm{HC}}_{-\bullet-1}^{(p-1)}({\mathcal A}_X)^{\ast}$ is the common eigenspace of the Adams operations $ \tilde{\Psi}^*_n $ with the eigenvalues $n^{p-1}$. On the other hand, $ \mathrm{HC}_{\bullet}^{(p)}(\mathfrak{a}_X) $ is the common eigenspace of the Adams operations $\tilde{\Psi}^n $ with the eigenvalues $ n^p $. Hence the first two isomorphisms of \eqref{mm2} intertwine $\tilde{\Psi}^n $ on $\overline{\mathrm{HC}}_{\bullet}(\mathcal U\mathfrak{a}_X)$ with $\, n \tilde{\Psi}^*_n \,$ on $\, \overline{\mathrm{HC}}_{-\bullet-1}^{(p-1)}({\mathcal A}_X)^{\ast}$. Now, by \eqref{ab3}, the map $\, (b_X^{-1})^{\ast}$ intertwines $ \tilde{\Psi}^*_n $ with $\tilde{\Phi}^*_n $, and under the last isomorphism in \eqref{mm2}, the dual power operation $\tilde{\Phi}^*_n $ on $ \rH^{\bullet}_{S^1}(\mathcal L_X)^{\ast} $ corresponds to $ \tilde{\Phi}^n $ on $ \rH^{S^1}_{\bullet}(\mathcal L X) $. Thus,
the composite map \eqref{mm2} intertwines $\tilde{\Psi}^n $ with $\,n\,\tilde{\Phi}^n $. This proves the second equality of Proposition~\ref{intert}. The proof of the first is similar.
\end{proof}
Theorem~\ref{top2} together with Theorem~\ref{hodgesbi} has the following important corollary.
\begin{corollary} \la{hsstring}
There are natural Hodge decompositions
$$
{\rH}_{\bullet}(\mathcal L X) \,\cong\, \bigoplus_{p=0}^{\infty} \overline{\mathrm H}^{(p)}_{\bullet}(\mathcal L X)\ ,\qquad
\rH^{S^1}_{\bullet}(\mathcal L X)\,\cong\,\bigoplus_{p=0}^{\infty} \overline{\mathrm H}^{S^1,\,(p)}_{\bullet}(\mathcal L X)\ .
$$
The Gysin sequence decomposes into a direct sum of Hodge components
$$
\ldots\, \xrightarrow{D} \,\overline{\mathrm H}^{S^1,\, (p+1)}_{n-1}(\mathcal L X) \,\to\,\overline{\mathrm H}^{(p)}_{n}(\mathcal L X) \,\to\, \overline{\mathrm H}^{S^1, (p)}_{n}(\mathcal L X) \,\xrightarrow{D}\, \overline{\mathrm H}^{S^1, \, (p+1)}_{n-2}(\mathcal L X)\,\to\, \ldots
$$
where $D$ stands for the Gysin map.
\end{corollary}
\subsection{Hodge filtration on string topology}
Let $M$ be a simply connected closed oriented manifold of dimension $d$.
A construction of Lambrechts and Stanley (see \cite{LS}, Theorem~1.1) provides a finite-dimensional commutative DG algebra $ \mathcal A$, which is a model for the singular cochain complex of $M$. This model comes equipped with a nondegenerate cyclic pairing of (cohomological) degree $ n = -d $. The linear dual of $ \mathcal A$ gives a cocommutative coalgebra model
$ C := {\mathcal A}^* $ for the singular chain complex of $M$ that has a cyclic pairing of (homological) degree $ n = -d $. It is known (and easy to check) that
$\boldsymbol{\Omega}_{\mathtt{Comm}}(C)$ is a Lie model of $M$; hence, by Theorem~\ref{top1}, we have
$$
\overline{\mathrm{HC}}_{\bullet}[\boldsymbol{\Omega}(C)]\,\cong\, \overline{\mathrm H}^{S^1}_{\bullet}(\mathcal L M)\,,\,\,\,\,\, \overline{{\rm{HH}}}_{\bullet}[\boldsymbol{\Omega}(C)]\,\cong\, \overline{\mathrm H}_{\bullet}(\mathcal L M) \ ,
$$
and the cyclic pairing on $C$ induces a derived Poisson structure on $\overline{\mathrm H}^{S^1}_{\bullet}(\mathcal L M)$ of degree $n+2$. The corresponding Lie bracket coincides with the Chas-Sullivan bracket \cite{ChS}, and the resulting Lie algebra is called the {\it string topology Lie algebra} of $M$ ({\it cf.}~\cite{CEG}). Further, by Theorem~\ref{liestronhom}, $\overline{\mathrm H}_{\bullet}(\mathcal L M)$ is a Lie module (of degree $n+2$) over the string topology Lie algebra of $M$, and the Gysin map $D\,:\, \overline{\mathrm H}^{S^1}_{\bullet}(\mathcal L M) \xrightarrow{} \overline{\mathrm H}^{S^1}_{\bullet-2}(\mathcal L M)$ is a Lie module homomorphism.
We may now apply in this situation the results of Section~\ref{Sect3} and Section~\ref{Sect4}:
as a consequence of Theorem~\ref{thodgefiltration}, Theorem~\ref{pstrhoch1}, Theorem~\ref{pstrhoch2} and Theorem~\ref{top2}, we get
\begin{theorem} \la{tstringhomology}
$(i)$
The string topology Lie algebra of a closed $d$-dimensional manifold $M$ is filtered as a Lie algebra with respect to the following {\rm Hodge filtration}
$$
F_p\overline{\mathrm H}^{S^1}_{\bullet}(\mathcal L M) \,:=\, \bigoplus_{q \leq p+1} \overline{\mathrm H}^{S^1,\, (q)}_{\bullet}(\mathcal L M)\,\text{.} $$
\noindent
$(ii)$ The homology of the free loop space $\overline{\mathrm H}_{\bullet}(\mathcal L M)$ is filtered as a Lie module over the string topology Lie algebra of $M$ with respect to the following Hodge filtration
$$ F_p\overline{\mathrm H}_{\bullet}(\mathcal L M) \,:=\, \bigoplus_{q \leq p+2} \overline{\mathrm H}^{(q)}_{\bullet}(\mathcal L M)\,\text{.} $$
\noindent
$(iii)$ The Chas-Sullivan bracket restricts to the first Hodge component $\, {\rH}^{S^1,\,(1)}_{\bullet}(\mathcal L M)$, making it a Lie algebra. Further, $ {\rH}^{S^1}_{\bullet}(\mathcal L M)$ is a graded Lie module over $\overline{\mathrm H}^{S^1,\,(1)}_{\bullet}(\mathcal L M)$ with the grading given by the Hodge decomposition of $\overline{\mathrm H}^{S^1}_{\bullet}(\mathcal L M)$.
\noindent
$(iv)$ For each $ p \ge 0 $, the Gysin map $\,D:\, \overline{\mathrm H}^{S^1,\,(p)}_{\bullet}(\mathcal L M) \xrightarrow{} \overline{\mathrm H}^{S^1,\,(p+1)}_{\bullet-2}(\mathcal L M)$ is a map of graded Lie modules over $\,\overline{\mathrm H}^{S^1, (1)}_{\bullet}(\mathcal L M)$.
\end{theorem}
The Lambrechts-Stanley Lie model $\boldsymbol{\Omega}_{\mathtt{Comm}}({\mathcal A}^*)$ and unimodular Lie algebras (in particular, abelian Lie algebras) are Koszul dual to ``Poincar\'{e} duality CDGC'', i.e. cocommutative DG coalgebras that are graded linear duals of Poincar\'{e} duality CDGA's in the sense of~\cite{LS}. This should be contrasted with
necklace Lie algebras, in which case the coproduct on the Koszul dual coalgebra and the cyclic pairing are unrelated. Given the example of symmetric algebras in Section~\ref{s2.2.2}, where the Hodge filtration actually becomes a Hodge decomposition, we expect the following conjecture to be true.
\begin{conjecture}
\la{conj1}
Let $\mathfrak{a}\,\in\, \mathtt{DGLA}_k$ be Koszul dual to a connected finite-dimensional $C\,\in\,\mathtt{DGCC}_{k/k}$ equipped with a Poincar\'{e} duality pairing of degree $n$. Then, the Hodge filtrations on the derived Poisson structures on $\overline{\mathrm{HC}}_{\bullet}(\mathfrak{a})$ and $\overline{{\rm{HH}}}_{\bullet}(\mathfrak{a})$ become direct Hodge decompositions, i.e. for all $p,q \ge 0 $,
$$\{\mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a}),\, \mathrm{HC}^{(q)}_{\bullet}(\mathfrak{a})\} \,\subseteq \, \mathrm{HC}^{(p+q-2)}_{\bullet}(\mathfrak{a})\ ,\quad \{\mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a}),\, {\rm{HH}}^{(q)}_{\bullet}(\mathfrak{a})\} \,\subseteq \, {\rm{HH}}^{(p+q-2)}_{\bullet}(\mathfrak{a})\ .
$$
In particular, we expect that the Chas-Sullivan bracket of a closed $d$-dimensional manifold satisfies
$$
\{\overline{\mathrm H}^{S^1,\,(p)}_{\bullet}(\mathcal L M), \ \overline{\mathrm H}^{S^1,\,(q)}_{\bullet}(\mathcal L M)\}\, \subseteq\, \overline{\mathrm H}^{S^1,\, (p+q-1)}_{\bullet}(\mathcal L M)\ .
$$
\end{conjecture}
\section{Relation to derived representation schemes} \la{s4}
In this section, we recall the construction of derived representation schemes of Lie algebras from~\cite[Section 6, Section 7]{bfprw1}. The main result of this section is Theorem~\ref{dreppoiss}, the proof of which we outline. The full details will appear in~\cite{Yin}.
\subsection{Derived representation schemes and Drinfeld traces}
Let $\g$ be a finite dimensional Lie algebra. Consider the functor
$$(\mbox{--})_{\g} \,:\, \mathtt{DGLA}_k \xrightarrow{} \mathtt{DGCA}_{k/k}\,, \,\,\,\,\mathfrak{a} \,\mapsto \, \mathfrak{a}_{\g}\,:=\, \frac{{\rm{Sym}}_k(\mathfrak{a} \otimes \g^{\ast})}{\langle\langle ( x \otimes \xi_1).(y \otimes \xi_2) -(y \otimes \xi_1).(x \otimes \xi_2) -[x,y] \otimes \xi \rangle \rangle} \,,$$
where $\g^{\ast}$ is the vector space dual to $\g$ and where $\xi \mapsto \xi_1 \wedge \xi_2$ is the map dual to the Lie bracket on $\g$. The augmentation on $\mathfrak{a}_\g$ is the one induced by the map taking the generators $\mathfrak{a} \otimes \g^{\ast}$ to $0$. Let $\g({\mbox{--}})\,:\, \mathtt{DGCA}_{k/k} \xrightarrow{} \mathtt{DGLA}_k$ denote the functor $B \mapsto \g(\bar{B})\,:=\, \g \otimes \bar{B}$. It is shown in~\cite[Section 6.3]{bfprw1} that the functors $(\mbox{--})_{\g}\,:\, \mathtt{DGLA}_k \rightleftarrows \mathtt{DGCA}_{k/k}\,:\, \g({\mbox{--}})$ form a (Quillen) adjoint pair.
Thus, $\mathfrak{a}_\g$ is the commutative (DG) algebra corresponding to the (DG) scheme $\Rep_\g(\mathfrak{a})$ parametrizing representations of $\mathfrak{a}$ in $\g$. Since the functor $(\mbox{--})_{\g}$ is left Quillen, it has a well behaved left derived functor
$$\boldsymbol{L}(\mbox{--})_{\g}\,:\,{\mathtt{Ho}}(\mathtt{DGLA}_k) \xrightarrow{} {\mathtt{Ho}}(\mathtt{DGCA}_{k/k})\,\text{.}$$
Like for any left derived functor, we have $\boldsymbol{L}(\mathfrak{a})_\g\,\cong\, \mathcal L_\g $ in ${\mathtt{Ho}}(\mathtt{DGCA}_{k/k})$, where $\mathcal L \stackrel{\sim}{\xrightarrow{}} \mathfrak{a}$ is any cofibrant resolution in $\mathtt{DGLA}_k$. We define
$$\DRep_\g(\mathfrak{a})\,:=\, \boldsymbol{L}(\mathfrak{a})_\g \, \text{ in } {\mathtt{Ho}}(\mathtt{DGCA}_{k/k})\,,\,\,\,\,\mathrm H_{\bullet}(\mathfrak{a},\g)\,:=\, \mathrm H_{\bullet}[\boldsymbol{L}(\mathfrak{a})_\g]\,\text{.}$$
$\DRep_\g(\mathfrak{a})$ is called the {\it derived representation algebra} for representations of $\mathfrak{a}$ in $\g$. The homology $\mathrm H_{\bullet}(\mathfrak{a},\g)$ is called the {\it representation homology} of $\mathfrak{a}$ in $\g$. It is not difficult to check that $\g$ acts naturally by derivations on the graded (commutative) algebra $\mathrm H_{\bullet}(\mathfrak{a},\g)$. We denote the corresponding (graded) subalgebra of $\g$-invariants by $\mathrm H_{\bullet}(\mathfrak{a},\g)^{\ad\,\g}$.
Let $\mathcal L \stackrel{\sim}{\xrightarrow{}} \mathfrak{a}$ be a cofibrant resolution. The unit of the adjunction $(\mbox{--})_{\g}\,:\, \mathtt{DGLA}_k \rightleftarrows \mathtt{DGCA}_{k/k}\,:\, \g({\mbox{--}})$ is the universal representation
$$ \pi_\g\,:\, \mathcal L \xrightarrow{} \g(\mathcal L_\g) \,\text{.}$$
Let the functor $\lambda^{(p)}$ be as in Section~\ref{s1.1}. There is a natural map $\lambda^{(p)}[\g(\mathcal L_\g)] \xrightarrow{} \mathcal L_\g \otimes \lambda^{(p)}(\g)$. For $P\,\in\, I^p(\g)\,:=\,{\rm{Sym}}^p(\g^{\ast})^{\ad\,\g}$, evaluation at $P$ gives a linear functional $\mathrm{ev}_P$ on $\lambda^{(p)}(\g)$. One thus has the composite map
$$\begin{diagram} \lambda^{(p)}(\mathcal L) & \rTo^{\lambda^{(p)}(\pi_\g)} & \lambda^{(p)}[\g(\mathcal L_\g)] & \rTo& \mathcal L_\g \otimes \lambda^{(p)}(\g) & \rTo^{{\rm{Id}} \otimes \mathrm{ev}_P} & \mathcal L_\g \end{diagram} $$
for $P\,\in\,I^p(\g)$. On homologies, this gives the map
$$ {\rm{Tr}}_\g(P, \mathfrak{a})\,:\, \mathrm{HC}^{(p)}_{\bullet}(\mathfrak{a}) \xrightarrow{} \mathrm H_{\bullet}(\mathfrak{a},\g)^{\ad\,\g}\,,$$
which we call the {\it Drinfeld trace map } associated to $P$ (see~\cite[Section 7]{bfprw1} for further details regarding this construction). If $\g$ is semisimple, the Killing form is a canonical element of $I^2(\g)$. We denote the associated Drinfeld trace by
$${\rm{Tr}}_\g(\mathfrak{a})\,:\,\mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a}) \xrightarrow{} \mathrm H_{\bullet}(\mathfrak{a},\g)^{\ad\,\g}\,\text{.}$$
Let $\mathfrak{a}\,\in\,\mathtt{DGLA}_k$ be Koszul dual to $C\,\in\,\mathtt{DGCC}_{k/k}$, where $C$ is $n$-cyclic.
\begin{theorem} \la{dreppoiss}
There is a Poisson structure on $\mathrm H_{\bullet}(\mathfrak{a},\g)^{\ad\,\g}$ such that the Drinfeld trace map ${\rm{Tr}}_\g(\mathfrak{a})$ is a graded Lie algebra homomorphism.
\end{theorem}
Theorem~\ref{dreppoiss} may thus be viewed as a generalization of~\cite[Theorem 6.7]{G} for the Lie operad (the latter result is proven only for free algebras over operads). It could also be seen as an analog of~\cite[Theorem 2]{BCER} in the Lie setting.
\subsection{Proof of Theorem~\ref{dreppoiss}}
Recall from~\cite[Section 4.5]{GK} that a cyclic pairing $\langle \mbox{--}, \mbox{--} \rangle$ on a (graded) Lie algebra $\mathfrak{a}$ is a symmetric pairing that is $\ad$-invariant. Equivalently, the map $\langle \mbox{--}, \mbox{--} \rangle \,:\, \mathfrak{a} \otimes \mathfrak{a} \xrightarrow{} k$ is a (graded) $\mathfrak{a}$-module homomorphism, where $k$ is equipped with the trivial action. Dually, a cyclic pairing on a (graded) Lie coalgebra $\mathfrak{G}$ is a pairing $\langle \mbox{--}, \mbox{--} \rangle\,:\, \mathfrak{G} \otimes \mathfrak{G} \xrightarrow{} k$ that is a (graded) $\mathfrak{G}$-comodule homomorphism, with $k$ equipped with the trivial coaction. Explicitly, if $\langle \mbox{--},\mbox{--} \rangle$ is a cyclic pairing of degree $n$ on a Lie coalgebra $\mathfrak{G}$, then if $\Delta(x) \,=\, x^1 \otimes x^2$, etc. in the Sweedler notation, we have
$$x^1 \langle x^2, y \rangle \pm x^2 \langle x^1, y \rangle \pm y^1 \langle x, y^2 \rangle \pm y^2 \langle x, y^1 \rangle \,=\,0\,, $$
where the signs are determined by the Koszul sign rule. If $\mathfrak{G}$ is differential graded, we further demand that a cyclic pairing of degree $n$ on $\mathfrak{G}$ be compatible with differential, i.e, that for all $x,y\,\in\,\mathfrak{G}$, $\langle \delta x, y \rangle +(-1)^{|x|+n}\langle x, \delta y \rangle =0$.
Further recall that for a DG Lie coalgebra $\mathfrak{G}$, one has the Chevalley-Eilenberg algebra $\mathcal{C}^c(\mathfrak{G};k)$ which is the construction formally dual to the Chevalley-Eilenberg coalgebra $\mathcal{C}(\mathfrak{a};k)$ os a DG Lie algebra.
\begin{lemma} \la{poissce}
If $\mathfrak{G}\,\in\,\mathtt{DGLC}_k$ is equipped with a cyclic pairing of degree $n$, then the Chevalley-Eilenberg algebra $\mathcal{C}^c(\mathfrak{G};k)$ acquires a DG Poisson structure.
\end{lemma}
\begin{proof}
Note that as a graded commutative algebra, $\mathcal{C}^c(\mathfrak{G};k)\,\cong\, {\rm{Sym}}(\mathfrak{G}[-1])$. The degree $n$-cyclic pairing on $\mathfrak{G}$ gives a skew symmetric degree pairing on $\mathfrak{G}[-1]$. This gives a graded Poisson structure on ${\rm{Sym}}(\mathfrak{G}[-1])$. The compatibility of this structure with respect to the differential on $\mathcal{C}^c(\mathfrak{G};k)$ follows from the fact that the pairing on $\mathfrak{G}$ is cyclic.
\end{proof}
Let $\mathfrak{a}\,\in\,\mathtt{DGLA}_k$ be Koszul dual to $C\,\in\,\mathtt{DGCC}_{k/k}$ where $C$ is $n$-cyclic. If $\g$ is semisimple, then the Killing form on $\g$ gives an isomorphism $\g \, \cong\,\g^{\ast}$. Under this isomorphism, the Killing form on $\g$ is identified with a cyclic pairing $\kappa$ on $\g^{\ast}$. Note that $\g^{\ast}(\bar{C})\,:=\,\g^{\ast} \otimes \bar{C}$ has the structure of a DG Lie coalgebra. Tensoring $\kappa$ with the pairing on $\bar{C}$, we obtain a degree $n$-cyclic pairing on $\g^{\ast}(\bar{C})$. By Lemma~\ref{poissce}, this gives a DG Poisson structure of degree on $\mathcal{C}^c(\g^{\ast}(C);k)$, which represents $\DRep_\g(\mathfrak{a})$ in ${\mathtt{Ho}}(\mathtt{DGCA}_{k/k})$ by~\cite[Theorem 6.5]{bfprw1}. On homologies, we obtain a graded Poisson structure on $\mathrm H_{\bullet}(\mathfrak{a},\g)$. It can be shown that the above Poisson structure on $\mathrm H_{\bullet}(\mathfrak{a},\g)$ restricts to a Poisson structure on $\mathrm H_{\bullet}(\mathfrak{a},\g)^{\ad\,\g}$.
On the other hand, the Killing form on $\g$ is an element of $I^2(\g)$. The associated Drinfeld trace is a map of graded vector spaces
${\rm{Tr}}_\g(\mathfrak{a})\,:\, \mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a}) \xrightarrow{} \mathrm H_{\bullet}(\mathfrak{a},\g)^{\ad\,\g}$. Note that by Theorem~\ref{thodgefiltration} (ii), $\mathrm{HC}^{(2)}_{\bullet}(\mathfrak{a})$ acquires the structure of a graded Lie algebra, namely, its derived Poisson structure. On the other hand, by Lemma~\ref{poissce}, the pairing on $\bar{C}$ induces a graded Poisson structure on $\mathrm H_{\bullet}(\mathfrak{a},\g)$. We now study the relation between these Poisson structures.
\subsubsection{} Let $\mathcal L\,:=\, \boldsymbol{\Omega}_{\mathtt{Comm}}(C)$. Thus, by~\cite[Theorem 6.5]{bfprw1}, $\mathcal L_\g\,\cong\, \mathcal{C}^c(\g^{\ast}(\bar{C});k)$. By Proposition~\ref{hodgebracketr} (ii), $\lambda^{(2)}(\mathcal L)$ is a DG Lie algebra. In addition, by Lemma~\ref{poissce}, $\mathcal L_\g$ has a DG Poisson structure. Let ${\rm{Tr}}_{\g}(\mathcal L)$ be the Drinfeld trace $\lambda^{(2)}(\mathcal L) \xrightarrow{} \mathcal L_\g$ associated with the Killing form on $\g$. Let $R\,:=\,\boldsymbol{\Omega}(C)$, which is isomorphic to $\mathcal U\mathcal L$ as a DG algebra. By Lemma~\ref{cycderhamhodge} and~\cite[Lemma A.1]{BR}, the map $\bar{\partial}\,:\, R \xrightarrow{} \Omega^1R_\n$ (see~\eqref{cyclicderham}) induces a map $\bar{\partial}\,:\, \lambda^{(2)}(\mathcal L) \xrightarrow{} \mathcal L \otimes V$, where $V\,:=\,\bar{C}[-1]$. Note that the DG-module of $\Omega^1_{\mathcal L_\g}$ of K\"{a}hler differentials on $\mathcal L_\g$ is isomorphic to $\mathcal L_\g \otimes \g^{\ast}(\bar{C})[-1]$ as a graded $\mathcal L_\g$-module. Let $d\,:\,\mathcal L_\g \xrightarrow{} \Omega^1_{\mathcal L_\g}$ be the de Rham differential. The prof of the following lemma will appear in~\cite{Yin}.
\begin{lemma} \la{twooneforms}
The following diagram commutes:
$$
\begin{diagram}
\lambda^{(2)}(\mathcal L) & \rTo^{\bar{\partial}} & \mathcal L \otimes V & \rTo^{\pi_{\g} \otimes {\rm{Id}}} & \mathcal L_\g \otimes \g \otimes V\\
& \rdTo_{{\rm{Tr}}_\g(\mathcal L)} & & & \dTo_{\cong} \\
& & \mathcal L_\g & \rTo^d & \Omega^1_{\mathcal L_\g}\\
\end{diagram}
$$
Here, the vertical isomorphism on the right identifies $\g$ with $\g^{\ast}$ through the Killing form.
\end{lemma}
It is not difficult to verify that the bracket on $\lambda^{(2)}(\mathcal L)$ is given by the composite map
$$ \begin{diagram} \lambda^{(2)}(\mathcal L) \otimes \lambda^{(2)}(\mathcal L) & \rTo^{\bar{\partial} \otimes \bar{\partial}} & (\mathcal L \otimes V) \otimes (\mathcal L \otimes V) & \rTo & (\mathcal L \otimes \mathcal L) \otimes (V \otimes V) & \rTo^{\mathrm{can} \otimes \langle \mbox{--}, \mbox{--} \rangle } & \lambda^{(2)}(\mathcal L) \end{diagram} \,,$$
where $\mathrm{can}\,:\, \mathcal L \otimes \mathcal L \xrightarrow{} \lambda^{(2)}(\mathcal L)$ is the canonical projection. By the construction of ${\rm{Tr}}_\g(\mathcal L)$, the following diagram commutes.
$$
\begin{diagram}
(\mathcal L \otimes V) \otimes (\mathcal L \otimes V) & \rTo^{(\pi_\g \otimes {\rm{Id}}) \otimes (\pi_\g \otimes {\rm{Id}})} & (\mathcal L_\g \otimes \g(\bar{C})[-1] ) \otimes (\mathcal L_\g \otimes \g(\bar{C})[-1] )\\
\dTo & & \dTo\\
(\mathcal L \otimes \mathcal L) \otimes (V \otimes V) & & (\mathcal L_\g \otimes \mathcal L_\g) \otimes (\g^{\ast}(\bar{C})[-1] \otimes \g^{\ast}(\bar{C})[-1])\\
\dTo^{\mathrm{can} \otimes \langle \mbox{--}, \mbox{--} \rangle} & & \dTo_{\mu \otimes \langle \mbox{--}, \mbox{--} \rangle}\\
\lambda^{(2)}(\mathcal L_\g) & \rTo^{{\rm{Tr}}_\g} & \mathcal L_\g\\
\end{diagram}
$$
Here, in the first vertical arrow on the right, $\g$ is identified with $\g^{\ast}$ via Killing form. $\mu$ is the product on $\mathcal L_\g$ and the pairing on $\g^{\ast}(\bar{C})[-1]$ is the one induced by the cyclic pairing on $\g^{\ast}(\bar{C})$. Composing the bottom arrow with the left vertical arrows composed with $\bar{\partial} \otimes \bar{\partial}$, we obtain ${\rm{Tr}}_\g(\{\mbox{--},\mbox{--}\})\,:\, \lambda^{(2)}(\mathcal L) \otimes \lambda^{(2)}(\mathcal L) \xrightarrow{} \mathcal L_\g$. On the other hand, by Lemma~\ref{twooneforms} composing the bottom arrow with the left vertical arrows composed with $\bar{\partial} \otimes \bar{\partial}$ gives us the map
$$ \lambda^{(2)}(\mathcal L) \otimes \lambda^{(2)}(\mathcal L) \xrightarrow{} \mathcal L_\g\,,\,\,\,\, (\alpha,\beta) \mapsto \langle d{\rm{Tr}}_\g(\alpha), d{\rm{Tr}}_\g(\beta) \rangle \,,$$
where the pairing $\langle \mbox{--},\mbox{--} \rangle \,:\, \Omega^1_{\mathcal L_\g} \otimes \Omega^1_{\mathcal L_\g} \xrightarrow{} \mathcal L_\g$ is induced by the pairing on $\g^{\ast}(\bar{C})[-1]$. This map is easily seen to be $\{{\rm{Tr}}_\g(\mbox{--}),{\rm{Tr}}_\g(\mbox{--})\}\,:\,\lambda^{(2)}(\mathcal L) \otimes \lambda^{(2)}(\mathcal L) \xrightarrow{} \mathcal L_\g$. Thus,
$${\rm{Tr}}_\g(\{\mbox{--},\mbox{--}\})\,=\, \{{\rm{Tr}}_\g(\mbox{--}),{\rm{Tr}}_\g(\mbox{--})\}\,:\,\lambda^{(2)}(\mathcal L) \otimes \lambda^{(2)}(\mathcal L) \xrightarrow{} \mathcal L_\g\,\text{.} $$
It is not difficult to verify that the DG Poisson bracket on $\mathcal L_\g$ restricts to a DG Poisson bracket on $\mathcal L_\g^{\ad\,\g}$. We thus obtain Theorem~\ref{dreppoiss} on homologies.
\section{Appendix: Free (Lie) algebras}
Let $V$ be a (homologically) graded $k$-vector space. Let $R\,:=\,T_kV$, the free graded $k$-algebra generated by $V$. Let $\mathcal L\,:=\, LV$, the free graded Lie algebra generated by $V$. It is well known that $R\,\cong\,\mathcal U\mathcal L$. Thus, the direct sum of symmetrization maps gives an isomorphism of graded $\mathcal L$-modules
$$ {\rm{Sym}}(\mathcal L) \,=\,\oplus_p {\rm{Sym}}^p(\mathcal L) \,\cong\, R \,\text{.}$$
In what follows, we shall view ${\rm{Sym}}^p(\mathcal L)$ as a graded subspace of $R$ via the symmetrization map. For $\beta_1,\ldots, \beta_p$ homogeneous in $R$, let $s(\beta_1,\ldots,\beta_p)$ denote the {\it symmetrization} of $\beta_1,\ldots, \beta_p$, i.e, the sum
$$s(\beta_1, \ldots,\beta_p) \,:=\, \sum_{ \sigma\,\in\,S_p} \pm \beta_{\sigma(1)} \cdot \ldots \cdot \beta_{\sigma(p)}\,,$$
where $\cdot$ denotes the product on $R$ and where the sign in front of each summand is the sign determined by the Koszul sign rule.
\begin{lemma} \la{derivations}
Let $\delta\,:\, R \xrightarrow{} R$ be a homogeneous derivation. Suppose that $\delta(V) \,\subseteq\, {\rm{Sym}}^q(\mathcal L)$.\\
(i) If $q>1$, then $\delta[{\rm{Sym}}^p(\mathcal L)] \,\subseteq\, \bigoplus_{r \leq p+q-1} {\rm{Sym}}^{r}(\mathcal L)$ for any $p \geq 1$.\\
(ii) If $q \leq 1$, then $\delta[{\rm{Sym}}^p(\mathcal L)] \,\subseteq\, {\rm{Sym}}^{p+q-1}(\mathcal L)$.
\end{lemma}
\begin{proof}
First, note that if $\alpha\,\in\,\mathcal L$ is homogeneous and if $\delta(\alpha)\,\in\, {\rm{Sym}}^q(\mathcal L)$, then for any homogeneous $v\,\in\,V$,
\begin{align*}
\delta([\alpha, v])\,&=\, \delta(\alpha \cdot v -(-1)^{|\alpha||v|} v \cdot \alpha)\\
&\,=\, \delta(\alpha) \cdot v +(-1)^{|\alpha||\delta|} \alpha \cdot \delta(v) -(-1)^{|\alpha||v|} \delta(v) \cdot \alpha -(-1)^{|\alpha||v|+|\delta||v|} v \cdot \delta(\alpha)\\
&\,=\, [\delta(\alpha), v] +(-1)^{|\alpha||\delta|} [\alpha, \delta(v)] \,\text{.}\\
\end{align*}
Since $[\mathcal L, {\rm{Sym}}^q(\mathcal L)] \,=\, [{\rm{Sym}}^q(\mathcal L), \mathcal L] \,\subseteq \, {\rm{Sym}}^q(\mathcal L)$ and $V \subseteq \mathcal L$, the element $\alpha_n := [v_1,[v_2, \ldots [v_{n-1}, v_n] \ldots ]]$ satisfies $\delta(\alpha_n) \subseteq {\rm{Sym}}^q(\mathcal L)$ for homogeneous $v_1,\ldots,v_n$ in $V$ by induction on $n$. It follows that $\delta(\mathcal L)\,\subseteq\,{\rm{Sym}}^q(\mathcal L)$. The desired lemma now follows from the fact that for $\beta_1,\ldots,\beta_p\,\in\,\mathcal L$,
$$ \delta[s(\beta_1,\ldots,\beta_p)]\,=\, \sum_{i=1}^p \pm s(\beta_1,\ldots, \delta(\beta_i),\ldots,\beta_p) \,\text{.}$$
Indeed, the symmetrization of $p-1$ elements of $\mathcal L$ with an element of ${\rm{Sym}}^q(\mathcal L)$ is in $\bigoplus_{r \leq p+q-1} {\rm{Sym}}^{r}(\mathcal L)$ if $q>1$ and in ${\rm{Sym}}^{p+q-1}(\mathcal L)$ if $q=0,1$.
\end{proof}
Identify $\Omega^1R_\n$ with $R \otimes V$ as in~\eqref{omeganat}. Let $\bar{\partial}\,:\, \bar{R} \xrightarrow{} \Omega^1R_{\n}$ be as in~\eqref{cyclicderham}.
\begin{lemma} \la{cycderhamhodge}
$\bar{\partial} [{\rm{Sym}}^p(\mathcal L)] \,\subseteq\, {\rm{Sym}}^{p-1}(\mathcal L) \otimes V$.
\end{lemma}
\begin{proof}
Let $\bar{\partial}'\,:\, \bar{R} \xrightarrow{} V \otimes \bar{R}$ be the operator
$$ (v_1,\ldots, v_n) \mapsto \sum_{i=1}^n (-1)^{(|v_1|+\ldots+|v_{i-1}|)(|v_i|+\ldots+|v_n|)} v_i \otimes (v_{i+1} ,\ldots, v_n,v_1, \ldots ,v_{i-1}) \,\text{.}$$
Note that there is a {\it right} action of $S_n$ on $V^{\otimes n}$. Explicitly, for $\sigma\,\in\,S_n$ and for $v_1, \ldots, v_n$ homogeneous, $(v_1 \ldots v_n) \cdot \sigma \,:=\, \pm (v_{\sigma(1)} \ldots v_{\sigma(n)})$, where the sign is determined by the Koszul sign rule. Let $\tau$ be the $n$-cycle $(1 2 \ldots n)$. Then, the restriction of $\bar{\partial}'$ to $V^{\otimes n}$ is given by the composite map
$$ \begin{diagram} V^{\otimes n} & \rTo^{(\mbox{--}) \cdot N} & V^{\otimes n} & \rTo & V \otimes V^{\otimes n-1} \end{diagram}\,,$$
where $N\,=\, \sum_{i=0}^{n-1} \tau^i$ and where the last arrow is the obvious isomorphism that permutes no factors. Further note that the above right action of $S_n$ on $V^{\otimes n}$ is dual to the left action of $S_n$ on $W^{\otimes n}$ used in~\cite[Section 4.5]{L} (with $W\,:=\,V^{\ast}$).
Let $S_{n,p}$ be the set of permutations in $S_n$ having $p-1$ descents in the sense of~\cite[Section 4.5.5]{L}. Let $l^P_n\,:=\, \sum_{\sigma\,\in\,S_{n,p}} \sigma$ and let $e^{(p)}_n$ be the Eulerian idempotent
$$e^{(p)}_n \,=\, \sum_{j=1}^n a^{p,j}_n l^j_n\,, $$
where the Stirling numbers $a^{p,j}_n$ are defined by the identity $\sum_{p=1}^n a^{p,j}_n X^p\,=\, \binom{X-j+n}{n}$.
By~\cite[Remark 2.10]{L2}, the right action of the Eulerian idempotent $e^{(p)}_n$ on $V^{\otimes n}$ is the projection from $V^{\otimes n}$ to ${\rm{Sym}}^p(\mathcal L) \cap V^{\otimes n}$. Let $S_{n-1}$ be viewed as the subgroup of $S_n$ fixing $1$. By the proof of~\cite[Theorem 4.6.6]{L} (more specifically, formula $(4.6.6.2)$ in {\it loc. cit.}),
$e^{(p)}_nN\,=\, Ne^{(p-1)}_{n-1}$. Hence, for any $\alpha\,\in\, {\rm{Sym}}^p(\mathcal L) \cap V^{\otimes n}$,
\begin{eqnarray*}
\alpha \cdot N \,=\, (\alpha \cdot e^{(p)}_n) \cdot N\,=\, \alpha \cdot (e^{(p)}_nN) \,=\, \alpha \cdot (Ne^{(p-1)}_{n-1}) \,=\, (\alpha \cdot N) \cdot e^{(p-1)}_{n-1}\,\text{.}
\end{eqnarray*}
It follows that $(\alpha \cdot N) \cdot e^{(i)}_{n-1}\,=\, \delta_{i,p-1} \alpha \cdot N$, where $\delta_{i,j}$ is the Kronecker delta. Thus,
$$\bar{\partial}' [{\rm{Sym}}^p(\mathcal L) \cap V^{\otimes n}] \,\subseteq\, V \otimes ({\rm{Sym}}^{p-1}(\mathcal L) \cap V^{\otimes n-1})\,\text{.} $$
The desired lemma follows once we observe that $\bar{\partial}$ is given by composing $\bar{\partial}'$ with the isomorphism $V \otimes R \,\cong\, {R} \otimes V$ that swaps factors.
\end{proof}
\subsection*{Acknowledgements}{\footnotesize
We would like to thank Mike Mandell for interesting discussions and suggestions.
Research of the first two authors was partially supported by the Simons Foundation Collaboration Grant 066274-00002B (`Representation Homology').}
|
\section{Introduction}
\label{intro}
Charge-independence breaking (CIB) is well established in the two-nucleon
(2N) system in the $^1S_0$ state as evidenced by the values of the scattering
lengths $-23.75 \pm 0.01$, $-17.3 \pm 0.8$, and $-18.5 \pm 0.3$~ fm
\cite{schori87,teramond87} for
the neutron-proton (np), proton-proton (pp)
(with the Coulomb force subtracted), and neutron-neutron (nn) systems,
respectively.
That knowledge of CIB is
incorporated into modern, high precision NN potentials, as exemplified
by the standard semi-phenomenological models:
Av18 \cite{av18}, CD~Bonn \cite{cdbonn}, or NijmI and
NijmII \cite{nijm}, as well as by the chiral
NN forces~\cite{epel6a,Epelbaum:2008ga,machl6b}. Treating neutrons and
protons as identical
particles requires
that nuclear systems are described not only in terms of the momentum and spin
but also isospin states.
The general classification of the isospin dependence of the NN force is
given in \cite{henley1979}.
The isospin violating 2N forces induce an admixture of
the total isospin $T={3/2}$ state
to the dominant $T={1/2}$ state in the three-nucleon (3N) system.
The CIB of the NN interaction thus affects 3N
observables. The detailed treatment of the 3N system with CIB NN forces
in the case of distinguishable or identical particles was formulated
and described
in \cite{wit91-cib}. We extend the investigation done in \cite{wit91-cib}
by including a three-nucleon force (3NF). In the calculations performed with
the standard semi-phenomenological potentials we use
the UrbanaIX (UIX) \cite{uIX} 3NF, while
the chiral N$^2$LO 3N force~\cite{epel2002} is used in addition to
the recent and most accurate chiral NN interactions \cite{epel1,epel2}.
In this paper, based on such dynamics, we discuss the role of the amplitude
with the
total three-nucleon (3N) isospin $T=3/2$ in elastic neutron-deuteron (nd)
scattering and in the corresponding breakup reaction.
In Sec.~\ref{section2} we briefly describe the formalism of
3N continuum Faddeev
calculations and the inclusion of CIB.
The results are presented in Sec.~ \ref{res}.
In Sec.~\ref{elastic} we discuss
our results for elastic
nd scattering and in Sec.~ \ref{breakup} describe our findings
for selected breakup configurations.
We summarize and conclude in Sec. ~\ref{summary}.
\section{3N scattering and charge independence breaking}
\label{section2}
Neutron-deuteron scattering with nucleons interacting
through a NN interaction $v_{NN}$ and a 3NF $V_{123}$, is
described in terms of a breakup operator $T$ satisfying the
Faddeev-type integral equation~\cite{wit88,glo96,hub97}
\begin{eqnarray}
T\vert \phi \rangle &=& t P \vert \phi \rangle +
(1+tG_0)V^{(1)}(1+P)\vert \phi \rangle + t P G_0 T \vert \phi \rangle \cr
&+&
(1+tG_0)V^{(1)}(1+P)G_0T \vert \phi \rangle \, .
\label{eq1a}
\end{eqnarray}
The two-nucleon $t$-matrix $t$ is the solution of the
Lippmann-Schwinger equation with the interaction
$v_{NN}$. $V^{(1)}$ is the part of a 3NF which is
symmetric under the interchange of nucleons $2$ and $3$: $V_{123}=V^{(1)}(1+P)$.
The permutation operator $P=P_{12}P_{23} +
P_{13}P_{23}$ is given in terms of the transposition operators,
$P_{ij}$, which interchange nucleons $i$ and $j$. The incoming state
$\vert \phi \rangle = \vert \mathbf{q}_0 \rangle \vert \phi_d \rangle$
describes the free relative motion of the neutron and the deuteron
with the relative momentum
$\mathbf{q}_0$ and contains the internal deuteron state $\vert \phi_d \rangle$.
Finally, $G_0$ is the resolvent of the three-body center-of-mass kinetic
energy.
The amplitude for elastic scattering leading to the corresponding
two-body final state $\vert \phi ' \rangle$ is then given by~\cite{glo96,hub97}
\begin{eqnarray}
\langle \phi' \vert U \vert \phi \rangle &=& \langle \phi'
\vert PG_0^{-1} \vert
\phi \rangle +
\langle \phi' \vert PT \vert \phi \rangle + \langle
\phi'\vert V^{(1)}(1+P)\vert \phi \rangle \cr
&+& \langle \phi' \vert V^{(1)}(1+P)G_0T\vert \phi \rangle,
\label{eq3}
\end{eqnarray}
while for the breakup reaction one has
\begin{eqnarray}
\langle \phi_0'\vert U_0 \vert \phi \rangle &=&\langle
\phi_0'\vert (1 + P)T\vert
\phi \rangle ,
\label{eq3_br}
\end{eqnarray}
where $\vert \phi_0' \rangle$ is the free three-body breakup channel state.
Solving Eq.(\ref{eq1a}) in the momentum-space partial wave basis, defined by
the magnitudes of the
3N Jacobi momenta $p$ and $q$ together with the angular momenta and isospin
quantum numbers $\alpha$ ($\beta$),
is performed by projecting Eq.~(\ref{eq1a}) onto two types of
basis states:
\begin{eqnarray}
\vert p q \alpha \rangle \equiv \vert p q ~{\rm angular~momenta} \rangle
\vert (t {1\over{2}})T={1\over {2}} M_T \rangle \, , ~ (t=0,1) \, ,
\label{eq4}
\end{eqnarray}
and
\begin{eqnarray}
\vert p q \beta \rangle \equiv \vert p q ~{\rm angular~momenta} \rangle
\vert (t {1\over{2}})T={3\over {2}} M_T \rangle \, , ~ (t=1) \, .
\label{eq5}
\end{eqnarray}
Assuming charge conservation and employing the notation
where the neutron (proton) isospin projection is
${1\over{2}}$ ($-{1\over{2}}$), the 2N t-operator in
the three-particle isospin
space can be decomposed for the nd system as~\cite{wit91-cib}:
\begin{eqnarray}
\langle (t{1\over{2}}) T M_T={1\over{2}} \vert t \vert
(t'{1\over{2}}) T' M_{T'}={1\over{2}} \rangle &=&
\delta_{tt'} \delta_{TT'}\delta_{T1/2} [ \delta_{t0} t^{t=0}_{np} +
\delta_{t1} ( {2 \over{3}} t^{t=1}_{nn} + {1 \over{3}} t^{t=1}_{np} ) ] \cr
&+& \delta_{tt'} \delta_{t1} (1-\delta_{TT'})
{\sqrt{2} \over {3}} ( t^{t=1}_{nn} - t^{t=1}_{np} ) \cr
&+&\delta_{tt'} \delta_{t1} \delta_{TT'}\delta_{T3/2}
( {1 \over{3}} t^{t=1}_{nn} + {2 \over{3}} t^{t=1}_{np} ) ~,
\label{eq6}
\end{eqnarray}
where $t_{nn}$ and $t_{np}$ are solutions of the Lippman-Schwinger equations
driven by the $v_{nn}$ and $v_{np}$ potentials, respectively.
As a result of solving Eq.~(\ref{eq1a}) one gets the amplitudes
$\langle p q \alpha \vert T \vert \phi \rangle$ and
$\langle p q \beta \vert T \vert \phi \rangle$,
which fulfill the following set of coupled integral equations:
\begin{eqnarray}
\langle p q \alpha \vert T\vert \phi \rangle &=&
\sum_{\alpha'} \int_{p'q'} \langle p q \alpha \vert t \vert p' q' \alpha'\rangle
\langle p' q' \alpha' \vert P \vert \phi \rangle \cr
&+&
\sum_{\alpha'} \int_{p'q'} \langle p q \alpha \vert V^{(1)}\vert p' q'
\alpha'\rangle
\langle p' q' \alpha' \vert (1+P)\vert \phi \rangle \cr
&+&
\sum_{\alpha'}\int_{p'q'} \langle p q \alpha \vert t \vert p' q' \alpha'\rangle
\langle p' q' \alpha' \vert G_0 V^{(1)} (1+P)\vert \phi \rangle \cr
&+&
\sum_{\alpha'}\int_{p'q'} \langle pq \alpha \vert t \vert p' q' \alpha'\rangle
\langle p' q' \alpha' \vert P G_0 T \vert \phi \rangle \cr
&+&
\sum_{\beta'} \int_{p'q'} \langle pq \alpha \vert t \vert p' q' \beta'\rangle
\langle p' q' \beta' \vert P G_0 T \vert \phi \rangle \cr
&+&
\sum_{\alpha'} \int_{p'q'} \langle p q \alpha \vert V^{(1)}\vert p' q'
\alpha'\rangle
\langle p' q' \alpha' \vert (1+P)G_0T \vert \phi \rangle \cr
&+&
\sum_{\alpha'} \int_{p'q'} \sum_{\alpha''} \int_{p''q''}
\langle p q \alpha \vert t \vert p' q' \alpha'\rangle
\langle p' q' \alpha' \vert G_0V^{(1)} \vert p''q''\alpha''\rangle \cr
&\times& \langle p''q''\alpha''\vert (1+P)G_0T \vert \phi \rangle \cr
&+&
\sum_{\beta'} \int_{p'q'} \sum_{\beta''} \int_{p''q''}
\langle p q \alpha \vert t \vert p' q' \beta'\rangle
\langle p' q' \beta' \vert G_0V^{(1)} \vert p''q''\beta''\rangle \cr
&\times& \langle p''q''\beta''\vert (1+P)G_0T \vert \phi \rangle \cr
\langle p q \beta \vert T\vert \phi \rangle &=&
\sum_{\alpha'} \int_{p'q'} \langle p q \beta \vert t \vert p' q' \alpha'\rangle
\langle p' q' \alpha' \vert P \vert \phi \rangle \cr
&+&
\sum_{\alpha'} \int_{p'q'} \langle p q \beta \vert t \vert p' q' \alpha'\rangle
\langle p' q' \alpha' \vert G_0 V^{(1)} (1+P)\vert \phi \rangle \cr
&+&
\sum_{\alpha'} \int_{p'q'} \langle p q \beta \vert t \vert p' q' \alpha'\rangle
\langle p' q' \alpha' \vert P G_0 T \vert \phi \rangle \cr
&+&
\sum_{\beta'} \int_{p'q'} \langle p q \beta \vert t \vert p' q' \beta'\rangle
\langle p' q' \beta' \vert P G_0 T \vert \phi \rangle \cr
&+&
\sum_{\beta'} \int_{p'q'} \langle p q \beta \vert V^{(1)}\vert p' q' \beta'\rangle
\langle p' q' \beta' \vert (1+P)G_0T \vert \phi \rangle \cr
&+&
\sum_{\alpha'} \int_{p'q'} \sum_{\alpha''} \int_{p''q''}
\langle p q \beta \vert t \vert p' q' \alpha'\rangle
\langle p' q' \alpha' \vert G_0V^{(1)} \vert p'' q''\alpha''\rangle \cr
&\times& \langle p'' q''\alpha''\vert
(1+P)G_0T \vert \phi \rangle \cr
&+&
\sum_{\beta'} \int_{p'q'} \sum_{\beta''} \int_{p''q''}
\langle p q \beta \vert t \vert p' q' \beta'\rangle
\langle p' q' \beta' \vert G_0V^{(1)} \vert p''q''\beta''\rangle \cr
&\times& \langle p'' q''\beta''\vert
(1+P) G_0T \vert \phi \rangle ~.
\label{eq7}
\end{eqnarray}
The form of the couplings in Eq.~(\ref{eq7}) follows from the fact that
the incoming neutron-deuteron state $\vert \phi \rangle$ is a
total isospin $T=1/2$ state, the permutation operator $P$ is diagonal
in the total isospin,
and the 3NF is assumed to conserve the total isospin $T$~\cite{epel2005}.
From Eq.~(\ref{eq7}) it is clear that only when the nn and np
interactions differ in the same orbital and spin angular-momentum states
with the 2N
subsystem isospin $t=1$ (CIB), then the amplitudes
$\langle p q \beta \vert T \vert \phi \rangle$ will be nonzero.
In addition, their magnitude is driven by the strength of the CIB
as given by the difference of the corresponding t-matrices
${\sqrt{2} \over {3}} ( t^{t=1}_{nn} - t^{t=1}_{np} )$ in Eq.~(\ref{eq6}).
In such a case, not only the magnitude of CIB decides about the importance of
the $\langle p q \beta \vert T \vert \phi \rangle$ contributions, but also the
isospin $T=3/2$ 3NF matrix elements, which participate in generating
the $\langle p q \beta \vert T \vert \phi \rangle$ amplitudes.
It is the set of equations Eq.~(\ref{eq7}) which we solve when we differentiate
between nn and np interactions and include both $T=1/2$ and $T=3/2$ 3NF
matrix elements.
In the case when the neglect of the $T=3/2$ amplitudes
$\langle p q \beta \vert T \vert \phi \rangle$ is
justified, the CIB can be taken care of by using
the effective two-body t-matrix generated with the
so-called $``2/3-1/3"$ rule,
$t_{\rm eff}={2 \over{3}} t^{t=1}_{nn} + {1 \over{3}} t^{t=1}_{np}$,
see Eq.(\ref{eq6}),
in the 2N subsystem isospin $t=1$ states
and restricting the treatment only to the amplitudes
$\langle p q \alpha \vert T \vert \phi \rangle$.
Since the final state $\vert \phi' \rangle$ in elastic nd scattering also
has the
total 3N isospin $T=1/2$, the amplitudes
$\langle p q \beta \vert T \vert \phi \rangle$ do
not contribute directly to this reaction. The $T=3/2$ admixture enters
in this case through a modification of the $T=1/2$ amplitudes
$\langle p q \alpha \vert T \vert \phi \rangle$ induced by the couplings given
in Eq.~(\ref{eq7}).
Contrary to that, for the nd breakup reaction both $T=1/2$
($\langle p q \alpha \vert T \vert \phi \rangle$) and $T=3/2$
($\langle p q \beta \vert T \vert \phi \rangle$) amplitudes contribute.
\section{Results}
\label{res}
In order to check the importance of the isospin $T=3/2$ contributions we solved
the 3N Faddeev equations Eq.~(\ref{eq7}) for four values of the
incoming neutron
laboratory energy: $E_{lab}=13$, $65$, $135$, and $250$~MeV.
As a NN potential we took the semi-locally regularized
N$^4$LO chiral potential of ref.~\cite{epel1,epel2,binder2016} with
the regulator $R=0.9$~fm, alone or
combined with the chiral
N$^2$LO 3NF~\cite{epel2002,heb1}, regularized with
the same regulator. We additionally regularized
matrix elements of that 3NF by multiplying it with a nonlocal regulator
$f(p,q)=\exp\{-(p^2+{3\over{4}}q^2)^3/{\Lambda}^6\}$
with large cut-off value $\Lambda=1000$~MeV.
This additional regulator is applied to the 3NF matrix elements only
for technical reasons. The practical calculation of the local 3NFs involve
the evaluation of convolution integrals whose calculation becomes
numerically unstable at very large momenta. The value of the cutoff scale
$\Lambda$ is chosen sufficiently large so that low-energy physics is not
affected by this additional regulator. In fact, we have checked
explicitly that the effects of this regulator for the chosen cutoff
value are negligible in three-body bound state and scattering calculations.
As a nn force we took the pp
version of that particular NN interaction (with the Coulomb force subtracted).
The low-energy constants
of the contact interactions in that 3NF were adjusted to the
triton binding energy
and we used $c_D=6.0$ for the one-pion exchange contribution
and $c_E=-1.0943$ for the 3N contact term (we are using the notation of
Ref.~\cite{epel2002} with $\Lambda=700$~MeV). That specific choice of
the $c_D$ and $c_E$ values does not only reproduce the experimental
triton binding energy when that N$^4$LO NN and N$^2$LO 3NF are combined
but also provides quite a good description of the nucleon-deuteron
elastic scattering
cross section data at higher energies. In order to provide
convergent predictions we solved Eq.~(\ref{eq7})
taking into account all partial wave states with the total 2N angular
momenta up to $j_{max}=5$ and 3N total angular momenta up to $J_{max}=25/2$.
The 3NF was included up to $J_{max}=7/2$.
We present results for that particular combination of the chiral
NN and 3N forces (of course, in the future a more consistent set of 2N
and 3N forces
needs to be employed, once the corresponding 3NFs are available) .
However, we checked using the example of
the Av18 and Urbana IX 3NF combination that the conclusions remain
unchanged when instead of the chiral forces so-called high
precision realistic forces are used.
Since CIB effects are driven by the difference between np and nn t-matrices we
display in Figs.~\ref{fig1} and \ref{fig2}
for the $^1S_0$ and $^3P_0$ NN partial waves, respectively, the np t-matrix
$t_{np}(p,p';E-{3\over{4}}q^2)$ (a) and
the difference $t_{np}(p,p';E-{3\over{4}}q^2)-t_{nn}(p,p';E-{3\over{4}}q^2)$ (b),
at the laboratory energy of the
incoming neutron $E_{lab}=13$~MeV. They are displayed as a function of
the NN relative momenta
$p$ and $p'$ for a chosen value of the spectator nucleon momentum
$q=0.528$~fm$^{-1}$ at which the 2N subsystem energy is equal to the
binding energy of the deuteron $E_d$: $E-{3\over{4}}q^2=E_d$.
The behaviour of the $t_{np}$
as well as of the difference $t_{np}-t_{nn}$ is similar at other energies.
It is interesting to note that the difference between the np and nn
t-matrices ranges up to $\approx 10~\%$.
In Tables~\ref{table1},~\ref{table2},~and~\ref{table3} we show at the chosen
energies the total cross section for the nd interaction, the total
nd elastic scattering cross section, and the total nd breakup
cross section, respectively, calculated with different underlying dynamics
based on the chiral N$^4$LO NN or/and N$^2$LO 3NF force.
Namely, the results in column~2 of those tables (no CIB,
$V_{123}=0$, $^1S_0$ np)
were obtained with the NN force only, assuming no CIB and using in all $t=1$
partial waves the effective t-matrix $t_{\rm eff}=(2/3)t_{nn}+(1/3)t_{np}$, with
the exception of the $^1S_0$ partial wave, where only the np force was taken.
In column~3 instead of the np $^1S_0$ NN force a nn one was taken
(no CIB, $V_{123}=0$, $^1S_0$ nn). In column~4 in all $t=1$ partial waves
(including also $^1S_0$ one) only $T=1/2$ was taken into account and
the effective t-matrix $t_{\rm eff}=(2/3)t_{nn}+(1/3)t_{np}$
was used (no CIB, $V_{123}=0$, $t_{\rm eff}$). In column~5 the NN interaction of
column~4 was combined with 3NF (no CIB, $V_{123}$, $t_{\rm eff}$).
In column~6 a proper treatment of the CIB in the $^1S_0$ partial wave
was performed by taking in that partial wave both np and nn interactions
and keeping in addition to the total isospin $T=1/2$ also $T=3/2$.
In all other $t=1$ states the effective t-matrix $t_{\rm eff}$ was used and
only states with $T=1/2$ were kept. No 3NF was
allowed ($^1S_0$ CIB, $V_{123}=0$).
Results, when in addition also the 3NF was active, are shown in
column~7 ($^1S_0$ CIB, $V_{123}$). The proper treatment of CIB in all states
with $t=1$, when both np and nn interactions were used and
both $T=1/2$ and $T=3/2$ states were kept, are shown in column~8 and 9
for the cases when NN interactions were used alone (CIB, $V_{123}=0$) and
combined with 3NF (CIB, $V_{123}$), respectively.
From Tables~\ref{table1},~\ref{table2},~and~\ref{table3} it is clear that
it is sufficient to use the effective t-matrix $t_{\rm eff}$ and to
neglect $T=3/2$
states completely to account exactly
for CIB effects in all three total cross sections. That is true in both cases,
when the 3NF is present or absent. In the case when the 3NF is not
included, the total cross
sections for the nd interaction, for elastic scattering, and for breakup,
depend
slightly on the $^1S_0$ t-matrix used in the calculations. Changing it
from $t_{np}$ to $t_{nn}$ leads to differences of the cross section values
up to $\approx 2\%$ (columns~2 and 4). Using in the $^1S_0$ channel the
effective t-matrix $t_{\rm eff}=(2/3)t_{nn}+(1/3)t_{np}$ accounts for all CIB
effects exactly, without the necessity to introduce the total
isospin $T=3/2$ components
in any of the $t=1$ partial wave states. Namely, the exact treatment
of CIB by using in all
$t=1$ states the $t_{np}$ and $t_{nn}$ t-matrices and both $T=1/2$ and $T=3/2$
partial wave states (column~8) gives the same value for all three total
cross sections. Also restricting the exact treatment of CIB to
the $^1S_0$ state
only (column~6) provides the same values for the total cross sections. It shows
that contribution of $T=3/2$ states, as far as the total cross sections are
concerned, can be neglected and all CIB effects properly taken into
account by restricting to the total isospin $T=1/2$ states only and using in
all $t=1$ channels the effective t-matrix generated according
to the $``2/3-1/3"$ rule.
The same is true when 3NF is included. In this case,
in all three total cross sections,
clear effects increasing with energy are seen. But again, all three
treatments of CIB yield the same numbers (columns 5, 7, and 9). Since in
the cases when $T=3/2$ states were included (columns 7 and 9) also the
corresponding $T=3/2$ matrix elements of a 3NF were used, we conclude that
their influence on the total cross sections is negligible.
\subsection{Elastic scattering}
\label{elastic}
In Figs.~\ref{fig3} and \ref{fig4} we display results for the nd elastic
scattering
angular distributions obtained using different assumptions about the
underlying
dynamics and treatment of CIB. As for the total cross sections, in case when
the 3NF is inactive, the three treatments of the CIB, namely
using $t_{\rm eff}$ and no
$T=3/2$ states, $T=3/2$ in the $^1S_0$ state with the $t_{np}$ and $t_{nn}$
t-matrices in that state, and $T=3/2$ in all $t=1$ states with
the corresponding np and nn t-matrices, provide the same elastic
scattering cross sections (dashed (cyan), dashed-dotted (magenta), and
dotted (maroon) lines, respectively). These lines overlap in
Figs.~\ref{fig3} and \ref{fig4} and the results are displayed in more detail
in b), c), and d) for particular ranges of angles).
We checked that also for
all nd elastic scattering spin observables, encompassing the neutron (vector)
and deuteron (vector and tensor) analyzing powers, the spin correlation
as well as spin transfer coefficients, the above three approaches lead to
the same results. Thus again the contribution of all $T=3/2$ states can
be neglected and CIB in elastic nd scattering treated exactly by
restricting only to $T=1/2$ states and using in
all $t=1$ states the $``2/3-1/3"$ rule to generate from $t_{np}$ and $t_{nn}$
the effective $t_{\rm eff}$ t-matrix.
Restricting to $t_{np}$ or $t_{nn}$ t-matrices in the $^1S_0$ channel and
neglecting all $T=3/2$ states changes the
elastic scattering cross sections and spin observables by up to
$\approx 1\%$ (see the solid (blue) and dashed (red) lines, respectively,
in Figs.~\ref{fig3} and \ref{fig4}).
Adding the 3NF changes the elastic scattering cross section. The 3NF effects
grow with the projectile energy and are especially large in the region of
intermediate and backward angles. But again the three approaches to CIB
provide the same
cross sections and spin observables (they are displayed for cross sections in
Figs.~\ref{fig3} and \ref{fig4} by overlapping lines: dotted (black),
dashed-double-dotted (green), and dotted-double-dashed (magenta)). That again
supports the conclusion that $T=3/2$ states can be neglected together with
the $T=3/2$ 3NF matrix elements for all nd elastic scattering observables.
\subsection{Breakup reaction}
\label{breakup}
In the final breakup state of three free nucleons both $T=1/2$ and $T=3/2$
total isospin components are allowed. Thus one would expect that here the
influence of the $T=3/2$ components will be better visible than in elastic
scattering.
We show in Figs.~\ref{fig5}-\ref{fig8} that this indeed is the case for the
example of
three kinematically complete breakup configurations: final-state-interaction
(FSI), quasi-free-scattering (QFS), and symmetrical-space-star (SST).
In the FSI configuration under the exact FSI condition, the two outgoing
nucleons have
equal momenta. Their strong interaction in the $^1S_0$ state leads to a
characteristic cross section maximum occurring at the exact FSI condition,
the magnitude of which is sensitive to the $^1S_0$ scattering length.
Since largest CIB effects are seen in the difference between np and nn (pp)
$^1S_0$ scattering lengths, the region of the FSI peak should reveal
largest CIB effects. That is clearly demonstrated in
Fig.~\ref{fig5}a for the nn and in Fig.~\ref{fig5}b for the
np FSI configuration, where
the solid (blue) and dashed (red) lines display FSI cross sections obtained
with the $^1S_0$ np and nn t-matrices, respectively, using in all other $t=1$
states the effective t-matrix $t_{\rm eff}$. Only $T=1/2$ states were used
and the 3NF was omitted. Using also in the $^1S_0$ state the effective
t-matrix leads to the dashed (cyan) line, which changes to the
dotted (black) line
when, keeping the rest unchanged, the 3NF is also included. Most
interesting is the effect of treating the CIB exactly in the state $^1S_0$ by
including the $T=3/2$ component and using both $t_{np}$ and $t_{nn}$ t-matrices:
dashed-dotted (magenta) and dashed-double-dotted (green) lines in case when
3NF is omitted and included, respectively. As expected, the inclusion
of the isospin
$T=3/2$ in the $^1S_0$ state brings the predictions, in the case when 3NF
is omitted, close to the nn prediction
(dashed (red) line in Fig.~\ref{fig5}a) for the nn FSI,
and close to the np prediction
(solid (blue) line in Fig.~\ref{fig5}b) for the np FSI. However, a significant
difference exists between that result and the pure nn or np ones as well as
when compared to
results obtained with the effective t-matrix $t_{\rm eff}$ (dashed (cyan) line in
Fig.~\ref{fig5}a and \ref{fig5}b). This shows that the proper treatment of CIB
in the FSI configurations of the nd breakup requires the inclusion of the
total isospin
$T=3/2$ component and using both $t_{np}$ and $t_{nn}$
t-matrices in the $^1S_0$ state.
This is also sufficient for the exact treatment of CIB as shown
by the results of the full CIB treatment, where in all $t=1$ partial waves also
isospin $T=3/2$ states are taken into account and
corresponding $t_{np}$ and $t_{nn}$ t-matrices are used, as shown
in Figs.~\ref{fig5}a and \ref{fig5}b
by the dotted (maroon) line for the NN
interaction acting alone and dotted-double-dashed
(magenta) line when combined with the 3NF, respectively.
These lines overlap with the
lines corresponding to the case when $T=3/2$ is included for the
state $^1S_0$. It is interesting to note that the $T=3/2$ component in the
$^1S_0$ state is important and provides FSI cross sections which are different
from the results obtained with particular NN $^1S_0$ interactions only
(np for np FSI and nn for nn FSI). It proves the importance of
including $T=3/2$ in the $^1S_0$ state and shows that both np and nn
interactions
have to be employed when
the FSI peaks are analyzed to extract the value of the corresponding
scattering length.
In order to see how the magnitude of the effects induced by
the $T=3/2$ $^1S_0$ component depends on
the particular FSI configuration, we present in Fig.~\ref{fig6} the
cross section in the maximum of the FSI peak
as a function of the laboratory production
angle of the final-state
interacting pair. Again it is clearly seen that restricting to $t_{\rm eff}$ only
and neglecting $T=3/2$ components is insufficient to include all CIB effects.
Inclusion of $T=3/2$ component only in the $^1S_0$ state is, however,
sufficient to
fully account for the CIB effects. The importance of that component
depends on the production angle. At the angles in the region around
$\approx 45^{o}$ the contribution of that component is tiny but becomes
significant at smaller and larger production angles.
For the QFS and SST configurations the picture is similar. Again in order
to fully
account for the CIB effects it is necessary and sufficient to include the total
isospin $T=3/2$ component in the $^1S_0$ state.
We exemplify this in Fig.~\ref{fig7} for the nn QFS and in Fig.~\ref{fig8}
for the SST
configurations. Again there is an angle around which the contribution
of that component is minimized. For the nn QFS it occurs around
$\theta^{lab}_1 \approx 28^o$ and for the SST around
$\theta^{c.m.}_{plane} \approx 90^o$.
\section{Summary}
\label{summary}
We investigated the importance of the scattering amplitude components
with the total 3N isospin $T=3/2$
in two 3N reactions. The inclusion of these components is required to
account for
CIB effects of the NN interaction. The difference between np and
nn (pp) forces leads to a situation in which also the matrix elements of the
3NF between $T=3/2$
states contribute to the considered 3N reactions. The modern NN
interactions, which describe existing pp and np data
with high precision, provide pp and np t-matrices
which differ up to $\approx 10 \%$. Such a magnitude of CIB requires
that the isospin $T=3/2$ components are included
in the calculation of the breakup reaction, especially for the
regions of the breakup phase-space close to the FSI condition.
However, in order to account for
all CIB effects it is sufficient to restrict the inclusion
of $T=3/2$ to the $^1S_0$
partial wave state
only instead of doing it in all $t=1$ states. For elastic scattering we found
that the $T=3/2$ components can be neglected completely and all CIB effects are
accounted for by restricting oneself to total 3N isospin $T=1/2$ partial
waves only and
using the effective t-matrix generated with the $``2/3-1/3''$ rule
$t_{\rm eff}=(2/3)t_{nn}+(1/3)t_{np}$.
These results allow one to reduce significantly the number of partial waves
in time-consuming 3N
calculations. This is of particular importance in view of the necessity
to fix the parameters of the higher-order chiral 3NF components
by fitting them to 3N scattering observables.
The presented results show that in 3N reactions the $T=3/2$ components
are overshadowed by the dominant
$T=1/2$ contributions.
It will be interesting to investigate reactions with three nucleons
in which only $T=3/2$ components contribute in the final state such as e.g.
$^3H + \pi^- \to n + n + n$. That will allow one to study the properties and
the importance of 3NFs in the $T=3/2$ states.
\begin{acknowledgements}
This work was performed by the LENPIC collaboration with support from
the Polish National Science Center
under Grant No. DEC-2013/10/M/ST2/00420 and
PRELUDIUM DEC-2013/11/N/ST2/03733, BMBF (contract
No. 05P2015 - NUSTAR R$\&$D), and ERC Grant No. 307986 STRONGINT.
The numerical calculations
have been performed on the supercomputer cluster of the JSC, J\"ulich,
Germany.
\end{acknowledgements}
|
\section{Introduction}
R Coronae Borealis (RCB) stars are defined by their light curves displaying sudden and large drops in brightness with slower recoveries to the baseline level, with these events occurring randomly in time. These spectacular dips are caused when the star forms dense dust clouds on the line of sight to Earth hiding the star. The RCB stars are all hydrogen deficient supergiants, with various abundance anomalies, including enriched nitrogen and carbon. RCB stars are rare, with only 76 known in our Milky Way. The evolutionary status of RCB stars is that they might be from recent coalescences of a double white dwarf binary or from a final helium shell flash in a born-again asymptotic giant branch (AGB) star. Clayton (1996; 2012) presents full reviews of RCB stars.
Most RCB stars are relatively cool, with surface temperatures of 5,000-7,000 K. However, four of the known RCB stars have a greatly hotter surface temperature, from 15,000-25,000 K, and these are called the `hot RCB stars'. The four known hot RCB stars are V348 Sgr, MV Sgr, and DY Cen in our Milky Way plus HV 2671 in the Large Magellanic Cloud. DY Cen and MV Sgr are hydrogen deficient (just like the cool RCB stars) and mostly composed of helium, while both have relatively infrequent drops in brightness. V348 Sgr and HV 2671 are very carbon-rich (55\% carbon, most of the rest helium), and elemental abundances like in central stars of planetary nebulae, while both have frequent episodes of brightness declines. A plausible idea is that the two greatly different composition indicate two formation mechanisms, with DY Cen and MV Sgr simply being the `progeny' of the normal RCB stars as they heat up, and with V348 Sgr and HV 2671 being somehow formed during a final helium-shell flash on post-AGB stars. Thus, the birth mechanism of the hot RCB stars could be either from born-again systems or white dwarf mergers. Nevertheless, it is still unclear as to the relationship between the hot RCB stars and the cool RCB stars, as well as to other classes of stars (the born-again stars and the Wolf-Rayet central stars of planetary nebulae). De Marco et al. (2002) present a full review of the hot RCB stars.
The key high-level science question is to understand the evolutionary state of the RCB stars, both hot and cold. For this, a critical piece of evidence is to watch them evolve in real time. Like for the born-again post-AGB stars (i.e., V605 Aql, Sakurai's object, and FG Sge), substantial movement across the HR diagram might be seen on time scales of a decade and a century. For this, De Marco et al. (2002) have pointed out that the baseline levels for MV Sgr, V348 Sgr, and DY Cen are apparently fading over the last century, and this would only be from movement from right-to-left across the top of the HR diagram.
These century-long light curves have inevitable big problems for two reasons. The light curves were compiled from magnitudes in the $V$, $m_{vis}$, and $m_{pg}$ systems, which makes for large systematic uncertainties between old and new magnitudes, with exactly this able to make apparent systematic declines when none are real. De Marco et al. did attempt to correct for such effects. The big effect not mentioned is that all the old photometry is always systematically in error by from 0.1 to $>$1.0 mag simply because the old standard stars had these errors. In the days before photoelectric photometers could reach to cover faint sequences (i.e., before the late 1970s), the standard stars and comparison sequences were all calibrated by photographic transfers from the Harvard-Groningen Selected Areas and the North Polar Sequence. These photographic transfers always had problems, due to effects like reciprocity failure, in that claimed magnitudes were always reported systematically different than what we take on the modern magnitude scales. In general, the errors are small for stars brighter than tenth magnitude and start increasing steeply as the stars get fainter. Thorough studies of errors for old standard fields and comparison stars are given by Sandage (2001) and Patat et al. (1997), while I have many examples of poor old calibrations (e.g., Schaefer 1994; 1995; 1996; 1998) The old `photographic magnitude system' (i.e., $m_{pg}$) is really just a poorly calibrated B magnitude.
All the early magnitudes of hot RCB stars are photographic magnitudes, all from the Harvard plates, all from old measures, and so there are inevitably possibly-large systematic errors in their long-term light curves. It might be that these systematic errors have created the apparent secular fading of the hot RCB stars, or it might be that the errors are such that the claim of secular fading is still correct.
In this paper, I will solve these problems for the old photometry, and answer the question of whether the hot RCB stars are fading or not. To do this, I visited the Harvard College Observatory (HCO) in October 2015, and remeasured many magnitudes from plates dated 1896 to 1989, all on the modern Johnson B magnitude system. For the modern portions of the light curves, I used a variety of sources from the literature and from the {\it American Association of Variable Star Observers} (AAVSO), all in the Johnson B magnitude system. My combined light curves, all in a single uniform system, have a larger time range and many more magnitudes than given in De Marco et al. I have further extended their work by adding the B-band light curve for the fourth and last-known hot RCB star, HV 2671.
\section{Photometry with Archival Plates}
From around 1890 to the late 1970s, a large part of astronomy was from photometry as based on photographic sky pictures. These pictures were recorded on blue-sensitive emulsion attached to one side of a glass plate, with the plate being exposed to star light in a special holder at the focus of the telescope. The developed emulsion had the sky being nearly transparent, while stars were round black points. Photographic emulsions have only a small dynamic range between the sky and saturation, so star images are almost all completely saturated (i.e., black) in the centers, with only a small annulus of grey in the outer tails of the star profile. With this, essentially, the only change in the star image as the magnitude varies is the diameter of the star image. Given the variability from plate-to-plate and the non-linearity of the emulsion, the only way to calibrate the image-diameter-versus-magnitude relation is to use comparison stars nearby on the plate. The procedure is to make some measure of the image diameter for the target star as well as for a sequence of nearby comparison stars with a tight spacing of magnitudes, both brighter and fainter than the target. With this, the magnitude of the target can be measured by interpolation in the image-diameter-versus-magnitude relation as determined for each plate from the comparison sequence.
The image radii vary such that the square of the radii or the logarithm of the radii are linear with the magnitude, depending on the brightness of the star (Schaefer 1981). In general, the full calibration curve (image radius versus magnitude) is nonlinear with either magnitude or flux. This condition violates one of the requirements ingrained in observers with CCDs, because then magnitudes cannot be calculated from any application of the magnitude equation with one comparison star. The long-standing traditional solution is to use a whole sequence of comparison stars, strung out over a wide range of brightnesses, so that the radius-versus-magnitude relation is empirically determined for the brightness of the target star.
The image diameters can be measured with machines called `iris diaphragm photometers' (first developed in the 1930s) and with photoelectric scanners (first developed in the 1970s). From the earliest days, the dominant method for measuring image diameters was simply for a trained observer to make size comparisons between the diameter of the target star and the diameters of nearby comparison stars. The human eye is remarkably accurate at side-by-side comparisons of the sizes of round objects. The procedure is to view the glass plate on a light table through a loupe or a low-power microscope, compare the target's size to the size of nearby stars, and judge by-eye the relative placement of the target star's diameter. To give a specific and typical example, if the target is judged to be halfway between two comparison stars of magnitude 12.2 and 12.6, then the target has a magnitude of 12.4. In practice, an inexperienced worker can produce magnitudes with an accuracy of 0.3 mag or so, while an experienced worker can produce magnitudes with a one-sigma uncertainty of 0.1 mag. For many situations, where magnitudes from many plates can be averaged together, the real accuracy of the light curves can be 0.02 mag or better. For an experienced worker, the by-eye method provides equal or better accuracy as compared to iris diaphragm photometers or scanning. The by-eye method is very simple, cheap, and fast, whereas the instrumented methods are always complex, slow, and costly.
Harvard College Observatory has a collection of roughly 500,000 glass plates recording the entire sky from 1889 to 1989. (There are few plates from 1954 to 1969 due to the notorious `Menzel Gap'.) These are almost all blue-sensitive emulsion on glass plates 8$\times$10 inch in size, stored in paper envelopes, and placed on shelves in time order. The plates were taken with a wide variety of telescope, from essentially camera lenses up to 24-inch apertures, with the plate sizes covering widths from 5$\degr$ to 42$\degr$. The limiting magnitudes for the normal-quality plates vary from B=12 to deeper than B=18. The Harvard plates have 1000-4000 plates covering any given position on the sky. Harvard has about half the existing archival plates in the world, and is nearly the only source for targets in the southern skies.
Historically, from 1890 to 1960, the Harvard plates dominated the world of variable stars for anything fainter than about eleventh magnitude. To take an example of the hot RCB stars, all four were discovered with the Harvard plates, and the only published information of any type from before the 1950s is the Harvard light curves. For many questions of modern astrophysics, light curves with 0.1 mag or 0.02 mag accuracy are more than adequate, so the accuracy attainable with CCDs is completely irrelevant. In the world of variable stars, the stars are displaying phenomena on all time scales. Modern studies can cover variable star phenomena on time scales faster than the duration of a single telescope run, and multiple telescope runs can be pasted together to get a picture of phenomena up to a decade in time scale. But to measure phenomena on time scales from a decade to a century, the only means is to use archival data. For most stars, the only source of archival information older than a decade or two is from archival photographic plates, and that largely means the Harvard plates. For the hot RCB stars, in looking for any secular trend (as associated with the evolution of the stars), the only solution is to get fully calibrated light curves from Harvard.
Historically, the Harvard plates were the predecessor of the Johnson B system through the North Polar Sequence. In modern times, the native magnitudes of the Harvard blue plates have always been found to have a near-zero color term with respect to the Johnson B system. This means that as long as the comparison star magnitudes are on the Johnson B system, then the resultant magnitudes are exactly in the Johnson B system.
For my measures of the Harvard plates, I have taken all my comparison star magnitudes from the B-band measures of the AAVSO Photometric All-Sky Survey ({\it APASS}, Henden \& Munari 2014). These magnitudes are tied to the Johnson B magnitudes with high accuracy (Munari et al. 2014), as calibrated from the standard stars of Landolt (2009). Thus, my modern measures of the Harvard plates are accurately in the Johnson B system.
Critically, both the very extensive {\it DASCH} program (Grindlay et al. 2012) and my own extensive measures prove that long term light curves from Harvard of normal (i.e., constant) stars do {\it not} produce any measurable slope or trend (i.e., typically $<$0.05 magnitude per century) over the last century. Further, these check star magnitudes are consistent with the modern measures. This is the proof that any observed secular trend is not some data or analysis artifact.
\section{Century-Long Light Curves For Hot RCB Stars}
The goal of this paper is to get the century-long light curves for all four known hot RCB stars so as to test for any secular fading in the maximum light. For this, the only way to get the old data is from Harvard, and these are only in the Johnson B-band. To minimize the mixing of bands, I will take the AAVSO and literature magnitudes for the B-band.
The Johnson B magnitudes for the four Hot RCB stars are listed in Table 1. These do not include the magnitudes where the star was substantially fainter than the maximum brightness.
\begin{table}
\centering
\caption{B Magnitudes from Harvard Plates.}
\label{tab:table1}
\begin{tabular}{llll}
\hline
Star & Julian Date & B (mag) & Plate\\
\hline
DY Cen & 2415898 & 12.7 & B29838 \\
DY Cen & 2416255 & 13.0 & B31827 \\
DY Cen & 2416959 & 13.3 & AM3470 \\
DY Cen & 2417257 & 12.6 & AM4107 \\
DY Cen & 2418405 & 12.8 & B40009 \\
DY Cen & 2418428 & 12.6 & AK286 \\
DY Cen & 2418437 & 12.6 & AM6102 \\
DY Cen & 2418507 & 12.7 & AM6372 \\
DY Cen & 2418869 & 13.3 & B41635 \\
DY Cen & 2420960 & 12.2 & AM11620 \\
DY Cen & 2421024 & 12.2 & AM11923 \\
DY Cen & 2421315 & 12.2 & AM12906 \\
DY Cen & 2421333 & 12.2 & AM12992 \\
DY Cen & 2421333 & 12.7 & AM13003 \\
DY Cen & 2421338 & 13.0 & AM13037 \\
DY Cen & 2421342 & 12.7 & AM13047 \\
DY Cen & 2422073 & 12.8 & AM14631 \\
DY Cen & 2422130 & 12.8 & AM14775 \\
DY Cen & 2422137 & 12.7 & AM14806 \\
DY Cen & 2422162 & 12.5 & AM14853 \\
DY Cen & 2422172 & 12.6 & AM14878 \\
DY Cen & 2422176 & 12.1 & AM14889 \\
DY Cen & 2422436 & 12.5 & AM15123 \\
DY Cen & 2422437 & 12.4 & AM15127 \\
DY Cen & 2422456 & 12.5 & AM15165 \\
DY Cen & 2422483 & 12.1 & AM15231 \\
DY Cen & 2422493 & 12.1 & AM15255 \\
DY Cen & 2422517 & 12.0 & AM15292 \\
DY Cen & 2422544 & 12.8 & AM15382 \\
DY Cen & 2423180 & 12.5 & AM15747 \\
DY Cen & 2426470 & 12.5 & RB1688 \\
DY Cen & 2426480 & 12.4 & RB1729 \\
DY Cen & 2426490 & 12.7 & RB1798 \\
DY Cen & 2426497 & 11.9 & RB1821 \\
DY Cen & 2426531 & 12.0 & RB1900 \\
DY Cen & 2426546 & 12.8 & RB1935 \\
DY Cen & 2426771 & 12.0 & RB2507 \\
DY Cen & 2426843 & 12.0 & RB2753 \\
DY Cen & 2426899 & 12.7 & RB3075 \\
DY Cen & 2431904 & 12.0 & RB14293 \\
DY Cen & 2431950 & 12.7 & RB14385 \\
DY Cen & 2432011 & 12.3 & RB14552 \\
DY Cen & 2432328 & 12.5 & RB15102 \\
DY Cen & 2432648 & 12.5 & RB15596 \\
DY Cen & 2432681 & 11.9 & RB15654 \\
DY Cen & 2432758 & 12.5 & RB15795 \\
DY Cen & 2433054 & 12.5 & RB16284 \\
DY Cen & 2445490 & 13.8 & DSB1047 \\
DY Cen & 2445813 & 13.5 & DSB1286 \\
DY Cen & 2445848 & 13.5 & DSB1325 \\
DY Cen & 2445872 & 13.1 & DSB1359 \\
DY Cen & 2445900 & 13.3 & DSB1388 \\
DY Cen & 2446243 & 13.5 & DSB1708 \\
DY Cen & 2446257 & 13.6 & DSB1711 \\
DY Cen & 2446291 & 13.5 & DSB1736 \\
DY Cen & 2446497 & 13.5 & DSB1903 \\
DY Cen & 2446527 & 13.2 & DSB1932 \\
DY Cen & 2446827 & 13.7 & DSB2153 \\
DY Cen & 2446945 & 13.5 & DSB2227 \\
DY Cen & 2447002 & 13.8 & DSB2300 \\
DY Cen & 2447022 & 13.8 & DSB2334 \\
DY Cen & 2447241 & 13.4 & DSB2493 \\
DY Cen & 2447267 & 13.3 & DSB2541 \\
\hline
\end{tabular}
\end{table}
\begin{table}
\centering
\contcaption{B Magnitudes from Harvard Plates.}
\label{tab:continued}
\begin{tabular}{llll}
\hline
Star & Julian Date & B (mag) & Plate\\
\hline
DY Cen & 2447298 & 13.7 & DSB2574 \\
DY Cen & 2447322 & 13.5 & DSB2588 \\
DY Cen & 2447357 & 13.3 & DSB2630 \\
DY Cen & 2447380 & 13.6 & DSB2653 \\
DY Cen & 2447590 & 13.5 & DSB2794 \\
DY Cen & 2447682 & 13.6 & DSB2827 \\
DY Cen & 2447761 & 13.6 & DSB2863 \\
MV Sgr & 2417077 & 12.6 & AM3801 \\
MV Sgr & 2422605 & 13.3 & MC16930 \\
MV Sgr & 2422606 & 13.0 & MC16931 \\
MV Sgr & 2425746 & 12.7 & RB334 \\
MV Sgr & 2425751 & 12.7 & RB345 \\
MV Sgr & 2425778 & 12.7 & RB402 \\
MV Sgr & 2428306 & 13.0 & MA5360 \\
MV Sgr & 2428672 & 12.9 & MA6375 \\
MV Sgr & 2429080 & 12.7 & MA7302 \\
MV Sgr & 2429429 & 12.1 & RB8838 \\
MV Sgr & 2429435 & 12.5 & RB8861 \\
MV Sgr & 2429441 & 12.6 & RB8891 \\
MV Sgr & 2429485 & 12.3 & RB9039 \\
MV Sgr & 2429547 & 12.0 & RB9181 \\
MV Sgr & 2429732 & 12.4 & RB9488 \\
MV Sgr & 2429793 & 12.4 & RB9691 \\
MV Sgr & 2429808 & 12.5 & RB9732 \\
MV Sgr & 2429811 & 12.4 & RB9755 \\
MV Sgr & 2429869 & 12.4 & RB9967 \\
MV Sgr & 2443616 & 13.5 & DSB512 \\
MV Sgr & 2444165 & 13.5 & DSB644 \\
MV Sgr & 2444821 & 13.5 & DSB769 \\
MV Sgr & 2445140 & 13.3 & DSB911 \\
MV Sgr & 2445173 & 13.3 & DSB919 \\
MV Sgr & 2445551 & 13.2 & DSB1087 \\
MV Sgr & 2445801 & 13.5 & DSB1274 \\
MV Sgr & 2445824 & 13.5 & DSB1305 \\
MV Sgr & 2445858 & 13.0 & DSB1346 \\
MV Sgr & 2445908 & 13.5 & DSB1408 \\
MV Sgr & 2446210 & 13.6 & DSB1672 \\
MV Sgr & 2446233 & 13.3 & DSB1702 \\
MV Sgr & 2446294 & 13.3 & DSB1755 \\
MV Sgr & 2446624 & 13.0 & DSB2022 \\
V348 Sgr & 2413724 & 12.1 & A1837 \\
V348 Sgr & 2415533 & 12.0 & AM808 \\
V348 Sgr & 2415576 & 11.7 & AM907 \\
V348 Sgr & 2415633 & 11.6 & AM1028 \\
V348 Sgr & 2415635 & 11.6 & AM1043 \\
V348 Sgr & 2417704 & 11.9 & AM4804 \\
V348 Sgr & 2417748 & 11.7 & AM4931 \\
V348 Sgr & 2417759 & 12.0 & AM4954 \\
V348 Sgr & 2417788 & 11.6 & AM5024 \\
V348 Sgr & 2417814 & 11.8 & AM5090 \\
V348 Sgr & 2417821 & 11.8 & AM5114 \\
V348 Sgr & 2418028 & 11.8 & AM5340 \\
V348 Sgr & 2418043 & 11.8 & AM5390 \\
V348 Sgr & 2418070 & 11.6 & AM5444 \\
V348 Sgr & 2418396 & 11.7 & AM6011 \\
V348 Sgr & 2418429 & 11.4 & AK288 \\
V348 Sgr & 2418439 & 11.6 & AM6114 \\
V348 Sgr & 2418454 & 11.7 & AM6170 \\
V348 Sgr & 2418502 & 11.8 & AM6347 \\
V348 Sgr & 2418532 & 11.7 & AM6454 \\
V348 Sgr & 2418822 & 11.8 & AM6952 \\
V348 Sgr & 2418849 & 11.8 & AM7033 \\
V348 Sgr & 2418856 & 11.6 & AM7063 \\
\hline
\end{tabular}
\end{table}
\begin{table}
\centering
\contcaption{B Magnitudes from Harvard Plates.}
\label{tab:continued}
\begin{tabular}{llll}
\hline
Star & Julian Date & B (mag) & Plate\\
\hline
V348 Sgr & 2419205 & 11.4 & AM7457 \\
V348 Sgr & 2419205 & 11.5 & AM7458 \\
V348 Sgr & 2419234 & 11.9 & AM7549 \\
V348 Sgr & 2419562 & 12.0 & AM8301 \\
V348 Sgr & 2419563 & 12.0 & AM8307 \\
V348 Sgr & 2419594 & 11.8 & AM8423 \\
V348 Sgr & 2419605 & 12.1 & AM8492 \\
V348 Sgr & 2419618 & 11.9 & AM8522 \\
V348 Sgr & 2419633 & 11.6 & AM8559 \\
V348 Sgr & 2422084 & 12.1 & AM14682 \\
V348 Sgr & 2422152 & 12.1 & MC16930 \\
V348 Sgr & 2422152 & 12.0 & MC16931 \\
V348 Sgr & 2422515 & 12.2 & MC16838 \\
V348 Sgr & 2422517 & 11.8 & AM15294 \\
V348 Sgr & 2422581 & 11.9 & AM15486 \\
V348 Sgr & 2422582 & 11.5 & MF07104 \\
V348 Sgr & 2423179 & 11.8 & AM15745 \\
V348 Sgr & 2423182 & 11.6 & AM15767 \\
V348 Sgr & 2423192 & 11.9 & A11979 \\
V348 Sgr & 2423195 & 11.8 & AM15795 \\
V348 Sgr & 2423210 & 11.5 & AM15842 \\
V348 Sgr & 2423223 & 11.7 & AM15869 \\
V348 Sgr & 2423236 & 11.8 & AM15909 \\
V348 Sgr & 2423248 & 11.5 & AM15931 \\
V348 Sgr & 2423249 & 11.6 & AM15934 \\
V348 Sgr & 2423347 & 11.9 & AM16120 \\
V348 Sgr & 2423663 & 11.5 & AM16382 \\
V348 Sgr & 2425706 & 12.0 & RB228 \\
V348 Sgr & 2425746 & 12.1 & RB334 \\
V348 Sgr & 2425751 & 11.9 & RB345 \\
V348 Sgr & 2425778 & 11.9 & RB402 \\
V348 Sgr & 2425798 & 11.9 & RB434 \\
V348 Sgr & 2426802 & 11.7 & RB2554 \\
V348 Sgr & 2426810 & 11.9 & RB2611 \\
V348 Sgr & 2426871 & 11.8 & RB2851 \\
V348 Sgr & 2426872 & 11.8 & RB2869 \\
V348 Sgr & 2427901 & 12.1 & RB6045 \\
V348 Sgr & 2428013 & 11.7 & RB6313 \\
V348 Sgr & 2428035 & 11.9 & AM17031 \\
V348 Sgr & 2428041 & 11.6 & AM17056 \\
V348 Sgr & 2429485 & 11.9 & RB9039 \\
V348 Sgr & 2429732 & 12.0 & RB9488 \\
V348 Sgr & 2430095 & 11.8 & RB10453 \\
V348 Sgr & 2430107 & 11.9 & AM21531 \\
V348 Sgr & 2430110 & 12.1 & AM21549 \\
V348 Sgr & 2430111 & 11.8 & RB10532 \\
V348 Sgr & 2430111 & 12.0 & RB10534 \\
V348 Sgr & 2430111 & 11.8 & RB10535 \\
V348 Sgr & 2430111 & 11.7 & RB10540 \\
V348 Sgr & 2430113 & 11.9 & RB10560 \\
V348 Sgr & 2430113 & 11.8 & RB10566 \\
V348 Sgr & 2430113 & 11.6 & RB10568 \\
V348 Sgr & 2430118 & 11.8 & RB10585 \\
V348 Sgr & 2430120 & 11.8 & RB10593 \\
V348 Sgr & 2430121 & 12.1 & RB10598 \\
V348 Sgr & 2430136 & 11.9 & RB10666 \\
V348 Sgr & 2430137 & 11.5 & RB10668 \\
V348 Sgr & 2430139 & 11.7 & RB10680 \\
V348 Sgr & 2430140 & 11.9 & RB10692 \\
V348 Sgr & 2430141 & 11.4 & RB10698 \\
V348 Sgr & 2430141 & 11.6 & RB10699 \\
V348 Sgr & 2430153 & 12.0 & RB10768 \\
V348 Sgr & 2430162 & 11.6 & RB10827 \\
V348 Sgr & 2430163 & 11.8 & RB10828 \\
\hline
\end{tabular}
\end{table}
\begin{table}
\centering
\contcaption{B Magnitudes from Harvard Plates.}
\label{tab:continued}
\begin{tabular}{llll}
\hline
Star & Julian Date & B (mag) & Plate\\
\hline
V348 Sgr & 2430163 & 11.7 & RB10829 \\
V348 Sgr & 2430220 & 11.7 & AM21975 \\
V348 Sgr & 2430221 & 11.9 & AX4088 \\
V348 Sgr & 2430299 & 11.9 & RB11123 \\
V348 Sgr & 2431158 & 12.1 & RB12537 \\
V348 Sgr & 2431172 & 11.6 & AM23552 \\
HV 2671 & 2413878 & 16.2 & A2172 \\
HV 2671 & 2414253 & 15.4 & B20843 \\
HV 2671 & 2416398 & 15.4 & B32728 \\
HV 2671 & 2416817 & 16.0 & A7098 \\
HV 2671 & 2423466 & 16.3 & A12286 \\
HV 2671 & 2423487 & 16.1 & A12288 \\
HV 2671 & 2423683 & 16.2 & A12699 \\
HV 2671 & 2423684 & 15.8 & A12700 \\
HV 2671 & 2423707 & 16.1 & A12788 \\
HV 2671 & 2423733 & 16.0 & A12830 \\
HV 2671 & 2423735 & 16.1 & A12834 \\
HV 2671 & 2423738 & 16.1 & A12848 \\
HV 2671 & 2423739 & 16.2 & A12851 \\
HV 2671 & 2423741 & 15.5 & A12855 \\
HV 2671 & 2423753 & 16.2 & A12865 \\
HV 2671 & 2425941 & 16.3 & A14366 \\
HV 2671 & 2426309 & 16.0 & A15041 \\
HV 2671 & 2426309 & 15.5 & MF15038 \\
HV 2671 & 2426322 & 16.3 & A15064 \\
HV 2671 & 2426328 & 15.9 & A15075 \\
HV 2671 & 2426335 & 16.1 & A15087 \\
HV 2671 & 2426410 & 16.1 & A15233 \\
HV 2671 & 2426412 & 15.9 & A15250 \\
HV 2671 & 2426413 & 16.1 & A15254 \\
HV 2671 & 2426414 & 16.2 & A15256 \\
HV 2671 & 2426421 & 16.1 & A15264 \\
HV 2671 & 2426426 & 16.1 & A15266 \\
HV 2671 & 2426441 & 16.1 & A15278 \\
HV 2671 & 2426444 & 16.2 & A15287 \\
HV 2671 & 2426452 & 16.3 & A15293 \\
HV 2671 & 2426453 & 16.1 & A15298 \\
HV 2671 & 2426454 & 15.9 & A15303 \\
HV 2671 & 2426455 & 16.1 & A15308 \\
HV 2671 & 2426456 & 16.3 & A15314 \\
HV 2671 & 2426566 & 16.3 & A15631 \\
HV 2671 & 2426568 & 16.4 & A15651 \\
HV 2671 & 2426569 & 16.3 & A15661 \\
HV 2671 & 2426572 & 16.4 & A15680 \\
HV 2671 & 2426573 & 16.2 & A15686 \\
HV 2671 & 2426578 & 16.4 & A15703 \\
HV 2671 & 2426606 & 15.2 & MF16077 \\
HV 2671 & 2426608 & 15.3 & MF16082 \\
HV 2671 & 2426636 & 16.4 & A15806 \\
HV 2671 & 2426657 & 15.7 & MF16170 \\
HV 2671 & 2426679 & 16.4 & A15838 \\
HV 2671 & 2426680 & 15.6 & MF16250 \\
HV 2671 & 2426684 & 16.3 & A15847 \\
HV 2671 & 2426687 & 16.0 & A15851 \\
HV 2671 & 2426687 & 15.7 & MF16282 \\
HV 2671 & 2426690 & 16.2 & A15858 \\
HV 2671 & 2426710 & 16.3 & A15872 \\
HV 2671 & 2426710 & 15.5 & MF16324 \\
HV 2671 & 2426720 & 15.4 & MF16389 \\
HV 2671 & 2426802 & 15.6 & MF16591 \\
HV 2671 & 2426931 & 15.3 & B56513 \\
HV 2671 & 2426946 & 15.6 & B56559 \\
HV 2671 & 2426947 & 16.2 & A16203 \\
HV 2671 & 2426950 & 15.7 & B56593 \\
\hline
\end{tabular}
\end{table}
\begin{table}
\centering
\contcaption{B Magnitudes from Harvard Plates.}
\label{tab:continued}
\begin{tabular}{llll}
\hline
Star & Julian Date & B (mag) & Plate\\
\hline
HV 2671 & 2426956 & 16.2 & B56627 \\
HV 2671 & 2426957 & 15.7 & B56637 \\
HV 2671 & 2426978 & 16.2 & A16254 \\
HV 2671 & 2427311 & 15.9 & A16561 \\
HV 2671 & 2427749 & 16.3 & A17232 \\
HV 2671 & 2427777 & 16.0 & A17258 \\
HV 2671 & 2427800 & 16.4 & A17287 \\
HV 2671 & 2427800 & 16.3 & A17288 \\
HV 2671 & 2427800 & 16.2 & A17289 \\
HV 2671 & 2427800 & 16.2 & A17290 \\
HV 2671 & 2427800 & 16.3 & A17291 \\
HV 2671 & 2427801 & 16.3 & A17295 \\
HV 2671 & 2427802 & 15.8 & A17298 \\
HV 2671 & 2427807 & 16.2 & A17302 \\
HV 2671 & 2427807 & 16.2 & A17303 \\
HV 2671 & 2427808 & 16.1 & A17307 \\
HV 2671 & 2427808 & 16.4 & A17308 \\
HV 2671 & 2427808 & 16.4 & A17309 \\
HV 2671 & 2427808 & 16.0 & A17311 \\
HV 2671 & 2427811 & 16.2 & A17315 \\
HV 2671 & 2429584 & 16.5 & A21491 \\
HV 2671 & 2429606 & 16.2 & B65009 \\
HV 2671 & 2429671 & 16.1 & B65083 \\
HV 2671 & 2429674 & 16.3 & A21606 \\
HV 2671 & 2429690 & 16.2 & A21621 \\
HV 2671 & 2429879 & 15.9 & B65919 \\
HV 2671 & 2429905 & 16.0 & A22207 \\
HV 2671 & 2429934 & 16.3 & A22269 \\
HV 2671 & 2429939 & 16.1 & A22277 \\
HV 2671 & 2429956 & 16.2 & A22305 \\
HV 2671 & 2429970 & 16.1 & A22330 \\
HV 2671 & 2429994 & 16.0 & A22340 \\
HV 2671 & 2430023 & 16.0 & B66141 \\
HV 2671 & 2430045 & 15.7 & MF28653 \\
HV 2671 & 2430057 & 16.2 & MF22404 \\
HV 2671 & 2430058 & 16.2 & A22409 \\
HV 2671 & 2430080 & 15.9 & B66300 \\
HV 2671 & 2430101 & 15.7 & MF28870 \\
HV 2671 & 2430110 & 15.8 & MF28953 \\
HV 2671 & 2430111 & 16.3 & MF28967 \\
HV 2671 & 2430112 & 16.2 & MF28976 \\
HV 2671 & 2430240 & 16.1 & B67078 \\
HV 2671 & 2430264 & 15.7 & B67149 \\
HV 2671 & 2430314 & 16.2 & A22980 \\
HV 2671 & 2430314 & 16.0 & B67253 \\
HV 2671 & 2430315 & 16.3 & A22987 \\
HV 2671 & 2430318 & 16.0 & A22992 \\
HV 2671 & 2430318 & 16.2 & A22994 \\
HV 2671 & 2430319 & 16.1 & A22995 \\
HV 2671 & 2430320 & 16.3 & A23002 \\
HV 2671 & 2430322 & 16.1 & A23007 \\
HV 2671 & 2430323 & 16.1 & A23008 \\
HV 2671 & 2430324 & 16.0 & A23011 \\
HV 2671 & 2430325 & 16.0 & A23017 \\
HV 2671 & 2430328 & 16.1 & A23020 \\
HV 2671 & 2430372 & 16.0 & B67322 \\
HV 2671 & 2430373 & 16.2 & A23044 \\
HV 2671 & 2430373 & 15.9 & B67325 \\
HV 2671 & 2430373 & 16.0 & B67327 \\
HV 2671 & 2430373 & 16.2 & A23046 \\
HV 2671 & 2430373 & 16.0 & A23047 \\
HV 2671 & 2430375 & 15.8 & B67328 \\
HV 2671 & 2430375 & 15.7 & B67330 \\
HV 2671 & 2430586 & 16.1 & A23415 \\
\hline
\end{tabular}
\end{table}
\begin{table}
\centering
\contcaption{B Magnitudes from Harvard Plates.}
\label{tab:continued}
\begin{tabular}{llll}
\hline
Star & Julian Date & B (mag) & Plate\\
\hline
HV 2671 & 2430591 & 16.1 & A23424 \\
HV 2671 & 2430591 & 15.9 & B67968 \\
HV 2671 & 2430594 & 16.0 & A23427 \\
HV 2671 & 2430606 & 16.2 & A23430 \\
HV 2671 & 2430621 & 16.1 & A23450 \\
HV 2671 & 2430621 & 15.8 & B68040 \\
HV 2671 & 2430625 & 16.1 & A23453 \\
HV 2671 & 2430640 & 16.3 & A23458 \\
HV 2671 & 2430641 & 16.3 & A23462 \\
HV 2671 & 2430642 & 16.0 & A23466 \\
HV 2671 & 2430648 & 16.1 & A23471 \\
HV 2671 & 2430666 & 16.1 & A23485 \\
HV 2671 & 2430673 & 16.2 & A23490 \\
HV 2671 & 2430696 & 16.3 & A23502 \\
HV 2671 & 2430713 & 16.2 & A23513 \\
HV 2671 & 2430749 & 15.9 & A23528 \\
HV 2671 & 2430750 & 16.2 & A23530 \\
HV 2671 & 2430766 & 15.4 & MF31282 \\
HV 2671 & 2430767 & 16.2 & A23570 \\
HV 2671 & 2430782 & 15.6 & MF31352 \\
HV 2671 & 2430791 & 15.1 & MF31364 \\
HV 2671 & 2430792 & 15.3 & MF31381 \\
HV 2671 & 2430793 & 15.5 & MF31390 \\
HV 2671 & 2430809 & 15.7 & B68351 \\
HV 2671 & 2431804 & 16.3 & A25189 \\
HV 2671 & 2431814 & 15.5 & B71365 \\
HV 2671 & 2431817 & 15.8 & MF35012 \\
HV 2671 & 2431823 & 16.1 & A25194 \\
HV 2671 & 2431873 & 16.2 & A25218 \\
HV 2671 & 2431874 & 16.0 & B71427 \\
HV 2671 & 2432070 & 16.3 & B72205 \\
HV 2671 & 2432070 & 16.3 & A25565 \\
HV 2671 & 2432940 & 15.6 & A26696 \\
HV 2671 & 2433161 & 15.9 & A26976 \\
HV 2671 & 2433181 & 16.2 & A26998 \\
\hline
\end{tabular}
\end{table}
The archived magnitudes from the AAVSO are freely available on-line. The AAVSO B-band measures are all taken with CCDs, with photometric uncertainties of 0.03 mag or better. Observers are designated with a three-letter designation, with HMB being Dr. Franz-Josef Hambsch in Belgium, DSI being Giorgio di Scala in Australia, and SXN being Michael Simonson in the United States. These are all calibrated with APASS comparison stars, and are thus in the Johnson B system.
I have also pulled a variety of magnitudes from the literature, and these are all CCD measures. (The one exception is the single magnitude from Herbig in 1964 for MV Sgr.) These have been calibrated ultimately from the Landolt fields, and thus are also in the Johnson B system. Intercomparison of modern published B magnitudes always shows that different sources disagree with each other up to $\sim$0.1 mag, even for known-constant stars and for effectively simultaneous measures of slow variable stars. This is likely being due to different color terms and calibrations between observers. Within each literature source, the quoted error bars are usually $\sim$0.01 mag, but these are always measurement errors and do not include systematic errors that will appear as a constant offset for each source. Fortunately, the hot RCB stars show variations that are greatly larger than these usual calibration problems, so the existence and slope of the trends remain unaffected. In all, to within the normal errors, all the literature magnitudes are on the modern Johnson B system.
The archival magnitudes in the literature (Hoffleit 1930; 1958; 1959; Woods 1928) are not used, because all have big photometric differences from the modern B magnitude system due to problems with the comparison star sequences, as was universal for the era. I have examined the exact same plates at Harvard, plus many more, all on the modern Johnson B system, so my magnitudes now supersede the old ones in the literature.
\begin{figure}
\includegraphics[width=1.1\columnwidth]{Fig1.pdf}
\caption{V348 Sgr in B from AAVSO in 2014-2015. The observer was Dr. Franz-Josef Hambsch, in Belgium with a 14-inch telescope. This Johnson B light curve has 483 points, for which 164 magnitudes have been selected as representing the star at maximum light, while Hambsch also has Johnson V magnitudes on all the same nights. This light curve illustrates that the complete recovery from a dip only asymptotically approaches some presumably-dust-free maximum. It also illustrates that the time duration when the star is in the dip but just below maximum is a very small fraction of the time. A further point is that we can confidently measure the magnitudes at maximum to better accuracy than the maximum can be defined. The main point of this figure is that the recent maximum of V348 Sgr is close to B=12.93, whereas the Harvard plates show a maximum light around B=11.8 over a century earlier, with this being proof of a secular decline.}
\label{fig:Fig. 1}
\end{figure}
A substantial problem is to select out the magnitudes taken when the star is at maximum light. Part of the problem is that the brightness recovery from a decline is only asymptotic, so all magnitudes will still have some residual dust to varying degrees. This is illustrated in Fig. 1, where V348 Sgr never quite recovers completely to some dust-free maximum brightness. An adequate solution is simply that this effect should be the same for old and new magnitudes, so there should be negligible effect on any trends. The biggest part of the problem is that most of the magnitudes are isolated in time, so we cannot recognize whether the star is at maximum or is in a dip. Magnitudes greatly fainter than some maximum are easily recognized and rejected, but magnitudes from the start or end of a dip, with the brightness only somewhat below the true maximum, can be included, resulting in an apparent fainter maximum. The inclusion of more or fewer in-decline magnitudes will make the star's maximum appear to be fainter or brighter. Fortunately, this problem is minimized by several means. First, dips are deep with fast drop offs, so there will be only a small fraction of the time during which the star will be close-but-below maximum light. That is, contaminated magnitudes must be rare and statistically negligible. Second, for DY Cen and MV Sgr, the dips are rare, so there is little opportunity for contamination. Third, I have rejected plates taken near times of known dips, whether or not the plate shows the star near a maximum. Fourth, for the AAVSO light curves, there are a high density of observations so that dips can be easily recognized (e.g., see Fig. 1) and avoided. Fifth, in generating a light curve at maximum, the effect of including magnitudes in dips) will only matter for measuring secular fading if the early and late measures have different inclusion fractions for dip-magnitudes, and this does not seem plausible. In general, operationally, when I have no additional information, I have tossed out magnitudes if they are more than a magnitude fainter than the maximum for that star and decade. There is a plausible chance that inclusion of the initial and final parts of dips has slightly lowered some of the averages over time. In general, the problem is likely to be minimal in the averages, and certainly the effect is smaller than the trends observed. Thus, I conclude that this problem is not a significant contributor to the observed trends for any of the hot RCB stars.
DY Cen has had no minimum from 1960 to 2016, as shown by the densely-sampled light curves from the Royal Astronomical Society of New Zealand and from the AAVSO (De Marco et al. 2002). With the Harvard plates, I can extend this back to 1935, although the interval from 1954 to 1960 is poorly covered due to the Menzel Gap. Before 1930, Hoffleit (1930) identified four dust dips with the Harvard plates, while I have added further dips. The known dips are in 1897, 1901, 1904.4, 1906.3-1908.5, 1914.5, 1915.3, 1918.5, 1924.1-1924.6, 1929.2-1929.5, 1931.2, 1932.6 and 1934.2. The coverage of the dips is patchy, but it appears that durations are a few months, other than the cases noted. Further dips are likely to have occurred, mainly during the part of the year when DY Cen is the closest to the Sun. We are left with a stark situation where DY Cen has many dust dips from 1895 to 1934, but none from 1935 to 2016.
With this, I have constructed a maximum light curve for each of the hot RCB stars with Harvard, AAVSO, and literature magnitudes, all in the Johnson B system. A simple plot of all these magnitudes shows the usual scatter, with this somewhat hiding secular trends. To solve this, I have averaged the magnitudes by source and time interval. The one-sigma uncertainty is taken to be the RMS scatter in the magnitudes divided by the square root of the number of magnitudes. These averages are presented in Table 2 and Fig. 2.
We see that all four hot RCB stars have an obvious secular decline of roughly one magnitude per century. This then provides the direct confirmation of the result in De Marco et al. (2002). The light curves show roughly linear declines. There is substantial scatter around these linear declines, with it being unclear whether this is due to real variation in the star's maximum light, due to inclusion of magnitudes just below maximum, or due to measurement error. The secular decline need not be linear or even monotonic. We can quantify the secular decline by an average decline rate, derived from a linear fit.
I have made chi-square fits for the light curves in Table 1 for a linear decline. The resultant fits have reduced chi-square values that are much greater than unity, pointing to the variations around the simple straight line being much larger than the nominal error bars. As such, the formal one-sigma error bars from the chi-square fits for the slope are not meaningful. The fitted slopes are 1.15 for DY Cen, 1.29 for MV Sgr, 1.29 for V348 Sgr, and 0.73 for HV 2671, all in units of magnitudes per century. For these four slopes, the average is 1.11 magnitude per century, with an RMS of 0.26 magnitude per century. The calculation of an averaged linear slope is making no implication that any of the stars has a constant linear slope, nor that the stars all have the same linear slope. Indeed, for DY Cen, the light curve appears to be more of a parabola than a line. Further, the Hot RCB stars are apparently a diverse class, so an average decline rate will be some sort of a mixture of rates for stars with different histories and masses. Still, the average fade rate of 1.11$\pm$0.26 magnitude per century has utility in expressing the typical decline rate and its variations, in quantitatively showing that the Hot RCB stars are fast fading, and in providing a representative rate for model calculations.
While still with only one band, the AAVSO visual light curves are long enough and with enough accuracy that we can get an independent measure of the secular fading rate. For DY Cen, 6438 visual magnitudes cover the time from April 1978 to October 2015 with no dips, with an average decline rate of 1.87 magnitudes per century. For V348 Sgr, the frequent dips make it harder to pick out a decline by eye from the full visual light curve, yet the maximum magnitudes are around 11.5 in the 1950's and around 12.2 for the last decade, for a decline rate of approximately 1.3 magnitudes per century.
\begin{table}
\centering
\caption{Hot RCB star light curves.}
\label{tab:table1}
\begin{tabular}{llll}
\hline
Star & Years & $\langle$B$\rangle$ (mag) & Source\\
\hline
DY Cen & 1902--1910 & 12.84 $\pm$ 0.10 & HCO (9 plates) \\
DY Cen & 1916--1922 & 12.46 $\pm$ 0.06 & HCO (21 plates) \\
DY Cen & 1931--1932 & 12.33 $\pm$ 0.12 & HCO (9 plates) \\
DY Cen & 1946--1949 & 12.36 $\pm$ 0.10 & HCO (8 plates) \\
DY Cen & 1970 & 12.62 $\pm$ 0.03 & Marino \& Walker (1971) \\
DY Cen & 1972 & 12.70 $\pm$ 0.03 & Sherwood (1975)$^a$ \\
DY Cen & 1983--1989 & 13.51 $\pm$ 0.04 & HCO (23 plates) \\
DY Cen & 1983 & 12.96 $\pm$ 0.02 & Kilkenny et al. (1985) \\
DY Cen & 1985 & 13.03 $\pm$ 0.02 & Goldsmith et al. (1990) \\
DY Cen & 1987 & 13.11 $\pm$ 0.02 & Pollacco \& Hill (1991) \\
DY Cen & 1988 & 13.22 $\pm$ 0.04 & Jones et al. (1989) \\
DY Cen & 2006--2007 & 13.45 $\pm$ 0.01 & AAVSO (DSI, 38 mags) \\
DY Cen & 2013--2015 & 13.82 $\pm$ 0.09 & AAVSO (SXN, 25 mags) \\
MV Sgr & 1905 & 12.60 $\pm$ 0.20 & HCO (1 plate) \\
MV Sgr & 1920 & 13.15 $\pm$ 0.20 & HCO (2 plates) \\
MV Sgr & 1929 & 12.70 $\pm$ 0.20 & HCO (3 plates) \\
MV Sgr & 1934-1940 & 12.48 $\pm$ 0.08 & HCO (13 plates) \\
MV Sgr & 1963 & 12.96 $\pm$ 0.10 & Herbig (1964) \\
MV Sgr & 1978--1986 & 13.36 $\pm$ 0.05 & HCO (14 plates) \\
MV Sgr & 1985 & 13.62 $\pm$ 0.03 & Goldsmith et al. (1990) \\
MV Sgr & 2006--2014 & 13.59 $\pm$ 0.01 & AAVSO (DSI, 76 mags) \\
MV Sgr & 2011--2015 & 13.90 $\pm$ 0.01 & AAVSO (SXN, 71 mags) \\
V348 Sgr & 1896--1901 & 11.80 $\pm$ 0.09 & HCO (5 plates) \\
V348 Sgr & 1907--1912 & 11.75 $\pm$ 0.03 & HCO (27 plates) \\
V348 Sgr & 1919--1923 & 11.79 $\pm$ 0.04 & HCO (18 plates) \\
V348 Sgr & 1929--1935 & 11.87 $\pm$ 0.03 & HCO (13 plates) \\
V348 Sgr & 1939--1944 & 11.81 $\pm$ 0.03 & HCO (30 plates) \\
V348 Sgr & 1970 & 12.50 $\pm$ 0.10 & Heck et al. (1985) \\
V348 Sgr & 1972--1974 & 12.78 $\pm$ 0.28 & Heck et al. (1985) \\
V348 Sgr & 1981 & 12.78 $\pm$ 0.01 & Heck et al. (1985) \\
V348 Sgr & 2014--2015 & 12.93 $\pm$ 0.02 & AAVSO (HMB, 164 mags) \\
HV 2671 & 1896--1904 & 15.76 $\pm$ 0.26 & HCO (4 plates) \\
HV 2671 & 1923 & 16.05 $\pm$ 0.10 & HCO (11 plates) \\
HV 2671 & 1929--1935 & 16.07 $\pm$ 0.04 & HCO (63 plates) \\
HV 2671 & 1939--1943 & 16.02 $\pm$ 0.03 & HCO (68 plates) \\
HV 2671 & 1945--1949 & 16.02 $\pm$ 0.08 & HCO (11 plates) \\
HV 2671 & 1993--1999 & 16.75 $\pm$ 0.1 & Alcock et al. (1996) \\
HV 2671 & 2001--2009 & 16.41 $\pm$ 0.1 & Soszynski et al. (2009) \\
\hline
\end{tabular}
$^a$As quoted in Rao et al. (1993)
\end{table}
\begin{figure}
\includegraphics[width=1.1\columnwidth]{Fig2a.pdf}
\includegraphics[width=1.1\columnwidth]{Fig2b.pdf}
\caption{Century-long Johnson B light curves for all four known hot RCB stars. The main point of this figure and this paper is that all four known hot RCB stars have obvious and significant secular declines. The thick lines are from the formal chi-square fit, which represent the average secular fading of the stars. The scatter around these best fits is much larger than the nominal error bars, and it is not clear whether this is due to the intrinsic variations of the maximum brightness, the inclusion of just-below-maximum in-dip magnitudes, or ordinary photometric errors. The four panels are for DY Cen, MV Sgr, V348 Sgr, and HV 2671. The fading of DY Cen is apparent only since 1960 or so, whereas the star was {\it brightening} before the 1930s.}
\label{fig:Fig. 2}
\end{figure}
\begin{figure}
\includegraphics[width=1.1\columnwidth]{Fig2c.pdf}
\includegraphics[width=1.1\columnwidth]{Fig2d.pdf}
\label{fig:Fig. 2cd}
\end{figure}
\section{DIscussion}
In essence, I have merely confirmed and extended the conclusion of De Marco et al. (2002) that the hot RCB stars are secularly fading. My improvements have been to use a single photometric system all with modern comparison stars, to collect many more magnitudes over a much wider time range, as well as to measure the decline rate for the fourth hot RCB star.
De Marco et al. (2002) have interpreted these secular declines as being due to the star evolving to hotter temperature at constant luminosity, such that the bolometric correction to the optical band makes for an apparent dimming. (The only other plausible explanation is some sort of general increase in the circumstellar dust density, but such would lead to color changes that are not observed in the cases of DY Cen and MV Sgr.) This interpretation matches with the general idea that the hot RCB stars are moving horizontally across the top of the HR diagram as part of their normal and fast evolution. Pandey et al. (2014) have explicitly tested this interpretation for DY Cen, where archival spectra give surface temperatures of 19,400$\pm$400 K in 1987, 23,000$\pm$300 K in 2002, and 24,800$\pm$600 K in 2010. This is 5,400 K in 23 years, or 23,500 K per century. This increase in the stellar temperature is confirmed and reflected in the dramatic change in the excitation of the nebula around DY Cen (Rao et al. 2013).
We can translate this rate of temperature change for DY Cen into a magnitude decline rate. The calculation of the change in bolometric corrections and the change of B-V color is presented in Fig. 1 of Pandey et al. (2014) for the relevant conditions. For a temperature of 19,400 K, they give V=12.78 and B-V=-0.80 (with an arbitrary zero point), for B=11.98. For a temperature of 24,800 K, they give V=13.38 and B-V=-0.85 (with the same arbitrary zero point), for B=12.53. Thus, the observed temperature decline in 23 years corresponds to a fading by 0.55 mag, for a decline rate of 2.39 magnitudes per century. This is close to the average decline rate for the years 1983 to 2015 (see Fig. 2a). So the observed temperature change is consistent with the observed decline rate.
For the evolution of DY Cen going back in time, the temperature must be relatively low in old times, resulting in a large bolometric correction. A simple extrapolation back to 1905 puts the temperature to near zero, so the temperature changes cannot be linear with time. Nevertheless, the temperature back in 1905 should be relatively quite cold. The bolometric correction for the B band is minimized for a stellar temperature of 7,500 K, so that for evolution at constant luminosity, the B magnitude will be brightest at that temperature and dimmer as the temperature departs from this value to both hotter and colder temperatures. So we then have a ready interpretation of the long-term trend in the maximum magnitude (Fig. 2a), with DY Cen starting in 1906 out colder than 7,500 K, heating up to 7,500 K in 1932 when the star was at its brightest, then continuing heating of the star makes it dim over the next decades. The correction from a constant luminosity to the B band can be taken from Table 15.7 of Cox et al. (2000), where the minimum correction is at 7,500 K, and where the corrections of 0.6 mag are for temperatures of 5,800 K and 11,000 K. With this, DY Cen had temperatures of 5,800 K around 1906, 7,500 K around 1932, and 11,000 K around the 1970s. So we now have a simple explanation for why the DY Cen maxima back around 1906 was substantially dimmer than around in 1932. In all, a continuous temperature increase from 5,800 K in 1905 to 24,800 K in 2010 can account for both the observed change in maximum magnitudes and the observed changes in the temperature.
I have made a crude model that accounts for the stellar temperature and maximum magnitude as a function of time. From the models of Saio (1988), I take the logarithm of the temperature to be linear with time, with this being approximately right for a given star under hot RCB conditions. I then set the linear relation by using the observed temperature in 2010 plus the 7,500 K condition for 1932. With these temperatures, I get the bolometric corrections and B-V colors for supergiants from Cox et al. (2000), and add a constant to get the B-band magnitude outside of a decline. This model light curve is displayed in Fig. 3. This model result is not perfect, with the worst problem being that the bolometric correction for 24,800 K should make DY Cen close to 2.0 mag fainter in 2010 than in 1932, whereas it is observed to be more like 1.3 mag fainter. This problem is easily solved if there is extra light in the DY Cen system, perhaps from a wide binary companion or from the circumstellar material. Nevertheless, it is clear that the model captures the essence of a normal RCB stars heating up from around 5,800 K in 1906 to 24,800 K in 2010.
\begin{figure}
\includegraphics[width=1.1\columnwidth]{Fig3.pdf}
\caption{DY Cen evolution from normal RCB to hot RCB to an extreme helium star. Over the last century, the maximum brightness has first brightened, come to a peak round 1932, then started a secular decline continuing to today. The temperature is observed to go from 19,400 K in 1987 to 24,800 K in 2010. This is all consistent with the expected evolution from right to left across the HR diagram at constant luminosity. DY Cen started in 1906 with a temperature of near 5,800 K as a normal RCB star, and rapidly heated up. As it heated up, the bolometric correction lessened, making the star appear brighter, until around 1932 when the bolometric correction is its smallest for a temperature near 7,500 K. As the star kept heating up, the bolometric correction got larger, making the maximum magnitude dim, with this continuing to today. A crude model of this is shown here, with the logarithm of the temperature assumed to change linearly with time, for comparison with the observed peak magnitude. The frequency of RCB dips has also change, with many known dips before 1934, but none known after 1934. All known dips are shown displayed onto the model light curve.}
\label{fig:Fig. 3}
\end{figure}
So we have actually watched DY Cen start out as an ordinary RCB star (with temperature around 5,800 K), and then heat up to become a hot RCB star, and now appear as an extreme helium star with no dust dips. This evolution has taken close to one century. So we have a real measure of the duration of the hot RCB stage, and it is about one century. This is a very fast phase of evolution. This explains why so few hot RCB stars have been seen in our Milky Way galaxy, despite them being supergiant stars.
The heating up of DY Cen is also associated with a sharp drop off in the frequency of dips. De Marco et al. (2002) note that DY Cen had only four knowns dips from 1897 to 1927, and zero known dips since 1960. With the Harvard plates, I have identified eight additional dips, all from 1904 to 1934. Apparently, the heating of the star's surface temperature is connected with the turn off of the dust formation, perhaps caused by a stoppage of pulsations as the star leaves some instability strip. We realize that there is only a narrow time window over which the hot RCB phenomenon can be recognized, with only a few decades from the time when DY Cen was sufficiently hotter than the upper limit for normal RCB stars up until the time when the dust dips turn off.
We can translate the typical decline rate of 1.11 magnitude per century into a temperature change rate. For a case with effective surface temperature of 15,000 K, Pandey et al. (2014) give V=12.15 and B-V=-0.75 (with the same arbitrary zero point), for B=11.40. For a temperature of 20,000 K, they give V=12.88 and B-V=-0.79, for B=12.09. For a 5,000 K temperature change over the range for hot RCB stars, the B magnitude changes by 0.69 magnitudes. If this change happens over 62 years, then the B magnitude will fade at the rate of 1.11 magnitude per century.
DY Cen is similar to extreme helium stars (supergiants composed mostly of helium with near one percent carbon and temperatures 9000-35,000 K). Jeffery et al. (2001) found that four out of twelve extreme helium stars are heating up with rates from 20 to 120 degrees K per year. (A useful program would be to search for B-band brightness changes from the 1890s to the present with the Harvard plates for the two stars with the fastest temperature changes; HD 160641 and BD -1$\degr$3438.) Such surface temperature changes are expected from models of extreme helium stars with masses of $\sim$0.9 M$_{\odot}$ (Saio 1988). The majority of stars with no measured change in surface temperature are presumably less massive, perhaps $\sim$0.7 M$_{\odot}$. DY Cen is changing at a rate of 235 K per year from 1987 to 2010. If the models of Saio (1988) are applicable to DY Cen, then this star would be $\sim$1.0 M$_{\odot}$.
With the realization as to how some `cold' RCB stars should evolve on a time scale of a century, we can look for the same changes amongst the known normal RCB stars. That is, the normal RCB stars are heating up, having their maximum magnitudes getting brighter, and their frequency of declines falling to near zero. But such changes have never been seen for any star that is now a `cold' RCB star. A small number of cold RCB stars have century-long light curves with no apparent change in their brightness at maximum, while R CrB itself has a 230 year record of unchanging peaks. So the heating up of the cold RCB stars must usually be too slow to produce observable effects. Still, some fraction of the now-cold RCB stars might be like DY Cen a century ago. Perhaps only the most-massive cold RCB stars will be evolving fast enough for the changes to be detectable.
A practical plan to search for fast evolving cold RCB stars is to construct a century-long light curve for many of them. This could show secular changes in the magnitude at maximum as well as in the frequencies of declines. In practice, the primary sources are archival data from AAVSO and Harvard. In any such study, care must be used to place all magnitudes onto a consistent magnitude system. (For example, old AAVSO magnitudes will require corrections for changes in the comparison sequences, and these can only be gotten from old charts archived at AAVSO Headquarters.) A substantial problem in seeking changes in the decline-frequency will be to adjust for the variations in time-coverage over the decades. With this, we have a plan for someone to make a systematic survey of century-long light curves for normal RCB stars so as to measure their evolution across the HR diagram.
In general, stars evolve on such slow time scales that astronomers have not been able to see the changes over time. Other than for supernovae, evolutionary changes have only been seen for a few post-AGB stars, including the born-again stars and the Stingray (Schaefer \& Edwards 2015). Now, we can add the four hot RCB stars, with observed temperature rises of 8,000 K or more over the last century.
|
\section{Introduction}
\looseness=-1
Chimera states in populations of coupled oscillators have attracted large
interest since their first observation and theoretical explanation
by Kuramoto and Battogtokh~\cite{Kuramoto-Battogtokh-02}. The essence
of chimera is in the spontaneous symmetry breaking: although a homogeneous
fully symmetric synchronous state exists, yet another nontrivial state combining
synchrony and asynchrony is possible and can even be stable. Chimeras
can be found at interaction of several populations of
oscillators~\cite{Abrams-Mirollo-Strogatz-Wiley-08,*Pikovsky-Rosenblum-08,%
*Tinsley_etal-12,*Martens_etal-13}, or
in an oscillatory
medium~\cite{Abrams-Strogatz-04,ShiKu04,Laing-09,Bordyugov-Pikovsky-Rosenblum-10},
the latter situation can be treated as a pattern formation
problem. Here the formulation in terms of the
coarse-grained complex order parameter indeed
allows one to reduce the problem to that of evolution of a complex
field~\cite{Laing-09,Bordyugov-Pikovsky-Rosenblum-10}. For a recent review see~\cite{Panaggio-Abrams-15}.
The goal of this paper is to develop a theory of chimera patterns in a one-dimensional (1D)
medium based on formulation of the problem as a set of partial differential equations (PDEs). This
allows us to represent
the chimera state as a solution of ordinary differential equations (ODEs), periodic in space
chimeras correspond to periodic orbits of these ODEs.
We show that in a limit of neutral coupling, these equations are integrable yielding singular
``one-point'' and ``two-point'' chimeras; for a weakly attracting coupling we find properties of chimera patterns
by virtue of perturbation analysis to these solutions.
Furthermore, we study stability of found chimera patterns
by employing a numerical method allowing to disentangle the essential continuous and
the discrete (point)
parts~\cite{Wolfrum_etal-11,Omelchenko-13} of the stability spectrum.\par
The original Kuramoto-Battogtokh (KB) model~\cite{Kuramoto-Battogtokh-02} is formulated as a 1D field of phase oscillators $\phi\left(x,t\right)$ evolving according to
\begin{equation}
\partial_{t}\phi=
\omega+\mathrm{Im}\left(e^{-i\left(\phi+\alpha\right)}\!\int\!G\left(x-\tilde{x}\right) e^{i\phi(\tilde{x},t)}d\tilde{x}\right),
\label{eq:kbmodel}
\end{equation}
with exponential kernel $G\!\left(y\right)=\kappa \exp\bigl(-\kappa\left|y\right|\bigr)\!\bigl/2\bigr.$. Coupling
is attractive if the phase shift $\alpha<\pi\bigl/2\bigr.$, then the synchronous state
where all the phases are equal is stable; $\alpha=\pi\bigl/2\bigr.$ corresponds to neutral coupling.
One can reformulate this setup as a 1D continuous oscillatory
medium~\cite{Laing-09,Bordyugov-Pikovsky-Rosenblum-10},
described by the complex field $Z\!\left(x,t\right)$, which represents a coarse-grained order parameter of the phases:
$Z\!\left(x,t\right)=\frac{1}{2\delta}\int_{x-\delta}^{x+\delta} e^{i\phi(\tilde{x},t)}d\tilde{x}$.
In the synchronous state $|Z|=1$, while for partial synchrony $0<|Z|<1$.
The dynamics of $Z\!\left(x,t\right)$ just follows locally the Ott-Antonsen equation~\cite{Ott-Antonsen-08,Panaggio-Abrams-15}
\begin{equation}
\partial_{t}Z=i\omega Z+\bigl(e^{-i\alpha}H-e^{i\alpha}H^{\ast}Z^{2}\bigr)\bigl/2\bigr.\;.
\label{eq:oa1}
\end{equation}
where a field $H\!\left(x,t\right)\!=\!\int \!G\!\left(x-\tilde{x}\right)Z\!\left(\tilde{x},t\right)d\tilde{x}$ describes~the force due to coupling.
This nonlocal coupling according stems from the following model
for the interaction of oscillators via the ``auxiliary''
field $H$ (cf. Refs.~\cite{ShiKu04,Laing-11,Laing-15}):
\begin{equation}
\tau\partial_{t}H=\kappa^{-2}\partial^{2}_{xx}H-H+Z\;.
\label{eq:difeq1}
\end{equation}
In the limit $\tau\to0$, Eq.~\eqref{eq:difeq1} reduces to an equation
\begin{equation}
\partial^{2}_{xx}H-\kappa^2 H =-\kappa^2 Z\;,
\label{eq:difeq2}
\end{equation}
solution of which depends on boundary conditions. In particular, in an infinite medium $\left|x\right|<\infty$,
the solution is $H\!\left(x,t\right)=\!\int \!\left(\kappa \exp\bigl(-\kappa\left|x-\tilde x\right|\bigr)\!\bigl/2\bigr.\right)
Z\!\left(\tilde{x},t\right)d\tilde{x}$ as in~\eqref{eq:kbmodel}.
Below we consider periodic in space medium with period $L$, in this case
the KB model exactly corresponds to Eqs.~\eqref{eq:oa1},{\,}\eqref{eq:difeq2} if integration
is performed in an infinite domain while the fields are assumed to have period $L$.
If an integration over the periodic domain of size $L$ is performed,
one should use the kernel
\begin{equation}
G(y)=\frac{\kappa}{2\sinh\bigl(\kappa L\bigl/2\bigr.\bigr) }\cosh\bigl(\kappa\left(\left|y\right|-L\bigl/2\bigr.\right)\bigr)\;,
\label{eq:kernel}
\end{equation}
which follows from the solution of \eqref{eq:difeq2} with periodic boundary conditions.
The formulated problem \eqref{eq:oa1},{\,}\eqref{eq:difeq2} contains two parameters having dimension of
length: $\kappa$ and $L$.
By rescaling coordinate $x$ we can set one of these parameters to one.
It is convenient to set $\kappa=1$, then the only parameter
is the size of the system $L$.\par
Our next goal is to find chimera states, which consist of synchronous and asynchronous parts. We look for rotating-wave solutions of system \eqref{eq:oa1},{\,}\eqref{eq:difeq2}, which are stationary
in a rotating reference frame:
$Z(x,t)=z(x)e^{i(\omega+\Omega)t}$, $H(x,t)=h(x)e^{i(\omega+\Omega)t}$,
where $\Omega$ is some unknown frequency to be defined below~\cite{bib:cpkbm_note}. Substituting this we get a system of an algebraic equation and an ODE for complex functions $z\!\left(x\right)$ and $h\!\left(x\right)$
\begin{gather}
e^{i\alpha}h^{\ast}z^2+2i\Omega z-e^{-i\alpha}h=0\;,\label{eq:zeq}\\
h''-h=-z\;.\label{eq:heq}
\end{gather}
Here and below prime denotes spatial derivative.
The first step is to express $z\!\left(x\right)$ from the quadratic equation~\eqref{eq:zeq}~(see~\cite{bib:sm}). This equation describes the order parameter $z\!\left(x\right)$ of
a set of oscillators driven by field $h\!\left(x\right)=r\!\left(x\right)e^{i\theta\left(x\right)}$, the solution at each point $x$ depends on the relation between $r$ and $\Omega$ (below for
simplicity of presentation we write the relations for $\Omega<0$). If $\left|r\right|\geq\left|\Omega\right|$, then the oscillators are locked and $|z|=1$, otherwise
the oscillators are partially synchronous with $0<|z|<1$. The solution reads
\begin{equation}
z\!=\!\begin{cases} -\left(i\Omega-\sqrt{r^2\!-\!\Omega^2}\right)r^{-1}e^{i(\theta-\alpha)},\hspace{2mm}\text{if}\hspace{2mm} \left|r\right|\!\geq\!\left|\Omega\right|,\vspace{1.0mm}\\
-i\left(\Omega+\sqrt{\Omega^2\!-\!r^2}\right)r^{-1}e^{i(\theta-\alpha)},\hspace{2mm}\text{if}\hspace{2mm} \left|r\right|\!<\!\left|\Omega\right|.
\end{cases}\label{eq:sas}
\end{equation}
We now substitute this solution in~\eqref{eq:heq}. Although $h\!\left(x\right)$ is complex, the resulting equation can be written,
due to gauge invariance $\theta\!\left(x\right)\!\to\!\theta\!\left(x\right)+\theta_{0}$, as a third-order system of ODEs for real functions $r\left(x\right)$ and $q\left(x\right)\!=\!r^2\left(x\right)\theta'\left(x\right)$
\begin{equation}
\begin{aligned}
r''&=r+r^{-3}q^{2}\!-r^{-1}\sqrt{r^2-\Omega^2}\cos\alpha+r^{-1}\Omega\sin\alpha\;,\\
q'&=\Omega\cos\alpha+\sqrt{r^2-\Omega^2}\sin\alpha\;,
\end{aligned}
\label{eq:rqsyn}
\end{equation}
in the domain where $\left|r\right|\!\geq\!\left|\Omega\right|$, and
\begin{equation}
\begin{aligned}
r''&=r+r^{-3}q^2+r^{-1}(\Omega+\sqrt{\Omega^2-r^2})\sin\alpha\;,\\
q'&=(\Omega+\sqrt{\Omega^2-r^2})\cos\alpha\;,
\end{aligned}
\label{eq:rqasyn}
\end{equation}
in the domain where $\left|r\right|\!<\!\left|\Omega\right|$.
Our goal is to find chimera patterns described by equations~\eqref{eq:rqsyn},{\,}\eqref{eq:rqasyn}, satisfying periodicity
condition $r(x+L)=r(x)$, $q(x+L)=q(x)$. It is more convenient not to fix the period $L$, but to fix the frequency of the rotating
chimera $\Omega$, then to find periodic solutions of \eqref{eq:rqsyn},{\,}\eqref{eq:rqasyn}, period $L$ of which depends on $\Omega$ (see~\cite{bib:sm}). This will
after inversion yield dependence $\Omega(L)$.
\begin{figure}[tb]
\centering
\includegraphics[width=\columnwidth]{SimplestChimeraStates}
\caption{(Color online) Profiles of simplest chimeras (with at most two SRs) for $\alpha\!=\!1.457$, $|z|$: solid lines, $|h|$: dotted lines.
$A$: the KB one-SR chimera for $\Omega\!=\!-0.648$,
$B$: symmetric two-SRs chimera for $\Omega\!=\!-0.558$,
$C$: asymmetric two-SRs chimera (here the widths of synchronous domains are different, and their phases differ not by $\pi$, like in case $B$) for $\Omega\!=\!-0.672$,
$D$: nearly synchronous one-SR chimera for $\Omega\!=\!-0.98762$.
Colors used correspond to coding of solutions in Fig.~\ref{fig:per}.}
\label{fig:ill}
\end{figure}
Before discussing numerical and analytical approaches, we illustrate in
Fig.~\ref{fig:ill} several solutions
for $\alpha=1.457$ (the value used in~\cite{Kuramoto-Battogtokh-02}) with
period $L\approx 11.2$. Presented solutions (types $A$ and $B$ have been already discussed in the
literature~\cite{Kuramoto-Battogtokh-02,Omelchenko-13,Panaggio-Abrams-15})
are just simplest possible chimeras
with at most two synchronous regions (SRs). Indeed,
the system \eqref{eq:rqsyn},{\,}\eqref{eq:rqasyn} is
a reversible (with respect to involution $r\to r$, $q\to -q$) third-order system of ODEs with a plethora of
solutions, including
chaotic ones. We illustrate this by constructing a two-dimensional Poincar\'e map in Fig.~\ref{fig:pm}a. It
shows typical for
nearly integrable Hamiltonian
systems picture of tori and periodic orbits of different periods. Not all points on the Poincar\'e surface lead to
physically meaningful solutions: we discarded the trajectories which resulted in values $\left|r\right|\!>\! 1$. The fixed point of the map Fig.~\ref{fig:pm}(a)
at $q=0$, $r\approx 0.84$ describes the one-hump chimera state $A$ in Fig.~\ref{fig:ill}.
The Poincar\'e map is constructed for a fixed value of $\Omega$, it provides several branches
of periodic orbits having different periods. Collecting solutions at a fixed period $L$, we obtain are many coexisting
chimera patterns; several three-SRs chimeras
are illustrated in Fig.~\ref{fig:pm}b. Our aim
in this study is not to follow
all possible periodic and chaotic solutions of this reversible system, below we focus on the simplest ones
illustrated in Fig.~\ref{fig:ill},
corresponding to fixed
points and period-two orbits of the Poincar\'e map.
\begin{figure}[tb]
\centering
\includegraphics[width=0.7\columnwidth]{pmap_2}
\includegraphics[width=0.7\columnwidth]{cfield_3_1457-4}
\caption{(Color online) (a) Poincar\'e map for system \eqref{eq:rqsyn},{\,}\eqref{eq:rqasyn}
for $\alpha=1.457$ and $\Omega=-0.8$.
The condition for the section:
$r'=0$, $r''<0$. (b) More complex patterns with three SRs
for $L\approx 15.1$ and $\Omega=-0.796$ (solid red line), $\Omega=-0.726$ (dashed blue line) and $\Omega=-0.674$ (dotted black line).}
\label{fig:pm}
\end{figure}
Remarkably, it is possible to describe basic chimera profiles semi-analytically, for
$\alpha\approx \pi\bigl/2\bigr.$.
Let us first consider the limiting case $\alpha=\pi\bigl/2\bigr.$. Here,
according to~\eqref{eq:rqsyn},{\,}\eqref{eq:rqasyn}, the derivative $q'\!\left(x\right)$
is non-negative in the synchronous state and vanishes
in the asynchronous state. Thus, a periodic solution with
$q(x)=q(x+L)$ should be everywhere asynchronous, except possibly for one
or two points at which $r(x)$ achieves an extremum $|r|=|\Omega|$. For this degenerate
chimera Eqs.~\eqref{eq:rqasyn} reduce to $q=0$ and an integrable second-order equation
\begin{equation}
\begin{gathered}
r''=-dU(r)\bigl/dr\bigr.\;,\\
U(r)=-r^2\bigl/2\bigr.-\sqrt{\Omega^2\!-\!r^2}-\Omega\ln\!\left(\sqrt{\Omega^{2}\!-\!r^{2}}\!-\!\Omega\right).
\end{gathered}
\label{eq:pot}
\end{equation}
In the potential $U\!\left(r\right)$, there are two types of trajectories, having maximum at $r_{max}=|\Omega|$, depending on the value of $\Omega$.
For $-1<\Omega<\Omega_{\ast}=2\left(\ln2-1\right)$ this is a periodic orbit with $0<r_{min}\leq |\Omega|$.
It reaches the boundary of the asynchronous region at one point and corresponds
to ``one-point chimera'', which can be considered as the limiting case
of curve $A$ in Fig.~\ref{fig:ill}, where the SR shrinks to a point.
For $\Omega_{\ast}<\Omega<0$ there is a symmetric periodic orbit (here it is convenient
to allow $r$ to change sign, this corresponds to a jump by $\pi$ in phase $\theta$
if $r$ is considered as positive like in Fig.~\ref{fig:ill}, curve $B$) with
$-|\Omega|\leq r \leq |\Omega|$. This ``two-point chimera'' corresponds to curve $B$ in
Fig.~\ref{fig:ill}. These two types of solutions merge in a homoclinic orbit with
infinite period at $\Omega=\Omega_{\ast}$, which can be named ``chimera soliton'' (one- or two-point,
depending on which side of the threshold the orbit is considered). The dependencies $\Omega\!\left(L\right)$
for these solutions are shown in Fig.~\ref{fig:per} as solid lines.
Note that, additionally, there is
a branch of synchronous solutions with $\Omega=-1$ which are steady states $r=1$.
The solutions above are degenerate chimeras, as the SR is
restricted to one or two points. Synchronous region becomes finite for $\alpha\lesssim \pi\bigl/2\bigr.$,
here one can develop a perturbation approach by introducing a
small parameter $\beta=\pi\bigl/2\bigr.-\alpha\ll 1$. Now $q\neq 0$, but because $q\sim \beta$,
we can neglect terms $\sim q^2$ in \eqref{eq:rqsyn},{\,}\eqref{eq:rqasyn}.
Then, the problem reduces to finding a periodic trajectory $r(x)$ of integrable equation, such that evolution of $q(x)$ is periodic:
\begin{equation}
\begin{gathered}
\hspace{-6.0mm}
\oint\!q'\!\left(x\right)dx \!=\!\!\int_{x:\,\left|r\left(x\right)\right|\geq\left|\Omega\right|}\! \left(\Omega\beta\!+\!\sqrt{r(x)^2\!-\!\Omega^2}\right)dx+{}\\{}\hspace{14.5mm}
\int_{x:\,\left|r\left(x\right)\right|\geq\left|\Omega\right|}\! \beta\!\left(\Omega\!+\!\sqrt{\Omega^{2}\!-\!r^{2}\!\left(x\right)}\right)dx\!=\!0\;.
\end{gathered}
\label{eq:intq}
\end{equation}
Detailed calculations are presented in \cite{bib:sm}. The result is that the size $L_{syn}$ of SR
becomes finite:
\begin{equation}
L_{syn}\approx\sqrt{\frac{8\beta}{\pi N_{SR}\sqrt{|\Omega|(1-|\Omega|)}}\oint\!\left(R'^{2}+R^{2}\right)dx}\;.
\label{eq:chreg}
\end{equation}
where $R(x)$ is the solution of \eqref{eq:pot} at $\beta=0$ and $N_{SR}$ is the number of SRs.
We compare the analytical approach above with the results
of direct numerical calculation, in the framework of \eqref{eq:rqsyn},{\,}\eqref{eq:rqasyn},
of periodic orbits in Fig.~\ref{fig:per},
for several values of $\alpha$. Panel (a) shows that for small $\beta$ chimera states (of types $A$ and $B$
of Fig.~\ref{fig:ill}) are close to
degenerate regimes at $\beta=0$. One can see in panels (a,b) that the two analytic
solutions at $\alpha=\pi\bigl/2\bigr.$ (the one-point chimera and the
synchronous state) merge into one branch at $\alpha\lesssim \pi\bigl/2\bigr.$ with a nonmonotonous
dependence $\Omega$ on $L$, cf. one-SR chimeras $A$ and $D$ in Fig.~\ref{fig:ill}. In panel (b) one can see
an additional branch corresponding to the two-SRs asymmetric chimera $C$ in Fig.~\ref{fig:ill}.
As a result, in (b) and (c) one has four solutions in some range of periods $L$.
Only two of them survive for small $\alpha$;
diagrams for $\alpha<0.9$ are qualitatively the same as panel (d) in Fig.~\ref{fig:per}.
\begin{figure}[tb]
\centering
\includegraphics[width=\columnwidth]{AlphaEllOmegaDependence}
\caption{(Color online) Periods of chimera states $L$ vs. parameter $\Omega$ for $\alpha=1.514$ (a),
$\alpha=1.457$ (b), $\alpha=1.229$ (c), and $\alpha=0.944$ (d).
Chimera states for $\alpha=\pi\bigl/2\bigr.$, obtained by integration
Eq.~\eqref{eq:pot}, are shown with violet solid lines. Different
markers correspond to chimera types depicted in Fig.~\ref{fig:ill}, as specified in panel (b). Cyan dashed lines
show frequency of the synchronous state $\Omega=-\sin\alpha$.}
\label{fig:per}
\end{figure}
\looseness=-1
Next, we discussed stability of the obtained chimera patterns.
For this goal we linearize Eq.~\eqref{eq:oa1},{\,}\eqref{eq:kernel} (see~\cite{bib:sm}).
Contrary to the problem of finding chimera solutions, this analysis cannot be reduced
to that of differential equations, rather we have to consider an integral-differential
equation~\eqref{eq:oa1},{\,}\eqref{eq:kernel} for
$Z(x,t)$. After spatial discretization we get a matrix eigenvalue problem. The difficulty here
is that, according to~\cite{Omelchenko-13,Xie_etal-14}, there is an essential continuous
$T$-shaped spectrum $\lambda_c$ consisting of
eigenvalues on the imaginary and the negative real axis, but stability is determined by the point spectrum
$\lambda_{p}$.
Unfortunately, it is not easy to discriminate these parts of the spectrum in the eigenvalues $\lambda$
of the approximate matrix, because the eigenvalues representing essential part of spectrum lie
not exactly on the imaginary axis.
We adopted the following procedure to select the point spectrum $\lambda_{p}$. For a chimera state in the domain
$0\leq x\leq L$, we can discretize the linearized system by using a set of points
$x_0+j \Delta$, \mbox{$j=0,1,\ldots,M-1$}, where $\Delta=L/M$ and $0\leq x_0\leq \Delta$
is an arbitrary continuous parameter.
This leads to an $2M\times 2M$ real matrix, eigenvalues $\lambda$
of which we obtained numerically.
Additionally, we varied the offset of discretization $x_0$. In numerics we used $M=2048$
and $N=64$ or $N=128$ equidistant values of $x_0$.
We have found that while the components of the essential
spectrum vary with $x_0$, the point spectrum $\lambda_{p}$ components vary extremely weakly with $x_0$ -- this allowed us to determine
point spectrum $\lambda_{p}$ reliably for most values of parameters.
\begin{figure}[tb]
\centering
\includegraphics[width=\columnwidth]{DiscreteSpectrum1}
\caption{(Color online) (a-d): Essential (blue markers) and point (red markers) spectra for chimera states at $\alpha=1.457$ and
four values of $\Omega$: (a) $\Omega=0.45$, (b) $\Omega=0.675$, (c) $\Omega=0.8$, (d) $\Omega=0.95$. In these plots all $2MN$ eigenvalues
with $M=2048$ and $N=128$ are plotted. (e,f): real and imaginary parts of point spectrum $\lambda_{p}$ for solutions $A,D$ (red circles)
and $B$ (blue diamonds) of Fig.~\ref{fig:per}(b).}
\label{fig:spec}
\end{figure}
Below we present stability analysis for $\alpha=1.457$, for branches $A,B,C,D$ (see Fig.~\ref{fig:per}(b)).
Four characteristic types of spectra are shown
in Fig.~\ref{fig:spec}(a-d). Only case (c) with point spectrum $\lambda_{p}$ having negative real part corresponds to a stable
chimera pattern, while all other patterns are unstable (oscillatory instability for cases (a,b) and monotonous instability
for case (d)). Dependence of the point spectrum $\lambda_{p}$ on parameter $\Omega$ for $\alpha=1.457$, branches $A,B,D$, is
shown in Fig.~\ref{fig:spec}(e,f). One can see that in the region $-0.68\lesssim \Omega \lesssim-0.64$ there are four points
of $\lambda_{p}$, for other values of $\Omega$ there is only one pair of eigenvalues
(or one real eigenvalue for branch $D$). This property may be attributed to the fact, that close to
the homoclinic orbit $\Omega\approx\Omega_{\ast}$ the length of the patterns is large so here two discrete modes
are possible. The only stable chimera state is of type $A$ (we refer here to Figs.~\ref{fig:ill} and \ref{fig:per}(b))
with $-0.91\lesssim \Omega \lesssim-0.69$.
On the contrary, chimera states with two SRs (type $B$) are unstable. Most difficult was analysis of two-SRs solutions of type~$C$~(Fig.~\ref{fig:dspec}), here the unstable branch of point spectrum $\lambda_{p}$ is real, and there is up to three stable complex pairs. In some cases only very fine discretization
with $M=6144$ allowed us to reveal unstable point eigenvalues~$\lambda_{p}$; we attribute this to a complex profile of this solution, requiring high resolution of perturbations.
\begin{figure}[tb]
\centering
\includegraphics[width=\columnwidth]{DiscreteSpectrum2}
\caption{(Color online) All eigenvalues [(a): $\Omega=-0.645$, (b): $\Omega=-0.735$, here $M=4096$, $N=64$]
and point spectra in dependence on $\Omega$ (c,d) for the asymmetric
branch $C$.}
\label{fig:dspec}
\end{figure}
Stability properties have been confirmed
by direct numerical simulations of the ensemble governed by Eq.~\eqref{eq:kbmodel},{\,}\eqref{eq:kernel},
see Fig.~\ref{fig:simul} for space-time plots of field $|H(k,t)|=|\sum_{j}G(|k-j|/KL)\exp[i\phi_j]|$. One can initialize
all the chimera patterns found above; in the unstable regions these patterns are destroyed, while stable chimera persists.
Remarkably, for weakly unstable two-SRs chimeras for $\Omega\approx -0.58$, where the real part of the point eigenvalue $\lambda_{p}$
has a minimum (see Fig.~\ref{fig:spec}(e)), the life time of prepared chimera is relatively large.
\begin{figure}[tb]
\centering
\includegraphics[width=\columnwidth]{DynamicsOfPhaseOscillators}
\caption{\looseness=-1 (Color online) Direct numerical simulations of stable and unstable chimeras for $\alpha\!=\!1.457$.
(a): Chimera of type $A$ for
$\Omega=-0.71377$, $L=6.03$;
(b): chimera of type $B$ for $\Omega=-0.575$, $L=12.06$; (c): chimera of type C, $\Omega=-0.6$, $L\!=\!12.06$.
The number of oscillators was $K\!=\!700$ per length unit.}
\label{fig:simul}
\end{figure}
Summarizing, in this Letter we reformulated the problem of chimera patterns in 1D medium
of coupled oscillators
as a system of PDEs. This allowed to find uniformly rotating chimera states
as solutions of an ODE. Although a large variety of patterns with large spatial
periods can be found, we restricted our attention in this Letter to the simplest ones, with at most
two synchronous domains.
Remarkably, these profiles can be explicitly described in the limit of neutral coupling between oscillators;
for coupling close to neutral one, a perturbation analysis yields approximate solutions. Exploring stability of the
found solutions appeared to be a nontrivial numerical problem. We suggested an approach to characterize
the essential and the point parts of the spectrum via finite discretizations. It appears that only chimeras
of the type originally studied by Kuramoto and Battogtokh are stable, while other found patterns are
linearly unstable.
The approach suggested could be extended in several directions. First, one can study general bifurcations
of chimera patterns. The difficulty here is that many tools for bifurcation analysis require sufficient smoothness
of the equations, what is not the case for chimera solutions. Stability analysis performed in this letter have been
restricted to perturbations with the same spatial period as the chimera itself, i.e. it describes stability for a medium
on a circle. Other unstable modes, e.g. of modulational instability type,
could appear if one formulates the stability problem for an infinite medium. Finally, the formulated PDEs
have been simplified using the separation of time scales; it would be interesting to study stability of chimeras in the full
Eqs.~\eqref{eq:oa1},{\,}\eqref{eq:difeq1} with $\tau\neq 0$.\par
\acknowledgments
We acknowledge discussions with O.~Omelchenko, M.~Wolfrum, and Yu.~Maistrenko.
L.\,S. was supported by ITN COSMOS (funded by
the European Union’s Horizon 2020 research and innovation
programme under the Marie Sklodowska-Curie grant agreement No 642563).
Numerical part of this work was
supported by the Russian Science Foundation
(Project No. 14-12-00811).\newpage
|
\section{Introduction}
The past two decades have seen tremendous progress in the description
of QCD with functional approaches such as the
functional renormalisation group (FRG), Dyson-Schwinger equations
(DSE), and n-particle irreducible methods ($n$PI). These approaches
constitute \emph{ab initio} descriptions of QCD in terms of quark and
gluon correlation functions. The full correlation functions satisfy a
hierarchy of loop equations that are derived from the functional FRG,
DSE and $n$PI relations for the respective generating functionals. By
now, systematic computational schemes are available, which can be
controlled by apparent convergence. In the present work on pure
Yang-Mills (YM) theory we complement the work in quenched QCD
\cite{Mitter:2014wpa}, where such a systematic expansion scheme has
been put forward within the FRG. Equipped with such a controlled
expansion, functional approaches to QCD are specifically interesting
at finite temperatures and large density, where reliable \emph{ab
initio} theoretical predictions and experimental results are missing
at present.
Most progress with functional approaches has been made in Landau gauge
QCD, which has many convenient properties for non-perturbative
numerical computations. Applications of functional methods include the
first-ever calculation of qualitative non-perturbative Landau gauge
propagators as well as investigations of the phase structure of
QCD. For reviews see
\cite{Berges:2000ew,Roberts:2000aa,Alkofer:2000wg,Pawlowski:2005xe,
Fischer:2006ub,Gies:2006wv,Schaefer:2006sr,Fischer:2008uz,Binosi:2009qm,%
Braun:2011pp,Maas:2011se,Sanchis-Alepuz:2015tha},
for applications to Yang-Mills theory see e.g.\
\cite{Ellwanger:1995qf,vonSmekal:1997ohs,Bergerhoff:1997cv,Gies:2002af,%
Pawlowski:2003hq,Fischer:2003rp,Fischer:2004uk,Aguilar:2008xm,%
Boucaud:2008ky,Tissier:2010ts,Quandt:2013wna,Quandt:2015aaa,Huber:2016tvc,Quandt:2016ykm},
and e.g.\ \cite{Feuchter:2004mk} for related studies in the Coulomb
gauge. The formal, algebraic, and numerical progress of the past
decades sets the stage for a systematic vertex expansion scheme of
Landau gauge QCD. Quantitative reliability is then obtained with
apparent convergence \cite{Mitter:2014wpa} as well as further
systematic error controls inherent to the method, see e.g.\
\cite{Litim:2000ci,Litim:2001up,Pawlowski:2005xe,Schnoerr:2013bk,Pawlowski:2015mlf}. In the aforementioned quenched QCD
investigation \cite{Mitter:2014wpa}, the gluon propagator was taken
from a separate FRG calculation in \cite{Fischer:2008uz,FPun}. This
gluon propagator shows quantitative agreement with the lattice
results, but has been obtained within an incomplete vertex expansion
scheme. Therefore, the results \cite{Fischer:2008uz,FPun} for the YM
correlations functions give no access to systematic error estimates.
In general, many applications of functional methods to bound states
and the QCD phase diagram use such mixed approaches, where part of the
correlation functions is deduced from phenomenological constraints or
other external input. Despite the huge success of mixed approaches, a
full \emph{ab initio} method is wanted for some of the most pressing
open questions of strongly-interacting matter. The phase structure of
QCD at large density is dominated by fluctuations and even a partial
phenomenological parameter fixing at vanishing density is bound to
lead to large systematic errors \cite{Helmboldt:2014iya}. The same
applies to the details of the hadron spectrum, in particular with
regard to the physics of the higher resonances, which requires
knowledge about correlation functions deep in the complex plane.
In the present work we perform a systematic vertex expansion of the
effective action of Landau gauge YM theory within the functional
renormalisation group approach, discussed in \Sec{sec:setup}. The
current approximation is summarised in \Sec{sec:expsch}, which also
includes a comparison to approximations used in other works. This
\emph{ab initio} approach starts from the classical action. Therefore
the only parameter is the strong coupling constant $\alpha_s$ at a
large, perturbative momentum scale. The most distinct feature of YM
theory is confinement, which is reflected by the creation of a gluon
mass gap in Landau gauge. We discuss the necessity of consistent
infrared irregularities as well as mechanisms for the generation of a
mass gap in \Sec{sec:massgap}. Numerical results from a
parameter-free, self-consistent calculation of propagators and
vertices are presented in \Sec{sec:mainresult}. Particular focus is
put on the importance of an accurate renormalisation of the relevant
vertices. We compare with corresponding DSE and lattice results and
discuss the apparent convergence of the vertex expansion. Finally, we
present numerical evidence for the dynamic mass gap generation in our
calculation. Further details, including a thorough discussion of the
necessary irregularities, can be found in the appendices.
\section{FRG flows for Yang-Mills theory in a vertex expansion}
\label{sec:setup}
Functional approaches to QCD and Yang-Mills theory are based on the
classical gauge fixed action of $SU(3)$ Yang-Mills theory. In general
covariant gauges in four dimensions it is given by
\begin{align}
S_{\rm cl}=\int_x\,\014 F_{\mu\nu}^a F_{\mu\nu}^a +\frac{1}{2\xi}\int_x\,(
\partial_\mu A^a_\mu)^2 -\int_x\,\bar c^a \partial_\mu D^{ab}_\mu
c^b\,.
\label{eq:sclassical}
\end{align}
Here, $\xi$ denotes the gauge fixing
parameter, which is taken to zero in the Landau gauge and $\int_x=\int \text{d}^4 x$.
The field
strength tensor and covariant derivative are given by
\begin{eqnarray}
F^a_{\mu\nu} &=& \partial_\mu A^a_\nu-\partial_\nu A^a_\mu+
g f^{abc}A_\mu^b A_\nu^c\, ,\nonumber \\[2ex]
D^{ab}_{\mu} & = &\delta^{ab}\partial_\mu-g f^{abc} A^c_\mu\,,
\end{eqnarray}
using the fundamental generators $T^a$, defined by
\begin{eqnarray}
\left[T^a,T^b\right] & =& if^{abc}T^c\,,\qquad
\text{tr}\left(T^aT^b\right) = \frac{1}{2}\delta^{ab}\,.
\end{eqnarray}
In general, our notation follows the one used in the works
\cite{Mitter:2014wpa,Braun:2014ata,Rennecke:2015eba} of the fQCD
collaboration \cite{fQCD}.
\subsection{Functional Renormalisation Group}
We use the functional renormalisation group approach as a
non-perturbative tool to investigate Yang-Mills theory. The FRG is
built on a flow equation for the one-particle irreducible ($1$PI)
effective action or free energy of the theory, the Wetterich equation
\cite{Wetterich:1992yh}. It is based on Wilson's idea of introducing
an infrared momentum cutoff scale $k$. Here, this infrared
regularisation of the gluon and ghost fluctuations is achieved by
modifying the action $S_{\rm cl} \to S_{\rm cl}+\Delta S_k$ with
\begin{align}
\label{eq:dSk}
\Delta S_k=\int_x\,
\012 A_\mu^a\, R_{k,\mu\nu}^{ab} \, A_\nu^b
+ \int_x\,\bar c^a\, R^{ab}_k\, c^b\,.
\end{align}
The regulator functions $R_k$ are momentum-dependent masses that
suppress the corresponding fluctuations below momentum scales
$p^2\approx k^2$ and vanish in the ultraviolet for momenta $p^2\gg
k^2$. See \App{app:regulator} for details on the regulators used
in the present work.
Consequently, the effective action, $\Gamma_k[\phi]\,$, is
infrared regularised, where $\phi$ denotes the superfield
\begin{align}
\phi=(A, c,\bar c)\,.
\end{align}
The fluctuations of the theory are then successively taken into
account by integrating the flow equation for the effective action, see
e.g.\ \cite{Fister:2011uw,Fister:2013bh},
\begin{align}
{\partial_t}\Gamma_k[\phi] = \int_p \,\012 \ G_{k,ab}^{\mu\nu}[\phi]
\,{\partial_t}R_{k,\mu\nu}^{ba} -\int_p
\,G^{ab}_k[\phi]\,{\partial_t}R^{ba}_k\,.
\label{eq:flow}
\end{align}
where $\int_p=\int \text{d}^4p/(2 \pi)^4\,$. Here we have introduced the
RG-time $t=\ln (k/\Lambda)$ with a reference scale $\Lambda\,$, which is
typically chosen as the initial UV cutoff $\Lambda\,$. Although this
flow equation comes in a simple one-loop form, it provides an exact
relation due to the presence of the full field-dependent propagator,
\begin{align}
G_k[\phi](p,q)= \01{\Gamma_k^{(2)}[\phi]+R_k }(p,q) \,,
\end{align}
on its right-hand side. Furthermore, the flow is infrared and
ultraviolet finite by construction. Via the integration of momentum
shells in the Wilsonian sense, it connects the ultraviolet, bare
action $S_{\rm cl}=\Gamma_{k\rightarrow\Lambda\rightarrow\infty}$ with
the full quantum effective action $\Gamma=\Gamma_{k\rightarrow0}\,$.
The flow equations for propagators and vertices are obtained by taking
functional derivatives of \eq{eq:flow}. At the vacuum expectation
values, these derivatives give equations for the $1$PI correlation
functions $\Gamma_k^{(n)}=\delta^n\Gamma_k/\delta\phi^n\,$, which
inherit the one-loop structure of \eq{eq:flow}. As the
cutoff-derivative of the regulator functions, $\partial_t R_k$, decays
sufficiently fast for large momenta, the momentum integration in
\eq{eq:flow} effectively receives only contributions for momenta
$p^2\lesssim k^2\,$. Furthermore, the flow depends solely on dressed
vertices and propagators, leading to a consistent RG and momentum
scaling for each diagram resulting from derivatives of \eq{eq:flow}.
Despite its simple structure, the resulting system of equations does
not close at a finite number of correlation functions. In general,
higher derivatives up to the order $\Gamma_k^{(n+2)}$ of the effective
action appear on the right hand side of the functional relations for
the correlation functions $\Gamma_k^{(n)}\,$.
\subsection{Vertex expansion of the effective action}
\label{sec:expsch}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{ym_quantitative_truncation_class}
\caption{ Approximation for the effective action. Only
classical tensor structures are included. See
\Fig{fig:diagrams} for diagrams that contribute to the
individual propagators and vertices. }
\label{fig:eff_action}
\end{figure}
\begin{figure}
\hspace*{-0.1cm}\includegraphics[width=0.49\textwidth]{ym_flow_equations}
\caption{Diagrams that contribute to the truncated flow of
propagators and vertices. Wiggly (dotted) lines correspond
to dressed gluon (ghost) propagators, filled circles denote
dressed ($1$PI) vertices and regulator insertions are
represented by crossed circles. Distinct permutations
include not only (anti-)symmetric permutations of external
legs but also permutations of the regulator insertions. }
\label{fig:diagrams}
\end{figure}
The structural form of the functional equations discussed in the
previous section necessitates the use of approximations in most
practical application. One systematic expansion scheme is the vertex
expansion, i.e.\ an expansion of the effective action in terms of
$1$PI Green's functions. This yields an infinite tower of coupled
equations for the correlation functions that has to be truncated at a
finite order. This expansion scheme allows a systematic error estimate
in terms of apparent convergence upon increasing the expansion order
or improving further approximations for example in the momentum
resolution or tensor structures of the included correlation functions.
We discuss the convergence of the vertex expansion in
\Sec{sec:truncationcheck}.
Here we calculate the effective action of $SU(3)$ Yang-Mills theory in
Landau gauge within a vertex expansion, see \Fig{fig:eff_action} for a
pictorial representation. The diagrams contributing to the resulting
equations of the constituents of our vertex expansion are summarised
graphically in \Fig{fig:diagrams}. The lowest order contributions
in this expansion are the inverse gluon and ghost propagators
parameterised via
\begin{align}
[\Gamma^{(2)}_{AA}]^{ab}_{\mu\nu}(p) &= Z_A(p)\, p^2\, \delta^{ab}\,
\Pi^{\bot}_{\mu\nu}(p) +\01\xi \delta^{ab} p_\mu p_\nu\ ,
\nonumber\\[2ex]
[\Gamma^{(2)}_{\bar c c}]^{ab}(p) &= Z_c(p)\, p^2\, \delta^{ab} \,,
\label{eq:propagators}
\end{align}
with dimensionless scalar dressing functions $1/Z_A$ and $1/Z_c$. Here,
$\Pi^{\bot}_{\mu\nu}(p)=\delta_{\mu\nu}-p_\mu p_\nu/p^2$ denotes the
corresponding projection operator.
We use this
splitting in tensor structures with canonical momentum dimension and
dimensionless dressings also for the higher order vertices.
On the three-point level we include the full transverse ghost-gluon
vertex and the classical tensor structure of the three-gluon vertex
\begin{align}
[\Gamma^{(3)}_{A \bar c c}]^{abc}_\mu (p,q) &= Z_{A \bar c c,\bot}(|p|,|q|,t)
[\mathcal{T}_{A \bar c c,{\rm cl}}]^{abc}_{\mu}(p,q)\,,\nonumber\\[2ex]
[\Gamma^{(3)}_{A^3}]^{abc}_{\mu\nu\rho} (p,q) &= Z_{A^3,\bot}(|p|,|q|,t)
[\mathcal{T}_{A^3,{\rm cl}}]^{abc}_{\mu\nu\rho}(p,q)\,.
\label{eq:threepoint}
\end{align}
Here, the momentum $p$ $(q)$ corresponds to the indices $a$ $(b)$ and
$t$ denotes the cosine of the angle between the momenta $p$ and $q\,$.
The classical tensor structure of the vertices has been summarised as
$\mathcal{T}_{A^3,{\rm cl}}$ and $\mathcal{T}_{A \bar c c,{\rm cl}}\,$,
which are listed explicitly in \App{app:tensorstructures}. In the
case of the transversally projected ghost-gluon vertex,
$\mathcal{T}_{A \bar c c,{\rm cl}}$ represents already a full basis
whereas a full basis for the transversally projected three-gluon
vertex consists of four elements. However, the effect of non-classical
tensor structures has been found to be subleading in this case
\cite{Eichmann:2014xya}.
The most important four-point function is given by the four-gluon
vertex, which appears already on the classical level. Similarly to the
three-gluon vertex, we approximate it with its classical tensor
structure
\begin{align}
[\Gamma_{A^4}^{(4)}]^{abcd}_{\mu\nu\rho\sigma}(p,q,r) &=
Z_{A^4,\bot}(\bar{p})
[\mathcal{T}_{A^4,{\rm cl}}]^{abcd}_{\mu\nu\rho\sigma}(p,q,r)\,,
\label{eq:fourgluon}
\end{align}
see \App{app:tensorstructures} for details. The dressing function of
the four-gluon vertex is approximated from its momentum dependence at
the symmetric point via the average momentum $\bar{p}\equiv
\sqrt{p^2+q^2+r^2+(p+q+r)^2}/2\,$, which has been shown to be a good
approximation of the full momentum dependence
\cite{Cyrol:2014kca,Cyrol:2014mt}. To improve this approximation
further, we additionally calculate the momentum dependence of the
four-gluon dressing function $Z_{A^4,\bot}(|p|,|q|,t)$ on the special
configuration $(p,q,r)=(p,q,-p)\,$. We use this special configuration
exclusively in the tadpole diagram of the gluon propagator equation,
cf.\ \Sec{sec:truncationcheck}. We show the difference between the
special configuration and the symmetric point approximation in the
appendix in \Fig{fig:fourGluonTadpoleDressing}. Although the
four-gluon vertex has been the subject of several studies
\cite{Kellermann:2008iw,Binosi:2014kka,Gracey:2014ola,Cyrol:2014kca,Cyrol:2014mt},
no fully conclusive statements about the importance of additional
non-classical tensors structures are available.
In summary we have taken into account the propagators and the fully
momentum-dependent classical tensor structures of the three-point
functions, as well as selected momentum-configurations of the gluon
four-point function, see the paragraph above, and
\App{app:tensorstructures}. For a comparison of the current approximation with that
used in other functional works one has to keep in mind that FRG,
Dyson-Schwinger or $n$PI equations implement different resummation
schemes. Thus, even on an identical approximation level of a
systematic vertex expansion, the included resummations differ.
In the present work we solve the coupled system of all
momentum-dependent classical vertex structures and propagators. In former
works with functional methods, see e.g.\
\cite{Ellwanger:1995qf,vonSmekal:1997ohs,Bergerhoff:1997cv,Gies:2002af,%
Pawlowski:2003hq,Fischer:2003rp,Fischer:2004uk,Kellermann:2008iw,Aguilar:2008xm,%
Boucaud:2008ky,Tissier:2010ts,Huber:2012kd,Aguilar:2013xqa,Pelaez:2013cpa,Blum:2014gna,Eichmann:2014xya,%
Gracey:2014mpa,Gracey:2014ola,Huber:2014isa,Williams:2015cvx,Binosi:2014kka,Cyrol:2014kca,Cyrol:2014mt},
only subsets of these correlation functions have
been coupled back. A notable exception is \cite{Huber:2016tvc}, where
a similar self-consistent approximation has been used for three-dimensional
Yang-Mills theory.
\subsection{Modified Slavnov-Taylor identities and transversality in
Landau gauge}
\label{sec:mSTIandVert_sub}
In Landau gauge, the dynamical system of correlation functions
consists only of the transversally projected correlators
\cite{Fischer:2008uz}. Those with at least one longitudinal gluon leg
do not feed back into the dynamics. To make these statements precise,
it is useful to split correlation functions into purely transverse
components and their complement with at least one longitudinal gluon
leg. The purely transverse vertices $\Gamma^{(n)}_{\bot}\,$, are defined
by attaching transverse projection operators to the corresponding
gluon legs,
\begin{align}
\label{eq:purely_transverse}
&\left[\Gamma^{(n)}_{\bot}\right]_{\mu_1\cdots\mu_{n_A}} \equiv
\Pi^{\bot}_{\mu_1\nu_1} \cdots \Pi^{\bot}_{\mu_{n_A}\nu_{n_A}}
\left[\Gamma^{(n)}\right]_{\nu_1\cdots\nu_{n_A}}\,,
\end{align}
where $n_A$ is the number of gluon legs and group indices and momentum
arguments have been suppressed for the sake of brevity. This defines
a unique decomposition of $n$-point functions into
\begin{align}
\label{eq:longitudinal}
\Gamma^{(n)}
=\Gamma^{(n)}_{\bot}+\Gamma^{(n)}_{\text{L}}\,,
\end{align}
where the longitudinal vertices $\Gamma^{(n)}_{\text{L}}$, have at
least one longitudinal gluon leg. Consequently, they are always
projected to zero by the purely transverse projection operators of
\eq{eq:purely_transverse}.
Functional equations for the transverse correlation functions close in
the Landau gauge, leading to the structure \cite{Fischer:2008uz},
\begin{align}\label{eq:closedFun}
\Gamma^{(n)}_{\bot}
={\rm Diag}[\{\Gamma^{(n)}_{\bot}\}]\,.
\end{align}
In \eq{eq:closedFun} Diag stands for diagrammatic expressions of
either integrated FRG, Dyson-Schwinger or $n$PI equations. Equation
\eq{eq:closedFun} follows from the fact that all internal legs are
transversally projected by the Landau gauge gluon propagator. Hence,
by using transverse projections for the external legs one obtains
\eq{eq:closedFun}. In contradistinction to this, the functional
equations for the vertices with at least one longitudinal gluon leg,
$\Gamma_{\rm L}^{(n)}$, are of the form
\begin{align}
\label{eq:closedFunL}
\Gamma^{(n)}_{\rm L}
={\rm Diag}[\{\Gamma^{(n)}_{\rm L}\}, \{\Gamma^{(n)}_{\bot}\}]\,.
\end{align}
In other words, the solution of the functional equations
\eq{eq:closedFunL} for $\Gamma^{(n)}_{\rm L}$ requires also the
solution of the transverse set of equations \eq{eq:closedFun}.
In the present setting, gauge invariance is encoded in modified
Slavnov-Taylor identities (mSTIs) and Ward-Takahashi identities
(mWTIs). They are derived from the standard Slavnov-Taylor identities
(STIs) by including the gauge or BRST variations of the regulator
terms, see
\cite{Ellwanger:1994iz,Ellwanger:1995qf,D'Attanasio:1996jd,Igarashi:2001mf,
Pawlowski:2005xe,Igarashi:2016gcf} for details. The mSTIs are of the
schematic form
\begin{align}
\label{eq:longFun}
\Gamma_{\rm L}^{(n)}&={\rm mSTI} [\{\Gamma_{\rm L}^{(n)}\}\,,\,
\{\Gamma_{\bot}^{(n)}\}\,,\,R_k]\,,
\end{align}
which reduce to the standard STIs in the limit of vanishing regulator,
$R_k\equiv 0$. The STIs and mSTIs have a similar structure as
\eq{eq:closedFunL} and can be used to obtain information about the
longitudinal part of the correlators. Alternatively, they provide a
non-trivial consistency check for approximate solutions of
\eq{eq:closedFunL}.
\subsubsection*{Consequences of the STIs \& mSTIs}
\label{sec:STImSTI}
For the purposes of this work, the most important effect of the
modification of the STIs due to the regulator term is that it leads to
a non-vanishing gluon mass parameter \cite{Ellwanger:1994iz},
\begin{align}
\Delta_{\rm mSTI} \left[
\Gamma_{AA}^{(2)}\right]_{\mu\nu}^{ab}\varpropto \delta^{ab}\,
\delta_{\mu\nu}\, \alpha(k)\, k^2\,.
\label{eq:mSTI_mass}
\end{align}
At $k=0$, where the regulators vanish, this modification disappears,
as the mSTIs reduce to the standard STIs. In particular, this entails
that, at $k=0$, the inverse longitudinal gluon propagator,
$\Gamma_{AA,\rm L}^{(2)}$, reduces to the classical one, solely
determined by the gauge fixing term
\begin{align}
\label{eq:LSTI}
p_\mu \left( [{\Gamma_{AA,\rm L}^{(2)}}]_{\mu\nu}^{ab}(p) - [{S_{AA,\rm
L}^{(2)}}]_{\mu\nu}^{ab}(p)\right) = 0\,.
\end{align}
This provides a unique condition for
determining the value of the gluon mass parameter \eq{eq:mSTI_mass}
at the ultraviolet initial scale $\Lambda$. However, it can
only serve its purpose, if the longitudinal system
is additionally solved.
One further conclusion from \eq{eq:longFun} is that the mSTIs do not
constrain the transverse correlation functions without further
input. This fact is not in tension with one of the main applications
of STIs in perturbation theory, i.e.\ relating the running of the
relevant vertices of Yang-Mills theory that require renormalisation.
As Yang-Mills theory is renormalisable, only the classical vertex
structures are renormalised and hence the renormalisation functions of
their transverse and longitudinal parts have to be identical.
As an instructive example we consider the ghost-gluon vertex. For this
example and the following discussions we evaluate the STIs within the
approximation used in the present work: on the right hand side of the
STIs we only consider contributions from the primitively divergent
vertices. In particular, this excludes contributions from the
two-ghost--two-gluon vertex. The ghost-gluon vertex can be
parameterised with two tensor structures,
\begin{align}
[\Gamma^{(3)}_{A\bar c c}]_{\mu}^{abc}(p,q) = \text{i} f^{abc}\Bigl[
q_\mu Z_{A \bar cc,\rm cl}(p,q) + p_\mu Z_{A\bar c c,\rm ncl}
(p,q)\Bigr]\,.
\label{eq:classquantsplit}
\end{align}
In \eq{eq:classquantsplit} we have introduced two dressing functions
$Z_{A \bar cc, \rm cl}$ and $Z_{A \bar cc,\rm ncl}$ as functions of
the gluon momentum $p$ and anti-ghost momentum $q\,$. In a general
covariant gauge only $Z_{A \bar cc,\rm cl}$ requires
renormalisation. Similar splittings into a classical tensor structure
and the rest can be used in other vertices. Trivially, this property
relates the perturbative RG-running of the transverse and longitudinal
projections of the classical tensor structures. Then, the STIs can be
used to determine the perturbative RG-running of the classical tensor
structures, leading to the well-known perturbative relations
\begin{equation}
\frac{Z_{A\bar c c, \rm cl}^2}{Z_c^2 Z_A}=
\frac{Z^2_{A^3,\rm cl}}{Z_A^3}=\frac{Z_{A^4,\rm cl}}{Z_A^2}\,,
\label{eq:RGrel}
\end{equation}
at the renormalisation scale $\mu$. Consequently, \eq{eq:RGrel} allows
for the definition of a unique renormalised two-loop coupling
$\alpha_s(\mu)$ from the vertices.
The momentum dependent STIs can also be used to make the relation
\eq{eq:RGrel} momentum-dependent. Keeping only the classical tensor
structures, we are led to the momentum dependent running couplings
\begin{align}
\alpha_{A\bar c c}(p) &= \0{1}{4 \pi}\,\frac{Z_{A\bar cc,\bot }^2(p)}
{ Z_A(p)\,Z_c^2(p)}\,,\nonumber\\[2ex]
\alpha_{A^3}(p) &= \0{1}{4 \pi} \,\frac{Z_{A^3,\bot }^2(p)}
{Z_A^3(p)}\ ,\nonumber\\[2ex]
\alpha_{A^4}(p) &= \0{1}{4 \pi}\, \frac{Z_{A^4,\bot }(p)}{
Z_A^2(p)}\,,
\label{eq:runcoup}
\end{align}
where the used transverse projection is indicated by the subscript
$\bot$, for details see \App{app:tensorstructures}. Additionally, the
vertices appearing in \eq{eq:runcoup} are evaluated at the symmetric
point, see \Sec{sec:truncationcheck} for the precise definition. The
STIs and two-loop universality demand that these running couplings
become degenerate at large perturbative momentum scales, where the
longitudinal and transverse parts of the vertices agree.
In Landau gauge, the ghost-gluon vertex is not renormalised on
specific momentum configurations, and we can alternatively define a
running coupling from the wave function renormalisation of ghost and
gluon \cite{vonSmekal:1997ohs,vonSmekal:2009ae},
\begin{align}
\label{eq:propcoupling}
\alpha_s(p)= \0{1}{4\pi} \,\0{g^2 }{Z_A(p) Z_c^2(p)}\,.
\end{align}
Note that the momentum-dependence of the running coupling
\eq{eq:propcoupling} does not coincide with that of the corresponding
running couplings obtained from other vertices, i.e.\ \eq{eq:runcoup}.
This is best seen in the ratio $\alpha_{A\bar c c}(p)/\alpha_s(p)=
Z_{A\bar c c,\bot}^{2}(p)/g^2\,$. In this context we also report on an important
result for the quark-gluon vertex coupling,
\begin{align}
\label{eq:quarkgluon}
\alpha_{A\bar q q}(p)=\0{1}{4\pi} \,\0{Z_{A\bar q q,\bot}(p)^2 }{Z_A(p)
Z_q(p)^2}\,,
\end{align}
with the dressing function of the classical tensor structure of the
quark-gluon vertex $Z_{A\bar q q,\bot}(p)^2\,$
and the quark dressing function $1/Z_q(p)$~\cite{Mitter:2014wpa}. The solution of the
corresponding STI reveals that the quark-gluon vertex coupling
$\alpha_{A\bar q q}$ agrees perturbatively with $\alpha_s(p)$ in
\eq{eq:propcoupling}, and hence it differs from the other vertex
couplings in \eq{eq:runcoup}. Note that the present truncation only
considers contributions from primitively divergent
vertices. Accordingly, the two-quark--two-ghost vertex contribution in
the STI for the quark-gluon vertex, see e.g.\ \cite{Alkofer:2000wg}, has
been dropped.
\section{Confinement, gluon mass gap, \newline and irregularities}
\label{sec:massgap}
It has been shown in
\cite{Braun:2007bx,Marhauser:2008fz,Braun:2009gm,Braun:2010cy,Fister:2013bh} that a
mass gap in the gluon propagator signals confinement in QCD in
covariant gauges. Furthermore, in Yang-Mills theory formulated in
covariant gauges, the gapping of the gluon relative to the ghost is
necessary and sufficient for producing a confining potential for the
corresponding order parameter, the Polyakov loop. Hence, understanding
the details of the dynamical generation of a gluon mass gap gives
insight into the confinement mechanism.
This relation holds for all potential infrared closures of the
perturbative Landau gauge. The standard infrared closure
corresponds to a full average over all Gribov regions. This leads to
the standard Zinn-Justin equation as used in the literature, e.g.\
\cite{Alkofer:2000wg}. In turn, the restriction to the first Gribov
regime can be implemented within the refined Gribov-Zwanziger
formalism, e.g.\
\cite{Dudal:2007cw,Dudal:2008sp,Dudal:2009xh,Dudal:2011gd,Capri:2015ixa},
that leads to infrared modifications of the STIs. In the following we
discuss the consequences of the standard STIs, a discussion of the
refined Gribov-Zwanziger formalism is deferred to future work.
\begin{figure*}
\centering
\includegraphics[width=0.48\textwidth]{gluon_prop_dressing_ind_sca}
\hfill
\includegraphics[width=0.489\textwidth]{ghost_prop_dressing_ind_sca}
\caption{Gluon dressing $1/Z_A$ (left) and ghost dressing $1/Z_c$
(right) in comparison to the lattice results
from \cite{Sternbeck:2006cg}. The scale setting and normalisation procedures are described in \App{app:rescaling}.}
\label{fig:main_result}
\end{figure*}
\subsection{Gluon mass gap and irregularities}
\label{sec:gluonmassirregularities}
In order to study the dynamical generation of the mass gap, we first
discuss the consequences of the STI for the longitudinal gluon two
point function \eq{eq:LSTI}. It states that no quantum fluctuations
contribute to the inverse longitudinal gluon propagator, i.e.\ the
longitudinal gluon propagator is defined by the gauge fixing
term. Therefore, the dynamical creation of a gluon mass gap requires
different diagrammatic contributions to the longitudinal and
transverse gluon mass parameter. The discussion of the prerequisites
for meeting this condition is qualitatively different for the scaling
and the decoupling solutions. Hence, these two cases are discussed
separately.
The scaling solution is characterised by the infrared behaviour
\cite{vonSmekal:1997ohs,Zwanziger:2001kw,Lerche:2002ep,Fischer:2002eq,%
Pawlowski:2003hq,Alkofer:2004it,Fischer:2006vf,Alkofer:2008jy,Fischer:2009tn}
\begin{align}\nonumber
\lim\limits_{p\rightarrow 0}Z_c(p^2)&\varpropto (p^2)^{\kappa}\,,\\[2ex]
\lim\limits_{p\rightarrow 0}Z_A(p^2)&\varpropto (p^2)^{-2\,
\kappa}\,,
\label{eq:scaling_sol}
\end{align}
with the scaling coefficient $1/2<\kappa<1$. A simple calculation
presented in \App{app:irregularities} shows that the ghost loop with
an infrared constant ghost-gluon vertex and scaling ghost propagator
is already capable of inducing a splitting in the longitudinal and
transverse gluon mass parameter.
Next we investigate the decoupling solution, e.g.\
\cite{Aguilar:2008xm,Boucaud:2008ky}, which scales with
\begin{align}\nonumber
\lim\limits_{p\rightarrow 0}Z_c(p^2)&\varpropto 1\, ,\\[2ex]
\lim\limits_{p\rightarrow 0}Z_A(p^2)&\varpropto (p^2)^{-1}\,,
\label{eq:decoupling_sol}
\end{align}
at small momenta. Assuming vertices that are regular in the limit of
one vanishing gluon momentum, one finds that all diagrammatic
contributions to the longitudinal and transverse gluon mass parameter
are identical. For example, if the ghost loop were to yield a
non-vanishing contribution to the gluon mass gap, the ghost-gluon
vertex would have to be a function of the angle $\theta=\arccos(t)$
between the gluon and anti-ghost momenta $p$ and $q$,
\begin{align}
\label{eq:irregular}
\lim\limits_{|p|\rightarrow 0}[\Gamma^{(3)}_{A\bar c
c}{}]_{\mu}^{abc}(|p|,|q|,t) = [\Gamma^{(3)}_{A\bar c
c}{}]_{\mu}^{abc}(0,|q|,t)\,,
\end{align}
even in the limit of vanishing gluon momentum $|p|\rightarrow
0\,$. Since the above limit depends on the angle, the vertex is
irregular. See \App{app:irregularities} for more details on this
particular case. Similar conclusions can be drawn for all vertices
appearing in the gluon propagator equation. Consequently, if all
vertices were regular, no gluon mass gap would be created. In
particular, regular vertices would entail the absence of
confinement. The necessity of irregularities for the creation of a
gluon mass gap was already realised by Cornwall~\cite{Cornwall:1981zr}.
In the light of these findings it is interesting to investigate the
consistency of irregularities with further Slavnov-Taylor
identities. Therefore, we consider the Slavnov-Taylor identity of the
three-gluon vertex, e.g.\ \cite{Alkofer:2000wg},
\begin{widetext}
\begin{align}
\label{eq:STIclass3gl}
\text{i} r_\rho
[\Gamma^{(3)}_{A^3}]{}^{abc}_{\mu\nu\rho}(p,q,r)\propto f^{abc}
\0{1}{Z_c(r^2)}\left( \tilde G_{\mu\sigma}(p,q) q^2
Z_A(q^2)\Pi^\bot_{\sigma\nu}(q) -\tilde G_{\nu\sigma}(q,p)
p^2 Z_A(p^2)\Pi^\bot_{\sigma\mu}(p) \right)\,,
\end{align}
\end{widetext}
where $\tilde G_{\mu \nu}$ relates to the ghost-gluon vertex via
\begin{align}
\label{eq:STIGtilde}
[\Gamma^{(3)}_{A \bar c c} ]_{\mu}^{abc}(p,q)=\text{i} g f^{abc} q_\nu
\tilde G_{\mu\nu}(p,q)\,.
\end{align}
For a regular $\tilde G_{\mu\nu}$ in the limit $p\rightarrow 0$
in \eq{eq:STIclass3gl}, the
scaling solution leads to a singular contribution of the type
\begin{align}\label{eq:3g_sing_scaling}
\lim\limits_{p\rightarrow 0} (p^2)^{1-2\kappa}\, \tilde
G_{\nu\sigma}(q,0)\, \Pi^\bot_{\sigma\mu}(p) + \rm regular\,,
\end{align}
where $\kappa$ is the scaling coefficient from
\eq{eq:scaling_sol}. This is consistent with the expected scaling
exponent of the three-gluon vertex in this limit
\cite{Alkofer:2008jy}. In the same limit, the decoupling solution
leads to a singular contribution of the form
\begin{align}
\label{eq:3g_sing_decoupling}
\lim\limits_{p\rightarrow 0} \tilde G_{\nu\sigma}(q,0)\,
\Pi^\bot_{\sigma\mu}(p) + \rm regular\,.
\end{align}
Since the transverse projector $\Pi^\bot_{\sigma\mu}(p)$ introduces an
angular dependence in the limit $p\rightarrow 0\,$, the STI again
demands an irregularity in limit of one vanishing momentum. Note that
this is just a statement about the three-gluon vertex projected with
one non-zero longitudinal leg $r_\rho\,$. Although this momentum
configuration does not enter the gluon mass gap directly, crossing
symmetry implies the necessary irregularity. In summary, these
arguments illustrate that also the three-gluon vertex STI is
consistent with the necessity of irregularities for both types of
solutions.
We close the discussion of vertex irregularities with the remark that
the infrared modification of the propagator-STI in the refined
Gribov-Zwanziger formalism may remove the necessity for irregularities
in the vertices.
\begin{figure*}
\centering
\includegraphics[width=0.48\textwidth]{gluon_prop_ind_sca}
\hfill
\includegraphics[width=0.496\textwidth]{running_couplings}
\caption{ Left: Gluon propagator in comparison to the lattice
results from \cite{Sternbeck:2006cg}. Right: Effective
running couplings defined in \eq{eq:runcoup} as obtained
from different Yang-Mills vertices as function of the
momentum.
}
\label{fig:main_result_II}
\end{figure*}
\subsection{Origin of irregularities}
As discussed in the previous section, self-consistency in terms of the
Slavnov-Taylor identities entails a correspondence between the
dynamical generation of a gluon mass gap and the presence of
irregularities. But the STIs do not provide a mechanism for the
creation of irregularities, the gluon mass gap, and in turn
confinement.
In the scaling solution, \eq{eq:scaling_sol}, the irregularities arise
naturally from the non-trivial scaling. Hence they are tightly linked
to the original Kugo-Ojima confinement scenario \cite{Kugo:1979gm}, that requires the
non-trivial scaling. Note, however, that this simply links different
signatures of confinement but does not reveal the mechanism at work.
For the decoupling solution \eq{eq:decoupling_sol}, we want to
discuss two possible scenarios. In the first scenario, the irregularities
are generated in the far infrared. A second possibility is that they are
triggered via a condensate and/or a resonance, providing a direct connection
of confinement and spontaneous symmetry breaking.
In the first scenario it is sufficient to focus on ghost loops as
possible sources of such irregularities, since the gluonic diagrams
decouple from the infrared dynamics due to the gluon mass gap. This
is a seemingly appealing scenario as it is the dynamical ghost that
distinguishes confining Yang-Mills theory from e.g.\ QED. However, in
the decoupling solution \eq{eq:decoupling_sol} both, the ghost-gluon
vertex as well as the ghost propagator, have infrared finite quantum
corrections: no ghost-loops contribute to their equation and
(infrared) constant dressing functions can be assumed for both. As a
consequence the ghost loop contributions to correlation functions have
the same infrared structure as perturbative ghost-loop
contributions. However, none of these perturbative ghost loops yields
the necessary irregularities, see \App{app:ghosttriangle} for an
explicit calculation.
In the second scenario, the generation of irregularities can be based
on the dynamical generation of a non-vanishing transverse background,
$F_{\mu\nu}^a F_{\mu\nu}^a \neq 0\,$, in the infrared. This gluon
condensate is the Savvidi vacuum \cite{Savvidy:1977as}, and its
generation in the present approach has been discussed in
\cite{Eichhorn:2010zc} with $F_{\mu\nu}^a F_{\mu\nu}^a \approx
\SI{1}{\GeV}^4\,$. Then, a vertex expansion about this non-trivial
IR-solution of the equation of motion introduces an IR-splitting of
transverse and longitudinal vertices due to the transversality of the
background field. This IR-splitting automatically implies
irregularities as discussed in \Sec{sec:gluonmassirregularities}, and
is sufficient for creating a physical mass gap in the gluon. This
scenario provides a direct relation of confinement and spontaneous
symmetry breaking. Therefore it is possibly connected to the presence
of resonances that are triggered in the longitudinal sector of the
theory, where they do not spoil the gapping of the completely
transverse sector. A purely longitudinal massless mode, as a source
for irregularities in the gluonic vertices, has been worked out in
\cite{Aguilar:2011ux,Aguilar:2011xe}, for a short summary see
\cite{Figueiredo:2016cvf}. As a consequence, an irregularity appears
in the purely longitudinal three-gluon vertex in a way that preserves
the corresponding Slavnov-Taylor identity. The creation of a purely
transverse background and the presence of longitudinal massless mode
would then be two sides of the same coin. Furthermore, the
longitudinal resonance has to occur at about the same scale as the
gluon condensate, in order to trigger the correct gluon mass gap. A
more detailed discussion and computation of this scenario cannot be
assessed in the purely transverse system and is therefore deferred to
future work.
\begin{figure*}
\centering
\includegraphics[width=0.488\textwidth]{AcbcSymPoint_clean}
\hfill
\includegraphics[width=0.48\textwidth]{AaaSymPoint_clean}
\caption{ Ghost-gluon vertex (left) and three-gluon
(right) vertex dressing functions $Z_{A \bar
cc,\bot }\left(p,\,p,\,-\frac{1}{2}\right)$ and
$Z_{A^3,\bot }\left(p,\,p,\,-\frac{1}{2}\right)$ in the
symmetric point configuration. More momentum
configurations and comparisons to Dyson-Schwinger
and lattice results can be found in
Figs.~\ref{fig:fourGluonTadpoleDressing}-\ref{fig:ThreePointVertices}.
In contrast to \Fig{fig:main_result}, the decoupling
dressings are normalised to the scaling solution
in the UV. }
\label{fig:main_result_III}
\end{figure*}
\subsection{The purely transverse system}
\label{sec:transverse}
In this work we restrict ourselves to a solution of the purely
transverse system \eq{eq:closedFun}, which is closed. The only
relevant UV parameters in this system are the strong coupling and the
transverse gluon mass parameter. In the UV the transverse mass
parameter agrees with the longitudinal one. The latter is fixed by the
mSTI for the longitudinal gluon propagator. Hence, the only
information needed from the longitudinal system is the initial value
for the transverse gluon mass parameter \eq{eq:mSTI_mass}. Note also
that there is at least one value for the initial gluon mass parameter
that yields a valid confining solution. In the following we vary the
gluon mass parameter and discuss the properties of the ensuing
solutions. We find a confining branch with both scaling and decoupling
solutions. In addition, we observe a transition to the deconfined
Higgs-type branch. No Coloumb branch is found. The unique scaling
solution satisfies the original Kugo-Ojima confinement criterion with
$Z_C(p=0)=0\,$. We emphasise that the existence of the scaling
solution is dynamically generated in a highly non-trivial way. The
details are discussed in \Sec{sec:gluonmassgap}.
\section{Numerical results}
\label{sec:mainresult}
We calculate Yang-Mills correlation functions by integrating the
self-consistent system of flow equations obtained from functional
derivatives of \eq{eq:flow}, see \Fig{fig:diagrams}
for diagrammatic representations. Technical details on the
numerical procedure are given in \App{app:technicalDetails}. We use
constant dressing functions as initial values for the $1$PI
correlators at the ultraviolet initial scale $\Lambda\,$. Consequently,
the initial action $\Gamma_\Lambda$ is given by the bare action of QCD
and the Slavnov-Taylor identities enforce relations between these
constant initial correlation functions. As is well-known, and also
discussed in \Sec{sec:mSTIandVert_sub}, the Landau gauge STIs leave
only three of the renormalisation constants independent, namely the
value of the strong running coupling and two trivial renormalisations
of the fields that drop out of any observable. To eliminate cutoff
effects, we choose the constant initial values for the vertex
dressings such that the momentum-dependent running couplings,
\eq{eq:runcoup} are degenerate at perturbative momentum scales $p$
with $\Lambda_\text{\tiny{QCD}}\ll p\ll \Lambda\,$ i.e. the
STIs \eq{eq:RGrel} are only fulfilled on scales considerably below
the UV cutoff scale. The modification of the Slavnov-Taylor
identity, caused by the regulator term, requires a non-physical gluonic mass
term $m_\Lambda^2$ at the cutoff $\Lambda$. The initial value for
the inverse gluon propagator is therefore taken as
\begin{align}
[\Gamma^{(2)}_{AA,\Lambda}]^{ab}_{\mu\nu}(p) &=
\left(Z_{A,\Lambda}\, p^2+ m_\Lambda^2\right)\,\delta^{ab}\,
\Pi^{\bot}_{\mu\nu}(p)\,.
\end{align}
The non-physical contribution $m_k^2$ to the gluon propagator vanishes
only as the renormalisation group scale, $k\,$, is lowered to zero,
where the mSTIs reduce to the usual Slavnov-Taylor identities. The
initial value $m_\Lambda^2$ can be uniquely fixed by demanding that
the resulting propagators and vertices are of the scaling
type. Consequently, the only parameter in this calculation is the
value of the strong running coupling at the renormalisation scale, as
initially stated. We also produce decoupling solutions by varying the
gluon mass parameter towards slightly larger values. Our reasoning
for their validity as confining solutions is presented in
\Sec{sec:gluonmassgap}.
\begin{figure*}
\centering
\includegraphics[width=0.497\textwidth]{AaaaSymPoint}
\hfill
\includegraphics[width=0.48\textwidth]{gluon_prop_dressing_truncations_pert_0D_fit}
\caption{ Left:
Four-gluon vertex dressing function as defined in
\eq{eq:fourgluon} at the symmetric point in comparison to
Dyson-Schwinger computations~\cite{Cyrol:2014kca}. We
normalised all curves to match the scaling result at
$p=\SI{2}{\GeV}$. Right: Gluon propagator dressings obtained with
different momentum approximations, see
\Sec{sec:truncationcheck} for details.
}
\label{fig:main_result_IIII}
\end{figure*}
\subsection{Correlation functions and running couplings}
\label{sec:corcoup}
We show our results for the Yang-Mills correlation functions as well
as the momentum-dependent transverse running couplings in
Figs.~\ref{fig:main_result}-\ref{fig:main_result_IIII}, see also
Figs.~\ref{fig:fourGluonTadpoleDressing}-\ref{fig:ThreePointVertices}
in the appendices for a comparison of the vertices to recent lattice
and DSE results. A discussion of truncation effects is deferred to
\Sec{sec:truncationcheck}. In order to be able to compare to results
from lattice simulations, we set the scale and normalise the dressings
as described in \App{app:rescaling}. At all momenta, where the
difference between the scaling (solid line) and decoupling (band
bounded by dashed-dot line) solutions is negligible, our results for
the correlations functions agree very well with the corresponding
lattice results. In the case of the scaling solution we find the
consistent scaling exponents
\begin{align}\nonumber
\kappa_\text{ghost}&=0.579\pm0.005\,,\\[2ex]
\kappa_\text{gluon}&=0.573\pm0.002\,,
\end{align}
where the uncertainties stem from a least square fit with the ansatz
\begin{align}\nonumber
Z_c(p)&\varpropto (p^2)^{ \kappa_\text{ghost}}\,,\\[2ex]
Z_A(p)&\varpropto (p^2)^{-2\, \kappa_\text{gluon}}\,.
\end{align}
As discussed in \Sec{sec:transverse}, the scaling solution is a
self-consistent solution of the purely transverse system in the used
approach, and has no systematic error related to the lack of solving
the longitudinal system. In turn, the presented decoupling solutions
suffer from the missing solution of the longitudinal system, leading
to a small additional systematic error. This argument already
suggests that it is the presented scaling solution that should agree
best with the lattice results in the regime $p\gtrsim \SI{1}{\GeV}\,$,
where the solutions show no sensitivity to the Gribov problem. This is
confirmed by the results, see in particular \Fig{fig:main_result}.
In the infrared regime, $p\lesssim \SI{1}{\GeV}\,$, the different
solutions approach their infrared asymptotics. In
\Fig{fig:main_result} and \Fig{fig:main_result_II} we compare the FRG
solutions with the lattice data from \cite{Sternbeck:2006cg}. In
agreement with other lattice results
\cite{Cucchieri:2007rg,Cucchieri:2008fc,Maas:2009ph} in four
dimensions, these propagators show a decoupling behaviour, for a
review see \cite{Maas:2011se}. Taking the IR behaviour of all
correlators into account, cf.\ also \Fig{fig:ThreePointVertices}, the
lattice solution \cite{Sternbeck:2006cg} is very close to the
decoupling solution (dot-dashed line) that is furthest from the
scaling solution (solid line). Note however, that the systematic
errors of both approaches, FRG computations and lattice simulations
increase towards the IR. While the FRG computations lack apparent
convergence in this regime, the lattice data are affected by the
non-perturbative gauge fixing procedure, i.e.\ the choice of Gribov
copies \cite{Maas:2009se,Sternbeck:2012mf,Maas:2015nva} and
discretisation artefacts \cite{Duarte:2016iko}. Consequently,
comparing the FRG IR band to the lattice propagators has to be taken
with a grain of salt. In the case of the vertices, we compare also to
results obtained within the Dyson-Schwinger equation approach
\cite{Huber:2012kd,Blum:2014gna,Cyrol:2014kca}, see
\Fig{fig:comparison} and \ref{fig:ThreePointVertices}. A comparison of
the different running vertex couplings is given in
\Sec{sec:comparison}.
We find that it is crucial to ensure the degeneracy in the different
running vertex couplings at perturbative momentum scales in order to
achieve quantitative accuracy, see also \Sec{sec:truncationcheck}.
The transverse effective running couplings, as defined in
\eq{eq:runcoup}, are shown in the right panel of
\Fig{fig:main_result_II}. To be able to cover a larger range of
momenta with manageable numerical effort, the shown running couplings
have been obtained within an approximation that takes only one
momentum variable into account in the vertices, see
\Sec{sec:truncationcheck}. At large perturbative momentum scales, we
find them to be perfectly degenerate, as is demanded by the
Slavnov-Taylor identities. The degeneracy of the running couplings is
lifted at a scale of roughly $\SI{2}{\GeV}\,$, which coincides with the
gapping scale of the gluon. Furthermore, the three-gluon vertex shows
a zero crossing at scales of \SIrange{0.1}{0.33}{GeV}, which is the
reason for the spike in the corresponding running coupling. This zero
crossing, which is caused by the infrared-dominant ghost-loop, is
well-known in the literature
\cite{Aguilar:2013vaa,Pelaez:2013cpa,Blum:2014gna,Eichmann:2014xya}.
Even though we are looking at the scaling solution, we find that the
running couplings defined from the purely gluonic vertices are still
strongly suppressed in the infrared. In particular the three-gluon
vertex running coupling becomes more strongly suppressed than the
four-gluon vertex running coupling. However, as demanded by scaling,
they seem to settle at tiny but finite fixed point values, which has
also been seen in Dyson-Schwinger studies
\cite{Eichmann:2014xya,Kellermann:2008iw,Cyrol:2014kca}.
\subsection{Quality of the approximation}
\label{sec:truncationcheck}
In \Fig{fig:main_result_IIII} (right panel), we show the scaling
solution for the propagators in different truncations. In all cases,
the full momentum dependence of the propagators is taken into account
whereas different approximations are used for the vertices. Including
only RG-scale-dependent constant vertex dressing functions is the
minimal approximation that can produce a scaling solution with a
physical gluon mass gap. The dot-dashed (magenta) line in
\Fig{fig:main_result_IIII} (right panel) corresponds to an
approximation with constant vertex dressing functions evaluated at the
symmetric configuration with momentum $\mathcal{O}(\SI{250}{\MeV})\,$.
Hence the vertices are only RG-scale-dependent vertices. For the
dashed blue results the dressing functions for the transversally
projected classical tensor structures have been approximated with a
single momentum variable $\bar{p}^2 \equiv \tfrac{1}{n}\sum_{i=1}^n
p_i^2\,$. Reducing the momentum dependence to a single variable
requires the definition of a momentum configuration to evaluate the
flow. Here, we use the symmetric point configuration, defined by $p_i
\cdot p_i = p^2$ and $p_i\cdot p_j = -1/(n-1)$ for $i\neq j\,$, where
$n=3\, (4)$ for the three(four)-gluon vertex. Finally, the solid red
line corresponds to our best truncation. As described in
\Sec{sec:expsch}, it takes into account the full momentum dependence
of the classical tensors structures of the three-point functions as
well as the four-gluon vertex in a symmetric point
approximation. Additionally, all (three-dimensional) momentum
configurations of the four-gluon vertex that are needed in the tadpole
diagram of the gluon propagator equation have been calculated and
coupled back in this diagram. The reliability of our approximation
can be assessed by comparing the two simpler truncations to the result
obtained in our best truncation scheme. We observe that our results
apparently converge towards the lattice result, as we improve the
momentum approximation for the vertices.
The effects of non-classical tensor structures and vertices are beyond
the scope of the current work and have to be checked in future
investigations, see however \cite{Eichmann:2014xya} for an
investigation of non-classical tensor structures of the three-gluon
vertex. Within the present work, the already very good agreement with
lattice results suggests, that their influence on the propagators is
small.
The final gluon propagator is sensitive to the correct renormalisation
of the vertices. For example, a one percent change of the three-gluon
vertex dressing at an UV scale of $\SI{20}{\GeV}$ magnifies by up to a
factor 10 in the final gluon propagator. Therefore, small errors in
the perturbative running of the vertices propagate, via
renormalisation, into the two-point functions. We expect a five
percent uncertainty in our results due to this.
Despite these uncertainties, we interpret the behaviour in
\Fig{fig:main_result_IIII} (right panel) as an indication for apparent
convergence.
\subsection{Comparison to other results}
\label{sec:comparison}
\begin{figure}
\includegraphics[width=0.48\textwidth]{running_couplings_comp_rescaled}
\caption{Running couplings \eq{eq:runcoup} in comparison with DSE running.
The grey band gives the spread of vertex
couplings from the FRG in the present work. The DSE results
are shown rescaled to fit our ghost-gluon vertex running
coupling at $\SI{10}{\GeV}$ to facilitate the comparison.
The inlay shows the unscaled couplings. Note that the FRG
running couplings naturally lie on top of each other and are
not depicted rescaled.}
\label{fig:comparison}
\end{figure}
In \Fig{fig:ThreePointVertices}, numerical results for the ghost-gluon and
three-gluon vertices are shown in comparison to other functional
methods as well as lattice results. In summary, the results from
various functional approaches and the lattice agree to a good degree.
But these correlation functions are not renormalisation group
invariant, and a fully meaningful comparison can only be made with
RG invariant quantities. Therefore, we compare our results for the
RG invariant running couplings with the respective results from DSE
computations. To be more precise, it is actually the $\beta$ functions
of the different vertices that are tied together by two-loop
universality in the sense that they should agree in the regime where
three-loop effects are negligible. Since constant factors drop out of
the $\beta$ functions, we have normalised the DSE running couplings to
the FRG result at large momentum scales in \Fig{fig:comparison}.
For the sake of visibility, we only have provided a band for
the spread of the FRG couplings as obtained from different vertices.
The shown DSE running couplings are based on a series of works
\cite{Huber:2012kd,Blum:2014gna,
Cyrol:2014kca,Williams:2014iea,Williams:2015cvx}, where the explicitly
shown results are taken from
\cite{Blum:2014gna,Cyrol:2014kca,Huber:unpublished,Williams:unpublished}.
Additionally, we provide the raw DSE running couplings that have not
been rescaled by a constant factor in the inlay.
\subsection{Mass gap, mSTIs and types of solutions}
\label{sec:gluonmassgap}
\begin{figure*}
\centering
\includegraphics[width=0.481\textwidth]{ghost_prop_dressing_tuning}
\hfill
\includegraphics[width=0.48\textwidth]{gluon_prop_tuning}
\caption{Ghost dressing functions $1/Z_c$ (left) and
gluon propagators (right) for different values of the
ultraviolet gluon mass parameter. Blue results correspond to
the Higgs-type branch and red results to the confined branch. The
solutions in all branches have been normalised to the
scaling solution in the UV. }
\label{fig:confinementVSHiggs}
\end{figure*}
As discussed in \Sec{sec:mSTIandVert_sub}, the introduction of the
regulator in the FRG leads to a modification of the Slavnov-Taylor
identities. In turn the inverse gluon propagator obtains a
contribution proportional to $\Delta\Gamma^{(2)}_{AA}\varpropto
k^2\alpha(k)$ for all $k>0\,$. Disentangling the physical mass gap
contribution from this mSTI contribution to the gluon mass parameter
is intricate, both conceptually and numerically. The resulting
numerical challenge is illustrated in the appendix in
\Fig{fig:flowQuantities}, where we show the $k$-running of the gluon
mass parameter. This is the analogue of the problem of quadratic
divergences in Dyson-Schwinger equations with a hard momentum cutoff,
see e.g.\ \cite{Huber:2014tva}. However, there has to exist at least
one choice for the gluon mass parameter $m_\Lambda^2$ that yields a
valid confining solution, see \Sec{sec:mSTIandVert_sub}. To resolve
the issue of finding this value, we first recall that a fully regular
solution has no confinement and necessarily shows a Higgs- or
Coulomb-type behaviour. Although we do not expect these branches to be
consistent solutions, we can trigger them by an appropriate choice of
the gluon mass parameter in the UV. The confinement branch then lies
between the Coulomb and the Higgs branch. We need, however, a
criterion for distinguishing between the confinement and the
Higgs-type branch.
To investigate the possible solutions in a controlled way, we start
deep in the Higgs-type branch: an asymptotically large initial gluon
mass parameter $m_\Lambda^2$ triggers an explicit mass term of the
gluon at $k=0\,$. If we could trigger this consistently in the present
$SU(3)$ theory, it would constitute a Higgs solution. Note that in the
current approximation it cannot be distinguished from massive
Yang-Mills theory, which has e.g.\ been considered in
\cite{Tissier:2010ts,Pelaez:2014mxa}. Starting from this Higgs-type
branch, we can then explore the limit of smaller initial mass
parameters. This finally leads us to the scaling solution, which
forms the boundary towards an unphysical region characterised by
Landau-pole-like singularities. It is left to distinguish between the
remaining confining and Higgs-type solutions, shown in
\Fig{fig:confinementVSHiggs}, without any information from the
longitudinal set of equations. For that purpose we use two criteria:
In the left panel of \Fig{fig:mass_parameter}, we show the mass gap of
the gluon, $m^{2}=\Gamma^{(2)}_{AA,k=0}(p=0)\,$, as a function of the
chosen initial value for the gluon mass parameter $m_{\Lambda}^2$
subtracted by the corresponding value for the scaling solution
$m_{\Lambda,\rm scaling}^2\,$. The latter solution corresponds to zero
on the x-axis in \Fig{fig:mass_parameter}. As mentioned before, going
beyond the scaling solution, $m_{\Lambda}^2<m_{\Lambda,\rm
scaling}^2\,$, leads to singularities. We interpret their presence as
a signal for the invalidity of the Coulomb branch as a possibly
realisation of non-Abelian Yang-Mills theory. The decisive feature of
the left panel of \Fig{fig:mass_parameter} is the presence of a
minimum at $m_{\rm min}^2\,$. If there were no dynamical mass gap
generation, $m^2$ would have to go to zero as we lower
$m_{\Lambda}^2\,$. In contrast to this, we find that the resulting
gluon mass gap is always larger than the value it takes at
$m_{\Lambda}^2=m_{\rm min}^2\,$. In particular, this entails that all
solutions to the left of the minimal value, $m_{\Lambda}^2<m_{\rm min}^2\,$,
are characterised by a large dynamical contribution to the
gluon mass gap, which we interpret as confinement.
\begin{figure*}
\centering
\includegraphics[width=0.485\textwidth]{mUV_m0_m}
\hfill
\includegraphics[width=0.48\textwidth]{order_param}
\caption{ Left: Gluon mass gap as a function of the gluon mass
parameter $m_\Lambda^2-m_\text{$\Lambda$, scaling}^2\,$,
where $m_\text{$\Lambda$, scaling}^2$ denotes the gluon mass
parameter that yields the scaling solution. Right: Momentum
value at which the gluon propagator assumes its maximum, as
a function of the gluon mass parameter
$m_\Lambda^2-m_\text{$\Lambda$, scaling}^2\,$. The inlay
exposes the power law behaviour of the gluon propagator
maximum in the vicinity of the transition region, see
\eq{eq:phaseTransitionFit}. Both plots were obtained from
our numerically less-demanding 1D approximation. We have
repeated this analysis in the transition regime from
Higgs-type to confinement branch also with the best
approximation and find the same behaviour. The shaded area
marks momentum scales that are not numerically resolved in
the present work. The points in this region rely on a
generic extrapolation. }
\label{fig:mass_parameter}
\end{figure*}
As a second criterion for differentiating between confining and Higgs
solutions, we use the presence of a maximum at non-vanishing momenta
in the gluon propagator, which signals positivity violation
\cite{Alkofer:2000wg}. In the right panel of \Fig{fig:mass_parameter},
we show the location of the maximum in the gluon propagator, again as
a function of the gluon mass parameter, $m_{\Lambda}^2-m_{\Lambda,\rm
scaling}^2\,$. We clearly see a region of confining solutions that
show a back-bending of the gluon propagator at small momenta,
see \Fig{fig:confinementVSHiggs}. The dashed line,
separating the shaded from the white region in the right panel of
\Fig{fig:mass_parameter}, indicates the smallest momentum value at
which the gluon propagator has been calculated. With this restriction
in mind, the fit in the inlay demonstrates that the location of the
maximum of the propagator scales to zero as one approaches the
critical value $m_{\rm c}^2\,$. We fit with
\begin{align}
\label{eq:phaseTransitionFit}
p_\text{max}(m^2_\Lambda) \propto
\left(\frac{m^2_\Lambda-m^2_{\rm c}}{m^2_{\rm
c}}\right)^\alpha\,,
\end{align}
which yields the critical exponent
\begin{align}
\label{eq:phaseTransitionFitResult}
\alpha = 1.95 \pm 0.6\,,
\end{align}
in the 1D approximation. Within the numerical accuracy, this boundary
value $m_{\rm c}^2$ is equivalent to the minimal value $m_{\rm min}^2$
of our first confinement criterion. Hence, the value of the UV mass
parameter that results in the minimal gluon mass gap, is also the one
that shows minimal back-bending. Note that the lattice simulations
show a gluon propagator that is at least very close to this minimal
mass gap.
As discussed in detail in \Sec{sec:massgap} and appendices
\ref{app:irregularities} and \ref{app:ghosttriangle}, a gluon mass gap
necessitates irregularities. The scaling solution by definition
contains these irregularities already in the propagators, cf.\
\eq{eq:scaling_sol}. For the decoupling-type solutions, we excluded
infrared irregularities of diagrammatic origin, see
\App{app:ghosttriangle}. Thus, for the decoupling-type solutions our
arguments for the validity of the solutions are weaker and remain to
be investigated in a solution including at least parts of the
longitudinal system, see the discussion in \Sec{sec:massgap}.
Additionally, it might be necessary to expand about the solution of
the equation of motion, see \cite{Eichhorn:2010zc}.
We summarise the findings of the present section. In the right panel
of \Fig{fig:mass_parameter} we can distinguish a confining branch with
positivity violation and a Higgs-type branch with a massive gluon
propagator. A Coulomb-type solution, on the other hand, can never be
produced with the functional renormalisation group since any attempt
to do so leads to Landau-pole-like singularities. The non-existence of
the Coulomb branch is tightly linked to the non-monotonous dependence
of the mass gap on the initial gluon mass parameter, see left panel of
\Fig{fig:mass_parameter}. This behaviour is of a dynamical origin that is
also responsible for the existence of the scaling solution for the
smallest possible UV gluon mass parameter.
\subsection{Discussion}
\label{sec:discussion}
As has been discussed already in \Sec{sec:corcoup}, one non-trivial
feature of the different vertex couplings is their quantitative
equivalence for momenta down to $p\approx \SI{2}{\GeV}\,$, see
\Fig{fig:main_result_II} (right panel). This property extends the
universal running of the couplings into the semi-perturbative regime.
On the other hand, the couplings violate universality in the
non-perturbative regime for $p\lesssim \SI{2}{\GeV}\,$. The
universality down to the semi-perturbative regime is a very welcome
feature of Landau gauge QCD, as it reduces the size of the
non-perturbative regime and hence the potential systematic errors. In
particular, one running coupling is sufficient to describe Landau
gauge Yang-Mills theory down to momentum scales of the order of the
gluon mass gap. This suggests to use the propagators together with
the ghost-gluon vertex for simple semi-quantitative calculations. The
above structure also explains and supports the semi-quantitative
nature of the results in low-order approximations.
This implies that self-consistent calculations of most or all vertices
have to reproduce this universality, in particular for momenta
$\SI{2}{\GeV}\lesssim p\lesssim \SI{10}{\GeV}\,$. When starting from
the value of the strong running coupling at perturbative momenta, we
find that a violation of the degeneracy of the running couplings,
\eq{eq:runcoup}, in this regime goes hand in hand with the loss of
even qualitative properties of the non-perturbative results in
self-consistent approximations. This surprising sensitivity to even
small deviations of the couplings from their universal running extends
to the fully dynamical system with quarks, see e.g.\
\cite{Mitter:2014wpa,CMPS:VC}. Note in this context that the
quark-gluon coupling $\alpha_{A\bar q q}\,$, \eq{eq:quarkgluon},
agrees with the ghost-gluon coupling $\alpha_s$ defined in
\eq{eq:propcoupling}, and not the vertex coupling $\alpha_{A\bar c
c}\,$, see \Sec{sec:STImSTI}. It can be shown in the full
QCD system, that deviations from universality on the percent level
have a qualitative impact on chiral symmetry breaking. The origin of
this is the sensitivity of chiral symmetry breaking to the correct
adjustment of physical scales, i.e.\ $\Lambda_{\text{\tiny QCD}}\,$,
in all subsystems. These observations underline the relevance of the
present results for the quantitative grip on chiral symmetry
breaking. A full analysis will be presented in a forthcoming work,
\cite{CMPS:VC}.
We close this discussion with the remark that universality in the
semi-perturbative regime is tightly linked with the consistent
renormalisation of all primitively divergent correlation functions.
We find it crucial to demand the validity of the STIs \eq{eq:RGrel}
only on momentum scales considerably below the ultraviolet cutoff
$\Lambda\,$. On the other hand, the relations \eq{eq:RGrel} are
violated close to the ultraviolet cutoff, due to the BPHZ-type
subtraction schemes. This constitutes no restriction to any practical
applications, since the cutoff can always be chosen large enough, such
that no violations effects can be found at momenta $p\ll \Lambda\,$.
One particular consequence of BPHZ-type subtraction schemes is then
that the calculated renormalisation constants necessarily have to
violate \eq{eq:RGrel}, since they contain contributions from momentum
regions close to the ultraviolet cutoff.
\section{Conclusion}
In this work we investigate correlation functions in Landau gauge
$SU(3)$ Yang-Mills theory. This analysis is performed in a vertex
expansion scheme for the effective action within the functional
renormalisation group approach. Besides the gluon and ghost
propagators, our approximation for the effective action includes the
self-consistent calculation of momentum-dependent dressings of the
transverse ghost-gluon, three-gluon and four-gluon vertices. Starting
from the gauge fixed tree-level perturbative action of Yang-Mills
theory, we obtain results for the correlators that are in very good
agreement with corresponding lattice QCD simulations. Furthermore, the
comparison of different vertex truncations indicates the apparent
convergence of the expansion scheme.
Special emphasis is put on the analysis of the dynamical creation of
the gluon mass gap at non-perturbative momenta. Self-consistency in
terms of the Slavnov-Taylor identities directly links this property to
the requirement of IR irregularities in the correlation functions. The
source of these irregularities is easily traced back to the
IR-divergent ghost propagator for the scaling solution. In the
decoupling-type solutions, the source of these irregularities is
harder to identify, where the creation of diagrammatic infrared
irregularities is ruled out by general arguments. Within our
truncation, we can exclude irregularities of non-diagrammatic origin
in the purely transverse subsystem. Hence it is necessary to solve
the longitudinal system to answer whether the required irregularities
are generated for decoupling-type solutions, which is not done in this
work. Nevertheless, we are able to produce decoupling-type solutions
by invoking two consistent criteria, which allow for the
differentiation between confining and Higgs-like solutions. The
decoupling-type solutions are bound by the solution that shows the
minimal mass gap, which is also the solution with minimal back-bending
of the gluon propagator.
\acknowledgments We thank Markus Q. Huber, Axel Maas, Fabian
Rennecke, Andre Sternbeck and Richard Williams for discussions as well
as providing unpublished data. This work is supported by EMMI, the
grant ERC-AdG-290623, the FWF through Erwin-Schr\"odinger-Stipendium
No.\ J3507-N27, the Studienstiftung des deutschen Volkes, the DFG
through grant STR 1462/1-1, and in part by the Office of Nuclear
Physics in the US Department of Energy's Office of Science
under Contract No. DE-AC02-05CH11231.
|
\section{Introduction}
Energy is a limited resource which faces additional challenges due to recent efficiency and de-carbonization goals worldwide. An important component of the ongoing process is the improvement in the energy management systems in residential and commercial buildings, which account for $30-40\%$ of the total energy demand in the developed world \cite{Nejat2015843}. Buildings are complex systems composed by a different number of devices and appliances, such as refrigerators, microwaves, cooking stoves, washing machines etc. However, there are also a number of sub-systems, e.g. electric heating, lighting. Even there are many influencing factors in building energy consumption, some patterns can be clearly identified and used further to improve demand side management systems and demand response (DR) programs~\cite{6687966}. Identifying and aggregating the flexibility resource at the community level can decrease the end-user energy bill. Concomitantly, as a long-term benefit, flexibility can lead also to emission reductions, and lower investments in transmission and distribution grid infrastructure. Therefore, the role of end-users and their available flexibility is becoming increasingly important in the Smart Grid context.
\footnote{\textbf{This article is a pre-print version.} Please cite this article as: E.~Mocanu, P.~H.~Nguyen and M.~Gibescu, \textit{Energy Disaggregation for Real-Time Building Flexibility Detection}, IEEE Power and Energy Society General Meeting, Boston, USA, 2016}
One possible way to detect building flexibility in real-time is by performing energy disaggregation. Disaggregation refers to the extraction of appliance level energy signals from an aggregate, or the whole-building, energy consumption signal. Often only this aggregated signal is made available via the smart meter infrastructure to the grid operator, due to privacy concerns of the end user. This new approach should open new paths towards better planning and operation of the smart grid, helping the transition of end-users from a passive to an active role. In addition, informing the end-user in real-time, or near real-time, about how much energy is used by each appliance can be a first step in voluntarily decreasing the overall energy consumption.
Introduced by W. Hart\cite{192069} in the early 1980s, the Non-Intrusive Load Monitoring (NILM) problem has nowadays several solutions for residential buildings. Traditional approaches for the energy disaggregation problem (or NILM problem) start by investigating if the device is turned on/off \cite{97667}, and followed by many steady-state methods \cite{1192027} and transient-state methods \cite{1192027} aiming to identify more complex appliance patterns. In the same time, advance building energy managements systems are looking beyond quantification of the energy consumption by including fusion information such as, the acoustic sensors to identify the operational state of the appliances \cite{Guvensan20131539}, the motion sensors, the frequency of the appliance used \cite{siam2011}, as well as time and appliance usage duration\cite{NIPS2010kolter, siam2011}. A more comprehensive discussion about these can be found in recent reviews, such as \cite{5618423,5735484,s121216838}. Moreover, new data analytics challenges arise in the context of an increasing number of smart meters, and consequently, a big volume of data, which highlights the need of more complex methods to analyze and take benefit of the fusion information \cite{6903193}. More recent researches have explored a wide range of different machine learnings methods, using both supervised and unsupervised learning, such us sparse coding \cite{NIPS2010kolter}, clustering \cite{5759180, 6939461} or different graphical models (e.g. Factorial Hidden Markov models (FHMM)\cite{siam2011}, Factorial Hidden Semi-Markov Model (FHSMM) \cite{siam2011}, Conditional FHMM \cite{siam2011}, Conditional Factorial Hidden Semi-Markov Model (CFHSMM)\cite{siam2011}, additive FHMM \cite{AISTATS2012_KolterJ12} or Bayesian Nonparametric Hidden Semi-Markov Models \cite{Johnson}) to perform energy disaggregation. Still, there is an evident challenge to develop an accurate solution that could perform well for every type of appliance.
In this paper, the aim is to perform real-time flexibility detection using energy disaggregation techniques. Therefore, the key methodological contribution of this paper is a machine learning based tool for exploiting the building energy disaggregation capabilities in an online manner. Our contributions can be summarized as follows. Firstly, we investigate the use of classification methods to perform energy disaggregation. Consequently, a comparison is performed between four widely-used classification methods, namely Naive Bayes (NB), K-Nearest Neighbors (KNN), Support Vector Machine (SVM) and AdaBoost. Secondly, we introduce a Restricted Boltzmann Machine (RBM) to perform automatic feature extraction in order to improve the performance of the four classification methods discussed. We validate our proposed approach by using a real measurement database, specifically conceived for energy disaggregation, i.e. the REDD~\cite{kolter2011}.
The remaining of this paper is organized as follows. Section~\ref{Sec:problem} introduces the problem description. Section~\ref{Sec:methods} describes our proposed approach for the energy disaggregation problem. In Section \ref{Sec:Exp} the experimental validation of the proposed methods is detailed and Section~\ref{Sec:con} concludes the paper.
\section{Problem Formulation and Methodology}\label{Sec:problem}
This section details the problem definition targeted in this paper. In one unified framework, we split the problem into two parts, where first the energy disaggregation problem is solved, and then an identification procedure is carried out to analyze the potential of building demand flexibility.
The proposed solution for energy disaggregation is addressed using four different classification methods. More formally, let us define an input space $\mathcal{D}$ and an output space (label space) $\mathcal{B}$. The question of learning is reduced to the question of estimating a functional relationship of the form $\mathcal{C} : D \rightarrow B$, that is a relationship between inputs and outputs. A classification algorithm is a procedure that takes the training data as input and outputs a classifier $\mathcal{C}$. The goal is then to find a $\mathcal{C}$ which makes “as few errors as possible”. Intuitively, the learned classifier should be based on \textit{enough} training examples, \textit{fit} the training example and should be \textit{simple}. Moreover, classification can be thought of as two separate problems: binary classification and multi-class classification.
\begin{figure}[h!]
\centering
\includegraphics[scale=0.46]{./classisfication.pdf}
\caption{Energy disaggregation}
\label{fig:class}
\end{figure}
In our specific case, the $B$ space is given by the electrical devices in the building, and the $D$ space is given by the aggregated electrical energy consumption of the building. In Figure~\ref{fig:class} the flow diagram of the energy disaggregation procedure is depicted. Firstly, using data from $n$ buildings we derive a corresponding model for each device inside them. Furthermore these binary classification models are used to automatically classify, whether a given device is active at any specific moment in time, by using the building's total electrical energy consumption profile.
\section{Proposed Methods}\label{Sec:methods}
In this section, we firstly briefly describe the four classification methods to perform energy disaggregation, these methods being part of the supervised learning paradigm. Secondly, we introduce the mathematical details of the Restricted Boltzmann Machine used to perform automatic features extraction, this method being part of the unsupervised learning paradigm.
\subsection{Classification methods}
For the classification problem, plenty of deterministic or probabilistic algorithms are known, where every observation is analyzed into a set of quantifiable properties, such as Naive Bayes~\cite{bishopbook}, Support Vector Machine~\cite{vapniksvm}, AdaBoost~\cite{Freund99ashort}, Random Forest Trees and so on. Prior studies tried to determine the most accurate classification method, as is shown in\cite{CaruanaNiculescu}, but currently there is not a general consensus in the favor of a particular method.
\subsubsection{Naive Bayes} is one of the most simple classification method based on a strong independence assumptions between the input features. Despite these relatively naive assumptions, with a training phase extremely easy to implement and fast computational time, Naive Bayes classifiers often outperform more sophisticated alternatives.
\subsubsection{k-Nearest Neighbors} is a non-parametric method used for classification.
The standard version of KNN used in this paper performs successively two steps. Specifically, the clusters are construct by partitioning the $k$-nearest neighbors based on a distance measure (i.e. Euclidean distance), followed by an update rule, such that the majority of those $k$-nearest neighbors decide the class of the next observations.
\subsubsection{AdaBoost}
it stands for Adaptive Boosting, and is a machine learning algorithm, which was proposed in the computational learning theory field by Y. Freund and R. Schapire~\cite{Freund99ashort}. AdaBoost method solves the classification problem using a linear combination of many weak classifiers into a single strong classifier. Acting as an expert, boosting often does not suffer from overfitting and it is worth to investigate in the context of our challenging dataset.
\subsubsection{Support Vector Machine (SVM)}
is introduced by Vapnik in 1995~\cite{vapniksvm} and becomes very popular for solving problems in classification, regression, and novelty detection. An important characteristic of SVM is that the determination of the model parameters corresponds to a convex optimization problem, and so any local solution is also a global optimum. This guarantee comes with some computational cost but also with a better robustness.
\subsection{Restricted Boltzmann Machine}\label{tab:rbm}
Restricted Boltzmann Machine is a two-layer generative stochastic neural network which is capable to learn a probability distribution over its set of inputs~\cite{originalrbm}. Such a model does not allow intra-layer connections between the units, and it allows just inter-layer connections. In fact, any unit from one layer has undirected connections to all the units from the other layers. Up to now, various types of restricted Boltzmann machines are already developed and successfully applied in different applications~\cite{eu}. Despite their differences, almost all of these architectures preserve RBMs characteristics. To formalize a restricted Boltzmann machine, and its variants, three main ingredients are required, namely an energy function providing scalar values for a given configuration of the network, the probabilistic inference and the learning rules required for fitting the free parameters.
Thus, a RBM consists in two binary layers, the visible layer, $\mathbf{v}=[v_1, v_2, .., v_{n_v}]$, in which each neuron represents one dimension (feature) of the input data and the hidden layer, $\mathbf{h}=[h_1, h_2, .., h_{n_h}]$, which represents hidden features extracted automatically by the RBM model from the input data, where $n_v$ is the number of visible neurons and $n_h$ is the number of the hidden neurons. Each visible neuron $i$ is connected to any hidden neuron $j$ by a weight, i.e. $W_{ij}$. All these weights are stored in a matrix $\mathbf{W}\in\mathcal{R}^{{n_v}\times{n_h}}$, where $\mathcal{R}$ is the set of real numbers, in which the rows represent the visible neurons and the columns the hidden ones. Finally, each visible neuron $i$ has associated a bias $a_i$ which is stored in a vector $\mathbf{a}=[a_1, a_2, .., a_{n_v}]$. Similarly, the hidden neurons have biases which are stored in a vector $\mathbf{b}=[b_1, b_2, .., b_{n_h}]$. Further on, we will note with $\mathbf{\Theta}=\{\mathbf{W},\mathbf{a},\mathbf{b}\}$ a set which represent the union of all free parameters of a RBM (i.e. weights and biases). Formally, the energy function of a RBM for any state $\{\mathbf{v},\mathbf{h}\}$ can be computed by summing over all possible interactions between neurons, weights and biases, as folows:
\begin{equation}
E(v,h)=-\sum_{i=1}^{n_v} \sum_{j=1}^{n_h} v_{i}h_{j}W_{ij}-\sum_{i=1}^{n_v} v_{i}a_{i}-\sum_{j=1}^{n_h}h_{j}b_{j}
\label{Eq:EnergyRBM}
\end{equation}
where the term $\sum\displaystyle_{i=1}^{n_v} \sum\displaystyle_{j=1}^{n_h} v_{i}h_{j}W_{ij}$ is given by the total energy between the neurons from different layers, while $\sum\displaystyle_{i=1}^{n_v}v_{i}a_{i}$ represents the energy of the visible neurons and $\sum\displaystyle_{j=1}^{n_h}h_{j}b_{j}$ is the energy of the hidden neurons.
The inference in a RBM means to determine two conditional distributions. For any hidden or visible neuron this can be done just by sampling from a sigmoid function, as shown below:
\begin{align}
p(h_j=1|\mathbf{v},\mathbf{\Theta})=\frac{1}{1+e^{-(b_j+\sum\displaystyle_{i=1}^{n_v} v_{i}w_{ij})}}\\
p(v_i=1|\mathbf{h},\mathbf{\Theta})=\frac{1}{1+e^{-(a_i+\sum\displaystyle_{j=1}^{n_h} h_{j}w_{ij})}}
\label{Eq:XBMinferencevisible}
\end{align}
To learn the parameters of a RBM model there are more variants in the literature (e.g. persistent contrastive divergence, parallel tempering~\cite{desjardins:aistats2010}, fast persistent contrastive divergence~\cite{Tieleman:2009:UFW:1553374.1553506}). Almost all of them being derived from the Contrastive Divergence (CD) method proposed by Hinton in~\cite{hintoncd}. For this reason, in this paper, we briefly describe and use just the original CD method. CD is an approximation of the maximum likelihood learning, which is practically intractable in a RBM. Thus, while in maximum likelihood the learning phase minimizes the Kullback-Leiber (KL) measure between the distribution of the input data and the model approximation, in CD the learning follows the gradient of:
\begin{equation}
CD_{n} \propto D_{KL}(p_{0}(\textbf{x})||p_{\infty}(\textbf{x}))-D_{KL}(p_{n}(\textbf{x})||p_{\infty}(\textbf{x}))
\end{equation}
where, $p_{n}(.)$ represents the resulting distribution of a Markov chain running for $n$ steps. Furthermore, the general update rule of the free parameters of a RBM model is given by:
\begin{equation}
\Delta\mathbf{\Theta}_{\tau+1}=\rho\Delta\mathbf{\Theta}_{\tau}+\alpha(\nabla\mathbf{\Theta}_{\tau+1}-\xi\mathbf{\Theta}_{\tau})
\label{eq:genuprule}
\end{equation}
where $\tau$, $\alpha$, $\rho$, and $\xi$ represent the update number, learning rate, momentum, and weights decay, respectively, as thoroughly discussed in~\cite{hintontrain}. Moreover, $\nabla\mathbf{\Theta}_{\tau+1}$ for each free parameter may be computed by deriving the energy function from Equation~\ref{Eq:EnergyRBM} with respect to that parameter, as detailed in~\cite{hintoncd}, yielding:
\begin{align}
\nabla w_{ij}&\propto{\langle}v_i h_j{\rangle}_{0}-{\langle}v_i h_j{\rangle}_{n}\\
\nabla a_{i}&\propto{\langle}{v_i}{\rangle}_{0}-{\langle}v_i{\rangle}_{n}\\
\nabla b_{j}&\propto{\langle}{h_j}{\rangle}_{0}-{\langle}h_j{\rangle}_{n}
\label{Eq:GXBMupdateb}
\end{align}
with $\langle\cdot\rangle_n$ being the distribution of the model obtained after $n$ steps of Gibbs sampling in a Markov Chain which starts from the original data distribution $\langle\cdot\rangle_0$.
\section{Experimental Results}\label{Sec:Exp}
In this section we analyze and validate our proposed approach using a real-world database, namely \textit{The Reference Energy Disaggregation Dataset} (REDD), described by Kolter and Johnson in~\cite{kolter2011}. This data was chosen as it is an open dataset\footnote{http://redd.csail.mit.edu/, Last visit November 5th, 2015} collected specifically for evaluating energy disaggregation methods. It contains aggregated data recorded from six buildings over few weeks sampled at 1 second resolution, together with the specific data for all appliances of each building at 3 seconds resolution.
In the first set of experiments, we study the performance of the classification methods (i.e. Naive Bayes, K-Nearest Neighbors, Support Vector Machine and AdaBoost) for detecting the activation of four appliances (i.e. refrigerator, electric heater, washer-dryer, dishwasher), specifically chosen for their ability to provide demand-side flexibility. Furthermore, in the second stage we demonstrate the improvement in the accuracy of the classification after a Restricted Boltzmann Machine is used for automatic feature extraction. Finally, assuming the aforementioned four appliances shiftable in time, we discuss the possible benefits of real-time flexibility detection.
The experiments were performed in the MATLAB\textsuperscript{\textregistered} environment using the methods described in Section~\ref{Sec:methods}. For the classification methods we have used the optimized parameters from the machine learning toolbox (e.g. SVM with radial kernel function). For each appliance we have built a separate binary classification model for every classification method. The input at every moment in time is given by a window of 10 consecutive time steps from the aggregated building consumption, while the output was represented by the activation of the appliance (i.e. on/off status). In all the experiments performed, we have trained the models on 5 buildings (i.e. 2, 3, 4, 5, and 6) and we have tested the models on a different building (i.e. 1). Also, as recommended in~\cite{6939461}, we have applied a median filter of 6 samples to make the data smoother.
For the feature extraction procedure we have implemented RBMs with the following parameters: 20 hidden neurons and 10 visible neurons (representing the time window of 10 consecutive time steps). After a short fine tuning procedure, the learning rate was set to $10^{-2}$, the momentum was set to 0.5, and the weight decay was set to 0.0002. We trained the RBM models for 25 epochs, and after that we have used the probabilities of the hidden neurons as inputs for the classification methods.
In order to characterize as fairly as possible the accuracy of the models proposed to classify the appliance activation we have calculated the classifier accuracy as follows:
\begin{equation}
Accuracy=\frac{\sum_{i=1}^{n} A_{ii} }{\sum_{i=1}^{n}\sum_{j=1}^{n} A_{ij}}
\end{equation}
where $A$ is the confusion matrix (also known as a contingency table or an error matrix), $A_{ii}$ represents the positive true value and the denominator represents the total number of data used in the classification procedure. This quantifies the proportion of the total number of instances that were correctly classified.
\subsection{Energy disaggregation}
In this subsection, we first perform a comparison between the four classification methods, namely Naive Bayes (NB), k-Nearest Neighbors (KNN), Support Vector Machine (SVM) and AdaBoost (AB). Table~\ref{tab:2class} summarizes the classification accuracy for different building electrical components, such as refrigerator, electric heater, washer-dryer and dishwasher. For a better insight into the results, an example of the energy consumption for the appliances corresponding to building 1 (the test data) is depicted in Figure~\ref{fig:devices}.
\begin{table}[ht!]
\caption{ Results showing accuracy [\%] for each of Naive Bayes, KNN, SVM and AdaBoost to classify an appliance versus all data.}
\centering
\begin{tabular}{|c|cccc|}
\hline
Appliance&NB&KNN&SVM&AdaBoost\\
\hline
refrigerator&52.18\%&67.36\%&67.45\%&\textbf{87.13\%}\\
electric heater&93.01\%&97.79\%&\textbf{98.84\%}&94.74\%\\
washer dryer&92.04\%&\textbf{96.17\%}&78.27\%&95.56\%\\
dishwasher&97.52\%&\textbf{98.11\%}&97.74\%&97.77\%\\
\hline
\end{tabular}
\label{tab:2class}
\end{table}
\begin{figure}[h!]
\centering
\includegraphics[scale=0.63]{./divices1-crop.pdf}
\caption{An example of energy consumption in Building 1 over 30 minutes for refrigerator, electric heater, washer dryer and dishwasher.}
\label{fig:devices}
\end{figure}
Furthermore, to improve the classification performance, we have employed the automatic features extraction procedure by using the Restricted Boltzmann Machine as described in Section\ref{tab:rbm}. Next, the extracted features are used as inputs for the classification methods.
We have tested and validated this approach on the same electrical appliances as before, as shown in Table~\ref{tab:2classrbm}.
\begin{table}[ht!]
\caption{ Results showing accuracy [\%] for each of Na$\ddot{\i}$ve Base, KNN, SVM and AdaBoost with RBM extension, to classify an appliance versus all data.}
\centering
\begin{tabular}{|c|cccc|}
\hline
Appliance&NB-RBM&KNN-RBM&SVM-RBM&AB-RBM\\
\hline
refrigerator&64.78\%&\textbf{96.72\%}&84.45\%&91.02\%\\
electric heater&99.13\%&99.81\%&\textbf{99.86}\%&99.84\%\\
washer dryer&99.14\%&97.31\%&89.23\%&\textbf{99.27}\%\\
dishwasher&97.64\%&98.43\%&\textbf{98.67}\%&97.82\% \\
\hline
\end{tabular}
\label{tab:2classrbm}
\end{table}
It can be observed that in all situations, the use of RBMs has improved the accuracy for each classifier. This culminates with an improvement of around 30\% for the case of the refrigerator classified with KNN, from 67.36\% initial accuracy, up to 96.72\% accuracy after the use of RBM. It is worth mentioning, that the imbalanced number of data points in every class suggests that a more deeper data mining analysis may be useful. In term of computational complexity the training time varies from the range of few seconds in the case of KNN up to few minutes in the case of SVM. In the testing phase, to classify all the data points considered (i.e. 745868 instances per year per appliance) each of the methods has ran in approximately 1 second, except SVM which ran in 4-5 seconds. Overall, this yields an execution time of a few microseconds per data point making the approach suitable for a large range of real-time applications.
\subsection{Flexibility detection}
The energy disaggregation results may be used further in a large number of applications, as reported in 2015 by the US Department of Energy in an extensive report \cite{Report} which aims to characterize the actual performance of energy disaggregation solutions used in both the academic research and in commercial products.
Most importantly, our results may be used to detect in real-time the building flexibility available. We observed that approximately 17\% of the total energy consumption for building 1 is used by the four disaggregated appliances, such as refrigerator 11.72\%, electric heater 5.08\%, washer-dryer 0.0007\% and dishwasher 0.9\% respectively. More statistical details about these appliances for building 1 are presented in Table~\ref{fig:devices}.
\begin{table}[ht!]\label{tab:summary}
\footnotesize
\centering
\caption{General characteristics of the building 1 appliances used in the experiments.}
\begin{tabular}{ccc}
&Mean&Standard deviation\\
\hline
\hline
refrigerator&56.41&86.65\\
electric heater&24.44&148.16\\
wash dryer&0.11&0.96\\
dishwasher &4.30&43.54\\
\hline
\end{tabular}
\label{tab:summary}
\vspace{-3mm}
\end{table}
A visual examination of the results, assuming that all the four appliances studied have smart time-shifting capabilities, and a detection accuracy of over 96\% in all the experiments, show a significant peak reduction. As by example, in Figure~\ref{fig:devices} the inflexible load is represented by the difference between the total energy consumption signal and the sum of our disaggregated signals over 24 hours. In this case, we may observe that the average buildings flexibility is 23.21\%.
\vspace{-3mm}
\begin{figure}[h!]
\centering
\includegraphics[scale=0.61]{./flex2-crop.pdf}
\caption{An example of electrical energy consumption in buildings over one day for inflexible load and flexible load (refrigerator, electric heater, washer dryer and dishwasher).}
\label{fig:devices}
\vspace{-5mm}
\end{figure}
\section{Conclusion}\label{Sec:con}
In this paper a novel tool capable to perform accurate energy disaggregation for real-time flexibility detection is proposed. A comparison between four existing classification methods was performed. Aiming at enhancing the quality of such estimates as well as at increasing the accuracy of energy disaggregation, a method for automatic features extraction is proposed, using Restricted Boltzmann Machines. By incorporating the RBM for feature extraction, each of the classification methods, i.e. Naive Bayes, k-Nearest Neighbors, Support Vector Machine and AdaBoost, has outperformed its non-preprocessed counterpart. The experimental validation performed on the REDD dataset shows that KNN- RBM has the best trade-off between accuracy and speed.
\section*{Acknowledgment}
This research has been funded by NL Enterprise Agency under the TKI Switch2SmartGrids project of Dutch Top Sector Energy.
\vspace{-1mm}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{intro}
Paracomplex geometry was introduced by Libermann \cite{Libermann:paracomplex1952}; in analogy with complex geometry, it is defined by a tensor $K$ with $K^2=\id$, whose eigenspaces are integrable distributions of dimension $n$. The local geometry is that of a product, but things become more complicated if a metric enters the picture. The natural compatibility condition to impose is that $K$ be an anti-isometry, so that $F=g(K\cdot, \cdot)$ defines a two-form; if this form is closed, the metric is said to be parak\"ahler. Such a metric is necessarily of neutral signature; its holonomy is contained in $\GL(n,\R)$, endowed with its standard action on $\R^n\oplus(\R^n)^*$.
Parak\"ahler manifolds carry a natural bilagrangian structure; as such, they form a natural object of study in symplectic geometry (see \cite{EtayoSantamaria,HarveyLawson}); they also provide a natural setting for the study of the mean curvature flow, which is proved to preserve Lagrangian submanifolds in \cite{ChursinSchaferSmoczyk}, under some assumptions on the curvature. Parak\"ahler geometry also finds applications in physics (\cite{CMMS:VectorI2004}) and in the study of optimal transport (\cite{KimMcCannWarren}). We refer to \cite{CruceanuFortunyGadea:survey1996} for a survey; a more recent reference is \cite{AlekseevskyMedoriTomassini:Homogeneous2009}.
Like in K\"ahler geometry, the Ricci tensor of a parak\"ahler manifold is given by $-\frac12 dd^c\log \abs{\phi}^2$, where $\phi$ is a local holomorphic volume form (see \cite{HarveyLawson}); this indicates that the existence of a parallel paracomplex volume form forces the Ricci to be zero. In terms of the restricted holonomy $\hol_0$, this condition is equivalent to $\hol_0\subset\SL(n,\R)$; thus, pseudoriemaniann metrics of neutral signature with (restricted) holonomy contained in $\SL(n,\R)$ are Ricci-flat.
This observation leads us to consider the smaller structure group $\SL(n,\R)$ rather than $\GL(n,\R)$; we do not impose integrability conditions, so that the eigendistributions of $K$ are not necessarily integrable, nor is $F$ necessarily closed. The failure of the holonomy condition is measured by a tensor called \emph{intrinsic torsion}; it follows from the above that the intrinsic torsion determines the Ricci tensor. An explicit formula to this effect is the main result of this paper (Theorems~\ref{thm:ricci}, \ref{thm:RicciS2}). We note that similar situations have been studied in \cite{Bryant:remarks} and \cite{BedulliVezzoni:SU3} for the Riemannian holonomy groups $\Gtwo$ and $\SU(3)$; see also \cite{CabreraSwann} for similar computations relative to the group $\Sp(n)\Sp(1)$. Our methods, however, are more akin to those of \cite{Conti:qc}, and they allow us to find an explicit formula valid in any dimension.
Intrinsic torsion relative to the structure group $\GL(n,\R)$ can be identified with the covariant derivative of $F$ under the Levi-Civita connection; the results of \cite{GadeaMasque} imply that, for $n\geq 3$, $\GL(n,\R)$-intrinsic torsion decomposes into eight components, corresponding to eight non-isomorphic $\GL(n,\R)$-modules $W_1\oplus\dotsb\oplus W_8$. Under $\SL(n,\R)$, these components remain irreducible, but two extra components appear, giving rise to ten intrinsic torsion classes. This situation is somewhat different from that of almost hermitian geometry, where in six dimensions the Gray-Hervella intrinsic torsion classes become reducible upon reducing from $\LieG{U}(3)$ to $\SU(3)$ (see \cite{GrayHervella,ChiossiSalamon}).
Beside the fundamental form $F$, a manifold with an $\SL(n,\R)$-structure carries decomposable $n$-forms $\alpha,\beta$, characterized as volume forms on the two distributions defined by $K$. In fact, the structure group $\SL(n,\R)$ is the largest subgroup of $\GL(2n,\R)$ that fixes the corresponding elements $F$, $\alpha$ and $\beta$ of $\Lambda(\R^{2n})$. These differential forms are closed if and only if the intrinsic torsion is zero (see Proposition~\ref{prop:forms}); structure groups satisfying this condition are known as strongly admissible (see \cite{Bryant}). This allows us to restate our formula for the Ricci tensor purely in terms of exterior derivatives (Theorem~\ref{thm:restateRicViaForms}).
The language of $\SL(n,\R)$-structure enables us to construct a cohomological invariant of paracomplex manifolds, analogous to the first Chern class in complex geometry, which obstructs the existence of Ricci-flat parak\"ahler metrics; we obtain a sufficient topological condition for its vanishing (Theorem~\ref{thm:aubinyau_dei_poveri}). One difference with the K\"ahler case is that compact parak\"ahler Einstein manifolds are necessarily Ricci-flat, as we show in Proposition~\ref{prop:pkeinsteincompact}.
Thus, unlike in the compact K\"ahler case, the invariant does not quite describe whether compatible parak\"ahler Einstein metrics have zero or non-zero scalar curvature. This is also not true in the non-compact setting, as we show by producing a non-compact parak\"ahler Einstein manifold with nonzero scalar curvature where the invariant is zero (Example~\ref{ex:Einstein}).
We then turn to the construction of Einstein examples. Examples of compact, Ricci-flat parak\"ahler manifolds appear in \cite{HarveyLawson}; these examples, however, are also flat. Non-flat examples on compact nilmanifolds were constructed by the second author in \cite{RossiPhD}; a similar example appears in Section~\ref{sec:examples}. These examples are homogeneous, marking a difference between Riemannian and pseudoriemannian geometry: by \cite{AlekseevskiKimelFel}, homogeneous Riemannian Ricci-flat manifolds are flat. Parak\"ahler Einstein manifolds with non-zero scalar curvature that are homogeneous under a semisimple Lie group are classified in \cite{AlekseevskyMedoriTomassini:Homogeneous2009}.
Outside the parak\"ahler setting, one significant intrinsic torsion class consists of nearly parak\"ahler structures, characterized by the fact that $\nabla K$ is skew-symmetric in the first two indices. In our language, this means that the $\GL(n,\R)$ intrinsic torsion is contained in $W_1\oplus W_5$ at each point; we say that its intrinsic torsion \emph{class} is $\mathcal{W}_1+\mathcal{W}_5$. Examples of Einstein and Ricci-flat nearly parak\"ahler metrics are constructed in \cite{IvanovZamkovoy:parahermitian}, \cite{CortesSchafer} and \cite{Schafer:conical}. In fact, it follows easily from our formula (see Corollary~\ref{cor:nk}) that nearly parak\"ahler manifolds of dimension six are automatically Einstein, as originally proved in \cite{IvanovZamkovoy:parahermitian}.
As a more restrictive condition, we study the intrinsic torsion class $\mathcal{W}_1$; we show that this is Ricci-flat in all dimensions. Considering left-invariant structures on nilpotent Lie groups, we obtain several explicit examples in dimension eight (Theorem~\ref{thm:nkricciflat}). These structures are automatically nearly parak\"ahler and Ricci-flat; they are also non-flat, and the underlying manifold is compact. We note that the previously known examples of Ricci-flat nearly parak\"ahler manifolds were either non-compact (\cite{Schafer:conical}) or flat (\cite{CortesSchafer}).
The intrinsic torsion class $\mathcal{W}_2$ is also Ricci-flat. This leads to a counterexample of the paracomplex version of the Goldberg conjecture as stated in \cite{Matsushita}, asserting that a compact, Einstein almost parak\"ahler manifold is necessarily parak\"ahler (Proposition~\ref{prop:goldberg}). Notice that the K\"ahler version of the conjecture is known to hold for non-negative scalar curvature (\cite{Sekigawa}); our example is Ricci-flat, showing that the paracomplex situation is different.
The intrinsic torsion class $\mathcal{W}_3$ is not Ricci-flat, but we are able to construct a compact Ricci-flat example on a nilmanifold. However, the class $\mathcal{W}_4$ is different: a nilpotent Lie group with an invariant structure with intrinsic torsion in $W_4+W_8$ is necessarily parak\"ahler (Proposition~\ref{prop:nilpotentw4w8}).
Observing that changing the sign of $K$ has the effect of swapping $W_i$ with $W_{i+4}$, this concludes the analysis of ``pure'' intrinsic torsion classes. By taking products, it follows that all intrinsic torsion classes that do not contain $\mathcal{W}_4$ or $\mathcal{W}_8$ can be realized as the intrinsic torsion class of a nilmanifold with a non-flat, Ricci-flat metric (Proposition~\ref{prop:ricciflatproducts}).
\section{The structure group $\GL(n,\R)$}
An almost paracomplex structure on a manifold of dimension $2n$ is a decomposition of the tangent space in two subbundles of rank $n$. The tangent space is then modeled on a direct sum
\begin{equation}
\label{eqn:linearmodel}
T=V\oplus H,
\end{equation}
where $V$ and $H$ are real vector spaces of dimension $n$; explicitly, we shall fix a basis $e_1,\dotsc, e_{2n}$ of $T$ with
\[V=\Span{e_1,\dotsc, e_n},\quad H=\Span{e_{n+1},\dotsc, e_{2n}},\]
and denote by $e^1,\dotsc, e^{2n}$ the dual basis of $T^*$. In these terms, we can think of an almost paracomplex structure as a $\GL(V)\times \GL(H)$-structure. In analogy with complex geometry, one considers an endomorphism of $T$ with $K^2=\id$, namely
\[K =\id_V - \id_H = e^i\otimes e_i - e^{n+j}\otimes e_{n+j};\]
here and in the sequel, summation over repeated indices is implied; we adopt the convention that lower case indices range from $1$ to $n$, and upper case indices range from $1$ to $2n$.
It is clear that a manifold of dimension $2n$ admits an almost paracomplex structure if and only if it admits a distribution of rank $n$; for example, $S^{2n}$ does not have an almost paracomplex structure \cite[Theorem~27.18]{Steenrod}. Thus, the long-standing problem of whether the six-dimensional sphere admits an integrable complex structure has a trivial answer in the paracomplex setting.
Like in almost complex geometry, differential forms on an almost paracomplex structure can be decomposed according to type via
\[\Lambda^k T^*=\bigoplus_{p+q=k} \Lambda^{p,q}, \quad \Lambda^{p,q} = \Lambda^pV^*\otimes \Lambda^qH^*;\]
in the literature, one also finds the notations $TM=T_+M\oplus T_-M$ and $\Lambda^{p,q}_\pm$.
An almost paracomplex structure is said to be \emph{paracomplex} or \emph{integrable} if the two rank $n$ distributions are integrable. By the Frobenius Theorem, this is equivalent to requiring that the exterior derivative have the form
\[d\colon \Lambda^{p,q}\to \Lambda^{p+1,q}+\Lambda^{p,q+1},\]
or to the vanishing of the Nijenhuis tensor
\[N(X,Y)=[X,Y]+[KX,KY]-K[KX,Y]-K[X,KY].\]
In the language of $G$-structures, integrability can be expressed in terms of \emph{intrinsic torsion}. Recall that the intrinsic torsion of a $G$-structure takes values in the cokernel of the linear map $\partial_G$, defined as the restriction to $T^*\otimes\lie{g}$ of the alternating map
\[\partial\colon T^*\otimes \gl(T)\to \Lambda^2T^*\otimes T, \quad e^i\otimes (e^j\otimes e_k) \to e^{ij}\otimes e_k;\]
more precisely, the intrinsic torsion is obtained by projecting on this space the torsion of any connection on the $G$-structure.
If $\partial_G$ has a left inverse $s$, any $G$-structure has a unique \emph{minimal} connection, namely one with torsion in $\ker s$. In the present case, the alternating map is not injective, but we still have the following:
\begin{proposition}
\label{prop:paracomplexit}
Every $\GL(V)\times \GL(H)$-structure admits a connection with torsion taking values in
\begin{equation}
\label{eqn:paracomplexit}
\Lambda^2V^*\otimes H+\Lambda^2H^*\otimes V;
\end{equation}
the torsion $\Theta$ of any such connection is related to the Nijenhuis tensor via
\[\Theta(X,Y)=-\frac14 N(X,Y).\]
\end{proposition}
\begin{proof}
The representation $\Lambda^2T^*\otimes T$ splits as
\[\Lambda^{2,0}\otimes V+\Lambda^{2,0}\otimes H+\Lambda^{1,1}\otimes V+\Lambda^{1,1}\otimes H+\Lambda^{0,2}\otimes V+\Lambda^{0,2}\otimes H,\]
and
\[T^*\otimes (\gl(V)+\gl(H))= (S^2V^*\otimes V+S^2H^*\otimes H) + (\Lambda^{2,0}\otimes V + \Lambda^{0,2}\otimes H+\Lambda^{1,1}\otimes T),\]
where the first component represents the kernel of the alternating map. It follows that the restriction of $\partial$ to the second component has a left inverse $s$ with kernel \eqref{eqn:paracomplexit}. Therefore, if $\overline\omega$ is any connection on the $\GL(V)\times \GL(H)$-structure with torsion $\overline\Theta$, the connection $\omega=\overline\omega-s(\overline\Theta)$
has torsion $\overline\Theta - \partial (s(\overline\Theta))\in \ker s.$
Applying the definition
$\Theta(X,Y)=\nabla_X Y-\nabla_Y X-[X,Y],$
and using the fact that $V$ and $H$ are preserved by the connection $\omega$, we obtain the required relation.
\end{proof}
\begin{remark}
In terms of an adapted coframe $e^1,\dotsc,e^{2n}$, the torsion of a connection such as in Proposition~\ref{prop:paracomplexit} may be written as
\[\Theta = (de^i)^{0,2}\otimes e_i + (de^{n+i})^{2,0}\otimes e_{n+i}.\]
\end{remark}
\smallskip
Given an almost paracomplex structure, a pseudoriemannian metric $g$ is called \emph{almost parahermitian} if
\[g(KX,KY)=-g(X,Y).\]
Such a metric is necessarily of neutral signature $(n,n)$; relative to \eqref{eqn:linearmodel}, it determines an isomorphism $V\cong H^*$. Accordingly, the structure group is reduced to $\GL(n,\R)$, and the tangent space is modeled on the representation
\[T=V\oplus V^*.\]
Alternatively, we may think of an almost parahermitian structure as determined by a non-degenerate two-form $F$ of type $(1,1)$; form and metric are related via
\[F(X,Y)=g(KX,Y).\]
In this context, the basis of $T$ can always be chosen so that
\[g=e^i\odot e^{n+i}, \quad F=e^{1,n+1}+\dots + e^{n,2n},\]
where $e^i\odot e^{n+i}$ stands for $e^i\otimes e^{n+i}+e^{n+i}\otimes e^i$ and $e^{i,n+i}$ for $e^i\wedge e^{n+i}$;
using the metric $g$, we will use the identification
\[V^*\ni e^i\mapsto e_{n+i}\in H.\]
Due to the existence of the volume form $F^n$, almost parahermitian manifolds are orientable.
Whilst a paracomplex manifold is always locally a product $M\times N$, one should not think of parahermitian geometry as a fancy way to describe Cartesian products. At the topological level, this can be seen from the following:
\begin{proposition}
\label{prop:noproduct}
Let $M$ and $N$ be manifolds of dimension $n$, and assume that $TM$ is not trivial. Then the product paracomplex structure on $M\times N$ does not admit a compatible parahermitian structure.
\end{proposition}
\begin{proof}
Let
\[\pi_1\colon M\times N\to M, \quad \pi_2\colon M\times N\to N\] denote the projections. The $K$-eigenspaces for the product paracomplex structure on $M\times N$ are $\pi_1^*TM$ and $\pi_2^*TN$. If a compatible parahermitian structure exists, the vector bundles
$\pi_1^*TM \cong (\pi_2^*TN)^*$
are isomorphic. Therefore, their restrictions to a submanifold $M\times\{y\}$ are also isomorphic. However, the restriction of $\pi_1^*TM$ is equivalent to $TM$, and the restriction of $(\pi_2^*TN)^*$ is trivial, which is absurd.
\end{proof}
At a point, we can think of the structure group $\GL(n,\R)$ as the stabilizer of $F$ in $\SO(n,n)$. At the Lie algebra level, this amounts to setting $B$ and $C$ to zero in
\[\so(n,n)=\biggl\{\begin{pmatrix} A & B \\ C & -\tran{A}\end{pmatrix}\mid B=-\tran{B}, C=-\tran{C}\biggr\}.\]
Having a metric at our disposal, we can write orthogonal decompositions such as
\[\so(n,n)=\gl(n,\R)\oplus \gl(n,\R)^\perp=\Sl(n,\R)\oplus \Sl(n,\R)^\perp,\]
where
\[
\Sl(n,\R)=\{Tr(A)=0, B=0=C\},\quad \Sl(n,\R)^\perp=\{A=\lambda I\}.\]
It will be convenient to fix the isomorphism
\begin{equation}
\label{eqn:lambdaisso}
\Lambda^2T^*\ni \alpha\mapsto M_\alpha\in \so(n,n), \quad \langle M_\alpha(v),w\rangle =\alpha(v,w).
\end{equation}
Explicitly,
\begin{gather*}
e^{ij}\mapsto e^i\otimes e_{n+j}-e^j\otimes e_{n+i}, \qquad e^{n+i,n+j}\mapsto e^{n+i}\otimes e_j-e^{n+j}\otimes e_i,\\
e^{i,n+j}\mapsto e^i\otimes e_j-e^{n+j}\otimes e_{n+i}.
\end{gather*}
\begin{lemma}
\label{lemma:glglscalar}
Through the identification \eqref{eqn:lambdaisso}, the Lie bracket on $\so(n,n)$ satisfies
\begin{equation*}
[a_{ij}e^{ij},b_{kl}e^{n+k,n+l}] =\frac2n a_{ij}(b_{ij}-b_{ji}) e^{k,n+k} \mod \Sl(n,\R) + \gl(n,\R)^\perp.
\end{equation*}
\end{lemma}
\begin{proof}
Follows from
\[[e^{ij},e^{n+k,n+l}] =
-\delta_{il}e^{j,n+k} +\delta_{jl}e^{i,n+k} +\delta_{ik}e^{j,n+l}-\delta_{jk}e^{i,n+l}.\qedhere\]
\end{proof}
An almost parahermitian structure is called \emph{parak\"ahler} if one (hence both) of $K$ and $F$ is parallel under the Levi-Civita connection. More generally, $\nabla F$ can be identified with the intrinsic torsion of a $\GL(n,\R)$-structure; the latter is known to decompose into eight components \cite{GadeaMasque}.
All finite-dimensional irreducible representations of $\GL(n,\R)$ appear inside some
\begin{equation}
\label{eqn:Vrs}
V^{\otimes^r}\otimes (V^*)^{\otimes ^s},
\end{equation}
(see e.g. \cite{FultonHarris}). Relative to the Cartan subalgebra of diagonal matrices, let $L_i$ denote the weight that maps $\operatorname{diag}(a_1,\dotsc, a_n)$ to $a_i$. For any $k+h\leq n$ and integers
$\lambda_1\geq \dotsc\geq \lambda_k$, $\lambda_n\geq \dotsc \geq \lambda_{n-h}$,
denote by $V_{\lambda_1,\dotsc, \lambda_k}^{\lambda_{n},\dotsc, \lambda_{n-h}}$ the representation with highest weight
\[\lambda_1 L_1 + \dotsc + \lambda_k L_k - (\lambda_{n-h}L_{n-h} + \dotsc + \lambda_n L_n).\]
These representations are also irreducible under $\SL(n,\R)$, but notice that under $\SL(n,\R)$ the numbers $\lambda_1,\dotsc, \lambda_n$ are not determined uniquely by the representation.
With this notation, $V=V_1$ and $V^*=V^1$. More generally,
\[\big(V_{\lambda_1,\dotsc, \lambda_k}^{\lambda_{n},\dotsc, \lambda_{n-h}}\big)^* =V^{\lambda_1,\dotsc, \lambda_k}_{\lambda_{n},\dotsc, \lambda_{n-h}};\]
in addition we have
\[\lie{sl}(n,\R)=V_1^1, \quad \Lambda^kV = V_{ \underbrace{\scriptstyle 1,\dots,1}_{\scriptstyle{k}}}, \quad S^kV=V_k.\]
We shall say that a representation has \emph{type} $(k,h)$ if it is the sum of irreducible representations of the form $V_{\lambda_1,\dotsc, \lambda_k}^{\lambda_{n},\dotsc, \lambda_{n-h}}$. In terms of Young diagrams, this says that the rows from $k+1$ to $n-h-1$ have the same length. It is clear that $V^{\otimes^r}$ has type $(r,0)$; dually, $(V^*)^{\otimes^s}$ has type $(0,s)$. It now follows easily from the Littlewood-Richardson rule that \eqref{eqn:Vrs} has type $(r,s)$.
We can think of a representation of type $(r,s)$ as a representation of $\GL(n,\R)$ for any choice of $n>r+s$. The decomposition into irreducible components is then independent of $n$. For instance, for $s\geq r$ we have
\begin{gather*}
\Lambda^rV\otimes \Lambda^s V^* = V_{ \underbrace{{\scriptstyle{1,\dots,1}}}_{\scriptstyle{r}}}^{\overbrace{\scriptstyle1,\dotsc, 1}^{\scriptstyle{s}}} + V_{ \underbrace{\scriptstyle1,\dots,1}_{\scriptstyle{r-1}}}^{\overbrace{\scriptstyle1,\dotsc, 1}^{\scriptstyle{s-1}}}+\dots + V^{\overbrace{\scriptstyle1,\dotsc, 1}^{\scriptstyle{s-r}}};
\end{gather*}
this also holds for $r+s=n$. Equivalently, we can write
\[\Lambda^{r,s}=\Lambda^{r,s}_0 + \{F\wedge\sigma,\sigma\in \Lambda^{r-1,s-1}\} \cong \Lambda^{r,s}_0 +\dots + \Lambda^{1,s-r+1}_0+ \Lambda^{0,s-r}.\]
In terms of an appropriate map $\Lambda\colon \Lambda^{r,s}\to\Lambda^{r-1,s-1}$, we have
\[\sigma=[\sigma]_0 + F\wedge \Lambda(\sigma), \quad \sigma\in\Lambda^{r,s}.\]
Explicitly,
\begin{equation}
\label{eqn:LambdaOn21}
\Lambda(\gamma)=-\frac1{n-1} e_i\hook e_{n+i}\hook\gamma, \quad \gamma\in\Lambda^{2,1}+\Lambda^{1,2}.
\end{equation}
As a consequence of the Littlewood-Richardson rule, one obtains:
\begin{proposition}[\cite{GadeaMasque}]
The intrinsic torsion of a $\GL(n,\R)$-structures lies in
\[T^*\otimes \lie{gl}(n,\R)^\perp=\coker \partial_{\GL(n,\R)}\cong W_1+\dots + W_8\]
where
\begin{gather*}
W_1=\Lambda^3V^*,\quad W_2=V^{2,1},\quad W_3=V_{1,1}^1 ,\quad W_4=V,\\
W_5=\Lambda^3V,\quad W_6=V_{2,1},\quad W_7=V^{1,1}_1 ,\quad W_8=V^*.
\end{gather*}
\end{proposition}
To compare with Proposition~\ref{prop:paracomplexit}, observe that the $\GL(V)\times\GL(H)$-intrinsic torsion corresponds to $W_1+W_2+W_5+W_6$; however, this space only splits in two irreducible components under the enlarged structure group.
\section{$\SL(n,\R)$-structures}
In this section we turn to the structure group $\SL(n,\R)$. The invariant elements of $\Lambda^*T$ under the action $\SL(n,\R)$ are generated by
\[F=e^{1,n+1}+\dots + e^{n,2n}, \quad \alpha=e^{1,\dots, n} , \quad \beta=e^{n+1,\dots, 2n};\]
conversely, $\SL(n,\R)$ is the largest group that fixes this subalgebra.
Given a $\GL(n,\R)$-structure, a reduction to $\SL(n,\R)$ is determined by a global, nowhere-vanishing form of type $(n,0)$. Clearly, a $\GL(n,\R)$-structure admits a reduction to $\SL(n,\R)$ if and only if it admits a reduction to $\GL_+(n,\R)$, i.e. the two rank $n$ distributions are orientable.
\begin{example}
Consider the standard parahermitian structure on $\R^4=\R^2\times \R^2$. This structure is preserved by the group $\Gamma$ generated by the diffeomorphism
\[(x,y,z,t)\mapsto (x+1,-y,z,-t).\]
Hence the quotient $\R^4/\Gamma$ has an induced parahermitian structure. In this case, the rank two subbundles are not orientable, and there is no global reduction to $\SL(n,\R)$.
\end{example}
\begin{remark}
Given an almost parahermitian structure on $M$, namely a $\GL(n,\R)$-structure $P$, it is always possible to find a $2:1$ cover of $M$ which admits a reduction to $\SL(n,\R)$. Indeed, the quotient $P/\GL_+(n,\R)$ is a $2:1$ cover of $M$ which admits a tautological $\GL_+(n,\R)$-structure.
\end{remark}
\smallskip
The intrinsic torsion of a $\SL(n,\R)$-structure takes values in the cokernel of the alternating map
\[ \partial_{\SL(n,\R)}\colon T^*\otimes \lie{sl}(n,\R)\to \Lambda^2T^*\otimes T;\]
since $\partial_{O(n,n)}$ is an isomorphism, we can identify this space with $T^*\otimes \Sl(n,\R)^\perp$.
\begin{proposition}
The intrinsic torsion of an $\SL(n,\R)$-structure lies in
\[T^*\otimes \Sl(n,\R)^\perp = T^*\otimes \gl(n,\R)^\perp \oplus W^{1,0}\oplus W^{0,1},\]
where \[W^{1,0}=\partial (V^*\otimes \R)\cong V^*, \quad W^{0,1}=\partial (V\otimes \R)\cong V.\]
\end{proposition}
The Levi-Civita connection can always be written in the form
\begin{equation}
\label{eqn:omegataulambda}
\omega^{LC}=\omega+\tau+\tilde\lambda,
\end{equation}
with $\tau \in T^*\otimes \lie{gl}(n,\R)^\perp$, $\omega$ is an almost-parahermitian connection and
\[\tilde\lambda=\lambda\otimes \begin{pmatrix} I & 0 \\ 0 & -I\end{pmatrix},\]
where $\lambda$ is a one-form. For future reference, we note that given a form $\sigma$,
\begin{equation}
\label{eqn:bracketlambda}
\partial(\tilde\lambda)\hook \sigma=\sum_{p,q} (p-q)\lambda\wedge \sigma^{p,q}.
\end{equation}
Thus, $\omega$ is the connection obtained from the Levi-Civita connection by projection on $\Sl(n,\R)$; we shall refer to it as the \emph{minimal} connection, and denote by $\nabla, D$ the corresponding covariant derivative and exterior covariant derivative. By construction, the torsion of $\omega$ is $\Theta=-\partial(\tau+\tilde\lambda)$. The component $\tau$ can be decomposed as the sum of
\[\tau=\tau_1+\dots + \tau_8,\]
with each $\tau_i$ corresponding to a section of the bundle associated to $W_i$. Relative to the action of $\R^*\subset\GL(n,\R)$, we can decompose $\tau$ into four components with weights $-3,1,3,-1$, namely
\begin{equation}
\label{eqn:tauindices}
\begin{gathered}
\tau_1+\tau_2= a_{ijk}e^i\otimes e^{jk}\in V^*\otimes \Lambda^2V^*,\quad
\tau_3+\tau_4= c_{ijk}e^i\otimes e_{jk}\in V^*\otimes \Lambda^2V,\\
\tau_5+\tau_6= b_{ijk}e^{n+i}\otimes e_{jk}\in V\otimes \Lambda^2V,\quad
\tau_7+\tau_8= d_{ijk}e^{n+i}\otimes e^{jk}\in V\otimes \Lambda^2V^*.
\end{gathered}
\end{equation}
Here, summation over all $i,j,k$ is implied, and we assume that $a_{ijk}=-a_{ikj}$.
The components $W_4$, $W_8$, $W^{1,0}$ and $W^{0,1}$ can be encoded in three one-forms
\begin{equation}
\label{eqn:pag7alto}
f_4 = a_ie^{n+i}, \quad f_8=b_ie^{i}, \quad \lambda=\lambda_I e^I
\end{equation}
characterized by
\begin{equation}
\label{eqn:tau4tau8lambda}
\tau_4 = a_i(e^k\otimes e_{ki}),\quad \tau_8=b_i(e^{n+k}\otimes e^{ki}), \quad \tilde\lambda = \lambda_I e^I\otimes (e^k\otimes e_k - e^{n+k}\otimes e_{n+k}).
\end{equation}
A useful symmetry arises as follows. Given an $\SL(n,\R)$-structure $P$, one can consider the $\SL(n,\R)$-structure $P\sigma$, where $\sigma=\left(\begin{smallmatrix} 0 & I \\ I & 0 \end{smallmatrix}\right);$
this amounts to interchanging $V$ and $H$. An adapted coframe $e^1,\dots, e^{2n}$ for $P$ determines an adapted coframe
\[e_\sigma^1,\dots, e_\sigma^{2n} = e^{n+1},\dots, e^{2n},e^1,\dots, e^n,\]
relative to which the intrinsic torsion has the form
\begin{gather*}
(a_\sigma)_{ijk} = b_{ijk},\quad(b_\sigma)_{ijk} = a_{ijk},\quad (c_\sigma)_{ijk} = d_{ijk},\quad (d_\sigma)_{ijk} = c_{ijk},\\
(\lambda_\sigma)_i = -\lambda_{n+i}, \quad (\lambda_\sigma)_{n+i}=-\lambda_i;
\end{gather*}
the minus sign originates from the action of $\sigma$ on $\left(\begin{smallmatrix} I & 0 \\ 0 & - I\end{smallmatrix}\right)$.
There are constraints on $\tau$ coming from the first Bianchi identity. Indeed, recall that given a tensorial $k$-form $\eta$, one has
\begin{equation}
\label{eqn:exteriorcovariant}
D\eta = \mathfrak{a}(\nabla\eta)+\Theta\hook\eta,
\end{equation}
where $\Theta$ is the torsion and
$\mathfrak{a}(\nabla\eta)=\langle \nabla\eta, \frac1{k!}\theta\wedge \dots \wedge \theta \rangle .$
In particular,
\[D\Theta = D(-\partial(\tau+\tilde\lambda))=\mathfrak{a}(-\nabla\partial(\tau+\tilde\lambda))+\partial(\tau+\tilde\lambda)\hook \partial(\tau+\tilde\lambda)\]
must satisfy
\[D\Theta=\Omega\wedge\theta\in \Lambda^2T^*\otimes \Sl(n,\R) \subset\Lambda^3T^*\otimes T.\]
Define the equivariant maps
\begin{align*}
p&\colon\Lambda^3T^*\otimes T\to \Lambda^{1,1}_0, & \eta\otimes v&\mapsto [v\hook \eta +(n-1) \Lambda\eta\wedge v\hook F]_{\Lambda^{1,1}_0};\\
q&\colon\Lambda^3T^*\otimes T\to \R, & \eta\otimes v&\mapsto \Lambda( v\hook \eta).
\end{align*}
\begin{proposition}
The intrinsic torsion of an $\SL(n,\R)$-structure satisfies
\label{prop:constraint}
\begin{multline*}
p\bigl(\mathfrak{a}(-\nabla\partial(\tau_3+\tau_4+\tau_7+\tau_8))+ \partial(\tau_1)\hook \partial(\tau_6)
+ \partial(\tau_2)\hook \partial(\tau_5)
+ \partial(\tau_6)\hook \partial(\tau_1)\\
+ \partial(\tau_5)\hook \partial(\tau_2)
+\partial(\tau_3+\tau_4+\tau_7+\tau_8)\hook \partial(\tilde\lambda )\bigr)
=0
\end{multline*}
\begin{equation*}
q(\mathfrak{a}(-\nabla\partial(\tau_4+\tau_8))+\partial(\tau_4+\tau_8)\hook \partial(\tilde\lambda )\bigr)=0.
\end{equation*}
\end{proposition}
\begin{proof}
As a first step, we prove that $p$ and $q$ kill $\Omega\wedge\theta$. Decomposing $\Lambda^3T^*$ under $\GL(n,\R)$, one readily sees that $p$ and $q$ are only non-trivial on $\Lambda^{2,1}\otimes V+\Lambda^{1,2}\otimes V^*$. Writing $\Omega=\Omega^{2,0}+\Omega^{1,1}+\Omega^{0,2}$, it follows that
\[p(\Omega\wedge\theta)=p(\Omega^{1,1}\wedge\theta).\]
By linearity, we can assume that $\Omega^{1,1}$ has the form $e^{i,n+j}\otimes (e^k\otimes e_{h}-e^{n+h}\otimes e_{n+k})$; using \eqref{eqn:LambdaOn21}, we obtain
\begin{multline*}
p(\Omega^{1,1}\wedge\theta)=p( e^{i,n+j,k}\otimes e_h - e^{i,n+j,n+h}\otimes e_{n+k})\\
=[\delta_{ih} e^{n+j,k} +e^{i,n+h}\delta_{kj}+e_j\hook e^{i,k,n+h} +e^{n+i}\hook e^{n+j,n+h,k}]_{\Lambda^{1,1}_0}=0.
\end{multline*}
The Bianchi identity now implies
\begin{equation*}
0=p(\Omega\wedge\theta)=p(\mathfrak{a}(-\nabla\partial(\tau+\tilde\lambda))+\partial(\tau+\tilde\lambda)\hook \partial(\tau+\tilde\lambda)),
\end{equation*}
and the same holds for $q$.
From $\lambda\wedge\lambda=0$, we obtain $\partial(\tilde\lambda)\hook\partial(\tilde\lambda)=0$; the component
$\partial(\tilde\lambda)\hook\partial(\tau)$
gives no contribution because of Schur's lemma and \eqref{eqn:bracketlambda}. Similarly, $\nabla(\partial(\tau_1+\tau_2+\tau_5+\tau_6))$ is in the kernel of $p$ and $q$ because it has no component in $\Lambda^{2,1}\otimes V+\Lambda^{1,2}\otimes V^*$.
Observe that
\[\mathfrak{a}(\nabla\partial\tilde\lambda) = d\lambda\wedge e^k\otimes e_k - d\lambda\wedge e^{n+k}\otimes e_{n+k},\]
hence, using \eqref{eqn:LambdaOn21} again,
\begin{multline*}
p(\mathfrak{a}(\nabla\partial\tilde\lambda)= \bigl[e_k\hook (d\lambda\wedge e^k) - e_{n+k}\hook (d\lambda\wedge e^{n+k}) \\
+(n-1)\Lambda(d\lambda \wedge e^k)\wedge e^{n+k}+(n-1) \Lambda(d\lambda\wedge e^{n+k})\wedge e^k)\bigr]_{\Lambda^{1,1}_0}\\
=\bigl[e_k\hook (d\lambda\wedge e^k) - e_{n+k}\hook (d\lambda\wedge e^{n+k})
-e_k\hook (d\lambda\wedge e^k)+e_{n+k}\hook (d\lambda\wedge e^{n+k})\bigr]_{\Lambda^{1,1}_0}=0;
\end{multline*}
a similar calculation shows that $q(\mathfrak{a}(\nabla\partial\tilde\lambda)=0$.
Writing
\begin{multline*}
\partial(\tau_3+\tau_4)\hook \partial(\tau_7+\tau_8)= 4c_{ijk}e^{i,n+j}\otimes e_{k} \hook (d_{hlm}e^{n+h,l}\otimes e_{n+m})\\
=-4c_{ijk}d_{hkm}e^{i,n+j, n+h}\otimes e_{n+m},
\end{multline*}
and symmetrically
$\partial(\tau_7+\tau_8)\hook \partial(\tau_3+\tau_4)= -4d_{ijk} c_{hkm}e^{n+i,j,h}\otimes e_m,$
we find that
\begin{multline*}
p(\partial(\tau_7+\tau_8)\hook \partial(\tau_3+\tau_4) + \partial(\tau_7+\tau_8)\hook \partial(\tau_3+\tau_4))\\
= \bigl[4c_{imk}d_{hkm}e^{i, n+h} -4c_{ijk}d_{hkh}e^{i,n+j} + 4c_{iik}d_{hkm}e^{ n+h,m}- 4c_{ijk}d_{ikm}e^{n+j,m}\\
+ 4d_{imk}c_{hkm}e^{n+i, h} -4d_{ijk}c_{hkh}e^{n+i,j} + 4d_{iik}c_{hkm}e^{h,n+m}- 4d_{ijk}c_{ikm}e^{j,n+m}\bigr]_{\Lambda^{1,1}_0}
\end{multline*}
is zero, and the same for $q$.
Explicit computations shows that $p$ and $q$ annihilate $\partial(\tau_2)\hook \partial(\tau_6)\allowbreak+\allowbreak\partial(\tau_6)\hook \partial(\tau_2)$ and
$\partial(\tau_1)\hook \partial(\tau_5)+\partial(\tau_5)\hook \partial(\tau_1)$.
Finally, observe that $\nabla\partial\tau_i$ lies in a module isomorphic to $T^*\otimes W^i$ which only contains a component isomorphic to $V^1_1$ for $i=3,4,7,8$, and to $\R$ for $i=4,8$. Similarly, $\partial(\tau_i)\hook\partial(\tau_j)$ lies in a module isomorphic to $W^i\otimes W^j$; the equivariance of $p$ and $q$ gives the statement.
\end{proof}
\begin{remark}
The component $\tau$ of the intrinsic torsion depends only on the $\GL(n,\R)$-struc\-ture; the component $\lambda$ only depends on the $\SL(n,\R)\times \SL(n,\R)$-structure. The components $\tau_1,\tau_2,\tau_5,\tau_6$ depend on the paracomplex structure (see Proposition~\ref{prop:paracomplexit}); more precisely, they are determined by
\[(de^i)^{0,2}, \quad (de^{n+i})^{2,0}.\]
\end{remark}
\section{Ricci curvature}
In this section we give a formula for the Ricci tensor of an $\SL(n,\R)$-structure, expressed in terms of its intrinsic torsion. Even though the structure group is $\SL(n,\R)$, all relevant representations have a natural action of $\GL(n,\R)$, and the maps we consider in this section are $\GL(n,\R)$-invariant; accordingly, we will regard two representations as isomorphic if they are under this larger group.
The basic idea is that, relative to the decomposition
\[\Lambda^2T^*\otimes \so(n,n)= \Lambda^2T^*\otimes (\Sl(n,\R)\oplus \gl(n,\R)^\perp\oplus \R),\]
the relevant part of the curvature of the Levi-Civita connection is determined by the last two components, which only depend on the intrinsic torsion. Indeed, writing the Levi-Civita connection form as in \eqref{eqn:omegataulambda}, its curvature decomposes as
\[\Omega^{LC}=(\Omega+\frac12[[\tau,\tau]]_{\Sl(n,\R)}) + (D\tau + [\tilde\lambda,\tau] )+ (d\lambda-\frac1n F(\tau,\tau)) \otimes \begin{pmatrix} I & 0 \\ 0 & -I\end{pmatrix},\]
where we have used Lemma \ref{lemma:glglscalar}. Here $D$ denotes the exterior covariant derivative of the minimal connection, and $F(\tau,\tau)$ denotes a $2$-form obtained by contracting $\tau$ with itself using $F$. More precisely, let $\langle\cdot,\cdot\rangle$ denote the natural pairing between $V$ and $V^*$, and consider the skew bilinear form on $\Lambda^2V+\Lambda^2V^*$ such that
\[\tilde F(\gamma,\sigma)=\langle \gamma,\sigma
\rangle=-\tilde F (\sigma,\gamma), \gamma\in\Lambda^2V, \sigma\in\Lambda^2V^*,\]
and zero otherwise, and define \[F(\eta\otimes \gamma, \eta'\otimes \gamma')=\tilde F(\gamma,\gamma') \eta\wedge \eta'.\]
Thus, for fixed $i,h,j\neq k$,
\[F(e^i\otimes e_{jk}, e^{n+h}\otimes e_{n+j,n+k}) = e^i\wedge e^{n+h}.\]
We shall also consider the $2$-form $\overline{F}(\tau,\tau)$ defined by
\[\overline{F}(e^I\otimes e^J\otimes e^K,e^H\otimes e^L \otimes e^M)=-F(e^I,e^H) F(e^J,e^L) e^K\wedge e^M.\]
We shall decompose the Ricci tensor of the Levi-Civita connection as
\[\ric=\ric'+\ric'',\quad \ric'\in V^*\otimes V, \quad \ric''\in S^2V\oplus S^2V^*.\]
Here, $V^*\otimes V$ represents a subspace of $S^2T^*$, i.e. $e^{i}\otimes e^{n+j}$ stands for $e^{i}\odot e^{n+j}$. We will also identify this space with $\Lambda^{1,1}$ through
\begin{equation}
\label{eqn:identify11}
V^*\otimes V\cong \Lambda^{1,1}, \quad e^i\otimes e^{n+j}\mapsto e^{i,n+j}.
\end{equation}
As a first approximation, the $V^*\otimes V$ part of the Ricci can be described as follows.
\begin{lemma}
The Ricci tensor of an $\SL(n,\R)$-structure satisfies
\label{lemma:ricprime}
\[\ric' = 2\ric' (D\tau+[\tilde\lambda, \tau])+n(d\lambda)^{1,1}-F( \tau,\tau)^{1,1}.\]
\end{lemma}
\begin{proof}
The Riemann tensor takes values in the kernel $\mathcal{R}$ of the skewing map
\begin{equation*}
S^2(\Lambda^2T^*)\to\Lambda^4T^*.
\end{equation*}
As a $\GL(n,\R)$-module, $S^2(\Lambda^2T^*)$ decomposes as
\[ S^2(\Lambda^2T^*)=3\R +3V_1^1+2V_{1,1}^{1,1} + V_2^2+U+U^*,\]
where $U=2\Lambda^2V+\Lambda^4V+V_{2,2}+V_{1,1,1}^1 + V_{2,1}^1 + S^2V$. Since the Ricci contraction is equivariant, and $\ric'$ takes values in $V^*\otimes V=V_1^1\oplus\R$, we only need to consider the components of $\mathcal{R}$ isomorphic to $V_1^1$ and $\R$.
The three components of $S^2(\Lambda^2T^*)$ isomorphic to $V^1_1$ contain the highest weight vectors
\[
v_1=
e^{n,k+n}\odot e^{k,n+1},\quad v_2=e^{n,n+1}\odot e^{k,n+k}, \quad v_3=e^{n+1,k+n}\odot e^{kn},
\]
where $nv_1-2v_2$ lies in $S^2(\Sl(n,\R))$, $v_2$ in $\Sl(n,\R)^\perp\otimes \Sl(n,\R)$ and $v_3$ in $ S^2(\Sl(n,\R)^\perp)$.
However $\mathcal{R}$ only contains two copies of $V_1^1$, generated by $v_1+v_2, v_1+v_3.$
Similarly, the components isomorphic to $\R$ in $S^2(\Lambda^2T^*)$ are generated by
\begin{gather*}
w_1= e^{n+i,n+j}\otimes e^{ij}+ e^{ij}\otimes e^{n+i,n+j}\in S^2(\gl(n,\R)^\perp),\\
w_2= e^{i,n+i}\otimes e^{j,n+j}\in S^2(\R)
\end{gather*}
and
\[w_3-\frac1nw_2\in S^2(\Sl(n,\R)), \quad w_3=\sum_{i,j} e^{j,n+i}\otimes e^{i,n+j}.\]
The vectors $w_1+2w_2$, $w_2+w_3$ generate the two copies of $\R$ in $\mathcal{R}$.
By equivariance, and neglecting components not isomorphic to $V_1^1$ and $\R$, which do not contribute to $\ric'$, we may assume that the Riemann tensor has the form
\[R=a(v_1+v_2)+b(v_1+v_3)+h(w_1+2w_2) + k(w_2+w_3).\]
The Ricci contraction of the fixed generators is given by
\begin{gather*}
\ric(w_1)=-2(n-1)g, \quad \ric(w_2)=g, \quad \ric(w_3)=ng,\\
\ric(v_1)=ne^n\odot e^{n+1}, \quad \ric(v_2)=2e^n\odot e^{n+1},\quad \ric(v_3)=(n-2)e^n\odot e^{n+1};
\end{gather*}
with our choice of $R$, the Ricci tensor is
\begin{equation}
\label{eqn:ricprime}
\ric' = (a(n+2)+b(2n-2))e^n\odot e^{n+1}+(2(2-n)h+(1+n)k)g.
\end{equation}
Consider the projections
\begin{gather*}
\pi_{\gl^\perp}\colon \mathcal{R}\to \Lambda^2T^*\otimes \gl(n,\R)^\perp,\qquad
\pi_{\Sl}\colon \mathcal{R}\to \Lambda^2T^*\otimes \Sl(n,\R),\\
\pi_{\R}\colon \mathcal{R}\to \Lambda^2T^*\otimes \R.
\end{gather*}
Since both the image of $\pi_{\gl^\perp}$ and the image of $\pi_\R$ contain $V_1^1\oplus\R$, it is possible to recover $\ric'$ by only considering these projections. Explicitly, we have
\begin{equation}
\label{eqn:projcurv}
D\tau + [\tilde\lambda,\tau]=\pi_{\gl^\perp}(R), \quad (d\lambda-\frac1n F(\tau,\tau)) \otimes \begin{pmatrix} I & 0 \\ 0 & -I\end{pmatrix}= \pi_{\R}(R).
\end{equation}
Then
\begin{gather*}
\pi_{\gl^\perp}(R)=bv_3+hw_1, \\
\pi_\R(R)=(2h+k(1+\frac1n))w_2+(a+\frac2n(a+b))e^{n,n+1}\otimes (e^k\otimes e_k - e^{n+k}\otimes e_{n+k}).
\end{gather*}
Thus \eqref{eqn:projcurv} gives
\begin{gather*}
\ric'(D\tau + [\tilde \lambda,\tau]) = b(n-2)e^n\odot e^{n+1} -2h(n-1)g,\\
\ric'((d\lambda-\frac1n F(\tau,\tau))\! \otimes\! \begin{pmatrix} I & 0 \\ 0 & -I\end{pmatrix})=(a\!+\!\frac2n(a\!+\!b))e^n\!\odot\! e^{n+1}+(2h\!+\!k(1\!+\!\frac1n))g.
\end{gather*}
It is now straightforward to verify that the linear combination
\[\ric'=2\ric'(D\tau + [\tilde \lambda,\tau]) +n\ric'(D\tilde\lambda+\frac12[[\tau,\tau]]_\R)\]
is consistent with \eqref{eqn:ricprime}; the statement follows observing that for any two-form $\eta$, through the identification \eqref{eqn:identify11},
\[\ric'\biggl(\eta\otimes\begin{pmatrix} I & 0 \\ 0 & -I\end{pmatrix}\biggr)=\eta^{1,1}.\qedhere\]
\end{proof}
It turns out that the explicit dependence on the minimal connection (i.e. the term $D\tau$) can be partly eliminated from the formula:
\begin{theorem}
\label{thm:ricci}
The Ricci tensor of an $\SL(n,\R)$-structure satisfies
\begin{equation*}
\begin{split}
\ric'&=2\ric'( \mathfrak{a}(\nabla\tau_3)+\lambda^{1,0}\wedge \tau_3) + 2(n-2)df_4^{1,1}+n(d\lambda)^{1,1}-2(n-1) f_4\wedge f_8\\
+&( 2n\Lambda(df_4) -4(n-1)\langle f_4,f_8\rangle ) F-10F(\tau_1,\tau_5)+2F(\tau_1,\tau_6)-4F(\tau_2,\tau_5)\\
-&2F(\tau_2,\tau_6)+2\overline F(\tau_2,\tau_6)-2F(\tau_3,\tau_7+\tau_8)-2(n-1)F(\tau_4,\tau_7).
\end{split}
\end{equation*}
\end{theorem}
\begin{proof}
As $[\tilde \lambda,\tau_i]$ depends equivariantly on an element of $T^*\otimes W_i$, by equivariance, the only contribution of $[\tilde\lambda,\tau ]$ to $\ric'$ comes from $[\tilde\lambda,\tau_3+\tau_4+\tau_7+\tau_8]$. Since $\tau_3+\tau_4\in V^*\otimes \Lambda^2V$ and $\tau_7+\tau_8\in V\otimes \Lambda^2V^*$,
\begin{equation}
\label{eqn:bracketlambdatau}
\ric'([\tilde\lambda,\tau])=\ric'(2\lambda\wedge (\tau_3+\tau_4)-2\lambda\wedge(\tau_7+\tau_8)).
\end{equation}
By \eqref{eqn:exteriorcovariant},
\[D(\tau_1+\tau_2) =-\partial(\tau+\tilde\lambda)\hook (\tau_1+\tau_2)\mod T^*\otimes (W_1+W_2);\]
and the same holds for $\tau_5+\tau_6$; by equivariance,
\[\ric'\!(D\tau) \!= \! \ric'\!(-\partial(\tau_5+\tau_6)\hook\!(\tau_1+\tau_2)
-\partial(\tau_1+\tau_2)\hook (\tau_5+\tau_6) + D\tau_3+D\tau_4+D\tau_7+D\tau_8).\]
Since the Ricci tensor is symmetric, we can identify $\ric'$ with its projection on $V^*\otimes V$; using Lemma~\ref{lemma:ricprime},
\[\ric'\!=2\ric'\!( D\tau_3+D\tau_4 -\partial(\tau_1+\tau_2)\hook\! (\tau_5+\tau_6)+2\lambda^{1,0}\wedge (\tau_3+\tau_4))+nd\lambda^{1,1}-F(\tau,\tau)^{1,1}.\]
Recalling that the contraction of $\partial(\tau_7+\tau_8)$ into $\tau_3+\tau_4$ is zero and \eqref{eqn:bracketlambda}, we obtain
\begin{align*}
\ric'(D\tau_3+D\tau_4)&=\ric'(\mathfrak{a}(\nabla\tau_3+\nabla\tau_4)-\partial(\tilde\lambda)\hook (\tau_3+\tau_4))\\
&=\ric'(\mathfrak{a}(\nabla\tau_3+\nabla\tau_4)-\lambda\wedge (\tau_3+ \tau_4) ).
\end{align*}
Therefore,
\begin{multline*}
\ric'=2\ric'( \mathfrak{a}(\nabla\tau_3+\nabla\tau_4) -\partial(\tau_1+\tau_2)\hook (\tau_5+\tau_6)+\lambda^{1,0}\wedge (\tau_3+\tau_4))\\
+n(d\lambda)^{1,1}-F(\tau,\tau)^{1,1}.
\end{multline*}
Writing $\tau_4$ as in \eqref{eqn:tau4tau8lambda},
\begin{align*}
\ric'(2\lambda^{1,0}\wedge\tau_4)&= 2\bigl(\lambda_i a_i e^k\otimes e^{n+k} +(n-2)\lambda^{1,0}\otimes a_ie^{n+i}\bigr)\\
&= 2(n-2)\lambda^{1,0}\otimes f_4 + 2\langle \lambda^{1,0},f_4\rangle F.
\end{align*}
In order to rewrite the term containing the covariant derivative of $\tau_4$, we may assume
$\nabla\tau_4 = a_{ij} e^j\otimes (e^k\otimes e_{ki});$
this implies
\[\ric(\mathfrak{a}(\nabla\tau_4) )= a_{ii}F + (n-2)a_{ij}e^j\otimes e^{n+i}= a_{ii}F +(n-2)\nabla f_4.\]
On the other hand we have
\begin{align*}
(df_4)^{1,1}& = \mathfrak{a}(\nabla f_4)^{1,1}-\partial(\tau_7+\tau_8)\hook f_4 + \lambda^{1,0}\wedge f_4\\
&= a_{ij}e^{j,n+i} +2d_{ijk}a_k e^{j,n+i} + \lambda^{1,0}\wedge f_4\\
&= a_{ij}e^{j,n+i} +F(\tau_7+\tau_8,\tau_4) + \lambda^{1,0}\wedge f_4,
\end{align*}
where
\[F(\tau_8,\tau_4)=\langle f_4,f_8\rangle F + f_4\wedge f_8.\]
In particular
\[\Lambda (df_4)=\frac1n a_{ii}+ (1-\frac1n)\langle f_4,f_8\rangle+\frac1n\langle \lambda^{1,0},f_4\rangle,\]
giving
\begin{multline*}
\ric(\mathfrak{a}(\nabla\tau_4) )= ( n\Lambda(df_4) -(n-1)\langle f_4,f_8\rangle-\langle \lambda^{1,0},f_4\rangle) F \\
+(n-2)(df_4^{1,1}-F(\tau_7+\tau_8,\tau_4) - \lambda^{1,0}\wedge f_4).
\end{multline*}
Decomposing $\tau$ as in \eqref{eqn:tauindices}, we find
\begin{align*}
F(\tau_1+\tau_2,\tau_5+\tau_6)&=- 2a_{ijk}b_{hjk} e^i\wedge e^{n+h}, \\
\overline{F}(\tau_1+\tau_2,\tau_5+\tau_6)&=- 4a_{ijk}b_{ijh} e^k\wedge e^{n+h};
\end{align*}
then
\begin{multline*}
\ric(\partial(\tau_1+\tau_2)\hook(\tau_5+\tau_6))= 4(a_{jik}b_{khj}-a_{ijk}b_{khj})e^i\otimes e^{n+h}\\
=4F(\tau_1,\tau_5)-2F(\tau_1,\tau_6)+ F(\tau_2,\tau_5)-\overline F(\tau_2,\tau_6).
\end{multline*}
By contrast,
\[F(\tau,\tau)=2F(\tau_1+\tau_2,\tau_5+\tau_6)+2F(\tau_3+\tau_4,\tau_7+\tau_8).\]
Summing up,
\begin{multline*}
\ric'=2\ric'( \mathfrak{a}(\nabla\tau_3)+\lambda^{1,0}\wedge \tau_3) + 2(n-2)(df_4-F(\tau_7+\tau_8,\tau_4)-\lambda^{1,0}\wedge f_4)\\
+2( n\Lambda(df_4) -(n-1)\langle f_4,f_8\rangle) F
-8F(\tau_1,\tau_5)+4F(\tau_1,\tau_6)-2F(\tau_2,\tau_5)+2\overline F(\tau_2,\tau_6)\\
+2(n-2)\lambda^{1,0}\wedge \tau_4+n(d\lambda)^{1,1}-2F(\tau_1+\tau_2,\tau_5+\tau_6)-2F(\tau_3+\tau_4,\tau_7+\tau_8),
\end{multline*}
from which we obtain the statement.
\end{proof}
As an $\SL(n,\R)$-module, $V^*\otimes V$ splits as the sum $V_1^1\oplus\R$; the two components of $\ric'$ in this decomposition can be written as $\ric'-sF$ and $s$, where $s$ denotes the scalar curvature
\[s=\frac1n\ric(e_i,e_{n+i}).\]
\begin{corollary}
\label{cor:s}
The scalar curvature of an $\SL(n,\R)$-structure is given by
\begin{multline*}
s=\frac{10}n\langle \tau_1,\tau_5\rangle -\frac2n \langle \tau_2,\tau_6\rangle
-\frac2n \langle \tau_3,\tau_7\rangle\\
+4(n-1)\Lambda df_4 - \frac{2(n-1)(2n-1)}{n}\langle f_4,f_8\rangle+n\Lambda (d\lambda).
\end{multline*}
\end{corollary}
\begin{proof}
By construction, $s=\Lambda \ric'$; in addition, the components which do not contain a copy of $\R$ give no contribution to the scalar curvature by equivariance. This gives
\begin{multline*}
s=(4n-4)\Lambda df_4-10\Lambda F(\tau_1,\tau_5)-2\Lambda F(\tau_2,\tau_6)+2\Lambda\overline F(\tau_2,\tau_6)\\
+n\Lambda(d\lambda)-2\Lambda F(\tau_3,\tau_7)-4(n-1)\langle f_4,f_8\rangle +2(n-1)\Lambda( f_8\wedge f_4).
\end{multline*}
A direct computation gives
\begin{align*}
\Lambda(F(\tau_1+\tau_2,\tau_5+\tau_6))&= -\frac1n \langle \tau_1,\tau_5 \rangle -\frac1n \langle \tau_2,\tau_6 \rangle, \\
\Lambda(\overline{F}(\tau_1+\tau_2,\tau_5+\tau_6))&= -\frac2n \langle \tau_1,\tau_5 \rangle -\frac2n \langle \tau_2,\tau_6 \rangle, \\
\Lambda(F(\tau_3+\tau_4,\tau_7+\tau_8))&= \frac1n \langle \tau_3,\tau_7 \rangle +\frac1n \langle \tau_4,\tau_8 \rangle,
\end{align*}
proving the statement.
\end{proof}
\begin{remark}
On a metric of neutral signature, the notion of ``positive'' scalar curvature is not meaningful: the pseudoriemannian metrics $g$ and $-g$ have the same Ricci tensor, but opposite scalar curvature. In our setup, this means that if we keep the splitting $V\oplus V^*$ but flip the sign of $F$, considering the $\SL(n,\R)$-structure determined by the adapted coframe
\[-e^1,\dotsc, -e^n,e^{n+1},\dotsc, e^{2n},\]
then the $\tau$ and the $\lambda$ stay the same, but $s$ changes its sign.
\end{remark}
\smallskip
The remaining part of the Ricci is given as follows. Denote by $\epsilon$ the symmetrization map
\[\epsilon\colon T^*\otimes T^*\to S^2T^*, \quad \eta\otimes \gamma \mapsto \eta\odot\gamma.\]
\begin{theorem}\label{thm:RicciS2}
The $\ric''$ component of the Ricci tensor of an $\SL(n,\R)$-structure satisfies
\begin{equation*}
\begin{split}
[\ric'']&_{S^2V} = \epsilon \Bigl((n-1)(\nabla f_4-f_4\otimes f_4+\lambda^{0,1}\otimes f_4) \\
&+\ric (\mathfrak{a}\nabla\tau_6-\partial(\tau_7)\hook(\tau_5 +\tau_6)-\partial(\tau_8)\hook \tau_6
-\partial(\tau_3)\hook \tau_3+ 3\lambda^{1,0}\wedge\tau_6)^{0,2}\Bigr),\\
[\ric'']&_{S^2V^*} = \epsilon \Bigl((n-1)(\nabla f_8- f_8\otimes f_8-\lambda^{1,0}\otimes f_8) \\
&+\ric (\mathfrak{a}\nabla\tau_2-\partial(\tau_3)\hook(\tau_1 +\tau_2)-\partial(\tau_4)\hook \tau_2-\partial(\tau_7)\hook \tau_7- 3\lambda^{0,1}\wedge\tau_2)^{2,0}\Bigr).
\end{split}
\end{equation*}
\end{theorem}
\begin{proof}
In the case we consider the action of $\GL(n,\R)$; two representations are isomorphic if they are under $\SL(n,\R)$ and they have the same weight under $\R^*$.
The space $\Lambda^2T^*\otimes \Lambda^2T^*$ contains two copies of $S^2V$, contained in $\Sl(n,\R)\otimes \Sl(n,\R)^\perp$ and $\Sl(n,\R)^\perp\otimes \Sl(n,\R)$ respectively. Since the Riemann tensor is symmetric, this means that the $S^2V$ part of the Ricci is entirely determined by its component in $\Sl(n,\R)\otimes \Sl(n,\R)^\perp$, i.e.
\[[\ric(\Omega)]_{S^2V} = \epsilon (\ric([D\tau + [\tilde\lambda,\tau]]_{S^2V})).\]
The only $W_i\otimes W_j$ that contain a copy of $S^2V$ are
\[W_5\otimes W_7,\ W_6\otimes W_7,\ W_6\otimes W_8,\ W_3\otimes W_3,\ W_4\otimes W_4.\]
Moreover, the two copies of $S^2V$ inside $T^*\otimes W_i$ are contained in $V\otimes W_4$ and $V^*\otimes W_6$. In consequence,
\begin{equation*}
\begin{split}
[D\tau]_{S^2V}& =
[\mathfrak{a}(\nabla\tau)-\partial(\tau)\hook\tau - \partial(\tilde\lambda)\hook\tau]_{S^2V}\\
&=
\bigl[\mathfrak{a}(\nabla\tau_4+\nabla\tau_6)-\partial(\tau_7)\hook(\tau_5 +\tau_6)-\partial(\tau_8)\hook \tau_6 \\
&\qquad -\partial(\tau_3)\hook \tau_3-\partial(\tau_4)\hook \tau_4-\lambda^{0,1}\wedge \tau_4 + \lambda^{1,0}\wedge\tau_6\bigr]_{S^2V}.
\end{split}
\end{equation*}
On the other hand \eqref{eqn:bracketlambdatau} gives
\[
[[\tilde\lambda,\tau]]_{S^2V} = [2\lambda^{0,1}\wedge \tau_4+2\lambda^{1,0}\wedge\tau_6]_{S^2V}.\]
Now
\[\ric(\partial(\tau_4) \hook\tau_4)= (n-1)a_ia_je^{n+i}\otimes e^{n+j} = (n-1)f_4\otimes f_4;\]
writing $a_{ij}e^{n+i}\otimes e^{n+j}$ for the $V\otimes V$ component of $\nabla f_4$, we obtain
\[\ric(\mathfrak{a}\nabla\tau_4)= (n-1)a_{ij}e^{n+i}\otimes e^{n+j}.\]
Moreover
\[\ric(\lambda^{0,1}\wedge\tau_4)= (n-1)\lambda^{0,1}\otimes f_4,\]
giving the first formula in the statement. The second formula is obtained applying the symmetry $\sigma$ that interchanges $V$ and $H$.
\end{proof}
\section{Forms}
In this section we find formulae that express the intrinsic torsion and the Ricci curvature in terms of exterior derivatives, rather than exterior covariant derivatives. In particular we relate the intrinsic torsion to the exterior derivatives of the forms $\alpha\in\Lambda^{n,0},\ \beta\in\Lambda^{0,n},\ F\in\Lambda^{1,1}$.
First, we observe that $v\mapsto v\wedge\alpha$, $v\mapsto v\wedge\beta$ induce isomorphisms
\[\pi_{0,1}\colon \Lambda^{n,1}\to \Lambda^{0,1},\quad \pi_{1,0}\colon \Lambda^{1,n}\to \Lambda^{1,0}.\]
\begin{proposition}
\label{prop:forms}
The intrinsic torsion of an $\SL(n,\R)$-structure determines $dF$, $d\alpha$ and $d\beta$ via
\begin{gather*}
(d F)^{3,0} = -\partial(\tau_1)\hook F, \qquad (dF)^{2,1}= -\partial(\tau_7)\hook F-2f_8\wedge F,\\
(d F)^{0,3} = -\partial(\tau_5)\hook F, \qquad (dF)^{1,2}= -\partial(\tau_3)\hook F -2f_4\wedge F,\\
(d\alpha)^{n,1}= -(n\lambda^{0,1} + (n-1)f_4)\wedge\alpha, \qquad
(d\alpha)^{n-1,2}= -\partial(\tau_5+\tau_6)\hook\alpha ,\\
(d\beta)^{1,n}= (n\lambda^{1,0} - (n-1)f_8)\wedge \beta , \qquad
(d\beta)^{2,n-1}= -\partial(\tau_1+\tau_2)\hook\beta.
\end{gather*}
Conversely,
\begin{gather*}
f_4=-\frac12\left(\Lambda (dF)\right)^{0,1},\qquad f_8=-\frac12\left(\Lambda (dF)\right)^{1,0},\\
\tau_1 = \frac16 e^i\otimes e_i\hook (d F)^{3,0},\qquad\tau_5 = -\frac16 e^{n+i}\otimes e_{n+i}\hook (d F)^{0,3},\\
\tau_2 =\frac12\langle ((d\beta)^{2,n-1}+\partial(\tau_1)\hook\beta),e^{n+j,n+k}\wedge (e_{i}\hook\alpha)\rangle e^{i}\otimes e^{j,k},\\
\tau_6 =\frac12\langle ((d\alpha)^{n-1,2}+\partial(\tau_5)\hook\alpha),e^{jk}\wedge (e_{n+i}\hook\beta)\rangle e^{n+i}\otimes e^{n+j,n+k},\\
\tau_3 = -\frac12 e^i\otimes e_i\hook (dF)^{1,2}-\tau_4,\quad
\tau_7 = \frac12 e^{n+i}\otimes e_{n+i}\hook (dF)^{2,1}-\tau_8,\\
\lambda^{0,1} = -\frac1n \pi_{0,1}(d \alpha)^{n,1}-\frac{n-1}{n}f_4,\qquad\lambda^{1,0} = \frac1n \pi_{1,0}(d \beta)^{1,n}+\frac{n-1}{n}f_8.
\end{gather*}
\end{proposition}
\begin{proof}
The usual formula \eqref{eqn:exteriorcovariant} gives
\[d\alpha = -\partial(\tau+\tilde\lambda)\hook\alpha = -\partial(\tau_3+\tau_4+\tau_5+\tau_6)\hook\alpha - n\lambda^{0,1}\wedge\alpha.\]
However, $-\partial(\tau_3)\hook \alpha$ is zero because it lies in $\Lambda^{n,1}\cong V$ but $W_3$ is not isomorphic to $V$; in addition:
\begin{gather*}\partial(\tau_4)\hook\alpha= -\sum_{i\neq k} a_i e^{k,n+i}\otimes e_k\hook\alpha = \sum_{i\neq k}a_ie^{n+i}\wedge \alpha=(n-1)f_4\wedge \alpha\\
d\alpha = -\partial(\tau_5+\tau_6)\hook\alpha -(n\lambda^{0,1} +(n-1)f_4)\wedge\alpha
\end{gather*}
where we have used \eqref{eqn:pag7alto}.
Similarly, $\partial(\tau_8)\hook\beta= (n-1)f_8\wedge \beta,$
\[d\beta = -\partial(\tau_1+\tau_2)\hook\beta +(n\lambda^{1,0}-(n-1)f_8)\wedge\beta.\]
For $F$ we compute
\[dF= -\partial(\tau+\tilde\lambda)\hook F=-2(f_4+f_8)\wedge F-\partial(\tau_1+\tau_3+\tau_5+\tau_7)\hook F\]
where we have used that $\partial(\tau_2+\tau_6)\hook F$ is zero because $W_2$ and $W_6$ are not isomorphic to any subspace of $\Lambda^3=\Lambda^{3,0}\oplus\Lambda^{2,1}\oplus\Lambda^{1,2}\oplus\Lambda^{0,3}$. By projecting this last formula on the different types of $(p,q)$-forms we get the statement.
We now prove the inverse formulae. From the last equation, we immediately get
\[-\frac12\Lambda (dF)=f_4+f_8\]
and the equations for $\lambda^{1,0}$ and $\lambda^{0,1}$ are obvious.
If we set $\tau_1=e^i\otimes e^{jk}+e^j\otimes e^{ki}+e^k\otimes e^{ij}$; then
\[(d F)^{3,0}=-\partial(\tau_1)\hook F = 6 e^{ijk}\]
and it follows that
\[\frac16 e^h\otimes e_h\hook (d F)^{3,0}=\tau_1.\]
Setting $\tau_3+\tau_4= e^i\otimes e^{n+j,n+k}$ we obtain
\[(d F)^{1,2}=-\partial(\tau_3+\tau_4)\hook F =-2 e^{i,n+j,n+k},\]
thus
\[\tau_3 + \tau_4 = -\frac12 e^i\otimes e_i\hook (dF)^{1,2}.\]
Finally if $\tau_2= a_{ijk}e^i\otimes e^{jk}$, for a fixed tensor $e^{n+h,n+l}\otimes e_m$ we have
\begin{align*}
\langle -\partial(\tau_2)&\hook\beta, e^{n+h,n+l}\otimes e_m\hook\alpha \rangle \\
&= \langle - a_{ijk} e^{ij}\otimes e_{n+k}\hook\beta+a_{ijk} e^{ik}\otimes e_{n+j}\hook\beta, e^{n+h,n+l}\otimes e_m\hook\alpha\rangle \\
&= -2a_{hlm}+2a_{lhm}= 2a_{mhl},
\end{align*}
since $a_{ijk}+a_{jki}+a_{kij}=0$ and $a_{ijk}=-a_{ikj}$. Thus,
\[\tau_2= \frac12\langle -\partial(\tau_2)\hook\beta, e^{n+j,n+k}\wedge e_i\hook\alpha \rangle e^i\otimes e^{jk}.\]
The remaining equations are proved in the same way.
\end{proof}
We can relate each component of the intrinsic torsion as we have done for $\tilde{\lambda}$, related to the 1-form $\lambda$, and $\tau_4$, $\tau_8$ (see \eqref{eqn:pag7alto}). Indeed, the component $\tau_1=a_{ijk}(e^i\otimes e^{jk}+e^j\otimes e^{ki}+e^k\otimes e^{ij})$ can be associated to the $(3,0)$-form $a_{ijk}e^{ijk}$, and analogously for $\tau_5$; we set
\[f_3=-\partial (\tau_3)\hook F,\qquad f_7=-\partial (\tau_7)\hook F.\]
Finally, if $\tau_2=a_{ijk} e^{i}\otimes e_{n+j,n+k}$, we set
\[f_2=a_{ijk} (e_{n+j,n+k}\hook \beta)\wedge e^{i}\in\Lambda^{1,n-2},\]
and if $\tau_6=b_{ijk} e^{n+i}\otimes e_{jk}$ we set
$f_6=b_{ijk} (e_{jk}\hook \alpha)\wedge e^{n+i}\in\Lambda^{n-2,1}$. We use the following convention: for any $p$-form $\sigma$, $p\geq 2$ and any bi-vector $e_{jk}$ the $(p-2)$-form $e_{jk}\hook \sigma$ is defined by
\[(e_{jk}\hook \sigma)(X_1,\dots,X_{p-2})=\sigma (e_j, e_k, X_1,\dots,X_{p-2});\]
equivalently, $e_{jk}\hook \sigma=e_k\hook(e_{j}\hook \sigma)$.
We can then restate the equations of the Ricci curvature by expressing the $\nabla \tau_i$ in terms of the exterior derivative of the forms $f_i$. To this purpose, we identify $V\otimes V$ with $\Lambda^{n-1,1}$ through
\begin{equation}
\label{eqn:VotimesVwithLambdan1}
v\otimes w\mapsto (v\hook\alpha)\wedge w,
\end{equation}
enabling us to identify a subspace of $\Lambda^{n-1,1}$ isomorphic to $S^2V$. Similarly, to obtain a subspace isomorphic to $S^2V^*$ we identify $V^*\otimes V^*$ with $\Lambda^{1,n-1}$ via the isomorphism $v \otimes w \mapsto (v\hook\beta)\wedge w$.
\begin{lemma}\label{lemma:f2f6equations}
The following equations hold for $f_2$ and $f_6$:
\begin{align*}
[\ric(\mathfrak{a}(\nabla \tau_6)-\partial(\tau_7+\tau_8)\hook\tau_6)+3\lambda^{1,0}\wedge \tau_6]_{S^2V}&= [df_6+n\lambda^{1,0}\wedge f_6]_{S^2V},\\
[\ric(\mathfrak{a}(\nabla \tau_2)-\partial(\tau_3+\tau_4)\hook\tau_2)-3\lambda^{0,1}\wedge \tau_2]_{S^2V^*}& = [df_2-n\lambda^{0,1}\wedge f_2]_{S^2V^*}
\end{align*}
and the following identities hold for $f_4$ and $f_8$:
\begin{equation*}
\begin{alignedat}{3} [\nabla f_4- f_4\otimes f_4+\lambda^{0,1}\otimes f_4]_{S^2V}&=[ (-1)^{n-1}&&d(f_4\hook\alpha)]_{S^2V} \\
& &&+ \frac12n\lambda^{0,1}\odot f_4 + (n-2)f_4\otimes f_4,\\
[\nabla f_8- f_8\otimes f_8-\lambda^{1,0}\otimes f_8]_{S^2V^*} &=[ (-1)^{n-1}&&d(f_8\hook\beta)]_{S^2V^*}\\
& &&- \frac12n\lambda^{1,0}\odot f_8 + (n-2)f_8\otimes f_8.
\end{alignedat}
\end{equation*}
\end{lemma}
\begin{proof}
We prove the first equation. As usual, we start from the following:
\[d f_6 = \mathfrak{a}(\nabla f_6) -\partial (\tau + \tilde{\lambda})\hook f_6.\]
If $\nabla \tau_6=a_{ijkh} e^i\otimes e^{n+j}\otimes e_{kh} +b_{ijkh} e^{n+i}\otimes e^{n+j}\otimes e_{kh}$,
then
\[ \nabla f_6=a_{ijkh} e^i\otimes(e_{kh}\hook\alpha)\wedge e^{n+j} + b_{ijkh} e^{n+i} \otimes(e_{kh}\hook\alpha)\wedge e^{n+j}.\]
We are interested in the $(n-1,1)$-component of $d f_6$, and more precisely in the $S^2(V)$ part. We get:
\[[df_6]^{n-1,1}=[\mathfrak{a}(\nabla f_6) -\partial (\tau_7 +\tau_8 + \tilde{\lambda})\hook f_6]^{n-1,1}.\]
Therefore,
\begin{equation*}
\begin{aligned}
[df_6]_{S^2V}&=[a_{ijkh} e^{i}\wedge (e_{kh}\hook\alpha)\wedge e^{n+j} - (n-3)\lambda^{1,0}\wedge f_6-\partial(\tau_7 +\tau_8)\hook f_6]_{S^2V}\\
&=[2a_{ijki} (e_{k}\hook\alpha)\wedge e^{n+j} - (n-3)\lambda^{1,0}\wedge f_6-\partial(\tau_7 +\tau_8)\hook f_6]_{S^2V}\\
&=a_{ijki} e^{n+k}\odot e^{n+j} - [(n-3)\lambda^{1,0}\wedge f_6+\partial(\tau_7 +\tau_8)\hook f_6]_{S^2V},
\end{aligned}
\end{equation*}
where we have used the identification \eqref{eqn:VotimesVwithLambdan1}. On the other hand, we have:
\[\ric (\mathfrak{a}(\nabla\tau_6)) = \ric(a_{ijkh} e^{i,n+j}\otimes (e^{n+k}\otimes e_h-e^{n+h}\otimes e_k))
=2a_{ijki} e^{n+j}\otimes e^{n+k}.
\]
Writing $\tau_6=b_{ijk} e^{n+i}\otimes e_{jk}$ and $\tau_7 +\tau_8= d_{ijk} e^{n+i}\otimes e^{jk}$,
we compute
\[[\ric(-\partial(\tau_7 +\tau_8)\hook\tau_6)]_{S^2V}= 2b_{khj} d_{ijk} e^{n+i}\odot e^{n+h}.\]
Moreover,
\[
-\partial(\tau_7 +\tau_8)\hook f_6 = 4b_{khj} d_{ijk} e_{h}\hook \alpha \wedge e^{n+i},\]
and the $S^2(V)$ component is $2b_{khj} d_{ijk} e^{n+h}\odot e^{n+i}$. Finally, note that
\[[\lambda^{1,0}\wedge f_6]_{S^2V} = \lambda_k b_{ijk}e^{n+j}\odot e^{n+i} = [\ric(\lambda^{1,0}\wedge \tau_6)]_{S^2V},\]
which concludes the proof of the first equation. The second one is proved in a similar way.
Consider the $(n-1,0)$-form $f_4\hook \alpha = a_ie_i\hook\alpha$. If $\nabla\tau_4=a_{ij} e^{n+i}\otimes e^k\otimes e_{kj}\allowbreak+b_{ij} e^i\otimes e^k\otimes e_{kj}$ then
\[\nabla (f_4\hook\alpha)=a_{ij}e^{n+i}\otimes e_j\hook\alpha + b_{ij}e^i\otimes e_j\hook\alpha.\]
As usual, $d (f_4\hook \alpha)=\mathfrak{a}(\nabla(f_4\hook\alpha))-\partial(\tau+\tilde{\lambda})\hook(f_4\hook\alpha)$, but we are interested in the $(n-1,1)$ part, and more precisely in the $S^2(V)$ component. We obtain:
\begin{equation*}
\begin{aligned}
\left[d (f_4\hook \alpha)\right]_{S^2(V)}&=[\mathfrak{a}\!\left(\nabla(f_4\hook\alpha)\right)-\partial(\tau_4)\hook (f_4\hook\alpha)-(n\!-\!1)\lambda^{0,1}\!\wedge\!(f_4\hook\alpha)]_{S^2(V)}\\
= (-1)^{n-1}[a_{ij}e_j\hook\alpha&\!\wedge\! e^{n+i}-(n\!-\!1)a_j e_j\hook\alpha\!\wedge\! a_i e^{n+i}-(n\!-\!1)a_j e_j\hook\alpha\!\wedge\!\lambda^{0,1}]_{S^2(V)}\\
&= (-1)^{n-1}[\nabla f_4 -(n-1) f_4 \otimes f_4-(n-1)f_4\otimes\lambda^{0,1}]_{S^2(V)}.
\end{aligned}
\end{equation*}
This ends the proof of the equation involving $f_4$. A similar argument proves the equation for $f_8$.
\end{proof}
\begin{lemma}\label{lemma:ric'nonabla}
The $V^1_1$-Ricci part of $\tau_3$ and $\tau_7$ can be related to $f_3$ and $f_7$ via the following equations:
\begin{gather*}
[\ric(\mathfrak{a}(\nabla\tau_3)+\lambda^{1,0}\wedge \tau_3)]_{V_1^1} =
\frac{2-n}{2}[\Lambda(df_3+\partial(\tau_7+\tau_8)\hook f_3)]_{V_1^1},\\
[\ric(\mathfrak{a}(\nabla\tau_7)-\lambda^{0,1}\wedge \tau_7)]_{V_1^1} =
\frac{2-n}{2}[\Lambda(df_7+\partial(\tau_3+\tau_4)\hook f_7)]_{V_1^1}.
\end{gather*}
\end{lemma}
\begin{proof}
Let $h_3$ be the equivariant map
\[h_3\colon \Lambda^{1,1}_0\to T^*\otimes W_3, \qquad e^{i,n+j}\mapsto \sum_k e^k\otimes e^i\otimes e^{n+k,n+j}-\frac1n (e^i\otimes e^k\otimes e^{n+k,n+j}),\]
and suppose that $\nabla \tau_3 =h_3(e^{n,n+1})$; then
\[\nabla f_3 = 2e^k\otimes e^{n+k,n,n+1} +\frac 2n e^n\otimes F\wedge e^{n+1}.\]
Using \eqref{eqn:exteriorcovariant} we have
\begin{equation*}
\begin{aligned}
d(f_3) &=\mathfrak{a}(\nabla f_3)-\partial(\tau+\tilde\lambda)\hook f_3\\
&= 2\frac{n+1}n F\wedge e^{n,n+1}-\partial(\tau)\hook f_3+\lambda\wedge f_3,
\end{aligned}
\end{equation*}
and an easy computation shows:
\[\Lambda(df_3+\partial(\tau_7+\tau_8)\hook f_3-\lambda^{1,0}\wedge f_3)^{2,2} = 2\frac{n+1}n e^{n,n+1}.\]
On the other hand we have:
\[\ric(\mathfrak{a}(\nabla\tau_3))= \frac{(2-n)(n+1)}n e^n\otimes e^{n+1};\]
thus,
\[[\ric(\mathfrak{a}(\nabla\tau_3))]_{V^1_1} = \frac{2-n}{2}\Lambda(df_3+\partial(\tau_7+\tau_8)\hook f_3-\lambda^{1,0}\wedge f_3)^{2,2}. \]
By the same token, if $\lambda^{1,0}\otimes \tau_3 = h_3(e^{n,n+1})$, we see that
\[\lambda^{1,0}\wedge f_3=2\frac{n+1}n F\wedge e^{n,n+1},\]
so
\[[\ric (\lambda^{1,0}\wedge \tau_3)]_{V_1^1} = \frac{(2-n)(n+1)}n \frac{n}{2(n+1)} \Lambda(\lambda^{1,0}\wedge f_3).\]
Summing up, we get
\[[\ric(\mathfrak{a}(\nabla\tau_3)+\lambda^{1,0}\wedge \tau_3)]_{V_1^1} =
\frac{2-n}{2}\Lambda(df_3+\partial(\tau_7+\tau_8)\hook f_3)^{2,2},\]
which proves the first of the two equations. By considering the equivariant map
\[h_7\colon \Lambda^{1,1}_0\to T^*\otimes W_7, \quad e^{i,n+j}\mapsto \sum_k e^{n+k}\otimes e^{n+i}\otimes e^{k,j}-\frac1n (e^{n+i}\otimes e^{n+k}\otimes e^{k,j})\]
and using a similar procedure, the other equation follows.
\end{proof}
Using Lemmas \ref{lemma:f2f6equations} and \ref{lemma:ric'nonabla} we are able to restate Theorems \ref{thm:ricci} and \ref{thm:RicciS2}, expressing the Ricci curvature of the Levi-Civita connection without using the covariant derivative $\nabla\tau$. Note that the projection of $\ric ( -\partial(\tau_7)\hook\tau_5 -\partial(\tau_3)\hook \tau_3)$ on tensors of type $(2,0)$ is now redundant and we can drop it; the same happens for the $S^2V^*$ component of the Ricci tensor.
\begin{theorem}\label{thm:restateRicViaForms}
The Ricci tensor of an $\SL(n,\R)$-structure satisfies:
\begin{equation*}
\begin{split}
\ric'&=(2-n)[\Lambda(df_3+\partial(\tau_7+\tau_8)\hook f_3)]_{V^1_1}+ 2(n-2)df_4^{1,1} -2(n-1) f_4\wedge f_8\\
&\!\!\!+( 2n\Lambda(df_4) -4(n-1)\langle f_4,f_8\rangle ) F-10F(\tau_1,\tau_5)+2F(\tau_1,\tau_6)-4F(\tau_2,\tau_5)\\
&\!\!\!-2F(\tau_2,\tau_6)+2\overline F(\tau_2,\tau_6)-2F(\tau_3,\tau_7+\tau_8)-2(n-1)F(\tau_4,\tau_7)+n(d\lambda)^{1,1},\\
[\ric'']&_{S^2V} = \epsilon \Bigl((n-1)([(-1)^{n-1}d(f_4\hook\alpha)]_{S^2V} + \frac12n\lambda^{0,1}\odot f_4 + (n-2)f_4\otimes f_4 )\\
&+[df_6+n\lambda^{1,0}\wedge f_6]_{S^2V}+\ric ( -\partial(\tau_7)\hook\tau_5 -\partial(\tau_3)\hook \tau_3)\Bigr),\\
[\ric'']&_{S^2V^*}\!=\! \epsilon \Bigl((n-1)([ (-1)^{n-1}d(f_8\hook\beta)]_{S^2V^*} - \frac12n\lambda^{1,0}\odot f_8 + (n-2)f_8\otimes f_8) \\
&+[df_2-n\lambda^{0,1}\wedge f_2]_{S^2V^*}+\ric (-\partial(\tau_3)\hook\tau_1 -\partial(\tau_7)\hook \tau_7)\Bigr).
\end{split}
\end{equation*}
\end{theorem}
\begin{remark}
In the notation of \cite{HarveyLawson}, a $D$-valued $(n,0)$-form has the form
\[\Phi=(\overline e p + e q)(\overline e \alpha+ e\beta) = \overline{e}(p\alpha) + e(q\beta) .\]
Assume that the structure is paracomplex; then this is holomorphic if and only if it is closed; this is equivalent to
\[d(\log p)^{0,1} = n\lambda^{0,1}+(n-1)f_4, \quad d(\log q)^{1,0} = -n\lambda^{1,0}+(n-1)f_8.\]
By \cite{HarveyLawson}, in the parak\"ahler case the Ricci form is given by
\[\ric=-\tau \partial\overline{\partial} \log \abs{\Phi}^2 = -\partial_+\partial_-\log \abs{\Phi}^2 .\]
In our case, $\abs{\Phi}^2=pq$, hence
\begin{align*}
-\partial_+\partial_- pq &= -\partial_+(n\lambda^{0,1}-(n-1)f_4))+ \partial_-(-n\lambda^{1,0}+(n-1)f_8) \\
&= (-nd\lambda + (n-1)(-df_4+df_8))^{1,1}.
\end{align*}
This is one component in our expression for the Ricci which corresponds to the curvature of the canonical bundle. In the parak\"ahler case, i.e. $\tau=0$, the expressions coincide.
\end{remark}
As an application, we recover a result of \cite{IvanovZamkovoy:parahermitian}, asserting that six-dimensional nearly parak\"ahler structures are Einstein. Recall that an almost parak\"ahler manifold is called nearly parak\"ahler if $\nabla K$ is skew-symmetric in the first two indices; in our language, this means that the intrinsic torsion lies in $W_1+W_5$ (see e.g. \cite{GadeaMasque}). From the paracomplex point of view, this means that under some identification $V\cong H^*$, the two components of the Nijenhuis tensor lie in
\[\Lambda^3V^*+\Lambda^3H^*;\]
i.e., the Nijenhuis tensor is totally skew-symmetric; this is the condition that ensures the existence of a connection with skew-symmetric torsion \cite{IvanovZamkovoy:parahermitian}.
With this definition, a parak\"ahler manifold is nearly parak\"ahler; we shall say a structure is \emph{strictly nearly parak\"ahler} if $\tau_1\neq0$ at each point. It was shown in \cite{SchaferSchulteHengesbach} that six-dimensional strictly nearly parak\"ahler structures can be characterized in terms of differential forms and exterior derivatives. In our language, we obtain:
\begin{corollary}
\label{cor:nk}
On a connected six dimensional manifold $M$, an almost parahermitian structure with $\tau_1\neq0$ is strictly nearly parak\"ahler if and only if it has a reduction to $\SL(3,\R)$ such that, up to rescaling by a constant,
either
\begin{equation*}
dF=\alpha + \beta,\quad d(\alpha-\beta)=-\frac13F^2
\end{equation*}
or
\begin{equation*}
dF=\alpha,\quad d\beta=\frac16F^2.
\end{equation*}
The metric is Einstein with $s=-\frac5{18}$ in the former case, and Ricci-flat in the latter.
\end{corollary}
\begin{proof}
It is clear from Proposition~\ref{prop:forms} that the two situations correspond to the intrinsic torsion classes $W_1+W_5$ or $W_1$.
Conversely, if the intrinsic torsion lies in $W_1+W_5$, with $\tau_1\neq0$, necessarily $dF$ has a component of type $(3,0)$; we can choose the reduction so that it equals $\alpha$. Then
\begin{equation*}
dF=\alpha + k\beta,\quad (d\beta)^{2,2}=\frac16F^2,
\end{equation*}
for some function $k$. If $k$ is not zero at some point, we have
\[(d\alpha)^{3,1}=0,\quad (d\alpha)^{2,2}=-\frac{k}6F^2,\quad (d\beta)^{1,3}=-d(\log k)\wedge \beta.\]
This implies that $d\lambda = 0$. Therefore, the metric is Einstein with curvature $-\frac{5k}{18}$; up to rescaling we can assume $k\equiv 1$.
Otherwise $k$ is identically zero; this implies that
\[0=d^2\beta =d(3\lambda\wedge\beta + \frac16F^2)= 3d\lambda\wedge \beta - \frac12\lambda\wedge F^2;\]
since $\lambda$ has type $(1,0)$, this implies that $\lambda=0$. Hence, the metric is Ricci-flat.
\end{proof}
The six-dimensional case is also special because a nearly parak\"ahler $6$-manifold gives rise to a nearly parallel $\Gtwo^*$-structure on a suitable warped product (\cite{CortesLeistner}).
\section{The Bott-Chern class}
Whilst the component $\tau$ of the intrinsic torsion depends only on the $\GL(n,\R)$-structure, the component $\lambda$ only depends on the $\SL(n,\R)\times \SL(n,\R)$-structure. In analogy with the complex case, it is possible to define an invariant that does not depend on a metric, playing the same role as the first Chern class.
Recall that the (para)Bott-Chern class cohomology spaces are defined as
\[H^{p,q}_{BC} =\frac{ \ker d\colon\Omega^{p,q}\to \Omega^{p+1,q}\oplus\Omega^{p,q+1} }{\im dd^c\colon \Omega^{p-1,q-1}\to \Omega^{p,q}},\]
where $d^c=K\circ d\circ K$. They are not finite-dimensional in general \cite{AngellaRossi:cohomology2012}.
\begin{proposition}
\label{prop:modlambda}
Let $(M,K)$ be a paracomplex manifold that admits a reduction to $\SL(n,\R)$, say $(F,\alpha,\beta)$. If $(\tilde F,\tilde\alpha, \tilde\beta)$ is another reduction to $\SL(n,\R)$ with
\[\tilde \alpha = e^{n(h+k)} \alpha, \quad \tilde\beta=e^{n(h-k)}\beta,\]
then its intrinsic torsion satisfies
\[\tilde\lambda = \lambda + d^ch -dk.\]
In particular, the Bott-Chern class $[d\lambda]$ only depends on $K$.
\end{proposition}
\begin{proof}
Straightforward from Proposition \ref{prop:forms}:
\begin{gather*}
-n\tilde\lambda\wedge\tilde \alpha = d(\tilde\alpha ) = n(dh+dk)\wedge \tilde\alpha -n \lambda^{0,1}\wedge \tilde\alpha\\
n\tilde\lambda\wedge\tilde \beta = d(\tilde\beta ) = n(dh-dk)\wedge \tilde\beta +n\lambda^{1,0}\wedge \tilde\beta.\qedhere\\
\end{gather*}
\end{proof}
\begin{proposition}
Let $(M,K)$ be a paracomplex manifold which admits a reduction to $\SL(n,\R)$. If $(M,K)$ admits a compatible Ricci-flat parak\"ahler metric, then $[d\lambda]=0$ in $H^{1,1}_{BC}$.
\end{proposition}
\begin{proof}
Theorem~\ref{thm:restateRicViaForms} gives $\ric'=d\lambda$. Since $[d\lambda]$ only depends on $K$, if a Ricci-flat parak\"ahler metric exists necessarily $d\lambda=0$.
\end{proof}
We do not know whether an analogue of the Aubin-Yau theorem holds~\cite{Aubin,Yau}: Proposition~\ref{prop:modlambda} only tells us what the volume of a potential Ricci-flat parak\"ahler metric should be, but does not guarantee existence.
On the other hand, we have the following:
\begin{proposition}
Let $(F,\alpha,\beta)$ be a paracomplex $\SL(n,\R)$-structure on $M$. Then $M$ admits a parak\"ahler-Einstein metric with $s\neq0$ compatible with the same paracomplex structure if and only if there exists a function $h$ such that
\[(d\lambda + dd^ch)^n = e^{2nh}F^n.\]
\end{proposition}
\begin{proof}
Any compatible $\SL(n,\R)$-structure has the form $(\tilde F,\tilde\alpha, \tilde\beta)$, with
\[\tilde \alpha = e^{n(h+k)} \alpha, \quad \tilde\beta=e^{n(h-k)}\beta.\]
Since $k$ does not affect the metric, it is no loss of generality to assume $k=0$. Then from Proposition \ref{prop:forms} we have:
\[d\tilde\lambda = d \lambda + dd^ch.\]
Set $\tilde F=d\tilde\lambda$; then $(\tilde F,\tilde \alpha,\tilde\beta)$ defines an $\SL(n,\R)$-structure if and only if
\[\tilde F^n = e^{2nh}F^n.\]
If this condition holds, the structure is automatically parak\"ahler and Einstein, because $\tilde F$ is closed and because of Theorem~\ref{thm:ricci}.
By the same token, a compatible parak\"ahler-Einstein metric necessarily satisfies $\tilde F=d\tilde\lambda$.
\end{proof}
This condition is the analogue of the Monge-Amp\`ere equation in complex geometry. A striking difference is that the volume form is automatically exact; this means that it makes no sense to assume the manifold is compact. In fact, \emph{compact} Einstein parak\"ahler manifolds are necessarily Ricci-flat:
\begin{proposition}
\label{prop:pkeinsteincompact}
If $M$ is a compact, parak\"ahler manifold, then
\[\int_M s F^n =0.\]
A left-invariant parak\"ahler structure on a unimodular Lie group has $s=0$.
\end{proposition}
\begin{proof}
By Corollary~\ref{cor:s}, $d\lambda\wedge F^{n-1} = \frac1n sF^n$; this form is exact because $F$ is symplectic. If $M$ is compact, Stokes' theorem gives the statement; the same argument applies to the case of a unimodular Lie group, since there are no exact invariant volume forms.
\end{proof}
\begin{remark}
More generally, this argument applies to \emph{balanced} almost parak\"ahler structures, i.e. those where $F^{n-1}$ is closed; then, in the case of an invariant structure on a unimodular Lie group, $\Lambda(d\lambda)=0$ and $\tau_4=0=\tau_8$. Therefore, the expression for the scalar curvature reduces to
\[
s=\frac{10}n\langle \tau_1,\tau_5\rangle -\frac2n \langle \tau_2,\tau_6\rangle-\frac2n \langle \tau_3,\tau_7\rangle.
\]
\end{remark}
\smallskip
Another difference with the (compact) complex case is that an Einstein para\-k\"ahler manifold with $s\neq0$ can have $[d\lambda]=0$; in fact, even $H^{1,1}_{BC}=0$ as in Example \ref{ex:Einstein}. On a paracomplex manifold $M$, decompose $d$ as $\partial_++ \partial_-$,
\[\partial_+\colon \Omega^{p,q}\to \Omega^{p+1,q}, \quad \partial_-\colon \Omega^{p,q}\to \Omega^{p,q+1}.\]
We can then define the (para)Dolbeault cohomology
\[H^{p,q}_+ = \frac{\ker \partial_+\colon \Omega^{p,q}\to \Omega^{p+1,q} }{\im \partial_+\colon \Omega^{p-1,q}\to \Omega^{p,q}},\quad H^{p,q}_- = \frac{\ker \partial_-\colon \Omega^{p,q}\to \Omega^{p,q+1} }{\im \partial_-\colon \Omega^{p,q-1}\to \Omega^{p,q}}.\]
\begin{theorem}
\label{thm:aubinyau_dei_poveri}
Let $(M,K)$ be a paracomplex manifold with
\[H^1(M)=0=H^2(M)=H^{1,0}_+(M)=H^{0,1}_-(M).
\]
Then $H^{1,1}_{BC}$ is zero.
\end{theorem}
\begin{proof}
Let $\mathcal{Z}^{1,1}$ be the sheaf of closed $(1,1)$-forms, and let $\mathcal{K}$ be the sheaf of $dd^c$-closed functions; we have a short exact sequence of sheaves
\[0\to \mathcal{K}\to \mathcal{A}^0\xrightarrow{dd^c} \mathcal{Z}^{1,1}\to 0.\]
This gives
\[0\to H^0(\mathcal{K})\to H^0(\mathcal{A}^0)\xrightarrow{dd^c} H^0( \mathcal{Z}^{1,1})\to H^1(\mathcal{K})\to 0,\]
where we have used the fact that $\mathcal{A}^0$ is acyclic. Under the present assumptions, the $dd^c$ lemma is equivalent to $H^1(\mathcal{K})=0$.
Now consider the sheaves
\[\mathcal{Z}^{p,q}_\pm =\ker \partial_{\pm}\colon \mathcal{A}^{p,q}\to \mathcal{A}^{p+1+q} ;\]
then the sequence
\[0\to\mathcal{Z}^0\to \mathcal{Z}^{0,0}_+\oplus \mathcal{Z}^{0,0}_-\to \mathcal{K}\to 0\]
is exact. Indeed, assume the first and second de Rham cohomology of an open set $U$ is zero; then
\[\gamma\in \mathcal{K}(U)\implies dd^{c}\gamma=0 \implies d^c\gamma = df\implies \partial_+(\gamma-f)=0,\ \partial_-(\gamma+f)=0;\]
thus $\gamma\in \mathcal{Z}^{0,0}_+(U)\oplus \mathcal{Z}^{0,0}_-(U)$.
Taking now the associated long exact sequence, and using the fact that $H^1(M)$ and $H^2(M)$ are zero, we deduce
\[H^1(\mathcal{K})=H^1( \mathcal{Z}^{0,0}_+)\oplus H^1( \mathcal{Z}^{0,0}_+).\]
Since a paracomplex manifold is locally a product, the following sequence is exact:
\[0\to \mathcal{Z}^{0,0}_+\to \mathcal{A}^0\xrightarrow{\partial_+}\mathcal{Z}^{1,0}_+\to 0,\]
giving the exact sequence in cohomology
\[H^0(\mathcal{A}^0)\xrightarrow{\partial_+}H^0(\mathcal{Z}^{1,0}_+)\to H^1(\mathcal{Z}^{0,0}_+)\to 0.\]
This shows that $H^1(\mathcal{Z}^{0,0}_+)=H^{1,0}_+(M)=0$, and a similar result applies to the minus component. Thus, $H^1(\mathcal{K})=0$, which is what we had to prove.
\end{proof}
\begin{example}\label{ex:Einstein}
On the solvable Lie group $G=\mathrm{Aff}_+(\R)\times \mathrm{Aff}_+(\R)\times \mathrm{Aff}_+(\R)$, consider a left-invariant basis of one-forms $e^1,\dotsc, e^6$ such that
\[de^1=de^4=e^{14},\quad de^2=de^5=e^{25},\quad de^3=de^6=e^{36}.\]
We consider the $\SL(3,\R)$-structure for which this is an adapted coframe; it is parak\"ahler and Einstein with $s\neq0$. This Lie group is diffeomorphic to $\R^6$; in addition, it is easy to check that the space of leaves for both foliations is smooth and diffeomorphic to $\R^3$, so that Theorem~\ref{thm:aubinyau_dei_poveri} applies and the Bott-Chern class $[d\lambda]$ is zero. Notice however that up to multiple, there is only one \emph{invariant} closed $(1,1)$-form: if a compatible Ricci-flat parak\"ahler metric exists, it cannot be invariant.
\end{example}
We conclude this section with a remark concerning the holonomy group of a parak\"ahler manifold. By construction, this is a subgroup of $\GL(n,\R)$. If it is also contained in $\GL_+(n,\R)$, the bundle $\Lambda^{n,0}$ is trivial and it is possible to fix a reduction to $\SL(n,\R)$. Ricci-flatness means that the \emph{restricted} holonomy is $\SL(n,\R)$; therefore, the full holonomy group has the form $G\times\GL(n,\R)$, with $G$ a discrete subgroup of $\GL_+(n,\R)$.
\begin{example}
Let $\lambda>0$ be a real constant and let
\[\Omega=\{(x,y)\in\R^n\times \R^n \mid x_1>\frac1{1-\lambda}\}.\]
Let $g$ be the affine transformation
\[(x,y)\mapsto (\lambda x + e_1, \lambda^{-1}y).\]
The group generated by $g$ acts on $\Omega$ in a free, proper discontinuous way. Let $M=\Omega/\langle g\rangle$ be the quotient.
The flat parak\"ahler structure
\[F=dx_1\wedge dy_1+\dots + dx_n\wedge dy_n, \quad h=\sum dx_i\odot dy_i\]
passes onto the quotient. It is easy to see that
\[\{(\lambda^k dx_1, \dots, \lambda^k dx_n, \lambda^{-k}dy_1,\dots, \lambda^{-k}dy_n) \mid k\in\Z\}\]
defines a parallel $\Z$-structure on $\Omega$; this structure also passes onto the quotient. Moreover, taking parallel transport on $\Omega$ and observing that
\[\pi_{*gp}(\D{x_i})=\lambda^{-1}\pi_{*p}(\D{x_i}), \quad \pi_{*gp}(\D{y_i})=\lambda\pi_{*p}(\D{y_i}), \]
one sees that the holonomy of $M$ is precisely $\Z$.
\end{example}
\section{Einstein examples}
\label{sec:examples}
In this section we show some examples of Einstein metrics associated to $\SL(n,\R)$-structures on Lie groups. With one exception (Example~\ref{example:non_nilpotent}), the Lie groups we consider are nilpotent with rational structure constants; therefore, each admits a compact quotient which carries an induced $\SL(n,\R)$-structure, also Einstein.
We will represent Lie groups by their structure constants; for instance the quadruplet $(0,0,0,12)$ represents a four-dimensional Lie group with a basis of left-invariant one-forms $e^1,e^2,e^3,e^4$ such that
\[de^1=0,\ de^2=0,\ de^3=0,\ de^4=e^{12};\]
the \emph{standard} $\SL(2,\R)$-structure on this Lie group is the one defined by the coframe $e^1,e^2,e^3,e^4$.
\begin{example}
\label{example:ricciflat}
A non-flat example of a parak\"ahler Ricci-flat manifold is the standard $\SL(n,\R)$-struc\-ture on the nilpotent Lie algebra $(24,0,0,0,0,35)$; in this case the curvature is
$-e^{34}\otimes e^{34}$.
\end{example}
\begin{example}
\label{example:non_nilpotent}
There also exist parak\"ahler Einstein Lie groups with non-zero scalar curvature: consider the one-parameter family of Lie algebras
\[(14,25,36,t14,t25,t36), \quad t\in\R;\]
notice that these are not nilpotent (in fact, not even unimodular). The standard $\SL(n,\R)$-structure is Einstein, and Ricci-flat when $t=0$. In addition, the parak\"ahler metric is unique once one fixes the paracomplex structure, if one requires it to be invariant.
\end{example}
\begin{example}
The Lie group $(0,0,0,-\frac13 e^{23},-\frac13 e^{31},-\frac13 e^{12})$ is flat and nearly para\-k\"ahler; in this case, the minimal connection is the flat connection.
We can modify the example in such a way that the minimal connection is a non-trivial element in the subspace of $T^*\otimes\Sl(3,\R)$ isomorphic to $V_2$; imposing $d^2=0$, we obtain the following Lie algebra:
\begin{gather*}
de^1=-2 \lambda_3 e^{12}-2 e^{23}+2 e^{13} \lambda_2\\
de^2=-2 \lambda_3 e^{12} \lambda_2-2 e^{23} \lambda_2+2 e^{13} \lambda_2^{2}\\
de^3=-2 \lambda_3 e^{23}+2 \lambda_3 e^{13} \lambda_2-2 \lambda_3^{2} e^{12}\\
de^4={(e^{36}-e^{25})} \lambda_3 \lambda_2- \lambda_3 e^{24}-\frac{1}{3} e^{23}+ e^{34} \lambda_2- \lambda_3^{2} e^{26}+ e^{35} \lambda_2^{2}\\
de^5=\frac{1}{3} e^{13}-e^{34}+ \lambda_3 e^{15} \lambda_2- e^{35} \lambda_2+ \lambda_3^{2} e^{16}- {(e^{36}-e^{14})} \lambda_3\\
de^6=\lambda_3 e^{26}-\frac{1}{3} e^{12}+e^{24}- \lambda_3 e^{16} \lambda_2+ {(e^{25}-e^{14})} \lambda_2- e^{15} \lambda_2^{2}.
\end{gather*}
The standard structure on this Lie algebra is again flat and nearly parak\"ahler. A suitable change of basis, namely
\[\{-e^2+ e^1 \lambda_2,e^3- \lambda_3 e^1,-\frac{1}{6} e^1+e^4+ e^5 \lambda_2+ \lambda_3 e^6,\frac{1}{3} e^1+e^4+ e^5 \lambda_2+ \lambda_3 e^6,- e^6,- e^5\},\]
shows the Lie algebra to be nilpotent and isomorphic to $(0,0,0,12,14,24)$.
\end{example}
It follows from Theorem~\ref{thm:restateRicViaForms} that if the intrinsic torsion is purely in $W_2$, then the metric is Ricci flat. This leads us to a counterexample to the paracomplex version of the Goldberg conjecture of almost-K\"ahler geometry. This ``para-Goldberg'' conjecture, as stated in \cite{Matsushita}, asserts that any compact almost parak\"ahler Einstein manifold is parak\"ahler.
\begin{proposition}
\label{prop:goldberg}
The standard $\SL(2,\R)$-structure on the nilmanifold $(0,0,0,12)$ defines a compact, almost parak\"ahler Ricci-flat manifold which is not parak\"ahler. Hence, the para-Goldberg conjecture does not hold.
\end{proposition}
\begin{proof}
One easily sees that $\lambda=0$ and the only component of the intrinsic torsion is $\tau_2$, implying Ricci-flatness (in fact, this metric is flat). The fact that $de^4$ has a component in $\Lambda^{2,0}$ shows that the paracomplex structure is not integrable.
\end{proof}
\begin{example}
Another example in the almost parak\"ahler class $W_2+W_6$ is the nilpotent Lie algebra
\[(0,0,46,0,12,0),\]
with the standard $\SL(3,\R)$-structure; its intrinsic torsion is
\[\tau_2 = e_2\otimes e^{12}, \quad \tau_6 = e_6\otimes e^{46}.\]
This metric is Ricci-flat, but the Levi-Civita curvature is $e^{26}\otimes e^{26}$.
\end{example}
\begin{example}
A Ricci-flat example with intrinsic torsion in $W_3$ is given by the standard $\SL(n,\R)$-structure on the nilpotent Lie algebra $(0,0,0,0,0,45)$; this structure is almost parak\"ahler and flat, with
\[\tau = \frac{1}{2} e_3\otimes e^{45}\in W_3.\]
A non-flat example is given by the standard $\SL(n,\R)$-structure on the $8$-dimensional nilmanifold $(0,0,0,0,0,0,56,57)$; then
\[\tau=\tau_3 = \frac{1}{2} e_3\otimes e^{56}+\frac{1}{2} e_4\otimes e^{57}\]
and the Ricci tensor is zero, but the curvature is $-\frac14 e^{56}\odot e^{45}$.
\end{example}
The nearly parak\"ahler examples we have shown so far are flat, and the geometry of flat nearly parak\"ahler structures is fully understood (see \cite{CortesSchafer}). We now illustrate a systematic approach to construct Ricci-flat nearly parak\"ahler metrics that are not flat. Our examples have intrinsic torsion in $W_1$; by Theorem~\ref{thm:restateRicViaForms}, this condition ensures Ricci-flatness.
The intrinsic torsion of an almost paracomplex structure can be identified with the Nijenhuis tensor, which has two components
\[N^H+N^V \in \Lambda^2V^*\otimes H+\Lambda^2H^*\otimes V.\]
Under the structure group $\GL(n,\R)$, these two irreducible modules decompose respectively into $W_1+W_2$ and $W_5+W_6$; by Proposition~\ref{prop:paracomplexit}, we can write
\[N^H=4\partial(\tau_1+\tau_2), \quad N^V=4\partial(\tau_5+\tau_6).\]
The key observation is that $N^H$ can be identified with a linear bundle map that does not depend on the almost paracomplex structure $K$ but only on the $1$-eigendistribution $\mathcal{V}$, namely
\[\tilde N^H\colon \mathcal{V}\ann\to \Lambda^2\mathcal{V}^*,\]
where $\mathcal{V}\ann$ is the subbundle of $T^*M$ whose fibre at $p$ is the annihilator of $\mathcal{V}_p$.
\begin{proposition}
\label{prop:almostW1}
Let $\lie{g}$ be a Lie algebra of dimension $2n$; let $V\subset\lie{g}$ be a subspace of dimension $n$ such that:
\begin{enumerate}
\item $\tilde N^H$ has rank three;
\item the image $W=\im \tilde N^H$ consists of simple forms, i.e. if $\eta\in W$ then $\eta\wedge\eta=0$;
\item there is no $\sigma\in V^*\setminus\{0\}$ such that $\sigma\wedge\eta=0$ for all $\eta\in W$.
\end{enumerate}
Then for any subalgebra $H\subset \lie{g}$ complementary to $V$ there is a $\GL(n,\R)$-structure compatible with the splitting $\lie{g}=V\oplus H$
such that
\[\tau_2=0=\tau_5=\tau_6.\]
\end{proposition}
\begin{proof}
Conditions 1 and 2 imply that $V^*$ is spanned by $e^{1},\dots, e^{n}$ such that $W$ is either of the form
\[\Span{e^{12},e^{13},e^{23}}\]
or
\[\Span{e^{14},e^{24},e^{34}};\]
the latter possibility is ruled out by Condition 3.
Having fixed the splitting $V\oplus H$, we can identify $V\ann$ with $H^*$; in particular, we can write
\[\tilde N^H=\eta^1\otimes e^{23}+\eta^2\otimes e^{31}+\eta^3\otimes e^{12},\]
where $\eta^1,\eta^2,\eta^3$ are linearly independent elements of $H^{**}$. Let $e^{n+1},\dots, e^{2n}$ be a basis of $H^*$ such that $\eta^i(e^{n+j})=\delta_{ij}$; relative to the coframe $e^1,\dotsc, e^{2n}$, we can write
\[N^H=e^{23}\otimes e_{n+1}+e^{31}\otimes e_{n+2}+e^{12}\otimes e_{n+3};\]
thus, this coframe defines an $\SL(n,\R)$-structure such that
\[\tau_1+\tau_2 =\frac18( e^1\otimes e^{23}+e^2\otimes e^{31}+e^3\otimes e^{12}),\] i.e. $\tau_2=0$. The fact that $\tau_5=0=\tau_6$ follows from $H$ being integrable.
\end{proof}
A similar result holds for distributions $\mathcal{V}$ and $\mathcal{H}$ on a manifold, if Conditions 1---3 are assumed to hold at each point, but we will only consider the case of Lie groups. Then, the condition $\tau_2=0$ can be expressed as a linear equation in $F$ by using the following contraction:
\begin{equation*}
c\colon (\Lambda^2V^*\otimes H)\otimes \Lambda^2T^*\to \Lambda^2V^*\otimes V^*, \quad (\eta\otimes h)\otimes F\mapsto \eta\otimes h\hook F.
\end{equation*}
\begin{corollary}
Let $\lie{g}=V\oplus H$ be a splitting such as in Proposition~\ref{prop:almostW1}, and assume that $d(\Lambda^{0,n})\subset\Lambda^{2,n-1}$. Then any non-degenerate form $F\in\Lambda^{1,1}$ such that
\begin{equation}
\label{eqn:linearricciW1}
dF\in\Lambda^{3,0}, \quad c(N^H,F)\in\Lambda^{3,0}
\end{equation}
defines a Ricci-flat $\GL(n,\R)$-structure with intrinsic torsion in $W_1$.
\end{corollary}
\begin{proof}
Fix a compatible reduction to $\SL(n,\R)$; then Proposition~\ref{prop:forms} implies
\[\lambda=0, \quad \tau_3=0=\tau_4=\tau_7=\tau_8.\]
By Proposition~\ref{prop:almostW1}, $\tau_2$, $\tau_5$ and $\tau_6$ vanish as well; thus the intrinsic torsion is in $W_1$, and by Theorem~\ref{thm:restateRicViaForms} the metric is Ricci-flat.
\end{proof}
This gives us an effective recipe to find Ricci-flat nearly parak\"ahler examples: for a given Lie algebra $\lie{g}$ of dimension $2n$, we seek a subspace $V\subset\lie{g}$ of dimension $n$ that satisfies the conditions of Proposition~\ref{prop:almostW1}, and then a complementary integrable distribution $H$ such that $d(\Lambda^{0,n})\subset\Lambda^{2,n-1}$. At this point it is only a matter of computing the space of forms of type $(1,1)$ that satisfy the linear equations \eqref{eqn:linearricciW1}, and verifying whether it contains a non-degenerate form. Implementing this strategy with a computer on Lie algebras of the form $\lie{h}\oplus\R$, where $\lie{h}$ ranges among $7$-dimensional nilpotent Lie algebras as classified by \cite{Gong}, gives the following:
\begin{theorem}
\label{thm:nkricciflat}
Each Lie algebra in Table~\ref{fig:nkricciflat} admits a family of non-flat, Ricci-flat nearly-parak\"ahler structures with intrinsic torsion $\tau_1\neq0$ and non-zero curvature, depending on parameters $\lambda,\mu\neq0$ and $k$.
\end{theorem}
We emphasize that this is not a classification; whilst our program did not find any other example, we do not know whether there exist other $8$-dimensional nilpotent Lie algebras with a strictly nearly parak\"ahler structure which is Ricci-flat but not flat.
\begin{table}[t]
\caption{Ricci-flat nearly-parak\"ahler structures on nilpotent Lie algebras}\label{fig:nkricciflat}
\resizebox{\textwidth}{!}{
\bgroup
\def\arraystretch{1.5}
\begin{tabular}{>{$}l<{$}|>{$}l<{$}}
\lie{h} & \lie{g} \text{ (in adapted coframe)}\\
\midrule
0,0,0,e^{12},e^{23}+e^{14},e^{24},e^{15}-e^{34}&0,0,-\lambda e^{18},0,-\mu e^{23} +\lambda e^{78}, \mu e^{13},- \mu e^{12} +\lambda\mu e^{38},0\\
0,0,0,e^{12},e^{23}+e^{14},e^{24}+e^{13},-e^{34}+e^{15}&0,0,-\lambda e^{18},0,\lambda e^{78}- \mu e^{23},-\lambda\mu e^{28} + \mu e^{13},\lambda \mu e^{38} - \mu e^{12},0\\
0,0,e^{12},0,e^{24}+e^{13},e^{23},e^{34}+e^{15}+e^{26}&0,-\lambda e^{18},0,0, \mu e^{23}+e^{17}+\lambda e^{68}-k\lambda e^{18} ,- \mu e^{13}+\lambda\mu e^{28}, \mu e^{12},0\\
0,0,0,e^{12},e^{23}+e^{14},-e^{24}+e^{13},-e^{34}+e^{15}&0,0,-\lambda e^{18},0,- \mu e^{23} +\lambda e^{78}, \mu e^{13} +\lambda \mu e^{28},- \mu e^{12}+\lambda\mu e^{38} ,0\\
0,0,e^{12},0,e^{13}+e^{24},e^{23},-e^{26}+e^{15}+e^{34}&0,-\lambda e^{18},0,0,-e^{17}+ \mu e^{23} +\lambda e^{68}+k\lambda e^{18},\lambda\mu e^{28} - \mu e^{13}, \mu e^{12},0
\end{tabular}
\egroup
}
\end{table}
\section{Non-existence results}
The examples of Section~\ref{sec:examples} correspond to left-invariant structures on nilpotent Lie groups; in this section we give some non-existence results in the same context.
\begin{lemma}
\label{lemma:paracomplexNilpotent}
Every left-invariant paracomplex $\SL(n,\R)$-structure on a nilpotent Lie group satisfies
\[\lambda=\frac{n-1}n (f_8-f_4).\]
\end{lemma}
\begin{proof}
Let $\lie{g}$ be a nilpotent Lie algebra with a fixed $\SL(n,\R)$-structure. For any $X\in\lie{g}$, $\ad X$ is nilpotent; the induced action on $\Lambda\lie{g}^*$ is also nilpotent.
By Proposition~\ref{prop:forms},
\[d\alpha = (-n\lambda-(n-1)f_4)\wedge \alpha;\]
whenever $X$ is in $H$, $\alpha$ is an eigenvector for the nilpotent endomorphism $\ad X$ with eigenvalue \[(-n\lambda-(n-1)f_4)(X);\] it follows that
\[n\lambda^{0,1}=-(n-1)f_4.\]
The same argument applied to $\beta$ yields $n\lambda^{1,0}=(n-1)f_8$.
\end{proof}
Theorem~\ref{thm:restateRicViaForms} immediately implies:
\begin{corollary}
Invariant parak\"ahler structures on nilpotent Lie groups are Ricci-flat.
\end{corollary}
In contrast with the Riemannian case, Ricci-flat nilpotent Lie groups are not necessarily abelian; Section~\ref{sec:examples} contains several examples. Alongside the parak\"ahler case (Example~\ref{example:ricciflat}), we have constructed Ricci-flat examples with intrinsic torsion in $W_1$, $W_2$ and $W_3$; the absence of the class $W_4$ can be explained by the following:
\begin{proposition}
\label{prop:nilpotentw4w8}
A nilpotent Lie group with an invariant, Einstein almost parak\"ahler structure with intrinsic torsion in $W_4+W_8$
is only Einstein if it is parak\"ahler, in which case it is Ricci-flat.
\end{proposition}
\begin{proof}
Fix a reduction to $\SL(n,\R)$; by Lemma~\ref{lemma:paracomplexNilpotent},
\[d\lambda = \frac{n-1}n (df_8-df_4).\]
On the other hand, Proposition~\ref{prop:forms} gives
$dF=-2(f_4+f_8)\wedge F;$
taking the exterior derivative,
$0=-2(df_4+df_8)\wedge F,$
and consequently $df_8=-df_4$. In addition, this two-form has type $(1,1)$ because the structure is paracomplex. Thus,
\[d\lambda = -2\frac{n-1}n df_4.\]
The formula for $\ric'$ gives
\[\ric'=2(n-2)df_4+nd\lambda-2(n-1)f_4\wedge f_8 \mod F\]
Thus, the metric is Einstein if there exists $k$ such that
\begin{equation}
\label{eqn:f4f8einstein}
df_4+(n-1)f_4\wedge f_8 =kF;
\end{equation}
taking the exterior derivative,
\begin{multline*}
(n-1)df_4\wedge f_8 +(n-1)f_4\wedge df_4=-2k(f_4+f_8)\wedge F \\
= -2(f_4+f_8)\wedge(df_4+(n-1)f_4\wedge f_8) =-2(f_4+f_8)\wedge df_4;
\end{multline*}
this gives
\[(n+1)df_4\wedge(f_4+f_8)=0,\]
and decomposing according to type
\[df_4\wedge f_4=0, \quad df_4\wedge f_8=0.\]
Therefore,
\[df_4\in \Span{f_4\wedge f_8};\]
this is only possible if $df_4$ is zero; by \eqref{eqn:f4f8einstein} this implies $f_4=f_8=0$, which in turn gives $\lambda=0$.
\end{proof}
\begin{remark}
Given an $\SL(n,\R)$-structure on a manifold, it follows from Proposition~\ref{prop:forms} that the intrinsic torsion is contained in $W_4+W_8+W^{1,0}+W^{0,1}$ if and only if $N=0$ and $dF=\theta\wedge F$ for some $1$-form $\theta$; such structures are known in the literature as locally conformally parak\"ahler \cite{GadeaMasque}. The result of Proposition~\ref{prop:nilpotentw4w8} can be rephrased by saying that an Einstein, left-invariant locally conformally parak\"ahler structure on a nilpotent Lie group is parak\"ahler.
\end{remark}
\section{Composite intrinsic torsion classes}
In this section we give a simple construction to build almost parahermitian manifolds in a prescribed intrinsic torsion class using almost parahermitian manifolds of lower dimension as building blocks.
We shall denote by $T_n$ the $2n$-dimensional $\GL(n,\R)$-module \eqref{eqn:linearmodel}; let $W_i(T_{n})$ denote the corresponding intrinsic torsion spaces. We will also use a subscript $n$ to denote the $2n$-dimensional forms $F_n$, $\alpha_n$ and $\beta_n$.
We identify $T_n\oplus T_m$ with $T_{n+m}$ by the isomorphism that maps the basis
\[\{e^1,\dots,e^n,\allowbreak f^1,\dots, f^m,\allowbreak e^{n+1},\dots,e^{2n},\allowbreak f^{m+1},\dots, f^{2m}\}\]
onto the basis
\[\{E^1,\dotsc, E^{2(m+n)}\};\]
here, $\{e_i\}$, $\{f_i\}$ and $\{E_i\}$ are the standard bases of $T_n$, $T_m$ and $T_{n+m}$ respectively. We obtain the following relations between the forms $F$, $\alpha$ and $\beta$:
\[F_{n+m}=F_n+F_m,\quad \alpha_{n+m}=\alpha_{n}\wedge\alpha_{m},\quad\beta_{n+m}=\beta_{n}\wedge\beta_{m}.\]
Moreover, we have the following lemma:
\begin{lemma}\label{lemma:higherdim}
The following relations hold:
\begin{gather*}
W_i(T_{n})\subset W_i(T_{n+m})\qquad\text{for}\ i\neq4,8,\\
W_4(T_{n})\subset W_3(T_{n+m})\oplus W_4(T_{n+m}),\qquad W_8(T_{n})\subset W_7(T_{n+m})\oplus W_8(T_{n+m}).
\end{gather*}
Moreover, the projections
\begin{align*}
W_3(T_n)\oplus W_4(T_n)\oplus W_3(T_m)\oplus W_4(T_m)&\to W_3(T_{n+m})\\
W_7(T_n)\oplus W_8(T_n)\oplus W_7(T_m)\oplus W_8(T_m)&\to W_7(T_{n+m})
\end{align*}
are injective.
\end{lemma}
\begin{proof}
Consider the commutative diagram
\begin{equation*}
\xymatrix{
\Lambda^2T^*_n\otimes T_n\ar@{->}[r]^f\ar@{->}[d] & \Lambda^2T^*_{n+m}\otimes T_{n+m}\ar@{->}[d] \\
\text{coker}(\partial_n)\ar@{->}[r]^{\hat{f}} & \text{coker}(\partial_{n+m})
}
\end{equation*}
where $f(a_{IJK}e^{IJ}\otimes e_K)=a_{IJK}e^{IJ}\otimes e_K$. The map $\hat{f}$ is injective because $f(\Lambda^2T^*_n\otimes T_n)\cap\partial(T^*_{n+m}\otimes \gl({n+m}))=\{0\}$.
For $i\neq4,8$, let $\tau_i$ denote an element of $W_i(T_n)$. It is obvious that $\Lambda^3V_n^*\subset\Lambda^3V_{n+m}^*$; hence, $\hat{f}(\tau_1)\in W_1(T_{n+m})$. Then the inclusion easily follows for $\tau_2$: in fact, $\hat{f}(\tau_1+\tau_2)\in V^*_{n+m}\otimes\Lambda^2V^*_{n+m}$ and $\hat{f}(\tau_2)$ has no component in $\Lambda^3V_{n+m}^*$ because $\tau_2$ has no component in $\Lambda^3V_{n}^*$. Now recall that $W_3(T_n)$ is the kernel of the contraction $V_n^*\otimes\Lambda^2V_n\to V_n$; it easily follows that $\hat{f}(\tau_3)$ belongs to the kernel of the corresponding contraction defined on $V^*_{n+m}\otimes\Lambda^2V_{n+m}$.
Finally, if $\tau_4=a_ie^k\otimes e_{ki}$, then $\hat{f}(\tau_4)\in V^*_{n+m}\otimes\Lambda^2V_{n+m}$ is given by
\[
\hat{f}(\tau_4)=a_ie^k\otimes e_{ki}
=\biggl(a_ie^k\otimes e_{ki}-\frac{n-1}{m}a_if^h\otimes f_h\wedge e_i\biggr)+\frac{n-1}{m}a_if^h\otimes f_h\wedge e_i,
\]
where the first summand lies in $W_3(T_{n+m})$ and the second one lies in $W_4(T_{n+m})$; a similar decomposition applies to any element $\tau_4'= a_i'f^k\otimes f_{ki}$ of $W_4(T_{n+m})$. Writing explicitly the image in $W_3(T_{n+m})$ of an element $\tau_3+\tau_4+\tau_3'+\tau_4'$ and imposing that the component in $V_m^*\otimes (V_m\wedge V_n)$ is zero we obtain $\tau_4=0$; similarly, one finds that $\tau_4'=0$. Injectivity follows.
The same arguments can be applied to $\tau_5, \tau_6,\tau_7$ and $\tau_8$.
\end{proof}
At the geometric level, given two almost parahermitian manifolds $(N,g_n,\allowbreak K_n,F_n)$ and $(M,g_m,K_m,F_m)$, on the product $N\times M$ we consider the natural almost parahermitian structure
\[(N\times M, g_n+g_m, K_n+K_m,F_n+F_m).\]
In our setting, an \emph{intrinsic torsion class} is a subset of $\{\mathcal{W}_1,\dots, \mathcal{W}_8\}$; it is customary to represent a subset $\{\mathcal{W}_{i_1},\dots, \mathcal{W}_{i_h}\}$ as a formal sum
\begin{equation}
\label{eqn:formalsum}
\mathcal{W}_{i_1}+ \dots + \mathcal{W}_{i_h}.
\end{equation}
Accordingly, the union of two intrinsic torsion classes $I$ and $J$ is written as $I+J$.
An almost parahermitian manifold $(M,g,K,F)$ is in the intrinsic torsion class \eqref{eqn:formalsum} if the components of the intrinsic torsion which are not identically zero are precisely $\tau_{i_1},\dots, \tau_{i_h}$; we write $\mathcal{W}(M)=\mathcal{W}_{i_1}+ \dots + \mathcal{W}_{i_h}$.
\begin{proposition}\label{prop:higherdim}
If $(N,g,K,F)$ is a $2n$-dimensional almost parahermitian manifold with $\tau_4=0=\tau_8$, then for each $m\in\N$ the natural almost parahermitian structure on the $2(m+n)$-dimensional manifold $N\times\R^{2m}$ is in the same intrinsic torsion class.
\end{proposition}
\begin{proof}
Follows from Proposition~\ref{prop:forms} and Lemma \ref{lemma:higherdim}.
\end{proof}
More generally, we can find non-trivial intrinsic torsion classes by combining two almost parahermitian manifolds. Indeed, by the same arguments we can prove the following:
\begin{proposition}
Let $M,N$ be two almost parahermitian manifolds of dimension respectively $2m$ and $2n$, each with $\tau_4=0=\tau_8$; then the intrinsic torsion class of the natural almost parahermitian structure on $M\times N$ is the union of the intrinsic torsion classes of $M$ and $N$.
\end{proposition}
The general case requires more notation. We say a map from the set of intrinsic torsion classes to itself is additive if it satisfies $g(I + J)=g(I) + g(J)$. Consider the additive map
\[g\colon \mathcal{W}_i\to \begin{cases} \mathcal{W}_i & i\neq 4,8\\ \mathcal{W}_3+\mathcal{W}_4 & i=4\\\mathcal{W}_7+\mathcal{W}_8 & i=8 \end{cases};\]
again by Lemma \ref{lemma:higherdim}, we find:
\begin{proposition}
\label{prop:products}
The intrinsic torsion of the natural almost parahermitian structure on a product $N\times M$ is
\[\mathcal{W}(N\times M)=g(\mathcal{W}(N)+\mathcal{W}(M)).\]
\end{proposition}
In the situation of Proposition~\ref{prop:products}, if $M$ and $N$ are Einstein with the same scalar curvature $s$, then the product $M\times N$ is also Einstein with scalar curvature $s$. In particular, the results of Section~\ref{sec:examples} imply:
\begin{proposition}
\label{prop:ricciflatproducts}
Each intrinsic torsion class involving only $\mathcal{W}_1$, $\mathcal{W}_2$, $\mathcal{W}_3$, $\mathcal{W}_5$, $\mathcal{W}_6$ and $\mathcal{W}_7$ contains compact manifolds with a Ricci-flat $\GL(n,\R)$-structure for $n\gg1$.
\end{proposition}
Unlike $\mathcal{W}_1$ and $\mathcal{W}_2$, the intrinsic torsion classes $\mathcal{W}_3$ and $\mathcal{W}_4$ are not Ricci-flat; this is indicated by the formulae of Theorem~\ref{thm:restateRicViaForms}, and can be verified through the following examples:
\begin{example}
In the class $\mathcal{W}_3$, the standard $\SL(3,\R)$-structure on the Lie algebra $(0,0,0,0,45,46)$ has intrinsic torsion
\[\tau_3 = \frac{1}{2} e_2\otimes e^{45}-\frac{1}{2} e_3\otimes e^{46};\]
the Ricci tensor is $-e^4\otimes e^4$.
\end{example}
\begin{example}
A non-Einstein example with intrinsic torsion contained in $W_4$ is the standard $\SL(3,\R)$-structure on the Lie algebra \[(-14,0,0, 0,45,46).\] In this case the Ricci curvature is $e^4\otimes e^4\in S^2V$.
\end{example}
\begin{remark}
Using this last example, together with those of Section \ref{sec:examples}, the construction of Proposition~\ref{prop:products} allows one to produce almost parahermitian structures in any given intrinsic torsion class.
\end{remark}
\bibliographystyle{plain}
|
\section{Introduction}
One of the earliest attempts to describe the nature of time and space comes from Parmenides ($\sim$ 500 BCE) \cite{Plato}. He and his pupil Zeno argued for monism---that there was only a single reality---and so to them time was a complete whole without division. They argued that this gives less absurdities than the opposing pluralistic view where multiple realities catered for different modes of being. Zeno's well-known paradoxes were attempts to illustrate the absurdities that would follow from pluralism. However, a new way of looking at nature, based on empirical observations and mathematical calculations, emerged in the European Renaissance period. The perceived difficulties associated with Zeno's paradoxes were largely swept aside with the development of calculus. Building on the work of Copernicus and Galileo, Newton proposed that an absolute time flows uniformly throughout an absolute space \cite{Novikov}. Newton's framework represents a kind of pluralism where each moment in time represents a separate reality. Then, about a century ago, James \cite{James} and McTaggart \cite{McTaggart} resurrected a monist view of time in the form of the block universe, where time is seen to be one structure without a present, past or future. The block universe represents, for the most part, the orthodoxy among physicists in modern times \cite{Zeh,Price}. Nevertheless, a new kind of pluralism will reemerge later in this chapter.
The impetus for abandoning Newton's framework of space and time in physics came from its failure to account for the propagation of light in the Michelson-Morley experiments of 1887. This anomaly led Einstein in 1905 to propose a new framework for space and time in his special theory of relativity. How we think of time and space today in terms of a background geometry is moulded by Einstein's relativistic \emph{spacetime}---an amalgamation of space and time into a single entity. An interval of either time or space for one reference frame can be an interval that extends over both time and space in another reference frame.\footnote{Appendix 1 discusses this in more detail.}
In this sense, one can say that space and time appear in special relativity on the same footing.
Yet time and space are quite different in other respects. For example, matter can be localised in a region of space but not in an interval of time. That is, a lump of matter---such as an atom, a coffee cup or even a galaxy---can exist in one region of space and no other, but conservation of mass\footnote{The terms mass and matter here can be taken to mean relativistic energy.} forbids matter from existing at one time interval and no other.
To exist only at one time interval, for one second after midday say, would mean the matter not existing before midday, existing only during the second after midday and vanishing at the end of the second. We avoid this drastic violation of mass conservation in conventional physics by insisting that matter follows an \emph{equation of motion} that translates it over all times.
The upshot is that matter is presumed to undergo continuous translation over time (as time evolution) but there is no corresponding presumption about the matter undergoing translations over space.
There is more to this---the presumed continuous translation over time occurs in a preferred direction and the direction is described by various arrows of time. The first to be named formally is the thermodynamic arrow \cite{Eddington} which points in the direction of increasing entropy. Other arrows include the cosmological arrow, which points away from the big bang, and the radiation arrow, which points in the direction of emission of waves \cite{Price}.
In contrast, space is isotropic.
There is another, quite subtle, difference between time and space that has largely escaped attention until recently: translations in time and space have very different discrete symmetry properties \cite{FPhys,FPhys2,PRSA}. The discrete symmetries here represent an invariance to the operations of charge conjugation (C), parity inversion (P) and time reversal (T). Although nature respects these symmetries in most situations, exceptions have been discovered in the last 60 years. The exceptions are observed as violations of particular combinations of the C, P and T symmetries in certain particle decays \cite{parity,CP,T1,T2,T3,T4}. The violations are independent of position in space, and so they occur over \emph{translations in time} (i.e. as a decay) and \emph{not translations in space}.
The fact that time and space have these differences does not, in itself, constitute a problem. On the surface, the differences don't appear to be pointing to a glaring anomaly that requires a reworking of the foundations of physics like the results of the Michelson-Morley experiment did. Yet there is an anomaly, one that has been around for so long that it risks being overlooked because of its familiarity. It is to do with the fact that there is no cause for the block universe to be anything other than symmetrical in time. In other words, there is no physical basis for one direction of time to be singled out \cite{Price}. This invites the question, so where is the other direction of time? It may be tempting to speculate that another part of the time axis may carry arrows pointing in the opposite direction. But this will not do, given the impact the discoveries of the violation of the discrete symmetries have for the Hamiltonian. The Hamiltonian is a mathematical object that defines the dynamics. The violation of time reversal symmetry, called T violation for short, implies that there is a different version of the Hamiltonian for each direction of time, yet we observe only one version in our universe and, not surprisingly, only the observed version of the Hamiltonian appears in conventional theories of physics. Where is the other direction of time and its concomitant version of the Hamiltonian? The fact that there is no answer in conventional physics constitutes a basic anomaly which calls for a fundamental shift in our thinking about time and space.
The purpose of this chapter is to expose the anomaly and then review my recent proposal \cite{PRSA} to resolve it through restructuring the way time and space appear in physical theory. The anomaly is articulated more precisely in \S2 and then \S 3 prepares for the required restructuring in terms of a goal and three basic principles. Following that, the principles are applied to non-relativistic quantum mechanics in \S 4 and the chapter ends with a discussion in \S 5. Additional background material and specific details are left to the appendices: special relativity in Appendix 1, generators and translations in time and space in Appendix 2, and quantum virtual paths in Appendix 3. Full details of my proposed resolution can be found in Ref.~\cite{PRSA}.
\section{An anomaly: missing direction of time and its Hamiltonian}
To expose the anomaly we must first lay to rest a common misconception that the arrows of time are melded in some way into the concept of time itself.
In particular, if the only thing that distinguishes the two directions of time is an increase of entropy in one direction, then perhaps one could be forgiven for succumbing to a conceptual shorthand and regarding the entropy increase as somehow causing the direction of time.
But, in truth, the arrows are only \emph{evidence} that time has a direction and there is simply no basis for claiming them as the \emph{cause} of that direction.
An analogy will help make the distinction between evidence and cause clearer. Imagine that the leaves falling from a tree are blown by a steady wind to land preferentially on the downwind side of the tree.
The pattern of leaves on the ground would then provide \emph{evidence} of the direction that the wind is blowing, but there would be no basis for claiming that the leaves \emph{cause} the wind to have any particular direction.
The same situation occurs with the direction of time: the arrows are patterns that provide evidence of the direction of the translations over time, but those patterns do not cause the translations themselves nor do they cause the translations to be in a particular direction.
Having laid bare the evidential nature of the arrows, we now examine the thermodynamic arrow in particular. This arrow, like all the arrows, is phenomenological in origin. It arises because thermodynamics was developed to be in accord with nature and thus it was intentionally structured to have an increasing entropy in the direction of time we refer to as ``forwards'' or the ``future''. However, as Loschmidt pointed out long ago, thermodynamics is consistent with time-symmetric physical laws, such as Newton's laws of motion, and so any prediction of an increase in entropy in one direction of time is, necessarily, a prediction of an increase in the opposite direction of time. To ignore this and claim that the thermodynamic arrow, or any of the arrows, explains the direction of time, is to commit what Price calls a double standard fallacy \cite{Price}. Avoiding the fallacy leaves us with the problem of a missing direction of time.
Its resolution calls for a time-symmetric model of nature that accounts for both directions of time---a model in which there are reasons for arrows to point in both directions. There have been admirable attempts along these lines by Carroll, Barbour and their co-workers \cite{Carroll,Barbour}, but there is something fundamental missing from their analyses because they only consider time-symmetric physical laws. The only fundamental law that is not time symmetric is usually dismissed as having little to do with large-scale effects \cite{Novikov,Price,Zeh,Carroll-book}. It is associated with the weak interaction, and its time asymmetry is observed as T violation in the decay of the K and B mesons \cite{T1,T2,T3,T4}. However, despite being previously overlooked, I have shown that T violation is capable of producing large-scale physical effects \cite{FPhys,FPhys2,PRSA}. Moreover, the experimentally observed T violation implies that the universe is described by two versions of the Hamiltonian, one for each direction of time. The double-headed arrows of Carroll, Barbour and co-workers do not account for this crucial fact.
The problem, then, is not only that there is a missing direction of time, but that the associated version of the Hamiltonian is missing along with it. The anomaly is the rather glaring absence of both directions of time and both versions of the Hamiltonian in conventional physical theories; it can be stated formally as follows.
\begin{anomaly}
There is no basis for nature to be asymmetric in time. Experiments in particle physics indicate that there are two versions of the Hamiltonian, one for each direction of time. A time-symmetric theory of nature must give an equal account of both directions of time and both versions of the Hamiltonian. Conventional theories fail in this regard because they can accommodate only one version of the Hamiltonian and one direction of time.
\end{anomaly}
The anomaly calls for a restructuring of the concepts of time and space in physics.
\section{The goal and basic principles}
The goal of the restructuring might appear to be to simply find a time-symmetric description that includes both versions of the Hamiltonians. However, aiming the goal directly at the anomaly like this misses an opportunity for rebuilding from a deeper level. For example, if Einstein had been satisfied with a description of the propagation of light that was consistent with the Michelson-Morley experiment, he may have settled on some aether-dragging model. Instead, his search for an indirect, but deeper, solution led to his special theory of relativity, a natural consequence of which was the resolution of the light-propagation anomaly. In the same way, we need to take a step back from the anomaly itself. We have seen that the differences between time and space are related by the fact that they involve translations: conservation laws and the equation of motion represent translations over time, the direction of time describes an asymmetry in translations over time, and the violation of the discrete symmetries is observed for translations over time. Our understanding of the relationship between time and space would be advanced significantly if all differences could be shown to have a common origin. The least understood among the differences is the C, P and T symmetry violations. Although the violations are generally considered to represent profound properties of nature, they don't play any significant role in conventional physics. Indeed, they stand out as having been overlooked. To address this situation, we undertake the more ambitious goal as follows:
\begin{goal}
To treat time and space on an equal footing at a fundamental level, and to allow their familiar differences to emerge phenomenologically from the discrete symmetry violations.
\end{goal}
If the violations deliver the differences between space and time then we will have found a theory that incorporates both versions of the Hamiltonian in a way that gives rise to the familiar direction of time. The anomaly would then be resolved as a natural consequence of the goal.
Having settled on the goal, we now turn to the basic principles needed to achieve it.
When the C, P and T symmetries are obeyed we want matter to be localisable both in time and space. This will require a formalism in which conservation laws do not apply and an equation of motion is not defined---this marks a serious departure from conventional physics. When the violation of the symmetries are introduced into the formalism, an effective equation of motion and conservation laws need to appear phenomenologically as a consequence---only then will it be in agreement with conventional physics. The symmetry violations clearly need to play a significant role in the formalism. The violations manifest as changes due to the C, P and T operations,\footnote{If the C, P and T operations do not change the system then the symmetries are obeyed. Violations represent the converse situation where changes result from the operations.} and so their impact would tend to be greater in a formalism in which the operations are more numerous. The P and T operations, in particular, are associated with reversing directions in space and time, respectively. It is clear from this that we need a formalism comprising paths in time and space which suffer innumerably-many reversals. A stochastic Wiener process involves paths of this kind in space. Feynman's path integral method \cite{Feynman} also involves similar kinds of paths over configuration space.
The important point about Feynman's method is that it underpins analytical mechanics in the limit that Planck's constant, $\hbar$, tends to zero. Indeed, his method shows that Hamilton's principle of least action arises as a consequence of destructive interference over all possible paths in configuration space between the initial and final points. But it stops short of considering paths that zigzag over time of the kind we need to consider here and, as a consequence, it stops short of considering the impact of the C, P and T symmetry violations that are the focus here. Nonetheless, it does demonstrate the importance of quantum path integrals for describing the universe on a large scale.
Although the paths need to comprise innumerably-many reversals, there are reasons to believe that there are physical limitations to the resolution of intervals in space and time \cite{Planck}. For example, the position of an object can be determined by observing the photons it scatters, but the accuracy of the result cannot be better than the Planck length $L_P=1.6\times 10^{-35}$~m \cite{Mead}. Correspondingly, the timing of the scattering events cannot be determined any better than the Planck time $L_T=L_P/c=5.4\times 10^{-44}$~s where $c$ is the speed of light. We will assume that fundamental resolution limits of this kind exist without specifying their value. It would be physically impossible to resolve the structure of paths with step sizes smaller than the resolution limit, and so we need to treat such paths as having equal physical status.
With these ideas in mind we formulate three principles on which to base the development of the new formalism:
\begin{principle} \label{p:qvp}
A quantum state is represented as a superposition of paths, each containing many reversals. We call these ``quantum virtual paths''.
\end{principle}
\begin{principle} \label{p:resolution}
There is a lower limit to the resolution of intervals in space and time. Quantum virtual paths with step sizes smaller that this limit have an equal physical status.
\end{principle}
\begin{principle} \label{p:phenomenological}
States have the same construction in both time and space. Any differences between space and time, such as dynamics and conservation laws, emerge phenomenologically as a result of the violation of discrete symmetries C, P and T.
\end{principle}
\section{Applying the principles}
We shall apply the three basic principles to represent the quantum state of an object.\footnote{We only use the static representation of a state from non-relativistic quantum mechanics. We do not apply an equation of motion nor do we impose conservation laws.} The object represents the only matter in space and time and it could be an atom, planet or galaxy. Its details are not important. We will refer to it as the ``galaxy'' in the following. The first task is to develop the formalism in general terms without referring specifically to time or space. For that let $w$ be a generic coordinate which will later be set to be either time or space. We want the galaxy to be localised with respect to $w$ such that the spread in $w$ is finite. The most general probability distribution with a finite spread has a bell-shape like $P(w)$ illustrated in Fig. \ref{fig:gaussian}.
\begin{figure}[t]
\sidecaption
\includegraphics[width=75mm]{Fig_Gaussian.pdf}
\caption{Bell-shaped probability distribution $P(w)$ representing an object localised in the vicinity of the origin of the $w$ coordinate. The standard deviation of the distribution is $\sigma_{\rm w}$.}\label{fig:gaussian}
\end{figure}
\subsection{Application of principle \ref{p:qvp}}
An equivalent representation is given by imagining that the galaxy takes a path that starts at the origin $w=0$ and randomly steps back and forth along the $w$ coordinate a number of times. Let there be $N$ steps in the path and let the magnitude of each step be $\delta w$.
For the final location of the galaxy to any value of $w$, the step size $\delta w$ needs to be infinitesimally small and $N$ needs to be correspondingly large. By setting
\begin{align} \label{eq:delta w}
\delta w = \frac{\sqrt{2}\sigma_{\rm w}}{\sqrt{N}}
\end{align}
and choosing a suitably-large value of $N$ we can make the step size, $\delta w$, as small as we like, and the maximum length of any path, $N\delta w$, correspondingly as large as we like, while keeping the standard deviation in the possible final locations fixed at $\sigma_{\rm w}$.
It needs to be emphasised that even though temporal references such as ``starts'', ``steps'' and ``final'' are used here, the paths do not represent actual movement over a time interval. Rather they represent the galaxy executing a sequence of \emph{virtual displacements} along $w$ without any reference to time at all.
That is, the galaxy is considered to be simply displaced from $w=0$ to the point represented by the end of the random path.
Virtual displacements arise in analytical mechanics when discussing constraints on motion \cite{Goldstein}; here the accumulation of many random virtual displacements give the possible values of $w$.
\begin{figure}[t]
\sidecaption
\includegraphics[width=75mm]{Fig_Random.pdf}
\caption{Conceptual sketch of a quantum virtual path. Each curve represents a random path of $N$ steps back and forth along the $w$ coordinate starting at $w=0$ and ending at a random value of $w$. The curves are displaced vertically to represent the relative density of paths. The inset illustrates the actions of the generators, $\hat W_F$ and $\hat W_B$, of translations in the $+w$ and $-w$ directions, respectively. }\label{fig:qvp - w - symmetry case}
\end{figure}
For the location of the galaxy to be described by the smooth bell-shaped distribution $P(w)$ we need not just one path and its end point, but infinitely many. We don't know which end point describes the location of the galaxy and so we have to allow for the possibility that it could be the end point of any one of many paths. Technically, this means we represent the location of the galaxy by a \emph{superposition} of the end points of all the paths.
The superposition is called a ``quantum virtual path'', where quantum refers to the fact that it is a quantum superposition \cite{PRSA}.
One can imagine a quantum virtual path for a specific value of $N$, say $N=600$, as the sum of the end points of all the paths illustrated in Fig. \ref{fig:qvp - w - symmetry case}. The step size $\delta w$ for each zigzag path in the figure is given by \eq{eq:delta w} for some fixed value of the standard deviation $\sigma_{\rm w}$. Another quantum virtual path can be constructed for $N=601$ in a similar way for a correspondingly smaller step size $\delta w$. Imagine that this has been done for every positive integer value of $N$.
As $N$ increases in this imagined process, the step size $\delta w$ reduces and the quantum virtual path represents an ever finer description of the state of the galaxy, eventually tending to the bell-shaped dashed curve shown in the figure. Each quantum virtual path so constructed represents a possible state of the galaxy in terms of its location along the $w$ coordinate.
Each step of $\delta w$ is produced using a particular operation called a ``generator'' of the translation. In particular, $\hat W_F$ is the generator for translations that increase the value of $w$ and $\hat W_B$ is the generator for ones that decrease its value, as illustrated in the inset of Fig.~\ref{fig:qvp - w - symmetry case}. If the generators are invariant to reversals of direction then they are equivalent, i.e. $\hat W_F=\hat W_B$. More will be said about this later.
A technical review of generators and translations is given in Appendix 2 and a brief discussion of how a quantum virtual path is related to the bell-shaped distribution $P(w)$ can be found in Appendix 3.
\subsection{Application of principle \ref{p:resolution}}
As the value of $N$ increases, the step size $\delta w$ from \eq{eq:delta w} becomes smaller. At some point $\delta w$ will be smaller than the resolution limit $\delta w_{\rm min}$ for the $w$ coordinate.
All quantum virtual paths with a step size $\delta w$ smaller than $\delta w_{\rm min}$ will give descriptions of equal status according to Principle \ref{p:resolution}.
For convenience, we shall collect the equivalent quantum virtual paths in a set called $\mathbf{G}$. Each quantum virtual path in this set equally represents the state of the galaxy in terms of its location along the $w$ coordinate. There are an infinite number of such quantum virtual paths in the set $\mathbf{G}$.
\subsection{Application of principle \ref{p:phenomenological}}
We now discuss space and time explicitly. First consider the spatial case which, for brevity, we limit to just the $x$ dimension. In this case the generic coordinate $w$ is replaced with $x$ and the generator of translations is replaced with $\hat p_x$, the component of momentum along the $x$ axis. There is only one generator for translations in both directions of the $x$ axis and so $\hat W_F=\hat W_B=\hat p_x$ here. Further technical details are given in Appendix 2. Fig.~\ref{fig:qvp - w - symmetry case} with $w$ replaced by $x$ illustrates a quantum virtual path over the $x$ axis. Collecting the quantum virtual paths with a step size smaller than some minimum resolution limit yields the set of states of equal status which we will call ${\bm{\upPsi}}$. All the quantum virtual paths in $\bm{\upPsi}$ are physically indistinguishable from the bell-shaped distribution $P(x)$ represented in Fig.~\ref{fig:gaussian} with $w$ replaced with $x$.
Next, we repeat the same exercise for time. In this case the coordinate is $w=t$ and, in general, there are two generators of translations given by the two versions of the Hamiltonian, i.e. $\hat W_F=\hat H_F$ and $\hat W_B=\hat H_B$ corresponding to the ``forwards'' and ``backwards'' directions of time, respectively. Technical details regarding these generators are given in Appendix 2. As with the spatial case, Fig.~\ref{fig:qvp - w - symmetry case} with $w$ replaced by $t$ illustrates a quantum virtual path over the $t$ axis, and
collecting the quantum virtual paths which have a step size smaller than some minimum resolution limit yields the set of states of equal status which we will call ${\bm{\upUpsilon}}$.
In a universe where the T symmetry holds, there is only one version of the Hamiltonian and so $\hat H_B=\hat H_F=\hat H$. In this case the galaxy is localised in time within a duration of the order of $\sigma_{\rm t}$ of the origin and all the states in $\bm{\upUpsilon}$ are physically indistinguishable from the bell-shaped distribution $P(t)$ represented in Fig.~\ref{fig:gaussian} with $w$ replaced with $t$.
The galaxy only exists in time for a relatively short duration at the origin $t=0$ and does not exist before or after this time. It can be imagined to come into existance momentarily and then promptly vanish.
Clearly, in this case, the galaxy has the same representation in time as in space---it is localised in both---and the formalism places time and space on the same footing in this respect. This is far removed from conventional quantum mechanics as there is no equation of motion and the mass of the galaxy is not conserved.
The converse case, where T symmetry is violated, is defined by $\hat H_B\ne\hat H_F$. The key point here is that multiple paths that zigzag in different ways from the origin to the same end point can interfere. The interference can be compared to the way waves travelling on the surface of water behave; if the trough of one wave occurs at the same point as the crest of another, the two waves will tend to cancel each other in a process called destructive interference, whereas if two troughs or two crests meet they tend to reinforce each other as deeper troughs or higher crests, respectively, in a process called constructive interference. In a similar way, multiple paths that end at the same point on the time axis interfere either destructively or constructively. The result is that instead of the probability distribution having a maximum at the origin, like the bell-shaped curve in Fig.~\ref{fig:gaussian}, destructive interference reduces the probability to zero in this region. This is compensated by constructive interference that yields two symmetrically positioned bell-shaped peaks further from the origin as illustrated in Fig.~\ref{fig:qvp - t - violation}.
In other words, each quantum virtual path is now composed of two bell-shaped peaks that represent the galaxy existing at two different times, $+t$ and $-t$, say. This situation is like Schr\"{o}dinger's cat that exists in a superposition of being both dead and alive simultaneously, except that here the galaxy is at two different times. As the value of $N$ increases, the two peaks become further separated as shown in Fig.~\ref{fig:qvp - t - violation}, and the galaxy shifts in time accordingly. Each quantum virtual path represents one of the states in the set ${\bm{\upUpsilon}}$, and according to Principle \ref{p:resolution}, has equal physical status. In terms of Fig.~\ref{fig:qvp - t - violation}, this means that each double-peaked curve equally represent the position of the galaxy in time.
The presence of T violation clearly has a dramatic affect on the temporal description of the galaxy. For example, consider the question, where in time is the galaxy likely to be found? Without T violation, the unequivocal answer is only near the origin in accordance with Fig.~\ref{fig:gaussian}, whereas with T violation, the answer implied by Fig.~\ref{fig:qvp - t - violation} would be at \emph{any} time $t$.
\begin{figure}[t]
\sidecaption
\includegraphics[width=75mm]{Fig_TV.pdf}
\caption{The probability distribution $P(t)$ for various values of $N$ in the case of T violation. Each curve has two bell-shaped peaks which move further apart as $N$ increases. }\label{fig:qvp - t - violation}
\end{figure}
\subsection{The origin of dynamics}
We now focus on the T violation case. According to Principle \ref{p:resolution}, all states in the set ${\bm{\upUpsilon}}$ have an equal status in representing the galaxy in time. For any given value of time, $t$, there is a corresponding state in ${\bm{\upUpsilon}}$ that represents the galaxy being in a superposition of the times $+t$ and $-t$.\footnote{In principle, the time $t$ could be chosen to be the current age of the universe, $13.8$ billion years. There is a state in ${\bm{\upUpsilon}}$ that represents the galaxy being in a superposition of the times 13.8 billion years and $-13.8$ billion years.} This implies that the galaxy exists at any time we wish to consider, and so its \emph{mass is conserved}. This conservation law has not been imposed on the formalism, as it would need to be in conventional theories, but rather it is phenomenology arising from T violation.
The corresponding equation of motion is found as follows. The two peaks $+t$ and $-t$ in each curve in Fig.~\ref{fig:qvp - t - violation} represent time-reversed versions of the galaxy.
An observer in the galaxy would not be able to distinguish between them and so we need only consider one, at $+t$ say. If the observer makes observations with a resolution in time that is broader than the width of the peak, the peak will appear to be instantaneous and a set of them will appear to form a continuous sequence. Under these circumstances, the observer would find evidence of an \emph{equation of motion} that is consistent with the Schr\"{o}dinger equation of conventional quantum mechanics. This equation has not been imposed on the formalism but rather it arises as phenomenology associated with T violation. This suggests that the origin of dynamics lies in T violation.
The remaining distinctive feature of time to consider is its direction, and the states in ${\bm{\upUpsilon}}$ have a time ordering in the following sense. According to the meaning of time evolution defined in Appendix 2, the peak labelled ``d'' Fig.~\ref{fig:qvp - t - violation} represents a state that has \emph{evolved in time} from the state represented by the peak labelled ``c'', and that state has evolved from the state represented by ``b'', which has evolved from the state represented by ``a'', but the converse is not true. This means that there is an arrow of time pointing in the direction of $+t$. The same argument applies to the time reversed states in regards to the $-t$ direction and so the \emph{arrow is double headed}, like those of Carroll and Barbour and co-workers \cite{Carroll,Barbour}. The important point here is that both versions of the Hamiltonian, $\hat H_F$ and $\hat H_B$, are included in the formalism.
We have now achieved our goal: we treated time and space on an equal footing and found their familiar differences to emerge phenomenologically from T violation.
\section{Discussion}
We began by identifying a fundamental anomaly in physics, viz. conventional theories fail to give a time symmetric description that accounts equally for both versions of the Hamiltonian and both directions of time. We have presented a new formalism for quantum mechanics that resolves this anomaly. The new formalism is based on three principles that allow quantum states in time and space to be treated on an equal footing in terms of quantum virtual paths. The distinctive features associated with time, i.e. conservation laws, equation of motion and the direction of time, are not imposed on the formalism but rather emerge phenomenologically as a result of T violation. These key differences between time and space follow from the fact that the generators of translations in space and time, the momentum operator and the Hamiltonian, respectively, have different symmetry properties: the momentum operator is invariant to the C, P and T symmetry operations whereas the Hamiltonian is not. Accounting for these differences gives the \emph{origin of dynamics}.
The new formalism also refines the meaning of time. In conventional theories, the word ``time'' refers to both a coordinate of a space-time \emph{background} as well as the parameter describing \emph{dynamical evolution}. Both concepts are firmly entwined by conservation laws. For example, the conservation of mass implies that a massive object will persist over all times and, accordingly, it is represented on a space-time background as existing at each time. The dynamical evolution of the object becomes the path of the object on the space-time background. Here, however, the two concepts of time as a background coordinate and as a dynamical parameter are distinct. Time and space have an equal footing as a background on which quantum states are represented. The states, as quantum virtual paths, represent objects that are localised in time and space: each state in the sets ${\bm{\upPsi}}$ and ${\bm{\upUpsilon}}$ represents a relatively-narrow bell-shaped distribution or a sum of two relatively-narrow bell-shaped distributions. In particular, mass is not conserved and there is no equation of motion for any \emph{individual state} in ${\bm{\upUpsilon}}$ (as illustrated by Fig.~\ref{fig:qvp - t - violation})---time appears only as a background coordinate. In contrast, mass conservation, the equation of motion and the direction of time, are properties of the\emph{ whole set} ${\bm{\upUpsilon}}$ where time appears as a dynamical parameter. In other words, time as a background coordinate and as a dynamical parameter apply to distinct constructs in the formalism.
It might appear unusual that a quantum formalism is being proposed to explain large scale structure of nature given that quantum effects are typically seen only in relatively small systems under controlled conditions. However, Feynman's path integral method has already demonstrated how quantum phenomena underpins Hamilton's least action principle in analytical mechanics \cite{Feynman} and thus large scale structure. In this regard, the new formalism should be considered as an extension of Feynman's method to encompass paths over time and the C, P and T symmetry violations and, thus, to apply to nature on a large scale as well.
Finally, the set of states ${\bm{\upUpsilon}}$ for T violation represents the galaxy at an infinite sequence of times. Each state in ${\bm{\upUpsilon}}$ may be viewed as representing a different reality. In this sense, the formalism resurrects a kind of pluralism. The monism-pluralism cycle for time turns once more.
\section*{Appendix 1}
\addcontentsline{toc}{section}{Appendix 1}
We briefly review here the Lorentz transformation in special relativity. A point in spacetime is referred to as an event; it is specified by four coordinates $x$, $y$, $z$, $t$ with respect to a reference frame.
The Lorentz transformation gives the relationship between the coordinates of two different inertial reference frames. In particular, consider two events that occur a distance of $\Delta x$ apart along the $x$ axis and separated by a duration of $\Delta t$ in time in the $x$, $y$, $z$, $t$ reference frame.
In the $x'$, $y'$, $z'$, $t'$ reference frame that is moving a constant speed $v$ along the $x$ axis of the first, the distance and duration along the $x'$ and $t'$ axes between the events are given by
\begin{align*}
\Delta x' &=\gamma(\Delta x-v\Delta t)\\
\Delta t' & =\gamma(\Delta t-v\Delta x/c^2)\ ,
\end{align*}
respectively, where $\gamma =\sqrt{1-v^2/c^2}$ and $c$ is the speed of light.
The important point here is that what is considered to be solely a spatial interval, $\Delta x$, in one reference frame becomes part of a temporal interval $\Delta t'$ as well as being part of a spatial interval $\Delta x'$ in the other reference frame. That is, space and time are interchangeable.
\section*{Appendix 2}
\addcontentsline{toc}{section}{Appendix 2}
Here, we briefly review translations and their generators. Recall that the Taylor expansion of a function $f(x)$,
\[
f(x+a) = f(x) + a\frac{\D}{\D x}f(x) + \frac{a^2}{2!} \frac{\D^2}{\D x^2}f(x) + \frac{a^3}{3!} \frac{\D^3}{\D x^3}f(x) + \ldots\ ,
\]
can be written compactly in exponential form as
\begin{align*}
f(x+a) = \E^{-\I a(\I\frac{\D}{\D x})}f(x)\ .
\end{align*}
When written in this form the differential operator $\I\frac{\D}{\D x}$ is said to be the generator of translations in $x$. The generator of spatial translations along the $x$ axis is $\hat p_x$, the operator representing the $x$ component of momentum. We need only consider one dimension of space for our purposes here. Thus we write
\begin{align} \label{eq:x - translate by a}
\ket{x+a}_{\rm x}=\E^{-\I a\hat p_x}\ket{x}_{\rm x}
\end{align}
where $\ket{x}_{\rm x}$ represents a state vector for position $x$ and, for convenience, we assume units in which $\hbar=1$. Similarly, the generator of translations in time $t$ is the Hamiltonian operator $\hat H$ and so
\begin{align} \label{eq:t - translate by a}
\ket{\psi(t+a)}_{\rm t}=\E^{-\I a\hat H}\ket{\psi(t)}_{\rm t}
\end{align}
where $\ket{\psi(t)}_{\rm t}$ represents a state at time $t$ and evolving in the $+t$ time direction.
The symmetry operations relevant to these translations are the parity inversion $\hat P$ and the time reversal $\hat T$ operations\footnote{We use the operator symbols $\hat P$ and $\hat T$ to represent the operations and the letters P and T to represent the corresponding symmetries. Thus, if the system is invariant to the $\hat P$ operation it obeys the P symmetry.} defined by Wigner \cite{Wigner}. Parity inversion interchanges $x$ with $-x$, $y$ with $-y$ and $z$ with $-z$. For example, $\hat P\ket{x}_{\rm x}=\ket{-x}_{\rm x}$ and $\hat T\ket{t}_{\rm t}=\ket{-t}_{\rm t}$. The reverse of the translation in \eq{eq:x - translate by a} can be written as
\begin{align*}
\ket{x-a}_{\rm x}&=\hat P\ket{-x+a}_{\rm x}=\hat P\E^{-\I a\hat p_x}\ket{-x}_{\rm x}=\hat P\E^{-\I a\hat p_x}\hat P^{-1}\ket{x}_{\rm x}\ .
\end{align*}
As $\hat P\hat p_x\hat P^{-1}=-\hat p_x$ we get
\begin{align*}
\ket{x-a}_{\rm x}=\E^{\I a\hat p_x}\ket{x}_{\rm x}
\end{align*}
as expected directly from \eq{eq:x - translate by a}. This shows that the generator of translations in either direction of the $x$ axis is the same. The reverse of the translation in \eq{eq:t - translate by a} is somewhat different, however. Consider
\begin{align}
\ket{\phi(t-a)}_{\rm t}&=\hat T\ket{\phi(-t+a)}_{\rm t}
=\hat T\E^{-\I a\hat H}\ket{\phi(-t)}_{\rm t}
=\hat T\E^{-\I a\hat H}\hat T^{-1}\ket{\phi(t)}_{\rm t}\nonumber\\
&=\E^{\I a\hat T\hat H\hat T^{-1}}\ket{\phi(t)}_{\rm t} \label{eq:t - translate by -a}
\end{align}
where $\ket{\phi(t)}_{\rm t}$ represents a state that evolves in the $-t$ direction and we have made use of the antiunitary nature of the time reversal operator, i.e. $\hat T\I\hat T^{-1}=-\I$, in the last line \cite{Wigner}. In general $\hat T\hat H\hat T^{-1}\ne\hat H$ and so we set, for convenience,
\begin{align*}
\hat H_B&=\hat T\hat H\hat T^{-1}\\
\hat H_F&=\hat H
\end{align*}
where the subscripts $F$ and $B$ refer to the ``forwards'' and ``backwards'' direction of time corresponding to the $+t$ and $-t$ time directions, respectively. If T symmetry is obeyed then
\begin{align*}
\hat H_B=\hat H_F=\hat H\ , \qquad\mbox{(T symmetry)}
\end{align*}
and so there is a unique version of the Hamiltonian, whereas for T violation there is a different version of the Hamiltonian for each direction of time,
\begin{align*}
\hat H_B\ne\hat H_F\ .\qquad\mbox{(T violation)}
\end{align*}
In general, we write \eq{eq:t - translate by a} and \eq{eq:t - translate by -a} as
\begin{align*}
\ket{\psi(t+a)}_{\rm t}&=\E^{-\I a\hat H_F}\ket{\psi(t)}_{\rm t}\\
\ket{\phi(t-a)}_{\rm t}&=\E^{\I a\hat H_B}\ket{\phi(t)}_{\rm t}\ .
\end{align*}
The key point to be made here is that the generator of translations in space, $\hat p_x$, is invariant (up to a sign change) under any of the C, P and T operations. In contrast, the generator of translations in time, $\hat H$, is not invariant to the C, P and T operations, in general. This underlies the statement in the Introduction that the symmetry violations occur over translations in time and not translations in space.
In the case of T violation we need to take care with using the correct Hamiltonian associated with each direction of time. In particular, we need to apply the following principle:
\begin{principle} \label{p:time evolution}
Physical time evolution is represented by the operators $\E^{-\I a\hat H_F}$ and $\E^{\I a\hat H_B}$ for the forward ($+t$) and backward ($-t$) directions of time, respectively. The operations $\E^{\I a\hat H_F}$ and $\E^{-\I a\hat H_B}$ represent the mathematical inverse operation of ``unwinding'' or ``backtracking'' the evolution produced by $\E^{-\I a\hat H_F}$ and $\E^{\I a\hat H_B}$, respectively.
\end{principle}
For example, $\E^{\I a\hat H_F}\ket{\psi(t+a)}_{\rm t}=\ket{\psi(t)}_{\rm t}$ represents unwinding the time evolution $\E^{-\I a\hat H_F}\ket{\psi(t)}_{\rm t}=\ket{\psi(t+a)}_{\rm t}$
whereas $\E^{\I a\hat H_B}\ket{\psi(t+a)}_{\rm t}$, which is not equal to $\ket{\psi(t)}_{\rm t}$ in general, represents time evolution of $\ket{\psi(t+a)}_{\rm t}$ in the $-t$ direction. More details are given in Ref. \cite{PRSA}.
\section*{Appendix 3}
\addcontentsline{toc}{section}{Appendix 3}
In this Appendix we briefly discuss the mathematical construction of quantum virtual paths for the generic coordinate $w$. Let the generators of translations be given by $\hat W_F$ and $\hat W_B$ for the $+w$ and $-w$ directions, respectively.
A quantum virtual path of the kind we want is given by \cite{PRSA}
\begin{align} \label{eq:qvp}
\ket{g}_N \propto
\frac{1}{2^N}\left(\E^{\I\hat W_B\delta w}\!+\!\E^{-\I\hat W_F\delta w}\right)^N\ket{0}_{\rm w}
\end{align}
where $\delta w$ is given by \eq{eq:delta w} and represents an increment in $w$ and $\ket{w}_{\rm w}$ represents a state for which $w$ is well-defined.\footnote{If $w$ represents a spatial coordinate then $\ket{w}_{\rm w}$ would be a corresponding spatial eigenstate. For the case where $w$ represents the time coordinate, however, we only need $\ket{w}_{\rm w}$ to represent a well-defined time. More details can be found in Ref.~\cite{PRSA}. } Expanding the power on the right side gives $2^N$ terms each with $N$ factors. Each term represents a path comprising $N$ steps of $\delta w$ over the $w$ coordinate. For example, a term of the form
\begin{align*}
\ldots\E^{\I\hat W_B\delta w}\E^{-\I\hat W_F\delta w} \E^{\I\hat W_B\delta w}\E^{\I\hat W_B\delta w}\E^{-\I\hat W_F\delta w}\ket{0}_{\rm w}
\end{align*}
represents the object starting at the origin $w=0$ and then undergoing virtual displacements to $w=\delta w$, $w=0$, $w=-\delta w$, $w=0$ $w=-\delta w$ and so on.
It is relatively straightforward to show that the state $ \ket{g}_N$ in \eq{eq:qvp} approaches a Gaussian state in the limit of large $N$ when the discrete symmetry holds. To see this set $\hat W_B=\hat W_F=\hat W$ and use
\begin{align*}
\exp(-A^2/2)=\lim_{N\to\infty}\cos^N(A/\sqrt{N})\
\end{align*}
to find
\begin{align*}
\lim_{N\to\infty}\ket{g}_N \propto \E^{-\hat W^2\sigma^2_{\rm w}}\ket{0}_{\rm w}\ ,
\end{align*}
and then, assuming that $\hat W$ has a complete orthonormal basis, rewrite this as the Fourier integral
\begin{align*}
\lim_{N\to\infty}\ket{g}_N & \propto\int \D w\ \E^{-w^2/4\sigma^2_{\rm w}} \E^{-\I \hat{W}w} \ket{0}_{\rm w}\\
&= \int g(w)\ket{w}_{\rm w} \D w
\end{align*}
where $g(w)$ is given by
\begin{align*}
g (w) = \E^{-w^2/4\sigma^2_{\rm w}}\ .
\end{align*}
The square of this, $g^2(w)$, is proportional to the bell-shaped probability distribution $P(w)$ represented in Fig.~\ref{fig:gaussian}.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.